Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * INET An implementation of the TCP/IP protocol suite for the LINUX |
| 3 | * operating system. INET is implemented using the BSD Socket |
| 4 | * interface as the means of communication with the user level. |
| 5 | * |
| 6 | * Implementation of the Transmission Control Protocol(TCP). |
| 7 | * |
Jesper Juhl | 02c30a8 | 2005-05-05 16:16:16 -0700 | [diff] [blame] | 8 | * Authors: Ross Biro |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
| 10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> |
| 11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> |
| 12 | * Florian La Roche, <flla@stud.uni-sb.de> |
| 13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> |
| 14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> |
| 15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> |
| 16 | * Matthew Dillon, <dillon@apollo.west.oic.com> |
| 17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> |
| 18 | * Jorge Cwik, <jorge@laser.satlink.net> |
| 19 | */ |
| 20 | |
| 21 | /* |
| 22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. |
| 23 | * : Fragmentation on mtu decrease |
| 24 | * : Segment collapse on retransmit |
| 25 | * : AF independence |
| 26 | * |
| 27 | * Linus Torvalds : send_delayed_ack |
| 28 | * David S. Miller : Charge memory using the right skb |
| 29 | * during syn/ack processing. |
| 30 | * David S. Miller : Output engine completely rewritten. |
| 31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. |
| 32 | * Cacophonix Gaul : draft-minshall-nagle-01 |
| 33 | * J Hadi Salim : ECN support |
| 34 | * |
| 35 | */ |
| 36 | |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 37 | #define pr_fmt(fmt) "TCP: " fmt |
| 38 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 39 | #include <net/tcp.h> |
| 40 | |
| 41 | #include <linux/compiler.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 42 | #include <linux/gfp.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 43 | #include <linux/module.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | |
| 45 | /* People can turn this off for buggy TCP's found in printers etc. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 47 | |
Jianjun Kong | 09cb105 | 2008-11-03 00:27:11 -0800 | [diff] [blame] | 48 | /* People can turn this on to work with those rare, broken TCPs that |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 49 | * interpret the window field as a signed quantity. |
| 50 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 52 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 53 | /* Default TSQ limit of two TSO segments */ |
| 54 | int sysctl_tcp_limit_output_bytes __read_mostly = 131072; |
| 55 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 56 | /* This limits the percentage of the congestion window which we |
| 57 | * will allow a single TSO frame to consume. Building TSO frames |
| 58 | * which are too large can cause TCP streams to be bursty. |
| 59 | */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 60 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 61 | |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 62 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
Shan Wei | 97b1ce25 | 2010-12-01 18:04:50 +0000 | [diff] [blame] | 63 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 64 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 65 | /* By default, RFC2861 behavior. */ |
Brian Haley | ab32ea5 | 2006-09-22 14:15:41 -0700 | [diff] [blame] | 66 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 67 | |
William Allen Simpson | 519855c | 2009-12-02 18:14:19 +0000 | [diff] [blame] | 68 | int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ |
David S. Miller | e6b09cc | 2009-12-02 22:18:58 -0800 | [diff] [blame] | 69 | EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); |
William Allen Simpson | 519855c | 2009-12-02 18:14:19 +0000 | [diff] [blame] | 70 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 71 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 72 | int push_one, gfp_t gfp); |
William Allen Simpson | 519855c | 2009-12-02 18:14:19 +0000 | [diff] [blame] | 73 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 74 | /* Account for new data that has been sent to the network. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 75 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
Ilpo Järvinen | 6ff03ac | 2007-08-24 22:44:06 -0700 | [diff] [blame] | 76 | { |
| 77 | struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 78 | unsigned int prior_packets = tp->packets_out; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 79 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 80 | tcp_advance_send_head(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 81 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 82 | |
Lucas De Marchi | 25985ed | 2011-03-30 22:57:33 -0300 | [diff] [blame] | 83 | /* Don't override Nagle indefinitely with F-RTO */ |
Ilpo Järvinen | 8512430 | 2007-11-26 20:17:38 +0800 | [diff] [blame] | 84 | if (tp->frto_counter == 2) |
| 85 | tp->frto_counter = 3; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 86 | |
| 87 | tp->packets_out += tcp_skb_pcount(skb); |
Yuchung Cheng | 750ea2b | 2012-05-02 13:30:04 +0000 | [diff] [blame] | 88 | if (!prior_packets || tp->early_retrans_delayed) |
| 89 | tcp_rearm_rto(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 90 | } |
| 91 | |
| 92 | /* SND.NXT, if window was not shrunk. |
| 93 | * If window has been shrunk, what should we make? It is not clear at all. |
| 94 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( |
| 95 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already |
| 96 | * invalid. OK, let's make this for now: |
| 97 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 98 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 99 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 100 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 101 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 102 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 103 | return tp->snd_nxt; |
| 104 | else |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 105 | return tcp_wnd_end(tp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 106 | } |
| 107 | |
| 108 | /* Calculate mss to advertise in SYN segment. |
| 109 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: |
| 110 | * |
| 111 | * 1. It is independent of path mtu. |
| 112 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. |
| 113 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of |
| 114 | * attached devices, because some buggy hosts are confused by |
| 115 | * large MSS. |
| 116 | * 4. We do not make 3, we advertise MSS, calculated from first |
| 117 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. |
| 118 | * This may be overridden via information stored in routing table. |
| 119 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, |
| 120 | * probably even Jumbo". |
| 121 | */ |
| 122 | static __u16 tcp_advertise_mss(struct sock *sk) |
| 123 | { |
| 124 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 125 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 126 | int mss = tp->advmss; |
| 127 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 128 | if (dst) { |
| 129 | unsigned int metric = dst_metric_advmss(dst); |
| 130 | |
| 131 | if (metric < mss) { |
| 132 | mss = metric; |
| 133 | tp->advmss = mss; |
| 134 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 135 | } |
| 136 | |
| 137 | return (__u16)mss; |
| 138 | } |
| 139 | |
| 140 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". |
| 141 | * This is the first part of cwnd validation mechanism. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 142 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 143 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 144 | struct tcp_sock *tp = tcp_sk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 145 | s32 delta = tcp_time_stamp - tp->lsndtime; |
| 146 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); |
| 147 | u32 cwnd = tp->snd_cwnd; |
| 148 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 149 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 150 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 151 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 152 | restart_cwnd = min(restart_cwnd, cwnd); |
| 153 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 154 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 155 | cwnd >>= 1; |
| 156 | tp->snd_cwnd = max(cwnd, restart_cwnd); |
| 157 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 158 | tp->snd_cwnd_used = 0; |
| 159 | } |
| 160 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 161 | /* Congestion state accounting after a packet has been sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 162 | static void tcp_event_data_sent(struct tcp_sock *tp, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 163 | struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 164 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 165 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 166 | const u32 now = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 167 | |
David S. Miller | 35089bb | 2006-06-13 22:33:04 -0700 | [diff] [blame] | 168 | if (sysctl_tcp_slow_start_after_idle && |
| 169 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 170 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | |
| 172 | tp->lsndtime = now; |
| 173 | |
| 174 | /* If it is a reply for ato after last received |
| 175 | * packet, enter pingpong mode. |
| 176 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 177 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
| 178 | icsk->icsk_ack.pingpong = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 179 | } |
| 180 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 181 | /* Account for an ACK we sent. */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 182 | 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] | 183 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 184 | tcp_dec_quickack_mode(sk, pkts); |
| 185 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 186 | } |
| 187 | |
| 188 | /* Determine a window scaling and initial window to offer. |
| 189 | * Based on the assumption that the given amount of space |
| 190 | * will be offered. Store the results in the tp structure. |
| 191 | * NOTE: for smooth operation initial space offering should |
| 192 | * be a multiple of mss if possible. We assume here that mss >= 1. |
| 193 | * This MUST be enforced by all callers. |
| 194 | */ |
| 195 | void tcp_select_initial_window(int __space, __u32 mss, |
| 196 | __u32 *rcv_wnd, __u32 *window_clamp, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 197 | int wscale_ok, __u8 *rcv_wscale, |
| 198 | __u32 init_rcv_wnd) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 199 | { |
| 200 | unsigned int space = (__space < 0 ? 0 : __space); |
| 201 | |
| 202 | /* If no clamp set the clamp to the max possible scaled window */ |
| 203 | if (*window_clamp == 0) |
| 204 | (*window_clamp) = (65535 << 14); |
| 205 | space = min(*window_clamp, space); |
| 206 | |
| 207 | /* Quantize space offering to a multiple of mss if possible. */ |
| 208 | if (space > mss) |
| 209 | space = (space / mss) * mss; |
| 210 | |
| 211 | /* NOTE: offering an initial window larger than 32767 |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 212 | * will break some buggy TCP stacks. If the admin tells us |
| 213 | * it is likely we could be speaking with such a buggy stack |
| 214 | * we will truncate our initial window offering to 32K-1 |
| 215 | * unless the remote has sent us a window scaling option, |
| 216 | * which we interpret as a sign the remote TCP is not |
| 217 | * misinterpreting the window field as a signed quantity. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 218 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 219 | if (sysctl_tcp_workaround_signed_windows) |
| 220 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
| 221 | else |
| 222 | (*rcv_wnd) = space; |
| 223 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 224 | (*rcv_wscale) = 0; |
| 225 | if (wscale_ok) { |
| 226 | /* Set window scaling on max possible window |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 227 | * See RFC1323 for an explanation of the limit to 14 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 228 | */ |
| 229 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); |
Stephen Hemminger | 316c159 | 2006-08-22 00:06:11 -0700 | [diff] [blame] | 230 | space = min_t(u32, space, *window_clamp); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 231 | while (space > 65535 && (*rcv_wscale) < 14) { |
| 232 | space >>= 1; |
| 233 | (*rcv_wscale)++; |
| 234 | } |
| 235 | } |
| 236 | |
Nandita Dukkipati | 356f039 | 2010-12-20 14:15:56 +0000 | [diff] [blame] | 237 | /* Set initial window to a value enough for senders starting with |
| 238 | * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place |
| 239 | * a limit on the initial window when mss is larger than 1460. |
| 240 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 241 | if (mss > (1 << *rcv_wscale)) { |
Nandita Dukkipati | 356f039 | 2010-12-20 14:15:56 +0000 | [diff] [blame] | 242 | int init_cwnd = TCP_DEFAULT_INIT_RCVWND; |
| 243 | if (mss > 1460) |
| 244 | init_cwnd = |
| 245 | max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 246 | /* when initializing use the value from init_rcv_wnd |
| 247 | * rather than the default from above |
| 248 | */ |
Nandita Dukkipati | b1afde6 | 2010-12-03 13:33:44 +0000 | [diff] [blame] | 249 | if (init_rcv_wnd) |
| 250 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); |
| 251 | else |
| 252 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 253 | } |
| 254 | |
| 255 | /* Set the clamp no higher than max representable value */ |
| 256 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); |
| 257 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 258 | EXPORT_SYMBOL(tcp_select_initial_window); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 259 | |
| 260 | /* Chose a new window to advertise, update state in tcp_sock for the |
| 261 | * socket, and return result with RFC1323 scaling applied. The return |
| 262 | * value can be stuffed directly into th->window for an outgoing |
| 263 | * frame. |
| 264 | */ |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 265 | static u16 tcp_select_window(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 266 | { |
| 267 | struct tcp_sock *tp = tcp_sk(sk); |
| 268 | u32 cur_win = tcp_receive_window(tp); |
| 269 | u32 new_win = __tcp_select_window(sk); |
| 270 | |
| 271 | /* Never shrink the offered window */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 272 | if (new_win < cur_win) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 273 | /* Danger Will Robinson! |
| 274 | * Don't update rcv_wup/rcv_wnd here or else |
| 275 | * we will not be able to advertise a zero |
| 276 | * window in time. --DaveM |
| 277 | * |
| 278 | * Relax Will Robinson. |
| 279 | */ |
Patrick McHardy | 607bfbf | 2008-03-20 16:11:27 -0700 | [diff] [blame] | 280 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 281 | } |
| 282 | tp->rcv_wnd = new_win; |
| 283 | tp->rcv_wup = tp->rcv_nxt; |
| 284 | |
| 285 | /* Make sure we do not exceed the maximum possible |
| 286 | * scaled window. |
| 287 | */ |
Rick Jones | 15d99e0 | 2006-03-20 22:40:29 -0800 | [diff] [blame] | 288 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 289 | new_win = min(new_win, MAX_TCP_WINDOW); |
| 290 | else |
| 291 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); |
| 292 | |
| 293 | /* RFC1323 scaling applied */ |
| 294 | new_win >>= tp->rx_opt.rcv_wscale; |
| 295 | |
| 296 | /* If we advertise zero window, disable fast path. */ |
| 297 | if (new_win == 0) |
| 298 | tp->pred_flags = 0; |
| 299 | |
| 300 | return new_win; |
| 301 | } |
| 302 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 303 | /* Packet ECN state for a SYN-ACK */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 304 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 305 | { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 306 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 307 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 308 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 309 | } |
| 310 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 311 | /* Packet ECN state for a SYN. */ |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 312 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
| 313 | { |
| 314 | struct tcp_sock *tp = tcp_sk(sk); |
| 315 | |
| 316 | tp->ecn_flags = 0; |
Hannes Frederic Sowa | 5d134f1 | 2013-01-05 16:10:48 +0000 | [diff] [blame^] | 317 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 318 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 319 | tp->ecn_flags = TCP_ECN_OK; |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | static __inline__ void |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 324 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 325 | { |
| 326 | if (inet_rsk(req)->ecn_ok) |
| 327 | th->ece = 1; |
| 328 | } |
| 329 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 330 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
| 331 | * be sent. |
| 332 | */ |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 333 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
| 334 | int tcp_header_len) |
| 335 | { |
| 336 | struct tcp_sock *tp = tcp_sk(sk); |
| 337 | |
| 338 | if (tp->ecn_flags & TCP_ECN_OK) { |
| 339 | /* Not-retransmitted data segment: set ECT and inject CWR. */ |
| 340 | if (skb->len != tcp_header_len && |
| 341 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { |
| 342 | INET_ECN_xmit(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 343 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
Ilpo Järvinen | bdf1ee5 | 2007-05-27 02:04:16 -0700 | [diff] [blame] | 344 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
| 345 | tcp_hdr(skb)->cwr = 1; |
| 346 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
| 347 | } |
| 348 | } else { |
| 349 | /* ACK or retransmitted segment: clear ECT|CE */ |
| 350 | INET_ECN_dontxmit(sk); |
| 351 | } |
| 352 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) |
| 353 | tcp_hdr(skb)->ece = 1; |
| 354 | } |
| 355 | } |
| 356 | |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 357 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
| 358 | * auto increment end seqno. |
| 359 | */ |
| 360 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) |
| 361 | { |
David S. Miller | 2e8e18e | 2010-04-08 11:32:30 -0700 | [diff] [blame] | 362 | skb->ip_summed = CHECKSUM_PARTIAL; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 363 | skb->csum = 0; |
| 364 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 365 | TCP_SKB_CB(skb)->tcp_flags = flags; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 366 | TCP_SKB_CB(skb)->sacked = 0; |
| 367 | |
| 368 | skb_shinfo(skb)->gso_segs = 1; |
| 369 | skb_shinfo(skb)->gso_size = 0; |
| 370 | skb_shinfo(skb)->gso_type = 0; |
| 371 | |
| 372 | TCP_SKB_CB(skb)->seq = seq; |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 373 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 374 | seq++; |
| 375 | TCP_SKB_CB(skb)->end_seq = seq; |
| 376 | } |
| 377 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 378 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 379 | { |
| 380 | return tp->snd_una != tp->snd_up; |
| 381 | } |
| 382 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 383 | #define OPTION_SACK_ADVERTISE (1 << 0) |
| 384 | #define OPTION_TS (1 << 1) |
| 385 | #define OPTION_MD5 (1 << 2) |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 386 | #define OPTION_WSCALE (1 << 3) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 387 | #define OPTION_COOKIE_EXTENSION (1 << 4) |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 388 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 389 | |
| 390 | struct tcp_out_options { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 391 | u16 options; /* bit field of OPTION_* */ |
| 392 | u16 mss; /* 0 to disable */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 393 | u8 ws; /* window scale, 0 to disable */ |
| 394 | u8 num_sack_blocks; /* number of SACK blocks to include */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 395 | u8 hash_size; /* bytes in hash_location */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 396 | __u8 *hash_location; /* temporary pointer, overloaded */ |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 397 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
| 398 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 399 | }; |
| 400 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 401 | /* The sysctl int routines are generic, so check consistency here. |
| 402 | */ |
| 403 | static u8 tcp_cookie_size_check(u8 desired) |
| 404 | { |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 405 | int cookie_size; |
| 406 | |
| 407 | if (desired > 0) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 408 | /* previously specified */ |
| 409 | return desired; |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 410 | |
| 411 | cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); |
| 412 | if (cookie_size <= 0) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 413 | /* no default specified */ |
| 414 | return 0; |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 415 | |
| 416 | if (cookie_size <= TCP_COOKIE_MIN) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 417 | /* value too small, specify minimum */ |
| 418 | return TCP_COOKIE_MIN; |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 419 | |
| 420 | if (cookie_size >= TCP_COOKIE_MAX) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 421 | /* value too large, specify maximum */ |
| 422 | return TCP_COOKIE_MAX; |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 423 | |
| 424 | if (cookie_size & 1) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 425 | /* 8-bit multiple, illegal, fix it */ |
Eric Dumazet | f198725 | 2010-12-07 12:20:47 +0000 | [diff] [blame] | 426 | cookie_size++; |
| 427 | |
| 428 | return (u8)cookie_size; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 429 | } |
| 430 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 431 | /* Write previously computed TCP options to the packet. |
| 432 | * |
| 433 | * Beware: Something in the Internet is very sensitive to the ordering of |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 434 | * TCP options, we learned this through the hard way, so be careful here. |
| 435 | * Luckily we can at least blame others for their non-compliance but from |
| 436 | * inter-operatibility perspective it seems that we're somewhat stuck with |
| 437 | * the ordering which we have been using if we want to keep working with |
| 438 | * those broken things (not that it currently hurts anybody as there isn't |
| 439 | * particular reason why the ordering would need to be changed). |
| 440 | * |
| 441 | * At least SACK_PERM as the first option is known to lead to a disaster |
| 442 | * (but it may well be that other scenarios fail similarly). |
| 443 | */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 444 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 445 | struct tcp_out_options *opts) |
| 446 | { |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 447 | u16 options = opts->options; /* mungable copy */ |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 448 | |
| 449 | /* Having both authentication and cookies for security is redundant, |
| 450 | * and there's certainly not enough room. Instead, the cookie-less |
| 451 | * extension variant is proposed. |
| 452 | * |
| 453 | * Consider the pessimal case with authentication. The options |
| 454 | * could look like: |
| 455 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 |
| 456 | */ |
| 457 | if (unlikely(OPTION_MD5 & options)) { |
| 458 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { |
| 459 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | |
| 460 | (TCPOLEN_COOKIE_BASE << 16) | |
| 461 | (TCPOPT_MD5SIG << 8) | |
| 462 | TCPOLEN_MD5SIG); |
| 463 | } else { |
| 464 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 465 | (TCPOPT_NOP << 16) | |
| 466 | (TCPOPT_MD5SIG << 8) | |
| 467 | TCPOLEN_MD5SIG); |
| 468 | } |
| 469 | options &= ~OPTION_COOKIE_EXTENSION; |
| 470 | /* overload cookie hash location */ |
| 471 | opts->hash_location = (__u8 *)ptr; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 472 | ptr += 4; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 473 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 474 | |
Ilpo Järvinen | fd6149d | 2008-10-23 14:06:35 -0700 | [diff] [blame] | 475 | if (unlikely(opts->mss)) { |
| 476 | *ptr++ = htonl((TCPOPT_MSS << 24) | |
| 477 | (TCPOLEN_MSS << 16) | |
| 478 | opts->mss); |
| 479 | } |
| 480 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 481 | if (likely(OPTION_TS & options)) { |
| 482 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 483 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
| 484 | (TCPOLEN_SACK_PERM << 16) | |
| 485 | (TCPOPT_TIMESTAMP << 8) | |
| 486 | TCPOLEN_TIMESTAMP); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 487 | options &= ~OPTION_SACK_ADVERTISE; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 488 | } else { |
| 489 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 490 | (TCPOPT_NOP << 16) | |
| 491 | (TCPOPT_TIMESTAMP << 8) | |
| 492 | TCPOLEN_TIMESTAMP); |
| 493 | } |
| 494 | *ptr++ = htonl(opts->tsval); |
| 495 | *ptr++ = htonl(opts->tsecr); |
| 496 | } |
| 497 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 498 | /* Specification requires after timestamp, so do it now. |
| 499 | * |
| 500 | * Consider the pessimal case without authentication. The options |
| 501 | * could look like: |
| 502 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 |
| 503 | */ |
| 504 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { |
| 505 | __u8 *cookie_copy = opts->hash_location; |
| 506 | u8 cookie_size = opts->hash_size; |
| 507 | |
| 508 | /* 8-bit multiple handled in tcp_cookie_size_check() above, |
| 509 | * and elsewhere. |
| 510 | */ |
| 511 | if (0x2 & cookie_size) { |
| 512 | __u8 *p = (__u8 *)ptr; |
| 513 | |
| 514 | /* 16-bit multiple */ |
| 515 | *p++ = TCPOPT_COOKIE; |
| 516 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; |
| 517 | *p++ = *cookie_copy++; |
| 518 | *p++ = *cookie_copy++; |
| 519 | ptr++; |
| 520 | cookie_size -= 2; |
| 521 | } else { |
| 522 | /* 32-bit multiple */ |
| 523 | *ptr++ = htonl(((TCPOPT_NOP << 24) | |
| 524 | (TCPOPT_NOP << 16) | |
| 525 | (TCPOPT_COOKIE << 8) | |
| 526 | TCPOLEN_COOKIE_BASE) + |
| 527 | cookie_size); |
| 528 | } |
| 529 | |
| 530 | if (cookie_size > 0) { |
| 531 | memcpy(ptr, cookie_copy, cookie_size); |
| 532 | ptr += (cookie_size / 4); |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 537 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 538 | (TCPOPT_NOP << 16) | |
| 539 | (TCPOPT_SACK_PERM << 8) | |
| 540 | TCPOLEN_SACK_PERM); |
| 541 | } |
| 542 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 543 | if (unlikely(OPTION_WSCALE & options)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 544 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 545 | (TCPOPT_WINDOW << 16) | |
| 546 | (TCPOLEN_WINDOW << 8) | |
| 547 | opts->ws); |
| 548 | } |
| 549 | |
| 550 | if (unlikely(opts->num_sack_blocks)) { |
| 551 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? |
| 552 | tp->duplicate_sack : tp->selective_acks; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 553 | int this_sack; |
| 554 | |
| 555 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
| 556 | (TCPOPT_NOP << 16) | |
| 557 | (TCPOPT_SACK << 8) | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 558 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 559 | TCPOLEN_SACK_PERBLOCK))); |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 560 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 561 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
| 562 | ++this_sack) { |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 563 | *ptr++ = htonl(sp[this_sack].start_seq); |
| 564 | *ptr++ = htonl(sp[this_sack].end_seq); |
| 565 | } |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 566 | |
Ilpo Järvinen | 5861f8e | 2009-03-14 14:23:01 +0000 | [diff] [blame] | 567 | tp->rx_opt.dsack = 0; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 568 | } |
Yuchung Cheng | 2100c8d | 2012-07-19 06:43:05 +0000 | [diff] [blame] | 569 | |
| 570 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { |
| 571 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; |
| 572 | |
| 573 | *ptr++ = htonl((TCPOPT_EXP << 24) | |
| 574 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | |
| 575 | TCPOPT_FASTOPEN_MAGIC); |
| 576 | |
| 577 | memcpy(ptr, foc->val, foc->len); |
| 578 | if ((foc->len & 3) == 2) { |
| 579 | u8 *align = ((u8 *)ptr) + foc->len; |
| 580 | align[0] = align[1] = TCPOPT_NOP; |
| 581 | } |
| 582 | ptr += (foc->len + 3) >> 2; |
| 583 | } |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 584 | } |
| 585 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 586 | /* Compute TCP options for SYN packets. This is not the final |
| 587 | * network wire format yet. |
| 588 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 589 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 590 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 591 | struct tcp_md5sig_key **md5) |
| 592 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 593 | struct tcp_sock *tp = tcp_sk(sk); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 594 | struct tcp_cookie_values *cvp = tp->cookie_values; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 595 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 596 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? |
| 597 | tcp_cookie_size_check(cvp->cookie_desired) : |
| 598 | 0; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 599 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 600 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 601 | #ifdef CONFIG_TCP_MD5SIG |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 602 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 603 | if (*md5) { |
| 604 | opts->options |= OPTION_MD5; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 605 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 606 | } |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 607 | #else |
| 608 | *md5 = NULL; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 609 | #endif |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 610 | |
| 611 | /* We always get an MSS option. The option bytes which will be seen in |
| 612 | * normal data packets should timestamps be used, must be in the MSS |
| 613 | * advertised. But we subtract them from tp->mss_cache so that |
| 614 | * calculations in tcp_sendmsg are simpler etc. So account for this |
| 615 | * fact here if necessary. If we don't do this correctly, as a |
| 616 | * receiver we won't recognize data packets as being full sized when we |
| 617 | * should, and thus we won't abide by the delayed ACK rules correctly. |
| 618 | * SACKs don't matter, we never delay an ACK when we have any of those |
| 619 | * going out. */ |
| 620 | opts->mss = tcp_advertise_mss(sk); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 621 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 622 | |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 623 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 624 | opts->options |= OPTION_TS; |
| 625 | opts->tsval = TCP_SKB_CB(skb)->when; |
| 626 | opts->tsecr = tp->rx_opt.ts_recent; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 627 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 628 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 629 | if (likely(sysctl_tcp_window_scaling)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 630 | opts->ws = tp->rx_opt.rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 631 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 632 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 633 | } |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 634 | if (likely(sysctl_tcp_sack)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 635 | opts->options |= OPTION_SACK_ADVERTISE; |
David S. Miller | b32d131 | 2008-07-21 18:45:34 -0700 | [diff] [blame] | 636 | if (unlikely(!(OPTION_TS & opts->options))) |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 637 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 638 | } |
| 639 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 640 | if (fastopen && fastopen->cookie.len >= 0) { |
| 641 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; |
| 642 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 643 | if (remaining >= need) { |
| 644 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 645 | opts->fastopen_cookie = &fastopen->cookie; |
| 646 | remaining -= need; |
| 647 | tp->syn_fastopen = 1; |
| 648 | } |
| 649 | } |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 650 | /* Note that timestamps are required by the specification. |
| 651 | * |
| 652 | * Odd numbers of bytes are prohibited by the specification, ensuring |
| 653 | * that the cookie is 16-bit aligned, and the resulting cookie pair is |
| 654 | * 32-bit aligned. |
| 655 | */ |
| 656 | if (*md5 == NULL && |
| 657 | (OPTION_TS & opts->options) && |
| 658 | cookie_size > 0) { |
| 659 | int need = TCPOLEN_COOKIE_BASE + cookie_size; |
| 660 | |
| 661 | if (0x2 & need) { |
| 662 | /* 32-bit multiple */ |
| 663 | need += 2; /* NOPs */ |
| 664 | |
| 665 | if (need > remaining) { |
| 666 | /* try shrinking cookie to fit */ |
| 667 | cookie_size -= 2; |
| 668 | need -= 4; |
| 669 | } |
| 670 | } |
| 671 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { |
| 672 | cookie_size -= 4; |
| 673 | need -= 4; |
| 674 | } |
| 675 | if (TCP_COOKIE_MIN <= cookie_size) { |
| 676 | opts->options |= OPTION_COOKIE_EXTENSION; |
| 677 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; |
| 678 | opts->hash_size = cookie_size; |
| 679 | |
| 680 | /* Remember for future incarnations. */ |
| 681 | cvp->cookie_desired = cookie_size; |
| 682 | |
| 683 | if (cvp->cookie_desired != cvp->cookie_pair_size) { |
| 684 | /* Currently use random bytes as a nonce, |
| 685 | * assuming these are completely unpredictable |
| 686 | * by hostile users of the same system. |
| 687 | */ |
| 688 | get_random_bytes(&cvp->cookie_pair[0], |
| 689 | cookie_size); |
| 690 | cvp->cookie_pair_size = cookie_size; |
| 691 | } |
| 692 | |
| 693 | remaining -= need; |
| 694 | } |
| 695 | } |
| 696 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 697 | } |
| 698 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 699 | /* Set up TCP options for SYN-ACKs. */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 700 | static unsigned int tcp_synack_options(struct sock *sk, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 701 | struct request_sock *req, |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 702 | unsigned int mss, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 703 | struct tcp_out_options *opts, |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 704 | struct tcp_md5sig_key **md5, |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 705 | struct tcp_extend_values *xvp, |
| 706 | struct tcp_fastopen_cookie *foc) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 707 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 708 | struct inet_request_sock *ireq = inet_rsk(req); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 709 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 710 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? |
| 711 | xvp->cookie_plus : |
| 712 | 0; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 713 | |
| 714 | #ifdef CONFIG_TCP_MD5SIG |
| 715 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
| 716 | if (*md5) { |
| 717 | opts->options |= OPTION_MD5; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 718 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
| 719 | |
| 720 | /* We can't fit any SACK blocks in a packet with MD5 + TS |
| 721 | * options. There was discussion about disabling SACK |
| 722 | * rather than TS in order to fit in better with old, |
| 723 | * buggy kernels, but that was deemed to be unnecessary. |
| 724 | */ |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 725 | ireq->tstamp_ok &= !ireq->sack_ok; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 726 | } |
| 727 | #else |
| 728 | *md5 = NULL; |
| 729 | #endif |
| 730 | |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 731 | /* We always send an MSS option. */ |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 732 | opts->mss = mss; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 733 | remaining -= TCPOLEN_MSS_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 734 | |
| 735 | if (likely(ireq->wscale_ok)) { |
| 736 | opts->ws = ireq->rcv_wscale; |
Ori Finkelman | 89e95a6 | 2009-10-01 06:41:59 +0000 | [diff] [blame] | 737 | opts->options |= OPTION_WSCALE; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 738 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 739 | } |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 740 | if (likely(ireq->tstamp_ok)) { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 741 | opts->options |= OPTION_TS; |
| 742 | opts->tsval = TCP_SKB_CB(skb)->when; |
| 743 | opts->tsecr = req->ts_recent; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 744 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 745 | } |
| 746 | if (likely(ireq->sack_ok)) { |
| 747 | opts->options |= OPTION_SACK_ADVERTISE; |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 748 | if (unlikely(!ireq->tstamp_ok)) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 749 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 750 | } |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 751 | if (foc != NULL) { |
| 752 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; |
| 753 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
| 754 | if (remaining >= need) { |
| 755 | opts->options |= OPTION_FAST_OPEN_COOKIE; |
| 756 | opts->fastopen_cookie = foc; |
| 757 | remaining -= need; |
| 758 | } |
| 759 | } |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 760 | /* Similar rationale to tcp_syn_options() applies here, too. |
| 761 | * If the <SYN> options fit, the same options should fit now! |
| 762 | */ |
| 763 | if (*md5 == NULL && |
Eric Dumazet | de213e5e | 2010-05-17 22:35:36 -0700 | [diff] [blame] | 764 | ireq->tstamp_ok && |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 765 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
| 766 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ |
| 767 | |
| 768 | if (0x2 & need) { |
| 769 | /* 32-bit multiple */ |
| 770 | need += 2; /* NOPs */ |
| 771 | } |
| 772 | if (need <= remaining) { |
| 773 | opts->options |= OPTION_COOKIE_EXTENSION; |
| 774 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; |
| 775 | remaining -= need; |
| 776 | } else { |
| 777 | /* There's no error return, so flag it. */ |
| 778 | xvp->cookie_out_never = 1; /* true */ |
| 779 | opts->hash_size = 0; |
| 780 | } |
| 781 | } |
| 782 | return MAX_TCP_OPTION_SPACE - remaining; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 783 | } |
| 784 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 785 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
| 786 | * final wire format yet. |
| 787 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 788 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 789 | struct tcp_out_options *opts, |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 790 | struct tcp_md5sig_key **md5) |
| 791 | { |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 792 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
| 793 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 794 | unsigned int size = 0; |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 795 | unsigned int eff_sacks; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 796 | |
| 797 | #ifdef CONFIG_TCP_MD5SIG |
| 798 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
| 799 | if (unlikely(*md5)) { |
| 800 | opts->options |= OPTION_MD5; |
| 801 | size += TCPOLEN_MD5SIG_ALIGNED; |
| 802 | } |
| 803 | #else |
| 804 | *md5 = NULL; |
| 805 | #endif |
| 806 | |
| 807 | if (likely(tp->rx_opt.tstamp_ok)) { |
| 808 | opts->options |= OPTION_TS; |
| 809 | opts->tsval = tcb ? tcb->when : 0; |
| 810 | opts->tsecr = tp->rx_opt.ts_recent; |
| 811 | size += TCPOLEN_TSTAMP_ALIGNED; |
| 812 | } |
| 813 | |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 814 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
| 815 | if (unlikely(eff_sacks)) { |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 816 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 817 | opts->num_sack_blocks = |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 818 | min_t(unsigned int, eff_sacks, |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 819 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
| 820 | TCPOLEN_SACK_PERBLOCK); |
| 821 | size += TCPOLEN_SACK_BASE_ALIGNED + |
| 822 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; |
| 823 | } |
| 824 | |
| 825 | return size; |
Stephen Hemminger | 40efc6f | 2006-01-03 16:03:49 -0800 | [diff] [blame] | 826 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 827 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 828 | |
| 829 | /* TCP SMALL QUEUES (TSQ) |
| 830 | * |
| 831 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) |
| 832 | * to reduce RTT and bufferbloat. |
| 833 | * We do this using a special skb destructor (tcp_wfree). |
| 834 | * |
| 835 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb |
| 836 | * needs to be reallocated in a driver. |
| 837 | * The invariant being skb->truesize substracted from sk->sk_wmem_alloc |
| 838 | * |
| 839 | * Since transmit from skb destructor is forbidden, we use a tasklet |
| 840 | * to process all sockets that eventually need to send more skbs. |
| 841 | * We use one tasklet per cpu, with its own queue of sockets. |
| 842 | */ |
| 843 | struct tsq_tasklet { |
| 844 | struct tasklet_struct tasklet; |
| 845 | struct list_head head; /* queue of tcp sockets */ |
| 846 | }; |
| 847 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); |
| 848 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 849 | static void tcp_tsq_handler(struct sock *sk) |
| 850 | { |
| 851 | if ((1 << sk->sk_state) & |
| 852 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | |
| 853 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) |
| 854 | tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); |
| 855 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 856 | /* |
| 857 | * One tasklest per cpu tries to send more skbs. |
| 858 | * We run in tasklet context but need to disable irqs when |
| 859 | * transfering tsq->head because tcp_wfree() might |
| 860 | * interrupt us (non NAPI drivers) |
| 861 | */ |
| 862 | static void tcp_tasklet_func(unsigned long data) |
| 863 | { |
| 864 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; |
| 865 | LIST_HEAD(list); |
| 866 | unsigned long flags; |
| 867 | struct list_head *q, *n; |
| 868 | struct tcp_sock *tp; |
| 869 | struct sock *sk; |
| 870 | |
| 871 | local_irq_save(flags); |
| 872 | list_splice_init(&tsq->head, &list); |
| 873 | local_irq_restore(flags); |
| 874 | |
| 875 | list_for_each_safe(q, n, &list) { |
| 876 | tp = list_entry(q, struct tcp_sock, tsq_node); |
| 877 | list_del(&tp->tsq_node); |
| 878 | |
| 879 | sk = (struct sock *)tp; |
| 880 | bh_lock_sock(sk); |
| 881 | |
| 882 | if (!sock_owned_by_user(sk)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 883 | tcp_tsq_handler(sk); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 884 | } else { |
| 885 | /* defer the work to tcp_release_cb() */ |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 886 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 887 | } |
| 888 | bh_unlock_sock(sk); |
| 889 | |
| 890 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); |
| 891 | sk_free(sk); |
| 892 | } |
| 893 | } |
| 894 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 895 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
| 896 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ |
Eric Dumazet | 563d34d | 2012-07-23 09:48:52 +0200 | [diff] [blame] | 897 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
| 898 | (1UL << TCP_MTU_REDUCED_DEFERRED)) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 899 | /** |
| 900 | * tcp_release_cb - tcp release_sock() callback |
| 901 | * @sk: socket |
| 902 | * |
| 903 | * called from release_sock() to perform protocol dependent |
| 904 | * actions before socket release. |
| 905 | */ |
| 906 | void tcp_release_cb(struct sock *sk) |
| 907 | { |
| 908 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 909 | unsigned long flags, nflags; |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 910 | |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 911 | /* perform an atomic operation only if at least one flag is set */ |
| 912 | do { |
| 913 | flags = tp->tsq_flags; |
| 914 | if (!(flags & TCP_DEFERRED_ALL)) |
| 915 | return; |
| 916 | nflags = flags & ~TCP_DEFERRED_ALL; |
| 917 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); |
| 918 | |
| 919 | if (flags & (1UL << TCP_TSQ_DEFERRED)) |
| 920 | tcp_tsq_handler(sk); |
| 921 | |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 922 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 923 | tcp_write_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 924 | __sock_put(sk); |
| 925 | } |
| 926 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { |
Eric Dumazet | 6f458df | 2012-07-20 05:45:50 +0000 | [diff] [blame] | 927 | tcp_delack_timer_handler(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 928 | __sock_put(sk); |
| 929 | } |
| 930 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { |
Eric Dumazet | 563d34d | 2012-07-23 09:48:52 +0200 | [diff] [blame] | 931 | sk->sk_prot->mtu_reduced(sk); |
Eric Dumazet | 144d56e | 2012-08-20 00:22:46 +0000 | [diff] [blame] | 932 | __sock_put(sk); |
| 933 | } |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 934 | } |
| 935 | EXPORT_SYMBOL(tcp_release_cb); |
| 936 | |
| 937 | void __init tcp_tasklet_init(void) |
| 938 | { |
| 939 | int i; |
| 940 | |
| 941 | for_each_possible_cpu(i) { |
| 942 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); |
| 943 | |
| 944 | INIT_LIST_HEAD(&tsq->head); |
| 945 | tasklet_init(&tsq->tasklet, |
| 946 | tcp_tasklet_func, |
| 947 | (unsigned long)tsq); |
| 948 | } |
| 949 | } |
| 950 | |
| 951 | /* |
| 952 | * Write buffer destructor automatically called from kfree_skb. |
| 953 | * We cant xmit new skbs from this context, as we might already |
| 954 | * hold qdisc lock. |
| 955 | */ |
Silviu-Mihai Popescu | 8e7dfbc | 2012-08-04 09:31:29 +0000 | [diff] [blame] | 956 | static void tcp_wfree(struct sk_buff *skb) |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 957 | { |
| 958 | struct sock *sk = skb->sk; |
| 959 | struct tcp_sock *tp = tcp_sk(sk); |
| 960 | |
| 961 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && |
| 962 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { |
| 963 | unsigned long flags; |
| 964 | struct tsq_tasklet *tsq; |
| 965 | |
| 966 | /* Keep a ref on socket. |
| 967 | * This last ref will be released in tcp_tasklet_func() |
| 968 | */ |
| 969 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); |
| 970 | |
| 971 | /* queue this socket to tasklet queue */ |
| 972 | local_irq_save(flags); |
| 973 | tsq = &__get_cpu_var(tsq_tasklet); |
| 974 | list_add(&tp->tsq_node, &tsq->head); |
| 975 | tasklet_schedule(&tsq->tasklet); |
| 976 | local_irq_restore(flags); |
| 977 | } else { |
| 978 | sock_wfree(skb); |
| 979 | } |
| 980 | } |
| 981 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 982 | /* This routine actually transmits TCP packets queued in by |
| 983 | * tcp_do_sendmsg(). This is used by both the initial |
| 984 | * transmission and possible later retransmissions. |
| 985 | * All SKB's seen here are completely headerless. It is our |
| 986 | * job to build the TCP header, and pass the packet down to |
| 987 | * IP so it can do the same plus pass the packet off to the |
| 988 | * device. |
| 989 | * |
| 990 | * We are working here with either a clone of the original |
| 991 | * SKB, or a fresh unique copy made by the retransmit engine. |
| 992 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 993 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
| 994 | gfp_t gfp_mask) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 995 | { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 996 | const struct inet_connection_sock *icsk = inet_csk(sk); |
| 997 | struct inet_sock *inet; |
| 998 | struct tcp_sock *tp; |
| 999 | struct tcp_skb_cb *tcb; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1000 | struct tcp_out_options opts; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1001 | unsigned int tcp_options_size, tcp_header_size; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1002 | struct tcp_md5sig_key *md5; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1003 | struct tcphdr *th; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1004 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1005 | |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1006 | BUG_ON(!skb || !tcp_skb_pcount(skb)); |
| 1007 | |
| 1008 | /* If congestion control is doing timestamping, we must |
| 1009 | * take such a timestamp before we potentially clone/copy. |
| 1010 | */ |
Stephen Hemminger | 164891a | 2007-04-23 22:26:16 -0700 | [diff] [blame] | 1011 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1012 | __net_timestamp(skb); |
| 1013 | |
| 1014 | if (likely(clone_it)) { |
| 1015 | if (unlikely(skb_cloned(skb))) |
| 1016 | skb = pskb_copy(skb, gfp_mask); |
| 1017 | else |
| 1018 | skb = skb_clone(skb, gfp_mask); |
| 1019 | if (unlikely(!skb)) |
| 1020 | return -ENOBUFS; |
| 1021 | } |
| 1022 | |
| 1023 | inet = inet_sk(sk); |
| 1024 | tp = tcp_sk(sk); |
| 1025 | tcb = TCP_SKB_CB(skb); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1026 | memset(&opts, 0, sizeof(opts)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1027 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1028 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1029 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
| 1030 | else |
| 1031 | tcp_options_size = tcp_established_options(sk, skb, &opts, |
| 1032 | &md5); |
| 1033 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1034 | |
Tom Herbert | 3853b58 | 2010-11-21 13:17:29 +0000 | [diff] [blame] | 1035 | if (tcp_packets_in_flight(tp) == 0) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1036 | tcp_ca_event(sk, CA_EVENT_TX_START); |
Tom Herbert | 3853b58 | 2010-11-21 13:17:29 +0000 | [diff] [blame] | 1037 | skb->ooo_okay = 1; |
| 1038 | } else |
| 1039 | skb->ooo_okay = 0; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1040 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 1041 | skb_push(skb, tcp_header_size); |
| 1042 | skb_reset_transport_header(skb); |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 1043 | |
| 1044 | skb_orphan(skb); |
| 1045 | skb->sk = sk; |
| 1046 | skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? |
| 1047 | tcp_wfree : sock_wfree; |
| 1048 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1049 | |
| 1050 | /* Build TCP header and checksum it. */ |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 1051 | th = tcp_hdr(skb); |
Eric Dumazet | c720c7e | 2009-10-15 06:30:45 +0000 | [diff] [blame] | 1052 | th->source = inet->inet_sport; |
| 1053 | th->dest = inet->inet_dport; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1054 | th->seq = htonl(tcb->seq); |
| 1055 | th->ack_seq = htonl(tp->rcv_nxt); |
Al Viro | df7a3b0 | 2006-09-27 18:38:52 -0700 | [diff] [blame] | 1056 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1057 | tcb->tcp_flags); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1058 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1059 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1060 | /* RFC1323: The window in SYN & SYN/ACK segments |
| 1061 | * is never scaled. |
| 1062 | */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 1063 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1064 | } else { |
| 1065 | th->window = htons(tcp_select_window(sk)); |
| 1066 | } |
| 1067 | th->check = 0; |
| 1068 | th->urg_ptr = 0; |
| 1069 | |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 1070 | /* The urg_mode check is necessary during a below snd_una win probe */ |
Herbert Xu | 7691367 | 2009-02-21 23:52:29 -0800 | [diff] [blame] | 1071 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
| 1072 | if (before(tp->snd_up, tcb->seq + 0x10000)) { |
| 1073 | th->urg_ptr = htons(tp->snd_up - tcb->seq); |
| 1074 | th->urg = 1; |
| 1075 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { |
Eric Dumazet | 0eae88f | 2010-04-20 19:06:52 -0700 | [diff] [blame] | 1076 | th->urg_ptr = htons(0xFFFF); |
Herbert Xu | 7691367 | 2009-02-21 23:52:29 -0800 | [diff] [blame] | 1077 | th->urg = 1; |
| 1078 | } |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1079 | } |
| 1080 | |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 1081 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1082 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1083 | TCP_ECN_send(sk, skb, tcp_header_size); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1084 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1085 | #ifdef CONFIG_TCP_MD5SIG |
| 1086 | /* Calculate the MD5 hash, as we have all we need now */ |
| 1087 | if (md5) { |
Eric Dumazet | a465419 | 2010-05-16 00:36:33 -0700 | [diff] [blame] | 1088 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 1089 | tp->af_specific->calc_md5_hash(opts.hash_location, |
Adam Langley | 49a72df | 2008-07-19 00:01:42 -0700 | [diff] [blame] | 1090 | md5, sk, NULL, skb); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1091 | } |
| 1092 | #endif |
| 1093 | |
Herbert Xu | bb29624 | 2010-04-11 02:15:55 +0000 | [diff] [blame] | 1094 | icsk->icsk_af_ops->send_check(sk, skb); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1095 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1096 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1097 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
| 1098 | |
| 1099 | if (skb->len != tcp_header_size) |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1100 | tcp_event_data_sent(tp, sk); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1101 | |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 1102 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
Tom Herbert | aa2ea05 | 2010-04-22 07:00:24 +0000 | [diff] [blame] | 1103 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
| 1104 | tcp_skb_pcount(skb)); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1105 | |
David S. Miller | d9d8da8 | 2011-05-06 22:23:20 -0700 | [diff] [blame] | 1106 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
Hua Zhong | 83de47c | 2006-04-28 15:26:50 -0700 | [diff] [blame] | 1107 | if (likely(err <= 0)) |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1108 | return err; |
| 1109 | |
Ilpo Järvinen | 3cfe3ba | 2007-02-27 10:09:49 -0800 | [diff] [blame] | 1110 | tcp_enter_cwr(sk, 1); |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 1111 | |
Gerrit Renker | b9df3cb | 2006-11-14 11:21:36 -0200 | [diff] [blame] | 1112 | return net_xmit_eval(err); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1113 | } |
| 1114 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1115 | /* This routine just queues the buffer for sending. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1116 | * |
| 1117 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, |
| 1118 | * otherwise socket can stall. |
| 1119 | */ |
| 1120 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 1121 | { |
| 1122 | struct tcp_sock *tp = tcp_sk(sk); |
| 1123 | |
| 1124 | /* Advance write_seq and place onto the write_queue. */ |
| 1125 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; |
| 1126 | skb_header_release(skb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1127 | tcp_add_write_queue_tail(sk, skb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1128 | sk->sk_wmem_queued += skb->truesize; |
| 1129 | sk_mem_charge(sk, skb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1130 | } |
| 1131 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1132 | /* Initialize TSO segments for a packet. */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1133 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1134 | unsigned int mss_now) |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1135 | { |
Herbert Xu | 8e5b9dd | 2009-06-28 18:03:30 +0000 | [diff] [blame] | 1136 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
| 1137 | skb->ip_summed == CHECKSUM_NONE) { |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1138 | /* Avoid the costly divide in the normal |
| 1139 | * non-TSO case. |
| 1140 | */ |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 1141 | skb_shinfo(skb)->gso_segs = 1; |
| 1142 | skb_shinfo(skb)->gso_size = 0; |
| 1143 | skb_shinfo(skb)->gso_type = 0; |
David S. Miller | f6302d1 | 2005-07-05 15:18:03 -0700 | [diff] [blame] | 1144 | } else { |
Ilpo Järvinen | 356f89e | 2007-08-24 23:00:31 -0700 | [diff] [blame] | 1145 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
Herbert Xu | 7967168 | 2006-06-22 02:40:14 -0700 | [diff] [blame] | 1146 | skb_shinfo(skb)->gso_size = mss_now; |
Herbert Xu | bcd7611 | 2006-06-30 13:36:35 -0700 | [diff] [blame] | 1147 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1148 | } |
| 1149 | } |
| 1150 | |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1151 | /* When a modification to fackets out becomes necessary, we need to check |
Ilpo Järvinen | 68f8353 | 2007-11-15 19:50:37 -0800 | [diff] [blame] | 1152 | * skb is counted to fackets_out or not. |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1153 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1154 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1155 | int decr) |
| 1156 | { |
Ilpo Järvinen | a47e5a9 | 2007-11-15 19:41:46 -0800 | [diff] [blame] | 1157 | struct tcp_sock *tp = tcp_sk(sk); |
| 1158 | |
Ilpo Järvinen | dc86967 | 2007-10-01 15:27:19 -0700 | [diff] [blame] | 1159 | if (!tp->sacked_out || tcp_is_reno(tp)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1160 | return; |
| 1161 | |
Ilpo Järvinen | 6859d49 | 2007-12-02 00:48:06 +0200 | [diff] [blame] | 1162 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1163 | tp->fackets_out -= decr; |
Ilpo Järvinen | 91fed7a | 2007-10-09 01:24:15 -0700 | [diff] [blame] | 1164 | } |
| 1165 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1166 | /* Pcount in the middle of the write queue got changed, we need to do various |
| 1167 | * tweaks to fix counters |
| 1168 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1169 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1170 | { |
| 1171 | struct tcp_sock *tp = tcp_sk(sk); |
| 1172 | |
| 1173 | tp->packets_out -= decr; |
| 1174 | |
| 1175 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
| 1176 | tp->sacked_out -= decr; |
| 1177 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) |
| 1178 | tp->retrans_out -= decr; |
| 1179 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
| 1180 | tp->lost_out -= decr; |
| 1181 | |
| 1182 | /* Reno case is special. Sigh... */ |
| 1183 | if (tcp_is_reno(tp) && decr > 0) |
| 1184 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); |
| 1185 | |
| 1186 | tcp_adjust_fackets_out(sk, skb, decr); |
| 1187 | |
| 1188 | if (tp->lost_skb_hint && |
| 1189 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && |
Ilpo Järvinen | 52cf3cc | 2009-04-18 05:48:48 +0000 | [diff] [blame] | 1190 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1191 | tp->lost_cnt_hint -= decr; |
| 1192 | |
| 1193 | tcp_verify_left_out(tp); |
| 1194 | } |
| 1195 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1196 | /* Function to create two new TCP segments. Shrinks the given segment |
| 1197 | * 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] | 1198 | * packet to the list. This won't be called frequently, I hope. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1199 | * Remember, these are still headerless SKBs at this point. |
| 1200 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1201 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
| 1202 | unsigned int mss_now) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1203 | { |
| 1204 | struct tcp_sock *tp = tcp_sk(sk); |
| 1205 | struct sk_buff *buff; |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1206 | int nsize, old_factor; |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1207 | int nlen; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1208 | u8 flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1209 | |
Ilpo Järvinen | 2fceec1 | 2011-04-01 21:47:41 -0700 | [diff] [blame] | 1210 | if (WARN_ON(len > skb->len)) |
| 1211 | return -EINVAL; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 1212 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1213 | nsize = skb_headlen(skb) - len; |
| 1214 | if (nsize < 0) |
| 1215 | nsize = 0; |
| 1216 | |
| 1217 | if (skb_cloned(skb) && |
| 1218 | skb_is_nonlinear(skb) && |
| 1219 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
| 1220 | return -ENOMEM; |
| 1221 | |
| 1222 | /* Get a new skb... force flag on. */ |
| 1223 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); |
| 1224 | if (buff == NULL) |
| 1225 | return -ENOMEM; /* We'll just try again later. */ |
Herbert Xu | ef5cb97 | 2006-04-18 13:24:14 -0700 | [diff] [blame] | 1226 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1227 | sk->sk_wmem_queued += buff->truesize; |
| 1228 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1229 | nlen = skb->len - len - nsize; |
| 1230 | buff->truesize += nlen; |
| 1231 | skb->truesize -= nlen; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1232 | |
| 1233 | /* Correct the sequence numbers. */ |
| 1234 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1235 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1236 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1237 | |
| 1238 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1239 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1240 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1241 | TCP_SKB_CB(buff)->tcp_flags = flags; |
Herbert Xu | e14c3ca | 2005-09-19 18:18:38 -0700 | [diff] [blame] | 1242 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1243 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1244 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1245 | /* Copy and checksum data tail into the new buffer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1246 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
| 1247 | skb_put(buff, nsize), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1248 | nsize, 0); |
| 1249 | |
| 1250 | skb_trim(skb, len); |
| 1251 | |
| 1252 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); |
| 1253 | } else { |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1254 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1255 | skb_split(skb, buff, len); |
| 1256 | } |
| 1257 | |
| 1258 | buff->ip_summed = skb->ip_summed; |
| 1259 | |
| 1260 | /* Looks stupid, but our code really uses when of |
| 1261 | * skbs, which it never sent before. --ANK |
| 1262 | */ |
| 1263 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; |
Patrick McHardy | a61bbcf | 2005-08-14 17:24:31 -0700 | [diff] [blame] | 1264 | buff->tstamp = skb->tstamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1265 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1266 | old_factor = tcp_skb_pcount(skb); |
| 1267 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1268 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1269 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1270 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1271 | |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1272 | /* If this packet has been sent out already, we must |
| 1273 | * adjust the various packet counters. |
| 1274 | */ |
Herbert Xu | cf0b450 | 2005-09-08 15:10:52 -0700 | [diff] [blame] | 1275 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
David S. Miller | 6475be1 | 2005-09-01 22:47:01 -0700 | [diff] [blame] | 1276 | int diff = old_factor - tcp_skb_pcount(skb) - |
| 1277 | tcp_skb_pcount(buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1278 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 1279 | if (diff) |
| 1280 | tcp_adjust_pcount(sk, skb, diff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1281 | } |
| 1282 | |
| 1283 | /* Link BUFF into the send queue. */ |
David S. Miller | f44b527 | 2005-07-05 15:18:34 -0700 | [diff] [blame] | 1284 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1285 | tcp_insert_write_queue_after(skb, buff, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1286 | |
| 1287 | return 0; |
| 1288 | } |
| 1289 | |
| 1290 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c |
| 1291 | * eventually). The difference is that pulled data not copied, but |
| 1292 | * immediately discarded. |
| 1293 | */ |
Herbert Xu ~{PmVHI~} | f291196 | 2006-06-05 15:03:37 -0700 | [diff] [blame] | 1294 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1295 | { |
| 1296 | int i, k, eat; |
| 1297 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1298 | eat = min_t(int, len, skb_headlen(skb)); |
| 1299 | if (eat) { |
| 1300 | __skb_pull(skb, eat); |
Eric Dumazet | 22b4a4f | 2012-04-18 10:14:23 +0000 | [diff] [blame] | 1301 | skb->avail_size -= eat; |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1302 | len -= eat; |
| 1303 | if (!len) |
| 1304 | return; |
| 1305 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1306 | eat = len; |
| 1307 | k = 0; |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1308 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1309 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
| 1310 | |
| 1311 | if (size <= eat) { |
Ian Campbell | aff65da | 2011-08-22 23:44:59 +0000 | [diff] [blame] | 1312 | skb_frag_unref(skb, i); |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1313 | eat -= size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1314 | } else { |
| 1315 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; |
| 1316 | if (eat) { |
| 1317 | skb_shinfo(skb)->frags[k].page_offset += eat; |
Eric Dumazet | 9e903e0 | 2011-10-18 21:00:24 +0000 | [diff] [blame] | 1318 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1319 | eat = 0; |
| 1320 | } |
| 1321 | k++; |
| 1322 | } |
| 1323 | } |
| 1324 | skb_shinfo(skb)->nr_frags = k; |
| 1325 | |
Arnaldo Carvalho de Melo | 27a884d | 2007-04-19 20:29:13 -0700 | [diff] [blame] | 1326 | skb_reset_tail_pointer(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1327 | skb->data_len -= len; |
| 1328 | skb->len = skb->data_len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1329 | } |
| 1330 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1331 | /* Remove acked data from a packet in the transmit queue. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1332 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
| 1333 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1334 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1335 | return -ENOMEM; |
| 1336 | |
Eric Dumazet | 4fa48bf | 2011-12-04 08:51:08 +0000 | [diff] [blame] | 1337 | __pskb_trim_head(skb, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1338 | |
| 1339 | TCP_SKB_CB(skb)->seq += len; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1340 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1341 | |
| 1342 | skb->truesize -= len; |
| 1343 | sk->sk_wmem_queued -= len; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1344 | sk_mem_uncharge(sk, len); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1345 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
| 1346 | |
Neal Cardwell | 5b35e1e | 2012-01-28 17:29:46 +0000 | [diff] [blame] | 1347 | /* Any change of skb->len requires recalculation of tso factor. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1348 | if (tcp_skb_pcount(skb) > 1) |
Neal Cardwell | 5b35e1e | 2012-01-28 17:29:46 +0000 | [diff] [blame] | 1349 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1350 | |
| 1351 | return 0; |
| 1352 | } |
| 1353 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1354 | /* Calculate MSS. Not accounting for SACKs here. */ |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1355 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1356 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1357 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1358 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1359 | int mss_now; |
| 1360 | |
| 1361 | /* Calculate base mss without TCP options: |
| 1362 | It is MMS_S - sizeof(tcphdr) of rfc1122 |
| 1363 | */ |
| 1364 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); |
| 1365 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1366 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1367 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1368 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1369 | |
| 1370 | if (dst && dst_allfrag(dst)) |
| 1371 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; |
| 1372 | } |
| 1373 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1374 | /* Clamp it (mss_clamp does not include tcp options) */ |
| 1375 | if (mss_now > tp->rx_opt.mss_clamp) |
| 1376 | mss_now = tp->rx_opt.mss_clamp; |
| 1377 | |
| 1378 | /* Now subtract optional transport overhead */ |
| 1379 | mss_now -= icsk->icsk_ext_hdr_len; |
| 1380 | |
| 1381 | /* Then reserve room for full set of TCP options and 8 bytes of data */ |
| 1382 | if (mss_now < 48) |
| 1383 | mss_now = 48; |
| 1384 | |
| 1385 | /* Now subtract TCP options size, not including SACKs */ |
| 1386 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); |
| 1387 | |
| 1388 | return mss_now; |
| 1389 | } |
| 1390 | |
| 1391 | /* Inverse of above */ |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1392 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1393 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1394 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1395 | const struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1396 | int mtu; |
| 1397 | |
| 1398 | mtu = mss + |
| 1399 | tp->tcp_header_len + |
| 1400 | icsk->icsk_ext_hdr_len + |
| 1401 | icsk->icsk_af_ops->net_header_len; |
| 1402 | |
Eric Dumazet | 6746960 | 2012-04-24 07:37:38 +0000 | [diff] [blame] | 1403 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
| 1404 | if (icsk->icsk_af_ops->net_frag_header_len) { |
| 1405 | const struct dst_entry *dst = __sk_dst_get(sk); |
| 1406 | |
| 1407 | if (dst && dst_allfrag(dst)) |
| 1408 | mtu += icsk->icsk_af_ops->net_frag_header_len; |
| 1409 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1410 | return mtu; |
| 1411 | } |
| 1412 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1413 | /* MTU probing init per socket */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1414 | void tcp_mtup_init(struct sock *sk) |
| 1415 | { |
| 1416 | struct tcp_sock *tp = tcp_sk(sk); |
| 1417 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1418 | |
| 1419 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; |
| 1420 | 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] | 1421 | icsk->icsk_af_ops->net_header_len; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1422 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
| 1423 | icsk->icsk_mtup.probe_size = 0; |
| 1424 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1425 | EXPORT_SYMBOL(tcp_mtup_init); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1426 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1427 | /* This function synchronize snd mss to current pmtu/exthdr set. |
| 1428 | |
| 1429 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts |
| 1430 | for TCP options, but includes only bare TCP header. |
| 1431 | |
| 1432 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 1433 | It is minimum of user_mss and mss received with SYN. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1434 | It also does not include TCP options. |
| 1435 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1436 | 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] | 1437 | |
| 1438 | tp->mss_cache is current effective sending mss, including |
| 1439 | all tcp options except for SACKs. It is evaluated, |
| 1440 | taking into account current pmtu, but never exceeds |
| 1441 | tp->rx_opt.mss_clamp. |
| 1442 | |
| 1443 | NOTE1. rfc1122 clearly states that advertised MSS |
| 1444 | DOES NOT include either tcp or ip options. |
| 1445 | |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1446 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
| 1447 | are READ ONLY outside this function. --ANK (980731) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1448 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1449 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
| 1450 | { |
| 1451 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1452 | struct inet_connection_sock *icsk = inet_csk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1453 | int mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1454 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1455 | if (icsk->icsk_mtup.search_high > pmtu) |
| 1456 | icsk->icsk_mtup.search_high = pmtu; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1457 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1458 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
Ilpo Järvinen | 409d22b | 2007-12-31 14:57:40 -0800 | [diff] [blame] | 1459 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1460 | |
| 1461 | /* And store cached results */ |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1462 | icsk->icsk_pmtu_cookie = pmtu; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1463 | if (icsk->icsk_mtup.enabled) |
| 1464 | 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] | 1465 | tp->mss_cache = mss_now; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1466 | |
| 1467 | return mss_now; |
| 1468 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 1469 | EXPORT_SYMBOL(tcp_sync_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1470 | |
| 1471 | /* Compute the current effective MSS, taking SACKs and IP options, |
| 1472 | * and even PMTU discovery events into account. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1473 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1474 | unsigned int tcp_current_mss(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1475 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1476 | const struct tcp_sock *tp = tcp_sk(sk); |
| 1477 | const struct dst_entry *dst = __sk_dst_get(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1478 | u32 mss_now; |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1479 | unsigned int header_len; |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1480 | struct tcp_out_options opts; |
| 1481 | struct tcp_md5sig_key *md5; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1482 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1483 | mss_now = tp->mss_cache; |
| 1484 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1485 | if (dst) { |
| 1486 | u32 mtu = dst_mtu(dst); |
Arnaldo Carvalho de Melo | d83d846 | 2005-12-13 23:26:10 -0800 | [diff] [blame] | 1487 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1488 | mss_now = tcp_sync_mss(sk, mtu); |
| 1489 | } |
| 1490 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 1491 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
| 1492 | sizeof(struct tcphdr); |
| 1493 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes |
| 1494 | * some common options. If this is an odd packet (because we have SACK |
| 1495 | * blocks etc) then our calculated header_len will be different, and |
| 1496 | * we have to adjust mss_now correspondingly */ |
| 1497 | if (header_len != tp->tcp_header_len) { |
| 1498 | int delta = (int) header_len - tp->tcp_header_len; |
| 1499 | mss_now -= delta; |
| 1500 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 1501 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1502 | return mss_now; |
| 1503 | } |
| 1504 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1505 | /* Congestion window validation. (RFC2861) */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1506 | static void tcp_cwnd_validate(struct sock *sk) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1507 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1508 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1509 | |
Ilpo Järvinen | d436d68 | 2007-12-31 14:58:00 -0800 | [diff] [blame] | 1510 | if (tp->packets_out >= tp->snd_cwnd) { |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 1511 | /* Network is feed fully. */ |
| 1512 | tp->snd_cwnd_used = 0; |
| 1513 | tp->snd_cwnd_stamp = tcp_time_stamp; |
| 1514 | } else { |
| 1515 | /* Network starves. */ |
| 1516 | if (tp->packets_out > tp->snd_cwnd_used) |
| 1517 | tp->snd_cwnd_used = tp->packets_out; |
| 1518 | |
David S. Miller | 15d33c0 | 2007-04-09 13:23:14 -0700 | [diff] [blame] | 1519 | if (sysctl_tcp_slow_start_after_idle && |
| 1520 | (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] | 1521 | tcp_cwnd_application_limited(sk); |
| 1522 | } |
| 1523 | } |
| 1524 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1525 | /* Returns the portion of skb which can be sent right away without |
| 1526 | * introducing MSS oddities to segment boundaries. In rare cases where |
| 1527 | * mss_now != mss_cache, we will request caller to create a small skb |
| 1528 | * per input skb which could be mostly avoided here (if desired). |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1529 | * |
| 1530 | * We explicitly want to create a request for splitting write queue tail |
| 1531 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, |
| 1532 | * thus all the complexity (cwnd_len is always MSS multiple which we |
| 1533 | * return whenever allowed by the other factors). Basically we need the |
| 1534 | * modulo only when the receiver window alone is the limiting factor or |
| 1535 | * 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] | 1536 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1537 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1538 | unsigned int mss_now, unsigned int max_segs) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1539 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1540 | const struct tcp_sock *tp = tcp_sk(sk); |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1541 | u32 needed, window, max_len; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1542 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1543 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1544 | max_len = mss_now * max_segs; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1545 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1546 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
| 1547 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1548 | |
Ilpo Järvinen | 5ea3a74 | 2008-03-11 17:55:27 -0700 | [diff] [blame] | 1549 | needed = min(skb->len, window); |
| 1550 | |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1551 | if (max_len <= needed) |
| 1552 | return max_len; |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1553 | |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 1554 | return needed - needed % mss_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1555 | } |
| 1556 | |
| 1557 | /* Can at least one segment of SKB be sent right now, according to the |
| 1558 | * congestion window rules? If so, return how many segments are allowed. |
| 1559 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1560 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
| 1561 | const struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1562 | { |
| 1563 | u32 in_flight, cwnd; |
| 1564 | |
| 1565 | /* Don't be strict about the congestion window for the final FIN. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1566 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
| 1567 | tcp_skb_pcount(skb) == 1) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1568 | return 1; |
| 1569 | |
| 1570 | in_flight = tcp_packets_in_flight(tp); |
| 1571 | cwnd = tp->snd_cwnd; |
| 1572 | if (in_flight < cwnd) |
| 1573 | return (cwnd - in_flight); |
| 1574 | |
| 1575 | return 0; |
| 1576 | } |
| 1577 | |
Uwe Kleine-König | b595076 | 2010-11-01 15:38:34 -0400 | [diff] [blame] | 1578 | /* Initialize TSO state of a skb. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1579 | * This must be invoked the first time we consider transmitting |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1580 | * SKB onto the wire. |
| 1581 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1582 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1583 | unsigned int mss_now) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1584 | { |
| 1585 | int tso_segs = tcp_skb_pcount(skb); |
| 1586 | |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1587 | 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] | 1588 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1589 | tso_segs = tcp_skb_pcount(skb); |
| 1590 | } |
| 1591 | return tso_segs; |
| 1592 | } |
| 1593 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1594 | /* Minshall's variant of the Nagle send check. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1595 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1596 | { |
Jianjun Kong | 09cb105 | 2008-11-03 00:27:11 -0800 | [diff] [blame] | 1597 | return after(tp->snd_sml, tp->snd_una) && |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1598 | !after(tp->snd_sml, tp->snd_nxt); |
| 1599 | } |
| 1600 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1601 | /* Return false, if packet can be sent now without violation Nagle's rules: |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1602 | * 1. It is full sized. |
| 1603 | * 2. Or it contains FIN. (already checked by caller) |
Feng King | 6d67e9b | 2011-11-05 04:23:23 +0000 | [diff] [blame] | 1604 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1605 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
| 1606 | * With Minshall's modification: all sent small packets are ACKed. |
| 1607 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1608 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 1609 | const struct sk_buff *skb, |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 1610 | unsigned int mss_now, int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1611 | { |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1612 | return skb->len < mss_now && |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1613 | ((nonagle & TCP_NAGLE_CORK) || |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1614 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1615 | } |
| 1616 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1617 | /* Return true if the Nagle test allows this packet to be |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1618 | * sent now. |
| 1619 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1620 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
| 1621 | unsigned int cur_mss, int nonagle) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1622 | { |
| 1623 | /* Nagle rule does not apply to frames, which sit in the middle of the |
| 1624 | * write_queue (they have no chances to get new data). |
| 1625 | * |
| 1626 | * This is implemented in the callers, where they modify the 'nonagle' |
| 1627 | * argument based upon the location of SKB in the send queue. |
| 1628 | */ |
| 1629 | if (nonagle & TCP_NAGLE_PUSH) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1630 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1631 | |
Ilpo Järvinen | d551e45 | 2007-04-30 00:42:20 -0700 | [diff] [blame] | 1632 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
| 1633 | * Nagle can be ignored during F-RTO too (see RFC4138). |
| 1634 | */ |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 1635 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1636 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1637 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1638 | |
| 1639 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1640 | return true; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1641 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1642 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1643 | } |
| 1644 | |
| 1645 | /* Does at least the first segment of SKB fit into the send window? */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1646 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
| 1647 | const struct sk_buff *skb, |
| 1648 | unsigned int cur_mss) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1649 | { |
| 1650 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1651 | |
| 1652 | if (skb->len > cur_mss) |
| 1653 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; |
| 1654 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1655 | return !after(end_seq, tcp_wnd_end(tp)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1656 | } |
| 1657 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1658 | /* 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] | 1659 | * should be put on the wire right now. If so, it returns the number of |
| 1660 | * packets allowed by the congestion window. |
| 1661 | */ |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1662 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1663 | unsigned int cur_mss, int nonagle) |
| 1664 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1665 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1666 | unsigned int cwnd_quota; |
| 1667 | |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1668 | tcp_init_tso_segs(sk, skb, cur_mss); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1669 | |
| 1670 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) |
| 1671 | return 0; |
| 1672 | |
| 1673 | cwnd_quota = tcp_cwnd_test(tp, skb); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1674 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1675 | cwnd_quota = 0; |
| 1676 | |
| 1677 | return cwnd_quota; |
| 1678 | } |
| 1679 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1680 | /* Test if sending is allowed right now. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1681 | bool tcp_may_send_now(struct sock *sk) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1682 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 1683 | const struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1684 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1685 | |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1686 | return skb && |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1687 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1688 | (tcp_skb_is_last(sk, skb) ? |
Eric Dumazet | a02cec2 | 2010-09-22 20:43:57 +0000 | [diff] [blame] | 1689 | tp->nonagle : TCP_NAGLE_PUSH)); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1690 | } |
| 1691 | |
| 1692 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
| 1693 | * which is put after SKB on the list. It is very much like |
| 1694 | * tcp_fragment() except that it may make several kinds of assumptions |
| 1695 | * in order to speed up the splitting operation. In particular, we |
| 1696 | * know that all the data is in scatter-gather pages, and that the |
| 1697 | * packet has never been sent out before (and thus is not cloned). |
| 1698 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1699 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 1700 | unsigned int mss_now, gfp_t gfp) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1701 | { |
| 1702 | struct sk_buff *buff; |
| 1703 | int nlen = skb->len - len; |
Ilpo Järvinen | 9ce0146 | 2009-02-28 04:44:42 +0000 | [diff] [blame] | 1704 | u8 flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1705 | |
| 1706 | /* All of a TSO frame must be composed of paged data. */ |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1707 | if (skb->len != skb->data_len) |
| 1708 | return tcp_fragment(sk, skb, len, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1709 | |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 1710 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1711 | if (unlikely(buff == NULL)) |
| 1712 | return -ENOMEM; |
| 1713 | |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1714 | sk->sk_wmem_queued += buff->truesize; |
| 1715 | sk_mem_charge(sk, buff->truesize); |
Herbert Xu | b60b49e | 2006-04-19 21:35:00 -0700 | [diff] [blame] | 1716 | buff->truesize += nlen; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1717 | skb->truesize -= nlen; |
| 1718 | |
| 1719 | /* Correct the sequence numbers. */ |
| 1720 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; |
| 1721 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; |
| 1722 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; |
| 1723 | |
| 1724 | /* PSH and FIN should only be set in the second packet. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1725 | flags = TCP_SKB_CB(skb)->tcp_flags; |
| 1726 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
| 1727 | TCP_SKB_CB(buff)->tcp_flags = flags; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1728 | |
| 1729 | /* This packet was never sent out yet, so no SACK bits. */ |
| 1730 | TCP_SKB_CB(buff)->sacked = 0; |
| 1731 | |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1732 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1733 | skb_split(skb, buff, len); |
| 1734 | |
| 1735 | /* Fix up tso_factor for both original and new SKB. */ |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 1736 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1737 | tcp_set_skb_tso_segs(sk, buff, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1738 | |
| 1739 | /* Link BUFF into the send queue. */ |
| 1740 | skb_header_release(buff); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1741 | tcp_insert_write_queue_after(skb, buff, sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1742 | |
| 1743 | return 0; |
| 1744 | } |
| 1745 | |
| 1746 | /* Try to defer sending, if possible, in order to minimize the amount |
| 1747 | * of TSO splitting we do. View it as a kind of TSO Nagle test. |
| 1748 | * |
| 1749 | * This algorithm is from John Heffner. |
| 1750 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1751 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1752 | { |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 1753 | struct tcp_sock *tp = tcp_sk(sk); |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1754 | const struct inet_connection_sock *icsk = inet_csk(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1755 | u32 send_win, cong_win, limit, in_flight; |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1756 | int win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1757 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1758 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1759 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1760 | |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 1761 | if (icsk->icsk_ca_state != TCP_CA_Open) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1762 | goto send_now; |
| 1763 | |
| 1764 | /* Defer for less than two clock ticks. */ |
Ilpo Järvinen | bd515c3 | 2007-12-20 20:36:03 -0800 | [diff] [blame] | 1765 | if (tp->tso_deferred && |
Ilpo Järvinen | a2acde0 | 2008-12-05 22:49:37 -0800 | [diff] [blame] | 1766 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1767 | goto send_now; |
David S. Miller | 908a75c | 2005-07-05 15:43:58 -0700 | [diff] [blame] | 1768 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1769 | in_flight = tcp_packets_in_flight(tp); |
| 1770 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1771 | 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] | 1772 | |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1773 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1774 | |
| 1775 | /* From in_flight test above, we know that cwnd > in_flight. */ |
| 1776 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; |
| 1777 | |
| 1778 | limit = min(send_win, cong_win); |
| 1779 | |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1780 | /* If a full-sized TSO skb can be sent, do it. */ |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 1781 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
| 1782 | sk->sk_gso_max_segs * tp->mss_cache)) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1783 | goto send_now; |
David S. Miller | ba244fe | 2006-03-11 18:51:49 -0800 | [diff] [blame] | 1784 | |
Ilpo Järvinen | 62ad276 | 2009-02-28 04:44:29 +0000 | [diff] [blame] | 1785 | /* Middle in queue won't get any more data, full sendable already? */ |
| 1786 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) |
| 1787 | goto send_now; |
| 1788 | |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1789 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
| 1790 | if (win_divisor) { |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1791 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
| 1792 | |
| 1793 | /* If at least some fraction of a window is available, |
| 1794 | * just use it. |
| 1795 | */ |
Eric Dumazet | ad9f4f5 | 2010-12-07 12:03:55 +0000 | [diff] [blame] | 1796 | chunk /= win_divisor; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1797 | if (limit >= chunk) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1798 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1799 | } else { |
| 1800 | /* Different approach, try not to defer past a single |
| 1801 | * ACK. Receiver should ACK every other full sized |
| 1802 | * frame, so if we have space for more than 3 frames |
| 1803 | * then send now. |
| 1804 | */ |
Neal Cardwell | 6b5a5c0 | 2011-11-21 17:15:14 +0000 | [diff] [blame] | 1805 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1806 | goto send_now; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1807 | } |
| 1808 | |
| 1809 | /* Ok, it looks like it is advisable to defer. */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1810 | tp->tso_deferred = 1 | (jiffies << 1); |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1811 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1812 | return true; |
John Heffner | ae8064a | 2006-10-18 20:36:48 -0700 | [diff] [blame] | 1813 | |
| 1814 | send_now: |
| 1815 | tp->tso_deferred = 0; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1816 | return false; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1817 | } |
| 1818 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1819 | /* Create a new MTU probe if we are ready. |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 1820 | * MTU probe is regularly attempting to increase the path MTU by |
| 1821 | * deliberately sending larger packets. This discovers routing |
| 1822 | * changes resulting in larger path MTUs. |
| 1823 | * |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1824 | * Returns 0 if we should wait to probe (no cwnd available), |
| 1825 | * 1 if a probe was sent, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1826 | * -1 otherwise |
| 1827 | */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1828 | static int tcp_mtu_probe(struct sock *sk) |
| 1829 | { |
| 1830 | struct tcp_sock *tp = tcp_sk(sk); |
| 1831 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 1832 | struct sk_buff *skb, *nskb, *next; |
| 1833 | int len; |
| 1834 | int probe_size; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1835 | int size_needed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1836 | int copy; |
| 1837 | int mss_now; |
| 1838 | |
| 1839 | /* Not currently probing/verifying, |
| 1840 | * not in recovery, |
| 1841 | * have enough cwnd, and |
| 1842 | * not SACKing (the variable headers throw things off) */ |
| 1843 | if (!icsk->icsk_mtup.enabled || |
| 1844 | icsk->icsk_mtup.probe_size || |
| 1845 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || |
| 1846 | tp->snd_cwnd < 11 || |
Ilpo Järvinen | cabeccb | 2009-02-28 04:44:38 +0000 | [diff] [blame] | 1847 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1848 | return -1; |
| 1849 | |
| 1850 | /* Very simple search strategy: just double the MSS. */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 1851 | mss_now = tcp_current_mss(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1852 | probe_size = 2 * tp->mss_cache; |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1853 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1854 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
| 1855 | /* TODO: set timer for probe_converge_event */ |
| 1856 | return -1; |
| 1857 | } |
| 1858 | |
| 1859 | /* Have enough data in the send queue to probe? */ |
Ilpo Järvinen | 7f9c33e | 2007-11-23 19:10:56 +0800 | [diff] [blame] | 1860 | if (tp->write_seq - tp->snd_nxt < size_needed) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1861 | return -1; |
| 1862 | |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1863 | if (tp->snd_wnd < size_needed) |
| 1864 | return -1; |
Ilpo Järvinen | 90840de | 2007-12-31 04:48:41 -0800 | [diff] [blame] | 1865 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
Ilpo Järvinen | 91cc17c | 2007-11-23 19:08:16 +0800 | [diff] [blame] | 1866 | return 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1867 | |
Ilpo Järvinen | d67c58e | 2007-12-02 00:48:01 +0200 | [diff] [blame] | 1868 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
| 1869 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { |
| 1870 | if (!tcp_packets_in_flight(tp)) |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1871 | return -1; |
| 1872 | else |
| 1873 | return 0; |
| 1874 | } |
| 1875 | |
| 1876 | /* We're allowed to probe. Build it now. */ |
| 1877 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) |
| 1878 | return -1; |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1879 | sk->sk_wmem_queued += nskb->truesize; |
| 1880 | sk_mem_charge(sk, nskb->truesize); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1881 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1882 | skb = tcp_send_head(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1883 | |
| 1884 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; |
| 1885 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1886 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1887 | TCP_SKB_CB(nskb)->sacked = 0; |
| 1888 | nskb->csum = 0; |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1889 | nskb->ip_summed = skb->ip_summed; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1890 | |
Ilpo Järvinen | 50c4817 | 2007-12-02 00:48:00 +0200 | [diff] [blame] | 1891 | tcp_insert_write_queue_before(nskb, skb, sk); |
| 1892 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1893 | len = 0; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1894 | tcp_for_write_queue_from_safe(skb, next, sk) { |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1895 | copy = min_t(int, skb->len, probe_size - len); |
| 1896 | if (nskb->ip_summed) |
| 1897 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
| 1898 | else |
| 1899 | nskb->csum = skb_copy_and_csum_bits(skb, 0, |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1900 | skb_put(nskb, copy), |
| 1901 | copy, nskb->csum); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1902 | |
| 1903 | if (skb->len <= copy) { |
| 1904 | /* We've eaten all the data from this skb. |
| 1905 | * Throw it away. */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1906 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1907 | tcp_unlink_write_queue(skb, sk); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 1908 | sk_wmem_free_skb(sk, skb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1909 | } else { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 1910 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 1911 | ~(TCPHDR_FIN|TCPHDR_PSH); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1912 | if (!skb_shinfo(skb)->nr_frags) { |
| 1913 | skb_pull(skb, copy); |
Patrick McHardy | 84fa793 | 2006-08-29 16:44:56 -0700 | [diff] [blame] | 1914 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1915 | skb->csum = csum_partial(skb->data, |
| 1916 | skb->len, 0); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1917 | } else { |
| 1918 | __pskb_trim_head(skb, copy); |
| 1919 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
| 1920 | } |
| 1921 | TCP_SKB_CB(skb)->seq += copy; |
| 1922 | } |
| 1923 | |
| 1924 | len += copy; |
Ilpo Järvinen | 234b686 | 2007-12-02 00:48:02 +0200 | [diff] [blame] | 1925 | |
| 1926 | if (len >= probe_size) |
| 1927 | break; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1928 | } |
| 1929 | tcp_init_tso_segs(sk, nskb, nskb->len); |
| 1930 | |
| 1931 | /* We're ready to send. If this fails, the probe will |
| 1932 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ |
| 1933 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; |
| 1934 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
| 1935 | /* Decrement cwnd here because we are sending |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 1936 | * effectively two packets. */ |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1937 | tp->snd_cwnd--; |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 1938 | tcp_event_new_data_sent(sk, nskb); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1939 | |
| 1940 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); |
John Heffner | 0e7b136 | 2006-03-20 21:32:58 -0800 | [diff] [blame] | 1941 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
| 1942 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1943 | |
| 1944 | return 1; |
| 1945 | } |
| 1946 | |
| 1947 | return -1; |
| 1948 | } |
| 1949 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1950 | /* This routine writes packets to the network. It advances the |
| 1951 | * send_head. This happens as incoming acks open up the remote |
| 1952 | * window for us. |
| 1953 | * |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 1954 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
| 1955 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into |
| 1956 | * account rare use of URG, this is not a big flaw. |
| 1957 | * |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1958 | * Returns true, if no segments are in flight and we have queued segments, |
| 1959 | * but cannot send anything now because of SWS or another problem. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1960 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1961 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
| 1962 | int push_one, gfp_t gfp) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1963 | { |
| 1964 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1965 | struct sk_buff *skb; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1966 | unsigned int tso_segs, sent_pkts; |
| 1967 | int cwnd_quota; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1968 | int result; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1969 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 1970 | sent_pkts = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1971 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1972 | if (!push_one) { |
| 1973 | /* Do MTU probing. */ |
| 1974 | result = tcp_mtu_probe(sk); |
| 1975 | if (!result) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 1976 | return false; |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 1977 | } else if (result > 0) { |
| 1978 | sent_pkts = 1; |
| 1979 | } |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 1980 | } |
| 1981 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 1982 | while ((skb = tcp_send_head(sk))) { |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 1983 | unsigned int limit; |
| 1984 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 1985 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1986 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1987 | BUG_ON(!tso_segs); |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 1988 | |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 1989 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) |
| 1990 | goto repair; /* Skip network transmission */ |
| 1991 | |
Herbert Xu | b68e9f8 | 2005-08-04 19:52:02 -0700 | [diff] [blame] | 1992 | cwnd_quota = tcp_cwnd_test(tp, skb); |
| 1993 | if (!cwnd_quota) |
| 1994 | break; |
| 1995 | |
| 1996 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) |
| 1997 | break; |
| 1998 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 1999 | if (tso_segs == 1) { |
| 2000 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
| 2001 | (tcp_skb_is_last(sk, skb) ? |
| 2002 | nonagle : TCP_NAGLE_PUSH)))) |
| 2003 | break; |
| 2004 | } else { |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2005 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2006 | break; |
| 2007 | } |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 2008 | |
Eric Dumazet | 46d3cea | 2012-07-11 05:50:31 +0000 | [diff] [blame] | 2009 | /* TSQ : sk_wmem_alloc accounts skb truesize, |
| 2010 | * including skb overhead. But thats OK. |
| 2011 | */ |
| 2012 | if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { |
| 2013 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
| 2014 | break; |
| 2015 | } |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 2016 | limit = mss_now; |
Ilpo Järvinen | f8269a4 | 2008-12-03 21:24:48 -0800 | [diff] [blame] | 2017 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
Ilpo Järvinen | 0e3a480 | 2007-12-24 21:33:45 -0800 | [diff] [blame] | 2018 | limit = tcp_mss_split_point(sk, skb, mss_now, |
Ben Hutchings | 1485348 | 2012-07-30 16:11:42 +0000 | [diff] [blame] | 2019 | min_t(unsigned int, |
| 2020 | cwnd_quota, |
| 2021 | sk->sk_gso_max_segs)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2022 | |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 2023 | if (skb->len > limit && |
Eric Dumazet | c4ead4c | 2010-06-24 01:00:22 +0000 | [diff] [blame] | 2024 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
Herbert Xu | c8ac377 | 2005-08-16 20:43:40 -0700 | [diff] [blame] | 2025 | break; |
| 2026 | |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2027 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2028 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2029 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2030 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2031 | |
Andrew Vagin | ec34232 | 2012-11-15 04:03:17 +0000 | [diff] [blame] | 2032 | repair: |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2033 | /* Advance the send_head. This one is sent out. |
| 2034 | * This call will increment packets_out. |
| 2035 | */ |
Ilpo Järvinen | 66f5fe6 | 2007-12-31 04:43:57 -0800 | [diff] [blame] | 2036 | tcp_event_new_data_sent(sk, skb); |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2037 | |
| 2038 | tcp_minshall_update(tp, mss_now, skb); |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2039 | sent_pkts += tcp_skb_pcount(skb); |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2040 | |
| 2041 | if (push_one) |
| 2042 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2043 | } |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2044 | |
David S. Miller | aa93466 | 2005-07-05 15:20:09 -0700 | [diff] [blame] | 2045 | if (likely(sent_pkts)) { |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2046 | if (tcp_in_cwnd_reduction(sk)) |
| 2047 | tp->prr_out += sent_pkts; |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2048 | tcp_cwnd_validate(sk); |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2049 | return false; |
David S. Miller | 92df7b5 | 2005-07-05 15:19:06 -0700 | [diff] [blame] | 2050 | } |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2051 | return !tp->packets_out && tcp_send_head(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2052 | } |
| 2053 | |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2054 | /* Push out any pending frames which were held back due to |
| 2055 | * TCP_CORK or attempt at coalescing tiny packets. |
| 2056 | * The socket must be locked by the caller. |
| 2057 | */ |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2058 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
| 2059 | int nonagle) |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2060 | { |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2061 | /* If we are closed, the bytes will have to remain here. |
| 2062 | * In time closedown will finish, we empty the write queue and |
| 2063 | * all will be happy. |
| 2064 | */ |
| 2065 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
| 2066 | return; |
| 2067 | |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 2068 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
| 2069 | sk_gfp_atomic(sk, GFP_ATOMIC))) |
Ilpo Järvinen | 726e07a | 2008-12-05 22:43:56 -0800 | [diff] [blame] | 2070 | tcp_check_probe_timer(sk); |
David S. Miller | a762a98 | 2005-07-05 15:18:51 -0700 | [diff] [blame] | 2071 | } |
| 2072 | |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2073 | /* Send _single_ skb sitting at the send head. This function requires |
| 2074 | * true push pending frames to setup probe timer etc. |
| 2075 | */ |
| 2076 | void tcp_push_one(struct sock *sk, unsigned int mss_now) |
| 2077 | { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2078 | struct sk_buff *skb = tcp_send_head(sk); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2079 | |
| 2080 | BUG_ON(!skb || skb->len < mss_now); |
| 2081 | |
Ilpo Järvinen | d5dd917 | 2008-12-05 22:48:55 -0800 | [diff] [blame] | 2082 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
David S. Miller | c1b4a7e | 2005-07-05 15:24:38 -0700 | [diff] [blame] | 2083 | } |
| 2084 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2085 | /* This function returns the amount that we can raise the |
| 2086 | * usable window based on the following constraints |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2087 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2088 | * 1. The window can never be shrunk once it is offered (RFC 793) |
| 2089 | * 2. We limit memory per socket |
| 2090 | * |
| 2091 | * RFC 1122: |
| 2092 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep |
| 2093 | * RECV.NEXT + RCV.WIN fixed until: |
| 2094 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" |
| 2095 | * |
| 2096 | * i.e. don't raise the right edge of the window until you can raise |
| 2097 | * it at least MSS bytes. |
| 2098 | * |
| 2099 | * Unfortunately, the recommended algorithm breaks header prediction, |
| 2100 | * since header prediction assumes th->window stays fixed. |
| 2101 | * |
| 2102 | * Strictly speaking, keeping th->window fixed violates the receiver |
| 2103 | * side SWS prevention criteria. The problem is that under this rule |
| 2104 | * a stream of single byte packets will cause the right side of the |
| 2105 | * window to always advance by a single byte. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2106 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2107 | * Of course, if the sender implements sender side SWS prevention |
| 2108 | * then this will not be a problem. |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2109 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2110 | * BSD seems to make the following compromise: |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2111 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2112 | * If the free space is less than the 1/4 of the maximum |
| 2113 | * space available and the free space is less than 1/2 mss, |
| 2114 | * then set the window to 0. |
| 2115 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] |
| 2116 | * Otherwise, just prevent the window from shrinking |
| 2117 | * and from being larger than the largest representable value. |
| 2118 | * |
| 2119 | * This prevents incremental opening of the window in the regime |
| 2120 | * where TCP is limited by the speed of the reader side taking |
| 2121 | * data out of the TCP receive queue. It does nothing about |
| 2122 | * those cases where the window is constrained on the sender side |
| 2123 | * because the pipeline is full. |
| 2124 | * |
| 2125 | * BSD also seems to "accidentally" limit itself to windows that are a |
| 2126 | * multiple of MSS, at least until the free space gets quite small. |
| 2127 | * This would appear to be a side effect of the mbuf implementation. |
| 2128 | * Combining these two algorithms results in the observed behavior |
| 2129 | * of having a fixed window size at almost all times. |
| 2130 | * |
| 2131 | * Below we obtain similar behavior by forcing the offered window to |
| 2132 | * a multiple of the mss when it is feasible to do so. |
| 2133 | * |
| 2134 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. |
| 2135 | * Regular options like TIMESTAMP are taken into account. |
| 2136 | */ |
| 2137 | u32 __tcp_select_window(struct sock *sk) |
| 2138 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2139 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2140 | struct tcp_sock *tp = tcp_sk(sk); |
Stephen Hemminger | caa20d9a | 2005-11-10 17:13:47 -0800 | [diff] [blame] | 2141 | /* MSS for the peer's data. Previous versions used mss_clamp |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2142 | * here. I don't know if the value based on our guesses |
| 2143 | * of peer's MSS is better for the performance. It's more correct |
| 2144 | * but may be worse for the performance because of rcv_mss |
| 2145 | * fluctuations. --SAW 1998/11/1 |
| 2146 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2147 | int mss = icsk->icsk_ack.rcv_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2148 | int free_space = tcp_space(sk); |
| 2149 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); |
| 2150 | int window; |
| 2151 | |
| 2152 | if (mss > full_space) |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2153 | mss = full_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2154 | |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2155 | if (free_space < (full_space >> 1)) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2156 | icsk->icsk_ack.quick = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2157 | |
Glauber Costa | 180d8cd | 2011-12-11 21:47:02 +0000 | [diff] [blame] | 2158 | if (sk_under_memory_pressure(sk)) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2159 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
| 2160 | 4U * tp->advmss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2161 | |
| 2162 | if (free_space < mss) |
| 2163 | return 0; |
| 2164 | } |
| 2165 | |
| 2166 | if (free_space > tp->rcv_ssthresh) |
| 2167 | free_space = tp->rcv_ssthresh; |
| 2168 | |
| 2169 | /* Don't do rounding if we are using window scaling, since the |
| 2170 | * scaled window will not line up with the MSS boundary anyway. |
| 2171 | */ |
| 2172 | window = tp->rcv_wnd; |
| 2173 | if (tp->rx_opt.rcv_wscale) { |
| 2174 | window = free_space; |
| 2175 | |
| 2176 | /* Advertise enough space so that it won't get scaled away. |
| 2177 | * Import case: prevent zero window announcement if |
| 2178 | * 1<<rcv_wscale > mss. |
| 2179 | */ |
| 2180 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) |
| 2181 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) |
| 2182 | << tp->rx_opt.rcv_wscale); |
| 2183 | } else { |
| 2184 | /* Get the largest window that is a nice multiple of mss. |
| 2185 | * Window clamp already applied above. |
| 2186 | * If our current window offering is within 1 mss of the |
| 2187 | * free space we just keep it. This prevents the divide |
| 2188 | * and multiply from happening most of the time. |
| 2189 | * We also don't do any window rounding when the free space |
| 2190 | * is too small. |
| 2191 | */ |
| 2192 | if (window <= free_space - mss || window > free_space) |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2193 | window = (free_space / mss) * mss; |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2194 | else if (mss == full_space && |
Eric Dumazet | b92edbe | 2007-12-20 21:48:32 -0800 | [diff] [blame] | 2195 | free_space > window + (full_space >> 1)) |
John Heffner | 8456507 | 2007-04-02 13:56:32 -0700 | [diff] [blame] | 2196 | window = free_space; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2197 | } |
| 2198 | |
| 2199 | return window; |
| 2200 | } |
| 2201 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2202 | /* Collapses two adjacent SKB's during retransmission. */ |
| 2203 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2204 | { |
| 2205 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2206 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2207 | int skb_size, next_skb_size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2208 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2209 | skb_size = skb->len; |
| 2210 | next_skb_size = next_skb->len; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2211 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2212 | 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] | 2213 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2214 | tcp_highest_sack_combine(sk, next_skb, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2215 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2216 | tcp_unlink_write_queue(next_skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2217 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2218 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
| 2219 | next_skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2220 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2221 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
| 2222 | skb->ip_summed = CHECKSUM_PARTIAL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2223 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2224 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
| 2225 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2226 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2227 | /* Update sequence range on original skb. */ |
| 2228 | 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] | 2229 | |
Ilpo Järvinen | e6c7d08 | 2009-02-28 04:44:35 +0000 | [diff] [blame] | 2230 | /* Merge over control information. This moves PSH/FIN etc. over */ |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2231 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2232 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2233 | /* All done, get rid of second SKB and account for it so |
| 2234 | * packet counting does not break. |
| 2235 | */ |
| 2236 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; |
Ilpo Järvinen | b768920 | 2007-09-20 11:37:19 -0700 | [diff] [blame] | 2237 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2238 | /* changed transmit queue under us so clear hints */ |
Ilpo Järvinen | ef9da47 | 2008-09-20 21:25:15 -0700 | [diff] [blame] | 2239 | tcp_clear_retrans_hints_partial(tp); |
| 2240 | if (next_skb == tp->retransmit_skb_hint) |
| 2241 | tp->retransmit_skb_hint = skb; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2242 | |
Ilpo Järvinen | 797108d | 2009-04-01 23:15:17 +0000 | [diff] [blame] | 2243 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
| 2244 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 2245 | sk_wmem_free_skb(sk, next_skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2246 | } |
| 2247 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2248 | /* Check if coalescing SKBs is legal. */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2249 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2250 | { |
| 2251 | if (tcp_skb_pcount(skb) > 1) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2252 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2253 | /* TODO: SACK collapsing could be used to remove this condition */ |
| 2254 | if (skb_shinfo(skb)->nr_frags != 0) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2255 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2256 | if (skb_cloned(skb)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2257 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2258 | if (skb == tcp_send_head(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2259 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2260 | /* Some heurestics for collapsing over SACK'd could be invented */ |
| 2261 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2262 | return false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2263 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2264 | return true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2265 | } |
| 2266 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2267 | /* Collapse packets in the retransmit queue to make to create |
| 2268 | * less packets on the wire. This is only done on retransmission. |
| 2269 | */ |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2270 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
| 2271 | int space) |
| 2272 | { |
| 2273 | struct tcp_sock *tp = tcp_sk(sk); |
| 2274 | struct sk_buff *skb = to, *tmp; |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2275 | bool first = true; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2276 | |
| 2277 | if (!sysctl_tcp_retrans_collapse) |
| 2278 | return; |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2279 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2280 | return; |
| 2281 | |
| 2282 | tcp_for_write_queue_from_safe(skb, tmp, sk) { |
| 2283 | if (!tcp_can_collapse(sk, skb)) |
| 2284 | break; |
| 2285 | |
| 2286 | space -= skb->len; |
| 2287 | |
| 2288 | if (first) { |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2289 | first = false; |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2290 | continue; |
| 2291 | } |
| 2292 | |
| 2293 | if (space < 0) |
| 2294 | break; |
| 2295 | /* Punt if not enough space exists in the first SKB for |
| 2296 | * the data in the second |
| 2297 | */ |
Eric Dumazet | a21d457 | 2012-04-10 20:30:48 +0000 | [diff] [blame] | 2298 | if (skb->len > skb_availroom(to)) |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2299 | break; |
| 2300 | |
| 2301 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) |
| 2302 | break; |
| 2303 | |
| 2304 | tcp_collapse_retrans(sk, to); |
| 2305 | } |
| 2306 | } |
| 2307 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2308 | /* This retransmits one SKB. Policy decisions and retransmit queue |
| 2309 | * state updates are done by the caller. Returns non-zero if an |
| 2310 | * error occurred which prevented the send. |
| 2311 | */ |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2312 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2313 | { |
| 2314 | struct tcp_sock *tp = tcp_sk(sk); |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2315 | struct inet_connection_sock *icsk = inet_csk(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2316 | unsigned int cur_mss; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2317 | |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2318 | /* Inconslusive MTU probe */ |
| 2319 | if (icsk->icsk_mtup.probe_size) { |
| 2320 | icsk->icsk_mtup.probe_size = 0; |
| 2321 | } |
| 2322 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2323 | /* 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] | 2324 | * copying overhead: fragmentation, tunneling, mangling etc. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2325 | */ |
| 2326 | if (atomic_read(&sk->sk_wmem_alloc) > |
| 2327 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) |
| 2328 | return -EAGAIN; |
| 2329 | |
| 2330 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
| 2331 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) |
| 2332 | BUG(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2333 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
| 2334 | return -ENOMEM; |
| 2335 | } |
| 2336 | |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2337 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
| 2338 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
| 2339 | |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2340 | cur_mss = tcp_current_mss(sk); |
Sridhar Samudrala | 7d227cd2 | 2008-05-21 16:42:20 -0700 | [diff] [blame] | 2341 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2342 | /* If receiver has shrunk his window, and skb is out of |
| 2343 | * new window, do not retransmit it. The exception is the |
| 2344 | * case, when window is shrunk to zero. In this case |
| 2345 | * our retransmit serves as a zero window probe. |
| 2346 | */ |
Joe Perches | 9d4fb27 | 2009-11-23 10:41:23 -0800 | [diff] [blame] | 2347 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
| 2348 | TCP_SKB_CB(skb)->seq != tp->snd_una) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2349 | return -EAGAIN; |
| 2350 | |
| 2351 | if (skb->len > cur_mss) { |
David S. Miller | 846998a | 2005-08-04 19:52:01 -0700 | [diff] [blame] | 2352 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2353 | return -ENOMEM; /* We'll try again later. */ |
Ilpo Järvinen | 02276f3 | 2009-02-28 04:44:31 +0000 | [diff] [blame] | 2354 | } else { |
Ilpo Järvinen | 9eb9362 | 2009-04-01 23:18:20 +0000 | [diff] [blame] | 2355 | int oldpcount = tcp_skb_pcount(skb); |
| 2356 | |
| 2357 | if (unlikely(oldpcount > 1)) { |
| 2358 | tcp_init_tso_segs(sk, skb, cur_mss); |
| 2359 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); |
| 2360 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2361 | } |
| 2362 | |
Ilpo Järvinen | 4a17fc3 | 2008-11-24 21:03:43 -0800 | [diff] [blame] | 2363 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2364 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2365 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
| 2366 | * retransmit when old data is attached. So strip it off |
| 2367 | * since it is cheap to do so and saves bytes on the network. |
| 2368 | */ |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 2369 | if (skb->len > 0 && |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2370 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
Stephen Hemminger | 2de979b | 2007-03-08 20:45:19 -0800 | [diff] [blame] | 2371 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2372 | if (!pskb_trim(skb, 0)) { |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2373 | /* Reuse, even though it does some unnecessary work */ |
| 2374 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2375 | TCP_SKB_CB(skb)->tcp_flags); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2376 | skb->ip_summed = CHECKSUM_NONE; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2377 | } |
| 2378 | } |
| 2379 | |
| 2380 | /* Make a copy, if the first transmission SKB clone we made |
| 2381 | * is still in somebody's hands, else make a clone. |
| 2382 | */ |
| 2383 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2384 | |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2385 | /* make sure skb->data is aligned on arches that require it */ |
| 2386 | if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { |
| 2387 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, |
| 2388 | GFP_ATOMIC); |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2389 | return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
| 2390 | -ENOBUFS; |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2391 | } else { |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2392 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Eric Dumazet | 117632e | 2011-12-03 21:39:53 +0000 | [diff] [blame] | 2393 | } |
Yuchung Cheng | 93b174a | 2012-12-06 08:45:32 +0000 | [diff] [blame] | 2394 | } |
| 2395 | |
| 2396 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
| 2397 | { |
| 2398 | struct tcp_sock *tp = tcp_sk(sk); |
| 2399 | int err = __tcp_retransmit_skb(sk, skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2400 | |
| 2401 | if (err == 0) { |
| 2402 | /* Update global TCP statistics. */ |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2403 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2404 | |
| 2405 | tp->total_retrans++; |
| 2406 | |
| 2407 | #if FASTRETRANS_DEBUG > 0 |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2408 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
Joe Perches | e87cc47 | 2012-05-13 21:56:26 +0000 | [diff] [blame] | 2409 | net_dbg_ratelimited("retrans_out leaked\n"); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2410 | } |
| 2411 | #endif |
Ilpo Järvinen | b08d6cb | 2007-10-11 17:36:13 -0700 | [diff] [blame] | 2412 | if (!tp->retrans_out) |
| 2413 | tp->lost_retrans_low = tp->snd_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2414 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
| 2415 | tp->retrans_out += tcp_skb_pcount(skb); |
| 2416 | |
| 2417 | /* Save stamp of the first retransmit. */ |
| 2418 | if (!tp->retrans_stamp) |
| 2419 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; |
| 2420 | |
Yuchung Cheng | c24f691 | 2011-02-07 12:57:04 +0000 | [diff] [blame] | 2421 | tp->undo_retrans += tcp_skb_pcount(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2422 | |
| 2423 | /* snd_nxt is stored to detect loss of retransmitted segment, |
| 2424 | * see tcp_input.c tcp_sacktag_write_queue(). |
| 2425 | */ |
| 2426 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; |
| 2427 | } |
| 2428 | return err; |
| 2429 | } |
| 2430 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2431 | /* Check if we forward retransmits are possible in the current |
| 2432 | * window/congestion state. |
| 2433 | */ |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2434 | static bool tcp_can_forward_retransmit(struct sock *sk) |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2435 | { |
| 2436 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2437 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2438 | |
| 2439 | /* Forward retransmissions are possible only during Recovery. */ |
| 2440 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2441 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2442 | |
| 2443 | /* No forward retransmissions in Reno are possible. */ |
| 2444 | if (tcp_is_reno(tp)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2445 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2446 | |
| 2447 | /* Yeah, we have to make difficult choice between forward transmission |
| 2448 | * and retransmission... Both ways have their merits... |
| 2449 | * |
| 2450 | * For now we do not retransmit anything, while we have some new |
| 2451 | * segments to send. In the other cases, follow rule 3 for |
| 2452 | * NextSeg() specified in RFC3517. |
| 2453 | */ |
| 2454 | |
| 2455 | if (tcp_may_send_now(sk)) |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2456 | return false; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2457 | |
Eric Dumazet | a2a385d | 2012-05-16 23:15:34 +0000 | [diff] [blame] | 2458 | return true; |
Ilpo Järvinen | b5afe7b | 2008-09-20 21:21:54 -0700 | [diff] [blame] | 2459 | } |
| 2460 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2461 | /* This gets called after a retransmit timeout, and the initially |
| 2462 | * retransmitted data is acknowledged. It tries to continue |
| 2463 | * resending the rest of the retransmit queue, until either |
| 2464 | * we've sent it all or the congestion window limit is reached. |
| 2465 | * If doing SACK, the first ACK which comes back for a timeout |
| 2466 | * based retransmit packet might feed us FACK information again. |
| 2467 | * If so, we use it to avoid unnecessarily retransmissions. |
| 2468 | */ |
| 2469 | void tcp_xmit_retransmit_queue(struct sock *sk) |
| 2470 | { |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 2471 | const struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2472 | struct tcp_sock *tp = tcp_sk(sk); |
| 2473 | struct sk_buff *skb; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2474 | struct sk_buff *hole = NULL; |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2475 | u32 last_lost; |
Ilpo Järvinen | 61eb55f | 2008-09-20 21:22:59 -0700 | [diff] [blame] | 2476 | int mib_idx; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2477 | int fwd_rexmitting = 0; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2478 | |
Ilpo Järvinen | 45e77d3 | 2010-07-19 01:16:18 +0000 | [diff] [blame] | 2479 | if (!tp->packets_out) |
| 2480 | return; |
| 2481 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2482 | if (!tp->lost_out) |
| 2483 | tp->retransmit_high = tp->snd_una; |
| 2484 | |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2485 | if (tp->retransmit_skb_hint) { |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2486 | skb = tp->retransmit_skb_hint; |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2487 | last_lost = TCP_SKB_CB(skb)->end_seq; |
| 2488 | if (after(last_lost, tp->retransmit_high)) |
| 2489 | last_lost = tp->retransmit_high; |
| 2490 | } else { |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2491 | skb = tcp_write_queue_head(sk); |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2492 | last_lost = tp->snd_una; |
| 2493 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2494 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2495 | tcp_for_write_queue_from(skb, sk) { |
| 2496 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2497 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2498 | if (skb == tcp_send_head(sk)) |
| 2499 | break; |
| 2500 | /* we could do better than to assign each time */ |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2501 | if (hole == NULL) |
| 2502 | tp->retransmit_skb_hint = skb; |
Stephen Hemminger | 6a438bb | 2005-11-10 17:14:59 -0800 | [diff] [blame] | 2503 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2504 | /* Assume this retransmit will generate |
| 2505 | * only one packet for congestion window |
| 2506 | * calculation purposes. This works because |
| 2507 | * tcp_retransmit_skb() will chop up the |
| 2508 | * packet to be MSS sized and all the |
| 2509 | * packet counting works out. |
| 2510 | */ |
| 2511 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) |
| 2512 | return; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2513 | |
| 2514 | if (fwd_rexmitting) { |
| 2515 | begin_fwd: |
| 2516 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
| 2517 | break; |
| 2518 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; |
| 2519 | |
| 2520 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2521 | tp->retransmit_high = last_lost; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2522 | if (!tcp_can_forward_retransmit(sk)) |
| 2523 | break; |
| 2524 | /* Backtrack if necessary to non-L'ed skb */ |
| 2525 | if (hole != NULL) { |
| 2526 | skb = hole; |
| 2527 | hole = NULL; |
| 2528 | } |
| 2529 | fwd_rexmitting = 1; |
| 2530 | goto begin_fwd; |
| 2531 | |
| 2532 | } else if (!(sacked & TCPCB_LOST)) { |
Ilpo Järvinen | ac11ba7 | 2009-02-28 04:44:27 +0000 | [diff] [blame] | 2533 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2534 | hole = skb; |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2535 | continue; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2536 | |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2537 | } else { |
Ilpo Järvinen | 618d9f2 | 2008-09-20 21:26:22 -0700 | [diff] [blame] | 2538 | last_lost = TCP_SKB_CB(skb)->end_seq; |
Ilpo Järvinen | 0e1c54c | 2008-09-20 21:24:21 -0700 | [diff] [blame] | 2539 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
| 2540 | mib_idx = LINUX_MIB_TCPFASTRETRANS; |
| 2541 | else |
| 2542 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; |
| 2543 | } |
| 2544 | |
| 2545 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2546 | continue; |
Pavel Emelyanov | 40b215e | 2008-07-03 01:05:41 -0700 | [diff] [blame] | 2547 | |
Eric Dumazet | 09e9b81 | 2012-01-25 04:44:20 +0000 | [diff] [blame] | 2548 | if (tcp_retransmit_skb(sk, skb)) { |
| 2549 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2550 | return; |
Eric Dumazet | 09e9b81 | 2012-01-25 04:44:20 +0000 | [diff] [blame] | 2551 | } |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2552 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
| 2553 | |
Yuchung Cheng | 684bad1 | 2012-09-02 17:38:04 +0000 | [diff] [blame] | 2554 | if (tcp_in_cwnd_reduction(sk)) |
Nandita Dukkipati | a262f0c | 2011-08-21 20:21:57 +0000 | [diff] [blame] | 2555 | tp->prr_out += tcp_skb_pcount(skb); |
| 2556 | |
Ilpo Järvinen | 08ebd17 | 2008-09-20 21:23:49 -0700 | [diff] [blame] | 2557 | if (skb == tcp_write_queue_head(sk)) |
| 2558 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 2559 | inet_csk(sk)->icsk_rto, |
| 2560 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2561 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2562 | } |
| 2563 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2564 | /* Send a fin. The caller locks the socket for us. This cannot be |
| 2565 | * allowed to fail queueing a FIN frame under any circumstances. |
| 2566 | */ |
| 2567 | void tcp_send_fin(struct sock *sk) |
| 2568 | { |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2569 | struct tcp_sock *tp = tcp_sk(sk); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2570 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2571 | int mss_now; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 2572 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2573 | /* Optimization, tack on the FIN if we have a queue of |
| 2574 | * unsent frames. But be careful about outgoing SACKS |
| 2575 | * and IP options. |
| 2576 | */ |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 2577 | mss_now = tcp_current_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2578 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2579 | if (tcp_send_head(sk) != NULL) { |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2580 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2581 | TCP_SKB_CB(skb)->end_seq++; |
| 2582 | tp->write_seq++; |
| 2583 | } else { |
| 2584 | /* Socket is locked, keep trying until memory is available. */ |
| 2585 | for (;;) { |
Wu Fengguang | aa13307 | 2009-09-02 23:45:45 -0700 | [diff] [blame] | 2586 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
| 2587 | sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2588 | if (skb) |
| 2589 | break; |
| 2590 | yield(); |
| 2591 | } |
| 2592 | |
| 2593 | /* Reserve space for headers and prepare control bits. */ |
| 2594 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2595 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2596 | tcp_init_nondata_skb(skb, tp->write_seq, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2597 | TCPHDR_ACK | TCPHDR_FIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2598 | tcp_queue_skb(sk, skb); |
| 2599 | } |
Ilpo Järvinen | 9e412ba | 2007-04-20 22:18:02 -0700 | [diff] [blame] | 2600 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2601 | } |
| 2602 | |
| 2603 | /* We get here when a process closes a file descriptor (either due to |
| 2604 | * an explicit close() or as a byproduct of exit()'ing) and there |
| 2605 | * was unread data in the receive queue. This behavior is recommended |
Gerrit Renker | 65bb723 | 2007-04-28 21:21:46 -0700 | [diff] [blame] | 2606 | * by RFC 2525, section 2.17. -DaveM |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2607 | */ |
Al Viro | dd0fc66 | 2005-10-07 07:46:04 +0100 | [diff] [blame] | 2608 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2609 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2610 | struct sk_buff *skb; |
| 2611 | |
| 2612 | /* NOTE: No TCP options attached and we never retransmit this. */ |
| 2613 | skb = alloc_skb(MAX_TCP_HEADER, priority); |
| 2614 | if (!skb) { |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2615 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2616 | return; |
| 2617 | } |
| 2618 | |
| 2619 | /* Reserve space for headers and prepare control bits. */ |
| 2620 | skb_reserve(skb, MAX_TCP_HEADER); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2621 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2622 | TCPHDR_ACK | TCPHDR_RST); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2623 | /* Send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2624 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2625 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
Pavel Emelyanov | 4e67344 | 2008-07-16 20:30:14 -0700 | [diff] [blame] | 2626 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
Sridhar Samudrala | 26af65c | 2008-06-04 15:19:35 -0700 | [diff] [blame] | 2627 | |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 2628 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2629 | } |
| 2630 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2631 | /* Send a crossed SYN-ACK during socket establishment. |
| 2632 | * WARNING: This routine must only be called when we have already sent |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2633 | * a SYN packet that crossed the incoming SYN that caused this routine |
| 2634 | * to get called. If this assumption fails then the initial rcv_wnd |
| 2635 | * and rcv_wscale values will not be correct. |
| 2636 | */ |
| 2637 | int tcp_send_synack(struct sock *sk) |
| 2638 | { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2639 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2640 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2641 | skb = tcp_write_queue_head(sk); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2642 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
Joe Perches | 91df42b | 2012-05-15 14:11:54 +0000 | [diff] [blame] | 2643 | pr_debug("%s: wrong queue state\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2644 | return -EFAULT; |
| 2645 | } |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2646 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2647 | if (skb_cloned(skb)) { |
| 2648 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); |
| 2649 | if (nskb == NULL) |
| 2650 | return -ENOMEM; |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2651 | tcp_unlink_write_queue(skb, sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2652 | skb_header_release(nskb); |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 2653 | __tcp_add_write_queue_head(sk, nskb); |
Hideo Aoki | 3ab224b | 2007-12-31 00:11:19 -0800 | [diff] [blame] | 2654 | sk_wmem_free_skb(sk, skb); |
| 2655 | sk->sk_wmem_queued += nskb->truesize; |
| 2656 | sk_mem_charge(sk, nskb->truesize); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2657 | skb = nskb; |
| 2658 | } |
| 2659 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 2660 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2661 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
| 2662 | } |
| 2663 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 2664 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2665 | } |
| 2666 | |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2667 | /** |
| 2668 | * tcp_make_synack - Prepare a SYN-ACK. |
| 2669 | * sk: listener socket |
| 2670 | * dst: dst entry attached to the SYNACK |
| 2671 | * req: request_sock pointer |
| 2672 | * rvp: request_values pointer |
| 2673 | * |
| 2674 | * Allocate one skb and build a SYNACK packet. |
| 2675 | * @dst is consumed : Caller should not use it again. |
| 2676 | */ |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 2677 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
William Allen Simpson | e6b4d11 | 2009-12-02 18:07:39 +0000 | [diff] [blame] | 2678 | struct request_sock *req, |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 2679 | struct request_values *rvp, |
| 2680 | struct tcp_fastopen_cookie *foc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2681 | { |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2682 | struct tcp_out_options opts; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2683 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2684 | struct inet_request_sock *ireq = inet_rsk(req); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2685 | struct tcp_sock *tp = tcp_sk(sk); |
Eric Dumazet | 28b2774 | 2010-03-08 11:32:01 -0800 | [diff] [blame] | 2686 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2687 | struct tcphdr *th; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2688 | struct sk_buff *skb; |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2689 | struct tcp_md5sig_key *md5; |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2690 | int tcp_header_size; |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2691 | int mss; |
Eric Dumazet | 28b2774 | 2010-03-08 11:32:01 -0800 | [diff] [blame] | 2692 | int s_data_desired = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2693 | |
Eric Dumazet | 28b2774 | 2010-03-08 11:32:01 -0800 | [diff] [blame] | 2694 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
| 2695 | s_data_desired = cvp->s_data_desired; |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 2696 | skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, |
| 2697 | sk_gfp_atomic(sk, GFP_ATOMIC)); |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2698 | if (unlikely(!skb)) { |
| 2699 | dst_release(dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2700 | return NULL; |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2701 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2702 | /* Reserve space for headers. */ |
| 2703 | skb_reserve(skb, MAX_TCP_HEADER); |
| 2704 | |
Eric Dumazet | 4aea39c | 2012-06-03 20:33:21 +0000 | [diff] [blame] | 2705 | skb_dst_set(skb, dst); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2706 | |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 2707 | mss = dst_metric_advmss(dst); |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2708 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
| 2709 | mss = tp->rx_opt.user_mss; |
| 2710 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2711 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
| 2712 | __u8 rcv_wscale; |
| 2713 | /* Set this up on the first call only */ |
| 2714 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); |
Hagen Paul Pfeifer | e88c64f | 2010-08-19 06:33:05 +0000 | [diff] [blame] | 2715 | |
| 2716 | /* limit the window selection if the user enforce a smaller rx buffer */ |
| 2717 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && |
| 2718 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) |
| 2719 | req->window_clamp = tcp_full_space(sk); |
| 2720 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2721 | /* tcp_full_space because it is guaranteed to be the first packet */ |
| 2722 | tcp_select_initial_window(tcp_full_space(sk), |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2723 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2724 | &req->rcv_wnd, |
| 2725 | &req->window_clamp, |
| 2726 | ireq->wscale_ok, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 2727 | &rcv_wscale, |
| 2728 | dst_metric(dst, RTAX_INITRWND)); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2729 | ireq->rcv_wscale = rcv_wscale; |
| 2730 | } |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2731 | |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2732 | memset(&opts, 0, sizeof(opts)); |
Florian Westphal | 8b5f12d | 2008-10-26 23:10:12 -0700 | [diff] [blame] | 2733 | #ifdef CONFIG_SYN_COOKIES |
| 2734 | if (unlikely(req->cookie_ts)) |
| 2735 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); |
| 2736 | else |
| 2737 | #endif |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2738 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2739 | tcp_header_size = tcp_synack_options(sk, req, mss, |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 2740 | skb, &opts, &md5, xvp, foc) |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2741 | + sizeof(*th); |
Adam Langley | 33ad798 | 2008-07-19 00:04:31 -0700 | [diff] [blame] | 2742 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2743 | skb_push(skb, tcp_header_size); |
| 2744 | skb_reset_transport_header(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2745 | |
Arnaldo Carvalho de Melo | aa8223c | 2007-04-10 21:04:22 -0700 | [diff] [blame] | 2746 | th = tcp_hdr(skb); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2747 | memset(th, 0, sizeof(struct tcphdr)); |
| 2748 | th->syn = 1; |
| 2749 | th->ack = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2750 | TCP_ECN_make_synack(req, th); |
KOVACS Krisztian | a3116ac5 | 2008-10-01 07:46:49 -0700 | [diff] [blame] | 2751 | th->source = ireq->loc_port; |
Arnaldo Carvalho de Melo | 2e6599c | 2005-06-18 22:46:52 -0700 | [diff] [blame] | 2752 | th->dest = ireq->rmt_port; |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 2753 | /* Setting of flags are superfluous here for callers (and ECE is |
| 2754 | * not even correctly set) |
| 2755 | */ |
| 2756 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 2757 | TCPHDR_SYN | TCPHDR_ACK); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2758 | |
| 2759 | if (OPTION_COOKIE_EXTENSION & opts.options) { |
Eric Dumazet | 28b2774 | 2010-03-08 11:32:01 -0800 | [diff] [blame] | 2760 | if (s_data_desired) { |
| 2761 | u8 *buf = skb_put(skb, s_data_desired); |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2762 | |
| 2763 | /* copy data directly from the listening socket. */ |
Eric Dumazet | 28b2774 | 2010-03-08 11:32:01 -0800 | [diff] [blame] | 2764 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
| 2765 | TCP_SKB_CB(skb)->end_seq += s_data_desired; |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2766 | } |
| 2767 | |
| 2768 | if (opts.hash_size > 0) { |
| 2769 | __u32 workspace[SHA_WORKSPACE_WORDS]; |
| 2770 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; |
| 2771 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; |
| 2772 | |
| 2773 | /* Secret recipe depends on the Timestamp, (future) |
| 2774 | * Sequence and Acknowledgment Numbers, Initiator |
| 2775 | * Cookie, and others handled by IP variant caller. |
| 2776 | */ |
| 2777 | *tail-- ^= opts.tsval; |
| 2778 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; |
| 2779 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; |
| 2780 | |
| 2781 | /* recommended */ |
Eric Dumazet | 0eae88f | 2010-04-20 19:06:52 -0700 | [diff] [blame] | 2782 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
David S. Miller | f9a2e69 | 2009-12-02 22:12:04 -0800 | [diff] [blame] | 2783 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
William Allen Simpson | 4957faade | 2009-12-02 18:25:27 +0000 | [diff] [blame] | 2784 | |
| 2785 | sha_transform((__u32 *)&xvp->cookie_bakery[0], |
| 2786 | (char *)mess, |
| 2787 | &workspace[0]); |
| 2788 | opts.hash_location = |
| 2789 | (__u8 *)&xvp->cookie_bakery[0]; |
| 2790 | } |
| 2791 | } |
| 2792 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2793 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
Jerry Chu | 8336886 | 2012-08-31 12:29:12 +0000 | [diff] [blame] | 2794 | /* XXX data is queued and acked as is. No buffer/window check */ |
| 2795 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2796 | |
| 2797 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ |
Ilpo Järvinen | 600ff0c | 2007-02-13 12:42:11 -0800 | [diff] [blame] | 2798 | th->window = htons(min(req->rcv_wnd, 65535U)); |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2799 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2800 | th->doff = (tcp_header_size >> 2); |
Tom Herbert | aa2ea05 | 2010-04-22 07:00:24 +0000 | [diff] [blame] | 2801 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2802 | |
| 2803 | #ifdef CONFIG_TCP_MD5SIG |
| 2804 | /* Okay, we have all we need - do the md5 hash if needed */ |
| 2805 | if (md5) { |
William Allen Simpson | bd0388a | 2009-12-02 18:23:05 +0000 | [diff] [blame] | 2806 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
Adam Langley | 49a72df | 2008-07-19 00:01:42 -0700 | [diff] [blame] | 2807 | md5, NULL, req, skb); |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2808 | } |
| 2809 | #endif |
| 2810 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2811 | return skb; |
| 2812 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 2813 | EXPORT_SYMBOL(tcp_make_synack); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2814 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2815 | /* Do all connect socket setups that can be done AF independent. */ |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 2816 | void tcp_connect_init(struct sock *sk) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2817 | { |
Eric Dumazet | cf533ea | 2011-10-21 05:22:42 -0400 | [diff] [blame] | 2818 | const struct dst_entry *dst = __sk_dst_get(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2819 | struct tcp_sock *tp = tcp_sk(sk); |
| 2820 | __u8 rcv_wscale; |
| 2821 | |
| 2822 | /* We'll fix this up when we get a response from the other end. |
| 2823 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. |
| 2824 | */ |
| 2825 | tp->tcp_header_len = sizeof(struct tcphdr) + |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 2826 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2827 | |
YOSHIFUJI Hideaki | cfb6eeb | 2006-11-14 19:07:45 -0800 | [diff] [blame] | 2828 | #ifdef CONFIG_TCP_MD5SIG |
| 2829 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) |
| 2830 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
| 2831 | #endif |
| 2832 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2833 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
| 2834 | if (tp->rx_opt.user_mss) |
| 2835 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2836 | tp->max_window = 0; |
John Heffner | 5d424d5 | 2006-03-20 17:53:41 -0800 | [diff] [blame] | 2837 | tcp_mtup_init(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2838 | tcp_sync_mss(sk, dst_mtu(dst)); |
| 2839 | |
| 2840 | if (!tp->window_clamp) |
| 2841 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); |
David S. Miller | 0dbaee3 | 2010-12-13 12:52:14 -0800 | [diff] [blame] | 2842 | tp->advmss = dst_metric_advmss(dst); |
Tom Quetchenbach | f5fff5d | 2008-09-21 00:21:51 -0700 | [diff] [blame] | 2843 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
| 2844 | tp->advmss = tp->rx_opt.user_mss; |
| 2845 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2846 | tcp_initialize_rcv_mss(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2847 | |
Hagen Paul Pfeifer | e88c64f | 2010-08-19 06:33:05 +0000 | [diff] [blame] | 2848 | /* limit the window selection if the user enforce a smaller rx buffer */ |
| 2849 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && |
| 2850 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) |
| 2851 | tp->window_clamp = tcp_full_space(sk); |
| 2852 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2853 | tcp_select_initial_window(tcp_full_space(sk), |
| 2854 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
| 2855 | &tp->rcv_wnd, |
| 2856 | &tp->window_clamp, |
David S. Miller | bb5b7c1 | 2009-12-15 20:56:42 -0800 | [diff] [blame] | 2857 | sysctl_tcp_window_scaling, |
laurent chavey | 31d1292 | 2009-12-15 11:15:28 +0000 | [diff] [blame] | 2858 | &rcv_wscale, |
| 2859 | dst_metric(dst, RTAX_INITRWND)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2860 | |
| 2861 | tp->rx_opt.rcv_wscale = rcv_wscale; |
| 2862 | tp->rcv_ssthresh = tp->rcv_wnd; |
| 2863 | |
| 2864 | sk->sk_err = 0; |
| 2865 | sock_reset_flag(sk, SOCK_DONE); |
| 2866 | tp->snd_wnd = 0; |
Hantzis Fotis | ee7537b | 2009-03-02 22:42:02 -0800 | [diff] [blame] | 2867 | tcp_init_wl(tp, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2868 | tp->snd_una = tp->write_seq; |
| 2869 | tp->snd_sml = tp->write_seq; |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 2870 | tp->snd_up = tp->write_seq; |
Pavel Emelyanov | 370816a | 2012-04-19 03:40:01 +0000 | [diff] [blame] | 2871 | tp->snd_nxt = tp->write_seq; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 2872 | |
| 2873 | if (likely(!tp->repair)) |
| 2874 | tp->rcv_nxt = 0; |
| 2875 | tp->rcv_wup = tp->rcv_nxt; |
| 2876 | tp->copied_seq = tp->rcv_nxt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2877 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 2878 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
| 2879 | inet_csk(sk)->icsk_retransmits = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2880 | tcp_clear_retrans(tp); |
| 2881 | } |
| 2882 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2883 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
| 2884 | { |
| 2885 | struct tcp_sock *tp = tcp_sk(sk); |
| 2886 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); |
| 2887 | |
| 2888 | tcb->end_seq += skb->len; |
| 2889 | skb_header_release(skb); |
| 2890 | __tcp_add_write_queue_tail(sk, skb); |
| 2891 | sk->sk_wmem_queued += skb->truesize; |
| 2892 | sk_mem_charge(sk, skb->truesize); |
| 2893 | tp->write_seq = tcb->end_seq; |
| 2894 | tp->packets_out += tcp_skb_pcount(skb); |
| 2895 | } |
| 2896 | |
| 2897 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, |
| 2898 | * queue a data-only packet after the regular SYN, such that regular SYNs |
| 2899 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges |
| 2900 | * only the SYN sequence, the data are retransmitted in the first ACK. |
| 2901 | * If cookie is not cached or other error occurs, falls back to send a |
| 2902 | * regular SYN with Fast Open cookie request option. |
| 2903 | */ |
| 2904 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) |
| 2905 | { |
| 2906 | struct tcp_sock *tp = tcp_sk(sk); |
| 2907 | struct tcp_fastopen_request *fo = tp->fastopen_req; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2908 | int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2909 | struct sk_buff *syn_data = NULL, *data; |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2910 | unsigned long last_syn_loss = 0; |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2911 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2912 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
Yuchung Cheng | aab4874 | 2012-07-19 06:43:10 +0000 | [diff] [blame] | 2913 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
| 2914 | &syn_loss, &last_syn_loss); |
| 2915 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ |
| 2916 | if (syn_loss > 1 && |
| 2917 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { |
| 2918 | fo->cookie.len = -1; |
| 2919 | goto fallback; |
| 2920 | } |
| 2921 | |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2922 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
| 2923 | fo->cookie.len = -1; |
| 2924 | else if (fo->cookie.len <= 0) |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2925 | goto fallback; |
| 2926 | |
| 2927 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and |
| 2928 | * user-MSS. Reserve maximum option space for middleboxes that add |
| 2929 | * private TCP options. The cost is reduced data space in SYN :( |
| 2930 | */ |
| 2931 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) |
| 2932 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; |
| 2933 | space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
| 2934 | MAX_TCP_OPTION_SPACE; |
| 2935 | |
| 2936 | syn_data = skb_copy_expand(syn, skb_headroom(syn), space, |
| 2937 | sk->sk_allocation); |
| 2938 | if (syn_data == NULL) |
| 2939 | goto fallback; |
| 2940 | |
| 2941 | for (i = 0; i < iovlen && syn_data->len < space; ++i) { |
| 2942 | struct iovec *iov = &fo->data->msg_iov[i]; |
| 2943 | unsigned char __user *from = iov->iov_base; |
| 2944 | int len = iov->iov_len; |
| 2945 | |
| 2946 | if (syn_data->len + len > space) |
| 2947 | len = space - syn_data->len; |
| 2948 | else if (i + 1 == iovlen) |
| 2949 | /* No more data pending in inet_wait_for_connect() */ |
| 2950 | fo->data = NULL; |
| 2951 | |
| 2952 | if (skb_add_data(syn_data, from, len)) |
| 2953 | goto fallback; |
| 2954 | } |
| 2955 | |
| 2956 | /* Queue a data-only packet after the regular SYN for retransmission */ |
| 2957 | data = pskb_copy(syn_data, sk->sk_allocation); |
| 2958 | if (data == NULL) |
| 2959 | goto fallback; |
| 2960 | TCP_SKB_CB(data)->seq++; |
| 2961 | TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; |
| 2962 | TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); |
| 2963 | tcp_connect_queue_skb(sk, data); |
| 2964 | fo->copied = data->len; |
| 2965 | |
| 2966 | if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { |
Yuchung Cheng | 67da22d | 2012-07-19 06:43:11 +0000 | [diff] [blame] | 2967 | tp->syn_data = (fo->copied > 0); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 2968 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
| 2969 | goto done; |
| 2970 | } |
| 2971 | syn_data = NULL; |
| 2972 | |
| 2973 | fallback: |
| 2974 | /* Send a regular SYN with Fast Open cookie request option */ |
| 2975 | if (fo->cookie.len > 0) |
| 2976 | fo->cookie.len = 0; |
| 2977 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); |
| 2978 | if (err) |
| 2979 | tp->syn_fastopen = 0; |
| 2980 | kfree_skb(syn_data); |
| 2981 | done: |
| 2982 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ |
| 2983 | return err; |
| 2984 | } |
| 2985 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 2986 | /* Build a SYN and send it off. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2987 | int tcp_connect(struct sock *sk) |
| 2988 | { |
| 2989 | struct tcp_sock *tp = tcp_sk(sk); |
| 2990 | struct sk_buff *buff; |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 2991 | int err; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 2992 | |
| 2993 | tcp_connect_init(sk); |
| 2994 | |
Andrey Vagin | 2b91647 | 2012-11-22 01:13:58 +0000 | [diff] [blame] | 2995 | if (unlikely(tp->repair)) { |
| 2996 | tcp_finish_connect(sk, NULL); |
| 2997 | return 0; |
| 2998 | } |
| 2999 | |
David S. Miller | d179cd1 | 2005-08-17 14:57:30 -0700 | [diff] [blame] | 3000 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3001 | if (unlikely(buff == NULL)) |
| 3002 | return -ENOBUFS; |
| 3003 | |
| 3004 | /* Reserve space for headers. */ |
| 3005 | skb_reserve(buff, MAX_TCP_HEADER); |
| 3006 | |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3007 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3008 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
| 3009 | tcp_connect_queue_skb(sk, buff); |
Ilpo Järvinen | e870a8e | 2008-01-03 20:39:01 -0800 | [diff] [blame] | 3010 | TCP_ECN_send_syn(sk, buff); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3011 | |
Yuchung Cheng | 783237e | 2012-07-19 06:43:07 +0000 | [diff] [blame] | 3012 | /* Send off SYN; include data in Fast Open. */ |
| 3013 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : |
| 3014 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
Eric Paris | ee58681 | 2010-11-16 11:52:49 +0000 | [diff] [blame] | 3015 | if (err == -ECONNREFUSED) |
| 3016 | return err; |
Wei Yongjun | bd37a08 | 2006-08-07 21:04:15 -0700 | [diff] [blame] | 3017 | |
| 3018 | /* We change tp->snd_nxt after the tcp_transmit_skb() call |
| 3019 | * in order to make this packet get counted in tcpOutSegs. |
| 3020 | */ |
| 3021 | tp->snd_nxt = tp->write_seq; |
| 3022 | tp->pushed_seq = tp->write_seq; |
Pavel Emelyanov | 81cc8a7 | 2008-07-16 20:22:04 -0700 | [diff] [blame] | 3023 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3024 | |
| 3025 | /* Timer for repeating the SYN until an answer. */ |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3026 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
| 3027 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3028 | return 0; |
| 3029 | } |
Eric Dumazet | 4bc2f18 | 2010-07-09 21:22:10 +0000 | [diff] [blame] | 3030 | EXPORT_SYMBOL(tcp_connect); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3031 | |
| 3032 | /* Send out a delayed ack, the caller does the policy checking |
| 3033 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() |
| 3034 | * for details. |
| 3035 | */ |
| 3036 | void tcp_send_delayed_ack(struct sock *sk) |
| 3037 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3038 | struct inet_connection_sock *icsk = inet_csk(sk); |
| 3039 | int ato = icsk->icsk_ack.ato; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3040 | unsigned long timeout; |
| 3041 | |
| 3042 | if (ato > TCP_DELACK_MIN) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3043 | const struct tcp_sock *tp = tcp_sk(sk); |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3044 | int max_ato = HZ / 2; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3045 | |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3046 | if (icsk->icsk_ack.pingpong || |
| 3047 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3048 | max_ato = TCP_DELACK_MAX; |
| 3049 | |
| 3050 | /* Slow path, intersegment interval is "high". */ |
| 3051 | |
| 3052 | /* 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] | 3053 | * 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] | 3054 | * directly. |
| 3055 | */ |
| 3056 | if (tp->srtt) { |
Ilpo Järvinen | 056834d | 2007-12-31 14:57:14 -0800 | [diff] [blame] | 3057 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3058 | |
| 3059 | if (rtt < max_ato) |
| 3060 | max_ato = rtt; |
| 3061 | } |
| 3062 | |
| 3063 | ato = min(ato, max_ato); |
| 3064 | } |
| 3065 | |
| 3066 | /* Stay within the limit we were given */ |
| 3067 | timeout = jiffies + ato; |
| 3068 | |
| 3069 | /* 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] | 3070 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3071 | /* If delack timer was blocked or is about to expire, |
| 3072 | * send ACK now. |
| 3073 | */ |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3074 | if (icsk->icsk_ack.blocked || |
| 3075 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3076 | tcp_send_ack(sk); |
| 3077 | return; |
| 3078 | } |
| 3079 | |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3080 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
| 3081 | timeout = icsk->icsk_ack.timeout; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3082 | } |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3083 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
| 3084 | icsk->icsk_ack.timeout = timeout; |
| 3085 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3086 | } |
| 3087 | |
| 3088 | /* This routine sends an ack and also updates the window. */ |
| 3089 | void tcp_send_ack(struct sock *sk) |
| 3090 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3091 | struct sk_buff *buff; |
| 3092 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3093 | /* If we have been reset, we may not send again. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3094 | if (sk->sk_state == TCP_CLOSE) |
| 3095 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3096 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3097 | /* We are not putting this on the write queue, so |
| 3098 | * tcp_transmit_skb() will set the ownership to this |
| 3099 | * sock. |
| 3100 | */ |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3101 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3102 | if (buff == NULL) { |
| 3103 | inet_csk_schedule_ack(sk); |
| 3104 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; |
| 3105 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
| 3106 | TCP_DELACK_MAX, TCP_RTO_MAX); |
| 3107 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3108 | } |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3109 | |
| 3110 | /* Reserve space for headers and prepare control bits. */ |
| 3111 | skb_reserve(buff, MAX_TCP_HEADER); |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3112 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3113 | |
| 3114 | /* Send it off, this clears delayed acks for us. */ |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3115 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3116 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3117 | } |
| 3118 | |
| 3119 | /* This routine sends a packet with an out of date sequence |
| 3120 | * number. It assumes the other end will try to ack it. |
| 3121 | * |
| 3122 | * Question: what should we make while urgent mode? |
| 3123 | * 4.4BSD forces sending single byte of data. We cannot send |
| 3124 | * out of window data, because we have SND.NXT==SND.MAX... |
| 3125 | * |
| 3126 | * Current solution: to send TWO zero-length segments in urgent mode: |
| 3127 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is |
| 3128 | * out-of-date with SND.UNA-1 to probe window. |
| 3129 | */ |
| 3130 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) |
| 3131 | { |
| 3132 | struct tcp_sock *tp = tcp_sk(sk); |
| 3133 | struct sk_buff *skb; |
| 3134 | |
| 3135 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ |
Mel Gorman | 99a1dec | 2012-07-31 16:44:14 -0700 | [diff] [blame] | 3136 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3137 | if (skb == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3138 | return -1; |
| 3139 | |
| 3140 | /* Reserve space for headers and set control bits. */ |
| 3141 | skb_reserve(skb, MAX_TCP_HEADER); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3142 | /* Use a previous sequence. This should cause the other |
| 3143 | * end to send an ack. Don't queue or clone SKB, just |
| 3144 | * send it. |
| 3145 | */ |
Changli Gao | a3433f3 | 2010-06-12 14:01:43 +0000 | [diff] [blame] | 3146 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3147 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
David S. Miller | dfb4b9d | 2005-12-06 16:24:52 -0800 | [diff] [blame] | 3148 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3149 | } |
| 3150 | |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3151 | void tcp_send_window_probe(struct sock *sk) |
| 3152 | { |
| 3153 | if (sk->sk_state == TCP_ESTABLISHED) { |
| 3154 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; |
Pavel Emelyanov | c0e88ff | 2012-04-19 03:41:01 +0000 | [diff] [blame] | 3155 | tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; |
Pavel Emelyanov | ee99528 | 2012-04-19 03:40:39 +0000 | [diff] [blame] | 3156 | tcp_xmit_probe_skb(sk, 0); |
| 3157 | } |
| 3158 | } |
| 3159 | |
Andi Kleen | 67edfef | 2009-07-21 23:00:40 +0000 | [diff] [blame] | 3160 | /* Initiate keepalive or window probe from timer. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3161 | int tcp_write_wakeup(struct sock *sk) |
| 3162 | { |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3163 | struct tcp_sock *tp = tcp_sk(sk); |
| 3164 | struct sk_buff *skb; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3165 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3166 | if (sk->sk_state == TCP_CLOSE) |
| 3167 | return -1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3168 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3169 | if ((skb = tcp_send_head(sk)) != NULL && |
| 3170 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { |
| 3171 | int err; |
Ilpo Järvinen | 0c54b85 | 2009-03-14 14:23:05 +0000 | [diff] [blame] | 3172 | unsigned int mss = tcp_current_mss(sk); |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3173 | 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] | 3174 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3175 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) |
| 3176 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3177 | |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3178 | /* We are probing the opening of a window |
| 3179 | * but the window size is != 0 |
| 3180 | * must have been a result SWS avoidance ( sender ) |
| 3181 | */ |
| 3182 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || |
| 3183 | skb->len > mss) { |
| 3184 | seg_size = min(seg_size, mss); |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3185 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3186 | if (tcp_fragment(sk, skb, seg_size, mss)) |
| 3187 | return -1; |
| 3188 | } else if (!tcp_skb_pcount(skb)) |
| 3189 | tcp_set_skb_tso_segs(sk, skb, mss); |
| 3190 | |
Eric Dumazet | 4de075e | 2011-09-27 13:25:05 -0400 | [diff] [blame] | 3191 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3192 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
| 3193 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
| 3194 | if (!err) |
| 3195 | tcp_event_new_data_sent(sk, skb); |
| 3196 | return err; |
| 3197 | } else { |
Ilpo Järvinen | 33f5f57 | 2008-10-07 14:43:06 -0700 | [diff] [blame] | 3198 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
Ilpo Järvinen | 058dc33 | 2007-12-31 04:51:11 -0800 | [diff] [blame] | 3199 | tcp_xmit_probe_skb(sk, 1); |
| 3200 | return tcp_xmit_probe_skb(sk, 0); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3201 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3202 | } |
| 3203 | |
| 3204 | /* A window probe timeout has occurred. If window is not closed send |
| 3205 | * a partial packet else a zero probe. |
| 3206 | */ |
| 3207 | void tcp_send_probe0(struct sock *sk) |
| 3208 | { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3209 | struct inet_connection_sock *icsk = inet_csk(sk); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3210 | struct tcp_sock *tp = tcp_sk(sk); |
| 3211 | int err; |
| 3212 | |
| 3213 | err = tcp_write_wakeup(sk); |
| 3214 | |
David S. Miller | fe067e8 | 2007-03-07 12:12:44 -0800 | [diff] [blame] | 3215 | if (tp->packets_out || !tcp_send_head(sk)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3216 | /* Cancel probe timer, if it is not required. */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3217 | icsk->icsk_probes_out = 0; |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3218 | icsk->icsk_backoff = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3219 | return; |
| 3220 | } |
| 3221 | |
| 3222 | if (err <= 0) { |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3223 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
| 3224 | icsk->icsk_backoff++; |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3225 | icsk->icsk_probes_out++; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3226 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3227 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
| 3228 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3229 | } else { |
| 3230 | /* If packet was not sent due to local congestion, |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3231 | * do not backoff and do not remember icsk_probes_out. |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3232 | * Let local senders to fight for local resources. |
| 3233 | * |
| 3234 | * Use accumulated backoff yet. |
| 3235 | */ |
Arnaldo Carvalho de Melo | 6687e98 | 2005-08-10 04:03:31 -0300 | [diff] [blame] | 3236 | if (!icsk->icsk_probes_out) |
| 3237 | icsk->icsk_probes_out = 1; |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 3238 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
Arnaldo Carvalho de Melo | 463c84b | 2005-08-09 20:10:42 -0700 | [diff] [blame] | 3239 | min(icsk->icsk_rto << icsk->icsk_backoff, |
Arnaldo Carvalho de Melo | 3f421ba | 2005-08-09 20:11:08 -0700 | [diff] [blame] | 3240 | TCP_RESOURCE_PROBE_INTERVAL), |
| 3241 | TCP_RTO_MAX); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 3242 | } |
| 3243 | } |