blob: 78fc89c1c43c7e16edb6b77bcc117518540b387a [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
Wei Liuc39c4c62015-06-03 11:10:42 +010053/* Default TSQ limit of four TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
Eric Dumazet46d3cea2012-07-11 05:50:31 +000055
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
David S. Miller35089bb2006-06-13 22:33:04 -070062/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070063int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070064
Eric Dumazetc9bee3b2013-07-22 20:27:07 -070065unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
67
Eric Dumazet46d3cea2012-07-11 05:50:31 +000068static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000070
Andi Kleen67edfef2009-07-21 23:00:40 +000071/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040072static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070073{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000074 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070075 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080076 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070077
David S. Millerfe067e82007-03-07 12:12:44 -080078 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080080
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080081 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000082 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000083 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000084 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000085 }
Yuchung Chengf19c29e2014-03-03 12:31:36 -080086
David S. Millerf7324ac2014-03-06 15:03:17 -050087 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -070089}
90
91/* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
96 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040097static inline __u32 tcp_acceptable_seq(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070098{
Eric Dumazetcf533ea2011-10-21 05:22:42 -040099 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700100
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102 return tp->snd_nxt;
103 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800104 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400124 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 int mss = tp->advmss;
126
David S. Miller0dbaee32010-12-13 12:52:14 -0800127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400141static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700143 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
147
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 restart_cwnd = min(restart_cwnd, cwnd);
152
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 cwnd >>= 1;
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
158}
159
Andi Kleen67edfef2009-07-21 23:00:40 +0000160/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800161static void tcp_event_data_sent(struct tcp_sock *tp,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400162 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
David S. Miller35089bb2006-06-13 22:33:04 -0700167 if (sysctl_tcp_slow_start_after_idle &&
168 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700169 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170
171 tp->lsndtime = now;
172
173 /* If it is a reply for ato after last received
174 * packet, enter pingpong mode.
175 */
Jon Maxwell2251ae42015-07-08 10:12:28 +1000176 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
177 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
179
Andi Kleen67edfef2009-07-21 23:00:40 +0000180/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800181static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700183 tcp_dec_quickack_mode(sk, pkts);
184 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185}
186
Yuchung Cheng85f16522013-06-11 15:35:32 -0700187
188u32 tcp_default_init_rwnd(u32 mss)
189{
190 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800191 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700192 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
193 * limit when mss is larger than 1460.
194 */
195 u32 init_rwnd = TCP_INIT_CWND * 2;
196
197 if (mss > 1460)
198 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
199 return init_rwnd;
200}
201
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202/* Determine a window scaling and initial window to offer.
203 * Based on the assumption that the given amount of space
204 * will be offered. Store the results in the tp structure.
205 * NOTE: for smooth operation initial space offering should
206 * be a multiple of mss if possible. We assume here that mss >= 1.
207 * This MUST be enforced by all callers.
208 */
209void tcp_select_initial_window(int __space, __u32 mss,
210 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000211 int wscale_ok, __u8 *rcv_wscale,
212 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213{
214 unsigned int space = (__space < 0 ? 0 : __space);
215
216 /* If no clamp set the clamp to the max possible scaled window */
217 if (*window_clamp == 0)
218 (*window_clamp) = (65535 << 14);
219 space = min(*window_clamp, space);
220
221 /* Quantize space offering to a multiple of mss if possible. */
222 if (space > mss)
223 space = (space / mss) * mss;
224
225 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800226 * will break some buggy TCP stacks. If the admin tells us
227 * it is likely we could be speaking with such a buggy stack
228 * we will truncate our initial window offering to 32K-1
229 * unless the remote has sent us a window scaling option,
230 * which we interpret as a sign the remote TCP is not
231 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 */
Rick Jones15d99e02006-03-20 22:40:29 -0800233 if (sysctl_tcp_workaround_signed_windows)
234 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 else
236 (*rcv_wnd) = space;
237
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 (*rcv_wscale) = 0;
239 if (wscale_ok) {
240 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900241 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242 */
243 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700244 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 while (space > 65535 && (*rcv_wscale) < 14) {
246 space >>= 1;
247 (*rcv_wscale)++;
248 }
249 }
250
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800251 if (mss > (1 << *rcv_wscale)) {
Yuchung Cheng85f16522013-06-11 15:35:32 -0700252 if (!init_rcv_wnd) /* Use default unless specified otherwise */
253 init_rcv_wnd = tcp_default_init_rwnd(mss);
254 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 }
256
257 /* Set the clamp no higher than max representable value */
258 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
259}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000260EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261
262/* Chose a new window to advertise, update state in tcp_sock for the
263 * socket, and return result with RFC1323 scaling applied. The return
264 * value can be stuffed directly into th->window for an outgoing
265 * frame.
266 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800267static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
269 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphal8e165e22014-02-25 14:34:32 +0100270 u32 old_win = tp->rcv_wnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 u32 cur_win = tcp_receive_window(tp);
272 u32 new_win = __tcp_select_window(sk);
273
274 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800275 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 /* Danger Will Robinson!
277 * Don't update rcv_wup/rcv_wnd here or else
278 * we will not be able to advertise a zero
279 * window in time. --DaveM
280 *
281 * Relax Will Robinson.
282 */
Florian Westphal8e165e22014-02-25 14:34:32 +0100283 if (new_win == 0)
284 NET_INC_STATS(sock_net(sk),
285 LINUX_MIB_TCPWANTZEROWINDOWADV);
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700286 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 }
288 tp->rcv_wnd = new_win;
289 tp->rcv_wup = tp->rcv_nxt;
290
291 /* Make sure we do not exceed the maximum possible
292 * scaled window.
293 */
Rick Jones15d99e02006-03-20 22:40:29 -0800294 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 new_win = min(new_win, MAX_TCP_WINDOW);
296 else
297 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
298
299 /* RFC1323 scaling applied */
300 new_win >>= tp->rx_opt.rcv_wscale;
301
302 /* If we advertise zero window, disable fast path. */
Florian Westphal8e165e22014-02-25 14:34:32 +0100303 if (new_win == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 tp->pred_flags = 0;
Florian Westphal8e165e22014-02-25 14:34:32 +0100305 if (old_win)
306 NET_INC_STATS(sock_net(sk),
307 LINUX_MIB_TCPTOZEROWINDOWADV);
308 } else if (old_win == 0) {
309 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
310 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
312 return new_win;
313}
314
Andi Kleen67edfef2009-07-21 23:00:40 +0000315/* Packet ECN state for a SYN-ACK */
Florian Westphal735d3832014-09-29 13:08:30 +0200316static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700317{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200318 const struct tcp_sock *tp = tcp_sk(sk);
319
Eric Dumazet4de075e2011-09-27 13:25:05 -0400320 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800321 if (!(tp->ecn_flags & TCP_ECN_OK))
Eric Dumazet4de075e2011-09-27 13:25:05 -0400322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200323 else if (tcp_ca_needs_ecn(sk))
324 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700325}
326
Andi Kleen67edfef2009-07-21 23:00:40 +0000327/* Packet ECN state for a SYN. */
Florian Westphal735d3832014-09-29 13:08:30 +0200328static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700329{
330 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100331 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
332 tcp_ca_needs_ecn(sk);
333
334 if (!use_ecn) {
335 const struct dst_entry *dst = __sk_dst_get(sk);
336
337 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
338 use_ecn = true;
339 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700340
341 tp->ecn_flags = 0;
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100342
343 if (use_ecn) {
Eric Dumazet4de075e2011-09-27 13:25:05 -0400344 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700345 tp->ecn_flags = TCP_ECN_OK;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200346 if (tcp_ca_needs_ecn(sk))
347 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700348 }
349}
350
Daniel Borkmann49213552015-05-19 21:04:22 +0200351static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
352{
353 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
354 /* tp->ecn_flags are cleared at a later point in time when
355 * SYN ACK is ultimatively being received.
356 */
357 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
358}
359
Florian Westphal735d3832014-09-29 13:08:30 +0200360static void
361tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200362 struct sock *sk)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700363{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200364 if (inet_rsk(req)->ecn_ok) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700365 th->ece = 1;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200366 if (tcp_ca_needs_ecn(sk))
367 INET_ECN_xmit(sk);
368 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700369}
370
Andi Kleen67edfef2009-07-21 23:00:40 +0000371/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
372 * be sent.
373 */
Florian Westphal735d3832014-09-29 13:08:30 +0200374static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700375 int tcp_header_len)
376{
377 struct tcp_sock *tp = tcp_sk(sk);
378
379 if (tp->ecn_flags & TCP_ECN_OK) {
380 /* Not-retransmitted data segment: set ECT and inject CWR. */
381 if (skb->len != tcp_header_len &&
382 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
383 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800384 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700385 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
386 tcp_hdr(skb)->cwr = 1;
387 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
388 }
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200389 } else if (!tcp_ca_needs_ecn(sk)) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700390 /* ACK or retransmitted segment: clear ECT|CE */
391 INET_ECN_dontxmit(sk);
392 }
393 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
394 tcp_hdr(skb)->ece = 1;
395 }
396}
397
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800398/* Constructs common control bits of non-data skb. If SYN/FIN is present,
399 * auto increment end seqno.
400 */
401static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
402{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700403 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800404 skb->csum = 0;
405
Eric Dumazet4de075e2011-09-27 13:25:05 -0400406 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800407 TCP_SKB_CB(skb)->sacked = 0;
408
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700409 tcp_skb_pcount_set(skb, 1);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800410
411 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000412 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800413 seq++;
414 TCP_SKB_CB(skb)->end_seq = seq;
415}
416
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000417static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700418{
419 return tp->snd_una != tp->snd_up;
420}
421
Adam Langley33ad7982008-07-19 00:04:31 -0700422#define OPTION_SACK_ADVERTISE (1 << 0)
423#define OPTION_TS (1 << 1)
424#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000425#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000426#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700427
428struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000429 u16 options; /* bit field of OPTION_* */
430 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700431 u8 ws; /* window scale, 0 to disable */
432 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000433 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000434 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000435 __u32 tsval, tsecr; /* need to include OPTION_TS */
436 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700437};
438
Andi Kleen67edfef2009-07-21 23:00:40 +0000439/* Write previously computed TCP options to the packet.
440 *
441 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700442 * TCP options, we learned this through the hard way, so be careful here.
443 * Luckily we can at least blame others for their non-compliance but from
stephen hemminger8e3bff92013-12-08 12:15:44 -0800444 * inter-operability perspective it seems that we're somewhat stuck with
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700445 * the ordering which we have been using if we want to keep working with
446 * those broken things (not that it currently hurts anybody as there isn't
447 * particular reason why the ordering would need to be changed).
448 *
449 * At least SACK_PERM as the first option is known to lead to a disaster
450 * (but it may well be that other scenarios fail similarly).
451 */
Adam Langley33ad7982008-07-19 00:04:31 -0700452static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000453 struct tcp_out_options *opts)
454{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000455 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000456
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000457 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000458 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
459 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000460 /* overload cookie hash location */
461 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700462 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800463 }
Adam Langley33ad7982008-07-19 00:04:31 -0700464
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700465 if (unlikely(opts->mss)) {
466 *ptr++ = htonl((TCPOPT_MSS << 24) |
467 (TCPOLEN_MSS << 16) |
468 opts->mss);
469 }
470
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000471 if (likely(OPTION_TS & options)) {
472 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700473 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
474 (TCPOLEN_SACK_PERM << 16) |
475 (TCPOPT_TIMESTAMP << 8) |
476 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000477 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700478 } else {
479 *ptr++ = htonl((TCPOPT_NOP << 24) |
480 (TCPOPT_NOP << 16) |
481 (TCPOPT_TIMESTAMP << 8) |
482 TCPOLEN_TIMESTAMP);
483 }
484 *ptr++ = htonl(opts->tsval);
485 *ptr++ = htonl(opts->tsecr);
486 }
487
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700489 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_NOP << 16) |
491 (TCPOPT_SACK_PERM << 8) |
492 TCPOLEN_SACK_PERM);
493 }
494
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000495 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700496 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_WINDOW << 16) |
498 (TCPOLEN_WINDOW << 8) |
499 opts->ws);
500 }
501
502 if (unlikely(opts->num_sack_blocks)) {
503 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
504 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800505 int this_sack;
506
507 *ptr++ = htonl((TCPOPT_NOP << 24) |
508 (TCPOPT_NOP << 16) |
509 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700510 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800511 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800512
Adam Langley33ad7982008-07-19 00:04:31 -0700513 for (this_sack = 0; this_sack < opts->num_sack_blocks;
514 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800515 *ptr++ = htonl(sp[this_sack].start_seq);
516 *ptr++ = htonl(sp[this_sack].end_seq);
517 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800518
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000519 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800520 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000521
522 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
523 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
Daniel Lee7f9b8382015-04-06 14:37:26 -0700524 u8 *p = (u8 *)ptr;
525 u32 len; /* Fast Open option length */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000526
Daniel Lee7f9b8382015-04-06 14:37:26 -0700527 if (foc->exp) {
528 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
529 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
530 TCPOPT_FASTOPEN_MAGIC);
531 p += TCPOLEN_EXP_FASTOPEN_BASE;
532 } else {
533 len = TCPOLEN_FASTOPEN_BASE + foc->len;
534 *p++ = TCPOPT_FASTOPEN;
535 *p++ = len;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000536 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700537
538 memcpy(p, foc->val, foc->len);
539 if ((len & 3) == 2) {
540 p[foc->len] = TCPOPT_NOP;
541 p[foc->len + 1] = TCPOPT_NOP;
542 }
543 ptr += (len + 3) >> 2;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000544 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800545}
546
Andi Kleen67edfef2009-07-21 23:00:40 +0000547/* Compute TCP options for SYN packets. This is not the final
548 * network wire format yet.
549 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000550static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700551 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400552 struct tcp_md5sig_key **md5)
553{
Adam Langley33ad7982008-07-19 00:04:31 -0700554 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000555 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000556 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700557
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800558#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700559 *md5 = tp->af_specific->md5_lookup(sk, sk);
560 if (*md5) {
561 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000562 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800563 }
Adam Langley33ad7982008-07-19 00:04:31 -0700564#else
565 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800566#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700567
568 /* We always get an MSS option. The option bytes which will be seen in
569 * normal data packets should timestamps be used, must be in the MSS
570 * advertised. But we subtract them from tp->mss_cache so that
571 * calculations in tcp_sendmsg are simpler etc. So account for this
572 * fact here if necessary. If we don't do this correctly, as a
573 * receiver we won't recognize data packets as being full sized when we
574 * should, and thus we won't abide by the delayed ACK rules correctly.
575 * SACKs don't matter, we never delay an ACK when we have any of those
576 * going out. */
577 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000578 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700579
Ian Morris51456b22015-04-03 09:17:26 +0100580 if (likely(sysctl_tcp_timestamps && !*md5)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700581 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700582 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700583 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000584 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700585 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800586 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700587 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000588 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000589 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700590 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800591 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700592 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700593 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000594 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700595 }
596
Yuchung Cheng783237e2012-07-19 06:43:07 +0000597 if (fastopen && fastopen->cookie.len >= 0) {
Daniel Lee2646c832015-04-06 14:37:27 -0700598 u32 need = fastopen->cookie.len;
599
600 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
601 TCPOLEN_FASTOPEN_BASE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000602 need = (need + 3) & ~3U; /* Align to 32 bits */
603 if (remaining >= need) {
604 opts->options |= OPTION_FAST_OPEN_COOKIE;
605 opts->fastopen_cookie = &fastopen->cookie;
606 remaining -= need;
607 tp->syn_fastopen = 1;
Daniel Lee2646c832015-04-06 14:37:27 -0700608 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000609 }
610 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000611
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000612 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700613}
614
Andi Kleen67edfef2009-07-21 23:00:40 +0000615/* Set up TCP options for SYN-ACKs. */
Eric Dumazet95c96172012-04-15 05:58:06 +0000616static unsigned int tcp_synack_options(struct sock *sk,
Adam Langley33ad7982008-07-19 00:04:31 -0700617 struct request_sock *req,
Eric Dumazet95c96172012-04-15 05:58:06 +0000618 unsigned int mss, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700619 struct tcp_out_options *opts,
Eric Dumazet80f03e22015-03-24 15:58:52 -0700620 const struct tcp_md5sig_key *md5,
Jerry Chu83368862012-08-31 12:29:12 +0000621 struct tcp_fastopen_cookie *foc)
William Allen Simpson4957faade2009-12-02 18:25:27 +0000622{
Adam Langley33ad7982008-07-19 00:04:31 -0700623 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet95c96172012-04-15 05:58:06 +0000624 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Adam Langley33ad7982008-07-19 00:04:31 -0700625
626#ifdef CONFIG_TCP_MD5SIG
Eric Dumazet80f03e22015-03-24 15:58:52 -0700627 if (md5) {
Adam Langley33ad7982008-07-19 00:04:31 -0700628 opts->options |= OPTION_MD5;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000629 remaining -= TCPOLEN_MD5SIG_ALIGNED;
630
631 /* We can't fit any SACK blocks in a packet with MD5 + TS
632 * options. There was discussion about disabling SACK
633 * rather than TS in order to fit in better with old,
634 * buggy kernels, but that was deemed to be unnecessary.
635 */
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700636 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700637 }
Adam Langley33ad7982008-07-19 00:04:31 -0700638#endif
639
William Allen Simpson4957faade2009-12-02 18:25:27 +0000640 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700641 opts->mss = mss;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000642 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700643
644 if (likely(ireq->wscale_ok)) {
645 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000646 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000647 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700648 }
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700649 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700650 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700651 opts->tsval = tcp_skb_timestamp(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700652 opts->tsecr = req->ts_recent;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000653 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700654 }
655 if (likely(ireq->sack_ok)) {
656 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700657 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faade2009-12-02 18:25:27 +0000658 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700659 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700660 if (foc != NULL && foc->len >= 0) {
661 u32 need = foc->len;
662
663 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
664 TCPOLEN_FASTOPEN_BASE;
Jerry Chu83368862012-08-31 12:29:12 +0000665 need = (need + 3) & ~3U; /* Align to 32 bits */
666 if (remaining >= need) {
667 opts->options |= OPTION_FAST_OPEN_COOKIE;
668 opts->fastopen_cookie = foc;
669 remaining -= need;
670 }
671 }
William Allen Simpson4957faade2009-12-02 18:25:27 +0000672
William Allen Simpson4957faade2009-12-02 18:25:27 +0000673 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700674}
675
Andi Kleen67edfef2009-07-21 23:00:40 +0000676/* Compute TCP options for ESTABLISHED sockets. This is not the
677 * final wire format yet.
678 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000679static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700680 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400681 struct tcp_md5sig_key **md5)
682{
Adam Langley33ad7982008-07-19 00:04:31 -0700683 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000684 unsigned int size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000685 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700686
Andi Kleen5843ef42013-09-30 13:29:11 -0700687 opts->options = 0;
688
Adam Langley33ad7982008-07-19 00:04:31 -0700689#ifdef CONFIG_TCP_MD5SIG
690 *md5 = tp->af_specific->md5_lookup(sk, sk);
691 if (unlikely(*md5)) {
692 opts->options |= OPTION_MD5;
693 size += TCPOLEN_MD5SIG_ALIGNED;
694 }
695#else
696 *md5 = NULL;
697#endif
698
699 if (likely(tp->rx_opt.tstamp_ok)) {
700 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700701 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700702 opts->tsecr = tp->rx_opt.ts_recent;
703 size += TCPOLEN_TSTAMP_ALIGNED;
704 }
705
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000706 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
707 if (unlikely(eff_sacks)) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000708 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
Adam Langley33ad7982008-07-19 00:04:31 -0700709 opts->num_sack_blocks =
Eric Dumazet95c96172012-04-15 05:58:06 +0000710 min_t(unsigned int, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700711 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
712 TCPOLEN_SACK_PERBLOCK);
713 size += TCPOLEN_SACK_BASE_ALIGNED +
714 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
715 }
716
717 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800718}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000720
721/* TCP SMALL QUEUES (TSQ)
722 *
723 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
724 * to reduce RTT and bufferbloat.
725 * We do this using a special skb destructor (tcp_wfree).
726 *
727 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
728 * needs to be reallocated in a driver.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800729 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000730 *
731 * Since transmit from skb destructor is forbidden, we use a tasklet
732 * to process all sockets that eventually need to send more skbs.
733 * We use one tasklet per cpu, with its own queue of sockets.
734 */
735struct tsq_tasklet {
736 struct tasklet_struct tasklet;
737 struct list_head head; /* queue of tcp sockets */
738};
739static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
740
Eric Dumazet6f458df2012-07-20 05:45:50 +0000741static void tcp_tsq_handler(struct sock *sk)
742{
743 if ((1 << sk->sk_state) &
744 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
745 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
John Ognessbf062002014-02-09 18:40:11 -0800746 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
747 0, GFP_ATOMIC);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000748}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000749/*
stephen hemminger8e3bff92013-12-08 12:15:44 -0800750 * One tasklet per cpu tries to send more skbs.
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000751 * We run in tasklet context but need to disable irqs when
stephen hemminger8e3bff92013-12-08 12:15:44 -0800752 * transferring tsq->head because tcp_wfree() might
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000753 * interrupt us (non NAPI drivers)
754 */
755static void tcp_tasklet_func(unsigned long data)
756{
757 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
758 LIST_HEAD(list);
759 unsigned long flags;
760 struct list_head *q, *n;
761 struct tcp_sock *tp;
762 struct sock *sk;
763
764 local_irq_save(flags);
765 list_splice_init(&tsq->head, &list);
766 local_irq_restore(flags);
767
768 list_for_each_safe(q, n, &list) {
769 tp = list_entry(q, struct tcp_sock, tsq_node);
770 list_del(&tp->tsq_node);
771
772 sk = (struct sock *)tp;
773 bh_lock_sock(sk);
774
775 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000776 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000777 } else {
778 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000779 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000780 }
781 bh_unlock_sock(sk);
782
783 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
784 sk_free(sk);
785 }
786}
787
Eric Dumazet6f458df2012-07-20 05:45:50 +0000788#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
789 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200790 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
791 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000792/**
793 * tcp_release_cb - tcp release_sock() callback
794 * @sk: socket
795 *
796 * called from release_sock() to perform protocol dependent
797 * actions before socket release.
798 */
799void tcp_release_cb(struct sock *sk)
800{
801 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000802 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000803
Eric Dumazet6f458df2012-07-20 05:45:50 +0000804 /* perform an atomic operation only if at least one flag is set */
805 do {
806 flags = tp->tsq_flags;
807 if (!(flags & TCP_DEFERRED_ALL))
808 return;
809 nflags = flags & ~TCP_DEFERRED_ALL;
810 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
811
812 if (flags & (1UL << TCP_TSQ_DEFERRED))
813 tcp_tsq_handler(sk);
814
Eric Dumazetc3f9b012014-03-10 09:50:11 -0700815 /* Here begins the tricky part :
816 * We are called from release_sock() with :
817 * 1) BH disabled
818 * 2) sk_lock.slock spinlock held
819 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 *
821 * But following code is meant to be called from BH handlers,
822 * so we should keep BH disabled, but early release socket ownership
823 */
824 sock_release_ownership(sk);
825
Eric Dumazet144d56e2012-08-20 00:22:46 +0000826 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000827 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000828 __sock_put(sk);
829 }
830 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000831 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000832 __sock_put(sk);
833 }
834 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Neal Cardwell4fab9072014-08-14 12:40:05 -0400835 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000836 __sock_put(sk);
837 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000838}
839EXPORT_SYMBOL(tcp_release_cb);
840
841void __init tcp_tasklet_init(void)
842{
843 int i;
844
845 for_each_possible_cpu(i) {
846 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
847
848 INIT_LIST_HEAD(&tsq->head);
849 tasklet_init(&tsq->tasklet,
850 tcp_tasklet_func,
851 (unsigned long)tsq);
852 }
853}
854
855/*
856 * Write buffer destructor automatically called from kfree_skb.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800857 * We can't xmit new skbs from this context, as we might already
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000858 * hold qdisc lock.
859 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000860void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000861{
862 struct sock *sk = skb->sk;
863 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700864 int wmem;
865
866 /* Keep one reference on sk_wmem_alloc.
867 * Will be released by sk_free() from here or tcp_tasklet_func()
868 */
869 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
870
871 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
872 * Wait until our queues (qdisc + devices) are drained.
873 * This gives :
874 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
875 * - chance for incoming ACK (processed by another cpu maybe)
876 * to migrate this flow (skb->ooo_okay will be eventually set)
877 */
878 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
879 goto out;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000880
881 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
882 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
883 unsigned long flags;
884 struct tsq_tasklet *tsq;
885
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000886 /* queue this socket to tasklet queue */
887 local_irq_save(flags);
Christoph Lameter903ceff2014-08-17 12:30:35 -0500888 tsq = this_cpu_ptr(&tsq_tasklet);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000889 list_add(&tp->tsq_node, &tsq->head);
890 tasklet_schedule(&tsq->tasklet);
891 local_irq_restore(flags);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700892 return;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000893 }
Eric Dumazet9b462d02014-10-13 06:27:47 -0700894out:
895 sk_free(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000896}
897
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898/* This routine actually transmits TCP packets queued in by
899 * tcp_do_sendmsg(). This is used by both the initial
900 * transmission and possible later retransmissions.
901 * All SKB's seen here are completely headerless. It is our
902 * job to build the TCP header, and pass the packet down to
903 * IP so it can do the same plus pass the packet off to the
904 * device.
905 *
906 * We are working here with either a clone of the original
907 * SKB, or a fresh unique copy made by the retransmit engine.
908 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800909static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
910 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800912 const struct inet_connection_sock *icsk = inet_csk(sk);
913 struct inet_sock *inet;
914 struct tcp_sock *tp;
915 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700916 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000917 unsigned int tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800918 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800919 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800920 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800922 BUG_ON(!skb || !tcp_skb_pcount(skb));
923
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700924 if (clone_it) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800925 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700926
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800927 if (unlikely(skb_cloned(skb)))
928 skb = pskb_copy(skb, gfp_mask);
929 else
930 skb = skb_clone(skb, gfp_mask);
931 if (unlikely(!skb))
932 return -ENOBUFS;
933 }
934
935 inet = inet_sk(sk);
936 tp = tcp_sk(sk);
937 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700938 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939
Eric Dumazet4de075e2011-09-27 13:25:05 -0400940 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700941 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
942 else
943 tcp_options_size = tcp_established_options(sk, skb, &opts,
944 &md5);
945 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900946
Eric Dumazet547669d2013-05-23 07:44:20 +0000947 if (tcp_packets_in_flight(tp) == 0)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800948 tcp_ca_event(sk, CA_EVENT_TX_START);
Eric Dumazet547669d2013-05-23 07:44:20 +0000949
950 /* if no packet is in qdisc/device queue, then allow XPS to select
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700951 * another queue. We can be called from tcp_tsq_handler()
952 * which holds one reference to sk_wmem_alloc.
953 *
954 * TODO: Ideally, in-flight pure ACK packets should not matter here.
955 * One way to get this would be to set skb->truesize = 2 on them.
Eric Dumazet547669d2013-05-23 07:44:20 +0000956 */
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700957 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800958
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700959 skb_push(skb, tcp_header_size);
960 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000961
962 skb_orphan(skb);
963 skb->sk = sk;
Eric Dumazet98781962015-02-03 18:31:53 -0800964 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
Tom Herbertb73c3d02014-07-01 21:32:17 -0700965 skb_set_hash_from_sk(skb, sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000966 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800967
968 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700969 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000970 th->source = inet->inet_sport;
971 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800972 th->seq = htonl(tcb->seq);
973 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700974 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400975 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800976
Eric Dumazet4de075e2011-09-27 13:25:05 -0400977 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800978 /* RFC1323: The window in SYN & SYN/ACK segments
979 * is never scaled.
980 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800981 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800982 } else {
983 th->window = htons(tcp_select_window(sk));
984 }
985 th->check = 0;
986 th->urg_ptr = 0;
987
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700988 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800989 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
990 if (before(tp->snd_up, tcb->seq + 0x10000)) {
991 th->urg_ptr = htons(tp->snd_up - tcb->seq);
992 th->urg = 1;
993 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700994 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800995 th->urg = 1;
996 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800997 }
998
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000999 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet51466a72015-06-11 09:15:16 -07001000 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001001 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
Florian Westphal735d3832014-09-29 13:08:30 +02001002 tcp_ecn_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001003
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001004#ifdef CONFIG_TCP_MD5SIG
1005 /* Calculate the MD5 hash, as we have all we need now */
1006 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -07001007 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00001008 tp->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07001009 md5, sk, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001010 }
1011#endif
1012
Herbert Xubb296242010-04-11 02:15:55 +00001013 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001014
Eric Dumazet4de075e2011-09-27 13:25:05 -04001015 if (likely(tcb->tcp_flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001016 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1017
1018 if (skb->len != tcp_header_size)
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001019 tcp_event_data_sent(tp, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001020
Wei Yongjunbd37a082006-08-07 21:04:15 -07001021 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +00001022 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1023 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001024
Marcelo Ricardo Leitner2efd0552015-05-20 16:35:41 -07001025 tp->segs_out += tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001026 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001027 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001028 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001029
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001030 /* Our usage of tstamp should remain private */
1031 skb->tstamp.tv64 = 0;
Eric Dumazet971f10e2014-09-27 09:50:57 -07001032
1033 /* Cleanup our debris for IP stacks */
1034 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1035 sizeof(struct inet6_skb_parm)));
1036
Eric Dumazetb0270e92014-04-15 12:58:34 -04001037 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001038
Hua Zhong83de47c2006-04-28 15:26:50 -07001039 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001040 return err;
1041
Christoph Paasch5ee2c942014-07-14 16:58:32 +02001042 tcp_enter_cwr(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001043
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -02001044 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045}
1046
Andi Kleen67edfef2009-07-21 23:00:40 +00001047/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 *
1049 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1050 * otherwise socket can stall.
1051 */
1052static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1053{
1054 struct tcp_sock *tp = tcp_sk(sk);
1055
1056 /* Advance write_seq and place onto the write_queue. */
1057 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001058 __skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -08001059 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001060 sk->sk_wmem_queued += skb->truesize;
1061 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062}
1063
Andi Kleen67edfef2009-07-21 23:00:40 +00001064/* Initialize TSO segments for a packet. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001065static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -07001066{
Eric Dumazet8f26fb12013-10-15 12:24:54 -07001067 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -07001068 /* Avoid the costly divide in the normal
1069 * non-TSO case.
1070 */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001071 tcp_skb_pcount_set(skb, 1);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001072 TCP_SKB_CB(skb)->tcp_gso_size = 0;
David S. Millerf6302d12005-07-05 15:18:03 -07001073 } else {
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001074 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
Eric Dumazetf69ad292015-06-11 09:15:18 -07001075 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 }
1077}
1078
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001079/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001080 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001081 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001082static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001083 int decr)
1084{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001085 struct tcp_sock *tp = tcp_sk(sk);
1086
Ilpo Järvinendc869672007-10-01 15:27:19 -07001087 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001088 return;
1089
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001090 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001091 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001092}
1093
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001094/* Pcount in the middle of the write queue got changed, we need to do various
1095 * tweaks to fix counters
1096 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001097static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001098{
1099 struct tcp_sock *tp = tcp_sk(sk);
1100
1101 tp->packets_out -= decr;
1102
1103 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1104 tp->sacked_out -= decr;
1105 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1106 tp->retrans_out -= decr;
1107 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1108 tp->lost_out -= decr;
1109
1110 /* Reno case is special. Sigh... */
1111 if (tcp_is_reno(tp) && decr > 0)
1112 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1113
1114 tcp_adjust_fackets_out(sk, skb, decr);
1115
1116 if (tp->lost_skb_hint &&
1117 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001118 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001119 tp->lost_cnt_hint -= decr;
1120
1121 tcp_verify_left_out(tp);
1122}
1123
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001124static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1125{
1126 struct skb_shared_info *shinfo = skb_shinfo(skb);
1127
1128 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1129 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1130 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1131 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1132
1133 shinfo->tx_flags &= ~tsflags;
1134 shinfo2->tx_flags |= tsflags;
1135 swap(shinfo->tskey, shinfo2->tskey);
1136 }
1137}
1138
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139/* Function to create two new TCP segments. Shrinks the given segment
1140 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001141 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 * Remember, these are still headerless SKBs at this point.
1143 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001144int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001145 unsigned int mss_now, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146{
1147 struct tcp_sock *tp = tcp_sk(sk);
1148 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001149 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001150 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001151 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001153 if (WARN_ON(len > skb->len))
1154 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001155
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 nsize = skb_headlen(skb) - len;
1157 if (nsize < 0)
1158 nsize = 0;
1159
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001160 if (skb_unclone(skb, gfp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 return -ENOMEM;
1162
1163 /* Get a new skb... force flag on. */
Eric Dumazeteb934472015-05-19 13:26:55 -07001164 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001165 if (!buff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001167
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001168 sk->sk_wmem_queued += buff->truesize;
1169 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001170 nlen = skb->len - len - nsize;
1171 buff->truesize += nlen;
1172 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173
1174 /* Correct the sequence numbers. */
1175 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1176 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1177 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1178
1179 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001180 flags = TCP_SKB_CB(skb)->tcp_flags;
1181 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1182 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001183 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
Patrick McHardy84fa7932006-08-29 16:44:56 -07001185 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001187 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1188 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 nsize, 0);
1190
1191 skb_trim(skb, len);
1192
1193 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1194 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001195 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 skb_split(skb, buff, len);
1197 }
1198
1199 buff->ip_summed = skb->ip_summed;
1200
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001201 buff->tstamp = skb->tstamp;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001202 tcp_fragment_tstamp(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203
David S. Miller6475be12005-09-01 22:47:01 -07001204 old_factor = tcp_skb_pcount(skb);
1205
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001207 tcp_set_skb_tso_segs(skb, mss_now);
1208 tcp_set_skb_tso_segs(buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209
David S. Miller6475be12005-09-01 22:47:01 -07001210 /* If this packet has been sent out already, we must
1211 * adjust the various packet counters.
1212 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001213 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001214 int diff = old_factor - tcp_skb_pcount(skb) -
1215 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001217 if (diff)
1218 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 }
1220
1221 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001222 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001223 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224
1225 return 0;
1226}
1227
1228/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1229 * eventually). The difference is that pulled data not copied, but
1230 * immediately discarded.
1231 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001232static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001234 struct skb_shared_info *shinfo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 int i, k, eat;
1236
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001237 eat = min_t(int, len, skb_headlen(skb));
1238 if (eat) {
1239 __skb_pull(skb, eat);
1240 len -= eat;
1241 if (!len)
1242 return;
1243 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 eat = len;
1245 k = 0;
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001246 shinfo = skb_shinfo(skb);
1247 for (i = 0; i < shinfo->nr_frags; i++) {
1248 int size = skb_frag_size(&shinfo->frags[i]);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001249
1250 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001251 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001252 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 } else {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001254 shinfo->frags[k] = shinfo->frags[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 if (eat) {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001256 shinfo->frags[k].page_offset += eat;
1257 skb_frag_size_sub(&shinfo->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 eat = 0;
1259 }
1260 k++;
1261 }
1262 }
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001263 shinfo->nr_frags = k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001265 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 skb->data_len -= len;
1267 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268}
1269
Andi Kleen67edfef2009-07-21 23:00:40 +00001270/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1272{
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001273 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 return -ENOMEM;
1275
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001276 __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
1278 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001279 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280
1281 skb->truesize -= len;
1282 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001283 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1285
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001286 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 if (tcp_skb_pcount(skb) > 1)
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001288 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289
1290 return 0;
1291}
1292
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001293/* Calculate MSS not accounting any TCP options. */
1294static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001295{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001296 const struct tcp_sock *tp = tcp_sk(sk);
1297 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001298 int mss_now;
1299
1300 /* Calculate base mss without TCP options:
1301 It is MMS_S - sizeof(tcphdr) of rfc1122
1302 */
1303 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1304
Eric Dumazet67469602012-04-24 07:37:38 +00001305 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1306 if (icsk->icsk_af_ops->net_frag_header_len) {
1307 const struct dst_entry *dst = __sk_dst_get(sk);
1308
1309 if (dst && dst_allfrag(dst))
1310 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1311 }
1312
John Heffner5d424d52006-03-20 17:53:41 -08001313 /* Clamp it (mss_clamp does not include tcp options) */
1314 if (mss_now > tp->rx_opt.mss_clamp)
1315 mss_now = tp->rx_opt.mss_clamp;
1316
1317 /* Now subtract optional transport overhead */
1318 mss_now -= icsk->icsk_ext_hdr_len;
1319
1320 /* Then reserve room for full set of TCP options and 8 bytes of data */
1321 if (mss_now < 48)
1322 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001323 return mss_now;
1324}
1325
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001326/* Calculate MSS. Not accounting for SACKs here. */
1327int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1328{
1329 /* Subtract TCP options size, not including SACKs */
1330 return __tcp_mtu_to_mss(sk, pmtu) -
1331 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1332}
1333
John Heffner5d424d52006-03-20 17:53:41 -08001334/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001335int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001336{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001337 const struct tcp_sock *tp = tcp_sk(sk);
1338 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001339 int mtu;
1340
1341 mtu = mss +
1342 tp->tcp_header_len +
1343 icsk->icsk_ext_hdr_len +
1344 icsk->icsk_af_ops->net_header_len;
1345
Eric Dumazet67469602012-04-24 07:37:38 +00001346 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1347 if (icsk->icsk_af_ops->net_frag_header_len) {
1348 const struct dst_entry *dst = __sk_dst_get(sk);
1349
1350 if (dst && dst_allfrag(dst))
1351 mtu += icsk->icsk_af_ops->net_frag_header_len;
1352 }
John Heffner5d424d52006-03-20 17:53:41 -08001353 return mtu;
1354}
1355
Andi Kleen67edfef2009-07-21 23:00:40 +00001356/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001357void tcp_mtup_init(struct sock *sk)
1358{
1359 struct tcp_sock *tp = tcp_sk(sk);
1360 struct inet_connection_sock *icsk = inet_csk(sk);
Fan Dub0f9ca52015-02-10 09:53:16 +08001361 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001362
Fan Dub0f9ca52015-02-10 09:53:16 +08001363 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
John Heffner5d424d52006-03-20 17:53:41 -08001364 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001365 icsk->icsk_af_ops->net_header_len;
Fan Dub0f9ca52015-02-10 09:53:16 +08001366 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
John Heffner5d424d52006-03-20 17:53:41 -08001367 icsk->icsk_mtup.probe_size = 0;
Fan Du05cbc0d2015-03-06 11:18:24 +08001368 if (icsk->icsk_mtup.enabled)
1369 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
John Heffner5d424d52006-03-20 17:53:41 -08001370}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001371EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001372
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373/* This function synchronize snd mss to current pmtu/exthdr set.
1374
1375 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1376 for TCP options, but includes only bare TCP header.
1377
1378 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001379 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 It also does not include TCP options.
1381
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001382 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383
1384 tp->mss_cache is current effective sending mss, including
1385 all tcp options except for SACKs. It is evaluated,
1386 taking into account current pmtu, but never exceeds
1387 tp->rx_opt.mss_clamp.
1388
1389 NOTE1. rfc1122 clearly states that advertised MSS
1390 DOES NOT include either tcp or ip options.
1391
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001392 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1393 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1396{
1397 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001398 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001399 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
John Heffner5d424d52006-03-20 17:53:41 -08001401 if (icsk->icsk_mtup.search_high > pmtu)
1402 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403
John Heffner5d424d52006-03-20 17:53:41 -08001404 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001405 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406
1407 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001408 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001409 if (icsk->icsk_mtup.enabled)
1410 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001411 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412
1413 return mss_now;
1414}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001415EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416
1417/* Compute the current effective MSS, taking SACKs and IP options,
1418 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001419 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001420unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001422 const struct tcp_sock *tp = tcp_sk(sk);
1423 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001424 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001425 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001426 struct tcp_out_options opts;
1427 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001429 mss_now = tp->mss_cache;
1430
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 if (dst) {
1432 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001433 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 mss_now = tcp_sync_mss(sk, mtu);
1435 }
1436
Adam Langley33ad7982008-07-19 00:04:31 -07001437 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1438 sizeof(struct tcphdr);
1439 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1440 * some common options. If this is an odd packet (because we have SACK
1441 * blocks etc) then our calculated header_len will be different, and
1442 * we have to adjust mss_now correspondingly */
1443 if (header_len != tp->tcp_header_len) {
1444 int delta = (int) header_len - tp->tcp_header_len;
1445 mss_now -= delta;
1446 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001447
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 return mss_now;
1449}
1450
Weiping Pan86fd14a2014-04-18 12:27:46 +08001451/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1452 * As additional protections, we do not touch cwnd in retransmission phases,
1453 * and if application hit its sndbuf limit recently.
1454 */
1455static void tcp_cwnd_application_limited(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001456{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001457 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001458
Weiping Pan86fd14a2014-04-18 12:27:46 +08001459 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1460 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1461 /* Limited by application or receiver window. */
1462 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1463 u32 win_used = max(tp->snd_cwnd_used, init_win);
1464 if (win_used < tp->snd_cwnd) {
1465 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1466 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1467 }
1468 tp->snd_cwnd_used = 0;
1469 }
1470 tp->snd_cwnd_stamp = tcp_time_stamp;
1471}
1472
Neal Cardwellca8a2262014-05-22 10:41:08 -04001473static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
David S. Millera762a982005-07-05 15:18:51 -07001474{
1475 struct tcp_sock *tp = tcp_sk(sk);
1476
Neal Cardwellca8a2262014-05-22 10:41:08 -04001477 /* Track the maximum number of outstanding packets in each
1478 * window, and remember whether we were cwnd-limited then.
1479 */
1480 if (!before(tp->snd_una, tp->max_packets_seq) ||
1481 tp->packets_out > tp->max_packets_out) {
1482 tp->max_packets_out = tp->packets_out;
1483 tp->max_packets_seq = tp->snd_nxt;
1484 tp->is_cwnd_limited = is_cwnd_limited;
1485 }
Eric Dumazete114a712014-04-30 11:58:13 -07001486
Eric Dumazet24901552014-05-02 21:18:05 -07001487 if (tcp_is_cwnd_limited(sk)) {
David S. Millera762a982005-07-05 15:18:51 -07001488 /* Network is feed fully. */
1489 tp->snd_cwnd_used = 0;
1490 tp->snd_cwnd_stamp = tcp_time_stamp;
1491 } else {
1492 /* Network starves. */
1493 if (tp->packets_out > tp->snd_cwnd_used)
1494 tp->snd_cwnd_used = tp->packets_out;
1495
David S. Miller15d33c02007-04-09 13:23:14 -07001496 if (sysctl_tcp_slow_start_after_idle &&
1497 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001498 tcp_cwnd_application_limited(sk);
1499 }
1500}
1501
Eric Dumazetd4589922013-12-13 13:51:23 -08001502/* Minshall's variant of the Nagle send check. */
1503static bool tcp_minshall_check(const struct tcp_sock *tp)
1504{
1505 return after(tp->snd_sml, tp->snd_una) &&
1506 !after(tp->snd_sml, tp->snd_nxt);
1507}
1508
1509/* Update snd_sml if this skb is under mss
1510 * Note that a TSO packet might end with a sub-mss segment
1511 * The test is really :
1512 * if ((skb->len % mss) != 0)
1513 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1514 * But we can avoid doing the divide again given we already have
1515 * skb_pcount = skb->len / mss_now
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001516 */
Eric Dumazetd4589922013-12-13 13:51:23 -08001517static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1518 const struct sk_buff *skb)
1519{
1520 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1521 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1522}
1523
1524/* Return false, if packet can be sent now without violation Nagle's rules:
1525 * 1. It is full sized. (provided by caller in %partial bool)
1526 * 2. Or it contains FIN. (already checked by caller)
1527 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1528 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1529 * With Minshall's modification: all sent small packets are ACKed.
1530 */
1531static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001532 int nonagle)
Eric Dumazetd4589922013-12-13 13:51:23 -08001533{
1534 return partial &&
1535 ((nonagle & TCP_NAGLE_CORK) ||
1536 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1537}
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001538
1539/* Return how many segs we'd like on a TSO packet,
1540 * to send one TSO packet per ms
1541 */
1542static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1543{
1544 u32 bytes, segs;
1545
1546 bytes = min(sk->sk_pacing_rate >> 10,
1547 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1548
1549 /* Goal is to send at least one packet per ms,
1550 * not one big TSO packet every 100 ms.
1551 * This preserves ACK clocking and is consistent
1552 * with tcp_tso_should_defer() heuristic.
1553 */
1554 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1555
1556 return min_t(u32, segs, sk->sk_gso_max_segs);
1557}
1558
Eric Dumazetd4589922013-12-13 13:51:23 -08001559/* Returns the portion of skb which can be sent right away */
1560static unsigned int tcp_mss_split_point(const struct sock *sk,
1561 const struct sk_buff *skb,
1562 unsigned int mss_now,
1563 unsigned int max_segs,
1564 int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001565{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001566 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetd4589922013-12-13 13:51:23 -08001567 u32 partial, needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001568
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001569 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001570 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001571
Ben Hutchings14853482012-07-30 16:11:42 +00001572 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1573 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001574
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001575 needed = min(skb->len, window);
1576
Ben Hutchings14853482012-07-30 16:11:42 +00001577 if (max_len <= needed)
1578 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001579
Eric Dumazetd4589922013-12-13 13:51:23 -08001580 partial = needed % mss_now;
1581 /* If last segment is not a full MSS, check if Nagle rules allow us
1582 * to include this last segment in this skb.
1583 * Otherwise, we'll split the skb at last MSS boundary
1584 */
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001585 if (tcp_nagle_check(partial != 0, tp, nonagle))
Eric Dumazetd4589922013-12-13 13:51:23 -08001586 return needed - partial;
1587
1588 return needed;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001589}
1590
1591/* Can at least one segment of SKB be sent right now, according to the
1592 * congestion window rules? If so, return how many segments are allowed.
1593 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001594static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1595 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001596{
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001597 u32 in_flight, cwnd, halfcwnd;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001598
1599 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001600 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1601 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001602 return 1;
1603
1604 in_flight = tcp_packets_in_flight(tp);
1605 cwnd = tp->snd_cwnd;
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001606 if (in_flight >= cwnd)
1607 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001608
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001609 /* For better scheduling, ensure we have at least
1610 * 2 GSO packets in flight.
1611 */
1612 halfcwnd = max(cwnd >> 1, 1U);
1613 return min(halfcwnd, cwnd - in_flight);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001614}
1615
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001616/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001617 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001618 * SKB onto the wire.
1619 */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001620static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001621{
1622 int tso_segs = tcp_skb_pcount(skb);
1623
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001624 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001625 tcp_set_skb_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001626 tso_segs = tcp_skb_pcount(skb);
1627 }
1628 return tso_segs;
1629}
1630
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001631
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001632/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001633 * sent now.
1634 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001635static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1636 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001637{
1638 /* Nagle rule does not apply to frames, which sit in the middle of the
1639 * write_queue (they have no chances to get new data).
1640 *
1641 * This is implemented in the callers, where they modify the 'nonagle'
1642 * argument based upon the location of SKB in the send queue.
1643 */
1644 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001645 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001646
Yuchung Cheng9b441902013-03-20 13:32:58 +00001647 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1648 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001649 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001650
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001651 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001652 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001653
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001654 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001655}
1656
1657/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001658static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1659 const struct sk_buff *skb,
1660 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001661{
1662 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1663
1664 if (skb->len > cur_mss)
1665 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1666
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001667 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001668}
1669
David S. Millerfe067e82007-03-07 12:12:44 -08001670/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001671 * should be put on the wire right now. If so, it returns the number of
1672 * packets allowed by the congestion window.
1673 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001674static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001675 unsigned int cur_mss, int nonagle)
1676{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001677 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001678 unsigned int cwnd_quota;
1679
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001680 tcp_init_tso_segs(skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001681
1682 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1683 return 0;
1684
1685 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001686 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001687 cwnd_quota = 0;
1688
1689 return cwnd_quota;
1690}
1691
Andi Kleen67edfef2009-07-21 23:00:40 +00001692/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001693bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001694{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001695 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001696 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001697
Eric Dumazeta02cec22010-09-22 20:43:57 +00001698 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001699 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001700 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001701 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001702}
1703
1704/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1705 * which is put after SKB on the list. It is very much like
1706 * tcp_fragment() except that it may make several kinds of assumptions
1707 * in order to speed up the splitting operation. In particular, we
1708 * know that all the data is in scatter-gather pages, and that the
1709 * packet has never been sent out before (and thus is not cloned).
1710 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001711static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001712 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001713{
1714 struct sk_buff *buff;
1715 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001716 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001717
1718 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001719 if (skb->len != skb->data_len)
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001720 return tcp_fragment(sk, skb, len, mss_now, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001721
Eric Dumazeteb934472015-05-19 13:26:55 -07001722 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001723 if (unlikely(!buff))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001724 return -ENOMEM;
1725
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001726 sk->sk_wmem_queued += buff->truesize;
1727 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001728 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001729 skb->truesize -= nlen;
1730
1731 /* Correct the sequence numbers. */
1732 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1733 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1734 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1735
1736 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001737 flags = TCP_SKB_CB(skb)->tcp_flags;
1738 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1739 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001740
1741 /* This packet was never sent out yet, so no SACK bits. */
1742 TCP_SKB_CB(buff)->sacked = 0;
1743
Patrick McHardy84fa7932006-08-29 16:44:56 -07001744 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001745 skb_split(skb, buff, len);
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001746 tcp_fragment_tstamp(skb, buff);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001747
1748 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001749 tcp_set_skb_tso_segs(skb, mss_now);
1750 tcp_set_skb_tso_segs(buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001751
1752 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001753 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001754 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001755
1756 return 0;
1757}
1758
1759/* Try to defer sending, if possible, in order to minimize the amount
1760 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1761 *
1762 * This algorithm is from John Heffner.
1763 */
Neal Cardwellca8a2262014-05-22 10:41:08 -04001764static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001765 bool *is_cwnd_limited, u32 max_segs)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001766{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001767 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet50c83392015-02-26 14:10:19 -08001768 u32 age, send_win, cong_win, limit, in_flight;
1769 struct tcp_sock *tp = tcp_sk(sk);
1770 struct skb_mstamp now;
1771 struct sk_buff *head;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001772 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001773
Eric Dumazet4de075e2011-09-27 13:25:05 -04001774 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001775 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001776
Eric Dumazet99d76622015-07-26 09:45:24 +02001777 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
John Heffnerae8064a2006-10-18 20:36:48 -07001778 goto send_now;
1779
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001780 /* Avoid bursty behavior by allowing defer
1781 * only if the last write was recent.
1782 */
1783 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
John Heffnerae8064a2006-10-18 20:36:48 -07001784 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001785
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001786 in_flight = tcp_packets_in_flight(tp);
1787
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001788 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001789
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001790 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001791
1792 /* From in_flight test above, we know that cwnd > in_flight. */
1793 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1794
1795 limit = min(send_win, cong_win);
1796
David S. Millerba244fe2006-03-11 18:51:49 -08001797 /* If a full-sized TSO skb can be sent, do it. */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001798 if (limit >= max_segs * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001799 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001800
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001801 /* Middle in queue won't get any more data, full sendable already? */
1802 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1803 goto send_now;
1804
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001805 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1806 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001807 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1808
1809 /* If at least some fraction of a window is available,
1810 * just use it.
1811 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001812 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001813 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001814 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001815 } else {
1816 /* Different approach, try not to defer past a single
1817 * ACK. Receiver should ACK every other full sized
1818 * frame, so if we have space for more than 3 frames
1819 * then send now.
1820 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001821 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001822 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001823 }
1824
Eric Dumazet50c83392015-02-26 14:10:19 -08001825 head = tcp_write_queue_head(sk);
1826 skb_mstamp_get(&now);
1827 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1828 /* If next ACK is likely to come too late (half srtt), do not defer */
1829 if (age < (tp->srtt_us >> 4))
1830 goto send_now;
1831
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001832 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001833
Neal Cardwellca8a2262014-05-22 10:41:08 -04001834 if (cong_win < send_win && cong_win < skb->len)
1835 *is_cwnd_limited = true;
1836
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001837 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001838
1839send_now:
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001840 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001841}
1842
Fan Du05cbc0d2015-03-06 11:18:24 +08001843static inline void tcp_mtu_check_reprobe(struct sock *sk)
1844{
1845 struct inet_connection_sock *icsk = inet_csk(sk);
1846 struct tcp_sock *tp = tcp_sk(sk);
1847 struct net *net = sock_net(sk);
1848 u32 interval;
1849 s32 delta;
1850
1851 interval = net->ipv4.sysctl_tcp_probe_interval;
1852 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1853 if (unlikely(delta >= interval * HZ)) {
1854 int mss = tcp_current_mss(sk);
1855
1856 /* Update current search range */
1857 icsk->icsk_mtup.probe_size = 0;
1858 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1859 sizeof(struct tcphdr) +
1860 icsk->icsk_af_ops->net_header_len;
1861 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1862
1863 /* Update probe time stamp */
1864 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1865 }
1866}
1867
John Heffner5d424d52006-03-20 17:53:41 -08001868/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001869 * MTU probe is regularly attempting to increase the path MTU by
1870 * deliberately sending larger packets. This discovers routing
1871 * changes resulting in larger path MTUs.
1872 *
John Heffner5d424d52006-03-20 17:53:41 -08001873 * Returns 0 if we should wait to probe (no cwnd available),
1874 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001875 * -1 otherwise
1876 */
John Heffner5d424d52006-03-20 17:53:41 -08001877static int tcp_mtu_probe(struct sock *sk)
1878{
1879 struct tcp_sock *tp = tcp_sk(sk);
1880 struct inet_connection_sock *icsk = inet_csk(sk);
1881 struct sk_buff *skb, *nskb, *next;
Fan Du6b58e0a2015-03-06 11:18:23 +08001882 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001883 int len;
1884 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001885 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001886 int copy;
1887 int mss_now;
Fan Du6b58e0a2015-03-06 11:18:23 +08001888 int interval;
John Heffner5d424d52006-03-20 17:53:41 -08001889
1890 /* Not currently probing/verifying,
1891 * not in recovery,
1892 * have enough cwnd, and
1893 * not SACKing (the variable headers throw things off) */
1894 if (!icsk->icsk_mtup.enabled ||
1895 icsk->icsk_mtup.probe_size ||
1896 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1897 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001898 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001899 return -1;
1900
Fan Du6b58e0a2015-03-06 11:18:23 +08001901 /* Use binary search for probe_size between tcp_mss_base,
1902 * and current mss_clamp. if (search_high - search_low)
1903 * smaller than a threshold, backoff from probing.
1904 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001905 mss_now = tcp_current_mss(sk);
Fan Du6b58e0a2015-03-06 11:18:23 +08001906 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1907 icsk->icsk_mtup.search_low) >> 1);
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001908 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
Fan Du6b58e0a2015-03-06 11:18:23 +08001909 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
Fan Du05cbc0d2015-03-06 11:18:24 +08001910 /* When misfortune happens, we are reprobing actively,
1911 * and then reprobe timer has expired. We stick with current
1912 * probing process by not resetting search range to its orignal.
1913 */
Fan Du6b58e0a2015-03-06 11:18:23 +08001914 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
Fan Du05cbc0d2015-03-06 11:18:24 +08001915 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1916 /* Check whether enough time has elaplased for
1917 * another round of probing.
1918 */
1919 tcp_mtu_check_reprobe(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001920 return -1;
1921 }
1922
1923 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001924 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001925 return -1;
1926
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001927 if (tp->snd_wnd < size_needed)
1928 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001929 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001930 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001931
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001932 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1933 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1934 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001935 return -1;
1936 else
1937 return 0;
1938 }
1939
1940 /* We're allowed to probe. Build it now. */
Eric Dumazeteb934472015-05-19 13:26:55 -07001941 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
Ian Morris51456b22015-04-03 09:17:26 +01001942 if (!nskb)
John Heffner5d424d52006-03-20 17:53:41 -08001943 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001944 sk->sk_wmem_queued += nskb->truesize;
1945 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001946
David S. Millerfe067e82007-03-07 12:12:44 -08001947 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001948
1949 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1950 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001951 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001952 TCP_SKB_CB(nskb)->sacked = 0;
1953 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001954 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001955
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001956 tcp_insert_write_queue_before(nskb, skb, sk);
1957
John Heffner5d424d52006-03-20 17:53:41 -08001958 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001959 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001960 copy = min_t(int, skb->len, probe_size - len);
1961 if (nskb->ip_summed)
1962 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1963 else
1964 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001965 skb_put(nskb, copy),
1966 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001967
1968 if (skb->len <= copy) {
1969 /* We've eaten all the data from this skb.
1970 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001971 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001972 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001973 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001974 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04001975 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001976 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001977 if (!skb_shinfo(skb)->nr_frags) {
1978 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001979 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001980 skb->csum = csum_partial(skb->data,
1981 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001982 } else {
1983 __pskb_trim_head(skb, copy);
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001984 tcp_set_skb_tso_segs(skb, mss_now);
John Heffner5d424d52006-03-20 17:53:41 -08001985 }
1986 TCP_SKB_CB(skb)->seq += copy;
1987 }
1988
1989 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001990
1991 if (len >= probe_size)
1992 break;
John Heffner5d424d52006-03-20 17:53:41 -08001993 }
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001994 tcp_init_tso_segs(nskb, nskb->len);
John Heffner5d424d52006-03-20 17:53:41 -08001995
1996 /* We're ready to send. If this fails, the probe will
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001997 * be resegmented into mss-sized pieces by tcp_write_xmit().
1998 */
John Heffner5d424d52006-03-20 17:53:41 -08001999 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2000 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002001 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08002002 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002003 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08002004
2005 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08002006 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2007 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08002008
2009 return 1;
2010 }
2011
2012 return -1;
2013}
2014
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015/* This routine writes packets to the network. It advances the
2016 * send_head. This happens as incoming acks open up the remote
2017 * window for us.
2018 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08002019 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2020 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2021 * account rare use of URG, this is not a big flaw.
2022 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002023 * Send at most one packet when push_one > 0. Temporarily ignore
2024 * cwnd limit to force at most one packet out when push_one == 2.
2025
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002026 * Returns true, if no segments are in flight and we have queued segments,
2027 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002029static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2030 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031{
2032 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07002033 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002034 unsigned int tso_segs, sent_pkts;
2035 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08002036 int result;
Neal Cardwellca8a2262014-05-22 10:41:08 -04002037 bool is_cwnd_limited = false;
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002038 u32 max_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039
David S. Miller92df7b52005-07-05 15:19:06 -07002040 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002041
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002042 if (!push_one) {
2043 /* Do MTU probing. */
2044 result = tcp_mtu_probe(sk);
2045 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002046 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002047 } else if (result > 0) {
2048 sent_pkts = 1;
2049 }
John Heffner5d424d52006-03-20 17:53:41 -08002050 }
2051
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002052 max_segs = tcp_tso_autosize(sk, mss_now);
David S. Millerfe067e82007-03-07 12:12:44 -08002053 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07002054 unsigned int limit;
2055
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002056 tso_segs = tcp_init_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002057 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07002058
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002059 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002060 /* "skb_mstamp" is used as a start point for the retransmit timer */
2061 skb_mstamp_get(&skb->skb_mstamp);
Andrew Vaginec342322012-11-15 04:03:17 +00002062 goto repair; /* Skip network transmission */
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002063 }
Andrew Vaginec342322012-11-15 04:03:17 +00002064
Herbert Xub68e9f82005-08-04 19:52:02 -07002065 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002066 if (!cwnd_quota) {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002067 is_cwnd_limited = true;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002068 if (push_one == 2)
2069 /* Force out a loss probe pkt. */
2070 cwnd_quota = 1;
2071 else
2072 break;
2073 }
Herbert Xub68e9f82005-08-04 19:52:02 -07002074
2075 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2076 break;
2077
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002078 if (tso_segs == 1) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002079 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2080 (tcp_skb_is_last(sk, skb) ?
2081 nonagle : TCP_NAGLE_PUSH))))
2082 break;
2083 } else {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002084 if (!push_one &&
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002085 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2086 max_segs))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002087 break;
2088 }
David S. Milleraa934662005-07-05 15:20:09 -07002089
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002090 limit = mss_now;
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002091 if (tso_segs > 1 && !tcp_urg_mode(tp))
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002092 limit = tcp_mss_split_point(sk, skb, mss_now,
2093 min_t(unsigned int,
2094 cwnd_quota,
2095 max_segs),
2096 nonagle);
2097
2098 if (skb->len > limit &&
2099 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2100 break;
2101
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002102 /* TCP Small Queues :
2103 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2104 * This allows for :
2105 * - better RTT estimation and ACK scheduling
2106 * - faster recovery
2107 * - high rates
Eric Dumazet98e09382013-11-13 06:32:54 -08002108 * Alas, some drivers / subsystems require a fair amount
2109 * of queued bytes to ensure line rate.
2110 * One example is wifi aggregation (802.11 AMPDU)
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002111 */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002112 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2113 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002114
2115 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002116 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
John Ognessbf062002014-02-09 18:40:11 -08002117 /* It is possible TX completion already happened
2118 * before we set TSQ_THROTTLED, so we must
2119 * test again the condition.
John Ognessbf062002014-02-09 18:40:11 -08002120 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002121 smp_mb__after_atomic();
John Ognessbf062002014-02-09 18:40:11 -08002122 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2123 break;
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002124 }
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002125
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002126 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07002127 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128
Andrew Vaginec342322012-11-15 04:03:17 +00002129repair:
David S. Miller92df7b52005-07-05 15:19:06 -07002130 /* Advance the send_head. This one is sent out.
2131 * This call will increment packets_out.
2132 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002133 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07002134
2135 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002136 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002137
2138 if (push_one)
2139 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 }
David S. Miller92df7b52005-07-05 15:19:06 -07002141
David S. Milleraa934662005-07-05 15:20:09 -07002142 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002143 if (tcp_in_cwnd_reduction(sk))
2144 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002145
2146 /* Send one loss probe per tail loss episode. */
2147 if (push_one != 2)
2148 tcp_schedule_loss_probe(sk);
Neal Cardwellca8a2262014-05-22 10:41:08 -04002149 tcp_cwnd_validate(sk, is_cwnd_limited);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002150 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07002151 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002152 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2153}
2154
2155bool tcp_schedule_loss_probe(struct sock *sk)
2156{
2157 struct inet_connection_sock *icsk = inet_csk(sk);
2158 struct tcp_sock *tp = tcp_sk(sk);
2159 u32 timeout, tlp_time_stamp, rto_time_stamp;
Eric Dumazet740b0f12014-02-26 14:02:48 -08002160 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002161
2162 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2163 return false;
2164 /* No consecutive loss probes. */
2165 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2166 tcp_rearm_rto(sk);
2167 return false;
2168 }
2169 /* Don't do any loss probe on a Fast Open connection before 3WHS
2170 * finishes.
2171 */
2172 if (sk->sk_state == TCP_SYN_RECV)
2173 return false;
2174
2175 /* TLP is only scheduled when next timer event is RTO. */
2176 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2177 return false;
2178
2179 /* Schedule a loss probe in 2*RTT for SACK capable connections
2180 * in Open state, that are either limited by cwnd or application.
2181 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002182 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002183 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2184 return false;
2185
2186 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2187 tcp_send_head(sk))
2188 return false;
2189
2190 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2191 * for delayed ack when there's one outstanding packet.
2192 */
2193 timeout = rtt << 1;
2194 if (tp->packets_out == 1)
2195 timeout = max_t(u32, timeout,
2196 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2197 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2198
2199 /* If RTO is shorter, just schedule TLP in its place. */
2200 tlp_time_stamp = tcp_time_stamp + timeout;
2201 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2202 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2203 s32 delta = rto_time_stamp - tcp_time_stamp;
2204 if (delta > 0)
2205 timeout = delta;
2206 }
2207
2208 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2209 TCP_RTO_MAX);
2210 return true;
2211}
2212
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002213/* Thanks to skb fast clones, we can detect if a prior transmit of
2214 * a packet is still in a qdisc or driver queue.
2215 * In this case, there is very little point doing a retransmit !
2216 * Note: This is called from BH context only.
2217 */
2218static bool skb_still_in_host_queue(const struct sock *sk,
2219 const struct sk_buff *skb)
2220{
Eric Dumazet39bb5e62014-10-30 10:32:34 -07002221 if (unlikely(skb_fclone_busy(sk, skb))) {
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002222 NET_INC_STATS_BH(sock_net(sk),
2223 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2224 return true;
2225 }
2226 return false;
2227}
2228
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002229/* When probe timeout (PTO) fires, send a new segment if one exists, else
2230 * retransmit the last segment.
2231 */
2232void tcp_send_loss_probe(struct sock *sk)
2233{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002234 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002235 struct sk_buff *skb;
2236 int pcount;
2237 int mss = tcp_current_mss(sk);
2238 int err = -1;
2239
Ian Morris00db4122015-04-03 09:17:27 +01002240 if (tcp_send_head(sk)) {
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002241 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2242 goto rearm_timer;
2243 }
2244
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002245 /* At most one outstanding TLP retransmission. */
2246 if (tp->tlp_high_seq)
2247 goto rearm_timer;
2248
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002249 /* Retransmit last segment. */
2250 skb = tcp_write_queue_tail(sk);
2251 if (WARN_ON(!skb))
2252 goto rearm_timer;
2253
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002254 if (skb_still_in_host_queue(sk, skb))
2255 goto rearm_timer;
2256
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002257 pcount = tcp_skb_pcount(skb);
2258 if (WARN_ON(!pcount))
2259 goto rearm_timer;
2260
2261 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002262 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2263 GFP_ATOMIC)))
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002264 goto rearm_timer;
2265 skb = tcp_write_queue_tail(sk);
2266 }
2267
2268 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2269 goto rearm_timer;
2270
Per Hurtigbef19092014-06-12 17:08:32 +02002271 err = __tcp_retransmit_skb(sk, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002272
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002273 /* Record snd_nxt for loss detection. */
2274 if (likely(!err))
2275 tp->tlp_high_seq = tp->snd_nxt;
2276
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002277rearm_timer:
Yuchung Chengfcd16c02015-08-12 11:18:18 -07002278 if (likely(!err)) {
2279 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2280 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2281 inet_csk(sk)->icsk_pending = 0;
2282 }
2283 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284}
2285
David S. Millera762a982005-07-05 15:18:51 -07002286/* Push out any pending frames which were held back due to
2287 * TCP_CORK or attempt at coalescing tiny packets.
2288 * The socket must be locked by the caller.
2289 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002290void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2291 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002292{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002293 /* If we are closed, the bytes will have to remain here.
2294 * In time closedown will finish, we empty the write queue and
2295 * all will be happy.
2296 */
2297 if (unlikely(sk->sk_state == TCP_CLOSE))
2298 return;
2299
Mel Gorman99a1dec2012-07-31 16:44:14 -07002300 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2301 sk_gfp_atomic(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002302 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002303}
2304
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002305/* Send _single_ skb sitting at the send head. This function requires
2306 * true push pending frames to setup probe timer etc.
2307 */
2308void tcp_push_one(struct sock *sk, unsigned int mss_now)
2309{
David S. Millerfe067e82007-03-07 12:12:44 -08002310 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002311
2312 BUG_ON(!skb || skb->len < mss_now);
2313
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002314 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002315}
2316
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317/* This function returns the amount that we can raise the
2318 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002319 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 * 1. The window can never be shrunk once it is offered (RFC 793)
2321 * 2. We limit memory per socket
2322 *
2323 * RFC 1122:
2324 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2325 * RECV.NEXT + RCV.WIN fixed until:
2326 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2327 *
2328 * i.e. don't raise the right edge of the window until you can raise
2329 * it at least MSS bytes.
2330 *
2331 * Unfortunately, the recommended algorithm breaks header prediction,
2332 * since header prediction assumes th->window stays fixed.
2333 *
2334 * Strictly speaking, keeping th->window fixed violates the receiver
2335 * side SWS prevention criteria. The problem is that under this rule
2336 * a stream of single byte packets will cause the right side of the
2337 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002338 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 * Of course, if the sender implements sender side SWS prevention
2340 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002341 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002343 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 * If the free space is less than the 1/4 of the maximum
2345 * space available and the free space is less than 1/2 mss,
2346 * then set the window to 0.
2347 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2348 * Otherwise, just prevent the window from shrinking
2349 * and from being larger than the largest representable value.
2350 *
2351 * This prevents incremental opening of the window in the regime
2352 * where TCP is limited by the speed of the reader side taking
2353 * data out of the TCP receive queue. It does nothing about
2354 * those cases where the window is constrained on the sender side
2355 * because the pipeline is full.
2356 *
2357 * BSD also seems to "accidentally" limit itself to windows that are a
2358 * multiple of MSS, at least until the free space gets quite small.
2359 * This would appear to be a side effect of the mbuf implementation.
2360 * Combining these two algorithms results in the observed behavior
2361 * of having a fixed window size at almost all times.
2362 *
2363 * Below we obtain similar behavior by forcing the offered window to
2364 * a multiple of the mss when it is feasible to do so.
2365 *
2366 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2367 * Regular options like TIMESTAMP are taken into account.
2368 */
2369u32 __tcp_select_window(struct sock *sk)
2370{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002371 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002373 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 * here. I don't know if the value based on our guesses
2375 * of peer's MSS is better for the performance. It's more correct
2376 * but may be worse for the performance because of rcv_mss
2377 * fluctuations. --SAW 1998/11/1
2378 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002379 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 int free_space = tcp_space(sk);
Florian Westphal86c1a042014-02-19 12:51:10 +01002381 int allowed_space = tcp_full_space(sk);
2382 int full_space = min_t(int, tp->window_clamp, allowed_space);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 int window;
2384
2385 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002386 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002388 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002389 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002391 if (tcp_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002392 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2393 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394
Florian Westphal86c1a042014-02-19 12:51:10 +01002395 /* free_space might become our new window, make sure we don't
2396 * increase it due to wscale.
2397 */
2398 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2399
2400 /* if free space is less than mss estimate, or is below 1/16th
2401 * of the maximum allowed, try to move to zero-window, else
2402 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2403 * new incoming data is dropped due to memory limits.
2404 * With large window, mss test triggers way too late in order
2405 * to announce zero window in time before rmem limit kicks in.
2406 */
2407 if (free_space < (allowed_space >> 4) || free_space < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 return 0;
2409 }
2410
2411 if (free_space > tp->rcv_ssthresh)
2412 free_space = tp->rcv_ssthresh;
2413
2414 /* Don't do rounding if we are using window scaling, since the
2415 * scaled window will not line up with the MSS boundary anyway.
2416 */
2417 window = tp->rcv_wnd;
2418 if (tp->rx_opt.rcv_wscale) {
2419 window = free_space;
2420
2421 /* Advertise enough space so that it won't get scaled away.
2422 * Import case: prevent zero window announcement if
2423 * 1<<rcv_wscale > mss.
2424 */
2425 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2426 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2427 << tp->rx_opt.rcv_wscale);
2428 } else {
2429 /* Get the largest window that is a nice multiple of mss.
2430 * Window clamp already applied above.
2431 * If our current window offering is within 1 mss of the
2432 * free space we just keep it. This prevents the divide
2433 * and multiply from happening most of the time.
2434 * We also don't do any window rounding when the free space
2435 * is too small.
2436 */
2437 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002438 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002439 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002440 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002441 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 }
2443
2444 return window;
2445}
2446
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002447/* Collapses two adjacent SKB's during retransmission. */
2448static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449{
2450 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002451 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002452 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002454 skb_size = skb->len;
2455 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002457 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002458
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002459 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002461 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002463 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2464 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002466 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2467 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002469 if (skb->ip_summed != CHECKSUM_PARTIAL)
2470 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002472 /* Update sequence range on original skb. */
2473 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002475 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002476 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002478 /* All done, get rid of second SKB and account for it so
2479 * packet counting does not break.
2480 */
2481 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002482
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002483 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002484 tcp_clear_retrans_hints_partial(tp);
2485 if (next_skb == tp->retransmit_skb_hint)
2486 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002487
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002488 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2489
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002490 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491}
2492
Andi Kleen67edfef2009-07-21 23:00:40 +00002493/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002494static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002495{
2496 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002497 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002498 /* TODO: SACK collapsing could be used to remove this condition */
2499 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002500 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002501 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002502 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002503 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002504 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002505 /* Some heurestics for collapsing over SACK'd could be invented */
2506 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002507 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002508
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002509 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002510}
2511
Andi Kleen67edfef2009-07-21 23:00:40 +00002512/* Collapse packets in the retransmit queue to make to create
2513 * less packets on the wire. This is only done on retransmission.
2514 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002515static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2516 int space)
2517{
2518 struct tcp_sock *tp = tcp_sk(sk);
2519 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002520 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002521
2522 if (!sysctl_tcp_retrans_collapse)
2523 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002524 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002525 return;
2526
2527 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2528 if (!tcp_can_collapse(sk, skb))
2529 break;
2530
2531 space -= skb->len;
2532
2533 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002534 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002535 continue;
2536 }
2537
2538 if (space < 0)
2539 break;
2540 /* Punt if not enough space exists in the first SKB for
2541 * the data in the second
2542 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002543 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002544 break;
2545
2546 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2547 break;
2548
2549 tcp_collapse_retrans(sk, to);
2550 }
2551}
2552
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553/* This retransmits one SKB. Policy decisions and retransmit queue
2554 * state updates are done by the caller. Returns non-zero if an
2555 * error occurred which prevented the send.
2556 */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002557int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558{
2559 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002560 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002561 unsigned int cur_mss;
Yuchung Chengc84a5712014-02-28 16:42:26 -08002562 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563
John Heffner5d424d52006-03-20 17:53:41 -08002564 /* Inconslusive MTU probe */
2565 if (icsk->icsk_mtup.probe_size) {
2566 icsk->icsk_mtup.probe_size = 0;
2567 }
2568
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002570 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 */
2572 if (atomic_read(&sk->sk_wmem_alloc) >
2573 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2574 return -EAGAIN;
2575
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002576 if (skb_still_in_host_queue(sk, skb))
2577 return -EBUSY;
2578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2580 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2581 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2583 return -ENOMEM;
2584 }
2585
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002586 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2587 return -EHOSTUNREACH; /* Routing failure or similar. */
2588
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002589 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002590
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 /* If receiver has shrunk his window, and skb is out of
2592 * new window, do not retransmit it. The exception is the
2593 * case, when window is shrunk to zero. In this case
2594 * our retransmit serves as a zero window probe.
2595 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002596 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2597 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 return -EAGAIN;
2599
2600 if (skb->len > cur_mss) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002601 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002603 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002604 int oldpcount = tcp_skb_pcount(skb);
2605
2606 if (unlikely(oldpcount > 1)) {
Eric Dumazetc52e24212013-10-15 11:54:30 -07002607 if (skb_unclone(skb, GFP_ATOMIC))
2608 return -ENOMEM;
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002609 tcp_init_tso_segs(skb, cur_mss);
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002610 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2611 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 }
2613
Daniel Borkmann49213552015-05-19 21:04:22 +02002614 /* RFC3168, section 6.1.1.1. ECN fallback */
2615 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2616 tcp_ecn_clear_syn(sk, skb);
2617
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002618 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 /* Make a copy, if the first transmission SKB clone we made
2621 * is still in somebody's hands, else make a clone.
2622 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
Thomas Graf50bceae2013-04-11 10:57:18 +00002624 /* make sure skb->data is aligned on arches that require it
2625 * and check if ack-trimming & collapsing extended the headroom
2626 * beyond what csum_start can cover.
2627 */
2628 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2629 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet117632e2011-12-03 21:39:53 +00002630 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2631 GFP_ATOMIC);
Yuchung Chengc84a5712014-02-28 16:42:26 -08002632 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2633 -ENOBUFS;
Eric Dumazet117632e2011-12-03 21:39:53 +00002634 } else {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002635 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002636 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002637
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002638 if (likely(!err)) {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002639 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002640 /* Update global TCP statistics. */
2641 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2642 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2643 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2644 tp->total_retrans++;
2645 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002646 return err;
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002647}
2648
2649int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2650{
2651 struct tcp_sock *tp = tcp_sk(sk);
2652 int err = __tcp_retransmit_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
2654 if (err == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002656 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002657 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 }
2659#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002660 if (!tp->retrans_out)
2661 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2663 tp->retrans_out += tcp_skb_pcount(skb);
2664
2665 /* Save stamp of the first retransmit. */
2666 if (!tp->retrans_stamp)
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002667 tp->retrans_stamp = tcp_skb_timestamp(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 /* snd_nxt is stored to detect loss of retransmitted segment,
2670 * see tcp_input.c tcp_sacktag_write_queue().
2671 */
2672 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002673 } else if (err != -EBUSY) {
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002674 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 }
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002676
2677 if (tp->undo_retrans < 0)
2678 tp->undo_retrans = 0;
2679 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 return err;
2681}
2682
Andi Kleen67edfef2009-07-21 23:00:40 +00002683/* Check if we forward retransmits are possible in the current
2684 * window/congestion state.
2685 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002686static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002687{
2688 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002689 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002690
2691 /* Forward retransmissions are possible only during Recovery. */
2692 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002693 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002694
2695 /* No forward retransmissions in Reno are possible. */
2696 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002697 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002698
2699 /* Yeah, we have to make difficult choice between forward transmission
2700 * and retransmission... Both ways have their merits...
2701 *
2702 * For now we do not retransmit anything, while we have some new
2703 * segments to send. In the other cases, follow rule 3 for
2704 * NextSeg() specified in RFC3517.
2705 */
2706
2707 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002708 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002709
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002710 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002711}
2712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713/* This gets called after a retransmit timeout, and the initially
2714 * retransmitted data is acknowledged. It tries to continue
2715 * resending the rest of the retransmit queue, until either
2716 * we've sent it all or the congestion window limit is reached.
2717 * If doing SACK, the first ACK which comes back for a timeout
2718 * based retransmit packet might feed us FACK information again.
2719 * If so, we use it to avoid unnecessarily retransmissions.
2720 */
2721void tcp_xmit_retransmit_queue(struct sock *sk)
2722{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002723 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 struct tcp_sock *tp = tcp_sk(sk);
2725 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002726 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002727 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002728 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002729 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002730
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002731 if (!tp->packets_out)
2732 return;
2733
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002734 if (!tp->lost_out)
2735 tp->retransmit_high = tp->snd_una;
2736
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002737 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002738 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002739 last_lost = TCP_SKB_CB(skb)->end_seq;
2740 if (after(last_lost, tp->retransmit_high))
2741 last_lost = tp->retransmit_high;
2742 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002743 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002744 last_lost = tp->snd_una;
2745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002747 tcp_for_write_queue_from(skb, sk) {
2748 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002750 if (skb == tcp_send_head(sk))
2751 break;
2752 /* we could do better than to assign each time */
Ian Morris51456b22015-04-03 09:17:26 +01002753 if (!hole)
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002754 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002755
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002756 /* Assume this retransmit will generate
2757 * only one packet for congestion window
2758 * calculation purposes. This works because
2759 * tcp_retransmit_skb() will chop up the
2760 * packet to be MSS sized and all the
2761 * packet counting works out.
2762 */
2763 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2764 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002765
2766 if (fwd_rexmitting) {
2767begin_fwd:
2768 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2769 break;
2770 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2771
2772 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002773 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002774 if (!tcp_can_forward_retransmit(sk))
2775 break;
2776 /* Backtrack if necessary to non-L'ed skb */
Ian Morris00db4122015-04-03 09:17:27 +01002777 if (hole) {
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002778 skb = hole;
2779 hole = NULL;
2780 }
2781 fwd_rexmitting = 1;
2782 goto begin_fwd;
2783
2784 } else if (!(sacked & TCPCB_LOST)) {
Ian Morris51456b22015-04-03 09:17:26 +01002785 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002786 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002787 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002789 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002790 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002791 if (icsk->icsk_ca_state != TCP_CA_Loss)
2792 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2793 else
2794 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2795 }
2796
2797 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002798 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002799
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002800 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002801 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002802
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002803 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2804
Yuchung Cheng684bad12012-09-02 17:38:04 +00002805 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002806 tp->prr_out += tcp_skb_pcount(skb);
2807
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002808 if (skb == tcp_write_queue_head(sk))
2809 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2810 inet_csk(sk)->icsk_rto,
2811 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813}
2814
Eric Dumazetd83769a2015-04-21 18:32:24 -07002815/* We allow to exceed memory limits for FIN packets to expedite
2816 * connection tear down and (memory) recovery.
Eric Dumazet845704a2015-04-23 10:42:39 -07002817 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2818 * or even be forced to close flow without any FIN.
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002819 * In general, we want to allow one skb per socket to avoid hangs
2820 * with edge trigger epoll()
Eric Dumazetd83769a2015-04-21 18:32:24 -07002821 */
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002822void sk_forced_mem_schedule(struct sock *sk, int size)
Eric Dumazetd83769a2015-04-21 18:32:24 -07002823{
2824 int amt, status;
2825
2826 if (size <= sk->sk_forward_alloc)
2827 return;
2828 amt = sk_mem_pages(size);
2829 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2830 sk_memory_allocated_add(sk, amt, &status);
2831}
2832
Eric Dumazet845704a2015-04-23 10:42:39 -07002833/* Send a FIN. The caller locks the socket for us.
2834 * We should try to send a FIN packet really hard, but eventually give up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 */
2836void tcp_send_fin(struct sock *sk)
2837{
Eric Dumazet845704a2015-04-23 10:42:39 -07002838 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002839 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002840
Eric Dumazet845704a2015-04-23 10:42:39 -07002841 /* Optimization, tack on the FIN if we have one skb in write queue and
2842 * this skb was not yet sent, or we are under memory pressure.
2843 * Note: in the latter case, FIN packet will be sent after a timeout,
2844 * as TCP stack thinks it has already been transmitted.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002846 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
Eric Dumazet845704a2015-04-23 10:42:39 -07002847coalesce:
2848 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2849 TCP_SKB_CB(tskb)->end_seq++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 tp->write_seq++;
Eric Dumazet845704a2015-04-23 10:42:39 -07002851 if (!tcp_send_head(sk)) {
2852 /* This means tskb was already sent.
2853 * Pretend we included the FIN on previous transmit.
2854 * We need to set tp->snd_nxt to the value it would have
2855 * if FIN had been sent. This is because retransmit path
2856 * does not change tp->snd_nxt.
2857 */
2858 tp->snd_nxt++;
2859 return;
2860 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 } else {
Eric Dumazet845704a2015-04-23 10:42:39 -07002862 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2863 if (unlikely(!skb)) {
2864 if (tskb)
2865 goto coalesce;
2866 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 }
Eric Dumazetd83769a2015-04-21 18:32:24 -07002868 skb_reserve(skb, MAX_TCP_HEADER);
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002869 sk_forced_mem_schedule(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002871 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002872 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 tcp_queue_skb(sk, skb);
2874 }
Eric Dumazet845704a2015-04-23 10:42:39 -07002875 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876}
2877
2878/* We get here when a process closes a file descriptor (either due to
2879 * an explicit close() or as a byproduct of exit()'ing) and there
2880 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002881 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 */
Al Virodd0fc662005-10-07 07:46:04 +01002883void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 struct sk_buff *skb;
2886
2887 /* NOTE: No TCP options attached and we never retransmit this. */
2888 skb = alloc_skb(MAX_TCP_HEADER, priority);
2889 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002890 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 return;
2892 }
2893
2894 /* Reserve space for headers and prepare control bits. */
2895 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002896 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002897 TCPHDR_ACK | TCPHDR_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 /* Send it off. */
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002899 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002900 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002901
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002902 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903}
2904
Andi Kleen67edfef2009-07-21 23:00:40 +00002905/* Send a crossed SYN-ACK during socket establishment.
2906 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 * a SYN packet that crossed the incoming SYN that caused this routine
2908 * to get called. If this assumption fails then the initial rcv_wnd
2909 * and rcv_wscale values will not be correct.
2910 */
2911int tcp_send_synack(struct sock *sk)
2912{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002913 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914
David S. Millerfe067e82007-03-07 12:12:44 -08002915 skb = tcp_write_queue_head(sk);
Ian Morris51456b22015-04-03 09:17:26 +01002916 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00002917 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 return -EFAULT;
2919 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04002920 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 if (skb_cloned(skb)) {
2922 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
Ian Morris51456b22015-04-03 09:17:26 +01002923 if (!nskb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002925 tcp_unlink_write_queue(skb, sk);
Eric Dumazetf4a775d2014-09-22 16:29:32 -07002926 __skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002927 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002928 sk_wmem_free_skb(sk, skb);
2929 sk->sk_wmem_queued += nskb->truesize;
2930 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 skb = nskb;
2932 }
2933
Eric Dumazet4de075e2011-09-27 13:25:05 -04002934 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Florian Westphal735d3832014-09-29 13:08:30 +02002935 tcp_ecn_send_synack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002937 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938}
2939
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002940/**
2941 * tcp_make_synack - Prepare a SYN-ACK.
2942 * sk: listener socket
2943 * dst: dst entry attached to the SYNACK
2944 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002945 *
2946 * Allocate one skb and build a SYNACK packet.
2947 * @dst is consumed : Caller should not use it again.
2948 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002949struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002950 struct request_sock *req,
Jerry Chu83368862012-08-31 12:29:12 +00002951 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002952{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002953 struct tcp_out_options opts;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002954 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 struct tcp_sock *tp = tcp_sk(sk);
2956 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957 struct sk_buff *skb;
Eric Dumazet80f03e22015-03-24 15:58:52 -07002958 struct tcp_md5sig_key *md5 = NULL;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002959 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002960 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961
Eric Dumazeta0b84862014-03-26 09:57:19 -07002962 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002963 if (unlikely(!skb)) {
2964 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002966 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 /* Reserve space for headers. */
2968 skb_reserve(skb, MAX_TCP_HEADER);
2969
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002970 skb_dst_set(skb, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971
David S. Miller0dbaee32010-12-13 12:52:14 -08002972 mss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002973 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2974 mss = tp->rx_opt.user_mss;
2975
Adam Langley33ad7982008-07-19 00:04:31 -07002976 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002977#ifdef CONFIG_SYN_COOKIES
2978 if (unlikely(req->cookie_ts))
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002979 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002980 else
2981#endif
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002982 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazet80f03e22015-03-24 15:58:52 -07002983
2984#ifdef CONFIG_TCP_MD5SIG
2985 rcu_read_lock();
Eric Dumazetfd3a1542015-03-24 15:58:56 -07002986 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
Eric Dumazet80f03e22015-03-24 15:58:52 -07002987#endif
2988 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00002989 foc) + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002990
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002991 skb_push(skb, tcp_header_size);
2992 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002994 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 memset(th, 0, sizeof(struct tcphdr));
2996 th->syn = 1;
2997 th->ack = 1;
Florian Westphal735d3832014-09-29 13:08:30 +02002998 tcp_ecn_make_synack(req, th, sk);
Eric Dumazetb44084c2013-10-10 00:04:37 -07002999 th->source = htons(ireq->ir_num);
Eric Dumazet634fb9792013-10-09 15:21:29 -07003000 th->dest = ireq->ir_rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08003001 /* Setting of flags are superfluous here for callers (and ECE is
3002 * not even correctly set)
3003 */
3004 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00003005 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faade2009-12-02 18:25:27 +00003006
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Jerry Chu83368862012-08-31 12:29:12 +00003008 /* XXX data is queued and acked as is. No buffer/window check */
3009 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010
3011 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08003012 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003013 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 th->doff = (tcp_header_size >> 2);
Eric Dumazeta0b84862014-03-26 09:57:19 -07003015 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003016
3017#ifdef CONFIG_TCP_MD5SIG
3018 /* Okay, we have all we need - do the md5 hash if needed */
Eric Dumazet80f03e22015-03-24 15:58:52 -07003019 if (md5)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003020 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07003021 md5, req_to_sk(req), skb);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003022 rcu_read_unlock();
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003023#endif
3024
Eric Dumazetb50edd72015-04-09 13:31:56 -07003025 /* Do not fool tcpdump (if any), clean our debris */
3026 skb->tstamp.tv64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 return skb;
3028}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003029EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030
Daniel Borkmann81164412015-01-05 23:57:48 +01003031static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3032{
3033 struct inet_connection_sock *icsk = inet_csk(sk);
3034 const struct tcp_congestion_ops *ca;
3035 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3036
3037 if (ca_key == TCP_CA_UNSPEC)
3038 return;
3039
3040 rcu_read_lock();
3041 ca = tcp_ca_find_key(ca_key);
3042 if (likely(ca && try_module_get(ca->owner))) {
3043 module_put(icsk->icsk_ca_ops->owner);
3044 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3045 icsk->icsk_ca_ops = ca;
3046 }
3047 rcu_read_unlock();
3048}
3049
Andi Kleen67edfef2009-07-21 23:00:40 +00003050/* Do all connect socket setups that can be done AF independent. */
stephen hemmingerf7e56a72013-12-29 11:39:51 -08003051static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003053 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 struct tcp_sock *tp = tcp_sk(sk);
3055 __u8 rcv_wscale;
3056
3057 /* We'll fix this up when we get a response from the other end.
3058 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3059 */
3060 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08003061 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003063#ifdef CONFIG_TCP_MD5SIG
Ian Morris00db4122015-04-03 09:17:27 +01003064 if (tp->af_specific->md5_lookup(sk, sk))
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003065 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3066#endif
3067
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 /* If user gave his TCP_MAXSEG, record it to clamp */
3069 if (tp->rx_opt.user_mss)
3070 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3071 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08003072 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 tcp_sync_mss(sk, dst_mtu(dst));
3074
Daniel Borkmann81164412015-01-05 23:57:48 +01003075 tcp_ca_dst_init(sk, dst);
3076
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 if (!tp->window_clamp)
3078 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08003079 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003080 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3081 tp->advmss = tp->rx_opt.user_mss;
3082
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00003085 /* limit the window selection if the user enforce a smaller rx buffer */
3086 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3087 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3088 tp->window_clamp = tcp_full_space(sk);
3089
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 tcp_select_initial_window(tcp_full_space(sk),
3091 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3092 &tp->rcv_wnd,
3093 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08003094 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00003095 &rcv_wscale,
3096 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003097
3098 tp->rx_opt.rcv_wscale = rcv_wscale;
3099 tp->rcv_ssthresh = tp->rcv_wnd;
3100
3101 sk->sk_err = 0;
3102 sock_reset_flag(sk, SOCK_DONE);
3103 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003104 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 tp->snd_una = tp->write_seq;
3106 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003107 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00003108 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003109
3110 if (likely(!tp->repair))
3111 tp->rcv_nxt = 0;
Andrew Vaginc7781a62013-08-27 12:20:40 +04003112 else
3113 tp->rcv_tstamp = tcp_time_stamp;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003114 tp->rcv_wup = tp->rcv_nxt;
3115 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003117 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3118 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 tcp_clear_retrans(tp);
3120}
3121
Yuchung Cheng783237e2012-07-19 06:43:07 +00003122static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3123{
3124 struct tcp_sock *tp = tcp_sk(sk);
3125 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3126
3127 tcb->end_seq += skb->len;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07003128 __skb_header_release(skb);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003129 __tcp_add_write_queue_tail(sk, skb);
3130 sk->sk_wmem_queued += skb->truesize;
3131 sk_mem_charge(sk, skb->truesize);
3132 tp->write_seq = tcb->end_seq;
3133 tp->packets_out += tcp_skb_pcount(skb);
3134}
3135
3136/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3137 * queue a data-only packet after the regular SYN, such that regular SYNs
3138 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3139 * only the SYN sequence, the data are retransmitted in the first ACK.
3140 * If cookie is not cached or other error occurs, falls back to send a
3141 * regular SYN with Fast Open cookie request option.
3142 */
3143static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3144{
3145 struct tcp_sock *tp = tcp_sk(sk);
3146 struct tcp_fastopen_request *fo = tp->fastopen_req;
Al Viro57be5bd2014-11-28 13:40:20 -05003147 int syn_loss = 0, space, err = 0, copied;
Yuchung Chengaab48742012-07-19 06:43:10 +00003148 unsigned long last_syn_loss = 0;
Eric Dumazet355a9012014-11-17 23:06:20 -08003149 struct sk_buff *syn_data;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003150
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003151 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00003152 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3153 &syn_loss, &last_syn_loss);
3154 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3155 if (syn_loss > 1 &&
3156 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3157 fo->cookie.len = -1;
3158 goto fallback;
3159 }
3160
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003161 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3162 fo->cookie.len = -1;
3163 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003164 goto fallback;
3165
3166 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3167 * user-MSS. Reserve maximum option space for middleboxes that add
3168 * private TCP options. The cost is reduced data space in SYN :(
3169 */
3170 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3171 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00003172 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00003173 MAX_TCP_OPTION_SPACE;
3174
Eric Dumazetf5ddcbb2014-02-20 10:09:18 -08003175 space = min_t(size_t, space, fo->size);
3176
3177 /* limit to order-0 allocations */
3178 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3179
Eric Dumazeteb934472015-05-19 13:26:55 -07003180 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
Eric Dumazet355a9012014-11-17 23:06:20 -08003181 if (!syn_data)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003182 goto fallback;
Eric Dumazet355a9012014-11-17 23:06:20 -08003183 syn_data->ip_summed = CHECKSUM_PARTIAL;
3184 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
Al Viro57be5bd2014-11-28 13:40:20 -05003185 copied = copy_from_iter(skb_put(syn_data, space), space,
3186 &fo->data->msg_iter);
3187 if (unlikely(!copied)) {
Eric Dumazet355a9012014-11-17 23:06:20 -08003188 kfree_skb(syn_data);
3189 goto fallback;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003190 }
Al Viro57be5bd2014-11-28 13:40:20 -05003191 if (copied != space) {
3192 skb_trim(syn_data, copied);
3193 space = copied;
3194 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003195
Eric Dumazet355a9012014-11-17 23:06:20 -08003196 /* No more data pending in inet_wait_for_connect() */
3197 if (space == fo->size)
3198 fo->data = NULL;
3199 fo->copied = space;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003200
Eric Dumazet355a9012014-11-17 23:06:20 -08003201 tcp_connect_queue_skb(sk, syn_data);
3202
3203 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3204
3205 syn->skb_mstamp = syn_data->skb_mstamp;
3206
3207 /* Now full SYN+DATA was cloned and sent (or not),
3208 * remove the SYN from the original skb (syn_data)
3209 * we keep in write queue in case of a retransmit, as we
3210 * also have the SYN packet (with no data) in the same queue.
Eric Dumazet431a9122014-03-09 17:36:02 -07003211 */
Eric Dumazet355a9012014-11-17 23:06:20 -08003212 TCP_SKB_CB(syn_data)->seq++;
3213 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3214 if (!err) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003215 tp->syn_data = (fo->copied > 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08003216 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003217 goto done;
3218 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003219
3220fallback:
3221 /* Send a regular SYN with Fast Open cookie request option */
3222 if (fo->cookie.len > 0)
3223 fo->cookie.len = 0;
3224 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3225 if (err)
3226 tp->syn_fastopen = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003227done:
3228 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3229 return err;
3230}
3231
Andi Kleen67edfef2009-07-21 23:00:40 +00003232/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233int tcp_connect(struct sock *sk)
3234{
3235 struct tcp_sock *tp = tcp_sk(sk);
3236 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00003237 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238
3239 tcp_connect_init(sk);
3240
Andrey Vagin2b916472012-11-22 01:13:58 +00003241 if (unlikely(tp->repair)) {
3242 tcp_finish_connect(sk, NULL);
3243 return 0;
3244 }
3245
Eric Dumazeteb934472015-05-19 13:26:55 -07003246 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
Eric Dumazet355a9012014-11-17 23:06:20 -08003247 if (unlikely(!buff))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003248 return -ENOBUFS;
3249
Changli Gaoa3433f32010-06-12 14:01:43 +00003250 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003251 tp->retrans_stamp = tcp_time_stamp;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003252 tcp_connect_queue_skb(sk, buff);
Florian Westphal735d3832014-09-29 13:08:30 +02003253 tcp_ecn_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254
Yuchung Cheng783237e2012-07-19 06:43:07 +00003255 /* Send off SYN; include data in Fast Open. */
3256 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3257 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00003258 if (err == -ECONNREFUSED)
3259 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07003260
3261 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3262 * in order to make this packet get counted in tcpOutSegs.
3263 */
3264 tp->snd_nxt = tp->write_seq;
3265 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003266 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003267
3268 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003269 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3270 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 return 0;
3272}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003273EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003274
3275/* Send out a delayed ack, the caller does the policy checking
3276 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3277 * for details.
3278 */
3279void tcp_send_delayed_ack(struct sock *sk)
3280{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003281 struct inet_connection_sock *icsk = inet_csk(sk);
3282 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283 unsigned long timeout;
3284
Florian Westphal98900922014-09-26 22:37:35 +02003285 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3286
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003288 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003289 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003291 if (icsk->icsk_ack.pingpong ||
3292 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 max_ato = TCP_DELACK_MAX;
3294
3295 /* Slow path, intersegment interval is "high". */
3296
3297 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003298 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299 * directly.
3300 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003301 if (tp->srtt_us) {
3302 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3303 TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304
3305 if (rtt < max_ato)
3306 max_ato = rtt;
3307 }
3308
3309 ato = min(ato, max_ato);
3310 }
3311
3312 /* Stay within the limit we were given */
3313 timeout = jiffies + ato;
3314
3315 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003316 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 /* If delack timer was blocked or is about to expire,
3318 * send ACK now.
3319 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003320 if (icsk->icsk_ack.blocked ||
3321 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 tcp_send_ack(sk);
3323 return;
3324 }
3325
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003326 if (!time_before(timeout, icsk->icsk_ack.timeout))
3327 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003329 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3330 icsk->icsk_ack.timeout = timeout;
3331 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332}
3333
3334/* This routine sends an ack and also updates the window. */
3335void tcp_send_ack(struct sock *sk)
3336{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003337 struct sk_buff *buff;
3338
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003340 if (sk->sk_state == TCP_CLOSE)
3341 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342
Florian Westphal98900922014-09-26 22:37:35 +02003343 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3344
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003345 /* We are not putting this on the write queue, so
3346 * tcp_transmit_skb() will set the ownership to this
3347 * sock.
3348 */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003349 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ian Morris51456b22015-04-03 09:17:26 +01003350 if (!buff) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003351 inet_csk_schedule_ack(sk);
3352 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3353 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3354 TCP_DELACK_MAX, TCP_RTO_MAX);
3355 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003356 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003357
3358 /* Reserve space for headers and prepare control bits. */
3359 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003360 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003361
Eric Dumazet98781962015-02-03 18:31:53 -08003362 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3363 * too much.
3364 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3365 * We also avoid tcp_wfree() overhead (cache line miss accessing
3366 * tp->tsq_flags) by using regular sock_wfree()
3367 */
3368 skb_set_tcp_pure_ack(buff);
3369
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003370 /* Send it off, this clears delayed acks for us. */
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003371 skb_mstamp_get(&buff->skb_mstamp);
Mel Gorman99a1dec2012-07-31 16:44:14 -07003372 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373}
Daniel Borkmanne3118e82014-09-26 22:37:36 +02003374EXPORT_SYMBOL_GPL(tcp_send_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375
3376/* This routine sends a packet with an out of date sequence
3377 * number. It assumes the other end will try to ack it.
3378 *
3379 * Question: what should we make while urgent mode?
3380 * 4.4BSD forces sending single byte of data. We cannot send
3381 * out of window data, because we have SND.NXT==SND.MAX...
3382 *
3383 * Current solution: to send TWO zero-length segments in urgent mode:
3384 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3385 * out-of-date with SND.UNA-1 to probe window.
3386 */
Eric Dumazete520af42015-05-06 14:26:25 -07003387static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388{
3389 struct tcp_sock *tp = tcp_sk(sk);
3390 struct sk_buff *skb;
3391
3392 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003393 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ian Morris51456b22015-04-03 09:17:26 +01003394 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395 return -1;
3396
3397 /* Reserve space for headers and set control bits. */
3398 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 /* Use a previous sequence. This should cause the other
3400 * end to send an ack. Don't queue or clone SKB, just
3401 * send it.
3402 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003403 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003404 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazete520af42015-05-06 14:26:25 -07003405 NET_INC_STATS_BH(sock_net(sk), mib);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003406 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407}
3408
Pavel Emelyanovee995282012-04-19 03:40:39 +00003409void tcp_send_window_probe(struct sock *sk)
3410{
3411 if (sk->sk_state == TCP_ESTABLISHED) {
3412 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
Eric Dumazete520af42015-05-06 14:26:25 -07003413 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
Pavel Emelyanovee995282012-04-19 03:40:39 +00003414 }
3415}
3416
Andi Kleen67edfef2009-07-21 23:00:40 +00003417/* Initiate keepalive or window probe from timer. */
Eric Dumazete520af42015-05-06 14:26:25 -07003418int tcp_write_wakeup(struct sock *sk, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003420 struct tcp_sock *tp = tcp_sk(sk);
3421 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003423 if (sk->sk_state == TCP_CLOSE)
3424 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
Ian Morris00db4122015-04-03 09:17:27 +01003426 skb = tcp_send_head(sk);
3427 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003428 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003429 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003430 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003432 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3433 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003435 /* We are probing the opening of a window
3436 * but the window size is != 0
3437 * must have been a result SWS avoidance ( sender )
3438 */
3439 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3440 skb->len > mss) {
3441 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003442 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Octavian Purdila6cc55e02014-06-06 17:32:37 +03003443 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003444 return -1;
3445 } else if (!tcp_skb_pcount(skb))
Eric Dumazet5bbb4322015-06-11 09:15:17 -07003446 tcp_set_skb_tso_segs(skb, mss);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003447
Eric Dumazet4de075e2011-09-27 13:25:05 -04003448 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003449 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3450 if (!err)
3451 tcp_event_new_data_sent(sk, skb);
3452 return err;
3453 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003454 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Eric Dumazete520af42015-05-06 14:26:25 -07003455 tcp_xmit_probe_skb(sk, 1, mib);
3456 return tcp_xmit_probe_skb(sk, 0, mib);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458}
3459
3460/* A window probe timeout has occurred. If window is not closed send
3461 * a partial packet else a zero probe.
3462 */
3463void tcp_send_probe0(struct sock *sk)
3464{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003465 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003467 unsigned long probe_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 int err;
3469
Eric Dumazete520af42015-05-06 14:26:25 -07003470 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471
David S. Millerfe067e82007-03-07 12:12:44 -08003472 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003474 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003475 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 return;
3477 }
3478
3479 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003480 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3481 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003482 icsk->icsk_probes_out++;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003483 probe_max = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 } else {
3485 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003486 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487 * Let local senders to fight for local resources.
3488 *
3489 * Use accumulated backoff yet.
3490 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003491 if (!icsk->icsk_probes_out)
3492 icsk->icsk_probes_out = 1;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003493 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 }
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003495 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003496 tcp_probe0_when(sk, probe_max),
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003497 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498}
Octavian Purdila5db92c92014-06-25 17:09:59 +03003499
3500int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3501{
3502 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3503 struct flowi fl;
3504 int res;
3505
3506 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3507 if (!res) {
3508 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3509 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3510 }
3511 return res;
3512}
3513EXPORT_SYMBOL(tcp_rtx_synack);