blob: 7d2dc015cd19a64505c968df43c544adabb65e0e [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 Dumazet46d3cea2012-07-11 05:50:31 +000065static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000067
Andi Kleen67edfef2009-07-21 23:00:40 +000068/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040069static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070070{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000071 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070072 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080073 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070074
David S. Millerfe067e82007-03-07 12:12:44 -080075 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080077
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080078 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000079 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000080 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000081 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000082 }
Yuchung Chengf19c29e2014-03-03 12:31:36 -080083
David S. Millerf7324ac2014-03-06 15:03:17 -050084 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -070086}
87
88/* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040094static inline __u32 tcp_acceptable_seq(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070095{
Eric Dumazetcf533ea2011-10-21 05:22:42 -040096 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070097
Ilpo Järvinen90840de2007-12-31 04:48:41 -080098 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 return tp->snd_nxt;
100 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800101 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102}
103
104/* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 *
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
111 * large MSS.
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
117 */
118static __u16 tcp_advertise_mss(struct sock *sk)
119{
120 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400121 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 int mss = tp->advmss;
123
David S. Miller0dbaee32010-12-13 12:52:14 -0800124 if (dst) {
125 unsigned int metric = dst_metric_advmss(dst);
126
127 if (metric < mss) {
128 mss = metric;
129 tp->advmss = mss;
130 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 }
132
133 return (__u16)mss;
134}
135
136/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
Eric Dumazet6f021c62015-08-21 12:30:00 -0700137 * This is the first part of cwnd validation mechanism.
138 */
139void tcp_cwnd_restart(struct sock *sk, s32 delta)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f021c62015-08-21 12:30:00 -0700142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 u32 cwnd = tp->snd_cwnd;
144
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 restart_cwnd = min(restart_cwnd, cwnd);
149
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 cwnd >>= 1;
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
155}
156
Andi Kleen67edfef2009-07-21 23:00:40 +0000157/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800158static void tcp_event_data_sent(struct tcp_sock *tp,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400159 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Neal Cardwell05c5a462015-09-09 21:54:37 -0700164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
166
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
Jon Maxwell2251ae42015-07-08 10:12:28 +1000172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174}
175
Andi Kleen67edfef2009-07-21 23:00:40 +0000176/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181}
182
Yuchung Cheng85f16522013-06-11 15:35:32 -0700183
184u32 tcp_default_init_rwnd(u32 mss)
185{
186 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800187 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
190 */
191 u32 init_rwnd = TCP_INIT_CWND * 2;
192
193 if (mss > 1460)
194 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195 return init_rwnd;
196}
197
Linus Torvalds1da177e2005-04-16 15:20:36 -0700198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (65535 << 14);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = (space / mss) * mss;
220
221 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 */
Rick Jones15d99e02006-03-20 22:40:29 -0800229 if (sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = space;
233
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 (*rcv_wscale) = 0;
235 if (wscale_ok) {
236 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900237 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 */
239 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700240 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 while (space > 65535 && (*rcv_wscale) < 14) {
242 space >>= 1;
243 (*rcv_wscale)++;
244 }
245 }
246
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800247 if (mss > (1 << *rcv_wscale)) {
Yuchung Cheng85f16522013-06-11 15:35:32 -0700248 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249 init_rcv_wnd = tcp_default_init_rwnd(mss);
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 }
252
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
255}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000256EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257
258/* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
261 * frame.
262 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800263static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264{
265 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphal8e165e22014-02-25 14:34:32 +0100266 u32 old_win = tp->rcv_wnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
269
270 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800271 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
276 *
277 * Relax Will Robinson.
278 */
Florian Westphal8e165e22014-02-25 14:34:32 +0100279 if (new_win == 0)
280 NET_INC_STATS(sock_net(sk),
281 LINUX_MIB_TCPWANTZEROWINDOWADV);
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 }
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
286
287 /* Make sure we do not exceed the maximum possible
288 * scaled window.
289 */
Rick Jones15d99e02006-03-20 22:40:29 -0800290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 new_win = min(new_win, MAX_TCP_WINDOW);
292 else
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
294
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
297
298 /* If we advertise zero window, disable fast path. */
Florian Westphal8e165e22014-02-25 14:34:32 +0100299 if (new_win == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 tp->pred_flags = 0;
Florian Westphal8e165e22014-02-25 14:34:32 +0100301 if (old_win)
302 NET_INC_STATS(sock_net(sk),
303 LINUX_MIB_TCPTOZEROWINDOWADV);
304 } else if (old_win == 0) {
305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
306 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307
308 return new_win;
309}
310
Andi Kleen67edfef2009-07-21 23:00:40 +0000311/* Packet ECN state for a SYN-ACK */
Florian Westphal735d3832014-09-29 13:08:30 +0200312static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700313{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200314 const struct tcp_sock *tp = tcp_sk(sk);
315
Eric Dumazet4de075e2011-09-27 13:25:05 -0400316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800317 if (!(tp->ecn_flags & TCP_ECN_OK))
Eric Dumazet4de075e2011-09-27 13:25:05 -0400318 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200319 else if (tcp_ca_needs_ecn(sk))
320 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700321}
322
Andi Kleen67edfef2009-07-21 23:00:40 +0000323/* Packet ECN state for a SYN. */
Florian Westphal735d3832014-09-29 13:08:30 +0200324static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700325{
326 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk);
329
330 if (!use_ecn) {
331 const struct dst_entry *dst = __sk_dst_get(sk);
332
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
334 use_ecn = true;
335 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700336
337 tp->ecn_flags = 0;
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100338
339 if (use_ecn) {
Eric Dumazet4de075e2011-09-27 13:25:05 -0400340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700341 tp->ecn_flags = TCP_ECN_OK;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200342 if (tcp_ca_needs_ecn(sk))
343 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700344 }
345}
346
Daniel Borkmann49213552015-05-19 21:04:22 +0200347static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348{
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
352 */
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
354}
355
Florian Westphal735d3832014-09-29 13:08:30 +0200356static void
Eric Dumazet6ac705b2015-09-25 07:39:18 -0700357tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700358{
Eric Dumazet6ac705b2015-09-25 07:39:18 -0700359 if (inet_rsk(req)->ecn_ok)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700360 th->ece = 1;
361}
362
Andi Kleen67edfef2009-07-21 23:00:40 +0000363/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
364 * be sent.
365 */
Florian Westphal735d3832014-09-29 13:08:30 +0200366static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700367 int tcp_header_len)
368{
369 struct tcp_sock *tp = tcp_sk(sk);
370
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 tcp_hdr(skb)->cwr = 1;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 }
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200381 } else if (!tcp_ca_needs_ecn(sk)) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
384 }
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386 tcp_hdr(skb)->ece = 1;
387 }
388}
389
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800390/* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
392 */
393static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700395 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800396 skb->csum = 0;
397
Eric Dumazet4de075e2011-09-27 13:25:05 -0400398 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800399 TCP_SKB_CB(skb)->sacked = 0;
400
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700401 tcp_skb_pcount_set(skb, 1);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800402
403 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000404 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800405 seq++;
406 TCP_SKB_CB(skb)->end_seq = seq;
407}
408
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000409static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700410{
411 return tp->snd_una != tp->snd_up;
412}
413
Adam Langley33ad7982008-07-19 00:04:31 -0700414#define OPTION_SACK_ADVERTISE (1 << 0)
415#define OPTION_TS (1 << 1)
416#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000417#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000418#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700419
420struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000425 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000426 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700429};
430
Andi Kleen67edfef2009-07-21 23:00:40 +0000431/* Write previously computed TCP options to the packet.
432 *
433 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
stephen hemminger8e3bff92013-12-08 12:15:44 -0800436 * inter-operability perspective it seems that we're somewhat stuck with
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
440 *
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
443 */
Adam Langley33ad7982008-07-19 00:04:31 -0700444static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000445 struct tcp_out_options *opts)
446{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000447 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000448
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000449 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700454 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800455 }
Adam Langley33ad7982008-07-19 00:04:31 -0700456
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
460 opts->mss);
461 }
462
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
468 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000469 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700470 } else {
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
472 (TCPOPT_NOP << 16) |
473 (TCPOPT_TIMESTAMP << 8) |
474 TCPOLEN_TIMESTAMP);
475 }
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
478 }
479
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_SACK_PERM << 8) |
484 TCPOLEN_SACK_PERM);
485 }
486
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000487 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
491 opts->ws);
492 }
493
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800497 int this_sack;
498
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
500 (TCPOPT_NOP << 16) |
501 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800503 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800504
Adam Langley33ad7982008-07-19 00:04:31 -0700505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
506 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
509 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800510
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000511 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800512 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000513
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
Daniel Lee7f9b8382015-04-06 14:37:26 -0700516 u8 *p = (u8 *)ptr;
517 u32 len; /* Fast Open option length */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000518
Daniel Lee7f9b8382015-04-06 14:37:26 -0700519 if (foc->exp) {
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
524 } else {
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
527 *p++ = len;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000528 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700529
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
534 }
535 ptr += (len + 3) >> 2;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000536 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800537}
538
Andi Kleen67edfef2009-07-21 23:00:40 +0000539/* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
541 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000542static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700543 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400544 struct tcp_md5sig_key **md5)
545{
Adam Langley33ad7982008-07-19 00:04:31 -0700546 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000547 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000548 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700549
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800550#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700551 *md5 = tp->af_specific->md5_lookup(sk, sk);
552 if (*md5) {
553 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000554 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800555 }
Adam Langley33ad7982008-07-19 00:04:31 -0700556#else
557 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800558#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700559
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
568 * going out. */
569 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000570 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700571
Ian Morris51456b22015-04-03 09:17:26 +0100572 if (likely(sysctl_tcp_timestamps && !*md5)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700573 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700574 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700575 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000576 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700577 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800578 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700579 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000580 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000581 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700582 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800583 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700584 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700585 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000586 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700587 }
588
Yuchung Cheng783237e2012-07-19 06:43:07 +0000589 if (fastopen && fastopen->cookie.len >= 0) {
Daniel Lee2646c832015-04-06 14:37:27 -0700590 u32 need = fastopen->cookie.len;
591
592 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593 TCPOLEN_FASTOPEN_BASE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000594 need = (need + 3) & ~3U; /* Align to 32 bits */
595 if (remaining >= need) {
596 opts->options |= OPTION_FAST_OPEN_COOKIE;
597 opts->fastopen_cookie = &fastopen->cookie;
598 remaining -= need;
599 tp->syn_fastopen = 1;
Daniel Lee2646c832015-04-06 14:37:27 -0700600 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000601 }
602 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000603
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000604 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700605}
606
Andi Kleen67edfef2009-07-21 23:00:40 +0000607/* Set up TCP options for SYN-ACKs. */
Eric Dumazet37bfbdd2015-09-25 07:39:17 -0700608static unsigned int tcp_synack_options(struct request_sock *req,
609 unsigned int mss, struct sk_buff *skb,
610 struct tcp_out_options *opts,
611 const struct tcp_md5sig_key *md5,
612 struct tcp_fastopen_cookie *foc)
William Allen Simpson4957faade2009-12-02 18:25:27 +0000613{
Adam Langley33ad7982008-07-19 00:04:31 -0700614 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet95c96172012-04-15 05:58:06 +0000615 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Adam Langley33ad7982008-07-19 00:04:31 -0700616
617#ifdef CONFIG_TCP_MD5SIG
Eric Dumazet80f03e22015-03-24 15:58:52 -0700618 if (md5) {
Adam Langley33ad7982008-07-19 00:04:31 -0700619 opts->options |= OPTION_MD5;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000620 remaining -= TCPOLEN_MD5SIG_ALIGNED;
621
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
626 */
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700627 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700628 }
Adam Langley33ad7982008-07-19 00:04:31 -0700629#endif
630
William Allen Simpson4957faade2009-12-02 18:25:27 +0000631 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700632 opts->mss = mss;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000633 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700634
635 if (likely(ireq->wscale_ok)) {
636 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000637 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000638 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700639 }
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700640 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700641 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700642 opts->tsval = tcp_skb_timestamp(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700643 opts->tsecr = req->ts_recent;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000644 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700645 }
646 if (likely(ireq->sack_ok)) {
647 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700648 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faade2009-12-02 18:25:27 +0000649 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700650 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700651 if (foc != NULL && foc->len >= 0) {
652 u32 need = foc->len;
653
654 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655 TCPOLEN_FASTOPEN_BASE;
Jerry Chu83368862012-08-31 12:29:12 +0000656 need = (need + 3) & ~3U; /* Align to 32 bits */
657 if (remaining >= need) {
658 opts->options |= OPTION_FAST_OPEN_COOKIE;
659 opts->fastopen_cookie = foc;
660 remaining -= need;
661 }
662 }
William Allen Simpson4957faade2009-12-02 18:25:27 +0000663
William Allen Simpson4957faade2009-12-02 18:25:27 +0000664 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700665}
666
Andi Kleen67edfef2009-07-21 23:00:40 +0000667/* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
669 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000670static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700671 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400672 struct tcp_md5sig_key **md5)
673{
Adam Langley33ad7982008-07-19 00:04:31 -0700674 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000675 unsigned int size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000676 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700677
Andi Kleen5843ef42013-09-30 13:29:11 -0700678 opts->options = 0;
679
Adam Langley33ad7982008-07-19 00:04:31 -0700680#ifdef CONFIG_TCP_MD5SIG
681 *md5 = tp->af_specific->md5_lookup(sk, sk);
682 if (unlikely(*md5)) {
683 opts->options |= OPTION_MD5;
684 size += TCPOLEN_MD5SIG_ALIGNED;
685 }
686#else
687 *md5 = NULL;
688#endif
689
690 if (likely(tp->rx_opt.tstamp_ok)) {
691 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700692 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700693 opts->tsecr = tp->rx_opt.ts_recent;
694 size += TCPOLEN_TSTAMP_ALIGNED;
695 }
696
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000697 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698 if (unlikely(eff_sacks)) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000699 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
Adam Langley33ad7982008-07-19 00:04:31 -0700700 opts->num_sack_blocks =
Eric Dumazet95c96172012-04-15 05:58:06 +0000701 min_t(unsigned int, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700702 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703 TCPOLEN_SACK_PERBLOCK);
704 size += TCPOLEN_SACK_BASE_ALIGNED +
705 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
706 }
707
708 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800709}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000711
712/* TCP SMALL QUEUES (TSQ)
713 *
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
717 *
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000721 *
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
725 */
726struct tsq_tasklet {
727 struct tasklet_struct tasklet;
728 struct list_head head; /* queue of tcp sockets */
729};
730static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
731
Eric Dumazet6f458df2012-07-20 05:45:50 +0000732static void tcp_tsq_handler(struct sock *sk)
733{
734 if ((1 << sk->sk_state) &
735 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
John Ognessbf062002014-02-09 18:40:11 -0800737 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
738 0, GFP_ATOMIC);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000739}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000740/*
stephen hemminger8e3bff92013-12-08 12:15:44 -0800741 * One tasklet per cpu tries to send more skbs.
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000742 * We run in tasklet context but need to disable irqs when
stephen hemminger8e3bff92013-12-08 12:15:44 -0800743 * transferring tsq->head because tcp_wfree() might
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000744 * interrupt us (non NAPI drivers)
745 */
746static void tcp_tasklet_func(unsigned long data)
747{
748 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
749 LIST_HEAD(list);
750 unsigned long flags;
751 struct list_head *q, *n;
752 struct tcp_sock *tp;
753 struct sock *sk;
754
755 local_irq_save(flags);
756 list_splice_init(&tsq->head, &list);
757 local_irq_restore(flags);
758
759 list_for_each_safe(q, n, &list) {
760 tp = list_entry(q, struct tcp_sock, tsq_node);
761 list_del(&tp->tsq_node);
762
763 sk = (struct sock *)tp;
764 bh_lock_sock(sk);
765
766 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000767 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000768 } else {
769 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000770 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000771 }
772 bh_unlock_sock(sk);
773
774 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
775 sk_free(sk);
776 }
777}
778
Eric Dumazet6f458df2012-07-20 05:45:50 +0000779#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
780 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200781 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
782 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000783/**
784 * tcp_release_cb - tcp release_sock() callback
785 * @sk: socket
786 *
787 * called from release_sock() to perform protocol dependent
788 * actions before socket release.
789 */
790void tcp_release_cb(struct sock *sk)
791{
792 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000793 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000794
Eric Dumazet6f458df2012-07-20 05:45:50 +0000795 /* perform an atomic operation only if at least one flag is set */
796 do {
797 flags = tp->tsq_flags;
798 if (!(flags & TCP_DEFERRED_ALL))
799 return;
800 nflags = flags & ~TCP_DEFERRED_ALL;
801 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
802
803 if (flags & (1UL << TCP_TSQ_DEFERRED))
804 tcp_tsq_handler(sk);
805
Eric Dumazetc3f9b012014-03-10 09:50:11 -0700806 /* Here begins the tricky part :
807 * We are called from release_sock() with :
808 * 1) BH disabled
809 * 2) sk_lock.slock spinlock held
810 * 3) socket owned by us (sk->sk_lock.owned == 1)
811 *
812 * But following code is meant to be called from BH handlers,
813 * so we should keep BH disabled, but early release socket ownership
814 */
815 sock_release_ownership(sk);
816
Eric Dumazet144d56e2012-08-20 00:22:46 +0000817 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000818 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000819 __sock_put(sk);
820 }
821 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000822 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000823 __sock_put(sk);
824 }
825 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Neal Cardwell4fab9072014-08-14 12:40:05 -0400826 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000827 __sock_put(sk);
828 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000829}
830EXPORT_SYMBOL(tcp_release_cb);
831
832void __init tcp_tasklet_init(void)
833{
834 int i;
835
836 for_each_possible_cpu(i) {
837 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
838
839 INIT_LIST_HEAD(&tsq->head);
840 tasklet_init(&tsq->tasklet,
841 tcp_tasklet_func,
842 (unsigned long)tsq);
843 }
844}
845
846/*
847 * Write buffer destructor automatically called from kfree_skb.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800848 * We can't xmit new skbs from this context, as we might already
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000849 * hold qdisc lock.
850 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000851void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000852{
853 struct sock *sk = skb->sk;
854 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700855 int wmem;
856
857 /* Keep one reference on sk_wmem_alloc.
858 * Will be released by sk_free() from here or tcp_tasklet_func()
859 */
860 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
861
862 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
863 * Wait until our queues (qdisc + devices) are drained.
864 * This gives :
865 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
866 * - chance for incoming ACK (processed by another cpu maybe)
867 * to migrate this flow (skb->ooo_okay will be eventually set)
868 */
869 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
870 goto out;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000871
872 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
873 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
874 unsigned long flags;
875 struct tsq_tasklet *tsq;
876
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000877 /* queue this socket to tasklet queue */
878 local_irq_save(flags);
Christoph Lameter903ceff2014-08-17 12:30:35 -0500879 tsq = this_cpu_ptr(&tsq_tasklet);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000880 list_add(&tp->tsq_node, &tsq->head);
881 tasklet_schedule(&tsq->tasklet);
882 local_irq_restore(flags);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700883 return;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000884 }
Eric Dumazet9b462d02014-10-13 06:27:47 -0700885out:
886 sk_free(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000887}
888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889/* This routine actually transmits TCP packets queued in by
890 * tcp_do_sendmsg(). This is used by both the initial
891 * transmission and possible later retransmissions.
892 * All SKB's seen here are completely headerless. It is our
893 * job to build the TCP header, and pass the packet down to
894 * IP so it can do the same plus pass the packet off to the
895 * device.
896 *
897 * We are working here with either a clone of the original
898 * SKB, or a fresh unique copy made by the retransmit engine.
899 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800900static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
901 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800903 const struct inet_connection_sock *icsk = inet_csk(sk);
904 struct inet_sock *inet;
905 struct tcp_sock *tp;
906 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700907 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000908 unsigned int tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800909 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800910 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800911 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800913 BUG_ON(!skb || !tcp_skb_pcount(skb));
914
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700915 if (clone_it) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800916 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700917
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800918 if (unlikely(skb_cloned(skb)))
919 skb = pskb_copy(skb, gfp_mask);
920 else
921 skb = skb_clone(skb, gfp_mask);
922 if (unlikely(!skb))
923 return -ENOBUFS;
924 }
925
926 inet = inet_sk(sk);
927 tp = tcp_sk(sk);
928 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700929 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930
Eric Dumazet4de075e2011-09-27 13:25:05 -0400931 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700932 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
933 else
934 tcp_options_size = tcp_established_options(sk, skb, &opts,
935 &md5);
936 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900937
Eric Dumazet547669d2013-05-23 07:44:20 +0000938 /* if no packet is in qdisc/device queue, then allow XPS to select
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700939 * another queue. We can be called from tcp_tsq_handler()
940 * which holds one reference to sk_wmem_alloc.
941 *
942 * TODO: Ideally, in-flight pure ACK packets should not matter here.
943 * One way to get this would be to set skb->truesize = 2 on them.
Eric Dumazet547669d2013-05-23 07:44:20 +0000944 */
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700945 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800946
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700947 skb_push(skb, tcp_header_size);
948 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000949
950 skb_orphan(skb);
951 skb->sk = sk;
Eric Dumazet98781962015-02-03 18:31:53 -0800952 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
Tom Herbertb73c3d02014-07-01 21:32:17 -0700953 skb_set_hash_from_sk(skb, sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000954 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800955
956 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700957 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000958 th->source = inet->inet_sport;
959 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800960 th->seq = htonl(tcb->seq);
961 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700962 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400963 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800964
Eric Dumazet4de075e2011-09-27 13:25:05 -0400965 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800966 /* RFC1323: The window in SYN & SYN/ACK segments
967 * is never scaled.
968 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800969 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800970 } else {
971 th->window = htons(tcp_select_window(sk));
972 }
973 th->check = 0;
974 th->urg_ptr = 0;
975
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700976 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800977 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
978 if (before(tp->snd_up, tcb->seq + 0x10000)) {
979 th->urg_ptr = htons(tp->snd_up - tcb->seq);
980 th->urg = 1;
981 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700982 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800983 th->urg = 1;
984 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800985 }
986
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000987 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet51466a72015-06-11 09:15:16 -0700988 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Eric Dumazet4de075e2011-09-27 13:25:05 -0400989 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
Florian Westphal735d3832014-09-29 13:08:30 +0200990 tcp_ecn_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800991
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800992#ifdef CONFIG_TCP_MD5SIG
993 /* Calculate the MD5 hash, as we have all we need now */
994 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -0700995 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000996 tp->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -0700997 md5, sk, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800998 }
999#endif
1000
Herbert Xubb296242010-04-11 02:15:55 +00001001 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001002
Eric Dumazet4de075e2011-09-27 13:25:05 -04001003 if (likely(tcb->tcp_flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001004 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1005
Martin KaFai Laua44d6ea2016-03-14 10:52:15 -07001006 if (skb->len != tcp_header_size) {
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001007 tcp_event_data_sent(tp, sk);
Martin KaFai Laua44d6ea2016-03-14 10:52:15 -07001008 tp->data_segs_out += tcp_skb_pcount(skb);
1009 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001010
Wei Yongjunbd37a082006-08-07 21:04:15 -07001011 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +00001012 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1013 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001014
Marcelo Ricardo Leitner2efd0552015-05-20 16:35:41 -07001015 tp->segs_out += tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001016 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001017 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001018 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001019
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001020 /* Our usage of tstamp should remain private */
1021 skb->tstamp.tv64 = 0;
Eric Dumazet971f10e2014-09-27 09:50:57 -07001022
1023 /* Cleanup our debris for IP stacks */
1024 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1025 sizeof(struct inet6_skb_parm)));
1026
Eric Dumazetb0270e92014-04-15 12:58:34 -04001027 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001028
Hua Zhong83de47c2006-04-28 15:26:50 -07001029 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001030 return err;
1031
Christoph Paasch5ee2c942014-07-14 16:58:32 +02001032 tcp_enter_cwr(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001033
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -02001034 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035}
1036
Andi Kleen67edfef2009-07-21 23:00:40 +00001037/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 *
1039 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1040 * otherwise socket can stall.
1041 */
1042static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1043{
1044 struct tcp_sock *tp = tcp_sk(sk);
1045
1046 /* Advance write_seq and place onto the write_queue. */
1047 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001048 __skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -08001049 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001050 sk->sk_wmem_queued += skb->truesize;
1051 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052}
1053
Andi Kleen67edfef2009-07-21 23:00:40 +00001054/* Initialize TSO segments for a packet. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001055static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -07001056{
Eric Dumazet8f26fb12013-10-15 12:24:54 -07001057 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -07001058 /* Avoid the costly divide in the normal
1059 * non-TSO case.
1060 */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001061 tcp_skb_pcount_set(skb, 1);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001062 TCP_SKB_CB(skb)->tcp_gso_size = 0;
David S. Millerf6302d12005-07-05 15:18:03 -07001063 } else {
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001064 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
Eric Dumazetf69ad292015-06-11 09:15:18 -07001065 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 }
1067}
1068
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001069/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001070 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001071 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001072static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001073 int decr)
1074{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001075 struct tcp_sock *tp = tcp_sk(sk);
1076
Ilpo Järvinendc869672007-10-01 15:27:19 -07001077 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001078 return;
1079
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001080 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001081 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001082}
1083
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001084/* Pcount in the middle of the write queue got changed, we need to do various
1085 * tweaks to fix counters
1086 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001087static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001088{
1089 struct tcp_sock *tp = tcp_sk(sk);
1090
1091 tp->packets_out -= decr;
1092
1093 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1094 tp->sacked_out -= decr;
1095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1096 tp->retrans_out -= decr;
1097 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1098 tp->lost_out -= decr;
1099
1100 /* Reno case is special. Sigh... */
1101 if (tcp_is_reno(tp) && decr > 0)
1102 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1103
1104 tcp_adjust_fackets_out(sk, skb, decr);
1105
1106 if (tp->lost_skb_hint &&
1107 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001108 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001109 tp->lost_cnt_hint -= decr;
1110
1111 tcp_verify_left_out(tp);
1112}
1113
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001114static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1115{
1116 struct skb_shared_info *shinfo = skb_shinfo(skb);
1117
1118 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1119 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1120 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1121 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1122
1123 shinfo->tx_flags &= ~tsflags;
1124 shinfo2->tx_flags |= tsflags;
1125 swap(shinfo->tskey, shinfo2->tskey);
1126 }
1127}
1128
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129/* Function to create two new TCP segments. Shrinks the given segment
1130 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001131 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 * Remember, these are still headerless SKBs at this point.
1133 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001134int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001135 unsigned int mss_now, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136{
1137 struct tcp_sock *tp = tcp_sk(sk);
1138 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001139 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001140 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001141 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001143 if (WARN_ON(len > skb->len))
1144 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001145
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 nsize = skb_headlen(skb) - len;
1147 if (nsize < 0)
1148 nsize = 0;
1149
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001150 if (skb_unclone(skb, gfp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 return -ENOMEM;
1152
1153 /* Get a new skb... force flag on. */
Eric Dumazeteb934472015-05-19 13:26:55 -07001154 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001155 if (!buff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001157
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001158 sk->sk_wmem_queued += buff->truesize;
1159 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001160 nlen = skb->len - len - nsize;
1161 buff->truesize += nlen;
1162 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163
1164 /* Correct the sequence numbers. */
1165 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1166 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1167 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1168
1169 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001170 flags = TCP_SKB_CB(skb)->tcp_flags;
1171 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1172 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001173 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174
Patrick McHardy84fa7932006-08-29 16:44:56 -07001175 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001177 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1178 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 nsize, 0);
1180
1181 skb_trim(skb, len);
1182
1183 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1184 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001185 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 skb_split(skb, buff, len);
1187 }
1188
1189 buff->ip_summed = skb->ip_summed;
1190
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001191 buff->tstamp = skb->tstamp;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001192 tcp_fragment_tstamp(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
David S. Miller6475be12005-09-01 22:47:01 -07001194 old_factor = tcp_skb_pcount(skb);
1195
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001197 tcp_set_skb_tso_segs(skb, mss_now);
1198 tcp_set_skb_tso_segs(buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
David S. Miller6475be12005-09-01 22:47:01 -07001200 /* If this packet has been sent out already, we must
1201 * adjust the various packet counters.
1202 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001203 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001204 int diff = old_factor - tcp_skb_pcount(skb) -
1205 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001206
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001207 if (diff)
1208 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 }
1210
1211 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001212 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001213 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214
1215 return 0;
1216}
1217
1218/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1219 * eventually). The difference is that pulled data not copied, but
1220 * immediately discarded.
1221 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001222static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001224 struct skb_shared_info *shinfo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 int i, k, eat;
1226
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001227 eat = min_t(int, len, skb_headlen(skb));
1228 if (eat) {
1229 __skb_pull(skb, eat);
1230 len -= eat;
1231 if (!len)
1232 return;
1233 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 eat = len;
1235 k = 0;
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001236 shinfo = skb_shinfo(skb);
1237 for (i = 0; i < shinfo->nr_frags; i++) {
1238 int size = skb_frag_size(&shinfo->frags[i]);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001239
1240 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001241 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001242 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 } else {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001244 shinfo->frags[k] = shinfo->frags[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 if (eat) {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001246 shinfo->frags[k].page_offset += eat;
1247 skb_frag_size_sub(&shinfo->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 eat = 0;
1249 }
1250 k++;
1251 }
1252 }
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001253 shinfo->nr_frags = k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001255 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 skb->data_len -= len;
1257 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258}
1259
Andi Kleen67edfef2009-07-21 23:00:40 +00001260/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1262{
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001263 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 return -ENOMEM;
1265
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001266 __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267
1268 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001269 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270
1271 skb->truesize -= len;
1272 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001273 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1275
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001276 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 if (tcp_skb_pcount(skb) > 1)
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001278 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
1280 return 0;
1281}
1282
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001283/* Calculate MSS not accounting any TCP options. */
1284static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001285{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001286 const struct tcp_sock *tp = tcp_sk(sk);
1287 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001288 int mss_now;
1289
1290 /* Calculate base mss without TCP options:
1291 It is MMS_S - sizeof(tcphdr) of rfc1122
1292 */
1293 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1294
Eric Dumazet67469602012-04-24 07:37:38 +00001295 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1296 if (icsk->icsk_af_ops->net_frag_header_len) {
1297 const struct dst_entry *dst = __sk_dst_get(sk);
1298
1299 if (dst && dst_allfrag(dst))
1300 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1301 }
1302
John Heffner5d424d52006-03-20 17:53:41 -08001303 /* Clamp it (mss_clamp does not include tcp options) */
1304 if (mss_now > tp->rx_opt.mss_clamp)
1305 mss_now = tp->rx_opt.mss_clamp;
1306
1307 /* Now subtract optional transport overhead */
1308 mss_now -= icsk->icsk_ext_hdr_len;
1309
1310 /* Then reserve room for full set of TCP options and 8 bytes of data */
1311 if (mss_now < 48)
1312 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001313 return mss_now;
1314}
1315
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001316/* Calculate MSS. Not accounting for SACKs here. */
1317int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1318{
1319 /* Subtract TCP options size, not including SACKs */
1320 return __tcp_mtu_to_mss(sk, pmtu) -
1321 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1322}
1323
John Heffner5d424d52006-03-20 17:53:41 -08001324/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001325int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001326{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001327 const struct tcp_sock *tp = tcp_sk(sk);
1328 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001329 int mtu;
1330
1331 mtu = mss +
1332 tp->tcp_header_len +
1333 icsk->icsk_ext_hdr_len +
1334 icsk->icsk_af_ops->net_header_len;
1335
Eric Dumazet67469602012-04-24 07:37:38 +00001336 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1337 if (icsk->icsk_af_ops->net_frag_header_len) {
1338 const struct dst_entry *dst = __sk_dst_get(sk);
1339
1340 if (dst && dst_allfrag(dst))
1341 mtu += icsk->icsk_af_ops->net_frag_header_len;
1342 }
John Heffner5d424d52006-03-20 17:53:41 -08001343 return mtu;
1344}
1345
Andi Kleen67edfef2009-07-21 23:00:40 +00001346/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001347void tcp_mtup_init(struct sock *sk)
1348{
1349 struct tcp_sock *tp = tcp_sk(sk);
1350 struct inet_connection_sock *icsk = inet_csk(sk);
Fan Dub0f9ca52015-02-10 09:53:16 +08001351 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001352
Fan Dub0f9ca52015-02-10 09:53:16 +08001353 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
John Heffner5d424d52006-03-20 17:53:41 -08001354 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001355 icsk->icsk_af_ops->net_header_len;
Fan Dub0f9ca52015-02-10 09:53:16 +08001356 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
John Heffner5d424d52006-03-20 17:53:41 -08001357 icsk->icsk_mtup.probe_size = 0;
Fan Du05cbc0d2015-03-06 11:18:24 +08001358 if (icsk->icsk_mtup.enabled)
1359 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
John Heffner5d424d52006-03-20 17:53:41 -08001360}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001361EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001362
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363/* This function synchronize snd mss to current pmtu/exthdr set.
1364
1365 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1366 for TCP options, but includes only bare TCP header.
1367
1368 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001369 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 It also does not include TCP options.
1371
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001372 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373
1374 tp->mss_cache is current effective sending mss, including
1375 all tcp options except for SACKs. It is evaluated,
1376 taking into account current pmtu, but never exceeds
1377 tp->rx_opt.mss_clamp.
1378
1379 NOTE1. rfc1122 clearly states that advertised MSS
1380 DOES NOT include either tcp or ip options.
1381
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001382 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1383 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386{
1387 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001388 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001389 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390
John Heffner5d424d52006-03-20 17:53:41 -08001391 if (icsk->icsk_mtup.search_high > pmtu)
1392 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
John Heffner5d424d52006-03-20 17:53:41 -08001394 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001395 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
1397 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001398 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001399 if (icsk->icsk_mtup.enabled)
1400 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001401 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402
1403 return mss_now;
1404}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001405EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406
1407/* Compute the current effective MSS, taking SACKs and IP options,
1408 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001410unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001412 const struct tcp_sock *tp = tcp_sk(sk);
1413 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001414 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001415 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001416 struct tcp_out_options opts;
1417 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001419 mss_now = tp->mss_cache;
1420
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 if (dst) {
1422 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001423 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 mss_now = tcp_sync_mss(sk, mtu);
1425 }
1426
Adam Langley33ad7982008-07-19 00:04:31 -07001427 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1428 sizeof(struct tcphdr);
1429 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1430 * some common options. If this is an odd packet (because we have SACK
1431 * blocks etc) then our calculated header_len will be different, and
1432 * we have to adjust mss_now correspondingly */
1433 if (header_len != tp->tcp_header_len) {
1434 int delta = (int) header_len - tp->tcp_header_len;
1435 mss_now -= delta;
1436 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001437
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 return mss_now;
1439}
1440
Weiping Pan86fd14a2014-04-18 12:27:46 +08001441/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1442 * As additional protections, we do not touch cwnd in retransmission phases,
1443 * and if application hit its sndbuf limit recently.
1444 */
1445static void tcp_cwnd_application_limited(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001446{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001447 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001448
Weiping Pan86fd14a2014-04-18 12:27:46 +08001449 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1450 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1451 /* Limited by application or receiver window. */
1452 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1453 u32 win_used = max(tp->snd_cwnd_used, init_win);
1454 if (win_used < tp->snd_cwnd) {
1455 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1456 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457 }
1458 tp->snd_cwnd_used = 0;
1459 }
1460 tp->snd_cwnd_stamp = tcp_time_stamp;
1461}
1462
Neal Cardwellca8a2262014-05-22 10:41:08 -04001463static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
David S. Millera762a982005-07-05 15:18:51 -07001464{
1465 struct tcp_sock *tp = tcp_sk(sk);
1466
Neal Cardwellca8a2262014-05-22 10:41:08 -04001467 /* Track the maximum number of outstanding packets in each
1468 * window, and remember whether we were cwnd-limited then.
1469 */
1470 if (!before(tp->snd_una, tp->max_packets_seq) ||
1471 tp->packets_out > tp->max_packets_out) {
1472 tp->max_packets_out = tp->packets_out;
1473 tp->max_packets_seq = tp->snd_nxt;
1474 tp->is_cwnd_limited = is_cwnd_limited;
1475 }
Eric Dumazete114a712014-04-30 11:58:13 -07001476
Eric Dumazet24901552014-05-02 21:18:05 -07001477 if (tcp_is_cwnd_limited(sk)) {
David S. Millera762a982005-07-05 15:18:51 -07001478 /* Network is feed fully. */
1479 tp->snd_cwnd_used = 0;
1480 tp->snd_cwnd_stamp = tcp_time_stamp;
1481 } else {
1482 /* Network starves. */
1483 if (tp->packets_out > tp->snd_cwnd_used)
1484 tp->snd_cwnd_used = tp->packets_out;
1485
David S. Miller15d33c02007-04-09 13:23:14 -07001486 if (sysctl_tcp_slow_start_after_idle &&
1487 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001488 tcp_cwnd_application_limited(sk);
1489 }
1490}
1491
Eric Dumazetd4589922013-12-13 13:51:23 -08001492/* Minshall's variant of the Nagle send check. */
1493static bool tcp_minshall_check(const struct tcp_sock *tp)
1494{
1495 return after(tp->snd_sml, tp->snd_una) &&
1496 !after(tp->snd_sml, tp->snd_nxt);
1497}
1498
1499/* Update snd_sml if this skb is under mss
1500 * Note that a TSO packet might end with a sub-mss segment
1501 * The test is really :
1502 * if ((skb->len % mss) != 0)
1503 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1504 * But we can avoid doing the divide again given we already have
1505 * skb_pcount = skb->len / mss_now
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001506 */
Eric Dumazetd4589922013-12-13 13:51:23 -08001507static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1508 const struct sk_buff *skb)
1509{
1510 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1511 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1512}
1513
1514/* Return false, if packet can be sent now without violation Nagle's rules:
1515 * 1. It is full sized. (provided by caller in %partial bool)
1516 * 2. Or it contains FIN. (already checked by caller)
1517 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1518 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1519 * With Minshall's modification: all sent small packets are ACKed.
1520 */
1521static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001522 int nonagle)
Eric Dumazetd4589922013-12-13 13:51:23 -08001523{
1524 return partial &&
1525 ((nonagle & TCP_NAGLE_CORK) ||
1526 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1527}
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001528
1529/* Return how many segs we'd like on a TSO packet,
1530 * to send one TSO packet per ms
1531 */
1532static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1533{
1534 u32 bytes, segs;
1535
1536 bytes = min(sk->sk_pacing_rate >> 10,
1537 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1538
1539 /* Goal is to send at least one packet per ms,
1540 * not one big TSO packet every 100 ms.
1541 * This preserves ACK clocking and is consistent
1542 * with tcp_tso_should_defer() heuristic.
1543 */
1544 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1545
1546 return min_t(u32, segs, sk->sk_gso_max_segs);
1547}
1548
Eric Dumazetd4589922013-12-13 13:51:23 -08001549/* Returns the portion of skb which can be sent right away */
1550static unsigned int tcp_mss_split_point(const struct sock *sk,
1551 const struct sk_buff *skb,
1552 unsigned int mss_now,
1553 unsigned int max_segs,
1554 int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001555{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001556 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetd4589922013-12-13 13:51:23 -08001557 u32 partial, needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001558
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001559 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001560 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001561
Ben Hutchings14853482012-07-30 16:11:42 +00001562 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1563 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001564
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001565 needed = min(skb->len, window);
1566
Ben Hutchings14853482012-07-30 16:11:42 +00001567 if (max_len <= needed)
1568 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001569
Eric Dumazetd4589922013-12-13 13:51:23 -08001570 partial = needed % mss_now;
1571 /* If last segment is not a full MSS, check if Nagle rules allow us
1572 * to include this last segment in this skb.
1573 * Otherwise, we'll split the skb at last MSS boundary
1574 */
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001575 if (tcp_nagle_check(partial != 0, tp, nonagle))
Eric Dumazetd4589922013-12-13 13:51:23 -08001576 return needed - partial;
1577
1578 return needed;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001579}
1580
1581/* Can at least one segment of SKB be sent right now, according to the
1582 * congestion window rules? If so, return how many segments are allowed.
1583 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001584static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1585 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001586{
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001587 u32 in_flight, cwnd, halfcwnd;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001588
1589 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001590 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1591 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001592 return 1;
1593
1594 in_flight = tcp_packets_in_flight(tp);
1595 cwnd = tp->snd_cwnd;
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001596 if (in_flight >= cwnd)
1597 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001598
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001599 /* For better scheduling, ensure we have at least
1600 * 2 GSO packets in flight.
1601 */
1602 halfcwnd = max(cwnd >> 1, 1U);
1603 return min(halfcwnd, cwnd - in_flight);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001604}
1605
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001606/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001607 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001608 * SKB onto the wire.
1609 */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001610static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001611{
1612 int tso_segs = tcp_skb_pcount(skb);
1613
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001614 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001615 tcp_set_skb_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001616 tso_segs = tcp_skb_pcount(skb);
1617 }
1618 return tso_segs;
1619}
1620
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001621
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001622/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001623 * sent now.
1624 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001625static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1626 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001627{
1628 /* Nagle rule does not apply to frames, which sit in the middle of the
1629 * write_queue (they have no chances to get new data).
1630 *
1631 * This is implemented in the callers, where they modify the 'nonagle'
1632 * argument based upon the location of SKB in the send queue.
1633 */
1634 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001635 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001636
Yuchung Cheng9b441902013-03-20 13:32:58 +00001637 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1638 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001639 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001640
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001641 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001642 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001643
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001644 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001645}
1646
1647/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001648static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1649 const struct sk_buff *skb,
1650 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001651{
1652 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1653
1654 if (skb->len > cur_mss)
1655 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1656
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001657 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001658}
1659
David S. Millerfe067e82007-03-07 12:12:44 -08001660/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001661 * should be put on the wire right now. If so, it returns the number of
1662 * packets allowed by the congestion window.
1663 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001664static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001665 unsigned int cur_mss, int nonagle)
1666{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001667 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001668 unsigned int cwnd_quota;
1669
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001670 tcp_init_tso_segs(skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001671
1672 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1673 return 0;
1674
1675 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001676 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001677 cwnd_quota = 0;
1678
1679 return cwnd_quota;
1680}
1681
Andi Kleen67edfef2009-07-21 23:00:40 +00001682/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001683bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001684{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001685 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001686 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001687
Eric Dumazeta02cec22010-09-22 20:43:57 +00001688 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001689 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001690 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001691 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001692}
1693
1694/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1695 * which is put after SKB on the list. It is very much like
1696 * tcp_fragment() except that it may make several kinds of assumptions
1697 * in order to speed up the splitting operation. In particular, we
1698 * know that all the data is in scatter-gather pages, and that the
1699 * packet has never been sent out before (and thus is not cloned).
1700 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001701static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001702 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001703{
1704 struct sk_buff *buff;
1705 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001706 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001707
1708 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001709 if (skb->len != skb->data_len)
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001710 return tcp_fragment(sk, skb, len, mss_now, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001711
Eric Dumazeteb934472015-05-19 13:26:55 -07001712 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001713 if (unlikely(!buff))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001714 return -ENOMEM;
1715
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001716 sk->sk_wmem_queued += buff->truesize;
1717 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001718 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001719 skb->truesize -= nlen;
1720
1721 /* Correct the sequence numbers. */
1722 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1723 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1724 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1725
1726 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001727 flags = TCP_SKB_CB(skb)->tcp_flags;
1728 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1729 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001730
1731 /* This packet was never sent out yet, so no SACK bits. */
1732 TCP_SKB_CB(buff)->sacked = 0;
1733
Patrick McHardy84fa7932006-08-29 16:44:56 -07001734 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001735 skb_split(skb, buff, len);
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001736 tcp_fragment_tstamp(skb, buff);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001737
1738 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001739 tcp_set_skb_tso_segs(skb, mss_now);
1740 tcp_set_skb_tso_segs(buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001741
1742 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001743 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001744 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001745
1746 return 0;
1747}
1748
1749/* Try to defer sending, if possible, in order to minimize the amount
1750 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1751 *
1752 * This algorithm is from John Heffner.
1753 */
Neal Cardwellca8a2262014-05-22 10:41:08 -04001754static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001755 bool *is_cwnd_limited, u32 max_segs)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001756{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001757 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet50c83392015-02-26 14:10:19 -08001758 u32 age, send_win, cong_win, limit, in_flight;
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct skb_mstamp now;
1761 struct sk_buff *head;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001762 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001763
Eric Dumazet4de075e2011-09-27 13:25:05 -04001764 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001765 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001766
Eric Dumazet99d76622015-07-26 09:45:24 +02001767 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
John Heffnerae8064a2006-10-18 20:36:48 -07001768 goto send_now;
1769
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001770 /* Avoid bursty behavior by allowing defer
1771 * only if the last write was recent.
1772 */
1773 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
John Heffnerae8064a2006-10-18 20:36:48 -07001774 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001775
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001776 in_flight = tcp_packets_in_flight(tp);
1777
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001778 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001779
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001780 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001781
1782 /* From in_flight test above, we know that cwnd > in_flight. */
1783 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1784
1785 limit = min(send_win, cong_win);
1786
David S. Millerba244fe2006-03-11 18:51:49 -08001787 /* If a full-sized TSO skb can be sent, do it. */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001788 if (limit >= max_segs * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001789 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001790
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001791 /* Middle in queue won't get any more data, full sendable already? */
1792 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1793 goto send_now;
1794
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001795 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1796 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001797 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1798
1799 /* If at least some fraction of a window is available,
1800 * just use it.
1801 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001802 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001803 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001804 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001805 } else {
1806 /* Different approach, try not to defer past a single
1807 * ACK. Receiver should ACK every other full sized
1808 * frame, so if we have space for more than 3 frames
1809 * then send now.
1810 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001811 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001812 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001813 }
1814
Eric Dumazet50c83392015-02-26 14:10:19 -08001815 head = tcp_write_queue_head(sk);
1816 skb_mstamp_get(&now);
1817 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1818 /* If next ACK is likely to come too late (half srtt), do not defer */
1819 if (age < (tp->srtt_us >> 4))
1820 goto send_now;
1821
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001822 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001823
Bendik Rønning Opstadd2e13392015-09-23 18:49:53 +02001824 if (cong_win < send_win && cong_win <= skb->len)
Neal Cardwellca8a2262014-05-22 10:41:08 -04001825 *is_cwnd_limited = true;
1826
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001827 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001828
1829send_now:
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001830 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001831}
1832
Fan Du05cbc0d2015-03-06 11:18:24 +08001833static inline void tcp_mtu_check_reprobe(struct sock *sk)
1834{
1835 struct inet_connection_sock *icsk = inet_csk(sk);
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct net *net = sock_net(sk);
1838 u32 interval;
1839 s32 delta;
1840
1841 interval = net->ipv4.sysctl_tcp_probe_interval;
1842 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1843 if (unlikely(delta >= interval * HZ)) {
1844 int mss = tcp_current_mss(sk);
1845
1846 /* Update current search range */
1847 icsk->icsk_mtup.probe_size = 0;
1848 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1849 sizeof(struct tcphdr) +
1850 icsk->icsk_af_ops->net_header_len;
1851 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1852
1853 /* Update probe time stamp */
1854 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1855 }
1856}
1857
John Heffner5d424d52006-03-20 17:53:41 -08001858/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001859 * MTU probe is regularly attempting to increase the path MTU by
1860 * deliberately sending larger packets. This discovers routing
1861 * changes resulting in larger path MTUs.
1862 *
John Heffner5d424d52006-03-20 17:53:41 -08001863 * Returns 0 if we should wait to probe (no cwnd available),
1864 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001865 * -1 otherwise
1866 */
John Heffner5d424d52006-03-20 17:53:41 -08001867static int tcp_mtu_probe(struct sock *sk)
1868{
1869 struct tcp_sock *tp = tcp_sk(sk);
1870 struct inet_connection_sock *icsk = inet_csk(sk);
1871 struct sk_buff *skb, *nskb, *next;
Fan Du6b58e0a2015-03-06 11:18:23 +08001872 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001873 int len;
1874 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001875 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001876 int copy;
1877 int mss_now;
Fan Du6b58e0a2015-03-06 11:18:23 +08001878 int interval;
John Heffner5d424d52006-03-20 17:53:41 -08001879
1880 /* Not currently probing/verifying,
1881 * not in recovery,
1882 * have enough cwnd, and
1883 * not SACKing (the variable headers throw things off) */
1884 if (!icsk->icsk_mtup.enabled ||
1885 icsk->icsk_mtup.probe_size ||
1886 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1887 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001888 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001889 return -1;
1890
Fan Du6b58e0a2015-03-06 11:18:23 +08001891 /* Use binary search for probe_size between tcp_mss_base,
1892 * and current mss_clamp. if (search_high - search_low)
1893 * smaller than a threshold, backoff from probing.
1894 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001895 mss_now = tcp_current_mss(sk);
Fan Du6b58e0a2015-03-06 11:18:23 +08001896 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1897 icsk->icsk_mtup.search_low) >> 1);
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001898 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
Fan Du6b58e0a2015-03-06 11:18:23 +08001899 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
Fan Du05cbc0d2015-03-06 11:18:24 +08001900 /* When misfortune happens, we are reprobing actively,
1901 * and then reprobe timer has expired. We stick with current
1902 * probing process by not resetting search range to its orignal.
1903 */
Fan Du6b58e0a2015-03-06 11:18:23 +08001904 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
Fan Du05cbc0d2015-03-06 11:18:24 +08001905 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1906 /* Check whether enough time has elaplased for
1907 * another round of probing.
1908 */
1909 tcp_mtu_check_reprobe(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001910 return -1;
1911 }
1912
1913 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001914 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001915 return -1;
1916
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001917 if (tp->snd_wnd < size_needed)
1918 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001919 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001920 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001921
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001922 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1923 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1924 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001925 return -1;
1926 else
1927 return 0;
1928 }
1929
1930 /* We're allowed to probe. Build it now. */
Eric Dumazeteb934472015-05-19 13:26:55 -07001931 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
Ian Morris51456b22015-04-03 09:17:26 +01001932 if (!nskb)
John Heffner5d424d52006-03-20 17:53:41 -08001933 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001934 sk->sk_wmem_queued += nskb->truesize;
1935 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001936
David S. Millerfe067e82007-03-07 12:12:44 -08001937 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001938
1939 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1940 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001941 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001942 TCP_SKB_CB(nskb)->sacked = 0;
1943 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001944 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001945
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001946 tcp_insert_write_queue_before(nskb, skb, sk);
1947
John Heffner5d424d52006-03-20 17:53:41 -08001948 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001949 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001950 copy = min_t(int, skb->len, probe_size - len);
1951 if (nskb->ip_summed)
1952 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1953 else
1954 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001955 skb_put(nskb, copy),
1956 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001957
1958 if (skb->len <= copy) {
1959 /* We've eaten all the data from this skb.
1960 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001961 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001962 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001963 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001964 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04001965 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001966 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001967 if (!skb_shinfo(skb)->nr_frags) {
1968 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001969 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001970 skb->csum = csum_partial(skb->data,
1971 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001972 } else {
1973 __pskb_trim_head(skb, copy);
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001974 tcp_set_skb_tso_segs(skb, mss_now);
John Heffner5d424d52006-03-20 17:53:41 -08001975 }
1976 TCP_SKB_CB(skb)->seq += copy;
1977 }
1978
1979 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001980
1981 if (len >= probe_size)
1982 break;
John Heffner5d424d52006-03-20 17:53:41 -08001983 }
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001984 tcp_init_tso_segs(nskb, nskb->len);
John Heffner5d424d52006-03-20 17:53:41 -08001985
1986 /* We're ready to send. If this fails, the probe will
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001987 * be resegmented into mss-sized pieces by tcp_write_xmit().
1988 */
John Heffner5d424d52006-03-20 17:53:41 -08001989 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1990 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001991 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001992 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001993 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001994
1995 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001996 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1997 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001998
1999 return 1;
2000 }
2001
2002 return -1;
2003}
2004
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005/* This routine writes packets to the network. It advances the
2006 * send_head. This happens as incoming acks open up the remote
2007 * window for us.
2008 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08002009 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2010 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2011 * account rare use of URG, this is not a big flaw.
2012 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002013 * Send at most one packet when push_one > 0. Temporarily ignore
2014 * cwnd limit to force at most one packet out when push_one == 2.
2015
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002016 * Returns true, if no segments are in flight and we have queued segments,
2017 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002019static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2020 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021{
2022 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07002023 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002024 unsigned int tso_segs, sent_pkts;
2025 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08002026 int result;
Neal Cardwellca8a2262014-05-22 10:41:08 -04002027 bool is_cwnd_limited = false;
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002028 u32 max_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029
David S. Miller92df7b52005-07-05 15:19:06 -07002030 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002031
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002032 if (!push_one) {
2033 /* Do MTU probing. */
2034 result = tcp_mtu_probe(sk);
2035 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002036 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002037 } else if (result > 0) {
2038 sent_pkts = 1;
2039 }
John Heffner5d424d52006-03-20 17:53:41 -08002040 }
2041
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002042 max_segs = tcp_tso_autosize(sk, mss_now);
David S. Millerfe067e82007-03-07 12:12:44 -08002043 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07002044 unsigned int limit;
2045
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002046 tso_segs = tcp_init_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002047 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07002048
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002049 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002050 /* "skb_mstamp" is used as a start point for the retransmit timer */
2051 skb_mstamp_get(&skb->skb_mstamp);
Andrew Vaginec342322012-11-15 04:03:17 +00002052 goto repair; /* Skip network transmission */
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002053 }
Andrew Vaginec342322012-11-15 04:03:17 +00002054
Herbert Xub68e9f82005-08-04 19:52:02 -07002055 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002056 if (!cwnd_quota) {
2057 if (push_one == 2)
2058 /* Force out a loss probe pkt. */
2059 cwnd_quota = 1;
2060 else
2061 break;
2062 }
Herbert Xub68e9f82005-08-04 19:52:02 -07002063
2064 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2065 break;
2066
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002067 if (tso_segs == 1) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002068 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2069 (tcp_skb_is_last(sk, skb) ?
2070 nonagle : TCP_NAGLE_PUSH))))
2071 break;
2072 } else {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002073 if (!push_one &&
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002074 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2075 max_segs))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002076 break;
2077 }
David S. Milleraa934662005-07-05 15:20:09 -07002078
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002079 limit = mss_now;
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002080 if (tso_segs > 1 && !tcp_urg_mode(tp))
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002081 limit = tcp_mss_split_point(sk, skb, mss_now,
2082 min_t(unsigned int,
2083 cwnd_quota,
2084 max_segs),
2085 nonagle);
2086
2087 if (skb->len > limit &&
2088 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2089 break;
2090
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002091 /* TCP Small Queues :
2092 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2093 * This allows for :
2094 * - better RTT estimation and ACK scheduling
2095 * - faster recovery
2096 * - high rates
Eric Dumazet98e09382013-11-13 06:32:54 -08002097 * Alas, some drivers / subsystems require a fair amount
2098 * of queued bytes to ensure line rate.
2099 * One example is wifi aggregation (802.11 AMPDU)
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002100 */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002101 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2102 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002103
2104 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002105 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
John Ognessbf062002014-02-09 18:40:11 -08002106 /* It is possible TX completion already happened
2107 * before we set TSQ_THROTTLED, so we must
2108 * test again the condition.
John Ognessbf062002014-02-09 18:40:11 -08002109 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002110 smp_mb__after_atomic();
John Ognessbf062002014-02-09 18:40:11 -08002111 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2112 break;
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002113 }
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002114
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002115 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07002116 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Andrew Vaginec342322012-11-15 04:03:17 +00002118repair:
David S. Miller92df7b52005-07-05 15:19:06 -07002119 /* Advance the send_head. This one is sent out.
2120 * This call will increment packets_out.
2121 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002122 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07002123
2124 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002125 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002126
2127 if (push_one)
2128 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 }
David S. Miller92df7b52005-07-05 15:19:06 -07002130
David S. Milleraa934662005-07-05 15:20:09 -07002131 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002132 if (tcp_in_cwnd_reduction(sk))
2133 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002134
2135 /* Send one loss probe per tail loss episode. */
2136 if (push_one != 2)
2137 tcp_schedule_loss_probe(sk);
Bendik Rønning Opstadd2e13392015-09-23 18:49:53 +02002138 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
Neal Cardwellca8a2262014-05-22 10:41:08 -04002139 tcp_cwnd_validate(sk, is_cwnd_limited);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002140 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07002141 }
Yuchung Chengb340b262015-08-12 11:18:19 -07002142 return !tp->packets_out && tcp_send_head(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002143}
2144
2145bool tcp_schedule_loss_probe(struct sock *sk)
2146{
2147 struct inet_connection_sock *icsk = inet_csk(sk);
2148 struct tcp_sock *tp = tcp_sk(sk);
2149 u32 timeout, tlp_time_stamp, rto_time_stamp;
Eric Dumazet740b0f12014-02-26 14:02:48 -08002150 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002151
2152 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2153 return false;
2154 /* No consecutive loss probes. */
2155 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2156 tcp_rearm_rto(sk);
2157 return false;
2158 }
2159 /* Don't do any loss probe on a Fast Open connection before 3WHS
2160 * finishes.
2161 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002162 if (tp->fastopen_rsk)
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002163 return false;
2164
2165 /* TLP is only scheduled when next timer event is RTO. */
2166 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2167 return false;
2168
2169 /* Schedule a loss probe in 2*RTT for SACK capable connections
2170 * in Open state, that are either limited by cwnd or application.
2171 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002172 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002173 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2174 return false;
2175
2176 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2177 tcp_send_head(sk))
2178 return false;
2179
2180 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
Yuchung Chengf9b99582015-09-18 11:40:33 -07002181 * for delayed ack when there's one outstanding packet. If no RTT
2182 * sample is available then probe after TCP_TIMEOUT_INIT.
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002183 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002184 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002185 if (tp->packets_out == 1)
2186 timeout = max_t(u32, timeout,
2187 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2188 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2189
2190 /* If RTO is shorter, just schedule TLP in its place. */
2191 tlp_time_stamp = tcp_time_stamp + timeout;
2192 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2193 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2194 s32 delta = rto_time_stamp - tcp_time_stamp;
2195 if (delta > 0)
2196 timeout = delta;
2197 }
2198
2199 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2200 TCP_RTO_MAX);
2201 return true;
2202}
2203
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002204/* Thanks to skb fast clones, we can detect if a prior transmit of
2205 * a packet is still in a qdisc or driver queue.
2206 * In this case, there is very little point doing a retransmit !
2207 * Note: This is called from BH context only.
2208 */
2209static bool skb_still_in_host_queue(const struct sock *sk,
2210 const struct sk_buff *skb)
2211{
Eric Dumazet39bb5e62014-10-30 10:32:34 -07002212 if (unlikely(skb_fclone_busy(sk, skb))) {
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002213 NET_INC_STATS_BH(sock_net(sk),
2214 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2215 return true;
2216 }
2217 return false;
2218}
2219
Yuchung Chengb340b262015-08-12 11:18:19 -07002220/* When probe timeout (PTO) fires, try send a new segment if possible, else
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002221 * retransmit the last segment.
2222 */
2223void tcp_send_loss_probe(struct sock *sk)
2224{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002225 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002226 struct sk_buff *skb;
2227 int pcount;
2228 int mss = tcp_current_mss(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002229
Yuchung Chengb340b262015-08-12 11:18:19 -07002230 skb = tcp_send_head(sk);
2231 if (skb) {
2232 if (tcp_snd_wnd_test(tp, skb, mss)) {
2233 pcount = tp->packets_out;
2234 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2235 if (tp->packets_out > pcount)
2236 goto probe_sent;
2237 goto rearm_timer;
2238 }
2239 skb = tcp_write_queue_prev(sk, skb);
2240 } else {
2241 skb = tcp_write_queue_tail(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002242 }
2243
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002244 /* At most one outstanding TLP retransmission. */
2245 if (tp->tlp_high_seq)
2246 goto rearm_timer;
2247
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002248 /* Retransmit last segment. */
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002249 if (WARN_ON(!skb))
2250 goto rearm_timer;
2251
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002252 if (skb_still_in_host_queue(sk, skb))
2253 goto rearm_timer;
2254
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002255 pcount = tcp_skb_pcount(skb);
2256 if (WARN_ON(!pcount))
2257 goto rearm_timer;
2258
2259 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002260 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2261 GFP_ATOMIC)))
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002262 goto rearm_timer;
Yuchung Chengb340b262015-08-12 11:18:19 -07002263 skb = tcp_write_queue_next(sk, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002264 }
2265
2266 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2267 goto rearm_timer;
2268
Yuchung Chengb340b262015-08-12 11:18:19 -07002269 if (__tcp_retransmit_skb(sk, skb))
2270 goto rearm_timer;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002271
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002272 /* Record snd_nxt for loss detection. */
Yuchung Chengb340b262015-08-12 11:18:19 -07002273 tp->tlp_high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002274
Yuchung Chengb340b262015-08-12 11:18:19 -07002275probe_sent:
2276 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2277 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2278 inet_csk(sk)->icsk_pending = 0;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002279rearm_timer:
Yuchung Chengfcd16c02015-08-12 11:18:18 -07002280 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281}
2282
David S. Millera762a982005-07-05 15:18:51 -07002283/* Push out any pending frames which were held back due to
2284 * TCP_CORK or attempt at coalescing tiny packets.
2285 * The socket must be locked by the caller.
2286 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002287void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2288 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002289{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002290 /* If we are closed, the bytes will have to remain here.
2291 * In time closedown will finish, we empty the write queue and
2292 * all will be happy.
2293 */
2294 if (unlikely(sk->sk_state == TCP_CLOSE))
2295 return;
2296
Mel Gorman99a1dec2012-07-31 16:44:14 -07002297 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
Eric Dumazet7450aaf2015-11-30 08:57:28 -08002298 sk_gfp_mask(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002299 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002300}
2301
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002302/* Send _single_ skb sitting at the send head. This function requires
2303 * true push pending frames to setup probe timer etc.
2304 */
2305void tcp_push_one(struct sock *sk, unsigned int mss_now)
2306{
David S. Millerfe067e82007-03-07 12:12:44 -08002307 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002308
2309 BUG_ON(!skb || skb->len < mss_now);
2310
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002311 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002312}
2313
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314/* This function returns the amount that we can raise the
2315 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002316 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 * 1. The window can never be shrunk once it is offered (RFC 793)
2318 * 2. We limit memory per socket
2319 *
2320 * RFC 1122:
2321 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2322 * RECV.NEXT + RCV.WIN fixed until:
2323 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2324 *
2325 * i.e. don't raise the right edge of the window until you can raise
2326 * it at least MSS bytes.
2327 *
2328 * Unfortunately, the recommended algorithm breaks header prediction,
2329 * since header prediction assumes th->window stays fixed.
2330 *
2331 * Strictly speaking, keeping th->window fixed violates the receiver
2332 * side SWS prevention criteria. The problem is that under this rule
2333 * a stream of single byte packets will cause the right side of the
2334 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002335 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 * Of course, if the sender implements sender side SWS prevention
2337 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002338 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002340 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 * If the free space is less than the 1/4 of the maximum
2342 * space available and the free space is less than 1/2 mss,
2343 * then set the window to 0.
2344 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2345 * Otherwise, just prevent the window from shrinking
2346 * and from being larger than the largest representable value.
2347 *
2348 * This prevents incremental opening of the window in the regime
2349 * where TCP is limited by the speed of the reader side taking
2350 * data out of the TCP receive queue. It does nothing about
2351 * those cases where the window is constrained on the sender side
2352 * because the pipeline is full.
2353 *
2354 * BSD also seems to "accidentally" limit itself to windows that are a
2355 * multiple of MSS, at least until the free space gets quite small.
2356 * This would appear to be a side effect of the mbuf implementation.
2357 * Combining these two algorithms results in the observed behavior
2358 * of having a fixed window size at almost all times.
2359 *
2360 * Below we obtain similar behavior by forcing the offered window to
2361 * a multiple of the mss when it is feasible to do so.
2362 *
2363 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2364 * Regular options like TIMESTAMP are taken into account.
2365 */
2366u32 __tcp_select_window(struct sock *sk)
2367{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002368 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002370 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 * here. I don't know if the value based on our guesses
2372 * of peer's MSS is better for the performance. It's more correct
2373 * but may be worse for the performance because of rcv_mss
2374 * fluctuations. --SAW 1998/11/1
2375 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002376 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 int free_space = tcp_space(sk);
Florian Westphal86c1a042014-02-19 12:51:10 +01002378 int allowed_space = tcp_full_space(sk);
2379 int full_space = min_t(int, tp->window_clamp, allowed_space);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 int window;
2381
2382 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002383 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002385 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002386 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002388 if (tcp_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002389 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2390 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Florian Westphal86c1a042014-02-19 12:51:10 +01002392 /* free_space might become our new window, make sure we don't
2393 * increase it due to wscale.
2394 */
2395 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2396
2397 /* if free space is less than mss estimate, or is below 1/16th
2398 * of the maximum allowed, try to move to zero-window, else
2399 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2400 * new incoming data is dropped due to memory limits.
2401 * With large window, mss test triggers way too late in order
2402 * to announce zero window in time before rmem limit kicks in.
2403 */
2404 if (free_space < (allowed_space >> 4) || free_space < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 return 0;
2406 }
2407
2408 if (free_space > tp->rcv_ssthresh)
2409 free_space = tp->rcv_ssthresh;
2410
2411 /* Don't do rounding if we are using window scaling, since the
2412 * scaled window will not line up with the MSS boundary anyway.
2413 */
2414 window = tp->rcv_wnd;
2415 if (tp->rx_opt.rcv_wscale) {
2416 window = free_space;
2417
2418 /* Advertise enough space so that it won't get scaled away.
2419 * Import case: prevent zero window announcement if
2420 * 1<<rcv_wscale > mss.
2421 */
2422 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2423 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2424 << tp->rx_opt.rcv_wscale);
2425 } else {
2426 /* Get the largest window that is a nice multiple of mss.
2427 * Window clamp already applied above.
2428 * If our current window offering is within 1 mss of the
2429 * free space we just keep it. This prevents the divide
2430 * and multiply from happening most of the time.
2431 * We also don't do any window rounding when the free space
2432 * is too small.
2433 */
2434 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002435 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002436 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002437 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002438 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 }
2440
2441 return window;
2442}
2443
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002444/* Collapses two adjacent SKB's during retransmission. */
2445static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446{
2447 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002448 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002449 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002451 skb_size = skb->len;
2452 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002454 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002455
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002456 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002458 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002460 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2461 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002463 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2464 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002466 if (skb->ip_summed != CHECKSUM_PARTIAL)
2467 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002469 /* Update sequence range on original skb. */
2470 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002472 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002473 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002475 /* All done, get rid of second SKB and account for it so
2476 * packet counting does not break.
2477 */
2478 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002479
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002480 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002481 tcp_clear_retrans_hints_partial(tp);
2482 if (next_skb == tp->retransmit_skb_hint)
2483 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002484
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002485 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2486
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002487 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488}
2489
Andi Kleen67edfef2009-07-21 23:00:40 +00002490/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002491static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002492{
2493 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002494 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002495 /* TODO: SACK collapsing could be used to remove this condition */
2496 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002497 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002498 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002499 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002500 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002501 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002502 /* Some heurestics for collapsing over SACK'd could be invented */
2503 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002504 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002505
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002506 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002507}
2508
Andi Kleen67edfef2009-07-21 23:00:40 +00002509/* Collapse packets in the retransmit queue to make to create
2510 * less packets on the wire. This is only done on retransmission.
2511 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002512static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2513 int space)
2514{
2515 struct tcp_sock *tp = tcp_sk(sk);
2516 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002517 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002518
2519 if (!sysctl_tcp_retrans_collapse)
2520 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002521 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002522 return;
2523
2524 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2525 if (!tcp_can_collapse(sk, skb))
2526 break;
2527
2528 space -= skb->len;
2529
2530 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002531 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002532 continue;
2533 }
2534
2535 if (space < 0)
2536 break;
2537 /* Punt if not enough space exists in the first SKB for
2538 * the data in the second
2539 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002540 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002541 break;
2542
2543 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2544 break;
2545
2546 tcp_collapse_retrans(sk, to);
2547 }
2548}
2549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550/* This retransmits one SKB. Policy decisions and retransmit queue
2551 * state updates are done by the caller. Returns non-zero if an
2552 * error occurred which prevented the send.
2553 */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002554int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555{
2556 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002557 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002558 unsigned int cur_mss;
Yuchung Chengc84a5712014-02-28 16:42:26 -08002559 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560
John Heffner5d424d52006-03-20 17:53:41 -08002561 /* Inconslusive MTU probe */
2562 if (icsk->icsk_mtup.probe_size) {
2563 icsk->icsk_mtup.probe_size = 0;
2564 }
2565
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002567 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568 */
2569 if (atomic_read(&sk->sk_wmem_alloc) >
2570 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2571 return -EAGAIN;
2572
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002573 if (skb_still_in_host_queue(sk, skb))
2574 return -EBUSY;
2575
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2577 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2578 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2580 return -ENOMEM;
2581 }
2582
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002583 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2584 return -EHOSTUNREACH; /* Routing failure or similar. */
2585
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002586 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002587
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 /* If receiver has shrunk his window, and skb is out of
2589 * new window, do not retransmit it. The exception is the
2590 * case, when window is shrunk to zero. In this case
2591 * our retransmit serves as a zero window probe.
2592 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002593 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2594 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 return -EAGAIN;
2596
2597 if (skb->len > cur_mss) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002598 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002600 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002601 int oldpcount = tcp_skb_pcount(skb);
2602
2603 if (unlikely(oldpcount > 1)) {
Eric Dumazetc52e24212013-10-15 11:54:30 -07002604 if (skb_unclone(skb, GFP_ATOMIC))
2605 return -ENOMEM;
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002606 tcp_init_tso_segs(skb, cur_mss);
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002607 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2608 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 }
2610
Daniel Borkmann49213552015-05-19 21:04:22 +02002611 /* RFC3168, section 6.1.1.1. ECN fallback */
2612 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2613 tcp_ecn_clear_syn(sk, skb);
2614
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002615 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 /* Make a copy, if the first transmission SKB clone we made
2618 * is still in somebody's hands, else make a clone.
2619 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620
Thomas Graf50bceae2013-04-11 10:57:18 +00002621 /* make sure skb->data is aligned on arches that require it
2622 * and check if ack-trimming & collapsing extended the headroom
2623 * beyond what csum_start can cover.
2624 */
2625 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2626 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet117632e2011-12-03 21:39:53 +00002627 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2628 GFP_ATOMIC);
Yuchung Chengc84a5712014-02-28 16:42:26 -08002629 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2630 -ENOBUFS;
Eric Dumazet117632e2011-12-03 21:39:53 +00002631 } else {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002632 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002633 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002634
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002635 if (likely(!err)) {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002636 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002637 /* Update global TCP statistics. */
2638 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2639 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2640 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2641 tp->total_retrans++;
2642 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002643 return err;
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002644}
2645
2646int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2647{
2648 struct tcp_sock *tp = tcp_sk(sk);
2649 int err = __tcp_retransmit_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650
2651 if (err == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002653 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002654 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655 }
2656#endif
2657 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2658 tp->retrans_out += tcp_skb_pcount(skb);
2659
2660 /* Save stamp of the first retransmit. */
2661 if (!tp->retrans_stamp)
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002662 tp->retrans_stamp = tcp_skb_timestamp(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002664 } else if (err != -EBUSY) {
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002665 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 }
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002667
2668 if (tp->undo_retrans < 0)
2669 tp->undo_retrans = 0;
2670 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 return err;
2672}
2673
Andi Kleen67edfef2009-07-21 23:00:40 +00002674/* Check if we forward retransmits are possible in the current
2675 * window/congestion state.
2676 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002677static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002678{
2679 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002680 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002681
2682 /* Forward retransmissions are possible only during Recovery. */
2683 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002684 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002685
2686 /* No forward retransmissions in Reno are possible. */
2687 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002688 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002689
2690 /* Yeah, we have to make difficult choice between forward transmission
2691 * and retransmission... Both ways have their merits...
2692 *
2693 * For now we do not retransmit anything, while we have some new
2694 * segments to send. In the other cases, follow rule 3 for
2695 * NextSeg() specified in RFC3517.
2696 */
2697
2698 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002699 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002700
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002701 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002702}
2703
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704/* This gets called after a retransmit timeout, and the initially
2705 * retransmitted data is acknowledged. It tries to continue
2706 * resending the rest of the retransmit queue, until either
2707 * we've sent it all or the congestion window limit is reached.
2708 * If doing SACK, the first ACK which comes back for a timeout
2709 * based retransmit packet might feed us FACK information again.
2710 * If so, we use it to avoid unnecessarily retransmissions.
2711 */
2712void tcp_xmit_retransmit_queue(struct sock *sk)
2713{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002714 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 struct tcp_sock *tp = tcp_sk(sk);
2716 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002717 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002718 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002719 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002720 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002721
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002722 if (!tp->packets_out)
2723 return;
2724
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002725 if (!tp->lost_out)
2726 tp->retransmit_high = tp->snd_una;
2727
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002728 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002729 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002730 last_lost = TCP_SKB_CB(skb)->end_seq;
2731 if (after(last_lost, tp->retransmit_high))
2732 last_lost = tp->retransmit_high;
2733 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002734 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002735 last_lost = tp->snd_una;
2736 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002738 tcp_for_write_queue_from(skb, sk) {
2739 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002741 if (skb == tcp_send_head(sk))
2742 break;
2743 /* we could do better than to assign each time */
Ian Morris51456b22015-04-03 09:17:26 +01002744 if (!hole)
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002745 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002746
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002747 /* Assume this retransmit will generate
2748 * only one packet for congestion window
2749 * calculation purposes. This works because
2750 * tcp_retransmit_skb() will chop up the
2751 * packet to be MSS sized and all the
2752 * packet counting works out.
2753 */
2754 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2755 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002756
2757 if (fwd_rexmitting) {
2758begin_fwd:
2759 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2760 break;
2761 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2762
2763 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002764 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002765 if (!tcp_can_forward_retransmit(sk))
2766 break;
2767 /* Backtrack if necessary to non-L'ed skb */
Ian Morris00db4122015-04-03 09:17:27 +01002768 if (hole) {
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002769 skb = hole;
2770 hole = NULL;
2771 }
2772 fwd_rexmitting = 1;
2773 goto begin_fwd;
2774
2775 } else if (!(sacked & TCPCB_LOST)) {
Ian Morris51456b22015-04-03 09:17:26 +01002776 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002777 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002778 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002780 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002781 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002782 if (icsk->icsk_ca_state != TCP_CA_Loss)
2783 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2784 else
2785 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2786 }
2787
2788 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002789 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002790
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002791 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002792 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002793
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002794 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2795
Yuchung Cheng684bad12012-09-02 17:38:04 +00002796 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002797 tp->prr_out += tcp_skb_pcount(skb);
2798
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002799 if (skb == tcp_write_queue_head(sk))
2800 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2801 inet_csk(sk)->icsk_rto,
2802 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804}
2805
Eric Dumazetd83769a2015-04-21 18:32:24 -07002806/* We allow to exceed memory limits for FIN packets to expedite
2807 * connection tear down and (memory) recovery.
Eric Dumazet845704a2015-04-23 10:42:39 -07002808 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2809 * or even be forced to close flow without any FIN.
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002810 * In general, we want to allow one skb per socket to avoid hangs
2811 * with edge trigger epoll()
Eric Dumazetd83769a2015-04-21 18:32:24 -07002812 */
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002813void sk_forced_mem_schedule(struct sock *sk, int size)
Eric Dumazetd83769a2015-04-21 18:32:24 -07002814{
Johannes Weinere8056052016-01-14 15:21:14 -08002815 int amt;
Eric Dumazetd83769a2015-04-21 18:32:24 -07002816
2817 if (size <= sk->sk_forward_alloc)
2818 return;
2819 amt = sk_mem_pages(size);
2820 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
Johannes Weinere8056052016-01-14 15:21:14 -08002821 sk_memory_allocated_add(sk, amt);
2822
Johannes Weinerbaac50b2016-01-14 15:21:17 -08002823 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2824 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
Eric Dumazetd83769a2015-04-21 18:32:24 -07002825}
2826
Eric Dumazet845704a2015-04-23 10:42:39 -07002827/* Send a FIN. The caller locks the socket for us.
2828 * We should try to send a FIN packet really hard, but eventually give up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 */
2830void tcp_send_fin(struct sock *sk)
2831{
Eric Dumazet845704a2015-04-23 10:42:39 -07002832 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002833 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002834
Eric Dumazet845704a2015-04-23 10:42:39 -07002835 /* Optimization, tack on the FIN if we have one skb in write queue and
2836 * this skb was not yet sent, or we are under memory pressure.
2837 * Note: in the latter case, FIN packet will be sent after a timeout,
2838 * as TCP stack thinks it has already been transmitted.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002840 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
Eric Dumazet845704a2015-04-23 10:42:39 -07002841coalesce:
2842 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2843 TCP_SKB_CB(tskb)->end_seq++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844 tp->write_seq++;
Eric Dumazet845704a2015-04-23 10:42:39 -07002845 if (!tcp_send_head(sk)) {
2846 /* This means tskb was already sent.
2847 * Pretend we included the FIN on previous transmit.
2848 * We need to set tp->snd_nxt to the value it would have
2849 * if FIN had been sent. This is because retransmit path
2850 * does not change tp->snd_nxt.
2851 */
2852 tp->snd_nxt++;
2853 return;
2854 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 } else {
Eric Dumazet845704a2015-04-23 10:42:39 -07002856 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2857 if (unlikely(!skb)) {
2858 if (tskb)
2859 goto coalesce;
2860 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 }
Eric Dumazetd83769a2015-04-21 18:32:24 -07002862 skb_reserve(skb, MAX_TCP_HEADER);
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002863 sk_forced_mem_schedule(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002865 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002866 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 tcp_queue_skb(sk, skb);
2868 }
Eric Dumazet845704a2015-04-23 10:42:39 -07002869 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870}
2871
2872/* We get here when a process closes a file descriptor (either due to
2873 * an explicit close() or as a byproduct of exit()'ing) and there
2874 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002875 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 */
Al Virodd0fc662005-10-07 07:46:04 +01002877void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879 struct sk_buff *skb;
2880
2881 /* NOTE: No TCP options attached and we never retransmit this. */
2882 skb = alloc_skb(MAX_TCP_HEADER, priority);
2883 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002884 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 return;
2886 }
2887
2888 /* Reserve space for headers and prepare control bits. */
2889 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002890 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002891 TCPHDR_ACK | TCPHDR_RST);
Eric Dumazet675ee232015-09-23 14:00:21 -07002892 skb_mstamp_get(&skb->skb_mstamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 /* Send it off. */
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002894 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002895 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002896
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002897 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898}
2899
Andi Kleen67edfef2009-07-21 23:00:40 +00002900/* Send a crossed SYN-ACK during socket establishment.
2901 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002902 * a SYN packet that crossed the incoming SYN that caused this routine
2903 * to get called. If this assumption fails then the initial rcv_wnd
2904 * and rcv_wscale values will not be correct.
2905 */
2906int tcp_send_synack(struct sock *sk)
2907{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002908 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909
David S. Millerfe067e82007-03-07 12:12:44 -08002910 skb = tcp_write_queue_head(sk);
Ian Morris51456b22015-04-03 09:17:26 +01002911 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00002912 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 return -EFAULT;
2914 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04002915 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 if (skb_cloned(skb)) {
2917 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
Ian Morris51456b22015-04-03 09:17:26 +01002918 if (!nskb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002920 tcp_unlink_write_queue(skb, sk);
Eric Dumazetf4a775d2014-09-22 16:29:32 -07002921 __skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002922 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002923 sk_wmem_free_skb(sk, skb);
2924 sk->sk_wmem_queued += nskb->truesize;
2925 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926 skb = nskb;
2927 }
2928
Eric Dumazet4de075e2011-09-27 13:25:05 -04002929 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Florian Westphal735d3832014-09-29 13:08:30 +02002930 tcp_ecn_send_synack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002932 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933}
2934
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002935/**
2936 * tcp_make_synack - Prepare a SYN-ACK.
2937 * sk: listener socket
2938 * dst: dst entry attached to the SYNACK
2939 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002940 *
2941 * Allocate one skb and build a SYNACK packet.
2942 * @dst is consumed : Caller should not use it again.
2943 */
Eric Dumazet5d062de2015-09-25 07:39:19 -07002944struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002945 struct request_sock *req,
Eric Dumazetca6fb062015-10-02 11:43:35 -07002946 struct tcp_fastopen_cookie *foc,
2947 bool attach_req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002949 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet5d062de2015-09-25 07:39:19 -07002950 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet80f03e22015-03-24 15:58:52 -07002951 struct tcp_md5sig_key *md5 = NULL;
Eric Dumazet5d062de2015-09-25 07:39:19 -07002952 struct tcp_out_options opts;
2953 struct sk_buff *skb;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002954 int tcp_header_size;
Eric Dumazet5d062de2015-09-25 07:39:19 -07002955 struct tcphdr *th;
2956 u16 user_mss;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002957 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958
Eric Dumazetca6fb062015-10-02 11:43:35 -07002959 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002960 if (unlikely(!skb)) {
2961 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002963 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 /* Reserve space for headers. */
2965 skb_reserve(skb, MAX_TCP_HEADER);
2966
Eric Dumazetca6fb062015-10-02 11:43:35 -07002967 if (attach_req) {
Eric Dumazet9e17f8a2015-11-01 15:36:55 -08002968 skb_set_owner_w(skb, req_to_sk(req));
Eric Dumazetca6fb062015-10-02 11:43:35 -07002969 } else {
2970 /* sk is a const pointer, because we want to express multiple
2971 * cpu might call us concurrently.
2972 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2973 */
2974 skb_set_owner_w(skb, (struct sock *)sk);
2975 }
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002976 skb_dst_set(skb, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977
David S. Miller0dbaee32010-12-13 12:52:14 -08002978 mss = dst_metric_advmss(dst);
Eric Dumazet5d062de2015-09-25 07:39:19 -07002979 user_mss = READ_ONCE(tp->rx_opt.user_mss);
2980 if (user_mss && user_mss < mss)
2981 mss = user_mss;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002982
Adam Langley33ad7982008-07-19 00:04:31 -07002983 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002984#ifdef CONFIG_SYN_COOKIES
2985 if (unlikely(req->cookie_ts))
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002986 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002987 else
2988#endif
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002989 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazet80f03e22015-03-24 15:58:52 -07002990
2991#ifdef CONFIG_TCP_MD5SIG
2992 rcu_read_lock();
Eric Dumazetfd3a1542015-03-24 15:58:56 -07002993 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
Eric Dumazet80f03e22015-03-24 15:58:52 -07002994#endif
Eric Dumazet58d607d2015-09-15 15:24:20 -07002995 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
Eric Dumazet37bfbdd2015-09-25 07:39:17 -07002996 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
2997 sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002998
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002999 skb_push(skb, tcp_header_size);
3000 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003002 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 memset(th, 0, sizeof(struct tcphdr));
3004 th->syn = 1;
3005 th->ack = 1;
Eric Dumazet6ac705b2015-09-25 07:39:18 -07003006 tcp_ecn_make_synack(req, th);
Eric Dumazetb44084c2013-10-10 00:04:37 -07003007 th->source = htons(ireq->ir_num);
Eric Dumazet634fb9792013-10-09 15:21:29 -07003008 th->dest = ireq->ir_rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08003009 /* Setting of flags are superfluous here for callers (and ECE is
3010 * not even correctly set)
3011 */
3012 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00003013 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faade2009-12-02 18:25:27 +00003014
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Jerry Chu83368862012-08-31 12:29:12 +00003016 /* XXX data is queued and acked as is. No buffer/window check */
3017 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018
3019 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Eric Dumazeted53d0a2015-10-08 19:33:23 -07003020 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
Eric Dumazet5d062de2015-09-25 07:39:19 -07003021 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 th->doff = (tcp_header_size >> 2);
Eric Dumazeta0b84862014-03-26 09:57:19 -07003023 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003024
3025#ifdef CONFIG_TCP_MD5SIG
3026 /* Okay, we have all we need - do the md5 hash if needed */
Eric Dumazet80f03e22015-03-24 15:58:52 -07003027 if (md5)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003028 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07003029 md5, req_to_sk(req), skb);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003030 rcu_read_unlock();
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003031#endif
3032
Eric Dumazetb50edd72015-04-09 13:31:56 -07003033 /* Do not fool tcpdump (if any), clean our debris */
3034 skb->tstamp.tv64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 return skb;
3036}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003037EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038
Daniel Borkmann81164412015-01-05 23:57:48 +01003039static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3040{
3041 struct inet_connection_sock *icsk = inet_csk(sk);
3042 const struct tcp_congestion_ops *ca;
3043 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3044
3045 if (ca_key == TCP_CA_UNSPEC)
3046 return;
3047
3048 rcu_read_lock();
3049 ca = tcp_ca_find_key(ca_key);
3050 if (likely(ca && try_module_get(ca->owner))) {
3051 module_put(icsk->icsk_ca_ops->owner);
3052 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3053 icsk->icsk_ca_ops = ca;
3054 }
3055 rcu_read_unlock();
3056}
3057
Andi Kleen67edfef2009-07-21 23:00:40 +00003058/* Do all connect socket setups that can be done AF independent. */
stephen hemmingerf7e56a72013-12-29 11:39:51 -08003059static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003061 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062 struct tcp_sock *tp = tcp_sk(sk);
3063 __u8 rcv_wscale;
3064
3065 /* We'll fix this up when we get a response from the other end.
3066 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3067 */
3068 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08003069 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003071#ifdef CONFIG_TCP_MD5SIG
Ian Morris00db4122015-04-03 09:17:27 +01003072 if (tp->af_specific->md5_lookup(sk, sk))
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003073 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3074#endif
3075
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076 /* If user gave his TCP_MAXSEG, record it to clamp */
3077 if (tp->rx_opt.user_mss)
3078 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3079 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08003080 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 tcp_sync_mss(sk, dst_mtu(dst));
3082
Daniel Borkmann81164412015-01-05 23:57:48 +01003083 tcp_ca_dst_init(sk, dst);
3084
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 if (!tp->window_clamp)
3086 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08003087 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003088 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3089 tp->advmss = tp->rx_opt.user_mss;
3090
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00003093 /* limit the window selection if the user enforce a smaller rx buffer */
3094 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3095 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3096 tp->window_clamp = tcp_full_space(sk);
3097
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 tcp_select_initial_window(tcp_full_space(sk),
3099 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3100 &tp->rcv_wnd,
3101 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08003102 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00003103 &rcv_wscale,
3104 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105
3106 tp->rx_opt.rcv_wscale = rcv_wscale;
3107 tp->rcv_ssthresh = tp->rcv_wnd;
3108
3109 sk->sk_err = 0;
3110 sock_reset_flag(sk, SOCK_DONE);
3111 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003112 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 tp->snd_una = tp->write_seq;
3114 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003115 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00003116 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003117
3118 if (likely(!tp->repair))
3119 tp->rcv_nxt = 0;
Andrew Vaginc7781a62013-08-27 12:20:40 +04003120 else
3121 tp->rcv_tstamp = tcp_time_stamp;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003122 tp->rcv_wup = tp->rcv_nxt;
3123 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003125 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3126 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 tcp_clear_retrans(tp);
3128}
3129
Yuchung Cheng783237e2012-07-19 06:43:07 +00003130static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3131{
3132 struct tcp_sock *tp = tcp_sk(sk);
3133 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3134
3135 tcb->end_seq += skb->len;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07003136 __skb_header_release(skb);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003137 __tcp_add_write_queue_tail(sk, skb);
3138 sk->sk_wmem_queued += skb->truesize;
3139 sk_mem_charge(sk, skb->truesize);
3140 tp->write_seq = tcb->end_seq;
3141 tp->packets_out += tcp_skb_pcount(skb);
3142}
3143
3144/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3145 * queue a data-only packet after the regular SYN, such that regular SYNs
3146 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3147 * only the SYN sequence, the data are retransmitted in the first ACK.
3148 * If cookie is not cached or other error occurs, falls back to send a
3149 * regular SYN with Fast Open cookie request option.
3150 */
3151static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3152{
3153 struct tcp_sock *tp = tcp_sk(sk);
3154 struct tcp_fastopen_request *fo = tp->fastopen_req;
Eric Dumazet07e100f2015-12-16 13:53:10 -08003155 int syn_loss = 0, space, err = 0;
Yuchung Chengaab48742012-07-19 06:43:10 +00003156 unsigned long last_syn_loss = 0;
Eric Dumazet355a9012014-11-17 23:06:20 -08003157 struct sk_buff *syn_data;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003158
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003159 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00003160 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3161 &syn_loss, &last_syn_loss);
3162 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3163 if (syn_loss > 1 &&
3164 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3165 fo->cookie.len = -1;
3166 goto fallback;
3167 }
3168
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003169 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3170 fo->cookie.len = -1;
3171 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003172 goto fallback;
3173
3174 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3175 * user-MSS. Reserve maximum option space for middleboxes that add
3176 * private TCP options. The cost is reduced data space in SYN :(
3177 */
3178 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3179 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00003180 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00003181 MAX_TCP_OPTION_SPACE;
3182
Eric Dumazetf5ddcbb2014-02-20 10:09:18 -08003183 space = min_t(size_t, space, fo->size);
3184
3185 /* limit to order-0 allocations */
3186 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3187
Eric Dumazeteb934472015-05-19 13:26:55 -07003188 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
Eric Dumazet355a9012014-11-17 23:06:20 -08003189 if (!syn_data)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003190 goto fallback;
Eric Dumazet355a9012014-11-17 23:06:20 -08003191 syn_data->ip_summed = CHECKSUM_PARTIAL;
3192 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
Eric Dumazet07e100f2015-12-16 13:53:10 -08003193 if (space) {
3194 int copied = copy_from_iter(skb_put(syn_data, space), space,
3195 &fo->data->msg_iter);
3196 if (unlikely(!copied)) {
3197 kfree_skb(syn_data);
3198 goto fallback;
3199 }
3200 if (copied != space) {
3201 skb_trim(syn_data, copied);
3202 space = copied;
3203 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003204 }
Eric Dumazet355a9012014-11-17 23:06:20 -08003205 /* No more data pending in inet_wait_for_connect() */
3206 if (space == fo->size)
3207 fo->data = NULL;
3208 fo->copied = space;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003209
Eric Dumazet355a9012014-11-17 23:06:20 -08003210 tcp_connect_queue_skb(sk, syn_data);
3211
3212 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3213
3214 syn->skb_mstamp = syn_data->skb_mstamp;
3215
3216 /* Now full SYN+DATA was cloned and sent (or not),
3217 * remove the SYN from the original skb (syn_data)
3218 * we keep in write queue in case of a retransmit, as we
3219 * also have the SYN packet (with no data) in the same queue.
Eric Dumazet431a9122014-03-09 17:36:02 -07003220 */
Eric Dumazet355a9012014-11-17 23:06:20 -08003221 TCP_SKB_CB(syn_data)->seq++;
3222 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3223 if (!err) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003224 tp->syn_data = (fo->copied > 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08003225 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003226 goto done;
3227 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003228
3229fallback:
3230 /* Send a regular SYN with Fast Open cookie request option */
3231 if (fo->cookie.len > 0)
3232 fo->cookie.len = 0;
3233 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3234 if (err)
3235 tp->syn_fastopen = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003236done:
3237 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3238 return err;
3239}
3240
Andi Kleen67edfef2009-07-21 23:00:40 +00003241/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242int tcp_connect(struct sock *sk)
3243{
3244 struct tcp_sock *tp = tcp_sk(sk);
3245 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00003246 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003247
3248 tcp_connect_init(sk);
3249
Andrey Vagin2b916472012-11-22 01:13:58 +00003250 if (unlikely(tp->repair)) {
3251 tcp_finish_connect(sk, NULL);
3252 return 0;
3253 }
3254
Eric Dumazeteb934472015-05-19 13:26:55 -07003255 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
Eric Dumazet355a9012014-11-17 23:06:20 -08003256 if (unlikely(!buff))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257 return -ENOBUFS;
3258
Changli Gaoa3433f32010-06-12 14:01:43 +00003259 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003260 tp->retrans_stamp = tcp_time_stamp;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003261 tcp_connect_queue_skb(sk, buff);
Florian Westphal735d3832014-09-29 13:08:30 +02003262 tcp_ecn_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263
Yuchung Cheng783237e2012-07-19 06:43:07 +00003264 /* Send off SYN; include data in Fast Open. */
3265 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3266 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00003267 if (err == -ECONNREFUSED)
3268 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07003269
3270 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3271 * in order to make this packet get counted in tcpOutSegs.
3272 */
3273 tp->snd_nxt = tp->write_seq;
3274 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003275 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276
3277 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003278 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3279 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280 return 0;
3281}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003282EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283
3284/* Send out a delayed ack, the caller does the policy checking
3285 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3286 * for details.
3287 */
3288void tcp_send_delayed_ack(struct sock *sk)
3289{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003290 struct inet_connection_sock *icsk = inet_csk(sk);
3291 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 unsigned long timeout;
3293
Florian Westphal98900922014-09-26 22:37:35 +02003294 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3295
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003297 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003298 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003300 if (icsk->icsk_ack.pingpong ||
3301 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 max_ato = TCP_DELACK_MAX;
3303
3304 /* Slow path, intersegment interval is "high". */
3305
3306 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003307 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 * directly.
3309 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003310 if (tp->srtt_us) {
3311 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3312 TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313
3314 if (rtt < max_ato)
3315 max_ato = rtt;
3316 }
3317
3318 ato = min(ato, max_ato);
3319 }
3320
3321 /* Stay within the limit we were given */
3322 timeout = jiffies + ato;
3323
3324 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003325 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 /* If delack timer was blocked or is about to expire,
3327 * send ACK now.
3328 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003329 if (icsk->icsk_ack.blocked ||
3330 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 tcp_send_ack(sk);
3332 return;
3333 }
3334
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003335 if (!time_before(timeout, icsk->icsk_ack.timeout))
3336 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003338 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3339 icsk->icsk_ack.timeout = timeout;
3340 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341}
3342
3343/* This routine sends an ack and also updates the window. */
3344void tcp_send_ack(struct sock *sk)
3345{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003346 struct sk_buff *buff;
3347
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003349 if (sk->sk_state == TCP_CLOSE)
3350 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351
Florian Westphal98900922014-09-26 22:37:35 +02003352 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3353
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003354 /* We are not putting this on the write queue, so
3355 * tcp_transmit_skb() will set the ownership to this
3356 * sock.
3357 */
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003358 buff = alloc_skb(MAX_TCP_HEADER,
3359 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3360 if (unlikely(!buff)) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003361 inet_csk_schedule_ack(sk);
3362 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3363 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3364 TCP_DELACK_MAX, TCP_RTO_MAX);
3365 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003367
3368 /* Reserve space for headers and prepare control bits. */
3369 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003370 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003371
Eric Dumazet98781962015-02-03 18:31:53 -08003372 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3373 * too much.
3374 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3375 * We also avoid tcp_wfree() overhead (cache line miss accessing
3376 * tp->tsq_flags) by using regular sock_wfree()
3377 */
3378 skb_set_tcp_pure_ack(buff);
3379
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003380 /* Send it off, this clears delayed acks for us. */
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003381 skb_mstamp_get(&buff->skb_mstamp);
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003382 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383}
Daniel Borkmanne3118e82014-09-26 22:37:36 +02003384EXPORT_SYMBOL_GPL(tcp_send_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385
3386/* This routine sends a packet with an out of date sequence
3387 * number. It assumes the other end will try to ack it.
3388 *
3389 * Question: what should we make while urgent mode?
3390 * 4.4BSD forces sending single byte of data. We cannot send
3391 * out of window data, because we have SND.NXT==SND.MAX...
3392 *
3393 * Current solution: to send TWO zero-length segments in urgent mode:
3394 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3395 * out-of-date with SND.UNA-1 to probe window.
3396 */
Eric Dumazete520af42015-05-06 14:26:25 -07003397static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398{
3399 struct tcp_sock *tp = tcp_sk(sk);
3400 struct sk_buff *skb;
3401
3402 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003403 skb = alloc_skb(MAX_TCP_HEADER,
3404 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
Ian Morris51456b22015-04-03 09:17:26 +01003405 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003406 return -1;
3407
3408 /* Reserve space for headers and set control bits. */
3409 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003410 /* Use a previous sequence. This should cause the other
3411 * end to send an ack. Don't queue or clone SKB, just
3412 * send it.
3413 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003414 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003415 skb_mstamp_get(&skb->skb_mstamp);
Renato Westphale2e80092015-10-19 18:51:34 -02003416 NET_INC_STATS(sock_net(sk), mib);
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003417 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418}
3419
Pavel Emelyanovee995282012-04-19 03:40:39 +00003420void tcp_send_window_probe(struct sock *sk)
3421{
3422 if (sk->sk_state == TCP_ESTABLISHED) {
3423 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
Eric Dumazete520af42015-05-06 14:26:25 -07003424 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
Pavel Emelyanovee995282012-04-19 03:40:39 +00003425 }
3426}
3427
Andi Kleen67edfef2009-07-21 23:00:40 +00003428/* Initiate keepalive or window probe from timer. */
Eric Dumazete520af42015-05-06 14:26:25 -07003429int tcp_write_wakeup(struct sock *sk, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003431 struct tcp_sock *tp = tcp_sk(sk);
3432 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003434 if (sk->sk_state == TCP_CLOSE)
3435 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436
Ian Morris00db4122015-04-03 09:17:27 +01003437 skb = tcp_send_head(sk);
3438 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003439 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003440 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003441 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003443 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3444 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003446 /* We are probing the opening of a window
3447 * but the window size is != 0
3448 * must have been a result SWS avoidance ( sender )
3449 */
3450 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3451 skb->len > mss) {
3452 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003453 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Octavian Purdila6cc55e02014-06-06 17:32:37 +03003454 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003455 return -1;
3456 } else if (!tcp_skb_pcount(skb))
Eric Dumazet5bbb4322015-06-11 09:15:17 -07003457 tcp_set_skb_tso_segs(skb, mss);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003458
Eric Dumazet4de075e2011-09-27 13:25:05 -04003459 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003460 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3461 if (!err)
3462 tcp_event_new_data_sent(sk, skb);
3463 return err;
3464 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003465 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Eric Dumazete520af42015-05-06 14:26:25 -07003466 tcp_xmit_probe_skb(sk, 1, mib);
3467 return tcp_xmit_probe_skb(sk, 0, mib);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469}
3470
3471/* A window probe timeout has occurred. If window is not closed send
3472 * a partial packet else a zero probe.
3473 */
3474void tcp_send_probe0(struct sock *sk)
3475{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003476 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisovc6214a92016-02-03 09:46:54 +02003478 struct net *net = sock_net(sk);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003479 unsigned long probe_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 int err;
3481
Eric Dumazete520af42015-05-06 14:26:25 -07003482 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
David S. Millerfe067e82007-03-07 12:12:44 -08003484 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003486 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003487 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 return;
3489 }
3490
3491 if (err <= 0) {
Nikolay Borisovc6214a92016-02-03 09:46:54 +02003492 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003493 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003494 icsk->icsk_probes_out++;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003495 probe_max = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 } else {
3497 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003498 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 * Let local senders to fight for local resources.
3500 *
3501 * Use accumulated backoff yet.
3502 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003503 if (!icsk->icsk_probes_out)
3504 icsk->icsk_probes_out = 1;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003505 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506 }
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003507 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003508 tcp_probe0_when(sk, probe_max),
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003509 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510}
Octavian Purdila5db92c92014-06-25 17:09:59 +03003511
Eric Dumazetea3bea32015-09-25 07:39:23 -07003512int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
Octavian Purdila5db92c92014-06-25 17:09:59 +03003513{
3514 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3515 struct flowi fl;
3516 int res;
3517
Eric Dumazet58d607d2015-09-15 15:24:20 -07003518 tcp_rsk(req)->txhash = net_tx_rndhash();
Eric Dumazetdc6ef6b2015-10-16 13:00:01 -07003519 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
Octavian Purdila5db92c92014-06-25 17:09:59 +03003520 if (!res) {
3521 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3522 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3523 }
3524 return res;
3525}
3526EXPORT_SYMBOL(tcp_rtx_synack);