blob: 86a0216fcaa1c54ed25a2a56bba2647e156a935c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * 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 Perches91df42b2012-05-15 14:11:54 +000037#define pr_fmt(fmt) "TCP: " fmt
38
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <net/tcp.h>
40
41#include <linux/compiler.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090042#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070046int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Jianjun Kong09cb1052008-11-03 00:27:11 -080048/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080049 * interpret the window field as a signed quantity.
50 */
Brian Haleyab32ea52006-09-22 14:15:41 -070051int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080052
Eric Dumazet46d3cea2012-07-11 05:50:31 +000053/* Default TSQ limit of two TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
55
Linus Torvalds1da177e2005-04-16 15:20:36 -070056/* 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 Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
Brian Haleyab32ea52006-09-22 14:15:41 -070062int sysctl_tcp_mtu_probing __read_mostly = 0;
Shan Wei97b1ce252010-12-01 18:04:50 +000063int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
John Heffner5d424d52006-03-20 17:53:41 -080064
David S. Miller35089bb2006-06-13 22:33:04 -070065/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070066int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070067
Eric Dumazetc9bee3b72013-07-22 20:27:07 -070068unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
69EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
70
Eric Dumazet46d3cea2012-07-11 05:50:31 +000071static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
72 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000073
Andi Kleen67edfef2009-07-21 23:00:40 +000074/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040075static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070076{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000077 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070078 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080079 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070080
David S. Millerfe067e82007-03-07 12:12:44 -080081 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070082 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080083
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080084 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000085 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000086 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000087 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000088 }
Yuchung Chengf19c29e2014-03-03 12:31:36 -080089
David S. Millerf7324ac2014-03-06 15:03:17 -050090 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
91 tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -070092}
93
94/* SND.NXT, if window was not shrunk.
95 * If window has been shrunk, what should we make? It is not clear at all.
96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
98 * invalid. OK, let's make this for now:
99 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400100static inline __u32 tcp_acceptable_seq(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101{
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400102 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700103
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800104 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105 return tp->snd_nxt;
106 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800107 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108}
109
110/* Calculate mss to advertise in SYN segment.
111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
112 *
113 * 1. It is independent of path mtu.
114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
116 * attached devices, because some buggy hosts are confused by
117 * large MSS.
118 * 4. We do not make 3, we advertise MSS, calculated from first
119 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
120 * This may be overridden via information stored in routing table.
121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
122 * probably even Jumbo".
123 */
124static __u16 tcp_advertise_mss(struct sock *sk)
125{
126 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400127 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 int mss = tp->advmss;
129
David S. Miller0dbaee32010-12-13 12:52:14 -0800130 if (dst) {
131 unsigned int metric = dst_metric_advmss(dst);
132
133 if (metric < mss) {
134 mss = metric;
135 tp->advmss = mss;
136 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 }
138
139 return (__u16)mss;
140}
141
142/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
143 * This is the first part of cwnd validation mechanism. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400144static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700146 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 s32 delta = tcp_time_stamp - tp->lsndtime;
148 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
149 u32 cwnd = tp->snd_cwnd;
150
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300151 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300153 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 restart_cwnd = min(restart_cwnd, cwnd);
155
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700156 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157 cwnd >>= 1;
158 tp->snd_cwnd = max(cwnd, restart_cwnd);
159 tp->snd_cwnd_stamp = tcp_time_stamp;
160 tp->snd_cwnd_used = 0;
161}
162
Andi Kleen67edfef2009-07-21 23:00:40 +0000163/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800164static void tcp_event_data_sent(struct tcp_sock *tp,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400165 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700167 struct inet_connection_sock *icsk = inet_csk(sk);
168 const u32 now = tcp_time_stamp;
Cong Wangbcefe172013-06-15 09:39:18 +0800169 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170
David S. Miller35089bb2006-06-13 22:33:04 -0700171 if (sysctl_tcp_slow_start_after_idle &&
172 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700173 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174
175 tp->lsndtime = now;
176
177 /* If it is a reply for ato after last received
178 * packet, enter pingpong mode.
179 */
Cong Wangbcefe172013-06-15 09:39:18 +0800180 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
181 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
182 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183}
184
Andi Kleen67edfef2009-07-21 23:00:40 +0000185/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800186static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700188 tcp_dec_quickack_mode(sk, pkts);
189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190}
191
Yuchung Cheng85f16522013-06-11 15:35:32 -0700192
193u32 tcp_default_init_rwnd(u32 mss)
194{
195 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800196 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700197 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
198 * limit when mss is larger than 1460.
199 */
200 u32 init_rwnd = TCP_INIT_CWND * 2;
201
202 if (mss > 1460)
203 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 return init_rwnd;
205}
206
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207/* Determine a window scaling and initial window to offer.
208 * Based on the assumption that the given amount of space
209 * will be offered. Store the results in the tp structure.
210 * NOTE: for smooth operation initial space offering should
211 * be a multiple of mss if possible. We assume here that mss >= 1.
212 * This MUST be enforced by all callers.
213 */
214void tcp_select_initial_window(int __space, __u32 mss,
215 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000216 int wscale_ok, __u8 *rcv_wscale,
217 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218{
219 unsigned int space = (__space < 0 ? 0 : __space);
220
221 /* If no clamp set the clamp to the max possible scaled window */
222 if (*window_clamp == 0)
223 (*window_clamp) = (65535 << 14);
224 space = min(*window_clamp, space);
225
226 /* Quantize space offering to a multiple of mss if possible. */
227 if (space > mss)
228 space = (space / mss) * mss;
229
230 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800231 * will break some buggy TCP stacks. If the admin tells us
232 * it is likely we could be speaking with such a buggy stack
233 * we will truncate our initial window offering to 32K-1
234 * unless the remote has sent us a window scaling option,
235 * which we interpret as a sign the remote TCP is not
236 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 */
Rick Jones15d99e02006-03-20 22:40:29 -0800238 if (sysctl_tcp_workaround_signed_windows)
239 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
240 else
241 (*rcv_wnd) = space;
242
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 (*rcv_wscale) = 0;
244 if (wscale_ok) {
245 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900246 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 */
248 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700249 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250 while (space > 65535 && (*rcv_wscale) < 14) {
251 space >>= 1;
252 (*rcv_wscale)++;
253 }
254 }
255
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800256 if (mss > (1 << *rcv_wscale)) {
Yuchung Cheng85f16522013-06-11 15:35:32 -0700257 if (!init_rcv_wnd) /* Use default unless specified otherwise */
258 init_rcv_wnd = tcp_default_init_rwnd(mss);
259 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 }
261
262 /* Set the clamp no higher than max representable value */
263 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
264}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000265EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266
267/* Chose a new window to advertise, update state in tcp_sock for the
268 * socket, and return result with RFC1323 scaling applied. The return
269 * value can be stuffed directly into th->window for an outgoing
270 * frame.
271 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800272static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273{
274 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphal8e165e22014-02-25 14:34:32 +0100275 u32 old_win = tp->rcv_wnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 u32 cur_win = tcp_receive_window(tp);
277 u32 new_win = __tcp_select_window(sk);
278
279 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800280 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 /* Danger Will Robinson!
282 * Don't update rcv_wup/rcv_wnd here or else
283 * we will not be able to advertise a zero
284 * window in time. --DaveM
285 *
286 * Relax Will Robinson.
287 */
Florian Westphal8e165e22014-02-25 14:34:32 +0100288 if (new_win == 0)
289 NET_INC_STATS(sock_net(sk),
290 LINUX_MIB_TCPWANTZEROWINDOWADV);
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700291 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 }
293 tp->rcv_wnd = new_win;
294 tp->rcv_wup = tp->rcv_nxt;
295
296 /* Make sure we do not exceed the maximum possible
297 * scaled window.
298 */
Rick Jones15d99e02006-03-20 22:40:29 -0800299 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 new_win = min(new_win, MAX_TCP_WINDOW);
301 else
302 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
303
304 /* RFC1323 scaling applied */
305 new_win >>= tp->rx_opt.rcv_wscale;
306
307 /* If we advertise zero window, disable fast path. */
Florian Westphal8e165e22014-02-25 14:34:32 +0100308 if (new_win == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 tp->pred_flags = 0;
Florian Westphal8e165e22014-02-25 14:34:32 +0100310 if (old_win)
311 NET_INC_STATS(sock_net(sk),
312 LINUX_MIB_TCPTOZEROWINDOWADV);
313 } else if (old_win == 0) {
314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316
317 return new_win;
318}
319
Andi Kleen67edfef2009-07-21 23:00:40 +0000320/* Packet ECN state for a SYN-ACK */
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200321static inline void TCP_ECN_send_synack(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700322{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200323 const struct tcp_sock *tp = tcp_sk(sk);
324
Eric Dumazet4de075e2011-09-27 13:25:05 -0400325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800326 if (!(tp->ecn_flags & TCP_ECN_OK))
Eric Dumazet4de075e2011-09-27 13:25:05 -0400327 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200328 else if (tcp_ca_needs_ecn(sk))
329 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700330}
331
Andi Kleen67edfef2009-07-21 23:00:40 +0000332/* Packet ECN state for a SYN. */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700333static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
334{
335 struct tcp_sock *tp = tcp_sk(sk);
336
337 tp->ecn_flags = 0;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200338 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
339 tcp_ca_needs_ecn(sk)) {
Eric Dumazet4de075e2011-09-27 13:25:05 -0400340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700341 tp->ecn_flags = TCP_ECN_OK;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200342 if (tcp_ca_needs_ecn(sk))
343 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700344 }
345}
346
347static __inline__ void
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200348TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th,
349 struct sock *sk)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700350{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200351 if (inet_rsk(req)->ecn_ok) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700352 th->ece = 1;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200353 if (tcp_ca_needs_ecn(sk))
354 INET_ECN_xmit(sk);
355 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700356}
357
Andi Kleen67edfef2009-07-21 23:00:40 +0000358/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
359 * be sent.
360 */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700361static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
362 int tcp_header_len)
363{
364 struct tcp_sock *tp = tcp_sk(sk);
365
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
370 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
373 tcp_hdr(skb)->cwr = 1;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
375 }
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200376 } else if (!tcp_ca_needs_ecn(sk)) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
379 }
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
381 tcp_hdr(skb)->ece = 1;
382 }
383}
384
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800385/* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
387 */
388static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
389{
Eric Dumazet7b7fc972013-12-05 22:31:30 -0800390 struct skb_shared_info *shinfo = skb_shinfo(skb);
391
David S. Miller2e8e18e2010-04-08 11:32:30 -0700392 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800393 skb->csum = 0;
394
Eric Dumazet4de075e2011-09-27 13:25:05 -0400395 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800396 TCP_SKB_CB(skb)->sacked = 0;
397
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700398 tcp_skb_pcount_set(skb, 1);
Eric Dumazet7b7fc972013-12-05 22:31:30 -0800399 shinfo->gso_size = 0;
400 shinfo->gso_type = 0;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800401
402 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000403 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800404 seq++;
405 TCP_SKB_CB(skb)->end_seq = seq;
406}
407
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000408static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700409{
410 return tp->snd_una != tp->snd_up;
411}
412
Adam Langley33ad7982008-07-19 00:04:31 -0700413#define OPTION_SACK_ADVERTISE (1 << 0)
414#define OPTION_TS (1 << 1)
415#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000416#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000417#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700418
419struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000420 u16 options; /* bit field of OPTION_* */
421 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700422 u8 ws; /* window scale, 0 to disable */
423 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000424 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000425 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000426 __u32 tsval, tsecr; /* need to include OPTION_TS */
427 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700428};
429
Andi Kleen67edfef2009-07-21 23:00:40 +0000430/* Write previously computed TCP options to the packet.
431 *
432 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700433 * TCP options, we learned this through the hard way, so be careful here.
434 * Luckily we can at least blame others for their non-compliance but from
stephen hemminger8e3bff92013-12-08 12:15:44 -0800435 * inter-operability perspective it seems that we're somewhat stuck with
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700436 * the ordering which we have been using if we want to keep working with
437 * those broken things (not that it currently hurts anybody as there isn't
438 * particular reason why the ordering would need to be changed).
439 *
440 * At least SACK_PERM as the first option is known to lead to a disaster
441 * (but it may well be that other scenarios fail similarly).
442 */
Adam Langley33ad7982008-07-19 00:04:31 -0700443static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000444 struct tcp_out_options *opts)
445{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000446 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000447
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000448 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000449 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
450 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000451 /* overload cookie hash location */
452 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700453 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800454 }
Adam Langley33ad7982008-07-19 00:04:31 -0700455
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700456 if (unlikely(opts->mss)) {
457 *ptr++ = htonl((TCPOPT_MSS << 24) |
458 (TCPOLEN_MSS << 16) |
459 opts->mss);
460 }
461
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000462 if (likely(OPTION_TS & options)) {
463 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700464 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
465 (TCPOLEN_SACK_PERM << 16) |
466 (TCPOPT_TIMESTAMP << 8) |
467 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000468 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700469 } else {
470 *ptr++ = htonl((TCPOPT_NOP << 24) |
471 (TCPOPT_NOP << 16) |
472 (TCPOPT_TIMESTAMP << 8) |
473 TCPOLEN_TIMESTAMP);
474 }
475 *ptr++ = htonl(opts->tsval);
476 *ptr++ = htonl(opts->tsecr);
477 }
478
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000479 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700480 *ptr++ = htonl((TCPOPT_NOP << 24) |
481 (TCPOPT_NOP << 16) |
482 (TCPOPT_SACK_PERM << 8) |
483 TCPOLEN_SACK_PERM);
484 }
485
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000486 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700487 *ptr++ = htonl((TCPOPT_NOP << 24) |
488 (TCPOPT_WINDOW << 16) |
489 (TCPOLEN_WINDOW << 8) |
490 opts->ws);
491 }
492
493 if (unlikely(opts->num_sack_blocks)) {
494 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
495 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800496 int this_sack;
497
498 *ptr++ = htonl((TCPOPT_NOP << 24) |
499 (TCPOPT_NOP << 16) |
500 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700501 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800502 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800503
Adam Langley33ad7982008-07-19 00:04:31 -0700504 for (this_sack = 0; this_sack < opts->num_sack_blocks;
505 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800506 *ptr++ = htonl(sp[this_sack].start_seq);
507 *ptr++ = htonl(sp[this_sack].end_seq);
508 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800509
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000510 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800511 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000512
513 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
514 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
515
516 *ptr++ = htonl((TCPOPT_EXP << 24) |
517 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
518 TCPOPT_FASTOPEN_MAGIC);
519
520 memcpy(ptr, foc->val, foc->len);
521 if ((foc->len & 3) == 2) {
522 u8 *align = ((u8 *)ptr) + foc->len;
523 align[0] = align[1] = TCPOPT_NOP;
524 }
525 ptr += (foc->len + 3) >> 2;
526 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800527}
528
Andi Kleen67edfef2009-07-21 23:00:40 +0000529/* Compute TCP options for SYN packets. This is not the final
530 * network wire format yet.
531 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000532static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700533 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400534 struct tcp_md5sig_key **md5)
535{
Adam Langley33ad7982008-07-19 00:04:31 -0700536 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000537 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000538 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700539
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800540#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700541 *md5 = tp->af_specific->md5_lookup(sk, sk);
542 if (*md5) {
543 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000544 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800545 }
Adam Langley33ad7982008-07-19 00:04:31 -0700546#else
547 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800548#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700549
550 /* We always get an MSS option. The option bytes which will be seen in
551 * normal data packets should timestamps be used, must be in the MSS
552 * advertised. But we subtract them from tp->mss_cache so that
553 * calculations in tcp_sendmsg are simpler etc. So account for this
554 * fact here if necessary. If we don't do this correctly, as a
555 * receiver we won't recognize data packets as being full sized when we
556 * should, and thus we won't abide by the delayed ACK rules correctly.
557 * SACKs don't matter, we never delay an ACK when we have any of those
558 * going out. */
559 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000560 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700561
David S. Millerbb5b7c12009-12-15 20:56:42 -0800562 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700563 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700564 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700565 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000566 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700567 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800568 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700569 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000570 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000571 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700572 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800573 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700574 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700575 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000576 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700577 }
578
Yuchung Cheng783237e2012-07-19 06:43:07 +0000579 if (fastopen && fastopen->cookie.len >= 0) {
580 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
581 need = (need + 3) & ~3U; /* Align to 32 bits */
582 if (remaining >= need) {
583 opts->options |= OPTION_FAST_OPEN_COOKIE;
584 opts->fastopen_cookie = &fastopen->cookie;
585 remaining -= need;
586 tp->syn_fastopen = 1;
587 }
588 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000589
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000590 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700591}
592
Andi Kleen67edfef2009-07-21 23:00:40 +0000593/* Set up TCP options for SYN-ACKs. */
Eric Dumazet95c96172012-04-15 05:58:06 +0000594static unsigned int tcp_synack_options(struct sock *sk,
Adam Langley33ad7982008-07-19 00:04:31 -0700595 struct request_sock *req,
Eric Dumazet95c96172012-04-15 05:58:06 +0000596 unsigned int mss, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700597 struct tcp_out_options *opts,
William Allen Simpson4957faade2009-12-02 18:25:27 +0000598 struct tcp_md5sig_key **md5,
Jerry Chu83368862012-08-31 12:29:12 +0000599 struct tcp_fastopen_cookie *foc)
William Allen Simpson4957faade2009-12-02 18:25:27 +0000600{
Adam Langley33ad7982008-07-19 00:04:31 -0700601 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet95c96172012-04-15 05:58:06 +0000602 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Adam Langley33ad7982008-07-19 00:04:31 -0700603
604#ifdef CONFIG_TCP_MD5SIG
605 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
606 if (*md5) {
607 opts->options |= OPTION_MD5;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000608 remaining -= TCPOLEN_MD5SIG_ALIGNED;
609
610 /* We can't fit any SACK blocks in a packet with MD5 + TS
611 * options. There was discussion about disabling SACK
612 * rather than TS in order to fit in better with old,
613 * buggy kernels, but that was deemed to be unnecessary.
614 */
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700615 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700616 }
617#else
618 *md5 = NULL;
619#endif
620
William Allen Simpson4957faade2009-12-02 18:25:27 +0000621 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700622 opts->mss = mss;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000623 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700624
625 if (likely(ireq->wscale_ok)) {
626 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000627 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000628 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700629 }
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700630 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700631 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700632 opts->tsval = tcp_skb_timestamp(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700633 opts->tsecr = req->ts_recent;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000634 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700635 }
636 if (likely(ireq->sack_ok)) {
637 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700638 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faade2009-12-02 18:25:27 +0000639 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700640 }
Yuchung Cheng89278c92014-05-11 20:22:10 -0700641 if (foc != NULL && foc->len >= 0) {
Jerry Chu83368862012-08-31 12:29:12 +0000642 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
643 need = (need + 3) & ~3U; /* Align to 32 bits */
644 if (remaining >= need) {
645 opts->options |= OPTION_FAST_OPEN_COOKIE;
646 opts->fastopen_cookie = foc;
647 remaining -= need;
648 }
649 }
William Allen Simpson4957faade2009-12-02 18:25:27 +0000650
William Allen Simpson4957faade2009-12-02 18:25:27 +0000651 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700652}
653
Andi Kleen67edfef2009-07-21 23:00:40 +0000654/* Compute TCP options for ESTABLISHED sockets. This is not the
655 * final wire format yet.
656 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000657static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700658 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400659 struct tcp_md5sig_key **md5)
660{
Adam Langley33ad7982008-07-19 00:04:31 -0700661 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000662 unsigned int size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000663 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700664
Andi Kleen5843ef42013-09-30 13:29:11 -0700665 opts->options = 0;
666
Adam Langley33ad7982008-07-19 00:04:31 -0700667#ifdef CONFIG_TCP_MD5SIG
668 *md5 = tp->af_specific->md5_lookup(sk, sk);
669 if (unlikely(*md5)) {
670 opts->options |= OPTION_MD5;
671 size += TCPOLEN_MD5SIG_ALIGNED;
672 }
673#else
674 *md5 = NULL;
675#endif
676
677 if (likely(tp->rx_opt.tstamp_ok)) {
678 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700679 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700680 opts->tsecr = tp->rx_opt.ts_recent;
681 size += TCPOLEN_TSTAMP_ALIGNED;
682 }
683
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000684 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
685 if (unlikely(eff_sacks)) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000686 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
Adam Langley33ad7982008-07-19 00:04:31 -0700687 opts->num_sack_blocks =
Eric Dumazet95c96172012-04-15 05:58:06 +0000688 min_t(unsigned int, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700689 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
690 TCPOLEN_SACK_PERBLOCK);
691 size += TCPOLEN_SACK_BASE_ALIGNED +
692 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
693 }
694
695 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800696}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000698
699/* TCP SMALL QUEUES (TSQ)
700 *
701 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
702 * to reduce RTT and bufferbloat.
703 * We do this using a special skb destructor (tcp_wfree).
704 *
705 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
706 * needs to be reallocated in a driver.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800707 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000708 *
709 * Since transmit from skb destructor is forbidden, we use a tasklet
710 * to process all sockets that eventually need to send more skbs.
711 * We use one tasklet per cpu, with its own queue of sockets.
712 */
713struct tsq_tasklet {
714 struct tasklet_struct tasklet;
715 struct list_head head; /* queue of tcp sockets */
716};
717static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
718
Eric Dumazet6f458df2012-07-20 05:45:50 +0000719static void tcp_tsq_handler(struct sock *sk)
720{
721 if ((1 << sk->sk_state) &
722 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
723 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
John Ognessbf062002014-02-09 18:40:11 -0800724 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
725 0, GFP_ATOMIC);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000726}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000727/*
stephen hemminger8e3bff92013-12-08 12:15:44 -0800728 * One tasklet per cpu tries to send more skbs.
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000729 * We run in tasklet context but need to disable irqs when
stephen hemminger8e3bff92013-12-08 12:15:44 -0800730 * transferring tsq->head because tcp_wfree() might
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000731 * interrupt us (non NAPI drivers)
732 */
733static void tcp_tasklet_func(unsigned long data)
734{
735 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
736 LIST_HEAD(list);
737 unsigned long flags;
738 struct list_head *q, *n;
739 struct tcp_sock *tp;
740 struct sock *sk;
741
742 local_irq_save(flags);
743 list_splice_init(&tsq->head, &list);
744 local_irq_restore(flags);
745
746 list_for_each_safe(q, n, &list) {
747 tp = list_entry(q, struct tcp_sock, tsq_node);
748 list_del(&tp->tsq_node);
749
750 sk = (struct sock *)tp;
751 bh_lock_sock(sk);
752
753 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000754 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000755 } else {
756 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000757 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000758 }
759 bh_unlock_sock(sk);
760
761 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
762 sk_free(sk);
763 }
764}
765
Eric Dumazet6f458df2012-07-20 05:45:50 +0000766#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
767 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200768 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
769 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000770/**
771 * tcp_release_cb - tcp release_sock() callback
772 * @sk: socket
773 *
774 * called from release_sock() to perform protocol dependent
775 * actions before socket release.
776 */
777void tcp_release_cb(struct sock *sk)
778{
779 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000780 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000781
Eric Dumazet6f458df2012-07-20 05:45:50 +0000782 /* perform an atomic operation only if at least one flag is set */
783 do {
784 flags = tp->tsq_flags;
785 if (!(flags & TCP_DEFERRED_ALL))
786 return;
787 nflags = flags & ~TCP_DEFERRED_ALL;
788 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
789
790 if (flags & (1UL << TCP_TSQ_DEFERRED))
791 tcp_tsq_handler(sk);
792
Eric Dumazetc3f9b012014-03-10 09:50:11 -0700793 /* Here begins the tricky part :
794 * We are called from release_sock() with :
795 * 1) BH disabled
796 * 2) sk_lock.slock spinlock held
797 * 3) socket owned by us (sk->sk_lock.owned == 1)
798 *
799 * But following code is meant to be called from BH handlers,
800 * so we should keep BH disabled, but early release socket ownership
801 */
802 sock_release_ownership(sk);
803
Eric Dumazet144d56e2012-08-20 00:22:46 +0000804 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000805 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000806 __sock_put(sk);
807 }
808 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000809 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000810 __sock_put(sk);
811 }
812 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Neal Cardwell4fab9072014-08-14 12:40:05 -0400813 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000814 __sock_put(sk);
815 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000816}
817EXPORT_SYMBOL(tcp_release_cb);
818
819void __init tcp_tasklet_init(void)
820{
821 int i;
822
823 for_each_possible_cpu(i) {
824 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
825
826 INIT_LIST_HEAD(&tsq->head);
827 tasklet_init(&tsq->tasklet,
828 tcp_tasklet_func,
829 (unsigned long)tsq);
830 }
831}
832
833/*
834 * Write buffer destructor automatically called from kfree_skb.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800835 * We can't xmit new skbs from this context, as we might already
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000836 * hold qdisc lock.
837 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000838void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000839{
840 struct sock *sk = skb->sk;
841 struct tcp_sock *tp = tcp_sk(sk);
842
843 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
844 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
845 unsigned long flags;
846 struct tsq_tasklet *tsq;
847
848 /* Keep a ref on socket.
849 * This last ref will be released in tcp_tasklet_func()
850 */
851 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
852
853 /* queue this socket to tasklet queue */
854 local_irq_save(flags);
855 tsq = &__get_cpu_var(tsq_tasklet);
856 list_add(&tp->tsq_node, &tsq->head);
857 tasklet_schedule(&tsq->tasklet);
858 local_irq_restore(flags);
859 } else {
860 sock_wfree(skb);
861 }
862}
863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864/* This routine actually transmits TCP packets queued in by
865 * tcp_do_sendmsg(). This is used by both the initial
866 * transmission and possible later retransmissions.
867 * All SKB's seen here are completely headerless. It is our
868 * job to build the TCP header, and pass the packet down to
869 * IP so it can do the same plus pass the packet off to the
870 * device.
871 *
872 * We are working here with either a clone of the original
873 * SKB, or a fresh unique copy made by the retransmit engine.
874 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800875static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
876 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800878 const struct inet_connection_sock *icsk = inet_csk(sk);
879 struct inet_sock *inet;
880 struct tcp_sock *tp;
881 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700882 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000883 unsigned int tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800884 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800885 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800886 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800888 BUG_ON(!skb || !tcp_skb_pcount(skb));
889
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700890 if (clone_it) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800891 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700892
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800893 if (unlikely(skb_cloned(skb)))
894 skb = pskb_copy(skb, gfp_mask);
895 else
896 skb = skb_clone(skb, gfp_mask);
897 if (unlikely(!skb))
898 return -ENOBUFS;
899 }
900
901 inet = inet_sk(sk);
902 tp = tcp_sk(sk);
903 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700904 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700905
Eric Dumazet4de075e2011-09-27 13:25:05 -0400906 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700907 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
908 else
909 tcp_options_size = tcp_established_options(sk, skb, &opts,
910 &md5);
911 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900912
Eric Dumazet547669d2013-05-23 07:44:20 +0000913 if (tcp_packets_in_flight(tp) == 0)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800914 tcp_ca_event(sk, CA_EVENT_TX_START);
Eric Dumazet547669d2013-05-23 07:44:20 +0000915
916 /* if no packet is in qdisc/device queue, then allow XPS to select
917 * another queue.
918 */
919 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800920
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700921 skb_push(skb, tcp_header_size);
922 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000923
924 skb_orphan(skb);
925 skb->sk = sk;
Eric Dumazetc9eeec22013-09-27 03:28:54 -0700926 skb->destructor = tcp_wfree;
Tom Herbertb73c3d02014-07-01 21:32:17 -0700927 skb_set_hash_from_sk(skb, sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000928 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800929
930 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700931 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000932 th->source = inet->inet_sport;
933 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800934 th->seq = htonl(tcb->seq);
935 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700936 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400937 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800938
Eric Dumazet4de075e2011-09-27 13:25:05 -0400939 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800940 /* RFC1323: The window in SYN & SYN/ACK segments
941 * is never scaled.
942 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800943 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800944 } else {
945 th->window = htons(tcp_select_window(sk));
946 }
947 th->check = 0;
948 th->urg_ptr = 0;
949
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700950 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800951 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
952 if (before(tp->snd_up, tcb->seq + 0x10000)) {
953 th->urg_ptr = htons(tp->snd_up - tcb->seq);
954 th->urg = 1;
955 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700956 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800957 th->urg = 1;
958 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800959 }
960
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000961 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet4de075e2011-09-27 13:25:05 -0400962 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700963 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800964
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800965#ifdef CONFIG_TCP_MD5SIG
966 /* Calculate the MD5 hash, as we have all we need now */
967 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -0700968 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000969 tp->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700970 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800971 }
972#endif
973
Herbert Xubb296242010-04-11 02:15:55 +0000974 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800975
Eric Dumazet4de075e2011-09-27 13:25:05 -0400976 if (likely(tcb->tcp_flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800977 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
978
979 if (skb->len != tcp_header_size)
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400980 tcp_event_data_sent(tp, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800981
Wei Yongjunbd37a082006-08-07 21:04:15 -0700982 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +0000983 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
984 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800985
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700986 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
987 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
988
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700989 /* Our usage of tstamp should remain private */
990 skb->tstamp.tv64 = 0;
Eric Dumazet971f10e2014-09-27 09:50:57 -0700991
992 /* Cleanup our debris for IP stacks */
993 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
994 sizeof(struct inet6_skb_parm)));
995
Eric Dumazetb0270e92014-04-15 12:58:34 -0400996 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700997
Hua Zhong83de47c2006-04-28 15:26:50 -0700998 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800999 return err;
1000
Christoph Paasch5ee2c942014-07-14 16:58:32 +02001001 tcp_enter_cwr(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001002
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -02001003 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004}
1005
Andi Kleen67edfef2009-07-21 23:00:40 +00001006/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 *
1008 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1009 * otherwise socket can stall.
1010 */
1011static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1012{
1013 struct tcp_sock *tp = tcp_sk(sk);
1014
1015 /* Advance write_seq and place onto the write_queue. */
1016 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001017 __skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -08001018 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001019 sk->sk_wmem_queued += skb->truesize;
1020 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021}
1022
Andi Kleen67edfef2009-07-21 23:00:40 +00001023/* Initialize TSO segments for a packet. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001024static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001025 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -07001026{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001027 struct skb_shared_info *shinfo = skb_shinfo(skb);
1028
Eric Dumazetc52e24212013-10-15 11:54:30 -07001029 /* Make sure we own this skb before messing gso_size/gso_segs */
1030 WARN_ON_ONCE(skb_cloned(skb));
1031
Eric Dumazet8f26fb12013-10-15 12:24:54 -07001032 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -07001033 /* Avoid the costly divide in the normal
1034 * non-TSO case.
1035 */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001036 tcp_skb_pcount_set(skb, 1);
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001037 shinfo->gso_size = 0;
1038 shinfo->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -07001039 } else {
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001040 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001041 shinfo->gso_size = mss_now;
1042 shinfo->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 }
1044}
1045
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001046/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001047 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001048 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001049static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001050 int decr)
1051{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001052 struct tcp_sock *tp = tcp_sk(sk);
1053
Ilpo Järvinendc869672007-10-01 15:27:19 -07001054 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001055 return;
1056
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001057 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001058 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001059}
1060
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001061/* Pcount in the middle of the write queue got changed, we need to do various
1062 * tweaks to fix counters
1063 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001064static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001065{
1066 struct tcp_sock *tp = tcp_sk(sk);
1067
1068 tp->packets_out -= decr;
1069
1070 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1071 tp->sacked_out -= decr;
1072 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1073 tp->retrans_out -= decr;
1074 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1075 tp->lost_out -= decr;
1076
1077 /* Reno case is special. Sigh... */
1078 if (tcp_is_reno(tp) && decr > 0)
1079 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1080
1081 tcp_adjust_fackets_out(sk, skb, decr);
1082
1083 if (tp->lost_skb_hint &&
1084 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001085 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001086 tp->lost_cnt_hint -= decr;
1087
1088 tcp_verify_left_out(tp);
1089}
1090
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001091static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1092{
1093 struct skb_shared_info *shinfo = skb_shinfo(skb);
1094
1095 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1096 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1097 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1098 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1099
1100 shinfo->tx_flags &= ~tsflags;
1101 shinfo2->tx_flags |= tsflags;
1102 swap(shinfo->tskey, shinfo2->tskey);
1103 }
1104}
1105
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106/* Function to create two new TCP segments. Shrinks the given segment
1107 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001108 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 * Remember, these are still headerless SKBs at this point.
1110 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001111int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001112 unsigned int mss_now, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
1114 struct tcp_sock *tp = tcp_sk(sk);
1115 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001116 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001117 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001118 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001120 if (WARN_ON(len > skb->len))
1121 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001122
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 nsize = skb_headlen(skb) - len;
1124 if (nsize < 0)
1125 nsize = 0;
1126
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001127 if (skb_unclone(skb, gfp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 return -ENOMEM;
1129
1130 /* Get a new skb... force flag on. */
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001131 buff = sk_stream_alloc_skb(sk, nsize, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 if (buff == NULL)
1133 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001134
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001135 sk->sk_wmem_queued += buff->truesize;
1136 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001137 nlen = skb->len - len - nsize;
1138 buff->truesize += nlen;
1139 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140
1141 /* Correct the sequence numbers. */
1142 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1143 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1144 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1145
1146 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001147 flags = TCP_SKB_CB(skb)->tcp_flags;
1148 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1149 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001150 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151
Patrick McHardy84fa7932006-08-29 16:44:56 -07001152 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001154 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1155 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 nsize, 0);
1157
1158 skb_trim(skb, len);
1159
1160 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1161 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001162 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 skb_split(skb, buff, len);
1164 }
1165
1166 buff->ip_summed = skb->ip_summed;
1167
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001168 buff->tstamp = skb->tstamp;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001169 tcp_fragment_tstamp(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170
David S. Miller6475be12005-09-01 22:47:01 -07001171 old_factor = tcp_skb_pcount(skb);
1172
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001174 tcp_set_skb_tso_segs(sk, skb, mss_now);
1175 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176
David S. Miller6475be12005-09-01 22:47:01 -07001177 /* If this packet has been sent out already, we must
1178 * adjust the various packet counters.
1179 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001180 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001181 int diff = old_factor - tcp_skb_pcount(skb) -
1182 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001184 if (diff)
1185 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 }
1187
1188 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001189 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001190 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191
1192 return 0;
1193}
1194
1195/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1196 * eventually). The difference is that pulled data not copied, but
1197 * immediately discarded.
1198 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001199static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001201 struct skb_shared_info *shinfo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 int i, k, eat;
1203
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001204 eat = min_t(int, len, skb_headlen(skb));
1205 if (eat) {
1206 __skb_pull(skb, eat);
1207 len -= eat;
1208 if (!len)
1209 return;
1210 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 eat = len;
1212 k = 0;
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001213 shinfo = skb_shinfo(skb);
1214 for (i = 0; i < shinfo->nr_frags; i++) {
1215 int size = skb_frag_size(&shinfo->frags[i]);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001216
1217 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001218 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001219 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 } else {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001221 shinfo->frags[k] = shinfo->frags[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 if (eat) {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001223 shinfo->frags[k].page_offset += eat;
1224 skb_frag_size_sub(&shinfo->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 eat = 0;
1226 }
1227 k++;
1228 }
1229 }
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001230 shinfo->nr_frags = k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001232 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 skb->data_len -= len;
1234 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235}
1236
Andi Kleen67edfef2009-07-21 23:00:40 +00001237/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1239{
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001240 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 return -ENOMEM;
1242
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001243 __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244
1245 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001246 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
1248 skb->truesize -= len;
1249 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001250 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1252
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001253 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 if (tcp_skb_pcount(skb) > 1)
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001255 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
1257 return 0;
1258}
1259
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001260/* Calculate MSS not accounting any TCP options. */
1261static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001262{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001263 const struct tcp_sock *tp = tcp_sk(sk);
1264 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001265 int mss_now;
1266
1267 /* Calculate base mss without TCP options:
1268 It is MMS_S - sizeof(tcphdr) of rfc1122
1269 */
1270 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1271
Eric Dumazet67469602012-04-24 07:37:38 +00001272 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1273 if (icsk->icsk_af_ops->net_frag_header_len) {
1274 const struct dst_entry *dst = __sk_dst_get(sk);
1275
1276 if (dst && dst_allfrag(dst))
1277 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1278 }
1279
John Heffner5d424d52006-03-20 17:53:41 -08001280 /* Clamp it (mss_clamp does not include tcp options) */
1281 if (mss_now > tp->rx_opt.mss_clamp)
1282 mss_now = tp->rx_opt.mss_clamp;
1283
1284 /* Now subtract optional transport overhead */
1285 mss_now -= icsk->icsk_ext_hdr_len;
1286
1287 /* Then reserve room for full set of TCP options and 8 bytes of data */
1288 if (mss_now < 48)
1289 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001290 return mss_now;
1291}
1292
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001293/* Calculate MSS. Not accounting for SACKs here. */
1294int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1295{
1296 /* Subtract TCP options size, not including SACKs */
1297 return __tcp_mtu_to_mss(sk, pmtu) -
1298 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1299}
1300
John Heffner5d424d52006-03-20 17:53:41 -08001301/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001302int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001303{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001304 const struct tcp_sock *tp = tcp_sk(sk);
1305 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001306 int mtu;
1307
1308 mtu = mss +
1309 tp->tcp_header_len +
1310 icsk->icsk_ext_hdr_len +
1311 icsk->icsk_af_ops->net_header_len;
1312
Eric Dumazet67469602012-04-24 07:37:38 +00001313 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1314 if (icsk->icsk_af_ops->net_frag_header_len) {
1315 const struct dst_entry *dst = __sk_dst_get(sk);
1316
1317 if (dst && dst_allfrag(dst))
1318 mtu += icsk->icsk_af_ops->net_frag_header_len;
1319 }
John Heffner5d424d52006-03-20 17:53:41 -08001320 return mtu;
1321}
1322
Andi Kleen67edfef2009-07-21 23:00:40 +00001323/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001324void tcp_mtup_init(struct sock *sk)
1325{
1326 struct tcp_sock *tp = tcp_sk(sk);
1327 struct inet_connection_sock *icsk = inet_csk(sk);
1328
1329 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1330 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001331 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -08001332 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1333 icsk->icsk_mtup.probe_size = 0;
1334}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001335EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001336
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337/* This function synchronize snd mss to current pmtu/exthdr set.
1338
1339 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1340 for TCP options, but includes only bare TCP header.
1341
1342 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001343 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 It also does not include TCP options.
1345
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001346 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 tp->mss_cache is current effective sending mss, including
1349 all tcp options except for SACKs. It is evaluated,
1350 taking into account current pmtu, but never exceeds
1351 tp->rx_opt.mss_clamp.
1352
1353 NOTE1. rfc1122 clearly states that advertised MSS
1354 DOES NOT include either tcp or ip options.
1355
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001356 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1357 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1360{
1361 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001362 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001363 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364
John Heffner5d424d52006-03-20 17:53:41 -08001365 if (icsk->icsk_mtup.search_high > pmtu)
1366 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
John Heffner5d424d52006-03-20 17:53:41 -08001368 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001369 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370
1371 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001372 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001373 if (icsk->icsk_mtup.enabled)
1374 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001375 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376
1377 return mss_now;
1378}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001379EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380
1381/* Compute the current effective MSS, taking SACKs and IP options,
1382 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001384unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001386 const struct tcp_sock *tp = tcp_sk(sk);
1387 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001388 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001389 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001390 struct tcp_out_options opts;
1391 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001393 mss_now = tp->mss_cache;
1394
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 if (dst) {
1396 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001397 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 mss_now = tcp_sync_mss(sk, mtu);
1399 }
1400
Adam Langley33ad7982008-07-19 00:04:31 -07001401 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1402 sizeof(struct tcphdr);
1403 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1404 * some common options. If this is an odd packet (because we have SACK
1405 * blocks etc) then our calculated header_len will be different, and
1406 * we have to adjust mss_now correspondingly */
1407 if (header_len != tp->tcp_header_len) {
1408 int delta = (int) header_len - tp->tcp_header_len;
1409 mss_now -= delta;
1410 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001411
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 return mss_now;
1413}
1414
Weiping Pan86fd14a2014-04-18 12:27:46 +08001415/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1416 * As additional protections, we do not touch cwnd in retransmission phases,
1417 * and if application hit its sndbuf limit recently.
1418 */
1419static void tcp_cwnd_application_limited(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001420{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001421 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001422
Weiping Pan86fd14a2014-04-18 12:27:46 +08001423 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1424 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1425 /* Limited by application or receiver window. */
1426 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1427 u32 win_used = max(tp->snd_cwnd_used, init_win);
1428 if (win_used < tp->snd_cwnd) {
1429 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1430 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1431 }
1432 tp->snd_cwnd_used = 0;
1433 }
1434 tp->snd_cwnd_stamp = tcp_time_stamp;
1435}
1436
Neal Cardwellca8a2262014-05-22 10:41:08 -04001437static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
David S. Millera762a982005-07-05 15:18:51 -07001438{
1439 struct tcp_sock *tp = tcp_sk(sk);
1440
Neal Cardwellca8a2262014-05-22 10:41:08 -04001441 /* Track the maximum number of outstanding packets in each
1442 * window, and remember whether we were cwnd-limited then.
1443 */
1444 if (!before(tp->snd_una, tp->max_packets_seq) ||
1445 tp->packets_out > tp->max_packets_out) {
1446 tp->max_packets_out = tp->packets_out;
1447 tp->max_packets_seq = tp->snd_nxt;
1448 tp->is_cwnd_limited = is_cwnd_limited;
1449 }
Eric Dumazete114a712014-04-30 11:58:13 -07001450
Eric Dumazet24901552014-05-02 21:18:05 -07001451 if (tcp_is_cwnd_limited(sk)) {
David S. Millera762a982005-07-05 15:18:51 -07001452 /* Network is feed fully. */
1453 tp->snd_cwnd_used = 0;
1454 tp->snd_cwnd_stamp = tcp_time_stamp;
1455 } else {
1456 /* Network starves. */
1457 if (tp->packets_out > tp->snd_cwnd_used)
1458 tp->snd_cwnd_used = tp->packets_out;
1459
David S. Miller15d33c02007-04-09 13:23:14 -07001460 if (sysctl_tcp_slow_start_after_idle &&
1461 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001462 tcp_cwnd_application_limited(sk);
1463 }
1464}
1465
Eric Dumazetd4589922013-12-13 13:51:23 -08001466/* Minshall's variant of the Nagle send check. */
1467static bool tcp_minshall_check(const struct tcp_sock *tp)
1468{
1469 return after(tp->snd_sml, tp->snd_una) &&
1470 !after(tp->snd_sml, tp->snd_nxt);
1471}
1472
1473/* Update snd_sml if this skb is under mss
1474 * Note that a TSO packet might end with a sub-mss segment
1475 * The test is really :
1476 * if ((skb->len % mss) != 0)
1477 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1478 * But we can avoid doing the divide again given we already have
1479 * skb_pcount = skb->len / mss_now
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001480 */
Eric Dumazetd4589922013-12-13 13:51:23 -08001481static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1482 const struct sk_buff *skb)
1483{
1484 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1485 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1486}
1487
1488/* Return false, if packet can be sent now without violation Nagle's rules:
1489 * 1. It is full sized. (provided by caller in %partial bool)
1490 * 2. Or it contains FIN. (already checked by caller)
1491 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1492 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1493 * With Minshall's modification: all sent small packets are ACKed.
1494 */
1495static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001496 int nonagle)
Eric Dumazetd4589922013-12-13 13:51:23 -08001497{
1498 return partial &&
1499 ((nonagle & TCP_NAGLE_CORK) ||
1500 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1501}
1502/* Returns the portion of skb which can be sent right away */
1503static unsigned int tcp_mss_split_point(const struct sock *sk,
1504 const struct sk_buff *skb,
1505 unsigned int mss_now,
1506 unsigned int max_segs,
1507 int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001508{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001509 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetd4589922013-12-13 13:51:23 -08001510 u32 partial, needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001511
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001512 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001513 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001514
Ben Hutchings14853482012-07-30 16:11:42 +00001515 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1516 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001517
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001518 needed = min(skb->len, window);
1519
Ben Hutchings14853482012-07-30 16:11:42 +00001520 if (max_len <= needed)
1521 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001522
Eric Dumazetd4589922013-12-13 13:51:23 -08001523 partial = needed % mss_now;
1524 /* If last segment is not a full MSS, check if Nagle rules allow us
1525 * to include this last segment in this skb.
1526 * Otherwise, we'll split the skb at last MSS boundary
1527 */
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001528 if (tcp_nagle_check(partial != 0, tp, nonagle))
Eric Dumazetd4589922013-12-13 13:51:23 -08001529 return needed - partial;
1530
1531 return needed;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001532}
1533
1534/* Can at least one segment of SKB be sent right now, according to the
1535 * congestion window rules? If so, return how many segments are allowed.
1536 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001537static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1538 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001539{
1540 u32 in_flight, cwnd;
1541
1542 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001543 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1544 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001545 return 1;
1546
1547 in_flight = tcp_packets_in_flight(tp);
1548 cwnd = tp->snd_cwnd;
1549 if (in_flight < cwnd)
1550 return (cwnd - in_flight);
1551
1552 return 0;
1553}
1554
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001555/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001556 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001557 * SKB onto the wire.
1558 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001559static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001560 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001561{
1562 int tso_segs = tcp_skb_pcount(skb);
1563
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001564 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001565 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001566 tso_segs = tcp_skb_pcount(skb);
1567 }
1568 return tso_segs;
1569}
1570
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001571
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001572/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001573 * sent now.
1574 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001575static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1576 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001577{
1578 /* Nagle rule does not apply to frames, which sit in the middle of the
1579 * write_queue (they have no chances to get new data).
1580 *
1581 * This is implemented in the callers, where they modify the 'nonagle'
1582 * argument based upon the location of SKB in the send queue.
1583 */
1584 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001585 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001586
Yuchung Cheng9b441902013-03-20 13:32:58 +00001587 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1588 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001589 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001590
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001591 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001592 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001593
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001594 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001595}
1596
1597/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001598static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1599 const struct sk_buff *skb,
1600 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001601{
1602 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1603
1604 if (skb->len > cur_mss)
1605 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1606
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001607 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001608}
1609
David S. Millerfe067e82007-03-07 12:12:44 -08001610/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001611 * should be put on the wire right now. If so, it returns the number of
1612 * packets allowed by the congestion window.
1613 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001614static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001615 unsigned int cur_mss, int nonagle)
1616{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001617 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001618 unsigned int cwnd_quota;
1619
David S. Miller846998a2005-08-04 19:52:01 -07001620 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001621
1622 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1623 return 0;
1624
1625 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001626 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001627 cwnd_quota = 0;
1628
1629 return cwnd_quota;
1630}
1631
Andi Kleen67edfef2009-07-21 23:00:40 +00001632/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001633bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001634{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001635 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001636 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001637
Eric Dumazeta02cec22010-09-22 20:43:57 +00001638 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001639 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001640 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001641 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001642}
1643
1644/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1645 * which is put after SKB on the list. It is very much like
1646 * tcp_fragment() except that it may make several kinds of assumptions
1647 * in order to speed up the splitting operation. In particular, we
1648 * know that all the data is in scatter-gather pages, and that the
1649 * packet has never been sent out before (and thus is not cloned).
1650 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001651static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001652 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001653{
1654 struct sk_buff *buff;
1655 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001656 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001657
1658 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001659 if (skb->len != skb->data_len)
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001660 return tcp_fragment(sk, skb, len, mss_now, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001661
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001662 buff = sk_stream_alloc_skb(sk, 0, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001663 if (unlikely(buff == NULL))
1664 return -ENOMEM;
1665
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001666 sk->sk_wmem_queued += buff->truesize;
1667 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001668 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001669 skb->truesize -= nlen;
1670
1671 /* Correct the sequence numbers. */
1672 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1673 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1674 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1675
1676 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001677 flags = TCP_SKB_CB(skb)->tcp_flags;
1678 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1679 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001680
1681 /* This packet was never sent out yet, so no SACK bits. */
1682 TCP_SKB_CB(buff)->sacked = 0;
1683
Patrick McHardy84fa7932006-08-29 16:44:56 -07001684 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001685 skb_split(skb, buff, len);
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001686 tcp_fragment_tstamp(skb, buff);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001687
1688 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001689 tcp_set_skb_tso_segs(sk, skb, mss_now);
1690 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001691
1692 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001693 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001694 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001695
1696 return 0;
1697}
1698
1699/* Try to defer sending, if possible, in order to minimize the amount
1700 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1701 *
1702 * This algorithm is from John Heffner.
1703 */
Neal Cardwellca8a2262014-05-22 10:41:08 -04001704static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1705 bool *is_cwnd_limited)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001706{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001707 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001708 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001709 u32 send_win, cong_win, limit, in_flight;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001710 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001711
Eric Dumazet4de075e2011-09-27 13:25:05 -04001712 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001713 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001714
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001715 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001716 goto send_now;
1717
1718 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001719 if (tp->tso_deferred &&
Ilpo Järvinena2acde02008-12-05 22:49:37 -08001720 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001721 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001722
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001723 in_flight = tcp_packets_in_flight(tp);
1724
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001725 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001726
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001727 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001728
1729 /* From in_flight test above, we know that cwnd > in_flight. */
1730 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1731
1732 limit = min(send_win, cong_win);
1733
David S. Millerba244fe2006-03-11 18:51:49 -08001734 /* If a full-sized TSO skb can be sent, do it. */
Ben Hutchings14853482012-07-30 16:11:42 +00001735 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
Eric Dumazet95bd09e2013-08-27 05:46:32 -07001736 tp->xmit_size_goal_segs * tp->mss_cache))
John Heffnerae8064a2006-10-18 20:36:48 -07001737 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001738
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001739 /* Middle in queue won't get any more data, full sendable already? */
1740 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1741 goto send_now;
1742
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001743 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1744 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001745 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1746
1747 /* If at least some fraction of a window is available,
1748 * just use it.
1749 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001750 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001751 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001752 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001753 } else {
1754 /* Different approach, try not to defer past a single
1755 * ACK. Receiver should ACK every other full sized
1756 * frame, so if we have space for more than 3 frames
1757 * then send now.
1758 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001759 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001760 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001761 }
1762
Eric Dumazetf4541d62013-03-21 17:36:09 +00001763 /* Ok, it looks like it is advisable to defer.
1764 * Do not rearm the timer if already set to not break TCP ACK clocking.
1765 */
1766 if (!tp->tso_deferred)
1767 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001768
Neal Cardwellca8a2262014-05-22 10:41:08 -04001769 if (cong_win < send_win && cong_win < skb->len)
1770 *is_cwnd_limited = true;
1771
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001772 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001773
1774send_now:
1775 tp->tso_deferred = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001776 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001777}
1778
John Heffner5d424d52006-03-20 17:53:41 -08001779/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001780 * MTU probe is regularly attempting to increase the path MTU by
1781 * deliberately sending larger packets. This discovers routing
1782 * changes resulting in larger path MTUs.
1783 *
John Heffner5d424d52006-03-20 17:53:41 -08001784 * Returns 0 if we should wait to probe (no cwnd available),
1785 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001786 * -1 otherwise
1787 */
John Heffner5d424d52006-03-20 17:53:41 -08001788static int tcp_mtu_probe(struct sock *sk)
1789{
1790 struct tcp_sock *tp = tcp_sk(sk);
1791 struct inet_connection_sock *icsk = inet_csk(sk);
1792 struct sk_buff *skb, *nskb, *next;
1793 int len;
1794 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001795 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001796 int copy;
1797 int mss_now;
1798
1799 /* Not currently probing/verifying,
1800 * not in recovery,
1801 * have enough cwnd, and
1802 * not SACKing (the variable headers throw things off) */
1803 if (!icsk->icsk_mtup.enabled ||
1804 icsk->icsk_mtup.probe_size ||
1805 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1806 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001807 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001808 return -1;
1809
1810 /* Very simple search strategy: just double the MSS. */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001811 mss_now = tcp_current_mss(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001812 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001813 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001814 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1815 /* TODO: set timer for probe_converge_event */
1816 return -1;
1817 }
1818
1819 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001820 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001821 return -1;
1822
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001823 if (tp->snd_wnd < size_needed)
1824 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001825 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001826 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001827
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001828 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1829 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1830 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001831 return -1;
1832 else
1833 return 0;
1834 }
1835
1836 /* We're allowed to probe. Build it now. */
1837 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1838 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001839 sk->sk_wmem_queued += nskb->truesize;
1840 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001841
David S. Millerfe067e82007-03-07 12:12:44 -08001842 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001843
1844 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1845 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001846 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001847 TCP_SKB_CB(nskb)->sacked = 0;
1848 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001849 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001850
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001851 tcp_insert_write_queue_before(nskb, skb, sk);
1852
John Heffner5d424d52006-03-20 17:53:41 -08001853 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001854 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001855 copy = min_t(int, skb->len, probe_size - len);
1856 if (nskb->ip_summed)
1857 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1858 else
1859 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001860 skb_put(nskb, copy),
1861 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001862
1863 if (skb->len <= copy) {
1864 /* We've eaten all the data from this skb.
1865 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001866 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001867 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001868 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001869 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04001870 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001871 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001872 if (!skb_shinfo(skb)->nr_frags) {
1873 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001874 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001875 skb->csum = csum_partial(skb->data,
1876 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001877 } else {
1878 __pskb_trim_head(skb, copy);
1879 tcp_set_skb_tso_segs(sk, skb, mss_now);
1880 }
1881 TCP_SKB_CB(skb)->seq += copy;
1882 }
1883
1884 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001885
1886 if (len >= probe_size)
1887 break;
John Heffner5d424d52006-03-20 17:53:41 -08001888 }
1889 tcp_init_tso_segs(sk, nskb, nskb->len);
1890
1891 /* We're ready to send. If this fails, the probe will
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001892 * be resegmented into mss-sized pieces by tcp_write_xmit().
1893 */
John Heffner5d424d52006-03-20 17:53:41 -08001894 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1895 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001896 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001897 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001898 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001899
1900 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001901 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1902 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001903
1904 return 1;
1905 }
1906
1907 return -1;
1908}
1909
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910/* This routine writes packets to the network. It advances the
1911 * send_head. This happens as incoming acks open up the remote
1912 * window for us.
1913 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001914 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1915 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1916 * account rare use of URG, this is not a big flaw.
1917 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001918 * Send at most one packet when push_one > 0. Temporarily ignore
1919 * cwnd limit to force at most one packet out when push_one == 2.
1920
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001921 * Returns true, if no segments are in flight and we have queued segments,
1922 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001924static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1925 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
1927 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001928 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001929 unsigned int tso_segs, sent_pkts;
1930 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001931 int result;
Neal Cardwellca8a2262014-05-22 10:41:08 -04001932 bool is_cwnd_limited = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933
David S. Miller92df7b52005-07-05 15:19:06 -07001934 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001935
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001936 if (!push_one) {
1937 /* Do MTU probing. */
1938 result = tcp_mtu_probe(sk);
1939 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001940 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001941 } else if (result > 0) {
1942 sent_pkts = 1;
1943 }
John Heffner5d424d52006-03-20 17:53:41 -08001944 }
1945
David S. Millerfe067e82007-03-07 12:12:44 -08001946 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001947 unsigned int limit;
1948
Herbert Xub68e9f82005-08-04 19:52:02 -07001949 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001950 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001951
Andrey Vagin9d186ca2014-08-13 16:03:10 +04001952 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001953 /* "skb_mstamp" is used as a start point for the retransmit timer */
1954 skb_mstamp_get(&skb->skb_mstamp);
Andrew Vaginec342322012-11-15 04:03:17 +00001955 goto repair; /* Skip network transmission */
Andrey Vagin9d186ca2014-08-13 16:03:10 +04001956 }
Andrew Vaginec342322012-11-15 04:03:17 +00001957
Herbert Xub68e9f82005-08-04 19:52:02 -07001958 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001959 if (!cwnd_quota) {
Neal Cardwellca8a2262014-05-22 10:41:08 -04001960 is_cwnd_limited = true;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001961 if (push_one == 2)
1962 /* Force out a loss probe pkt. */
1963 cwnd_quota = 1;
1964 else
1965 break;
1966 }
Herbert Xub68e9f82005-08-04 19:52:02 -07001967
1968 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1969 break;
1970
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001971 if (tso_segs == 1) {
1972 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1973 (tcp_skb_is_last(sk, skb) ?
1974 nonagle : TCP_NAGLE_PUSH))))
1975 break;
1976 } else {
Neal Cardwellca8a2262014-05-22 10:41:08 -04001977 if (!push_one &&
1978 tcp_tso_should_defer(sk, skb, &is_cwnd_limited))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001979 break;
1980 }
David S. Milleraa934662005-07-05 15:20:09 -07001981
Eric Dumazetc9eeec22013-09-27 03:28:54 -07001982 /* TCP Small Queues :
1983 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1984 * This allows for :
1985 * - better RTT estimation and ACK scheduling
1986 * - faster recovery
1987 * - high rates
Eric Dumazet98e09382013-11-13 06:32:54 -08001988 * Alas, some drivers / subsystems require a fair amount
1989 * of queued bytes to ensure line rate.
1990 * One example is wifi aggregation (802.11 AMPDU)
Eric Dumazet46d3cea2012-07-11 05:50:31 +00001991 */
Eric Dumazet98e09382013-11-13 06:32:54 -08001992 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1993 sk->sk_pacing_rate >> 10);
Eric Dumazetc9eeec22013-09-27 03:28:54 -07001994
1995 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
Eric Dumazet46d3cea2012-07-11 05:50:31 +00001996 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
John Ognessbf062002014-02-09 18:40:11 -08001997 /* It is possible TX completion already happened
1998 * before we set TSQ_THROTTLED, so we must
1999 * test again the condition.
John Ognessbf062002014-02-09 18:40:11 -08002000 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002001 smp_mb__after_atomic();
John Ognessbf062002014-02-09 18:40:11 -08002002 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2003 break;
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002004 }
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002005
Herbert Xuc8ac3772005-08-16 20:43:40 -07002006 limit = mss_now;
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08002007 if (tso_segs > 1 && !tcp_urg_mode(tp))
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08002008 limit = tcp_mss_split_point(sk, skb, mss_now,
Ben Hutchings14853482012-07-30 16:11:42 +00002009 min_t(unsigned int,
2010 cwnd_quota,
Eric Dumazetd4589922013-12-13 13:51:23 -08002011 sk->sk_gso_max_segs),
2012 nonagle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
Herbert Xuc8ac3772005-08-16 20:43:40 -07002014 if (skb->len > limit &&
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00002015 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
Herbert Xuc8ac3772005-08-16 20:43:40 -07002016 break;
2017
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002018 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07002019 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020
Andrew Vaginec342322012-11-15 04:03:17 +00002021repair:
David S. Miller92df7b52005-07-05 15:19:06 -07002022 /* Advance the send_head. This one is sent out.
2023 * This call will increment packets_out.
2024 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002025 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07002026
2027 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002028 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002029
2030 if (push_one)
2031 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 }
David S. Miller92df7b52005-07-05 15:19:06 -07002033
David S. Milleraa934662005-07-05 15:20:09 -07002034 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002035 if (tcp_in_cwnd_reduction(sk))
2036 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002037
2038 /* Send one loss probe per tail loss episode. */
2039 if (push_one != 2)
2040 tcp_schedule_loss_probe(sk);
Neal Cardwellca8a2262014-05-22 10:41:08 -04002041 tcp_cwnd_validate(sk, is_cwnd_limited);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002042 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07002043 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002044 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2045}
2046
2047bool tcp_schedule_loss_probe(struct sock *sk)
2048{
2049 struct inet_connection_sock *icsk = inet_csk(sk);
2050 struct tcp_sock *tp = tcp_sk(sk);
2051 u32 timeout, tlp_time_stamp, rto_time_stamp;
Eric Dumazet740b0f12014-02-26 14:02:48 -08002052 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002053
2054 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2055 return false;
2056 /* No consecutive loss probes. */
2057 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2058 tcp_rearm_rto(sk);
2059 return false;
2060 }
2061 /* Don't do any loss probe on a Fast Open connection before 3WHS
2062 * finishes.
2063 */
2064 if (sk->sk_state == TCP_SYN_RECV)
2065 return false;
2066
2067 /* TLP is only scheduled when next timer event is RTO. */
2068 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2069 return false;
2070
2071 /* Schedule a loss probe in 2*RTT for SACK capable connections
2072 * in Open state, that are either limited by cwnd or application.
2073 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002074 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002075 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2076 return false;
2077
2078 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2079 tcp_send_head(sk))
2080 return false;
2081
2082 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2083 * for delayed ack when there's one outstanding packet.
2084 */
2085 timeout = rtt << 1;
2086 if (tp->packets_out == 1)
2087 timeout = max_t(u32, timeout,
2088 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2089 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2090
2091 /* If RTO is shorter, just schedule TLP in its place. */
2092 tlp_time_stamp = tcp_time_stamp + timeout;
2093 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2094 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2095 s32 delta = rto_time_stamp - tcp_time_stamp;
2096 if (delta > 0)
2097 timeout = delta;
2098 }
2099
2100 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2101 TCP_RTO_MAX);
2102 return true;
2103}
2104
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002105/* Thanks to skb fast clones, we can detect if a prior transmit of
2106 * a packet is still in a qdisc or driver queue.
2107 * In this case, there is very little point doing a retransmit !
2108 * Note: This is called from BH context only.
2109 */
2110static bool skb_still_in_host_queue(const struct sock *sk,
2111 const struct sk_buff *skb)
2112{
2113 const struct sk_buff *fclone = skb + 1;
2114
2115 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
2116 fclone->fclone == SKB_FCLONE_CLONE)) {
2117 NET_INC_STATS_BH(sock_net(sk),
2118 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2119 return true;
2120 }
2121 return false;
2122}
2123
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002124/* When probe timeout (PTO) fires, send a new segment if one exists, else
2125 * retransmit the last segment.
2126 */
2127void tcp_send_loss_probe(struct sock *sk)
2128{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002129 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002130 struct sk_buff *skb;
2131 int pcount;
2132 int mss = tcp_current_mss(sk);
2133 int err = -1;
2134
2135 if (tcp_send_head(sk) != NULL) {
2136 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2137 goto rearm_timer;
2138 }
2139
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002140 /* At most one outstanding TLP retransmission. */
2141 if (tp->tlp_high_seq)
2142 goto rearm_timer;
2143
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002144 /* Retransmit last segment. */
2145 skb = tcp_write_queue_tail(sk);
2146 if (WARN_ON(!skb))
2147 goto rearm_timer;
2148
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002149 if (skb_still_in_host_queue(sk, skb))
2150 goto rearm_timer;
2151
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002152 pcount = tcp_skb_pcount(skb);
2153 if (WARN_ON(!pcount))
2154 goto rearm_timer;
2155
2156 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002157 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2158 GFP_ATOMIC)))
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002159 goto rearm_timer;
2160 skb = tcp_write_queue_tail(sk);
2161 }
2162
2163 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2164 goto rearm_timer;
2165
Per Hurtigbef19092014-06-12 17:08:32 +02002166 err = __tcp_retransmit_skb(sk, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002167
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002168 /* Record snd_nxt for loss detection. */
2169 if (likely(!err))
2170 tp->tlp_high_seq = tp->snd_nxt;
2171
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002172rearm_timer:
2173 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2174 inet_csk(sk)->icsk_rto,
2175 TCP_RTO_MAX);
2176
2177 if (likely(!err))
2178 NET_INC_STATS_BH(sock_net(sk),
2179 LINUX_MIB_TCPLOSSPROBES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180}
2181
David S. Millera762a982005-07-05 15:18:51 -07002182/* Push out any pending frames which were held back due to
2183 * TCP_CORK or attempt at coalescing tiny packets.
2184 * The socket must be locked by the caller.
2185 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002186void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2187 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002188{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002189 /* If we are closed, the bytes will have to remain here.
2190 * In time closedown will finish, we empty the write queue and
2191 * all will be happy.
2192 */
2193 if (unlikely(sk->sk_state == TCP_CLOSE))
2194 return;
2195
Mel Gorman99a1dec2012-07-31 16:44:14 -07002196 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2197 sk_gfp_atomic(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002198 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002199}
2200
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002201/* Send _single_ skb sitting at the send head. This function requires
2202 * true push pending frames to setup probe timer etc.
2203 */
2204void tcp_push_one(struct sock *sk, unsigned int mss_now)
2205{
David S. Millerfe067e82007-03-07 12:12:44 -08002206 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002207
2208 BUG_ON(!skb || skb->len < mss_now);
2209
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002210 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002211}
2212
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213/* This function returns the amount that we can raise the
2214 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002215 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * 1. The window can never be shrunk once it is offered (RFC 793)
2217 * 2. We limit memory per socket
2218 *
2219 * RFC 1122:
2220 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2221 * RECV.NEXT + RCV.WIN fixed until:
2222 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2223 *
2224 * i.e. don't raise the right edge of the window until you can raise
2225 * it at least MSS bytes.
2226 *
2227 * Unfortunately, the recommended algorithm breaks header prediction,
2228 * since header prediction assumes th->window stays fixed.
2229 *
2230 * Strictly speaking, keeping th->window fixed violates the receiver
2231 * side SWS prevention criteria. The problem is that under this rule
2232 * a stream of single byte packets will cause the right side of the
2233 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002234 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 * Of course, if the sender implements sender side SWS prevention
2236 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002237 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002239 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 * If the free space is less than the 1/4 of the maximum
2241 * space available and the free space is less than 1/2 mss,
2242 * then set the window to 0.
2243 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2244 * Otherwise, just prevent the window from shrinking
2245 * and from being larger than the largest representable value.
2246 *
2247 * This prevents incremental opening of the window in the regime
2248 * where TCP is limited by the speed of the reader side taking
2249 * data out of the TCP receive queue. It does nothing about
2250 * those cases where the window is constrained on the sender side
2251 * because the pipeline is full.
2252 *
2253 * BSD also seems to "accidentally" limit itself to windows that are a
2254 * multiple of MSS, at least until the free space gets quite small.
2255 * This would appear to be a side effect of the mbuf implementation.
2256 * Combining these two algorithms results in the observed behavior
2257 * of having a fixed window size at almost all times.
2258 *
2259 * Below we obtain similar behavior by forcing the offered window to
2260 * a multiple of the mss when it is feasible to do so.
2261 *
2262 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2263 * Regular options like TIMESTAMP are taken into account.
2264 */
2265u32 __tcp_select_window(struct sock *sk)
2266{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002267 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002269 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 * here. I don't know if the value based on our guesses
2271 * of peer's MSS is better for the performance. It's more correct
2272 * but may be worse for the performance because of rcv_mss
2273 * fluctuations. --SAW 1998/11/1
2274 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002275 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 int free_space = tcp_space(sk);
Florian Westphal86c1a042014-02-19 12:51:10 +01002277 int allowed_space = tcp_full_space(sk);
2278 int full_space = min_t(int, tp->window_clamp, allowed_space);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 int window;
2280
2281 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002282 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002284 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002285 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Glauber Costa180d8cd2011-12-11 21:47:02 +00002287 if (sk_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002288 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2289 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290
Florian Westphal86c1a042014-02-19 12:51:10 +01002291 /* free_space might become our new window, make sure we don't
2292 * increase it due to wscale.
2293 */
2294 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2295
2296 /* if free space is less than mss estimate, or is below 1/16th
2297 * of the maximum allowed, try to move to zero-window, else
2298 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2299 * new incoming data is dropped due to memory limits.
2300 * With large window, mss test triggers way too late in order
2301 * to announce zero window in time before rmem limit kicks in.
2302 */
2303 if (free_space < (allowed_space >> 4) || free_space < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 return 0;
2305 }
2306
2307 if (free_space > tp->rcv_ssthresh)
2308 free_space = tp->rcv_ssthresh;
2309
2310 /* Don't do rounding if we are using window scaling, since the
2311 * scaled window will not line up with the MSS boundary anyway.
2312 */
2313 window = tp->rcv_wnd;
2314 if (tp->rx_opt.rcv_wscale) {
2315 window = free_space;
2316
2317 /* Advertise enough space so that it won't get scaled away.
2318 * Import case: prevent zero window announcement if
2319 * 1<<rcv_wscale > mss.
2320 */
2321 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2322 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2323 << tp->rx_opt.rcv_wscale);
2324 } else {
2325 /* Get the largest window that is a nice multiple of mss.
2326 * Window clamp already applied above.
2327 * If our current window offering is within 1 mss of the
2328 * free space we just keep it. This prevents the divide
2329 * and multiply from happening most of the time.
2330 * We also don't do any window rounding when the free space
2331 * is too small.
2332 */
2333 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002334 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002335 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002336 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002337 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 }
2339
2340 return window;
2341}
2342
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002343/* Collapses two adjacent SKB's during retransmission. */
2344static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345{
2346 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002347 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002348 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002350 skb_size = skb->len;
2351 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002353 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002354
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002355 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002357 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002359 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2360 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002362 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2363 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002365 if (skb->ip_summed != CHECKSUM_PARTIAL)
2366 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002368 /* Update sequence range on original skb. */
2369 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002371 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002372 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002374 /* All done, get rid of second SKB and account for it so
2375 * packet counting does not break.
2376 */
2377 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002378
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002379 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002380 tcp_clear_retrans_hints_partial(tp);
2381 if (next_skb == tp->retransmit_skb_hint)
2382 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002383
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002384 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2385
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002386 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387}
2388
Andi Kleen67edfef2009-07-21 23:00:40 +00002389/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002390static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002391{
2392 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002393 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002394 /* TODO: SACK collapsing could be used to remove this condition */
2395 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002396 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002397 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002398 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002399 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002400 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002401 /* Some heurestics for collapsing over SACK'd could be invented */
2402 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002403 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002404
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002405 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002406}
2407
Andi Kleen67edfef2009-07-21 23:00:40 +00002408/* Collapse packets in the retransmit queue to make to create
2409 * less packets on the wire. This is only done on retransmission.
2410 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002411static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2412 int space)
2413{
2414 struct tcp_sock *tp = tcp_sk(sk);
2415 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002416 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002417
2418 if (!sysctl_tcp_retrans_collapse)
2419 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002420 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002421 return;
2422
2423 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2424 if (!tcp_can_collapse(sk, skb))
2425 break;
2426
2427 space -= skb->len;
2428
2429 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002430 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002431 continue;
2432 }
2433
2434 if (space < 0)
2435 break;
2436 /* Punt if not enough space exists in the first SKB for
2437 * the data in the second
2438 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002439 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002440 break;
2441
2442 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2443 break;
2444
2445 tcp_collapse_retrans(sk, to);
2446 }
2447}
2448
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449/* This retransmits one SKB. Policy decisions and retransmit queue
2450 * state updates are done by the caller. Returns non-zero if an
2451 * error occurred which prevented the send.
2452 */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002453int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454{
2455 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002456 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002457 unsigned int cur_mss;
Yuchung Chengc84a5712014-02-28 16:42:26 -08002458 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459
John Heffner5d424d52006-03-20 17:53:41 -08002460 /* Inconslusive MTU probe */
2461 if (icsk->icsk_mtup.probe_size) {
2462 icsk->icsk_mtup.probe_size = 0;
2463 }
2464
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002466 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 */
2468 if (atomic_read(&sk->sk_wmem_alloc) >
2469 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2470 return -EAGAIN;
2471
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002472 if (skb_still_in_host_queue(sk, skb))
2473 return -EBUSY;
2474
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2476 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2477 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2479 return -ENOMEM;
2480 }
2481
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002482 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2483 return -EHOSTUNREACH; /* Routing failure or similar. */
2484
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002485 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002486
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 /* If receiver has shrunk his window, and skb is out of
2488 * new window, do not retransmit it. The exception is the
2489 * case, when window is shrunk to zero. In this case
2490 * our retransmit serves as a zero window probe.
2491 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002492 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2493 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494 return -EAGAIN;
2495
2496 if (skb->len > cur_mss) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002497 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002499 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002500 int oldpcount = tcp_skb_pcount(skb);
2501
2502 if (unlikely(oldpcount > 1)) {
Eric Dumazetc52e24212013-10-15 11:54:30 -07002503 if (skb_unclone(skb, GFP_ATOMIC))
2504 return -ENOMEM;
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002505 tcp_init_tso_segs(sk, skb, cur_mss);
2506 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2507 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002508 }
2509
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002510 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 /* Make a copy, if the first transmission SKB clone we made
2513 * is still in somebody's hands, else make a clone.
2514 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
Thomas Graf50bceae2013-04-11 10:57:18 +00002516 /* make sure skb->data is aligned on arches that require it
2517 * and check if ack-trimming & collapsing extended the headroom
2518 * beyond what csum_start can cover.
2519 */
2520 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2521 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet117632e2011-12-03 21:39:53 +00002522 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2523 GFP_ATOMIC);
Yuchung Chengc84a5712014-02-28 16:42:26 -08002524 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2525 -ENOBUFS;
Eric Dumazet117632e2011-12-03 21:39:53 +00002526 } else {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002527 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002528 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002529
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002530 if (likely(!err)) {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002531 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002532 /* Update global TCP statistics. */
2533 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2534 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2535 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2536 tp->total_retrans++;
2537 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002538 return err;
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002539}
2540
2541int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2542{
2543 struct tcp_sock *tp = tcp_sk(sk);
2544 int err = __tcp_retransmit_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545
2546 if (err == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002548 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002549 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 }
2551#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002552 if (!tp->retrans_out)
2553 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2555 tp->retrans_out += tcp_skb_pcount(skb);
2556
2557 /* Save stamp of the first retransmit. */
2558 if (!tp->retrans_stamp)
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002559 tp->retrans_stamp = tcp_skb_timestamp(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 /* snd_nxt is stored to detect loss of retransmitted segment,
2562 * see tcp_input.c tcp_sacktag_write_queue().
2563 */
2564 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002565 } else if (err != -EBUSY) {
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002566 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 }
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002568
2569 if (tp->undo_retrans < 0)
2570 tp->undo_retrans = 0;
2571 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 return err;
2573}
2574
Andi Kleen67edfef2009-07-21 23:00:40 +00002575/* Check if we forward retransmits are possible in the current
2576 * window/congestion state.
2577 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002578static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002579{
2580 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002581 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002582
2583 /* Forward retransmissions are possible only during Recovery. */
2584 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002585 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002586
2587 /* No forward retransmissions in Reno are possible. */
2588 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002589 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002590
2591 /* Yeah, we have to make difficult choice between forward transmission
2592 * and retransmission... Both ways have their merits...
2593 *
2594 * For now we do not retransmit anything, while we have some new
2595 * segments to send. In the other cases, follow rule 3 for
2596 * NextSeg() specified in RFC3517.
2597 */
2598
2599 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002600 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002601
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002602 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002603}
2604
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605/* This gets called after a retransmit timeout, and the initially
2606 * retransmitted data is acknowledged. It tries to continue
2607 * resending the rest of the retransmit queue, until either
2608 * we've sent it all or the congestion window limit is reached.
2609 * If doing SACK, the first ACK which comes back for a timeout
2610 * based retransmit packet might feed us FACK information again.
2611 * If so, we use it to avoid unnecessarily retransmissions.
2612 */
2613void tcp_xmit_retransmit_queue(struct sock *sk)
2614{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002615 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 struct tcp_sock *tp = tcp_sk(sk);
2617 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002618 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002619 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002620 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002621 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002622
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002623 if (!tp->packets_out)
2624 return;
2625
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002626 if (!tp->lost_out)
2627 tp->retransmit_high = tp->snd_una;
2628
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002629 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002630 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002631 last_lost = TCP_SKB_CB(skb)->end_seq;
2632 if (after(last_lost, tp->retransmit_high))
2633 last_lost = tp->retransmit_high;
2634 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002635 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002636 last_lost = tp->snd_una;
2637 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002639 tcp_for_write_queue_from(skb, sk) {
2640 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002642 if (skb == tcp_send_head(sk))
2643 break;
2644 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002645 if (hole == NULL)
2646 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002647
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002648 /* Assume this retransmit will generate
2649 * only one packet for congestion window
2650 * calculation purposes. This works because
2651 * tcp_retransmit_skb() will chop up the
2652 * packet to be MSS sized and all the
2653 * packet counting works out.
2654 */
2655 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2656 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002657
2658 if (fwd_rexmitting) {
2659begin_fwd:
2660 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2661 break;
2662 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2663
2664 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002665 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002666 if (!tcp_can_forward_retransmit(sk))
2667 break;
2668 /* Backtrack if necessary to non-L'ed skb */
2669 if (hole != NULL) {
2670 skb = hole;
2671 hole = NULL;
2672 }
2673 fwd_rexmitting = 1;
2674 goto begin_fwd;
2675
2676 } else if (!(sacked & TCPCB_LOST)) {
Ilpo Järvinenac11ba72009-02-28 04:44:27 +00002677 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002678 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002679 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002681 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002682 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002683 if (icsk->icsk_ca_state != TCP_CA_Loss)
2684 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2685 else
2686 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2687 }
2688
2689 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002690 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002691
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002692 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002693 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002694
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002695 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2696
Yuchung Cheng684bad12012-09-02 17:38:04 +00002697 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002698 tp->prr_out += tcp_skb_pcount(skb);
2699
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002700 if (skb == tcp_write_queue_head(sk))
2701 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2702 inet_csk(sk)->icsk_rto,
2703 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705}
2706
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707/* Send a fin. The caller locks the socket for us. This cannot be
2708 * allowed to fail queueing a FIN frame under any circumstances.
2709 */
2710void tcp_send_fin(struct sock *sk)
2711{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002712 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002713 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002715
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 /* Optimization, tack on the FIN if we have a queue of
2717 * unsent frames. But be careful about outgoing SACKS
2718 * and IP options.
2719 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002720 mss_now = tcp_current_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
David S. Millerfe067e82007-03-07 12:12:44 -08002722 if (tcp_send_head(sk) != NULL) {
Eric Dumazet4de075e2011-09-27 13:25:05 -04002723 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 TCP_SKB_CB(skb)->end_seq++;
2725 tp->write_seq++;
2726 } else {
2727 /* Socket is locked, keep trying until memory is available. */
2728 for (;;) {
Wu Fengguangaa133072009-09-02 23:45:45 -07002729 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2730 sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 if (skb)
2732 break;
2733 yield();
2734 }
2735
2736 /* Reserve space for headers and prepare control bits. */
2737 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002739 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002740 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741 tcp_queue_skb(sk, skb);
2742 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002743 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744}
2745
2746/* We get here when a process closes a file descriptor (either due to
2747 * an explicit close() or as a byproduct of exit()'ing) and there
2748 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002749 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 */
Al Virodd0fc662005-10-07 07:46:04 +01002751void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 struct sk_buff *skb;
2754
2755 /* NOTE: No TCP options attached and we never retransmit this. */
2756 skb = alloc_skb(MAX_TCP_HEADER, priority);
2757 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002758 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 return;
2760 }
2761
2762 /* Reserve space for headers and prepare control bits. */
2763 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002764 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002765 TCPHDR_ACK | TCPHDR_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 /* Send it off. */
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002767 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002768 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002769
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002770 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771}
2772
Andi Kleen67edfef2009-07-21 23:00:40 +00002773/* Send a crossed SYN-ACK during socket establishment.
2774 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 * a SYN packet that crossed the incoming SYN that caused this routine
2776 * to get called. If this assumption fails then the initial rcv_wnd
2777 * and rcv_wscale values will not be correct.
2778 */
2779int tcp_send_synack(struct sock *sk)
2780{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002781 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782
David S. Millerfe067e82007-03-07 12:12:44 -08002783 skb = tcp_write_queue_head(sk);
Eric Dumazet4de075e2011-09-27 13:25:05 -04002784 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00002785 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 return -EFAULT;
2787 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04002788 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 if (skb_cloned(skb)) {
2790 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2791 if (nskb == NULL)
2792 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002793 tcp_unlink_write_queue(skb, sk);
Eric Dumazetf4a775d2014-09-22 16:29:32 -07002794 __skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002795 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002796 sk_wmem_free_skb(sk, skb);
2797 sk->sk_wmem_queued += nskb->truesize;
2798 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 skb = nskb;
2800 }
2801
Eric Dumazet4de075e2011-09-27 13:25:05 -04002802 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Daniel Borkmann30e502a2014-09-26 22:37:33 +02002803 TCP_ECN_send_synack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002805 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806}
2807
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002808/**
2809 * tcp_make_synack - Prepare a SYN-ACK.
2810 * sk: listener socket
2811 * dst: dst entry attached to the SYNACK
2812 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002813 *
2814 * Allocate one skb and build a SYNACK packet.
2815 * @dst is consumed : Caller should not use it again.
2816 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002817struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002818 struct request_sock *req,
Jerry Chu83368862012-08-31 12:29:12 +00002819 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002821 struct tcp_out_options opts;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002822 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 struct tcp_sock *tp = tcp_sk(sk);
2824 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002826 struct tcp_md5sig_key *md5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002827 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002828 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829
Eric Dumazeta0b84862014-03-26 09:57:19 -07002830 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002831 if (unlikely(!skb)) {
2832 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002834 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 /* Reserve space for headers. */
2836 skb_reserve(skb, MAX_TCP_HEADER);
2837
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002838 skb_dst_set(skb, dst);
Eric Dumazetca10b9e2013-04-08 17:58:11 +00002839 security_skb_owned_by(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840
David S. Miller0dbaee32010-12-13 12:52:14 -08002841 mss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002842 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2843 mss = tp->rx_opt.user_mss;
2844
Adam Langley33ad7982008-07-19 00:04:31 -07002845 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002846#ifdef CONFIG_SYN_COOKIES
2847 if (unlikely(req->cookie_ts))
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002848 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002849 else
2850#endif
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002851 skb_mstamp_get(&skb->skb_mstamp);
Christoph Paasch1a2c6182013-03-17 08:23:34 +00002852 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2853 foc) + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002854
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002855 skb_push(skb, tcp_header_size);
2856 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002858 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 memset(th, 0, sizeof(struct tcphdr));
2860 th->syn = 1;
2861 th->ack = 1;
Daniel Borkmann30e502a2014-09-26 22:37:33 +02002862 TCP_ECN_make_synack(req, th, sk);
Eric Dumazetb44084c2013-10-10 00:04:37 -07002863 th->source = htons(ireq->ir_num);
Eric Dumazet634fb9792013-10-09 15:21:29 -07002864 th->dest = ireq->ir_rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002865 /* Setting of flags are superfluous here for callers (and ECE is
2866 * not even correctly set)
2867 */
2868 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00002869 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faade2009-12-02 18:25:27 +00002870
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Jerry Chu83368862012-08-31 12:29:12 +00002872 /* XXX data is queued and acked as is. No buffer/window check */
2873 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874
2875 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002876 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002877 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 th->doff = (tcp_header_size >> 2);
Eric Dumazeta0b84862014-03-26 09:57:19 -07002879 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002880
2881#ifdef CONFIG_TCP_MD5SIG
2882 /* Okay, we have all we need - do the md5 hash if needed */
2883 if (md5) {
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002884 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002885 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002886 }
2887#endif
2888
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 return skb;
2890}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002891EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892
Andi Kleen67edfef2009-07-21 23:00:40 +00002893/* Do all connect socket setups that can be done AF independent. */
stephen hemmingerf7e56a72013-12-29 11:39:51 -08002894static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002896 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 struct tcp_sock *tp = tcp_sk(sk);
2898 __u8 rcv_wscale;
2899
2900 /* We'll fix this up when we get a response from the other end.
2901 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2902 */
2903 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08002904 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002906#ifdef CONFIG_TCP_MD5SIG
2907 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2908 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2909#endif
2910
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911 /* If user gave his TCP_MAXSEG, record it to clamp */
2912 if (tp->rx_opt.user_mss)
2913 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2914 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002915 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 tcp_sync_mss(sk, dst_mtu(dst));
2917
2918 if (!tp->window_clamp)
2919 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08002920 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002921 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2922 tp->advmss = tp->rx_opt.user_mss;
2923
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00002926 /* limit the window selection if the user enforce a smaller rx buffer */
2927 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2928 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2929 tp->window_clamp = tcp_full_space(sk);
2930
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 tcp_select_initial_window(tcp_full_space(sk),
2932 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2933 &tp->rcv_wnd,
2934 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08002935 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00002936 &rcv_wscale,
2937 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938
2939 tp->rx_opt.rcv_wscale = rcv_wscale;
2940 tp->rcv_ssthresh = tp->rcv_wnd;
2941
2942 sk->sk_err = 0;
2943 sock_reset_flag(sk, SOCK_DONE);
2944 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08002945 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 tp->snd_una = tp->write_seq;
2947 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002948 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00002949 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00002950
2951 if (likely(!tp->repair))
2952 tp->rcv_nxt = 0;
Andrew Vaginc7781a62013-08-27 12:20:40 +04002953 else
2954 tp->rcv_tstamp = tcp_time_stamp;
Pavel Emelyanovee995282012-04-19 03:40:39 +00002955 tp->rcv_wup = tp->rcv_nxt;
2956 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002958 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2959 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960 tcp_clear_retrans(tp);
2961}
2962
Yuchung Cheng783237e2012-07-19 06:43:07 +00002963static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2964{
2965 struct tcp_sock *tp = tcp_sk(sk);
2966 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2967
2968 tcb->end_seq += skb->len;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07002969 __skb_header_release(skb);
Yuchung Cheng783237e2012-07-19 06:43:07 +00002970 __tcp_add_write_queue_tail(sk, skb);
2971 sk->sk_wmem_queued += skb->truesize;
2972 sk_mem_charge(sk, skb->truesize);
2973 tp->write_seq = tcb->end_seq;
2974 tp->packets_out += tcp_skb_pcount(skb);
2975}
2976
2977/* Build and send a SYN with data and (cached) Fast Open cookie. However,
2978 * queue a data-only packet after the regular SYN, such that regular SYNs
2979 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2980 * only the SYN sequence, the data are retransmitted in the first ACK.
2981 * If cookie is not cached or other error occurs, falls back to send a
2982 * regular SYN with Fast Open cookie request option.
2983 */
2984static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2985{
2986 struct tcp_sock *tp = tcp_sk(sk);
2987 struct tcp_fastopen_request *fo = tp->fastopen_req;
Yuchung Chengaab48742012-07-19 06:43:10 +00002988 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
Yuchung Cheng783237e2012-07-19 06:43:07 +00002989 struct sk_buff *syn_data = NULL, *data;
Yuchung Chengaab48742012-07-19 06:43:10 +00002990 unsigned long last_syn_loss = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00002991
Yuchung Cheng67da22d2012-07-19 06:43:11 +00002992 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00002993 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2994 &syn_loss, &last_syn_loss);
2995 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2996 if (syn_loss > 1 &&
2997 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2998 fo->cookie.len = -1;
2999 goto fallback;
3000 }
3001
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003002 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3003 fo->cookie.len = -1;
3004 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003005 goto fallback;
3006
3007 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3008 * user-MSS. Reserve maximum option space for middleboxes that add
3009 * private TCP options. The cost is reduced data space in SYN :(
3010 */
3011 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3012 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00003013 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00003014 MAX_TCP_OPTION_SPACE;
3015
Eric Dumazetf5ddcbb2014-02-20 10:09:18 -08003016 space = min_t(size_t, space, fo->size);
3017
3018 /* limit to order-0 allocations */
3019 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3020
3021 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
Yuchung Cheng783237e2012-07-19 06:43:07 +00003022 sk->sk_allocation);
3023 if (syn_data == NULL)
3024 goto fallback;
3025
3026 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3027 struct iovec *iov = &fo->data->msg_iov[i];
3028 unsigned char __user *from = iov->iov_base;
3029 int len = iov->iov_len;
3030
3031 if (syn_data->len + len > space)
3032 len = space - syn_data->len;
3033 else if (i + 1 == iovlen)
3034 /* No more data pending in inet_wait_for_connect() */
3035 fo->data = NULL;
3036
3037 if (skb_add_data(syn_data, from, len))
3038 goto fallback;
3039 }
3040
3041 /* Queue a data-only packet after the regular SYN for retransmission */
3042 data = pskb_copy(syn_data, sk->sk_allocation);
3043 if (data == NULL)
3044 goto fallback;
3045 TCP_SKB_CB(data)->seq++;
3046 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3047 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3048 tcp_connect_queue_skb(sk, data);
3049 fo->copied = data->len;
3050
Eric Dumazet431a9122014-03-09 17:36:02 -07003051 /* syn_data is about to be sent, we need to take current time stamps
3052 * for the packets that are in write queue : SYN packet and DATA
3053 */
3054 skb_mstamp_get(&syn->skb_mstamp);
3055 data->skb_mstamp = syn->skb_mstamp;
3056
Yuchung Cheng783237e2012-07-19 06:43:07 +00003057 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003058 tp->syn_data = (fo->copied > 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08003059 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003060 goto done;
3061 }
3062 syn_data = NULL;
3063
3064fallback:
3065 /* Send a regular SYN with Fast Open cookie request option */
3066 if (fo->cookie.len > 0)
3067 fo->cookie.len = 0;
3068 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3069 if (err)
3070 tp->syn_fastopen = 0;
3071 kfree_skb(syn_data);
3072done:
3073 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3074 return err;
3075}
3076
Andi Kleen67edfef2009-07-21 23:00:40 +00003077/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078int tcp_connect(struct sock *sk)
3079{
3080 struct tcp_sock *tp = tcp_sk(sk);
3081 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00003082 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083
3084 tcp_connect_init(sk);
3085
Andrey Vagin2b916472012-11-22 01:13:58 +00003086 if (unlikely(tp->repair)) {
3087 tcp_finish_connect(sk, NULL);
3088 return 0;
3089 }
3090
David S. Millerd179cd12005-08-17 14:57:30 -07003091 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 if (unlikely(buff == NULL))
3093 return -ENOBUFS;
3094
3095 /* Reserve space for headers. */
3096 skb_reserve(buff, MAX_TCP_HEADER);
3097
Changli Gaoa3433f32010-06-12 14:01:43 +00003098 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003099 tp->retrans_stamp = tcp_time_stamp;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003100 tcp_connect_queue_skb(sk, buff);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08003101 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102
Yuchung Cheng783237e2012-07-19 06:43:07 +00003103 /* Send off SYN; include data in Fast Open. */
3104 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3105 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00003106 if (err == -ECONNREFUSED)
3107 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07003108
3109 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3110 * in order to make this packet get counted in tcpOutSegs.
3111 */
3112 tp->snd_nxt = tp->write_seq;
3113 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003114 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115
3116 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003117 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3118 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 return 0;
3120}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003121EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122
3123/* Send out a delayed ack, the caller does the policy checking
3124 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3125 * for details.
3126 */
3127void tcp_send_delayed_ack(struct sock *sk)
3128{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003129 struct inet_connection_sock *icsk = inet_csk(sk);
3130 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 unsigned long timeout;
3132
Florian Westphal98900922014-09-26 22:37:35 +02003133 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3134
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003136 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003137 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003139 if (icsk->icsk_ack.pingpong ||
3140 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 max_ato = TCP_DELACK_MAX;
3142
3143 /* Slow path, intersegment interval is "high". */
3144
3145 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003146 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 * directly.
3148 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003149 if (tp->srtt_us) {
3150 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3151 TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152
3153 if (rtt < max_ato)
3154 max_ato = rtt;
3155 }
3156
3157 ato = min(ato, max_ato);
3158 }
3159
3160 /* Stay within the limit we were given */
3161 timeout = jiffies + ato;
3162
3163 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003164 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 /* If delack timer was blocked or is about to expire,
3166 * send ACK now.
3167 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003168 if (icsk->icsk_ack.blocked ||
3169 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 tcp_send_ack(sk);
3171 return;
3172 }
3173
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003174 if (!time_before(timeout, icsk->icsk_ack.timeout))
3175 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003177 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3178 icsk->icsk_ack.timeout = timeout;
3179 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180}
3181
3182/* This routine sends an ack and also updates the window. */
3183void tcp_send_ack(struct sock *sk)
3184{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003185 struct sk_buff *buff;
3186
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003188 if (sk->sk_state == TCP_CLOSE)
3189 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190
Florian Westphal98900922014-09-26 22:37:35 +02003191 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3192
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003193 /* We are not putting this on the write queue, so
3194 * tcp_transmit_skb() will set the ownership to this
3195 * sock.
3196 */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003197 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003198 if (buff == NULL) {
3199 inet_csk_schedule_ack(sk);
3200 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3201 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3202 TCP_DELACK_MAX, TCP_RTO_MAX);
3203 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003205
3206 /* Reserve space for headers and prepare control bits. */
3207 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003208 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003209
3210 /* Send it off, this clears delayed acks for us. */
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003211 skb_mstamp_get(&buff->skb_mstamp);
Mel Gorman99a1dec2012-07-31 16:44:14 -07003212 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003213}
Daniel Borkmanne3118e82014-09-26 22:37:36 +02003214EXPORT_SYMBOL_GPL(tcp_send_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
3216/* This routine sends a packet with an out of date sequence
3217 * number. It assumes the other end will try to ack it.
3218 *
3219 * Question: what should we make while urgent mode?
3220 * 4.4BSD forces sending single byte of data. We cannot send
3221 * out of window data, because we have SND.NXT==SND.MAX...
3222 *
3223 * Current solution: to send TWO zero-length segments in urgent mode:
3224 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3225 * out-of-date with SND.UNA-1 to probe window.
3226 */
3227static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3228{
3229 struct tcp_sock *tp = tcp_sk(sk);
3230 struct sk_buff *skb;
3231
3232 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003233 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003234 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 return -1;
3236
3237 /* Reserve space for headers and set control bits. */
3238 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 /* Use a previous sequence. This should cause the other
3240 * end to send an ack. Don't queue or clone SKB, just
3241 * send it.
3242 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003243 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003244 skb_mstamp_get(&skb->skb_mstamp);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003245 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246}
3247
Pavel Emelyanovee995282012-04-19 03:40:39 +00003248void tcp_send_window_probe(struct sock *sk)
3249{
3250 if (sk->sk_state == TCP_ESTABLISHED) {
3251 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3252 tcp_xmit_probe_skb(sk, 0);
3253 }
3254}
3255
Andi Kleen67edfef2009-07-21 23:00:40 +00003256/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257int tcp_write_wakeup(struct sock *sk)
3258{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003259 struct tcp_sock *tp = tcp_sk(sk);
3260 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003262 if (sk->sk_state == TCP_CLOSE)
3263 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003265 if ((skb = tcp_send_head(sk)) != NULL &&
3266 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3267 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003268 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003269 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003271 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3272 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003274 /* We are probing the opening of a window
3275 * but the window size is != 0
3276 * must have been a result SWS avoidance ( sender )
3277 */
3278 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3279 skb->len > mss) {
3280 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003281 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Octavian Purdila6cc55e02014-06-06 17:32:37 +03003282 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003283 return -1;
3284 } else if (!tcp_skb_pcount(skb))
3285 tcp_set_skb_tso_segs(sk, skb, mss);
3286
Eric Dumazet4de075e2011-09-27 13:25:05 -04003287 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003288 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3289 if (!err)
3290 tcp_event_new_data_sent(sk, skb);
3291 return err;
3292 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003293 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003294 tcp_xmit_probe_skb(sk, 1);
3295 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297}
3298
3299/* A window probe timeout has occurred. If window is not closed send
3300 * a partial packet else a zero probe.
3301 */
3302void tcp_send_probe0(struct sock *sk)
3303{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003304 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003306 unsigned long probe_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003307 int err;
3308
3309 err = tcp_write_wakeup(sk);
3310
David S. Millerfe067e82007-03-07 12:12:44 -08003311 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003313 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003314 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 return;
3316 }
3317
3318 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003319 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3320 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003321 icsk->icsk_probes_out++;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003322 probe_max = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 } else {
3324 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003325 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 * Let local senders to fight for local resources.
3327 *
3328 * Use accumulated backoff yet.
3329 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003330 if (!icsk->icsk_probes_out)
3331 icsk->icsk_probes_out = 1;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003332 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 }
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003334 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3335 inet_csk_rto_backoff(icsk, probe_max),
3336 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337}
Octavian Purdila5db92c92014-06-25 17:09:59 +03003338
3339int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3340{
3341 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3342 struct flowi fl;
3343 int res;
3344
3345 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3346 if (!res) {
3347 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3348 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3349 }
3350 return res;
3351}
3352EXPORT_SYMBOL(tcp_rtx_synack);