Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * INET An implementation of the TCP/IP protocol suite for the LINUX |
| 3 | * operating system. INET is implemented using the BSD Socket |
| 4 | * interface as the means of communication with the user level. |
| 5 | * |
| 6 | * Implementation of the Transmission Control Protocol(TCP). |
| 7 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 12 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 16 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 18 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. |
| 23 | * : Fragmentation on mtu decrease |
| 24 | * : Segment collapse on retransmit |
| 25 | * : AF independence |
| 26 | * |
| 27 | * Linus Torvalds : send_delayed_ack |
| 28 | * David S. Miller : Charge memory using the right skb |
| 29 | * during syn/ack processing. |
| 30 | * David S. Miller : Output engine completely rewritten. |
| 31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. |
| 32 | * Cacophonix Gaul : draft-minshall-nagle-01 |
| 33 | * J Hadi Salim : ECN support |
| 34 | * |
| 35 | */ |
| 36 | |
| 37 | #include <net/tcp.h> |
| 38 | |
| 39 | #include <linux/compiler.h> |
| 40 | #include <linux/module.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 41 | |
| 42 | /* 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] | 43 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 45 | /* People can turn this on to work with those rare, broken TCPs that |
| 46 | * interpret the window field as a signed quantity. |
| 47 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 48 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 49 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 50 | /* This limits the percentage of the congestion window which we |
| 51 | * will allow a single TSO frame to consume. Building TSO frames |
| 52 | * which are too large can cause TCP streams to be bursty. |
| 53 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 54 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 55 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 56 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
| 57 | int sysctl_tcp_base_mss __read_mostly = 512; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 58 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 59 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 60 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 61 | |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 62 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 63 | { |
| 64 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 65 | unsigned int prior_packets = tp->packets_out; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 66 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 67 | tcp_advance_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 68 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 69 | |
| 70 | /* Don't override Nagle indefinately with F-RTO */ |
| 71 | if (tp->frto_counter == 2) |
| 72 | tp->frto_counter = 3; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 73 | |
| 74 | tp->packets_out += tcp_skb_pcount(skb); |
| 75 | if (!prior_packets) |
| 76 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 77 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 78 | } |
| 79 | |
| 80 | /* SND.NXT, if window was not shrunk. |
| 81 | * If window has been shrunk, what should we make? It is not clear at all. |
| 82 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 83 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 84 | * invalid. OK, let's make this for now: |
| 85 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 86 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 87 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 88 | struct tcp_sock *tp = tcp_sk(sk); |
| 89 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 90 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 91 | return tp->snd_nxt; |
| 92 | else |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 93 | return tcp_wnd_end(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 94 | } |
| 95 | |
| 96 | /* Calculate mss to advertise in SYN segment. |
| 97 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 98 | * |
| 99 | * 1. It is independent of path mtu. |
| 100 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 101 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 102 | * attached devices, because some buggy hosts are confused by |
| 103 | * large MSS. |
| 104 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 105 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 106 | * This may be overridden via information stored in routing table. |
| 107 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 108 | * probably even Jumbo". |
| 109 | */ |
| 110 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 111 | { |
| 112 | struct tcp_sock *tp = tcp_sk(sk); |
| 113 | struct dst_entry *dst = __sk_dst_get(sk); |
| 114 | int mss = tp->advmss; |
| 115 | |
| 116 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { |
| 117 | mss = dst_metric(dst, RTAX_ADVMSS); |
| 118 | tp->advmss = mss; |
| 119 | } |
| 120 | |
| 121 | return (__u16)mss; |
| 122 | } |
| 123 | |
| 124 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
| 125 | * This is the first part of cwnd validation mechanism. */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 126 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 127 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 128 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | s32 delta = tcp_time_stamp - tp->lsndtime; |
| 130 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); |
| 131 | u32 cwnd = tp->snd_cwnd; |
| 132 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 133 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 134 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 135 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 136 | restart_cwnd = min(restart_cwnd, cwnd); |
| 137 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 138 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 139 | cwnd >>= 1; |
| 140 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 141 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 142 | tp->snd_cwnd_used = 0; |
| 143 | } |
| 144 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 145 | static void tcp_event_data_sent(struct tcp_sock *tp, |
| 146 | struct sk_buff *skb, struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 147 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 148 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 149 | const u32 now = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 150 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 151 | if (sysctl_tcp_slow_start_after_idle && |
| 152 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 153 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 154 | |
| 155 | tp->lsndtime = now; |
| 156 | |
| 157 | /* If it is a reply for ato after last received |
| 158 | * packet, enter pingpong mode. |
| 159 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 160 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
| 161 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | } |
| 163 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 164 | 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] | 165 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 166 | tcp_dec_quickack_mode(sk, pkts); |
| 167 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 168 | } |
| 169 | |
| 170 | /* Determine a window scaling and initial window to offer. |
| 171 | * Based on the assumption that the given amount of space |
| 172 | * will be offered. Store the results in the tp structure. |
| 173 | * NOTE: for smooth operation initial space offering should |
| 174 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 175 | * This MUST be enforced by all callers. |
| 176 | */ |
| 177 | void tcp_select_initial_window(int __space, __u32 mss, |
| 178 | __u32 *rcv_wnd, __u32 *window_clamp, |
| 179 | int wscale_ok, __u8 *rcv_wscale) |
| 180 | { |
| 181 | unsigned int space = (__space < 0 ? 0 : __space); |
| 182 | |
| 183 | /* If no clamp set the clamp to the max possible scaled window */ |
| 184 | if (*window_clamp == 0) |
| 185 | (*window_clamp) = (65535 << 14); |
| 186 | space = min(*window_clamp, space); |
| 187 | |
| 188 | /* Quantize space offering to a multiple of mss if possible. */ |
| 189 | if (space > mss) |
| 190 | space = (space / mss) * mss; |
| 191 | |
| 192 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 193 | * will break some buggy TCP stacks. If the admin tells us |
| 194 | * it is likely we could be speaking with such a buggy stack |
| 195 | * we will truncate our initial window offering to 32K-1 |
| 196 | * unless the remote has sent us a window scaling option, |
| 197 | * which we interpret as a sign the remote TCP is not |
| 198 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 199 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 200 | if (sysctl_tcp_workaround_signed_windows) |
| 201 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 202 | else |
| 203 | (*rcv_wnd) = space; |
| 204 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 205 | (*rcv_wscale) = 0; |
| 206 | if (wscale_ok) { |
| 207 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 208 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 209 | */ |
| 210 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 211 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 212 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 213 | space >>= 1; |
| 214 | (*rcv_wscale)++; |
| 215 | } |
| 216 | } |
| 217 | |
| 218 | /* Set initial window to value enough for senders, |
David S. Miller | 6b25185 | 2005-09-28 16:31:48 -0700 | [diff] [blame] | 219 | * following RFC2414. Senders, not following this RFC, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 220 | * will be satisfied with 2. |
| 221 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 222 | if (mss > (1 << *rcv_wscale)) { |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 223 | int init_cwnd = 4; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 224 | if (mss > 1460 * 3) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 225 | init_cwnd = 2; |
David S. Miller | 01ff367 | 2005-09-29 17:07:20 -0700 | [diff] [blame] | 226 | else if (mss > 1460) |
| 227 | init_cwnd = 3; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 228 | if (*rcv_wnd > init_cwnd * mss) |
| 229 | *rcv_wnd = init_cwnd * mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 230 | } |
| 231 | |
| 232 | /* Set the clamp no higher than max representable value */ |
| 233 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 234 | } |
| 235 | |
| 236 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 237 | * socket, and return result with RFC1323 scaling applied. The return |
| 238 | * value can be stuffed directly into th->window for an outgoing |
| 239 | * frame. |
| 240 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 241 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 242 | { |
| 243 | struct tcp_sock *tp = tcp_sk(sk); |
| 244 | u32 cur_win = tcp_receive_window(tp); |
| 245 | u32 new_win = __tcp_select_window(sk); |
| 246 | |
| 247 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 248 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 249 | /* Danger Will Robinson! |
| 250 | * Don't update rcv_wup/rcv_wnd here or else |
| 251 | * we will not be able to advertise a zero |
| 252 | * window in time. --DaveM |
| 253 | * |
| 254 | * Relax Will Robinson. |
| 255 | */ |
Patrick McHardy | 607bfbf | 2008-03-20 16:11:27 -0700 | [diff] [blame] | 256 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 257 | } |
| 258 | tp->rcv_wnd = new_win; |
| 259 | tp->rcv_wup = tp->rcv_nxt; |
| 260 | |
| 261 | /* Make sure we do not exceed the maximum possible |
| 262 | * scaled window. |
| 263 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 264 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 265 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 266 | else |
| 267 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 268 | |
| 269 | /* RFC1323 scaling applied */ |
| 270 | new_win >>= tp->rx_opt.rcv_wscale; |
| 271 | |
| 272 | /* If we advertise zero window, disable fast path. */ |
| 273 | if (new_win == 0) |
| 274 | tp->pred_flags = 0; |
| 275 | |
| 276 | return new_win; |
| 277 | } |
| 278 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 279 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 280 | { |
| 281 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 282 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 283 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; |
| 284 | } |
| 285 | |
| 286 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
| 287 | { |
| 288 | struct tcp_sock *tp = tcp_sk(sk); |
| 289 | |
| 290 | tp->ecn_flags = 0; |
| 291 | if (sysctl_tcp_ecn) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 292 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 293 | tp->ecn_flags = TCP_ECN_OK; |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | static __inline__ void |
| 298 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) |
| 299 | { |
| 300 | if (inet_rsk(req)->ecn_ok) |
| 301 | th->ece = 1; |
| 302 | } |
| 303 | |
| 304 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
| 305 | int tcp_header_len) |
| 306 | { |
| 307 | struct tcp_sock *tp = tcp_sk(sk); |
| 308 | |
| 309 | if (tp->ecn_flags & TCP_ECN_OK) { |
| 310 | /* Not-retransmitted data segment: set ECT and inject CWR. */ |
| 311 | if (skb->len != tcp_header_len && |
| 312 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { |
| 313 | INET_ECN_xmit(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 314 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 315 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
| 316 | tcp_hdr(skb)->cwr = 1; |
| 317 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
| 318 | } |
| 319 | } else { |
| 320 | /* ACK or retransmitted segment: clear ECT|CE */ |
| 321 | INET_ECN_dontxmit(sk); |
| 322 | } |
| 323 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) |
| 324 | tcp_hdr(skb)->ece = 1; |
| 325 | } |
| 326 | } |
| 327 | |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 328 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
| 329 | * auto increment end seqno. |
| 330 | */ |
| 331 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) |
| 332 | { |
| 333 | skb->csum = 0; |
| 334 | |
| 335 | TCP_SKB_CB(skb)->flags = flags; |
| 336 | TCP_SKB_CB(skb)->sacked = 0; |
| 337 | |
| 338 | skb_shinfo(skb)->gso_segs = 1; |
| 339 | skb_shinfo(skb)->gso_size = 0; |
| 340 | skb_shinfo(skb)->gso_type = 0; |
| 341 | |
| 342 | TCP_SKB_CB(skb)->seq = seq; |
| 343 | if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) |
| 344 | seq++; |
| 345 | TCP_SKB_CB(skb)->end_seq = seq; |
| 346 | } |
| 347 | |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 348 | 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] | 349 | __u32 tstamp, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 350 | { |
| 351 | if (tp->rx_opt.tstamp_ok) { |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 352 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 353 | (TCPOPT_NOP << 16) | |
| 354 | (TCPOPT_TIMESTAMP << 8) | |
| 355 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 356 | *ptr++ = htonl(tstamp); |
| 357 | *ptr++ = htonl(tp->rx_opt.ts_recent); |
| 358 | } |
| 359 | if (tp->rx_opt.eff_sacks) { |
| 360 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; |
| 361 | int this_sack; |
| 362 | |
| 363 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 364 | (TCPOPT_NOP << 16) | |
| 365 | (TCPOPT_SACK << 8) | |
| 366 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * |
| 367 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 368 | |
| 369 | 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] | 370 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 371 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 372 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 373 | |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 374 | if (tp->rx_opt.dsack) { |
| 375 | tp->rx_opt.dsack = 0; |
| 376 | tp->rx_opt.eff_sacks--; |
| 377 | } |
| 378 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 379 | #ifdef CONFIG_TCP_MD5SIG |
| 380 | if (md5_hash) { |
| 381 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 382 | (TCPOPT_NOP << 16) | |
| 383 | (TCPOPT_MD5SIG << 8) | |
| 384 | TCPOLEN_MD5SIG); |
| 385 | *md5_hash = (__u8 *)ptr; |
| 386 | } |
| 387 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 388 | } |
| 389 | |
| 390 | /* Construct a tcp options header for a SYN or SYN_ACK packet. |
| 391 | * If this is every changed make sure to change the definition of |
| 392 | * MAX_SYN_SIZE to match the new maximum number of options that you |
| 393 | * can generate. |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 394 | * |
| 395 | * Note - that with the RFC2385 TCP option, we make room for the |
| 396 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the |
| 397 | * location to be filled is passed back up. |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 398 | */ |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 399 | 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] | 400 | int offer_wscale, int wscale, __u32 tstamp, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 401 | __u32 ts_recent, __u8 **md5_hash) |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 402 | { |
| 403 | /* We always get an MSS option. |
| 404 | * The option bytes which will be seen in normal data |
| 405 | * packets should timestamps be used, must be in the MSS |
| 406 | * advertised. But we subtract them from tp->mss_cache so |
| 407 | * that calculations in tcp_sendmsg are simpler etc. |
| 408 | * So account for this fact here if necessary. If we |
| 409 | * don't do this correctly, as a receiver we won't |
| 410 | * recognize data packets as being full sized when we |
| 411 | * should, and thus we won't abide by the delayed ACK |
| 412 | * rules correctly. |
| 413 | * SACKs don't matter, we never delay an ACK when we |
| 414 | * have any of those going out. |
| 415 | */ |
| 416 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); |
| 417 | if (ts) { |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 418 | if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 419 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 420 | (TCPOLEN_SACK_PERM << 16) | |
| 421 | (TCPOPT_TIMESTAMP << 8) | |
| 422 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 423 | else |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 424 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 425 | (TCPOPT_NOP << 16) | |
| 426 | (TCPOPT_TIMESTAMP << 8) | |
| 427 | TCPOLEN_TIMESTAMP); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 428 | *ptr++ = htonl(tstamp); /* TSVAL */ |
| 429 | *ptr++ = htonl(ts_recent); /* TSECR */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 430 | } else if (sack) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 431 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 432 | (TCPOPT_NOP << 16) | |
| 433 | (TCPOPT_SACK_PERM << 8) | |
| 434 | TCPOLEN_SACK_PERM); |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 435 | if (offer_wscale) |
YOSHIFUJI Hideaki | 496c98d | 2006-10-10 19:41:21 -0700 | [diff] [blame] | 436 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 437 | (TCPOPT_WINDOW << 16) | |
| 438 | (TCPOLEN_WINDOW << 8) | |
| 439 | (wscale)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 440 | #ifdef CONFIG_TCP_MD5SIG |
| 441 | /* |
| 442 | * If MD5 is enabled, then we set the option, and include the size |
| 443 | * (always 18). The actual MD5 hash is added just before the |
| 444 | * packet is sent. |
| 445 | */ |
| 446 | if (md5_hash) { |
| 447 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 448 | (TCPOPT_NOP << 16) | |
| 449 | (TCPOPT_MD5SIG << 8) | |
| 450 | TCPOLEN_MD5SIG); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 451 | *md5_hash = (__u8 *)ptr; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 452 | } |
| 453 | #endif |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 454 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 455 | |
| 456 | /* This routine actually transmits TCP packets queued in by |
| 457 | * tcp_do_sendmsg(). This is used by both the initial |
| 458 | * transmission and possible later retransmissions. |
| 459 | * All SKB's seen here are completely headerless. It is our |
| 460 | * job to build the TCP header, and pass the packet down to |
| 461 | * IP so it can do the same plus pass the packet off to the |
| 462 | * device. |
| 463 | * |
| 464 | * We are working here with either a clone of the original |
| 465 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 466 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 467 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
| 468 | gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 469 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 470 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 471 | struct inet_sock *inet; |
| 472 | struct tcp_sock *tp; |
| 473 | struct tcp_skb_cb *tcb; |
| 474 | int tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 475 | #ifdef CONFIG_TCP_MD5SIG |
| 476 | struct tcp_md5sig_key *md5; |
| 477 | __u8 *md5_hash_location; |
| 478 | #endif |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 479 | struct tcphdr *th; |
| 480 | int sysctl_flags; |
| 481 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 482 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 483 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
| 484 | |
| 485 | /* If congestion control is doing timestamping, we must |
| 486 | * take such a timestamp before we potentially clone/copy. |
| 487 | */ |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 488 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 489 | __net_timestamp(skb); |
| 490 | |
| 491 | if (likely(clone_it)) { |
| 492 | if (unlikely(skb_cloned(skb))) |
| 493 | skb = pskb_copy(skb, gfp_mask); |
| 494 | else |
| 495 | skb = skb_clone(skb, gfp_mask); |
| 496 | if (unlikely(!skb)) |
| 497 | return -ENOBUFS; |
| 498 | } |
| 499 | |
| 500 | inet = inet_sk(sk); |
| 501 | tp = tcp_sk(sk); |
| 502 | tcb = TCP_SKB_CB(skb); |
| 503 | tcp_header_size = tp->tcp_header_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 504 | |
| 505 | #define SYSCTL_FLAG_TSTAMPS 0x1 |
| 506 | #define SYSCTL_FLAG_WSCALE 0x2 |
| 507 | #define SYSCTL_FLAG_SACK 0x4 |
| 508 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 509 | sysctl_flags = 0; |
| 510 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 511 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 512 | if (sysctl_tcp_timestamps) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 513 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
| 514 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 515 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 516 | if (sysctl_tcp_window_scaling) { |
| 517 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; |
| 518 | sysctl_flags |= SYSCTL_FLAG_WSCALE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 519 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 520 | if (sysctl_tcp_sack) { |
| 521 | sysctl_flags |= SYSCTL_FLAG_SACK; |
| 522 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) |
| 523 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 524 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 525 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
| 526 | /* A SACK is 2 pad bytes, a 2 byte header, plus |
| 527 | * 2 32-bit sequence numbers for each SACK block. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 528 | */ |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 529 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + |
| 530 | (tp->rx_opt.eff_sacks * |
| 531 | TCPOLEN_SACK_PERBLOCK)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 532 | } |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 533 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 534 | if (tcp_packets_in_flight(tp) == 0) |
| 535 | tcp_ca_event(sk, CA_EVENT_TX_START); |
| 536 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 537 | #ifdef CONFIG_TCP_MD5SIG |
| 538 | /* |
| 539 | * Are we doing MD5 on this segment? If so - make |
| 540 | * room for it. |
| 541 | */ |
| 542 | md5 = tp->af_specific->md5_lookup(sk, sk); |
| 543 | if (md5) |
| 544 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 545 | #endif |
| 546 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 547 | skb_push(skb, tcp_header_size); |
| 548 | skb_reset_transport_header(skb); |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 549 | skb_set_owner_w(skb, sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 550 | |
| 551 | /* Build TCP header and checksum it. */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 552 | th = tcp_hdr(skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 553 | th->source = inet->sport; |
| 554 | th->dest = inet->dport; |
| 555 | th->seq = htonl(tcb->seq); |
| 556 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 557 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 558 | tcb->flags); |
| 559 | |
| 560 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
| 561 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 562 | * is never scaled. |
| 563 | */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 564 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 565 | } else { |
| 566 | th->window = htons(tcp_select_window(sk)); |
| 567 | } |
| 568 | th->check = 0; |
| 569 | th->urg_ptr = 0; |
| 570 | |
| 571 | if (unlikely(tp->urg_mode && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 572 | between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { |
| 573 | th->urg_ptr = htons(tp->snd_up - tcb->seq); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 574 | th->urg = 1; |
| 575 | } |
| 576 | |
| 577 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 578 | tcp_syn_build_options((__be32 *)(th + 1), |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 579 | tcp_advertise_mss(sk), |
| 580 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), |
| 581 | (sysctl_flags & SYSCTL_FLAG_SACK), |
| 582 | (sysctl_flags & SYSCTL_FLAG_WSCALE), |
| 583 | tp->rx_opt.rcv_wscale, |
| 584 | tcb->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 585 | tp->rx_opt.ts_recent, |
| 586 | |
| 587 | #ifdef CONFIG_TCP_MD5SIG |
| 588 | md5 ? &md5_hash_location : |
| 589 | #endif |
| 590 | NULL); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 591 | } else { |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 592 | tcp_build_and_update_options((__be32 *)(th + 1), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 593 | tp, tcb->when, |
| 594 | #ifdef CONFIG_TCP_MD5SIG |
| 595 | md5 ? &md5_hash_location : |
| 596 | #endif |
| 597 | NULL); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 598 | TCP_ECN_send(sk, skb, tcp_header_size); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 599 | } |
| 600 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 601 | #ifdef CONFIG_TCP_MD5SIG |
| 602 | /* Calculate the MD5 hash, as we have all we need now */ |
| 603 | if (md5) { |
| 604 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 605 | md5, |
| 606 | sk, NULL, NULL, |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 607 | tcp_hdr(skb), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 608 | skb->len); |
| 609 | } |
| 610 | #endif |
| 611 | |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 612 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 613 | |
| 614 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
| 615 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 616 | |
| 617 | if (skb->len != tcp_header_size) |
| 618 | tcp_event_data_sent(tp, skb, sk); |
| 619 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 620 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
| 621 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 622 | |
David S. Miller | e89862f | 2007-01-26 01:04:55 -0800 | [diff] [blame] | 623 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 624 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 625 | return err; |
| 626 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 627 | tcp_enter_cwr(sk, 1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 628 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 629 | return net_xmit_eval(err); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 630 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 631 | #undef SYSCTL_FLAG_TSTAMPS |
| 632 | #undef SYSCTL_FLAG_WSCALE |
| 633 | #undef SYSCTL_FLAG_SACK |
| 634 | } |
| 635 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 636 | /* This routine just queue's the buffer |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 637 | * |
| 638 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 639 | * otherwise socket can stall. |
| 640 | */ |
| 641 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 642 | { |
| 643 | struct tcp_sock *tp = tcp_sk(sk); |
| 644 | |
| 645 | /* Advance write_seq and place onto the write_queue. */ |
| 646 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
| 647 | skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 648 | tcp_add_write_queue_tail(sk, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 649 | sk->sk_wmem_queued += skb->truesize; |
| 650 | sk_mem_charge(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 651 | } |
| 652 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 653 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
| 654 | unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 655 | { |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 656 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 657 | /* Avoid the costly divide in the normal |
| 658 | * non-TSO case. |
| 659 | */ |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 660 | skb_shinfo(skb)->gso_segs = 1; |
| 661 | skb_shinfo(skb)->gso_size = 0; |
| 662 | skb_shinfo(skb)->gso_type = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 663 | } else { |
Ilpo Järvinen | 356f89e | 2007-08-24 23:00:31 -0700 | [diff] [blame] | 664 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 665 | skb_shinfo(skb)->gso_size = mss_now; |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 666 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 667 | } |
| 668 | } |
| 669 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 670 | /* When a modification to fackets out becomes necessary, we need to check |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 671 | * skb is counted to fackets_out or not. |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 672 | */ |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 673 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 674 | int decr) |
| 675 | { |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 676 | struct tcp_sock *tp = tcp_sk(sk); |
| 677 | |
Ilpo Järvinen | dc86967 | 2007-10-01 15:27:19 -0700 | [diff] [blame] | 678 | if (!tp->sacked_out || tcp_is_reno(tp)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 679 | return; |
| 680 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 681 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 682 | tp->fackets_out -= decr; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 683 | } |
| 684 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 685 | /* Function to create two new TCP segments. Shrinks the given segment |
| 686 | * 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] | 687 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 688 | * Remember, these are still headerless SKBs at this point. |
| 689 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 690 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
| 691 | unsigned int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 692 | { |
| 693 | struct tcp_sock *tp = tcp_sk(sk); |
| 694 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 695 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 696 | int nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 697 | u16 flags; |
| 698 | |
Herbert Xu | b2cc99f0 | 2005-10-20 17:13:13 -0200 | [diff] [blame] | 699 | BUG_ON(len > skb->len); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 700 | |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 701 | tcp_clear_retrans_hints_partial(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 702 | nsize = skb_headlen(skb) - len; |
| 703 | if (nsize < 0) |
| 704 | nsize = 0; |
| 705 | |
| 706 | if (skb_cloned(skb) && |
| 707 | skb_is_nonlinear(skb) && |
| 708 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 709 | return -ENOMEM; |
| 710 | |
| 711 | /* Get a new skb... force flag on. */ |
| 712 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); |
| 713 | if (buff == NULL) |
| 714 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 715 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 716 | sk->sk_wmem_queued += buff->truesize; |
| 717 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 718 | nlen = skb->len - len - nsize; |
| 719 | buff->truesize += nlen; |
| 720 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 721 | |
| 722 | /* Correct the sequence numbers. */ |
| 723 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 724 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 725 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 726 | |
| 727 | /* PSH and FIN should only be set in the second packet. */ |
| 728 | flags = TCP_SKB_CB(skb)->flags; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 729 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 730 | TCP_SKB_CB(buff)->flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 731 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 732 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 733 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 734 | /* Copy and checksum data tail into the new buffer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 735 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
| 736 | skb_put(buff, nsize), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 737 | nsize, 0); |
| 738 | |
| 739 | skb_trim(skb, len); |
| 740 | |
| 741 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 742 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 743 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 744 | skb_split(skb, buff, len); |
| 745 | } |
| 746 | |
| 747 | buff->ip_summed = skb->ip_summed; |
| 748 | |
| 749 | /* Looks stupid, but our code really uses when of |
| 750 | * skbs, which it never sent before. --ANK |
| 751 | */ |
| 752 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 753 | buff->tstamp = skb->tstamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 754 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 755 | old_factor = tcp_skb_pcount(skb); |
| 756 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 757 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 758 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 759 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 760 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 761 | /* If this packet has been sent out already, we must |
| 762 | * adjust the various packet counters. |
| 763 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 764 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 765 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 766 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 767 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 768 | tp->packets_out -= diff; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 769 | |
| 770 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 771 | tp->sacked_out -= diff; |
| 772 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 773 | tp->retrans_out -= diff; |
| 774 | |
Ilpo Järvinen | b5860bb | 2007-08-09 14:33:18 +0300 | [diff] [blame] | 775 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 776 | tp->lost_out -= diff; |
Herbert Xu | 83ca28b | 2005-09-22 23:32:56 -0700 | [diff] [blame] | 777 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 778 | /* Adjust Reno SACK estimate. */ |
| 779 | if (tcp_is_reno(tp) && diff > 0) { |
| 780 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); |
| 781 | tcp_verify_left_out(tp); |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 782 | } |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 783 | tcp_adjust_fackets_out(sk, skb, diff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 784 | } |
| 785 | |
| 786 | /* Link BUFF into the send queue. */ |
David S. Miller | f44b527 | 2005-07-05 15:18:34 -0700 | [diff] [blame] | 787 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 788 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 789 | |
| 790 | return 0; |
| 791 | } |
| 792 | |
| 793 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 794 | * eventually). The difference is that pulled data not copied, but |
| 795 | * immediately discarded. |
| 796 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 797 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 798 | { |
| 799 | int i, k, eat; |
| 800 | |
| 801 | eat = len; |
| 802 | k = 0; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 803 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 804 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
| 805 | put_page(skb_shinfo(skb)->frags[i].page); |
| 806 | eat -= skb_shinfo(skb)->frags[i].size; |
| 807 | } else { |
| 808 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; |
| 809 | if (eat) { |
| 810 | skb_shinfo(skb)->frags[k].page_offset += eat; |
| 811 | skb_shinfo(skb)->frags[k].size -= eat; |
| 812 | eat = 0; |
| 813 | } |
| 814 | k++; |
| 815 | } |
| 816 | } |
| 817 | skb_shinfo(skb)->nr_frags = k; |
| 818 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 819 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 820 | skb->data_len -= len; |
| 821 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 822 | } |
| 823 | |
| 824 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 825 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 826 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 827 | return -ENOMEM; |
| 828 | |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 829 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
| 830 | if (unlikely(len < skb_headlen(skb))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 831 | __skb_pull(skb, len); |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 832 | else |
| 833 | __pskb_trim_head(skb, len - skb_headlen(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 834 | |
| 835 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 836 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 837 | |
| 838 | skb->truesize -= len; |
| 839 | sk->sk_wmem_queued -= len; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 840 | sk_mem_uncharge(sk, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 841 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 842 | |
| 843 | /* Any change of skb->len requires recalculation of tso |
| 844 | * factor and mss. |
| 845 | */ |
| 846 | if (tcp_skb_pcount(skb) > 1) |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 847 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 848 | |
| 849 | return 0; |
| 850 | } |
| 851 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 852 | /* Not accounting for SACKs here. */ |
| 853 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
| 854 | { |
| 855 | struct tcp_sock *tp = tcp_sk(sk); |
| 856 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 857 | int mss_now; |
| 858 | |
| 859 | /* Calculate base mss without TCP options: |
| 860 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 861 | */ |
| 862 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 863 | |
| 864 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 865 | if (mss_now > tp->rx_opt.mss_clamp) |
| 866 | mss_now = tp->rx_opt.mss_clamp; |
| 867 | |
| 868 | /* Now subtract optional transport overhead */ |
| 869 | mss_now -= icsk->icsk_ext_hdr_len; |
| 870 | |
| 871 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 872 | if (mss_now < 48) |
| 873 | mss_now = 48; |
| 874 | |
| 875 | /* Now subtract TCP options size, not including SACKs */ |
| 876 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); |
| 877 | |
| 878 | return mss_now; |
| 879 | } |
| 880 | |
| 881 | /* Inverse of above */ |
| 882 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
| 883 | { |
| 884 | struct tcp_sock *tp = tcp_sk(sk); |
| 885 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 886 | int mtu; |
| 887 | |
| 888 | mtu = mss + |
| 889 | tp->tcp_header_len + |
| 890 | icsk->icsk_ext_hdr_len + |
| 891 | icsk->icsk_af_ops->net_header_len; |
| 892 | |
| 893 | return mtu; |
| 894 | } |
| 895 | |
| 896 | void tcp_mtup_init(struct sock *sk) |
| 897 | { |
| 898 | struct tcp_sock *tp = tcp_sk(sk); |
| 899 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 900 | |
| 901 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; |
| 902 | 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] | 903 | icsk->icsk_af_ops->net_header_len; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 904 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
| 905 | icsk->icsk_mtup.probe_size = 0; |
| 906 | } |
| 907 | |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 908 | /* Bound MSS / TSO packet size with the half of the window */ |
| 909 | static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) |
| 910 | { |
| 911 | if (tp->max_window && pktsize > (tp->max_window >> 1)) |
| 912 | return max(tp->max_window >> 1, 68U - tp->tcp_header_len); |
| 913 | else |
| 914 | return pktsize; |
| 915 | } |
| 916 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 917 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 918 | |
| 919 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 920 | for TCP options, but includes only bare TCP header. |
| 921 | |
| 922 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 923 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 924 | It also does not include TCP options. |
| 925 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 926 | 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] | 927 | |
| 928 | tp->mss_cache is current effective sending mss, including |
| 929 | all tcp options except for SACKs. It is evaluated, |
| 930 | taking into account current pmtu, but never exceeds |
| 931 | tp->rx_opt.mss_clamp. |
| 932 | |
| 933 | NOTE1. rfc1122 clearly states that advertised MSS |
| 934 | DOES NOT include either tcp or ip options. |
| 935 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 936 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 937 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 938 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 939 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 940 | { |
| 941 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 942 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 943 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 944 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 945 | if (icsk->icsk_mtup.search_high > pmtu) |
| 946 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 947 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 948 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 949 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 950 | |
| 951 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 952 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 953 | if (icsk->icsk_mtup.enabled) |
| 954 | 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] | 955 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 956 | |
| 957 | return mss_now; |
| 958 | } |
| 959 | |
| 960 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 961 | * and even PMTU discovery events into account. |
| 962 | * |
| 963 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up |
| 964 | * cannot be large. However, taking into account rare use of URG, this |
| 965 | * is not a big flaw. |
| 966 | */ |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 967 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 968 | { |
| 969 | struct tcp_sock *tp = tcp_sk(sk); |
| 970 | struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 971 | u32 mss_now; |
| 972 | u16 xmit_size_goal; |
| 973 | int doing_tso = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 974 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 975 | mss_now = tp->mss_cache; |
| 976 | |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 977 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 978 | doing_tso = 1; |
| 979 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 980 | if (dst) { |
| 981 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 982 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 983 | mss_now = tcp_sync_mss(sk, mtu); |
| 984 | } |
| 985 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 986 | if (tp->rx_opt.eff_sacks) |
| 987 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + |
| 988 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 989 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 990 | #ifdef CONFIG_TCP_MD5SIG |
| 991 | if (tp->af_specific->md5_lookup(sk, sk)) |
| 992 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; |
| 993 | #endif |
| 994 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 995 | xmit_size_goal = mss_now; |
| 996 | |
| 997 | if (doing_tso) { |
Peter P Waskiewicz Jr | 82cc1a7 | 2008-03-21 03:43:19 -0700 | [diff] [blame] | 998 | xmit_size_goal = ((sk->sk_gso_max_size - 1) - |
Arnaldo Carvalho de Melo | 8292a17 | 2005-12-13 23:15:52 -0800 | [diff] [blame] | 999 | inet_csk(sk)->icsk_af_ops->net_header_len - |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1000 | inet_csk(sk)->icsk_ext_hdr_len - |
| 1001 | tp->tcp_header_len); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1002 | |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 1003 | xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1004 | xmit_size_goal -= (xmit_size_goal % mss_now); |
| 1005 | } |
| 1006 | tp->xmit_size_goal = xmit_size_goal; |
| 1007 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1008 | return mss_now; |
| 1009 | } |
| 1010 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1011 | /* Congestion window validation. (RFC2861) */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1012 | static void tcp_cwnd_validate(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1013 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1014 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1015 | |
Ilpo Järvinen | d436d68 | 2007-12-31 14:58:00 -0800 | [diff] [blame] | 1016 | if (tp->packets_out >= tp->snd_cwnd) { |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1017 | /* Network is feed fully. */ |
| 1018 | tp->snd_cwnd_used = 0; |
| 1019 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1020 | } else { |
| 1021 | /* Network starves. */ |
| 1022 | if (tp->packets_out > tp->snd_cwnd_used) |
| 1023 | tp->snd_cwnd_used = tp->packets_out; |
| 1024 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 1025 | if (sysctl_tcp_slow_start_after_idle && |
| 1026 | (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] | 1027 | tcp_cwnd_application_limited(sk); |
| 1028 | } |
| 1029 | } |
| 1030 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1031 | /* Returns the portion of skb which can be sent right away without |
| 1032 | * introducing MSS oddities to segment boundaries. In rare cases where |
| 1033 | * mss_now != mss_cache, we will request caller to create a small skb |
| 1034 | * per input skb which could be mostly avoided here (if desired). |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1035 | * |
| 1036 | * We explicitly want to create a request for splitting write queue tail |
| 1037 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, |
| 1038 | * thus all the complexity (cwnd_len is always MSS multiple which we |
| 1039 | * return whenever allowed by the other factors). Basically we need the |
| 1040 | * modulo only when the receiver window alone is the limiting factor or |
| 1041 | * when we would be allowed to send the split-due-to-Nagle skb fully. |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1042 | */ |
| 1043 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1044 | unsigned int mss_now, unsigned int cwnd) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1045 | { |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1046 | struct tcp_sock *tp = tcp_sk(sk); |
| 1047 | u32 needed, window, cwnd_len; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1048 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1049 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1050 | cwnd_len = mss_now * cwnd; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1051 | |
| 1052 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) |
| 1053 | return cwnd_len; |
| 1054 | |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1055 | needed = min(skb->len, window); |
| 1056 | |
Ilpo Järvinen | 1751540 | 2008-04-15 20:36:55 -0700 | [diff] [blame] | 1057 | if (cwnd_len <= needed) |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1058 | return cwnd_len; |
| 1059 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1060 | return needed - needed % mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1061 | } |
| 1062 | |
| 1063 | /* Can at least one segment of SKB be sent right now, according to the |
| 1064 | * congestion window rules? If so, return how many segments are allowed. |
| 1065 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1066 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
| 1067 | struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1068 | { |
| 1069 | u32 in_flight, cwnd; |
| 1070 | |
| 1071 | /* Don't be strict about the congestion window for the final FIN. */ |
John Heffner | 104439a | 2007-02-05 17:53:11 -0800 | [diff] [blame] | 1072 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 1073 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1074 | return 1; |
| 1075 | |
| 1076 | in_flight = tcp_packets_in_flight(tp); |
| 1077 | cwnd = tp->snd_cwnd; |
| 1078 | if (in_flight < cwnd) |
| 1079 | return (cwnd - in_flight); |
| 1080 | |
| 1081 | return 0; |
| 1082 | } |
| 1083 | |
| 1084 | /* This must be invoked the first time we consider transmitting |
| 1085 | * SKB onto the wire. |
| 1086 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1087 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
| 1088 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1089 | { |
| 1090 | int tso_segs = tcp_skb_pcount(skb); |
| 1091 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1092 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1093 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1094 | tso_segs = tcp_skb_pcount(skb); |
| 1095 | } |
| 1096 | return tso_segs; |
| 1097 | } |
| 1098 | |
| 1099 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
| 1100 | { |
| 1101 | return after(tp->snd_sml,tp->snd_una) && |
| 1102 | !after(tp->snd_sml, tp->snd_nxt); |
| 1103 | } |
| 1104 | |
| 1105 | /* Return 0, if packet can be sent now without violation Nagle's rules: |
| 1106 | * 1. It is full sized. |
| 1107 | * 2. Or it contains FIN. (already checked by caller) |
| 1108 | * 3. Or TCP_NODELAY was set. |
| 1109 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1110 | * With Minshall's modification: all sent small packets are ACKed. |
| 1111 | */ |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1112 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1113 | const struct sk_buff *skb, |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1114 | unsigned mss_now, int nonagle) |
| 1115 | { |
| 1116 | return (skb->len < mss_now && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1117 | ((nonagle & TCP_NAGLE_CORK) || |
| 1118 | (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1119 | } |
| 1120 | |
| 1121 | /* Return non-zero if the Nagle test allows this packet to be |
| 1122 | * sent now. |
| 1123 | */ |
| 1124 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, |
| 1125 | unsigned int cur_mss, int nonagle) |
| 1126 | { |
| 1127 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1128 | * write_queue (they have no chances to get new data). |
| 1129 | * |
| 1130 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1131 | * argument based upon the location of SKB in the send queue. |
| 1132 | */ |
| 1133 | if (nonagle & TCP_NAGLE_PUSH) |
| 1134 | return 1; |
| 1135 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 1136 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
| 1137 | * Nagle can be ignored during F-RTO too (see RFC4138). |
| 1138 | */ |
| 1139 | if (tp->urg_mode || (tp->frto_counter == 2) || |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1140 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
| 1141 | return 1; |
| 1142 | |
| 1143 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) |
| 1144 | return 1; |
| 1145 | |
| 1146 | return 0; |
| 1147 | } |
| 1148 | |
| 1149 | /* Does at least the first segment of SKB fit into the send window? */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1150 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
| 1151 | unsigned int cur_mss) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1152 | { |
| 1153 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1154 | |
| 1155 | if (skb->len > cur_mss) |
| 1156 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1157 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1158 | return !after(end_seq, tcp_wnd_end(tp)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1159 | } |
| 1160 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1161 | /* 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] | 1162 | * should be put on the wire right now. If so, it returns the number of |
| 1163 | * packets allowed by the congestion window. |
| 1164 | */ |
| 1165 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, |
| 1166 | unsigned int cur_mss, int nonagle) |
| 1167 | { |
| 1168 | struct tcp_sock *tp = tcp_sk(sk); |
| 1169 | unsigned int cwnd_quota; |
| 1170 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1171 | tcp_init_tso_segs(sk, skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1172 | |
| 1173 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1174 | return 0; |
| 1175 | |
| 1176 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1177 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1178 | cwnd_quota = 0; |
| 1179 | |
| 1180 | return cwnd_quota; |
| 1181 | } |
| 1182 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1183 | int tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1184 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1185 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1186 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1187 | |
| 1188 | return (skb && |
| 1189 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), |
| 1190 | (tcp_skb_is_last(sk, skb) ? |
Ilpo Järvinen | 4e67d87 | 2007-12-05 02:25:32 -0800 | [diff] [blame] | 1191 | tp->nonagle : TCP_NAGLE_PUSH))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1192 | } |
| 1193 | |
| 1194 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1195 | * which is put after SKB on the list. It is very much like |
| 1196 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1197 | * in order to speed up the splitting operation. In particular, we |
| 1198 | * know that all the data is in scatter-gather pages, and that the |
| 1199 | * packet has never been sent out before (and thus is not cloned). |
| 1200 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1201 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
| 1202 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1203 | { |
| 1204 | struct sk_buff *buff; |
| 1205 | int nlen = skb->len - len; |
| 1206 | u16 flags; |
| 1207 | |
| 1208 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1209 | if (skb->len != skb->data_len) |
| 1210 | return tcp_fragment(sk, skb, len, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1211 | |
Pavel Emelyanov | df97c70 | 2007-11-29 21:22:33 +1100 | [diff] [blame] | 1212 | buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1213 | if (unlikely(buff == NULL)) |
| 1214 | return -ENOMEM; |
| 1215 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1216 | sk->sk_wmem_queued += buff->truesize; |
| 1217 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1218 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1219 | skb->truesize -= nlen; |
| 1220 | |
| 1221 | /* Correct the sequence numbers. */ |
| 1222 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1223 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1224 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1225 | |
| 1226 | /* PSH and FIN should only be set in the second packet. */ |
| 1227 | flags = TCP_SKB_CB(skb)->flags; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1228 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1229 | TCP_SKB_CB(buff)->flags = flags; |
| 1230 | |
| 1231 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1232 | TCP_SKB_CB(buff)->sacked = 0; |
| 1233 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1234 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1235 | skb_split(skb, buff, len); |
| 1236 | |
| 1237 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1238 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1239 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1240 | |
| 1241 | /* Link BUFF into the send queue. */ |
| 1242 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1243 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1244 | |
| 1245 | return 0; |
| 1246 | } |
| 1247 | |
| 1248 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1249 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1250 | * |
| 1251 | * This algorithm is from John Heffner. |
| 1252 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1253 | 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] | 1254 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1255 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1256 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1257 | u32 send_win, cong_win, limit, in_flight; |
| 1258 | |
| 1259 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1260 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1261 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1262 | if (icsk->icsk_ca_state != TCP_CA_Open) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1263 | goto send_now; |
| 1264 | |
| 1265 | /* Defer for less than two clock ticks. */ |
Ilpo Järvinen | bd515c3 | 2007-12-20 20:36:03 -0800 | [diff] [blame] | 1266 | if (tp->tso_deferred && |
| 1267 | ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1268 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1269 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1270 | in_flight = tcp_packets_in_flight(tp); |
| 1271 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1272 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1273 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1274 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1275 | |
| 1276 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1277 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1278 | |
| 1279 | limit = min(send_win, cong_win); |
| 1280 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1281 | /* If a full-sized TSO skb can be sent, do it. */ |
Peter P Waskiewicz Jr | 82cc1a7 | 2008-03-21 03:43:19 -0700 | [diff] [blame] | 1282 | if (limit >= sk->sk_gso_max_size) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1283 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1284 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1285 | if (sysctl_tcp_tso_win_divisor) { |
| 1286 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1287 | |
| 1288 | /* If at least some fraction of a window is available, |
| 1289 | * just use it. |
| 1290 | */ |
| 1291 | chunk /= sysctl_tcp_tso_win_divisor; |
| 1292 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1293 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1294 | } else { |
| 1295 | /* Different approach, try not to defer past a single |
| 1296 | * ACK. Receiver should ACK every other full sized |
| 1297 | * frame, so if we have space for more than 3 frames |
| 1298 | * then send now. |
| 1299 | */ |
| 1300 | if (limit > tcp_max_burst(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1301 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1302 | } |
| 1303 | |
| 1304 | /* Ok, it looks like it is advisable to defer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1305 | tp->tso_deferred = 1 | (jiffies << 1); |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1306 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1307 | return 1; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1308 | |
| 1309 | send_now: |
| 1310 | tp->tso_deferred = 0; |
| 1311 | return 0; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1312 | } |
| 1313 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1314 | /* Create a new MTU probe if we are ready. |
| 1315 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1316 | * 1 if a probe was sent, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1317 | * -1 otherwise |
| 1318 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1319 | static int tcp_mtu_probe(struct sock *sk) |
| 1320 | { |
| 1321 | struct tcp_sock *tp = tcp_sk(sk); |
| 1322 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1323 | struct sk_buff *skb, *nskb, *next; |
| 1324 | int len; |
| 1325 | int probe_size; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1326 | int size_needed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1327 | int copy; |
| 1328 | int mss_now; |
| 1329 | |
| 1330 | /* Not currently probing/verifying, |
| 1331 | * not in recovery, |
| 1332 | * have enough cwnd, and |
| 1333 | * not SACKing (the variable headers throw things off) */ |
| 1334 | if (!icsk->icsk_mtup.enabled || |
| 1335 | icsk->icsk_mtup.probe_size || |
| 1336 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1337 | tp->snd_cwnd < 11 || |
| 1338 | tp->rx_opt.eff_sacks) |
| 1339 | return -1; |
| 1340 | |
| 1341 | /* Very simple search strategy: just double the MSS. */ |
| 1342 | mss_now = tcp_current_mss(sk, 0); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1343 | probe_size = 2 * tp->mss_cache; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1344 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1345 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
| 1346 | /* TODO: set timer for probe_converge_event */ |
| 1347 | return -1; |
| 1348 | } |
| 1349 | |
| 1350 | /* Have enough data in the send queue to probe? */ |
Ilpo Järvinen | 7f9c33e | 2007-11-23 19:10:56 +0800 | [diff] [blame] | 1351 | if (tp->write_seq - tp->snd_nxt < size_needed) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1352 | return -1; |
| 1353 | |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1354 | if (tp->snd_wnd < size_needed) |
| 1355 | return -1; |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1356 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1357 | return 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1358 | |
Ilpo Järvinen | d67c58e | 2007-12-02 00:48:01 +0200 | [diff] [blame] | 1359 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
| 1360 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { |
| 1361 | if (!tcp_packets_in_flight(tp)) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1362 | return -1; |
| 1363 | else |
| 1364 | return 0; |
| 1365 | } |
| 1366 | |
| 1367 | /* We're allowed to probe. Build it now. */ |
| 1368 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) |
| 1369 | return -1; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1370 | sk->sk_wmem_queued += nskb->truesize; |
| 1371 | sk_mem_charge(sk, nskb->truesize); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1372 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1373 | skb = tcp_send_head(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1374 | |
| 1375 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1376 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; |
| 1377 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; |
| 1378 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1379 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1380 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1381 | |
Ilpo Järvinen | 50c4817 | 2007-12-02 00:48:00 +0200 | [diff] [blame] | 1382 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1383 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1384 | len = 0; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1385 | tcp_for_write_queue_from_safe(skb, next, sk) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1386 | copy = min_t(int, skb->len, probe_size - len); |
| 1387 | if (nskb->ip_summed) |
| 1388 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1389 | else |
| 1390 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1391 | skb_put(nskb, copy), |
| 1392 | copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1393 | |
| 1394 | if (skb->len <= copy) { |
| 1395 | /* We've eaten all the data from this skb. |
| 1396 | * Throw it away. */ |
| 1397 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1398 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1399 | sk_wmem_free_skb(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1400 | } else { |
| 1401 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1402 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1403 | if (!skb_shinfo(skb)->nr_frags) { |
| 1404 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1405 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1406 | skb->csum = csum_partial(skb->data, |
| 1407 | skb->len, 0); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1408 | } else { |
| 1409 | __pskb_trim_head(skb, copy); |
| 1410 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1411 | } |
| 1412 | TCP_SKB_CB(skb)->seq += copy; |
| 1413 | } |
| 1414 | |
| 1415 | len += copy; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1416 | |
| 1417 | if (len >= probe_size) |
| 1418 | break; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1419 | } |
| 1420 | tcp_init_tso_segs(sk, nskb, nskb->len); |
| 1421 | |
| 1422 | /* We're ready to send. If this fails, the probe will |
| 1423 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ |
| 1424 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; |
| 1425 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 1426 | /* Decrement cwnd here because we are sending |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1427 | * effectively two packets. */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1428 | tp->snd_cwnd--; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1429 | tcp_event_new_data_sent(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1430 | |
| 1431 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 1432 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 1433 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1434 | |
| 1435 | return 1; |
| 1436 | } |
| 1437 | |
| 1438 | return -1; |
| 1439 | } |
| 1440 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1441 | /* This routine writes packets to the network. It advances the |
| 1442 | * send_head. This happens as incoming acks open up the remote |
| 1443 | * window for us. |
| 1444 | * |
| 1445 | * Returns 1, if no segments are in flight and we have queued segments, but |
| 1446 | * cannot send anything now because of SWS or another problem. |
| 1447 | */ |
David S. Miller | a2e2a59 | 2005-07-05 15:19:23 -0700 | [diff] [blame] | 1448 | 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] | 1449 | { |
| 1450 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1451 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1452 | unsigned int tso_segs, sent_pkts; |
| 1453 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1454 | int result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1455 | |
| 1456 | /* If we are closed, the bytes will have to remain here. |
| 1457 | * In time closedown will finish, we empty the write queue and all |
| 1458 | * will be happy. |
| 1459 | */ |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1460 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 1461 | return 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1462 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1463 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1464 | |
| 1465 | /* Do MTU probing. */ |
| 1466 | if ((result = tcp_mtu_probe(sk)) == 0) { |
| 1467 | return 0; |
| 1468 | } else if (result > 0) { |
| 1469 | sent_pkts = 1; |
| 1470 | } |
| 1471 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1472 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1473 | unsigned int limit; |
| 1474 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1475 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1476 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1477 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1478 | cwnd_quota = tcp_cwnd_test(tp, skb); |
| 1479 | if (!cwnd_quota) |
| 1480 | break; |
| 1481 | |
| 1482 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 1483 | break; |
| 1484 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1485 | if (tso_segs == 1) { |
| 1486 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 1487 | (tcp_skb_is_last(sk, skb) ? |
| 1488 | nonagle : TCP_NAGLE_PUSH)))) |
| 1489 | break; |
| 1490 | } else { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1491 | if (tcp_tso_should_defer(sk, skb)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1492 | break; |
| 1493 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1494 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1495 | limit = mss_now; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1496 | if (tso_segs > 1) |
| 1497 | limit = tcp_mss_split_point(sk, skb, mss_now, |
| 1498 | cwnd_quota); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1499 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1500 | if (skb->len > limit && |
| 1501 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1502 | break; |
| 1503 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1504 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1505 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1506 | if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1507 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1508 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1509 | /* Advance the send_head. This one is sent out. |
| 1510 | * This call will increment packets_out. |
| 1511 | */ |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1512 | tcp_event_new_data_sent(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1513 | |
| 1514 | tcp_minshall_update(tp, mss_now, skb); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1515 | sent_pkts++; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1516 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1517 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1518 | if (likely(sent_pkts)) { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1519 | tcp_cwnd_validate(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1520 | return 0; |
| 1521 | } |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1522 | return !tp->packets_out && tcp_send_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1523 | } |
| 1524 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1525 | /* Push out any pending frames which were held back due to |
| 1526 | * TCP_CORK or attempt at coalescing tiny packets. |
| 1527 | * The socket must be locked by the caller. |
| 1528 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1529 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 1530 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1531 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1532 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1533 | |
| 1534 | if (skb) { |
David S. Miller | 55c97f3 | 2005-07-05 15:19:38 -0700 | [diff] [blame] | 1535 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1536 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1537 | } |
| 1538 | } |
| 1539 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1540 | /* Send _single_ skb sitting at the send head. This function requires |
| 1541 | * true push pending frames to setup probe timer etc. |
| 1542 | */ |
| 1543 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 1544 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1545 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1546 | unsigned int tso_segs, cwnd_quota; |
| 1547 | |
| 1548 | BUG_ON(!skb || skb->len < mss_now); |
| 1549 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1550 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1551 | cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); |
| 1552 | |
| 1553 | if (likely(cwnd_quota)) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1554 | unsigned int limit; |
| 1555 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1556 | BUG_ON(!tso_segs); |
| 1557 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1558 | limit = mss_now; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1559 | if (tso_segs > 1) |
| 1560 | limit = tcp_mss_split_point(sk, skb, mss_now, |
| 1561 | cwnd_quota); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1562 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1563 | if (skb->len > limit && |
| 1564 | unlikely(tso_fragment(sk, skb, limit, mss_now))) |
| 1565 | return; |
| 1566 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1567 | /* Send it out now. */ |
| 1568 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 1569 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1570 | if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1571 | tcp_event_new_data_sent(sk, skb); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1572 | tcp_cwnd_validate(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1573 | return; |
| 1574 | } |
| 1575 | } |
| 1576 | } |
| 1577 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1578 | /* This function returns the amount that we can raise the |
| 1579 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1580 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1581 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 1582 | * 2. We limit memory per socket |
| 1583 | * |
| 1584 | * RFC 1122: |
| 1585 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 1586 | * RECV.NEXT + RCV.WIN fixed until: |
| 1587 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 1588 | * |
| 1589 | * i.e. don't raise the right edge of the window until you can raise |
| 1590 | * it at least MSS bytes. |
| 1591 | * |
| 1592 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 1593 | * since header prediction assumes th->window stays fixed. |
| 1594 | * |
| 1595 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 1596 | * side SWS prevention criteria. The problem is that under this rule |
| 1597 | * a stream of single byte packets will cause the right side of the |
| 1598 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1599 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1600 | * Of course, if the sender implements sender side SWS prevention |
| 1601 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1602 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1603 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1604 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1605 | * If the free space is less than the 1/4 of the maximum |
| 1606 | * space available and the free space is less than 1/2 mss, |
| 1607 | * then set the window to 0. |
| 1608 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 1609 | * Otherwise, just prevent the window from shrinking |
| 1610 | * and from being larger than the largest representable value. |
| 1611 | * |
| 1612 | * This prevents incremental opening of the window in the regime |
| 1613 | * where TCP is limited by the speed of the reader side taking |
| 1614 | * data out of the TCP receive queue. It does nothing about |
| 1615 | * those cases where the window is constrained on the sender side |
| 1616 | * because the pipeline is full. |
| 1617 | * |
| 1618 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 1619 | * multiple of MSS, at least until the free space gets quite small. |
| 1620 | * This would appear to be a side effect of the mbuf implementation. |
| 1621 | * Combining these two algorithms results in the observed behavior |
| 1622 | * of having a fixed window size at almost all times. |
| 1623 | * |
| 1624 | * Below we obtain similar behavior by forcing the offered window to |
| 1625 | * a multiple of the mss when it is feasible to do so. |
| 1626 | * |
| 1627 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 1628 | * Regular options like TIMESTAMP are taken into account. |
| 1629 | */ |
| 1630 | u32 __tcp_select_window(struct sock *sk) |
| 1631 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1632 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1633 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1634 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1635 | * here. I don't know if the value based on our guesses |
| 1636 | * of peer's MSS is better for the performance. It's more correct |
| 1637 | * but may be worse for the performance because of rcv_mss |
| 1638 | * fluctuations. --SAW 1998/11/1 |
| 1639 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1640 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1641 | int free_space = tcp_space(sk); |
| 1642 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); |
| 1643 | int window; |
| 1644 | |
| 1645 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1646 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1647 | |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 1648 | if (free_space < (full_space >> 1)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1649 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1650 | |
| 1651 | if (tcp_memory_pressure) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1652 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
| 1653 | 4U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1654 | |
| 1655 | if (free_space < mss) |
| 1656 | return 0; |
| 1657 | } |
| 1658 | |
| 1659 | if (free_space > tp->rcv_ssthresh) |
| 1660 | free_space = tp->rcv_ssthresh; |
| 1661 | |
| 1662 | /* Don't do rounding if we are using window scaling, since the |
| 1663 | * scaled window will not line up with the MSS boundary anyway. |
| 1664 | */ |
| 1665 | window = tp->rcv_wnd; |
| 1666 | if (tp->rx_opt.rcv_wscale) { |
| 1667 | window = free_space; |
| 1668 | |
| 1669 | /* Advertise enough space so that it won't get scaled away. |
| 1670 | * Import case: prevent zero window announcement if |
| 1671 | * 1<<rcv_wscale > mss. |
| 1672 | */ |
| 1673 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 1674 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 1675 | << tp->rx_opt.rcv_wscale); |
| 1676 | } else { |
| 1677 | /* Get the largest window that is a nice multiple of mss. |
| 1678 | * Window clamp already applied above. |
| 1679 | * If our current window offering is within 1 mss of the |
| 1680 | * free space we just keep it. This prevents the divide |
| 1681 | * and multiply from happening most of the time. |
| 1682 | * We also don't do any window rounding when the free space |
| 1683 | * is too small. |
| 1684 | */ |
| 1685 | if (window <= free_space - mss || window > free_space) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1686 | window = (free_space / mss) * mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 1687 | else if (mss == full_space && |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 1688 | free_space > window + (full_space >> 1)) |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 1689 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1690 | } |
| 1691 | |
| 1692 | return window; |
| 1693 | } |
| 1694 | |
| 1695 | /* Attempt to collapse two adjacent SKB's during retransmission. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1696 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, |
| 1697 | int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1698 | { |
| 1699 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1700 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1701 | int skb_size, next_skb_size; |
| 1702 | u16 flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1703 | |
| 1704 | /* The first test we must make is that neither of these two |
| 1705 | * SKB's are still referenced by someone else. |
| 1706 | */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1707 | if (skb_cloned(skb) || skb_cloned(next_skb)) |
| 1708 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1709 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1710 | skb_size = skb->len; |
| 1711 | next_skb_size = next_skb->len; |
| 1712 | flags = TCP_SKB_CB(skb)->flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1713 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1714 | /* Also punt if next skb has been SACK'd. */ |
| 1715 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) |
| 1716 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1717 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1718 | /* Next skb is out of window. */ |
| 1719 | if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp))) |
| 1720 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1721 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1722 | /* Punt if not enough space exists in the first SKB for |
| 1723 | * the data in the second, or the total combined payload |
| 1724 | * would exceed the MSS. |
| 1725 | */ |
| 1726 | if ((next_skb_size > skb_tailroom(skb)) || |
| 1727 | ((skb_size + next_skb_size) > mss_now)) |
| 1728 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1729 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1730 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
Ilpo Järvinen | a6963a6 | 2007-09-25 22:44:14 -0700 | [diff] [blame] | 1731 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1732 | tcp_highest_sack_combine(sk, next_skb, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1733 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1734 | /* Ok. We will be able to collapse the packet. */ |
| 1735 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1736 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1737 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
| 1738 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1739 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1740 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 1741 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1742 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1743 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
| 1744 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1745 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1746 | /* Update sequence range on original skb. */ |
| 1747 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1748 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1749 | /* Merge over control information. */ |
| 1750 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ |
| 1751 | TCP_SKB_CB(skb)->flags = flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1752 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1753 | /* All done, get rid of second SKB and account for it so |
| 1754 | * packet counting does not break. |
| 1755 | */ |
| 1756 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; |
| 1757 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1758 | tp->retrans_out -= tcp_skb_pcount(next_skb); |
| 1759 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) |
| 1760 | tp->lost_out -= tcp_skb_pcount(next_skb); |
| 1761 | /* Reno case is special. Sigh... */ |
| 1762 | if (tcp_is_reno(tp) && tp->sacked_out) |
| 1763 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1764 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1765 | tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); |
| 1766 | tp->packets_out -= tcp_skb_pcount(next_skb); |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 1767 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 1768 | /* changed transmit queue under us so clear hints */ |
| 1769 | tcp_clear_retrans_hints_partial(tp); |
| 1770 | |
| 1771 | sk_wmem_free_skb(sk, next_skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1772 | } |
| 1773 | |
| 1774 | /* Do a simple retransmit without using the backoff mechanisms in |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1775 | * tcp_timer. This is used for path mtu discovery. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1776 | * The socket is already locked here. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1777 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1778 | void tcp_simple_retransmit(struct sock *sk) |
| 1779 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1780 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1781 | struct tcp_sock *tp = tcp_sk(sk); |
| 1782 | struct sk_buff *skb; |
| 1783 | unsigned int mss = tcp_current_mss(sk, 0); |
| 1784 | int lost = 0; |
| 1785 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1786 | tcp_for_write_queue(skb, sk) { |
| 1787 | if (skb == tcp_send_head(sk)) |
| 1788 | break; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1789 | if (skb->len > mss && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1790 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
| 1791 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1792 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
| 1793 | tp->retrans_out -= tcp_skb_pcount(skb); |
| 1794 | } |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1795 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1796 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
| 1797 | tp->lost_out += tcp_skb_pcount(skb); |
| 1798 | lost = 1; |
| 1799 | } |
| 1800 | } |
| 1801 | } |
| 1802 | |
Ilpo Järvinen | 5af4ec2 | 2007-09-20 11:30:48 -0700 | [diff] [blame] | 1803 | tcp_clear_all_retrans_hints(tp); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1804 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1805 | if (!lost) |
| 1806 | return; |
| 1807 | |
Ilpo Järvinen | 882beba | 2008-04-07 22:33:07 -0700 | [diff] [blame] | 1808 | if (tcp_is_reno(tp)) |
| 1809 | tcp_limit_reno_sacked(tp); |
| 1810 | |
Ilpo Järvinen | 005903b | 2007-08-09 14:44:16 +0300 | [diff] [blame] | 1811 | tcp_verify_left_out(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1812 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1813 | /* Don't muck with the congestion window here. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1814 | * Reason is that we do not increase amount of _data_ |
| 1815 | * in network, but units changed and effective |
| 1816 | * cwnd/ssthresh really reduced now. |
| 1817 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1818 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1819 | tp->high_seq = tp->snd_nxt; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1820 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1821 | tp->prior_ssthresh = 0; |
| 1822 | tp->undo_marker = 0; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1823 | tcp_set_ca_state(sk, TCP_CA_Loss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1824 | } |
| 1825 | tcp_xmit_retransmit_queue(sk); |
| 1826 | } |
| 1827 | |
| 1828 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 1829 | * state updates are done by the caller. Returns non-zero if an |
| 1830 | * error occurred which prevented the send. |
| 1831 | */ |
| 1832 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
| 1833 | { |
| 1834 | struct tcp_sock *tp = tcp_sk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1835 | struct inet_connection_sock *icsk = inet_csk(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 1836 | unsigned int cur_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1837 | int err; |
| 1838 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1839 | /* Inconslusive MTU probe */ |
| 1840 | if (icsk->icsk_mtup.probe_size) { |
| 1841 | icsk->icsk_mtup.probe_size = 0; |
| 1842 | } |
| 1843 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1844 | /* 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] | 1845 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1846 | */ |
| 1847 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 1848 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 1849 | return -EAGAIN; |
| 1850 | |
| 1851 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 1852 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 1853 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1854 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 1855 | return -ENOMEM; |
| 1856 | } |
| 1857 | |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 1858 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
| 1859 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 1860 | |
| 1861 | cur_mss = tcp_current_mss(sk, 0); |
| 1862 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1863 | /* If receiver has shrunk his window, and skb is out of |
| 1864 | * new window, do not retransmit it. The exception is the |
| 1865 | * case, when window is shrunk to zero. In this case |
| 1866 | * our retransmit serves as a zero window probe. |
| 1867 | */ |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1868 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1869 | && TCP_SKB_CB(skb)->seq != tp->snd_una) |
| 1870 | return -EAGAIN; |
| 1871 | |
| 1872 | if (skb->len > cur_mss) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1873 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1874 | return -ENOMEM; /* We'll try again later. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1875 | } |
| 1876 | |
| 1877 | /* Collapse two adjacent packets if worthwhile and we can. */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1878 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && |
| 1879 | (skb->len < (cur_mss >> 1)) && |
| 1880 | (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && |
| 1881 | (!tcp_skb_is_last(sk, skb)) && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1882 | (skb_shinfo(skb)->nr_frags == 0 && |
| 1883 | skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && |
| 1884 | (tcp_skb_pcount(skb) == 1 && |
| 1885 | tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1886 | (sysctl_tcp_retrans_collapse != 0)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1887 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
| 1888 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1889 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
| 1890 | * retransmit when old data is attached. So strip it off |
| 1891 | * since it is cheap to do so and saves bytes on the network. |
| 1892 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 1893 | if (skb->len > 0 && |
| 1894 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
| 1895 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1896 | if (!pskb_trim(skb, 0)) { |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 1897 | /* Reuse, even though it does some unnecessary work */ |
| 1898 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, |
| 1899 | TCP_SKB_CB(skb)->flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1900 | skb->ip_summed = CHECKSUM_NONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1901 | } |
| 1902 | } |
| 1903 | |
| 1904 | /* Make a copy, if the first transmission SKB clone we made |
| 1905 | * is still in somebody's hands, else make a clone. |
| 1906 | */ |
| 1907 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1908 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1909 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1910 | |
| 1911 | if (err == 0) { |
| 1912 | /* Update global TCP statistics. */ |
| 1913 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); |
| 1914 | |
| 1915 | tp->total_retrans++; |
| 1916 | |
| 1917 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1918 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1919 | if (net_ratelimit()) |
| 1920 | printk(KERN_DEBUG "retrans_out leaked.\n"); |
| 1921 | } |
| 1922 | #endif |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 1923 | if (!tp->retrans_out) |
| 1924 | tp->lost_retrans_low = tp->snd_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1925 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 1926 | tp->retrans_out += tcp_skb_pcount(skb); |
| 1927 | |
| 1928 | /* Save stamp of the first retransmit. */ |
| 1929 | if (!tp->retrans_stamp) |
| 1930 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; |
| 1931 | |
| 1932 | tp->undo_retrans++; |
| 1933 | |
| 1934 | /* snd_nxt is stored to detect loss of retransmitted segment, |
| 1935 | * see tcp_input.c tcp_sacktag_write_queue(). |
| 1936 | */ |
| 1937 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; |
| 1938 | } |
| 1939 | return err; |
| 1940 | } |
| 1941 | |
| 1942 | /* This gets called after a retransmit timeout, and the initially |
| 1943 | * retransmitted data is acknowledged. It tries to continue |
| 1944 | * resending the rest of the retransmit queue, until either |
| 1945 | * we've sent it all or the congestion window limit is reached. |
| 1946 | * If doing SACK, the first ACK which comes back for a timeout |
| 1947 | * based retransmit packet might feed us FACK information again. |
| 1948 | * If so, we use it to avoid unnecessarily retransmissions. |
| 1949 | */ |
| 1950 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 1951 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1952 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1953 | struct tcp_sock *tp = tcp_sk(sk); |
| 1954 | struct sk_buff *skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1955 | int packet_cnt; |
| 1956 | |
| 1957 | if (tp->retransmit_skb_hint) { |
| 1958 | skb = tp->retransmit_skb_hint; |
| 1959 | packet_cnt = tp->retransmit_cnt_hint; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1960 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1961 | skb = tcp_write_queue_head(sk); |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1962 | packet_cnt = 0; |
| 1963 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1964 | |
| 1965 | /* First pass: retransmit lost packets. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1966 | if (tp->lost_out) { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1967 | tcp_for_write_queue_from(skb, sk) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1968 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
| 1969 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1970 | if (skb == tcp_send_head(sk)) |
| 1971 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1972 | /* we could do better than to assign each time */ |
| 1973 | tp->retransmit_skb_hint = skb; |
| 1974 | tp->retransmit_cnt_hint = packet_cnt; |
| 1975 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1976 | /* Assume this retransmit will generate |
| 1977 | * only one packet for congestion window |
| 1978 | * calculation purposes. This works because |
| 1979 | * tcp_retransmit_skb() will chop up the |
| 1980 | * packet to be MSS sized and all the |
| 1981 | * packet counting works out. |
| 1982 | */ |
| 1983 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 1984 | return; |
| 1985 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1986 | if (sacked & TCPCB_LOST) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1987 | if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1988 | if (tcp_retransmit_skb(sk, skb)) { |
| 1989 | tp->retransmit_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1990 | return; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1991 | } |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1992 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1993 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); |
| 1994 | else |
| 1995 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); |
| 1996 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1997 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 1998 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 1999 | inet_csk(sk)->icsk_rto, |
| 2000 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2001 | } |
| 2002 | |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2003 | packet_cnt += tcp_skb_pcount(skb); |
| 2004 | if (packet_cnt >= tp->lost_out) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2005 | break; |
| 2006 | } |
| 2007 | } |
| 2008 | } |
| 2009 | |
| 2010 | /* OK, demanded retransmission is finished. */ |
| 2011 | |
| 2012 | /* Forward retransmissions are possible only during Recovery. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2013 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2014 | return; |
| 2015 | |
| 2016 | /* No forward retransmissions in Reno are possible. */ |
Ilpo Järvinen | e60402d | 2007-08-09 15:14:46 +0300 | [diff] [blame] | 2017 | if (tcp_is_reno(tp)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2018 | return; |
| 2019 | |
| 2020 | /* Yeah, we have to make difficult choice between forward transmission |
| 2021 | * and retransmission... Both ways have their merits... |
| 2022 | * |
| 2023 | * For now we do not retransmit anything, while we have some new |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2024 | * segments to send. In the other cases, follow rule 3 for |
| 2025 | * NextSeg() specified in RFC3517. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2026 | */ |
| 2027 | |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2028 | if (tcp_may_send_now(sk)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2029 | return; |
| 2030 | |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2031 | /* If nothing is SACKed, highest_sack in the loop won't be valid */ |
| 2032 | if (!tp->sacked_out) |
| 2033 | return; |
| 2034 | |
| 2035 | if (tp->forward_skb_hint) |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2036 | skb = tp->forward_skb_hint; |
Ilpo Järvinen | 539d243 | 2007-05-27 02:03:20 -0700 | [diff] [blame] | 2037 | else |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2038 | skb = tcp_write_queue_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2039 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2040 | tcp_for_write_queue_from(skb, sk) { |
| 2041 | if (skb == tcp_send_head(sk)) |
| 2042 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2043 | tp->forward_skb_hint = skb; |
| 2044 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 2045 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2046 | break; |
| 2047 | |
| 2048 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 2049 | break; |
| 2050 | |
| 2051 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) |
| 2052 | continue; |
| 2053 | |
| 2054 | /* Ok, retransmit it. */ |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2055 | if (tcp_retransmit_skb(sk, skb)) { |
| 2056 | tp->forward_skb_hint = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2057 | break; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2058 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2059 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2060 | if (skb == tcp_write_queue_head(sk)) |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2061 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2062 | inet_csk(sk)->icsk_rto, |
| 2063 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2064 | |
| 2065 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); |
| 2066 | } |
| 2067 | } |
| 2068 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2069 | /* Send a fin. The caller locks the socket for us. This cannot be |
| 2070 | * allowed to fail queueing a FIN frame under any circumstances. |
| 2071 | */ |
| 2072 | void tcp_send_fin(struct sock *sk) |
| 2073 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2074 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2075 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2076 | int mss_now; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2077 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2078 | /* Optimization, tack on the FIN if we have a queue of |
| 2079 | * unsent frames. But be careful about outgoing SACKS |
| 2080 | * and IP options. |
| 2081 | */ |
| 2082 | mss_now = tcp_current_mss(sk, 1); |
| 2083 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2084 | if (tcp_send_head(sk) != NULL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2085 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; |
| 2086 | TCP_SKB_CB(skb)->end_seq++; |
| 2087 | tp->write_seq++; |
| 2088 | } else { |
| 2089 | /* Socket is locked, keep trying until memory is available. */ |
| 2090 | for (;;) { |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2091 | skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2092 | if (skb) |
| 2093 | break; |
| 2094 | yield(); |
| 2095 | } |
| 2096 | |
| 2097 | /* Reserve space for headers and prepare control bits. */ |
| 2098 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2099 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2100 | tcp_init_nondata_skb(skb, tp->write_seq, |
| 2101 | TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2102 | tcp_queue_skb(sk, skb); |
| 2103 | } |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2104 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2105 | } |
| 2106 | |
| 2107 | /* We get here when a process closes a file descriptor (either due to |
| 2108 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2109 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2110 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2111 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2112 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2113 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2114 | struct sk_buff *skb; |
| 2115 | |
| 2116 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2117 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2118 | if (!skb) { |
| 2119 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
| 2120 | return; |
| 2121 | } |
| 2122 | |
| 2123 | /* Reserve space for headers and prepare control bits. */ |
| 2124 | skb_reserve(skb, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2125 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
| 2126 | TCPCB_FLAG_ACK | TCPCB_FLAG_RST); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2127 | /* Send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2128 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2129 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2130 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
Sridhar Samudrala | 26af65c | 2008-06-04 15:19:35 -0700 | [diff] [blame] | 2131 | |
| 2132 | TCP_INC_STATS(TCP_MIB_OUTRSTS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2133 | } |
| 2134 | |
| 2135 | /* WARNING: This routine must only be called when we have already sent |
| 2136 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2137 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2138 | * and rcv_wscale values will not be correct. |
| 2139 | */ |
| 2140 | int tcp_send_synack(struct sock *sk) |
| 2141 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2142 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2143 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2144 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2145 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2146 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
| 2147 | return -EFAULT; |
| 2148 | } |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2149 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2150 | if (skb_cloned(skb)) { |
| 2151 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
| 2152 | if (nskb == NULL) |
| 2153 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2154 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2155 | skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2156 | __tcp_add_write_queue_head(sk, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2157 | sk_wmem_free_skb(sk, skb); |
| 2158 | sk->sk_wmem_queued += nskb->truesize; |
| 2159 | sk_mem_charge(sk, nskb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2160 | skb = nskb; |
| 2161 | } |
| 2162 | |
| 2163 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; |
| 2164 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
| 2165 | } |
| 2166 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2167 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2168 | } |
| 2169 | |
| 2170 | /* |
| 2171 | * Prepare a SYN-ACK. |
| 2172 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2173 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
| 2174 | struct request_sock *req) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2175 | { |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2176 | struct inet_request_sock *ireq = inet_rsk(req); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2177 | struct tcp_sock *tp = tcp_sk(sk); |
| 2178 | struct tcphdr *th; |
| 2179 | int tcp_header_size; |
| 2180 | struct sk_buff *skb; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2181 | #ifdef CONFIG_TCP_MD5SIG |
| 2182 | struct tcp_md5sig_key *md5; |
| 2183 | __u8 *md5_hash_location; |
| 2184 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2185 | |
| 2186 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); |
| 2187 | if (skb == NULL) |
| 2188 | return NULL; |
| 2189 | |
| 2190 | /* Reserve space for headers. */ |
| 2191 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2192 | |
| 2193 | skb->dst = dst_clone(dst); |
| 2194 | |
| 2195 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2196 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
| 2197 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2198 | /* 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] | 2199 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2200 | |
| 2201 | #ifdef CONFIG_TCP_MD5SIG |
| 2202 | /* Are we doing MD5 on this segment? If so - make room for it */ |
| 2203 | md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
| 2204 | if (md5) |
| 2205 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; |
| 2206 | #endif |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2207 | skb_push(skb, tcp_header_size); |
| 2208 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2209 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2210 | th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2211 | memset(th, 0, sizeof(struct tcphdr)); |
| 2212 | th->syn = 1; |
| 2213 | th->ack = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2214 | TCP_ECN_make_synack(req, th); |
| 2215 | th->source = inet_sk(sk)->sport; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2216 | th->dest = ireq->rmt_port; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2217 | /* Setting of flags are superfluous here for callers (and ECE is |
| 2218 | * not even correctly set) |
| 2219 | */ |
| 2220 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, |
| 2221 | TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2222 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2223 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2224 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2225 | __u8 rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2226 | /* Set this up on the first call only */ |
| 2227 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); |
| 2228 | /* tcp_full_space because it is guaranteed to be the first packet */ |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2229 | tcp_select_initial_window(tcp_full_space(sk), |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2230 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2231 | &req->rcv_wnd, |
| 2232 | &req->window_clamp, |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2233 | ireq->wscale_ok, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2234 | &rcv_wscale); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2235 | ireq->rcv_wscale = rcv_wscale; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2236 | } |
| 2237 | |
| 2238 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 2239 | th->window = htons(min(req->rcv_wnd, 65535U)); |
Florian Westphal | 4dfc281 | 2008-04-10 03:12:40 -0700 | [diff] [blame] | 2240 | #ifdef CONFIG_SYN_COOKIES |
| 2241 | if (unlikely(req->cookie_ts)) |
| 2242 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); |
| 2243 | else |
| 2244 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2245 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 2246 | 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] | 2247 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2248 | TCP_SKB_CB(skb)->when, |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2249 | req->ts_recent, |
| 2250 | ( |
| 2251 | #ifdef CONFIG_TCP_MD5SIG |
| 2252 | md5 ? &md5_hash_location : |
| 2253 | #endif |
| 2254 | NULL) |
| 2255 | ); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2256 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2257 | th->doff = (tcp_header_size >> 2); |
| 2258 | TCP_INC_STATS(TCP_MIB_OUTSEGS); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2259 | |
| 2260 | #ifdef CONFIG_TCP_MD5SIG |
| 2261 | /* Okay, we have all we need - do the md5 hash if needed */ |
| 2262 | if (md5) { |
| 2263 | tp->af_specific->calc_md5_hash(md5_hash_location, |
| 2264 | md5, |
| 2265 | NULL, dst, req, |
YOSHIFUJI Hideaki | 076fb72 | 2008-04-17 12:48:12 +0900 | [diff] [blame] | 2266 | tcp_hdr(skb), |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2267 | skb->len); |
| 2268 | } |
| 2269 | #endif |
| 2270 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2271 | return skb; |
| 2272 | } |
| 2273 | |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2274 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2275 | * Do all connect socket setups that can be done AF independent. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2276 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 2277 | static void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2278 | { |
| 2279 | struct dst_entry *dst = __sk_dst_get(sk); |
| 2280 | struct tcp_sock *tp = tcp_sk(sk); |
| 2281 | __u8 rcv_wscale; |
| 2282 | |
| 2283 | /* We'll fix this up when we get a response from the other end. |
| 2284 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 2285 | */ |
| 2286 | tp->tcp_header_len = sizeof(struct tcphdr) + |
| 2287 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
| 2288 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2289 | #ifdef CONFIG_TCP_MD5SIG |
| 2290 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) |
| 2291 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 2292 | #endif |
| 2293 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2294 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 2295 | if (tp->rx_opt.user_mss) |
| 2296 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2297 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2298 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2299 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 2300 | |
| 2301 | if (!tp->window_clamp) |
| 2302 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
| 2303 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); |
| 2304 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2305 | |
| 2306 | tcp_select_initial_window(tcp_full_space(sk), |
| 2307 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 2308 | &tp->rcv_wnd, |
| 2309 | &tp->window_clamp, |
| 2310 | sysctl_tcp_window_scaling, |
| 2311 | &rcv_wscale); |
| 2312 | |
| 2313 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 2314 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 2315 | |
| 2316 | sk->sk_err = 0; |
| 2317 | sock_reset_flag(sk, SOCK_DONE); |
| 2318 | tp->snd_wnd = 0; |
| 2319 | tcp_init_wl(tp, tp->write_seq, 0); |
| 2320 | tp->snd_una = tp->write_seq; |
| 2321 | tp->snd_sml = tp->write_seq; |
| 2322 | tp->rcv_nxt = 0; |
| 2323 | tp->rcv_wup = 0; |
| 2324 | tp->copied_seq = 0; |
| 2325 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2326 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 2327 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2328 | tcp_clear_retrans(tp); |
| 2329 | } |
| 2330 | |
| 2331 | /* |
| 2332 | * Build a SYN and send it off. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2333 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2334 | int tcp_connect(struct sock *sk) |
| 2335 | { |
| 2336 | struct tcp_sock *tp = tcp_sk(sk); |
| 2337 | struct sk_buff *buff; |
| 2338 | |
| 2339 | tcp_connect_init(sk); |
| 2340 | |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 2341 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2342 | if (unlikely(buff == NULL)) |
| 2343 | return -ENOBUFS; |
| 2344 | |
| 2345 | /* Reserve space for headers. */ |
| 2346 | skb_reserve(buff, MAX_TCP_HEADER); |
| 2347 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2348 | tp->snd_nxt = tp->write_seq; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2349 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); |
| 2350 | TCP_ECN_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2351 | |
| 2352 | /* Send it off. */ |
| 2353 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2354 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; |
| 2355 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2356 | __tcp_add_write_queue_tail(sk, buff); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2357 | sk->sk_wmem_queued += buff->truesize; |
| 2358 | sk_mem_charge(sk, buff->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2359 | tp->packets_out += tcp_skb_pcount(buff); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2360 | tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 2361 | |
| 2362 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 2363 | * in order to make this packet get counted in tcpOutSegs. |
| 2364 | */ |
| 2365 | tp->snd_nxt = tp->write_seq; |
| 2366 | tp->pushed_seq = tp->write_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2367 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); |
| 2368 | |
| 2369 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2370 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2371 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2372 | return 0; |
| 2373 | } |
| 2374 | |
| 2375 | /* Send out a delayed ack, the caller does the policy checking |
| 2376 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 2377 | * for details. |
| 2378 | */ |
| 2379 | void tcp_send_delayed_ack(struct sock *sk) |
| 2380 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2381 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 2382 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2383 | unsigned long timeout; |
| 2384 | |
| 2385 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2386 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2387 | int max_ato = HZ / 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2388 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2389 | if (icsk->icsk_ack.pingpong || |
| 2390 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2391 | max_ato = TCP_DELACK_MAX; |
| 2392 | |
| 2393 | /* Slow path, intersegment interval is "high". */ |
| 2394 | |
| 2395 | /* 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] | 2396 | * 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] | 2397 | * directly. |
| 2398 | */ |
| 2399 | if (tp->srtt) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2400 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2401 | |
| 2402 | if (rtt < max_ato) |
| 2403 | max_ato = rtt; |
| 2404 | } |
| 2405 | |
| 2406 | ato = min(ato, max_ato); |
| 2407 | } |
| 2408 | |
| 2409 | /* Stay within the limit we were given */ |
| 2410 | timeout = jiffies + ato; |
| 2411 | |
| 2412 | /* 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] | 2413 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2414 | /* If delack timer was blocked or is about to expire, |
| 2415 | * send ACK now. |
| 2416 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2417 | if (icsk->icsk_ack.blocked || |
| 2418 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2419 | tcp_send_ack(sk); |
| 2420 | return; |
| 2421 | } |
| 2422 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2423 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 2424 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2425 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2426 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 2427 | icsk->icsk_ack.timeout = timeout; |
| 2428 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2429 | } |
| 2430 | |
| 2431 | /* This routine sends an ack and also updates the window. */ |
| 2432 | void tcp_send_ack(struct sock *sk) |
| 2433 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2434 | struct sk_buff *buff; |
| 2435 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2436 | /* If we have been reset, we may not send again. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2437 | if (sk->sk_state == TCP_CLOSE) |
| 2438 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2439 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2440 | /* We are not putting this on the write queue, so |
| 2441 | * tcp_transmit_skb() will set the ownership to this |
| 2442 | * sock. |
| 2443 | */ |
| 2444 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
| 2445 | if (buff == NULL) { |
| 2446 | inet_csk_schedule_ack(sk); |
| 2447 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
| 2448 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 2449 | TCP_DELACK_MAX, TCP_RTO_MAX); |
| 2450 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2451 | } |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2452 | |
| 2453 | /* Reserve space for headers and prepare control bits. */ |
| 2454 | skb_reserve(buff, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2455 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2456 | |
| 2457 | /* Send it off, this clears delayed acks for us. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2458 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 2459 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2460 | } |
| 2461 | |
| 2462 | /* This routine sends a packet with an out of date sequence |
| 2463 | * number. It assumes the other end will try to ack it. |
| 2464 | * |
| 2465 | * Question: what should we make while urgent mode? |
| 2466 | * 4.4BSD forces sending single byte of data. We cannot send |
| 2467 | * out of window data, because we have SND.NXT==SND.MAX... |
| 2468 | * |
| 2469 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 2470 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 2471 | * out-of-date with SND.UNA-1 to probe window. |
| 2472 | */ |
| 2473 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) |
| 2474 | { |
| 2475 | struct tcp_sock *tp = tcp_sk(sk); |
| 2476 | struct sk_buff *skb; |
| 2477 | |
| 2478 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
| 2479 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2480 | if (skb == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2481 | return -1; |
| 2482 | |
| 2483 | /* Reserve space for headers and set control bits. */ |
| 2484 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2485 | /* Use a previous sequence. This should cause the other |
| 2486 | * end to send an ack. Don't queue or clone SKB, just |
| 2487 | * send it. |
| 2488 | */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2489 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2490 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2491 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2492 | } |
| 2493 | |
| 2494 | int tcp_write_wakeup(struct sock *sk) |
| 2495 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2496 | struct tcp_sock *tp = tcp_sk(sk); |
| 2497 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2498 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2499 | if (sk->sk_state == TCP_CLOSE) |
| 2500 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2501 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2502 | if ((skb = tcp_send_head(sk)) != NULL && |
| 2503 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { |
| 2504 | int err; |
| 2505 | unsigned int mss = tcp_current_mss(sk, 0); |
| 2506 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2507 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2508 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 2509 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2510 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2511 | /* We are probing the opening of a window |
| 2512 | * but the window size is != 0 |
| 2513 | * must have been a result SWS avoidance ( sender ) |
| 2514 | */ |
| 2515 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 2516 | skb->len > mss) { |
| 2517 | seg_size = min(seg_size, mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2518 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2519 | if (tcp_fragment(sk, skb, seg_size, mss)) |
| 2520 | return -1; |
| 2521 | } else if (!tcp_skb_pcount(skb)) |
| 2522 | tcp_set_skb_tso_segs(sk, skb, mss); |
| 2523 | |
| 2524 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; |
| 2525 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 2526 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
| 2527 | if (!err) |
| 2528 | tcp_event_new_data_sent(sk, skb); |
| 2529 | return err; |
| 2530 | } else { |
| 2531 | if (tp->urg_mode && |
| 2532 | between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
| 2533 | tcp_xmit_probe_skb(sk, 1); |
| 2534 | return tcp_xmit_probe_skb(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2535 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2536 | } |
| 2537 | |
| 2538 | /* A window probe timeout has occurred. If window is not closed send |
| 2539 | * a partial packet else a zero probe. |
| 2540 | */ |
| 2541 | void tcp_send_probe0(struct sock *sk) |
| 2542 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2543 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2544 | struct tcp_sock *tp = tcp_sk(sk); |
| 2545 | int err; |
| 2546 | |
| 2547 | err = tcp_write_wakeup(sk); |
| 2548 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2549 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2550 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2551 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2552 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2553 | return; |
| 2554 | } |
| 2555 | |
| 2556 | if (err <= 0) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2557 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
| 2558 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2559 | icsk->icsk_probes_out++; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2560 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2561 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 2562 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2563 | } else { |
| 2564 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2565 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2566 | * Let local senders to fight for local resources. |
| 2567 | * |
| 2568 | * Use accumulated backoff yet. |
| 2569 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2570 | if (!icsk->icsk_probes_out) |
| 2571 | icsk->icsk_probes_out = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2572 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2573 | min(icsk->icsk_rto << icsk->icsk_backoff, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 2574 | TCP_RESOURCE_PROBE_INTERVAL), |
| 2575 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2576 | } |
| 2577 | } |
| 2578 | |
Glenn Griffin | c6aefaf | 2008-02-07 21:49:26 -0800 | [diff] [blame] | 2579 | EXPORT_SYMBOL(tcp_select_initial_window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2580 | EXPORT_SYMBOL(tcp_connect); |
| 2581 | EXPORT_SYMBOL(tcp_make_synack); |
| 2582 | EXPORT_SYMBOL(tcp_simple_retransmit); |
| 2583 | EXPORT_SYMBOL(tcp_sync_mss); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2584 | EXPORT_SYMBOL(tcp_mtup_init); |