blob: a369e8a70b2c775bfee94d7f329ee892c2cdc895 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
Joe Perches91df42b2012-05-15 14:11:54 +000037#define pr_fmt(fmt) "TCP: " fmt
38
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <net/tcp.h>
40
41#include <linux/compiler.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090042#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070046int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Jianjun Kong09cb1052008-11-03 00:27:11 -080048/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080049 * interpret the window field as a signed quantity.
50 */
Brian Haleyab32ea52006-09-22 14:15:41 -070051int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080052
Eric Dumazet46d3cea2012-07-11 05:50:31 +000053/* Default TSQ limit of two TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
55
Linus Torvalds1da177e2005-04-16 15:20:36 -070056/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
Brian Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
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 Dumazetc9bee3b72013-07-22 20:27:07 -070065unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
67
Eric Dumazet46d3cea2012-07-11 05:50:31 +000068static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000070
Andi Kleen67edfef2009-07-21 23:00:40 +000071/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040072static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070073{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000074 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070075 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080076 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070077
David S. Millerfe067e82007-03-07 12:12:44 -080078 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080080
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080081 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000082 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000083 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000084 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000085 }
Yuchung Chengf19c29e2014-03-03 12:31:36 -080086
David S. Millerf7324ac2014-03-06 15:03:17 -050087 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -070089}
90
91/* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
96 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040097static inline __u32 tcp_acceptable_seq(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070098{
Eric Dumazetcf533ea2011-10-21 05:22:42 -040099 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700100
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102 return tp->snd_nxt;
103 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800104 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400124 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 int mss = tp->advmss;
126
David S. Miller0dbaee32010-12-13 12:52:14 -0800127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400141static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700143 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
147
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 restart_cwnd = min(restart_cwnd, cwnd);
152
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 cwnd >>= 1;
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
158}
159
Andi Kleen67edfef2009-07-21 23:00:40 +0000160/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800161static void tcp_event_data_sent(struct tcp_sock *tp,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400162 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
Cong Wangbcefe172013-06-15 09:39:18 +0800166 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167
David S. Miller35089bb2006-06-13 22:33:04 -0700168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
172 tp->lsndtime = now;
173
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
176 */
Cong Wangbcefe172013-06-15 09:39:18 +0800177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180}
181
Andi Kleen67edfef2009-07-21 23:00:40 +0000182/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800183static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
Yuchung Cheng85f16522013-06-11 15:35:32 -0700189
190u32 tcp_default_init_rwnd(u32 mss)
191{
192 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800193 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
196 */
197 u32 init_rwnd = TCP_INIT_CWND * 2;
198
199 if (mss > 1460)
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
201 return init_rwnd;
202}
203
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204/* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
210 */
211void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000213 int wscale_ok, __u8 *rcv_wscale,
214 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215{
216 unsigned int space = (__space < 0 ? 0 : __space);
217
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (65535 << 14);
221 space = min(*window_clamp, space);
222
223 /* Quantize space offering to a multiple of mss if possible. */
224 if (space > mss)
225 space = (space / mss) * mss;
226
227 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 */
Rick Jones15d99e02006-03-20 22:40:29 -0800235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
237 else
238 (*rcv_wnd) = space;
239
Linus Torvalds1da177e2005-04-16 15:20:36 -0700240 (*rcv_wscale) = 0;
241 if (wscale_ok) {
242 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900243 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 */
245 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700246 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 while (space > 65535 && (*rcv_wscale) < 14) {
248 space >>= 1;
249 (*rcv_wscale)++;
250 }
251 }
252
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800253 if (mss > (1 << *rcv_wscale)) {
Yuchung Cheng85f16522013-06-11 15:35:32 -0700254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 }
258
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
261}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000262EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
264/* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
267 * frame.
268 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800269static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270{
271 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphal8e165e22014-02-25 14:34:32 +0100272 u32 old_win = tp->rcv_wnd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273 u32 cur_win = tcp_receive_window(tp);
274 u32 new_win = __tcp_select_window(sk);
275
276 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800277 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 /* Danger Will Robinson!
279 * Don't update rcv_wup/rcv_wnd here or else
280 * we will not be able to advertise a zero
281 * window in time. --DaveM
282 *
283 * Relax Will Robinson.
284 */
Florian Westphal8e165e22014-02-25 14:34:32 +0100285 if (new_win == 0)
286 NET_INC_STATS(sock_net(sk),
287 LINUX_MIB_TCPWANTZEROWINDOWADV);
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700288 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 }
290 tp->rcv_wnd = new_win;
291 tp->rcv_wup = tp->rcv_nxt;
292
293 /* Make sure we do not exceed the maximum possible
294 * scaled window.
295 */
Rick Jones15d99e02006-03-20 22:40:29 -0800296 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700297 new_win = min(new_win, MAX_TCP_WINDOW);
298 else
299 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
300
301 /* RFC1323 scaling applied */
302 new_win >>= tp->rx_opt.rcv_wscale;
303
304 /* If we advertise zero window, disable fast path. */
Florian Westphal8e165e22014-02-25 14:34:32 +0100305 if (new_win == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 tp->pred_flags = 0;
Florian Westphal8e165e22014-02-25 14:34:32 +0100307 if (old_win)
308 NET_INC_STATS(sock_net(sk),
309 LINUX_MIB_TCPTOZEROWINDOWADV);
310 } else if (old_win == 0) {
311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
312 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
314 return new_win;
315}
316
Andi Kleen67edfef2009-07-21 23:00:40 +0000317/* Packet ECN state for a SYN-ACK */
Florian Westphal735d3832014-09-29 13:08:30 +0200318static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700319{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200320 const struct tcp_sock *tp = tcp_sk(sk);
321
Eric Dumazet4de075e2011-09-27 13:25:05 -0400322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800323 if (!(tp->ecn_flags & TCP_ECN_OK))
Eric Dumazet4de075e2011-09-27 13:25:05 -0400324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200325 else if (tcp_ca_needs_ecn(sk))
326 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700327}
328
Andi Kleen67edfef2009-07-21 23:00:40 +0000329/* Packet ECN state for a SYN. */
Florian Westphal735d3832014-09-29 13:08:30 +0200330static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700331{
332 struct tcp_sock *tp = tcp_sk(sk);
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100333 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
334 tcp_ca_needs_ecn(sk);
335
336 if (!use_ecn) {
337 const struct dst_entry *dst = __sk_dst_get(sk);
338
339 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 use_ecn = true;
341 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700342
343 tp->ecn_flags = 0;
Florian Westphalf7b3bec2014-11-03 17:35:03 +0100344
345 if (use_ecn) {
Eric Dumazet4de075e2011-09-27 13:25:05 -0400346 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700347 tp->ecn_flags = TCP_ECN_OK;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200348 if (tcp_ca_needs_ecn(sk))
349 INET_ECN_xmit(sk);
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700350 }
351}
352
Florian Westphal735d3832014-09-29 13:08:30 +0200353static void
354tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200355 struct sock *sk)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700356{
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200357 if (inet_rsk(req)->ecn_ok) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700358 th->ece = 1;
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200359 if (tcp_ca_needs_ecn(sk))
360 INET_ECN_xmit(sk);
361 }
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700362}
363
Andi Kleen67edfef2009-07-21 23:00:40 +0000364/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 * be sent.
366 */
Florian Westphal735d3832014-09-29 13:08:30 +0200367static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700368 int tcp_header_len)
369{
370 struct tcp_sock *tp = tcp_sk(sk);
371
372 if (tp->ecn_flags & TCP_ECN_OK) {
373 /* Not-retransmitted data segment: set ECT and inject CWR. */
374 if (skb->len != tcp_header_len &&
375 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800377 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700378 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 tcp_hdr(skb)->cwr = 1;
380 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 }
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200382 } else if (!tcp_ca_needs_ecn(sk)) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700383 /* ACK or retransmitted segment: clear ECT|CE */
384 INET_ECN_dontxmit(sk);
385 }
386 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387 tcp_hdr(skb)->ece = 1;
388 }
389}
390
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800391/* Constructs common control bits of non-data skb. If SYN/FIN is present,
392 * auto increment end seqno.
393 */
394static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395{
Eric Dumazet7b7fc972013-12-05 22:31:30 -0800396 struct skb_shared_info *shinfo = skb_shinfo(skb);
397
David S. Miller2e8e18e2010-04-08 11:32:30 -0700398 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800399 skb->csum = 0;
400
Eric Dumazet4de075e2011-09-27 13:25:05 -0400401 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800402 TCP_SKB_CB(skb)->sacked = 0;
403
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700404 tcp_skb_pcount_set(skb, 1);
Eric Dumazet7b7fc972013-12-05 22:31:30 -0800405 shinfo->gso_size = 0;
406 shinfo->gso_type = 0;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800407
408 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000409 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800410 seq++;
411 TCP_SKB_CB(skb)->end_seq = seq;
412}
413
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000414static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700415{
416 return tp->snd_una != tp->snd_up;
417}
418
Adam Langley33ad7982008-07-19 00:04:31 -0700419#define OPTION_SACK_ADVERTISE (1 << 0)
420#define OPTION_TS (1 << 1)
421#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000422#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000423#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700424
425struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000426 u16 options; /* bit field of OPTION_* */
427 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700428 u8 ws; /* window scale, 0 to disable */
429 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000430 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000431 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000432 __u32 tsval, tsecr; /* need to include OPTION_TS */
433 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700434};
435
Andi Kleen67edfef2009-07-21 23:00:40 +0000436/* Write previously computed TCP options to the packet.
437 *
438 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700439 * TCP options, we learned this through the hard way, so be careful here.
440 * Luckily we can at least blame others for their non-compliance but from
stephen hemminger8e3bff92013-12-08 12:15:44 -0800441 * inter-operability perspective it seems that we're somewhat stuck with
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700442 * the ordering which we have been using if we want to keep working with
443 * those broken things (not that it currently hurts anybody as there isn't
444 * particular reason why the ordering would need to be changed).
445 *
446 * At least SACK_PERM as the first option is known to lead to a disaster
447 * (but it may well be that other scenarios fail similarly).
448 */
Adam Langley33ad7982008-07-19 00:04:31 -0700449static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000450 struct tcp_out_options *opts)
451{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000452 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000453
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000454 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000455 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
456 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000457 /* overload cookie hash location */
458 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700459 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800460 }
Adam Langley33ad7982008-07-19 00:04:31 -0700461
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700462 if (unlikely(opts->mss)) {
463 *ptr++ = htonl((TCPOPT_MSS << 24) |
464 (TCPOLEN_MSS << 16) |
465 opts->mss);
466 }
467
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000468 if (likely(OPTION_TS & options)) {
469 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700470 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
471 (TCPOLEN_SACK_PERM << 16) |
472 (TCPOPT_TIMESTAMP << 8) |
473 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000474 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700475 } else {
476 *ptr++ = htonl((TCPOPT_NOP << 24) |
477 (TCPOPT_NOP << 16) |
478 (TCPOPT_TIMESTAMP << 8) |
479 TCPOLEN_TIMESTAMP);
480 }
481 *ptr++ = htonl(opts->tsval);
482 *ptr++ = htonl(opts->tsecr);
483 }
484
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000485 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700486 *ptr++ = htonl((TCPOPT_NOP << 24) |
487 (TCPOPT_NOP << 16) |
488 (TCPOPT_SACK_PERM << 8) |
489 TCPOLEN_SACK_PERM);
490 }
491
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000492 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700493 *ptr++ = htonl((TCPOPT_NOP << 24) |
494 (TCPOPT_WINDOW << 16) |
495 (TCPOLEN_WINDOW << 8) |
496 opts->ws);
497 }
498
499 if (unlikely(opts->num_sack_blocks)) {
500 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
501 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800502 int this_sack;
503
504 *ptr++ = htonl((TCPOPT_NOP << 24) |
505 (TCPOPT_NOP << 16) |
506 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700507 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800508 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800509
Adam Langley33ad7982008-07-19 00:04:31 -0700510 for (this_sack = 0; this_sack < opts->num_sack_blocks;
511 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800512 *ptr++ = htonl(sp[this_sack].start_seq);
513 *ptr++ = htonl(sp[this_sack].end_seq);
514 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800515
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000516 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800517 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000518
519 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
520 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
Daniel Lee7f9b8382015-04-06 14:37:26 -0700521 u8 *p = (u8 *)ptr;
522 u32 len; /* Fast Open option length */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000523
Daniel Lee7f9b8382015-04-06 14:37:26 -0700524 if (foc->exp) {
525 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
526 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
527 TCPOPT_FASTOPEN_MAGIC);
528 p += TCPOLEN_EXP_FASTOPEN_BASE;
529 } else {
530 len = TCPOLEN_FASTOPEN_BASE + foc->len;
531 *p++ = TCPOPT_FASTOPEN;
532 *p++ = len;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000533 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700534
535 memcpy(p, foc->val, foc->len);
536 if ((len & 3) == 2) {
537 p[foc->len] = TCPOPT_NOP;
538 p[foc->len + 1] = TCPOPT_NOP;
539 }
540 ptr += (len + 3) >> 2;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000541 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800542}
543
Andi Kleen67edfef2009-07-21 23:00:40 +0000544/* Compute TCP options for SYN packets. This is not the final
545 * network wire format yet.
546 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000547static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700548 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400549 struct tcp_md5sig_key **md5)
550{
Adam Langley33ad7982008-07-19 00:04:31 -0700551 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000552 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000553 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700554
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800555#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700556 *md5 = tp->af_specific->md5_lookup(sk, sk);
557 if (*md5) {
558 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000559 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800560 }
Adam Langley33ad7982008-07-19 00:04:31 -0700561#else
562 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800563#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700564
565 /* We always get an MSS option. The option bytes which will be seen in
566 * normal data packets should timestamps be used, must be in the MSS
567 * advertised. But we subtract them from tp->mss_cache so that
568 * calculations in tcp_sendmsg are simpler etc. So account for this
569 * fact here if necessary. If we don't do this correctly, as a
570 * receiver we won't recognize data packets as being full sized when we
571 * should, and thus we won't abide by the delayed ACK rules correctly.
572 * SACKs don't matter, we never delay an ACK when we have any of those
573 * going out. */
574 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000575 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700576
Ian Morris51456b22015-04-03 09:17:26 +0100577 if (likely(sysctl_tcp_timestamps && !*md5)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700578 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700579 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700580 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000581 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700582 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800583 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700584 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000585 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000586 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700587 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800588 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700589 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700590 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000591 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700592 }
593
Yuchung Cheng783237e2012-07-19 06:43:07 +0000594 if (fastopen && fastopen->cookie.len >= 0) {
Daniel Lee2646c832015-04-06 14:37:27 -0700595 u32 need = fastopen->cookie.len;
596
597 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
598 TCPOLEN_FASTOPEN_BASE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000599 need = (need + 3) & ~3U; /* Align to 32 bits */
600 if (remaining >= need) {
601 opts->options |= OPTION_FAST_OPEN_COOKIE;
602 opts->fastopen_cookie = &fastopen->cookie;
603 remaining -= need;
604 tp->syn_fastopen = 1;
Daniel Lee2646c832015-04-06 14:37:27 -0700605 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000606 }
607 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000608
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000609 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700610}
611
Andi Kleen67edfef2009-07-21 23:00:40 +0000612/* Set up TCP options for SYN-ACKs. */
Eric Dumazet95c96172012-04-15 05:58:06 +0000613static unsigned int tcp_synack_options(struct sock *sk,
Adam Langley33ad7982008-07-19 00:04:31 -0700614 struct request_sock *req,
Eric Dumazet95c96172012-04-15 05:58:06 +0000615 unsigned int mss, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700616 struct tcp_out_options *opts,
Eric Dumazet80f03e22015-03-24 15:58:52 -0700617 const struct tcp_md5sig_key *md5,
Jerry Chu83368862012-08-31 12:29:12 +0000618 struct tcp_fastopen_cookie *foc)
William Allen Simpson4957faade2009-12-02 18:25:27 +0000619{
Adam Langley33ad7982008-07-19 00:04:31 -0700620 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet95c96172012-04-15 05:58:06 +0000621 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Adam Langley33ad7982008-07-19 00:04:31 -0700622
623#ifdef CONFIG_TCP_MD5SIG
Eric Dumazet80f03e22015-03-24 15:58:52 -0700624 if (md5) {
Adam Langley33ad7982008-07-19 00:04:31 -0700625 opts->options |= OPTION_MD5;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000626 remaining -= TCPOLEN_MD5SIG_ALIGNED;
627
628 /* We can't fit any SACK blocks in a packet with MD5 + TS
629 * options. There was discussion about disabling SACK
630 * rather than TS in order to fit in better with old,
631 * buggy kernels, but that was deemed to be unnecessary.
632 */
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700633 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700634 }
Adam Langley33ad7982008-07-19 00:04:31 -0700635#endif
636
William Allen Simpson4957faade2009-12-02 18:25:27 +0000637 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700638 opts->mss = mss;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000639 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700640
641 if (likely(ireq->wscale_ok)) {
642 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000643 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000644 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700645 }
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700646 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700647 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700648 opts->tsval = tcp_skb_timestamp(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700649 opts->tsecr = req->ts_recent;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000650 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700651 }
652 if (likely(ireq->sack_ok)) {
653 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700654 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faade2009-12-02 18:25:27 +0000655 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700656 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700657 if (foc != NULL && foc->len >= 0) {
658 u32 need = foc->len;
659
660 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
661 TCPOLEN_FASTOPEN_BASE;
Jerry Chu83368862012-08-31 12:29:12 +0000662 need = (need + 3) & ~3U; /* Align to 32 bits */
663 if (remaining >= need) {
664 opts->options |= OPTION_FAST_OPEN_COOKIE;
665 opts->fastopen_cookie = foc;
666 remaining -= need;
667 }
668 }
William Allen Simpson4957faade2009-12-02 18:25:27 +0000669
William Allen Simpson4957faade2009-12-02 18:25:27 +0000670 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700671}
672
Andi Kleen67edfef2009-07-21 23:00:40 +0000673/* Compute TCP options for ESTABLISHED sockets. This is not the
674 * final wire format yet.
675 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000676static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700677 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400678 struct tcp_md5sig_key **md5)
679{
Adam Langley33ad7982008-07-19 00:04:31 -0700680 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000681 unsigned int size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000682 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700683
Andi Kleen5843ef42013-09-30 13:29:11 -0700684 opts->options = 0;
685
Adam Langley33ad7982008-07-19 00:04:31 -0700686#ifdef CONFIG_TCP_MD5SIG
687 *md5 = tp->af_specific->md5_lookup(sk, sk);
688 if (unlikely(*md5)) {
689 opts->options |= OPTION_MD5;
690 size += TCPOLEN_MD5SIG_ALIGNED;
691 }
692#else
693 *md5 = NULL;
694#endif
695
696 if (likely(tp->rx_opt.tstamp_ok)) {
697 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700698 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700699 opts->tsecr = tp->rx_opt.ts_recent;
700 size += TCPOLEN_TSTAMP_ALIGNED;
701 }
702
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000703 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
704 if (unlikely(eff_sacks)) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000705 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
Adam Langley33ad7982008-07-19 00:04:31 -0700706 opts->num_sack_blocks =
Eric Dumazet95c96172012-04-15 05:58:06 +0000707 min_t(unsigned int, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700708 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
709 TCPOLEN_SACK_PERBLOCK);
710 size += TCPOLEN_SACK_BASE_ALIGNED +
711 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
712 }
713
714 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800715}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000717
718/* TCP SMALL QUEUES (TSQ)
719 *
720 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
721 * to reduce RTT and bufferbloat.
722 * We do this using a special skb destructor (tcp_wfree).
723 *
724 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
725 * needs to be reallocated in a driver.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800726 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000727 *
728 * Since transmit from skb destructor is forbidden, we use a tasklet
729 * to process all sockets that eventually need to send more skbs.
730 * We use one tasklet per cpu, with its own queue of sockets.
731 */
732struct tsq_tasklet {
733 struct tasklet_struct tasklet;
734 struct list_head head; /* queue of tcp sockets */
735};
736static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
737
Eric Dumazet6f458df2012-07-20 05:45:50 +0000738static void tcp_tsq_handler(struct sock *sk)
739{
740 if ((1 << sk->sk_state) &
741 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
742 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
John Ognessbf062002014-02-09 18:40:11 -0800743 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
744 0, GFP_ATOMIC);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000745}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000746/*
stephen hemminger8e3bff92013-12-08 12:15:44 -0800747 * One tasklet per cpu tries to send more skbs.
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000748 * We run in tasklet context but need to disable irqs when
stephen hemminger8e3bff92013-12-08 12:15:44 -0800749 * transferring tsq->head because tcp_wfree() might
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000750 * interrupt us (non NAPI drivers)
751 */
752static void tcp_tasklet_func(unsigned long data)
753{
754 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
755 LIST_HEAD(list);
756 unsigned long flags;
757 struct list_head *q, *n;
758 struct tcp_sock *tp;
759 struct sock *sk;
760
761 local_irq_save(flags);
762 list_splice_init(&tsq->head, &list);
763 local_irq_restore(flags);
764
765 list_for_each_safe(q, n, &list) {
766 tp = list_entry(q, struct tcp_sock, tsq_node);
767 list_del(&tp->tsq_node);
768
769 sk = (struct sock *)tp;
770 bh_lock_sock(sk);
771
772 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000773 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000774 } else {
775 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000776 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000777 }
778 bh_unlock_sock(sk);
779
780 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
781 sk_free(sk);
782 }
783}
784
Eric Dumazet6f458df2012-07-20 05:45:50 +0000785#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
786 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200787 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
788 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000789/**
790 * tcp_release_cb - tcp release_sock() callback
791 * @sk: socket
792 *
793 * called from release_sock() to perform protocol dependent
794 * actions before socket release.
795 */
796void tcp_release_cb(struct sock *sk)
797{
798 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000799 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000800
Eric Dumazet6f458df2012-07-20 05:45:50 +0000801 /* perform an atomic operation only if at least one flag is set */
802 do {
803 flags = tp->tsq_flags;
804 if (!(flags & TCP_DEFERRED_ALL))
805 return;
806 nflags = flags & ~TCP_DEFERRED_ALL;
807 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
808
809 if (flags & (1UL << TCP_TSQ_DEFERRED))
810 tcp_tsq_handler(sk);
811
Eric Dumazetc3f9b012014-03-10 09:50:11 -0700812 /* Here begins the tricky part :
813 * We are called from release_sock() with :
814 * 1) BH disabled
815 * 2) sk_lock.slock spinlock held
816 * 3) socket owned by us (sk->sk_lock.owned == 1)
817 *
818 * But following code is meant to be called from BH handlers,
819 * so we should keep BH disabled, but early release socket ownership
820 */
821 sock_release_ownership(sk);
822
Eric Dumazet144d56e2012-08-20 00:22:46 +0000823 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000824 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000825 __sock_put(sk);
826 }
827 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000828 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000829 __sock_put(sk);
830 }
831 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Neal Cardwell4fab9072014-08-14 12:40:05 -0400832 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000833 __sock_put(sk);
834 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000835}
836EXPORT_SYMBOL(tcp_release_cb);
837
838void __init tcp_tasklet_init(void)
839{
840 int i;
841
842 for_each_possible_cpu(i) {
843 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
844
845 INIT_LIST_HEAD(&tsq->head);
846 tasklet_init(&tsq->tasklet,
847 tcp_tasklet_func,
848 (unsigned long)tsq);
849 }
850}
851
852/*
853 * Write buffer destructor automatically called from kfree_skb.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800854 * We can't xmit new skbs from this context, as we might already
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000855 * hold qdisc lock.
856 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000857void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000858{
859 struct sock *sk = skb->sk;
860 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700861 int wmem;
862
863 /* Keep one reference on sk_wmem_alloc.
864 * Will be released by sk_free() from here or tcp_tasklet_func()
865 */
866 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
867
868 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
869 * Wait until our queues (qdisc + devices) are drained.
870 * This gives :
871 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
872 * - chance for incoming ACK (processed by another cpu maybe)
873 * to migrate this flow (skb->ooo_okay will be eventually set)
874 */
875 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
876 goto out;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000877
878 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
879 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
880 unsigned long flags;
881 struct tsq_tasklet *tsq;
882
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000883 /* queue this socket to tasklet queue */
884 local_irq_save(flags);
Christoph Lameter903ceff2014-08-17 12:30:35 -0500885 tsq = this_cpu_ptr(&tsq_tasklet);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000886 list_add(&tp->tsq_node, &tsq->head);
887 tasklet_schedule(&tsq->tasklet);
888 local_irq_restore(flags);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700889 return;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000890 }
Eric Dumazet9b462d02014-10-13 06:27:47 -0700891out:
892 sk_free(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000893}
894
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895/* This routine actually transmits TCP packets queued in by
896 * tcp_do_sendmsg(). This is used by both the initial
897 * transmission and possible later retransmissions.
898 * All SKB's seen here are completely headerless. It is our
899 * job to build the TCP header, and pass the packet down to
900 * IP so it can do the same plus pass the packet off to the
901 * device.
902 *
903 * We are working here with either a clone of the original
904 * SKB, or a fresh unique copy made by the retransmit engine.
905 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800906static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
907 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800909 const struct inet_connection_sock *icsk = inet_csk(sk);
910 struct inet_sock *inet;
911 struct tcp_sock *tp;
912 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700913 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000914 unsigned int tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800915 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800916 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800917 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800919 BUG_ON(!skb || !tcp_skb_pcount(skb));
920
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700921 if (clone_it) {
Eric Dumazet740b0f12014-02-26 14:02:48 -0800922 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700923
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800924 if (unlikely(skb_cloned(skb)))
925 skb = pskb_copy(skb, gfp_mask);
926 else
927 skb = skb_clone(skb, gfp_mask);
928 if (unlikely(!skb))
929 return -ENOBUFS;
930 }
931
932 inet = inet_sk(sk);
933 tp = tcp_sk(sk);
934 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700935 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
Eric Dumazet4de075e2011-09-27 13:25:05 -0400937 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700938 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
939 else
940 tcp_options_size = tcp_established_options(sk, skb, &opts,
941 &md5);
942 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900943
Eric Dumazet547669d2013-05-23 07:44:20 +0000944 if (tcp_packets_in_flight(tp) == 0)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800945 tcp_ca_event(sk, CA_EVENT_TX_START);
Eric Dumazet547669d2013-05-23 07:44:20 +0000946
947 /* if no packet is in qdisc/device queue, then allow XPS to select
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700948 * another queue. We can be called from tcp_tsq_handler()
949 * which holds one reference to sk_wmem_alloc.
950 *
951 * TODO: Ideally, in-flight pure ACK packets should not matter here.
952 * One way to get this would be to set skb->truesize = 2 on them.
Eric Dumazet547669d2013-05-23 07:44:20 +0000953 */
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700954 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800955
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700956 skb_push(skb, tcp_header_size);
957 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000958
959 skb_orphan(skb);
960 skb->sk = sk;
Eric Dumazet98781962015-02-03 18:31:53 -0800961 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
Tom Herbertb73c3d02014-07-01 21:32:17 -0700962 skb_set_hash_from_sk(skb, sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000963 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800964
965 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700966 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000967 th->source = inet->inet_sport;
968 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800969 th->seq = htonl(tcb->seq);
970 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700971 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400972 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800973
Eric Dumazet4de075e2011-09-27 13:25:05 -0400974 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800975 /* RFC1323: The window in SYN & SYN/ACK segments
976 * is never scaled.
977 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800978 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800979 } else {
980 th->window = htons(tcp_select_window(sk));
981 }
982 th->check = 0;
983 th->urg_ptr = 0;
984
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700985 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800986 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
987 if (before(tp->snd_up, tcb->seq + 0x10000)) {
988 th->urg_ptr = htons(tp->snd_up - tcb->seq);
989 th->urg = 1;
990 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700991 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800992 th->urg = 1;
993 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800994 }
995
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000996 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet4de075e2011-09-27 13:25:05 -0400997 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
Florian Westphal735d3832014-09-29 13:08:30 +0200998 tcp_ecn_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800999
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001000#ifdef CONFIG_TCP_MD5SIG
1001 /* Calculate the MD5 hash, as we have all we need now */
1002 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -07001003 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00001004 tp->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07001005 md5, sk, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001006 }
1007#endif
1008
Herbert Xubb296242010-04-11 02:15:55 +00001009 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001010
Eric Dumazet4de075e2011-09-27 13:25:05 -04001011 if (likely(tcb->tcp_flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001012 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1013
1014 if (skb->len != tcp_header_size)
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001015 tcp_event_data_sent(tp, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001016
Wei Yongjunbd37a082006-08-07 21:04:15 -07001017 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +00001018 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1019 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001020
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001021 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
1022 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1023
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001024 /* Our usage of tstamp should remain private */
1025 skb->tstamp.tv64 = 0;
Eric Dumazet971f10e2014-09-27 09:50:57 -07001026
1027 /* Cleanup our debris for IP stacks */
1028 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1029 sizeof(struct inet6_skb_parm)));
1030
Eric Dumazetb0270e92014-04-15 12:58:34 -04001031 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001032
Hua Zhong83de47c2006-04-28 15:26:50 -07001033 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001034 return err;
1035
Christoph Paasch5ee2c942014-07-14 16:58:32 +02001036 tcp_enter_cwr(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001037
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -02001038 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001039}
1040
Andi Kleen67edfef2009-07-21 23:00:40 +00001041/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 *
1043 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1044 * otherwise socket can stall.
1045 */
1046static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1047{
1048 struct tcp_sock *tp = tcp_sk(sk);
1049
1050 /* Advance write_seq and place onto the write_queue. */
1051 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001052 __skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -08001053 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001054 sk->sk_wmem_queued += skb->truesize;
1055 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056}
1057
Andi Kleen67edfef2009-07-21 23:00:40 +00001058/* Initialize TSO segments for a packet. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001059static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001060 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -07001061{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001062 struct skb_shared_info *shinfo = skb_shinfo(skb);
1063
Eric Dumazetc52e24212013-10-15 11:54:30 -07001064 /* Make sure we own this skb before messing gso_size/gso_segs */
1065 WARN_ON_ONCE(skb_cloned(skb));
1066
Eric Dumazet8f26fb12013-10-15 12:24:54 -07001067 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -07001068 /* Avoid the costly divide in the normal
1069 * non-TSO case.
1070 */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001071 tcp_skb_pcount_set(skb, 1);
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001072 shinfo->gso_size = 0;
1073 shinfo->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -07001074 } else {
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001075 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001076 shinfo->gso_size = mss_now;
1077 shinfo->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 }
1079}
1080
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001081/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001082 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001083 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001084static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001085 int decr)
1086{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001087 struct tcp_sock *tp = tcp_sk(sk);
1088
Ilpo Järvinendc869672007-10-01 15:27:19 -07001089 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001090 return;
1091
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001092 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001093 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001094}
1095
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001096/* Pcount in the middle of the write queue got changed, we need to do various
1097 * tweaks to fix counters
1098 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001099static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001100{
1101 struct tcp_sock *tp = tcp_sk(sk);
1102
1103 tp->packets_out -= decr;
1104
1105 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1106 tp->sacked_out -= decr;
1107 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1108 tp->retrans_out -= decr;
1109 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1110 tp->lost_out -= decr;
1111
1112 /* Reno case is special. Sigh... */
1113 if (tcp_is_reno(tp) && decr > 0)
1114 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1115
1116 tcp_adjust_fackets_out(sk, skb, decr);
1117
1118 if (tp->lost_skb_hint &&
1119 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001120 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001121 tp->lost_cnt_hint -= decr;
1122
1123 tcp_verify_left_out(tp);
1124}
1125
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001126static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1127{
1128 struct skb_shared_info *shinfo = skb_shinfo(skb);
1129
1130 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1131 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1132 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1133 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1134
1135 shinfo->tx_flags &= ~tsflags;
1136 shinfo2->tx_flags |= tsflags;
1137 swap(shinfo->tskey, shinfo2->tskey);
1138 }
1139}
1140
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141/* Function to create two new TCP segments. Shrinks the given segment
1142 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001143 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 * Remember, these are still headerless SKBs at this point.
1145 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001146int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001147 unsigned int mss_now, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148{
1149 struct tcp_sock *tp = tcp_sk(sk);
1150 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001151 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001152 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001153 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001155 if (WARN_ON(len > skb->len))
1156 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001157
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 nsize = skb_headlen(skb) - len;
1159 if (nsize < 0)
1160 nsize = 0;
1161
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001162 if (skb_unclone(skb, gfp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 return -ENOMEM;
1164
1165 /* Get a new skb... force flag on. */
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001166 buff = sk_stream_alloc_skb(sk, nsize, gfp);
Ian Morris51456b22015-04-03 09:17:26 +01001167 if (!buff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001169
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001170 sk->sk_wmem_queued += buff->truesize;
1171 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001172 nlen = skb->len - len - nsize;
1173 buff->truesize += nlen;
1174 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175
1176 /* Correct the sequence numbers. */
1177 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1178 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1179 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1180
1181 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001182 flags = TCP_SKB_CB(skb)->tcp_flags;
1183 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1184 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001185 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
Patrick McHardy84fa7932006-08-29 16:44:56 -07001187 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001189 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1190 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191 nsize, 0);
1192
1193 skb_trim(skb, len);
1194
1195 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1196 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001197 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 skb_split(skb, buff, len);
1199 }
1200
1201 buff->ip_summed = skb->ip_summed;
1202
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001203 buff->tstamp = skb->tstamp;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001204 tcp_fragment_tstamp(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205
David S. Miller6475be12005-09-01 22:47:01 -07001206 old_factor = tcp_skb_pcount(skb);
1207
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001209 tcp_set_skb_tso_segs(sk, skb, mss_now);
1210 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211
David S. Miller6475be12005-09-01 22:47:01 -07001212 /* If this packet has been sent out already, we must
1213 * adjust the various packet counters.
1214 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001215 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001216 int diff = old_factor - tcp_skb_pcount(skb) -
1217 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001219 if (diff)
1220 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 }
1222
1223 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001224 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001225 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226
1227 return 0;
1228}
1229
1230/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1231 * eventually). The difference is that pulled data not copied, but
1232 * immediately discarded.
1233 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001234static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001236 struct skb_shared_info *shinfo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 int i, k, eat;
1238
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001239 eat = min_t(int, len, skb_headlen(skb));
1240 if (eat) {
1241 __skb_pull(skb, eat);
1242 len -= eat;
1243 if (!len)
1244 return;
1245 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 eat = len;
1247 k = 0;
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001248 shinfo = skb_shinfo(skb);
1249 for (i = 0; i < shinfo->nr_frags; i++) {
1250 int size = skb_frag_size(&shinfo->frags[i]);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001251
1252 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001253 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001254 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001255 } else {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001256 shinfo->frags[k] = shinfo->frags[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 if (eat) {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001258 shinfo->frags[k].page_offset += eat;
1259 skb_frag_size_sub(&shinfo->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 eat = 0;
1261 }
1262 k++;
1263 }
1264 }
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001265 shinfo->nr_frags = k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001267 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 skb->data_len -= len;
1269 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270}
1271
Andi Kleen67edfef2009-07-21 23:00:40 +00001272/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1274{
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001275 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 return -ENOMEM;
1277
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001278 __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
1280 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001281 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282
1283 skb->truesize -= len;
1284 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001285 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1287
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001288 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 if (tcp_skb_pcount(skb) > 1)
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001290 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291
1292 return 0;
1293}
1294
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001295/* Calculate MSS not accounting any TCP options. */
1296static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001297{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001298 const struct tcp_sock *tp = tcp_sk(sk);
1299 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001300 int mss_now;
1301
1302 /* Calculate base mss without TCP options:
1303 It is MMS_S - sizeof(tcphdr) of rfc1122
1304 */
1305 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1306
Eric Dumazet67469602012-04-24 07:37:38 +00001307 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1308 if (icsk->icsk_af_ops->net_frag_header_len) {
1309 const struct dst_entry *dst = __sk_dst_get(sk);
1310
1311 if (dst && dst_allfrag(dst))
1312 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1313 }
1314
John Heffner5d424d52006-03-20 17:53:41 -08001315 /* Clamp it (mss_clamp does not include tcp options) */
1316 if (mss_now > tp->rx_opt.mss_clamp)
1317 mss_now = tp->rx_opt.mss_clamp;
1318
1319 /* Now subtract optional transport overhead */
1320 mss_now -= icsk->icsk_ext_hdr_len;
1321
1322 /* Then reserve room for full set of TCP options and 8 bytes of data */
1323 if (mss_now < 48)
1324 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001325 return mss_now;
1326}
1327
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001328/* Calculate MSS. Not accounting for SACKs here. */
1329int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1330{
1331 /* Subtract TCP options size, not including SACKs */
1332 return __tcp_mtu_to_mss(sk, pmtu) -
1333 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1334}
1335
John Heffner5d424d52006-03-20 17:53:41 -08001336/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001337int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001338{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001339 const struct tcp_sock *tp = tcp_sk(sk);
1340 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001341 int mtu;
1342
1343 mtu = mss +
1344 tp->tcp_header_len +
1345 icsk->icsk_ext_hdr_len +
1346 icsk->icsk_af_ops->net_header_len;
1347
Eric Dumazet67469602012-04-24 07:37:38 +00001348 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1349 if (icsk->icsk_af_ops->net_frag_header_len) {
1350 const struct dst_entry *dst = __sk_dst_get(sk);
1351
1352 if (dst && dst_allfrag(dst))
1353 mtu += icsk->icsk_af_ops->net_frag_header_len;
1354 }
John Heffner5d424d52006-03-20 17:53:41 -08001355 return mtu;
1356}
1357
Andi Kleen67edfef2009-07-21 23:00:40 +00001358/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001359void tcp_mtup_init(struct sock *sk)
1360{
1361 struct tcp_sock *tp = tcp_sk(sk);
1362 struct inet_connection_sock *icsk = inet_csk(sk);
Fan Dub0f9ca52015-02-10 09:53:16 +08001363 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001364
Fan Dub0f9ca52015-02-10 09:53:16 +08001365 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
John Heffner5d424d52006-03-20 17:53:41 -08001366 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001367 icsk->icsk_af_ops->net_header_len;
Fan Dub0f9ca52015-02-10 09:53:16 +08001368 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
John Heffner5d424d52006-03-20 17:53:41 -08001369 icsk->icsk_mtup.probe_size = 0;
Fan Du05cbc0d2015-03-06 11:18:24 +08001370 if (icsk->icsk_mtup.enabled)
1371 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
John Heffner5d424d52006-03-20 17:53:41 -08001372}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001373EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001374
Linus Torvalds1da177e2005-04-16 15:20:36 -07001375/* This function synchronize snd mss to current pmtu/exthdr set.
1376
1377 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1378 for TCP options, but includes only bare TCP header.
1379
1380 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001381 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 It also does not include TCP options.
1383
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001384 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
1386 tp->mss_cache is current effective sending mss, including
1387 all tcp options except for SACKs. It is evaluated,
1388 taking into account current pmtu, but never exceeds
1389 tp->rx_opt.mss_clamp.
1390
1391 NOTE1. rfc1122 clearly states that advertised MSS
1392 DOES NOT include either tcp or ip options.
1393
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001394 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1395 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1398{
1399 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001400 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001401 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402
John Heffner5d424d52006-03-20 17:53:41 -08001403 if (icsk->icsk_mtup.search_high > pmtu)
1404 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405
John Heffner5d424d52006-03-20 17:53:41 -08001406 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001407 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
1409 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001410 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001411 if (icsk->icsk_mtup.enabled)
1412 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001413 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001414
1415 return mss_now;
1416}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001417EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
1419/* Compute the current effective MSS, taking SACKs and IP options,
1420 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001422unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001424 const struct tcp_sock *tp = tcp_sk(sk);
1425 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001426 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001427 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001428 struct tcp_out_options opts;
1429 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001431 mss_now = tp->mss_cache;
1432
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433 if (dst) {
1434 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001435 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 mss_now = tcp_sync_mss(sk, mtu);
1437 }
1438
Adam Langley33ad7982008-07-19 00:04:31 -07001439 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1440 sizeof(struct tcphdr);
1441 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1442 * some common options. If this is an odd packet (because we have SACK
1443 * blocks etc) then our calculated header_len will be different, and
1444 * we have to adjust mss_now correspondingly */
1445 if (header_len != tp->tcp_header_len) {
1446 int delta = (int) header_len - tp->tcp_header_len;
1447 mss_now -= delta;
1448 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001449
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 return mss_now;
1451}
1452
Weiping Pan86fd14a2014-04-18 12:27:46 +08001453/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1454 * As additional protections, we do not touch cwnd in retransmission phases,
1455 * and if application hit its sndbuf limit recently.
1456 */
1457static void tcp_cwnd_application_limited(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001458{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001459 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001460
Weiping Pan86fd14a2014-04-18 12:27:46 +08001461 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1462 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1463 /* Limited by application or receiver window. */
1464 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1465 u32 win_used = max(tp->snd_cwnd_used, init_win);
1466 if (win_used < tp->snd_cwnd) {
1467 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1468 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1469 }
1470 tp->snd_cwnd_used = 0;
1471 }
1472 tp->snd_cwnd_stamp = tcp_time_stamp;
1473}
1474
Neal Cardwellca8a2262014-05-22 10:41:08 -04001475static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
David S. Millera762a982005-07-05 15:18:51 -07001476{
1477 struct tcp_sock *tp = tcp_sk(sk);
1478
Neal Cardwellca8a2262014-05-22 10:41:08 -04001479 /* Track the maximum number of outstanding packets in each
1480 * window, and remember whether we were cwnd-limited then.
1481 */
1482 if (!before(tp->snd_una, tp->max_packets_seq) ||
1483 tp->packets_out > tp->max_packets_out) {
1484 tp->max_packets_out = tp->packets_out;
1485 tp->max_packets_seq = tp->snd_nxt;
1486 tp->is_cwnd_limited = is_cwnd_limited;
1487 }
Eric Dumazete114a712014-04-30 11:58:13 -07001488
Eric Dumazet24901552014-05-02 21:18:05 -07001489 if (tcp_is_cwnd_limited(sk)) {
David S. Millera762a982005-07-05 15:18:51 -07001490 /* Network is feed fully. */
1491 tp->snd_cwnd_used = 0;
1492 tp->snd_cwnd_stamp = tcp_time_stamp;
1493 } else {
1494 /* Network starves. */
1495 if (tp->packets_out > tp->snd_cwnd_used)
1496 tp->snd_cwnd_used = tp->packets_out;
1497
David S. Miller15d33c02007-04-09 13:23:14 -07001498 if (sysctl_tcp_slow_start_after_idle &&
1499 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001500 tcp_cwnd_application_limited(sk);
1501 }
1502}
1503
Eric Dumazetd4589922013-12-13 13:51:23 -08001504/* Minshall's variant of the Nagle send check. */
1505static bool tcp_minshall_check(const struct tcp_sock *tp)
1506{
1507 return after(tp->snd_sml, tp->snd_una) &&
1508 !after(tp->snd_sml, tp->snd_nxt);
1509}
1510
1511/* Update snd_sml if this skb is under mss
1512 * Note that a TSO packet might end with a sub-mss segment
1513 * The test is really :
1514 * if ((skb->len % mss) != 0)
1515 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1516 * But we can avoid doing the divide again given we already have
1517 * skb_pcount = skb->len / mss_now
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001518 */
Eric Dumazetd4589922013-12-13 13:51:23 -08001519static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1520 const struct sk_buff *skb)
1521{
1522 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1523 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1524}
1525
1526/* Return false, if packet can be sent now without violation Nagle's rules:
1527 * 1. It is full sized. (provided by caller in %partial bool)
1528 * 2. Or it contains FIN. (already checked by caller)
1529 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1530 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1531 * With Minshall's modification: all sent small packets are ACKed.
1532 */
1533static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001534 int nonagle)
Eric Dumazetd4589922013-12-13 13:51:23 -08001535{
1536 return partial &&
1537 ((nonagle & TCP_NAGLE_CORK) ||
1538 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1539}
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001540
1541/* Return how many segs we'd like on a TSO packet,
1542 * to send one TSO packet per ms
1543 */
1544static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1545{
1546 u32 bytes, segs;
1547
1548 bytes = min(sk->sk_pacing_rate >> 10,
1549 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1550
1551 /* Goal is to send at least one packet per ms,
1552 * not one big TSO packet every 100 ms.
1553 * This preserves ACK clocking and is consistent
1554 * with tcp_tso_should_defer() heuristic.
1555 */
1556 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1557
1558 return min_t(u32, segs, sk->sk_gso_max_segs);
1559}
1560
Eric Dumazetd4589922013-12-13 13:51:23 -08001561/* Returns the portion of skb which can be sent right away */
1562static unsigned int tcp_mss_split_point(const struct sock *sk,
1563 const struct sk_buff *skb,
1564 unsigned int mss_now,
1565 unsigned int max_segs,
1566 int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001567{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001568 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetd4589922013-12-13 13:51:23 -08001569 u32 partial, needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001570
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001571 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001572 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001573
Ben Hutchings14853482012-07-30 16:11:42 +00001574 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1575 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001576
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001577 needed = min(skb->len, window);
1578
Ben Hutchings14853482012-07-30 16:11:42 +00001579 if (max_len <= needed)
1580 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001581
Eric Dumazetd4589922013-12-13 13:51:23 -08001582 partial = needed % mss_now;
1583 /* If last segment is not a full MSS, check if Nagle rules allow us
1584 * to include this last segment in this skb.
1585 * Otherwise, we'll split the skb at last MSS boundary
1586 */
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001587 if (tcp_nagle_check(partial != 0, tp, nonagle))
Eric Dumazetd4589922013-12-13 13:51:23 -08001588 return needed - partial;
1589
1590 return needed;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001591}
1592
1593/* Can at least one segment of SKB be sent right now, according to the
1594 * congestion window rules? If so, return how many segments are allowed.
1595 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001596static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1597 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001598{
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001599 u32 in_flight, cwnd, halfcwnd;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001600
1601 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001602 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1603 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001604 return 1;
1605
1606 in_flight = tcp_packets_in_flight(tp);
1607 cwnd = tp->snd_cwnd;
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001608 if (in_flight >= cwnd)
1609 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001610
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001611 /* For better scheduling, ensure we have at least
1612 * 2 GSO packets in flight.
1613 */
1614 halfcwnd = max(cwnd >> 1, 1U);
1615 return min(halfcwnd, cwnd - in_flight);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001616}
1617
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001618/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001619 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001620 * SKB onto the wire.
1621 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001622static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001623 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001624{
1625 int tso_segs = tcp_skb_pcount(skb);
1626
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001627 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001628 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001629 tso_segs = tcp_skb_pcount(skb);
1630 }
1631 return tso_segs;
1632}
1633
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001634
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001635/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001636 * sent now.
1637 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001638static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1639 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001640{
1641 /* Nagle rule does not apply to frames, which sit in the middle of the
1642 * write_queue (they have no chances to get new data).
1643 *
1644 * This is implemented in the callers, where they modify the 'nonagle'
1645 * argument based upon the location of SKB in the send queue.
1646 */
1647 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001648 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001649
Yuchung Cheng9b441902013-03-20 13:32:58 +00001650 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1651 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001652 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001653
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001654 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001655 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001656
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001657 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001658}
1659
1660/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001661static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1662 const struct sk_buff *skb,
1663 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001664{
1665 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1666
1667 if (skb->len > cur_mss)
1668 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1669
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001670 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001671}
1672
David S. Millerfe067e82007-03-07 12:12:44 -08001673/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001674 * should be put on the wire right now. If so, it returns the number of
1675 * packets allowed by the congestion window.
1676 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001677static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001678 unsigned int cur_mss, int nonagle)
1679{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001680 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001681 unsigned int cwnd_quota;
1682
David S. Miller846998a2005-08-04 19:52:01 -07001683 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001684
1685 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1686 return 0;
1687
1688 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001689 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001690 cwnd_quota = 0;
1691
1692 return cwnd_quota;
1693}
1694
Andi Kleen67edfef2009-07-21 23:00:40 +00001695/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001696bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001697{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001698 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001699 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001700
Eric Dumazeta02cec22010-09-22 20:43:57 +00001701 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001702 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001703 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001704 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001705}
1706
1707/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1708 * which is put after SKB on the list. It is very much like
1709 * tcp_fragment() except that it may make several kinds of assumptions
1710 * in order to speed up the splitting operation. In particular, we
1711 * know that all the data is in scatter-gather pages, and that the
1712 * packet has never been sent out before (and thus is not cloned).
1713 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001714static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001715 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001716{
1717 struct sk_buff *buff;
1718 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001719 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001720
1721 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001722 if (skb->len != skb->data_len)
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001723 return tcp_fragment(sk, skb, len, mss_now, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001724
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001725 buff = sk_stream_alloc_skb(sk, 0, gfp);
Ian Morris51456b22015-04-03 09:17:26 +01001726 if (unlikely(!buff))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001727 return -ENOMEM;
1728
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001729 sk->sk_wmem_queued += buff->truesize;
1730 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001731 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001732 skb->truesize -= nlen;
1733
1734 /* Correct the sequence numbers. */
1735 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1736 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1737 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1738
1739 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001740 flags = TCP_SKB_CB(skb)->tcp_flags;
1741 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1742 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001743
1744 /* This packet was never sent out yet, so no SACK bits. */
1745 TCP_SKB_CB(buff)->sacked = 0;
1746
Patrick McHardy84fa7932006-08-29 16:44:56 -07001747 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001748 skb_split(skb, buff, len);
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001749 tcp_fragment_tstamp(skb, buff);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001750
1751 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001752 tcp_set_skb_tso_segs(sk, skb, mss_now);
1753 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001754
1755 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001756 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001757 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001758
1759 return 0;
1760}
1761
1762/* Try to defer sending, if possible, in order to minimize the amount
1763 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1764 *
1765 * This algorithm is from John Heffner.
1766 */
Neal Cardwellca8a2262014-05-22 10:41:08 -04001767static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001768 bool *is_cwnd_limited, u32 max_segs)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001769{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001770 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet50c83392015-02-26 14:10:19 -08001771 u32 age, send_win, cong_win, limit, in_flight;
1772 struct tcp_sock *tp = tcp_sk(sk);
1773 struct skb_mstamp now;
1774 struct sk_buff *head;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001775 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001776
Eric Dumazet4de075e2011-09-27 13:25:05 -04001777 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001778 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001779
Eric Dumazeta0ea7002015-02-26 14:10:20 -08001780 if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR)))
John Heffnerae8064a2006-10-18 20:36:48 -07001781 goto send_now;
1782
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001783 /* Avoid bursty behavior by allowing defer
1784 * only if the last write was recent.
1785 */
1786 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
John Heffnerae8064a2006-10-18 20:36:48 -07001787 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001788
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001789 in_flight = tcp_packets_in_flight(tp);
1790
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001791 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001792
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001793 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001794
1795 /* From in_flight test above, we know that cwnd > in_flight. */
1796 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1797
1798 limit = min(send_win, cong_win);
1799
David S. Millerba244fe2006-03-11 18:51:49 -08001800 /* If a full-sized TSO skb can be sent, do it. */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001801 if (limit >= max_segs * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001802 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001803
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001804 /* Middle in queue won't get any more data, full sendable already? */
1805 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1806 goto send_now;
1807
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001808 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1809 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001810 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1811
1812 /* If at least some fraction of a window is available,
1813 * just use it.
1814 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001815 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001816 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001817 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001818 } else {
1819 /* Different approach, try not to defer past a single
1820 * ACK. Receiver should ACK every other full sized
1821 * frame, so if we have space for more than 3 frames
1822 * then send now.
1823 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001824 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001825 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001826 }
1827
Eric Dumazet50c83392015-02-26 14:10:19 -08001828 head = tcp_write_queue_head(sk);
1829 skb_mstamp_get(&now);
1830 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1831 /* If next ACK is likely to come too late (half srtt), do not defer */
1832 if (age < (tp->srtt_us >> 4))
1833 goto send_now;
1834
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001835 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001836
Neal Cardwellca8a2262014-05-22 10:41:08 -04001837 if (cong_win < send_win && cong_win < skb->len)
1838 *is_cwnd_limited = true;
1839
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001840 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001841
1842send_now:
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001843 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001844}
1845
Fan Du05cbc0d2015-03-06 11:18:24 +08001846static inline void tcp_mtu_check_reprobe(struct sock *sk)
1847{
1848 struct inet_connection_sock *icsk = inet_csk(sk);
1849 struct tcp_sock *tp = tcp_sk(sk);
1850 struct net *net = sock_net(sk);
1851 u32 interval;
1852 s32 delta;
1853
1854 interval = net->ipv4.sysctl_tcp_probe_interval;
1855 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1856 if (unlikely(delta >= interval * HZ)) {
1857 int mss = tcp_current_mss(sk);
1858
1859 /* Update current search range */
1860 icsk->icsk_mtup.probe_size = 0;
1861 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1862 sizeof(struct tcphdr) +
1863 icsk->icsk_af_ops->net_header_len;
1864 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1865
1866 /* Update probe time stamp */
1867 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1868 }
1869}
1870
John Heffner5d424d52006-03-20 17:53:41 -08001871/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001872 * MTU probe is regularly attempting to increase the path MTU by
1873 * deliberately sending larger packets. This discovers routing
1874 * changes resulting in larger path MTUs.
1875 *
John Heffner5d424d52006-03-20 17:53:41 -08001876 * Returns 0 if we should wait to probe (no cwnd available),
1877 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001878 * -1 otherwise
1879 */
John Heffner5d424d52006-03-20 17:53:41 -08001880static int tcp_mtu_probe(struct sock *sk)
1881{
1882 struct tcp_sock *tp = tcp_sk(sk);
1883 struct inet_connection_sock *icsk = inet_csk(sk);
1884 struct sk_buff *skb, *nskb, *next;
Fan Du6b58e0a2015-03-06 11:18:23 +08001885 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001886 int len;
1887 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001888 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001889 int copy;
1890 int mss_now;
Fan Du6b58e0a2015-03-06 11:18:23 +08001891 int interval;
John Heffner5d424d52006-03-20 17:53:41 -08001892
1893 /* Not currently probing/verifying,
1894 * not in recovery,
1895 * have enough cwnd, and
1896 * not SACKing (the variable headers throw things off) */
1897 if (!icsk->icsk_mtup.enabled ||
1898 icsk->icsk_mtup.probe_size ||
1899 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1900 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001901 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001902 return -1;
1903
Fan Du6b58e0a2015-03-06 11:18:23 +08001904 /* Use binary search for probe_size between tcp_mss_base,
1905 * and current mss_clamp. if (search_high - search_low)
1906 * smaller than a threshold, backoff from probing.
1907 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001908 mss_now = tcp_current_mss(sk);
Fan Du6b58e0a2015-03-06 11:18:23 +08001909 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1910 icsk->icsk_mtup.search_low) >> 1);
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001911 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
Fan Du6b58e0a2015-03-06 11:18:23 +08001912 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
Fan Du05cbc0d2015-03-06 11:18:24 +08001913 /* When misfortune happens, we are reprobing actively,
1914 * and then reprobe timer has expired. We stick with current
1915 * probing process by not resetting search range to its orignal.
1916 */
Fan Du6b58e0a2015-03-06 11:18:23 +08001917 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
Fan Du05cbc0d2015-03-06 11:18:24 +08001918 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1919 /* Check whether enough time has elaplased for
1920 * another round of probing.
1921 */
1922 tcp_mtu_check_reprobe(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001923 return -1;
1924 }
1925
1926 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001927 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001928 return -1;
1929
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001930 if (tp->snd_wnd < size_needed)
1931 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001932 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001933 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001934
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001935 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1936 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1937 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001938 return -1;
1939 else
1940 return 0;
1941 }
1942
1943 /* We're allowed to probe. Build it now. */
Ian Morris51456b22015-04-03 09:17:26 +01001944 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC);
1945 if (!nskb)
John Heffner5d424d52006-03-20 17:53:41 -08001946 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001947 sk->sk_wmem_queued += nskb->truesize;
1948 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001949
David S. Millerfe067e82007-03-07 12:12:44 -08001950 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001951
1952 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1953 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001954 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001955 TCP_SKB_CB(nskb)->sacked = 0;
1956 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001957 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001958
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001959 tcp_insert_write_queue_before(nskb, skb, sk);
1960
John Heffner5d424d52006-03-20 17:53:41 -08001961 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001962 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001963 copy = min_t(int, skb->len, probe_size - len);
1964 if (nskb->ip_summed)
1965 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1966 else
1967 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001968 skb_put(nskb, copy),
1969 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001970
1971 if (skb->len <= copy) {
1972 /* We've eaten all the data from this skb.
1973 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001974 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001975 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001976 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001977 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04001978 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001979 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001980 if (!skb_shinfo(skb)->nr_frags) {
1981 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001982 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001983 skb->csum = csum_partial(skb->data,
1984 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001985 } else {
1986 __pskb_trim_head(skb, copy);
1987 tcp_set_skb_tso_segs(sk, skb, mss_now);
1988 }
1989 TCP_SKB_CB(skb)->seq += copy;
1990 }
1991
1992 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001993
1994 if (len >= probe_size)
1995 break;
John Heffner5d424d52006-03-20 17:53:41 -08001996 }
1997 tcp_init_tso_segs(sk, nskb, nskb->len);
1998
1999 /* We're ready to send. If this fails, the probe will
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002000 * be resegmented into mss-sized pieces by tcp_write_xmit().
2001 */
John Heffner5d424d52006-03-20 17:53:41 -08002002 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2003 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002004 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08002005 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002006 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08002007
2008 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08002009 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2010 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08002011
2012 return 1;
2013 }
2014
2015 return -1;
2016}
2017
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018/* This routine writes packets to the network. It advances the
2019 * send_head. This happens as incoming acks open up the remote
2020 * window for us.
2021 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08002022 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2023 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2024 * account rare use of URG, this is not a big flaw.
2025 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002026 * Send at most one packet when push_one > 0. Temporarily ignore
2027 * cwnd limit to force at most one packet out when push_one == 2.
2028
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002029 * Returns true, if no segments are in flight and we have queued segments,
2030 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002032static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2033 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034{
2035 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07002036 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002037 unsigned int tso_segs, sent_pkts;
2038 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08002039 int result;
Neal Cardwellca8a2262014-05-22 10:41:08 -04002040 bool is_cwnd_limited = false;
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002041 u32 max_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042
David S. Miller92df7b52005-07-05 15:19:06 -07002043 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002044
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002045 if (!push_one) {
2046 /* Do MTU probing. */
2047 result = tcp_mtu_probe(sk);
2048 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002049 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002050 } else if (result > 0) {
2051 sent_pkts = 1;
2052 }
John Heffner5d424d52006-03-20 17:53:41 -08002053 }
2054
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002055 max_segs = tcp_tso_autosize(sk, mss_now);
David S. Millerfe067e82007-03-07 12:12:44 -08002056 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07002057 unsigned int limit;
2058
Herbert Xub68e9f82005-08-04 19:52:02 -07002059 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002060 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07002061
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002062 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002063 /* "skb_mstamp" is used as a start point for the retransmit timer */
2064 skb_mstamp_get(&skb->skb_mstamp);
Andrew Vaginec342322012-11-15 04:03:17 +00002065 goto repair; /* Skip network transmission */
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002066 }
Andrew Vaginec342322012-11-15 04:03:17 +00002067
Herbert Xub68e9f82005-08-04 19:52:02 -07002068 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002069 if (!cwnd_quota) {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002070 is_cwnd_limited = true;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002071 if (push_one == 2)
2072 /* Force out a loss probe pkt. */
2073 cwnd_quota = 1;
2074 else
2075 break;
2076 }
Herbert Xub68e9f82005-08-04 19:52:02 -07002077
2078 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2079 break;
2080
Herbert Xu843925f2015-01-01 00:39:23 +11002081 if (tso_segs == 1 || !max_segs) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002082 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2083 (tcp_skb_is_last(sk, skb) ?
2084 nonagle : TCP_NAGLE_PUSH))))
2085 break;
2086 } else {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002087 if (!push_one &&
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002088 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2089 max_segs))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002090 break;
2091 }
David S. Milleraa934662005-07-05 15:20:09 -07002092
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002093 limit = mss_now;
Herbert Xu843925f2015-01-01 00:39:23 +11002094 if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp))
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002095 limit = tcp_mss_split_point(sk, skb, mss_now,
2096 min_t(unsigned int,
2097 cwnd_quota,
2098 max_segs),
2099 nonagle);
2100
2101 if (skb->len > limit &&
2102 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2103 break;
2104
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002105 /* TCP Small Queues :
2106 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2107 * This allows for :
2108 * - better RTT estimation and ACK scheduling
2109 * - faster recovery
2110 * - high rates
Eric Dumazet98e09382013-11-13 06:32:54 -08002111 * Alas, some drivers / subsystems require a fair amount
2112 * of queued bytes to ensure line rate.
2113 * One example is wifi aggregation (802.11 AMPDU)
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002114 */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002115 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2116 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002117
2118 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002119 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
John Ognessbf062002014-02-09 18:40:11 -08002120 /* It is possible TX completion already happened
2121 * before we set TSQ_THROTTLED, so we must
2122 * test again the condition.
John Ognessbf062002014-02-09 18:40:11 -08002123 */
Peter Zijlstra4e857c52014-03-17 18:06:10 +01002124 smp_mb__after_atomic();
John Ognessbf062002014-02-09 18:40:11 -08002125 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2126 break;
Eric Dumazet46d3cea2012-07-11 05:50:31 +00002127 }
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002128
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002129 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07002130 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
Andrew Vaginec342322012-11-15 04:03:17 +00002132repair:
David S. Miller92df7b52005-07-05 15:19:06 -07002133 /* Advance the send_head. This one is sent out.
2134 * This call will increment packets_out.
2135 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002136 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07002137
2138 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002139 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002140
2141 if (push_one)
2142 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 }
David S. Miller92df7b52005-07-05 15:19:06 -07002144
David S. Milleraa934662005-07-05 15:20:09 -07002145 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002146 if (tcp_in_cwnd_reduction(sk))
2147 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002148
2149 /* Send one loss probe per tail loss episode. */
2150 if (push_one != 2)
2151 tcp_schedule_loss_probe(sk);
Neal Cardwellca8a2262014-05-22 10:41:08 -04002152 tcp_cwnd_validate(sk, is_cwnd_limited);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002153 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07002154 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002155 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2156}
2157
2158bool tcp_schedule_loss_probe(struct sock *sk)
2159{
2160 struct inet_connection_sock *icsk = inet_csk(sk);
2161 struct tcp_sock *tp = tcp_sk(sk);
2162 u32 timeout, tlp_time_stamp, rto_time_stamp;
Eric Dumazet740b0f12014-02-26 14:02:48 -08002163 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002164
2165 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2166 return false;
2167 /* No consecutive loss probes. */
2168 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2169 tcp_rearm_rto(sk);
2170 return false;
2171 }
2172 /* Don't do any loss probe on a Fast Open connection before 3WHS
2173 * finishes.
2174 */
2175 if (sk->sk_state == TCP_SYN_RECV)
2176 return false;
2177
2178 /* TLP is only scheduled when next timer event is RTO. */
2179 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2180 return false;
2181
2182 /* Schedule a loss probe in 2*RTT for SACK capable connections
2183 * in Open state, that are either limited by cwnd or application.
2184 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08002185 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002186 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2187 return false;
2188
2189 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2190 tcp_send_head(sk))
2191 return false;
2192
2193 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2194 * for delayed ack when there's one outstanding packet.
2195 */
2196 timeout = rtt << 1;
2197 if (tp->packets_out == 1)
2198 timeout = max_t(u32, timeout,
2199 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2200 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2201
2202 /* If RTO is shorter, just schedule TLP in its place. */
2203 tlp_time_stamp = tcp_time_stamp + timeout;
2204 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2205 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2206 s32 delta = rto_time_stamp - tcp_time_stamp;
2207 if (delta > 0)
2208 timeout = delta;
2209 }
2210
2211 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2212 TCP_RTO_MAX);
2213 return true;
2214}
2215
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002216/* Thanks to skb fast clones, we can detect if a prior transmit of
2217 * a packet is still in a qdisc or driver queue.
2218 * In this case, there is very little point doing a retransmit !
2219 * Note: This is called from BH context only.
2220 */
2221static bool skb_still_in_host_queue(const struct sock *sk,
2222 const struct sk_buff *skb)
2223{
Eric Dumazet39bb5e62014-10-30 10:32:34 -07002224 if (unlikely(skb_fclone_busy(sk, skb))) {
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002225 NET_INC_STATS_BH(sock_net(sk),
2226 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2227 return true;
2228 }
2229 return false;
2230}
2231
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002232/* When probe timeout (PTO) fires, send a new segment if one exists, else
2233 * retransmit the last segment.
2234 */
2235void tcp_send_loss_probe(struct sock *sk)
2236{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002237 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002238 struct sk_buff *skb;
2239 int pcount;
2240 int mss = tcp_current_mss(sk);
2241 int err = -1;
2242
Ian Morris00db4122015-04-03 09:17:27 +01002243 if (tcp_send_head(sk)) {
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002244 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2245 goto rearm_timer;
2246 }
2247
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002248 /* At most one outstanding TLP retransmission. */
2249 if (tp->tlp_high_seq)
2250 goto rearm_timer;
2251
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002252 /* Retransmit last segment. */
2253 skb = tcp_write_queue_tail(sk);
2254 if (WARN_ON(!skb))
2255 goto rearm_timer;
2256
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002257 if (skb_still_in_host_queue(sk, skb))
2258 goto rearm_timer;
2259
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002260 pcount = tcp_skb_pcount(skb);
2261 if (WARN_ON(!pcount))
2262 goto rearm_timer;
2263
2264 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002265 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2266 GFP_ATOMIC)))
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002267 goto rearm_timer;
2268 skb = tcp_write_queue_tail(sk);
2269 }
2270
2271 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2272 goto rearm_timer;
2273
Per Hurtigbef19092014-06-12 17:08:32 +02002274 err = __tcp_retransmit_skb(sk, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002275
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002276 /* Record snd_nxt for loss detection. */
2277 if (likely(!err))
2278 tp->tlp_high_seq = tp->snd_nxt;
2279
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002280rearm_timer:
2281 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2282 inet_csk(sk)->icsk_rto,
2283 TCP_RTO_MAX);
2284
2285 if (likely(!err))
2286 NET_INC_STATS_BH(sock_net(sk),
2287 LINUX_MIB_TCPLOSSPROBES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288}
2289
David S. Millera762a982005-07-05 15:18:51 -07002290/* Push out any pending frames which were held back due to
2291 * TCP_CORK or attempt at coalescing tiny packets.
2292 * The socket must be locked by the caller.
2293 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002294void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2295 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002296{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002297 /* If we are closed, the bytes will have to remain here.
2298 * In time closedown will finish, we empty the write queue and
2299 * all will be happy.
2300 */
2301 if (unlikely(sk->sk_state == TCP_CLOSE))
2302 return;
2303
Mel Gorman99a1dec2012-07-31 16:44:14 -07002304 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2305 sk_gfp_atomic(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002306 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002307}
2308
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002309/* Send _single_ skb sitting at the send head. This function requires
2310 * true push pending frames to setup probe timer etc.
2311 */
2312void tcp_push_one(struct sock *sk, unsigned int mss_now)
2313{
David S. Millerfe067e82007-03-07 12:12:44 -08002314 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002315
2316 BUG_ON(!skb || skb->len < mss_now);
2317
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002318 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002319}
2320
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321/* This function returns the amount that we can raise the
2322 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002323 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 * 1. The window can never be shrunk once it is offered (RFC 793)
2325 * 2. We limit memory per socket
2326 *
2327 * RFC 1122:
2328 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2329 * RECV.NEXT + RCV.WIN fixed until:
2330 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2331 *
2332 * i.e. don't raise the right edge of the window until you can raise
2333 * it at least MSS bytes.
2334 *
2335 * Unfortunately, the recommended algorithm breaks header prediction,
2336 * since header prediction assumes th->window stays fixed.
2337 *
2338 * Strictly speaking, keeping th->window fixed violates the receiver
2339 * side SWS prevention criteria. The problem is that under this rule
2340 * a stream of single byte packets will cause the right side of the
2341 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002342 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 * Of course, if the sender implements sender side SWS prevention
2344 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002345 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002347 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 * If the free space is less than the 1/4 of the maximum
2349 * space available and the free space is less than 1/2 mss,
2350 * then set the window to 0.
2351 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2352 * Otherwise, just prevent the window from shrinking
2353 * and from being larger than the largest representable value.
2354 *
2355 * This prevents incremental opening of the window in the regime
2356 * where TCP is limited by the speed of the reader side taking
2357 * data out of the TCP receive queue. It does nothing about
2358 * those cases where the window is constrained on the sender side
2359 * because the pipeline is full.
2360 *
2361 * BSD also seems to "accidentally" limit itself to windows that are a
2362 * multiple of MSS, at least until the free space gets quite small.
2363 * This would appear to be a side effect of the mbuf implementation.
2364 * Combining these two algorithms results in the observed behavior
2365 * of having a fixed window size at almost all times.
2366 *
2367 * Below we obtain similar behavior by forcing the offered window to
2368 * a multiple of the mss when it is feasible to do so.
2369 *
2370 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2371 * Regular options like TIMESTAMP are taken into account.
2372 */
2373u32 __tcp_select_window(struct sock *sk)
2374{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002375 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002377 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 * here. I don't know if the value based on our guesses
2379 * of peer's MSS is better for the performance. It's more correct
2380 * but may be worse for the performance because of rcv_mss
2381 * fluctuations. --SAW 1998/11/1
2382 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002383 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 int free_space = tcp_space(sk);
Florian Westphal86c1a042014-02-19 12:51:10 +01002385 int allowed_space = tcp_full_space(sk);
2386 int full_space = min_t(int, tp->window_clamp, allowed_space);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 int window;
2388
2389 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002390 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002392 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002393 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394
Glauber Costa180d8cd2011-12-11 21:47:02 +00002395 if (sk_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002396 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2397 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398
Florian Westphal86c1a042014-02-19 12:51:10 +01002399 /* free_space might become our new window, make sure we don't
2400 * increase it due to wscale.
2401 */
2402 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2403
2404 /* if free space is less than mss estimate, or is below 1/16th
2405 * of the maximum allowed, try to move to zero-window, else
2406 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2407 * new incoming data is dropped due to memory limits.
2408 * With large window, mss test triggers way too late in order
2409 * to announce zero window in time before rmem limit kicks in.
2410 */
2411 if (free_space < (allowed_space >> 4) || free_space < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 return 0;
2413 }
2414
2415 if (free_space > tp->rcv_ssthresh)
2416 free_space = tp->rcv_ssthresh;
2417
2418 /* Don't do rounding if we are using window scaling, since the
2419 * scaled window will not line up with the MSS boundary anyway.
2420 */
2421 window = tp->rcv_wnd;
2422 if (tp->rx_opt.rcv_wscale) {
2423 window = free_space;
2424
2425 /* Advertise enough space so that it won't get scaled away.
2426 * Import case: prevent zero window announcement if
2427 * 1<<rcv_wscale > mss.
2428 */
2429 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2430 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2431 << tp->rx_opt.rcv_wscale);
2432 } else {
2433 /* Get the largest window that is a nice multiple of mss.
2434 * Window clamp already applied above.
2435 * If our current window offering is within 1 mss of the
2436 * free space we just keep it. This prevents the divide
2437 * and multiply from happening most of the time.
2438 * We also don't do any window rounding when the free space
2439 * is too small.
2440 */
2441 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002442 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002443 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002444 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002445 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 }
2447
2448 return window;
2449}
2450
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002451/* Collapses two adjacent SKB's during retransmission. */
2452static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453{
2454 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002455 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002456 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002458 skb_size = skb->len;
2459 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002461 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002462
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002463 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002465 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002467 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2468 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002470 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2471 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002473 if (skb->ip_summed != CHECKSUM_PARTIAL)
2474 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002476 /* Update sequence range on original skb. */
2477 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002479 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002480 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002482 /* All done, get rid of second SKB and account for it so
2483 * packet counting does not break.
2484 */
2485 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002486
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002487 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002488 tcp_clear_retrans_hints_partial(tp);
2489 if (next_skb == tp->retransmit_skb_hint)
2490 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002491
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002492 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2493
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002494 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495}
2496
Andi Kleen67edfef2009-07-21 23:00:40 +00002497/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002498static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002499{
2500 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002501 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002502 /* TODO: SACK collapsing could be used to remove this condition */
2503 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002504 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002505 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002506 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002507 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002508 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002509 /* Some heurestics for collapsing over SACK'd could be invented */
2510 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002511 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002512
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002513 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002514}
2515
Andi Kleen67edfef2009-07-21 23:00:40 +00002516/* Collapse packets in the retransmit queue to make to create
2517 * less packets on the wire. This is only done on retransmission.
2518 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002519static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2520 int space)
2521{
2522 struct tcp_sock *tp = tcp_sk(sk);
2523 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002524 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002525
2526 if (!sysctl_tcp_retrans_collapse)
2527 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002528 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002529 return;
2530
2531 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2532 if (!tcp_can_collapse(sk, skb))
2533 break;
2534
2535 space -= skb->len;
2536
2537 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002538 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002539 continue;
2540 }
2541
2542 if (space < 0)
2543 break;
2544 /* Punt if not enough space exists in the first SKB for
2545 * the data in the second
2546 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002547 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002548 break;
2549
2550 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2551 break;
2552
2553 tcp_collapse_retrans(sk, to);
2554 }
2555}
2556
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557/* This retransmits one SKB. Policy decisions and retransmit queue
2558 * state updates are done by the caller. Returns non-zero if an
2559 * error occurred which prevented the send.
2560 */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002561int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562{
2563 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002564 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002565 unsigned int cur_mss;
Yuchung Chengc84a5712014-02-28 16:42:26 -08002566 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567
John Heffner5d424d52006-03-20 17:53:41 -08002568 /* Inconslusive MTU probe */
2569 if (icsk->icsk_mtup.probe_size) {
2570 icsk->icsk_mtup.probe_size = 0;
2571 }
2572
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002574 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 */
2576 if (atomic_read(&sk->sk_wmem_alloc) >
2577 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2578 return -EAGAIN;
2579
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002580 if (skb_still_in_host_queue(sk, skb))
2581 return -EBUSY;
2582
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2584 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2585 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2587 return -ENOMEM;
2588 }
2589
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002590 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2591 return -EHOSTUNREACH; /* Routing failure or similar. */
2592
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002593 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002594
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 /* If receiver has shrunk his window, and skb is out of
2596 * new window, do not retransmit it. The exception is the
2597 * case, when window is shrunk to zero. In this case
2598 * our retransmit serves as a zero window probe.
2599 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002600 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2601 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 return -EAGAIN;
2603
2604 if (skb->len > cur_mss) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002605 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002607 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002608 int oldpcount = tcp_skb_pcount(skb);
2609
2610 if (unlikely(oldpcount > 1)) {
Eric Dumazetc52e24212013-10-15 11:54:30 -07002611 if (skb_unclone(skb, GFP_ATOMIC))
2612 return -ENOMEM;
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002613 tcp_init_tso_segs(sk, skb, cur_mss);
2614 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2615 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 }
2617
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002618 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 /* Make a copy, if the first transmission SKB clone we made
2621 * is still in somebody's hands, else make a clone.
2622 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
Thomas Graf50bceae2013-04-11 10:57:18 +00002624 /* make sure skb->data is aligned on arches that require it
2625 * and check if ack-trimming & collapsing extended the headroom
2626 * beyond what csum_start can cover.
2627 */
2628 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2629 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet117632e2011-12-03 21:39:53 +00002630 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2631 GFP_ATOMIC);
Yuchung Chengc84a5712014-02-28 16:42:26 -08002632 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2633 -ENOBUFS;
Eric Dumazet117632e2011-12-03 21:39:53 +00002634 } else {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002635 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002636 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002637
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002638 if (likely(!err)) {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002639 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002640 /* Update global TCP statistics. */
2641 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2642 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2643 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2644 tp->total_retrans++;
2645 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002646 return err;
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002647}
2648
2649int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2650{
2651 struct tcp_sock *tp = tcp_sk(sk);
2652 int err = __tcp_retransmit_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
2654 if (err == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002656 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002657 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 }
2659#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002660 if (!tp->retrans_out)
2661 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2663 tp->retrans_out += tcp_skb_pcount(skb);
2664
2665 /* Save stamp of the first retransmit. */
2666 if (!tp->retrans_stamp)
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002667 tp->retrans_stamp = tcp_skb_timestamp(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 /* snd_nxt is stored to detect loss of retransmitted segment,
2670 * see tcp_input.c tcp_sacktag_write_queue().
2671 */
2672 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002673 } else if (err != -EBUSY) {
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002674 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675 }
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002676
2677 if (tp->undo_retrans < 0)
2678 tp->undo_retrans = 0;
2679 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 return err;
2681}
2682
Andi Kleen67edfef2009-07-21 23:00:40 +00002683/* Check if we forward retransmits are possible in the current
2684 * window/congestion state.
2685 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002686static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002687{
2688 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002689 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002690
2691 /* Forward retransmissions are possible only during Recovery. */
2692 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002693 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002694
2695 /* No forward retransmissions in Reno are possible. */
2696 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002697 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002698
2699 /* Yeah, we have to make difficult choice between forward transmission
2700 * and retransmission... Both ways have their merits...
2701 *
2702 * For now we do not retransmit anything, while we have some new
2703 * segments to send. In the other cases, follow rule 3 for
2704 * NextSeg() specified in RFC3517.
2705 */
2706
2707 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002708 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002709
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002710 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002711}
2712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713/* This gets called after a retransmit timeout, and the initially
2714 * retransmitted data is acknowledged. It tries to continue
2715 * resending the rest of the retransmit queue, until either
2716 * we've sent it all or the congestion window limit is reached.
2717 * If doing SACK, the first ACK which comes back for a timeout
2718 * based retransmit packet might feed us FACK information again.
2719 * If so, we use it to avoid unnecessarily retransmissions.
2720 */
2721void tcp_xmit_retransmit_queue(struct sock *sk)
2722{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002723 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 struct tcp_sock *tp = tcp_sk(sk);
2725 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002726 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002727 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002728 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002729 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002730
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002731 if (!tp->packets_out)
2732 return;
2733
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002734 if (!tp->lost_out)
2735 tp->retransmit_high = tp->snd_una;
2736
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002737 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002738 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002739 last_lost = TCP_SKB_CB(skb)->end_seq;
2740 if (after(last_lost, tp->retransmit_high))
2741 last_lost = tp->retransmit_high;
2742 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002743 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002744 last_lost = tp->snd_una;
2745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002747 tcp_for_write_queue_from(skb, sk) {
2748 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002750 if (skb == tcp_send_head(sk))
2751 break;
2752 /* we could do better than to assign each time */
Ian Morris51456b22015-04-03 09:17:26 +01002753 if (!hole)
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002754 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002755
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002756 /* Assume this retransmit will generate
2757 * only one packet for congestion window
2758 * calculation purposes. This works because
2759 * tcp_retransmit_skb() will chop up the
2760 * packet to be MSS sized and all the
2761 * packet counting works out.
2762 */
2763 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2764 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002765
2766 if (fwd_rexmitting) {
2767begin_fwd:
2768 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2769 break;
2770 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2771
2772 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002773 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002774 if (!tcp_can_forward_retransmit(sk))
2775 break;
2776 /* Backtrack if necessary to non-L'ed skb */
Ian Morris00db4122015-04-03 09:17:27 +01002777 if (hole) {
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002778 skb = hole;
2779 hole = NULL;
2780 }
2781 fwd_rexmitting = 1;
2782 goto begin_fwd;
2783
2784 } else if (!(sacked & TCPCB_LOST)) {
Ian Morris51456b22015-04-03 09:17:26 +01002785 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002786 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002787 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002789 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002790 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002791 if (icsk->icsk_ca_state != TCP_CA_Loss)
2792 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2793 else
2794 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2795 }
2796
2797 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002798 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002799
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002800 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002801 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002802
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002803 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2804
Yuchung Cheng684bad12012-09-02 17:38:04 +00002805 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002806 tp->prr_out += tcp_skb_pcount(skb);
2807
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002808 if (skb == tcp_write_queue_head(sk))
2809 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2810 inet_csk(sk)->icsk_rto,
2811 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813}
2814
Eric Dumazetd83769a2015-04-21 18:32:24 -07002815/* We allow to exceed memory limits for FIN packets to expedite
2816 * connection tear down and (memory) recovery.
Eric Dumazet845704a2015-04-23 10:42:39 -07002817 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2818 * or even be forced to close flow without any FIN.
Eric Dumazetd83769a2015-04-21 18:32:24 -07002819 */
2820static void sk_forced_wmem_schedule(struct sock *sk, int size)
2821{
2822 int amt, status;
2823
2824 if (size <= sk->sk_forward_alloc)
2825 return;
2826 amt = sk_mem_pages(size);
2827 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2828 sk_memory_allocated_add(sk, amt, &status);
2829}
2830
Eric Dumazet845704a2015-04-23 10:42:39 -07002831/* Send a FIN. The caller locks the socket for us.
2832 * We should try to send a FIN packet really hard, but eventually give up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 */
2834void tcp_send_fin(struct sock *sk)
2835{
Eric Dumazet845704a2015-04-23 10:42:39 -07002836 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002837 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002838
Eric Dumazet845704a2015-04-23 10:42:39 -07002839 /* Optimization, tack on the FIN if we have one skb in write queue and
2840 * this skb was not yet sent, or we are under memory pressure.
2841 * Note: in the latter case, FIN packet will be sent after a timeout,
2842 * as TCP stack thinks it has already been transmitted.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 */
Eric Dumazet845704a2015-04-23 10:42:39 -07002844 if (tskb && (tcp_send_head(sk) || sk_under_memory_pressure(sk))) {
2845coalesce:
2846 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2847 TCP_SKB_CB(tskb)->end_seq++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 tp->write_seq++;
Eric Dumazet845704a2015-04-23 10:42:39 -07002849 if (!tcp_send_head(sk)) {
2850 /* This means tskb was already sent.
2851 * Pretend we included the FIN on previous transmit.
2852 * We need to set tp->snd_nxt to the value it would have
2853 * if FIN had been sent. This is because retransmit path
2854 * does not change tp->snd_nxt.
2855 */
2856 tp->snd_nxt++;
2857 return;
2858 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 } else {
Eric Dumazet845704a2015-04-23 10:42:39 -07002860 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2861 if (unlikely(!skb)) {
2862 if (tskb)
2863 goto coalesce;
2864 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 }
Eric Dumazetd83769a2015-04-21 18:32:24 -07002866 skb_reserve(skb, MAX_TCP_HEADER);
2867 sk_forced_wmem_schedule(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002869 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002870 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 tcp_queue_skb(sk, skb);
2872 }
Eric Dumazet845704a2015-04-23 10:42:39 -07002873 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874}
2875
2876/* We get here when a process closes a file descriptor (either due to
2877 * an explicit close() or as a byproduct of exit()'ing) and there
2878 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002879 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 */
Al Virodd0fc662005-10-07 07:46:04 +01002881void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 struct sk_buff *skb;
2884
2885 /* NOTE: No TCP options attached and we never retransmit this. */
2886 skb = alloc_skb(MAX_TCP_HEADER, priority);
2887 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002888 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 return;
2890 }
2891
2892 /* Reserve space for headers and prepare control bits. */
2893 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002894 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002895 TCPHDR_ACK | TCPHDR_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 /* Send it off. */
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002897 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002898 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002899
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002900 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901}
2902
Andi Kleen67edfef2009-07-21 23:00:40 +00002903/* Send a crossed SYN-ACK during socket establishment.
2904 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 * a SYN packet that crossed the incoming SYN that caused this routine
2906 * to get called. If this assumption fails then the initial rcv_wnd
2907 * and rcv_wscale values will not be correct.
2908 */
2909int tcp_send_synack(struct sock *sk)
2910{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002911 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912
David S. Millerfe067e82007-03-07 12:12:44 -08002913 skb = tcp_write_queue_head(sk);
Ian Morris51456b22015-04-03 09:17:26 +01002914 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00002915 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 return -EFAULT;
2917 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04002918 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 if (skb_cloned(skb)) {
2920 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
Ian Morris51456b22015-04-03 09:17:26 +01002921 if (!nskb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002923 tcp_unlink_write_queue(skb, sk);
Eric Dumazetf4a775d2014-09-22 16:29:32 -07002924 __skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002925 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002926 sk_wmem_free_skb(sk, skb);
2927 sk->sk_wmem_queued += nskb->truesize;
2928 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929 skb = nskb;
2930 }
2931
Eric Dumazet4de075e2011-09-27 13:25:05 -04002932 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Florian Westphal735d3832014-09-29 13:08:30 +02002933 tcp_ecn_send_synack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002935 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936}
2937
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002938/**
2939 * tcp_make_synack - Prepare a SYN-ACK.
2940 * sk: listener socket
2941 * dst: dst entry attached to the SYNACK
2942 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002943 *
2944 * Allocate one skb and build a SYNACK packet.
2945 * @dst is consumed : Caller should not use it again.
2946 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002947struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002948 struct request_sock *req,
Jerry Chu83368862012-08-31 12:29:12 +00002949 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002951 struct tcp_out_options opts;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002952 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 struct tcp_sock *tp = tcp_sk(sk);
2954 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 struct sk_buff *skb;
Eric Dumazet80f03e22015-03-24 15:58:52 -07002956 struct tcp_md5sig_key *md5 = NULL;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002957 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002958 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959
Eric Dumazeta0b84862014-03-26 09:57:19 -07002960 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002961 if (unlikely(!skb)) {
2962 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002964 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 /* Reserve space for headers. */
2966 skb_reserve(skb, MAX_TCP_HEADER);
2967
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002968 skb_dst_set(skb, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969
David S. Miller0dbaee32010-12-13 12:52:14 -08002970 mss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002971 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2972 mss = tp->rx_opt.user_mss;
2973
Adam Langley33ad7982008-07-19 00:04:31 -07002974 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002975#ifdef CONFIG_SYN_COOKIES
2976 if (unlikely(req->cookie_ts))
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002977 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002978 else
2979#endif
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002980 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazet80f03e22015-03-24 15:58:52 -07002981
2982#ifdef CONFIG_TCP_MD5SIG
2983 rcu_read_lock();
Eric Dumazetfd3a1542015-03-24 15:58:56 -07002984 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
Eric Dumazet80f03e22015-03-24 15:58:52 -07002985#endif
2986 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
Christoph Paasch1a2c6182013-03-17 08:23:34 +00002987 foc) + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002988
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002989 skb_push(skb, tcp_header_size);
2990 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002992 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 memset(th, 0, sizeof(struct tcphdr));
2994 th->syn = 1;
2995 th->ack = 1;
Florian Westphal735d3832014-09-29 13:08:30 +02002996 tcp_ecn_make_synack(req, th, sk);
Eric Dumazetb44084c2013-10-10 00:04:37 -07002997 th->source = htons(ireq->ir_num);
Eric Dumazet634fb9792013-10-09 15:21:29 -07002998 th->dest = ireq->ir_rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002999 /* Setting of flags are superfluous here for callers (and ECE is
3000 * not even correctly set)
3001 */
3002 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00003003 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faade2009-12-02 18:25:27 +00003004
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Jerry Chu83368862012-08-31 12:29:12 +00003006 /* XXX data is queued and acked as is. No buffer/window check */
3007 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008
3009 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08003010 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003011 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 th->doff = (tcp_header_size >> 2);
Eric Dumazeta0b84862014-03-26 09:57:19 -07003013 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003014
3015#ifdef CONFIG_TCP_MD5SIG
3016 /* Okay, we have all we need - do the md5 hash if needed */
Eric Dumazet80f03e22015-03-24 15:58:52 -07003017 if (md5)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003018 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07003019 md5, req_to_sk(req), skb);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003020 rcu_read_unlock();
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003021#endif
3022
Eric Dumazetb50edd72015-04-09 13:31:56 -07003023 /* Do not fool tcpdump (if any), clean our debris */
3024 skb->tstamp.tv64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 return skb;
3026}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003027EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028
Daniel Borkmann81164412015-01-05 23:57:48 +01003029static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3030{
3031 struct inet_connection_sock *icsk = inet_csk(sk);
3032 const struct tcp_congestion_ops *ca;
3033 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3034
3035 if (ca_key == TCP_CA_UNSPEC)
3036 return;
3037
3038 rcu_read_lock();
3039 ca = tcp_ca_find_key(ca_key);
3040 if (likely(ca && try_module_get(ca->owner))) {
3041 module_put(icsk->icsk_ca_ops->owner);
3042 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3043 icsk->icsk_ca_ops = ca;
3044 }
3045 rcu_read_unlock();
3046}
3047
Andi Kleen67edfef2009-07-21 23:00:40 +00003048/* Do all connect socket setups that can be done AF independent. */
stephen hemmingerf7e56a72013-12-29 11:39:51 -08003049static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003051 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052 struct tcp_sock *tp = tcp_sk(sk);
3053 __u8 rcv_wscale;
3054
3055 /* We'll fix this up when we get a response from the other end.
3056 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3057 */
3058 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08003059 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003061#ifdef CONFIG_TCP_MD5SIG
Ian Morris00db4122015-04-03 09:17:27 +01003062 if (tp->af_specific->md5_lookup(sk, sk))
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003063 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3064#endif
3065
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 /* If user gave his TCP_MAXSEG, record it to clamp */
3067 if (tp->rx_opt.user_mss)
3068 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3069 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08003070 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 tcp_sync_mss(sk, dst_mtu(dst));
3072
Daniel Borkmann81164412015-01-05 23:57:48 +01003073 tcp_ca_dst_init(sk, dst);
3074
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 if (!tp->window_clamp)
3076 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08003077 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003078 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3079 tp->advmss = tp->rx_opt.user_mss;
3080
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00003083 /* limit the window selection if the user enforce a smaller rx buffer */
3084 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3085 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3086 tp->window_clamp = tcp_full_space(sk);
3087
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 tcp_select_initial_window(tcp_full_space(sk),
3089 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3090 &tp->rcv_wnd,
3091 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08003092 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00003093 &rcv_wscale,
3094 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095
3096 tp->rx_opt.rcv_wscale = rcv_wscale;
3097 tp->rcv_ssthresh = tp->rcv_wnd;
3098
3099 sk->sk_err = 0;
3100 sock_reset_flag(sk, SOCK_DONE);
3101 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003102 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103 tp->snd_una = tp->write_seq;
3104 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003105 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00003106 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003107
3108 if (likely(!tp->repair))
3109 tp->rcv_nxt = 0;
Andrew Vaginc7781a62013-08-27 12:20:40 +04003110 else
3111 tp->rcv_tstamp = tcp_time_stamp;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003112 tp->rcv_wup = tp->rcv_nxt;
3113 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003115 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3116 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 tcp_clear_retrans(tp);
3118}
3119
Yuchung Cheng783237e2012-07-19 06:43:07 +00003120static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3121{
3122 struct tcp_sock *tp = tcp_sk(sk);
3123 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3124
3125 tcb->end_seq += skb->len;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07003126 __skb_header_release(skb);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003127 __tcp_add_write_queue_tail(sk, skb);
3128 sk->sk_wmem_queued += skb->truesize;
3129 sk_mem_charge(sk, skb->truesize);
3130 tp->write_seq = tcb->end_seq;
3131 tp->packets_out += tcp_skb_pcount(skb);
3132}
3133
3134/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3135 * queue a data-only packet after the regular SYN, such that regular SYNs
3136 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3137 * only the SYN sequence, the data are retransmitted in the first ACK.
3138 * If cookie is not cached or other error occurs, falls back to send a
3139 * regular SYN with Fast Open cookie request option.
3140 */
3141static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3142{
3143 struct tcp_sock *tp = tcp_sk(sk);
3144 struct tcp_fastopen_request *fo = tp->fastopen_req;
Al Viro57be5bd2014-11-28 13:40:20 -05003145 int syn_loss = 0, space, err = 0, copied;
Yuchung Chengaab48742012-07-19 06:43:10 +00003146 unsigned long last_syn_loss = 0;
Eric Dumazet355a9012014-11-17 23:06:20 -08003147 struct sk_buff *syn_data;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003148
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003149 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00003150 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3151 &syn_loss, &last_syn_loss);
3152 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3153 if (syn_loss > 1 &&
3154 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3155 fo->cookie.len = -1;
3156 goto fallback;
3157 }
3158
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003159 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3160 fo->cookie.len = -1;
3161 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003162 goto fallback;
3163
3164 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3165 * user-MSS. Reserve maximum option space for middleboxes that add
3166 * private TCP options. The cost is reduced data space in SYN :(
3167 */
3168 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3169 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00003170 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00003171 MAX_TCP_OPTION_SPACE;
3172
Eric Dumazetf5ddcbb2014-02-20 10:09:18 -08003173 space = min_t(size_t, space, fo->size);
3174
3175 /* limit to order-0 allocations */
3176 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3177
Eric Dumazet355a9012014-11-17 23:06:20 -08003178 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation);
3179 if (!syn_data)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003180 goto fallback;
Eric Dumazet355a9012014-11-17 23:06:20 -08003181 syn_data->ip_summed = CHECKSUM_PARTIAL;
3182 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
Al Viro57be5bd2014-11-28 13:40:20 -05003183 copied = copy_from_iter(skb_put(syn_data, space), space,
3184 &fo->data->msg_iter);
3185 if (unlikely(!copied)) {
Eric Dumazet355a9012014-11-17 23:06:20 -08003186 kfree_skb(syn_data);
3187 goto fallback;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003188 }
Al Viro57be5bd2014-11-28 13:40:20 -05003189 if (copied != space) {
3190 skb_trim(syn_data, copied);
3191 space = copied;
3192 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003193
Eric Dumazet355a9012014-11-17 23:06:20 -08003194 /* No more data pending in inet_wait_for_connect() */
3195 if (space == fo->size)
3196 fo->data = NULL;
3197 fo->copied = space;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003198
Eric Dumazet355a9012014-11-17 23:06:20 -08003199 tcp_connect_queue_skb(sk, syn_data);
3200
3201 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3202
3203 syn->skb_mstamp = syn_data->skb_mstamp;
3204
3205 /* Now full SYN+DATA was cloned and sent (or not),
3206 * remove the SYN from the original skb (syn_data)
3207 * we keep in write queue in case of a retransmit, as we
3208 * also have the SYN packet (with no data) in the same queue.
Eric Dumazet431a9122014-03-09 17:36:02 -07003209 */
Eric Dumazet355a9012014-11-17 23:06:20 -08003210 TCP_SKB_CB(syn_data)->seq++;
3211 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3212 if (!err) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003213 tp->syn_data = (fo->copied > 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08003214 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003215 goto done;
3216 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003217
3218fallback:
3219 /* Send a regular SYN with Fast Open cookie request option */
3220 if (fo->cookie.len > 0)
3221 fo->cookie.len = 0;
3222 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3223 if (err)
3224 tp->syn_fastopen = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003225done:
3226 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3227 return err;
3228}
3229
Andi Kleen67edfef2009-07-21 23:00:40 +00003230/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003231int tcp_connect(struct sock *sk)
3232{
3233 struct tcp_sock *tp = tcp_sk(sk);
3234 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00003235 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236
3237 tcp_connect_init(sk);
3238
Andrey Vagin2b916472012-11-22 01:13:58 +00003239 if (unlikely(tp->repair)) {
3240 tcp_finish_connect(sk, NULL);
3241 return 0;
3242 }
3243
Eric Dumazet355a9012014-11-17 23:06:20 -08003244 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
3245 if (unlikely(!buff))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 return -ENOBUFS;
3247
Changli Gaoa3433f32010-06-12 14:01:43 +00003248 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003249 tp->retrans_stamp = tcp_time_stamp;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003250 tcp_connect_queue_skb(sk, buff);
Florian Westphal735d3832014-09-29 13:08:30 +02003251 tcp_ecn_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252
Yuchung Cheng783237e2012-07-19 06:43:07 +00003253 /* Send off SYN; include data in Fast Open. */
3254 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3255 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00003256 if (err == -ECONNREFUSED)
3257 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07003258
3259 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3260 * in order to make this packet get counted in tcpOutSegs.
3261 */
3262 tp->snd_nxt = tp->write_seq;
3263 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003264 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265
3266 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003267 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3268 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269 return 0;
3270}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003271EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272
3273/* Send out a delayed ack, the caller does the policy checking
3274 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3275 * for details.
3276 */
3277void tcp_send_delayed_ack(struct sock *sk)
3278{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003279 struct inet_connection_sock *icsk = inet_csk(sk);
3280 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 unsigned long timeout;
3282
Florian Westphal98900922014-09-26 22:37:35 +02003283 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3284
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003286 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003287 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003289 if (icsk->icsk_ack.pingpong ||
3290 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 max_ato = TCP_DELACK_MAX;
3292
3293 /* Slow path, intersegment interval is "high". */
3294
3295 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003296 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 * directly.
3298 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003299 if (tp->srtt_us) {
3300 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3301 TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
3303 if (rtt < max_ato)
3304 max_ato = rtt;
3305 }
3306
3307 ato = min(ato, max_ato);
3308 }
3309
3310 /* Stay within the limit we were given */
3311 timeout = jiffies + ato;
3312
3313 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003314 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 /* If delack timer was blocked or is about to expire,
3316 * send ACK now.
3317 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003318 if (icsk->icsk_ack.blocked ||
3319 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320 tcp_send_ack(sk);
3321 return;
3322 }
3323
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003324 if (!time_before(timeout, icsk->icsk_ack.timeout))
3325 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003327 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3328 icsk->icsk_ack.timeout = timeout;
3329 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330}
3331
3332/* This routine sends an ack and also updates the window. */
3333void tcp_send_ack(struct sock *sk)
3334{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003335 struct sk_buff *buff;
3336
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003338 if (sk->sk_state == TCP_CLOSE)
3339 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003340
Florian Westphal98900922014-09-26 22:37:35 +02003341 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3342
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003343 /* We are not putting this on the write queue, so
3344 * tcp_transmit_skb() will set the ownership to this
3345 * sock.
3346 */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003347 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ian Morris51456b22015-04-03 09:17:26 +01003348 if (!buff) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003349 inet_csk_schedule_ack(sk);
3350 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3351 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3352 TCP_DELACK_MAX, TCP_RTO_MAX);
3353 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003355
3356 /* Reserve space for headers and prepare control bits. */
3357 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003358 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003359
Eric Dumazet98781962015-02-03 18:31:53 -08003360 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3361 * too much.
3362 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3363 * We also avoid tcp_wfree() overhead (cache line miss accessing
3364 * tp->tsq_flags) by using regular sock_wfree()
3365 */
3366 skb_set_tcp_pure_ack(buff);
3367
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003368 /* Send it off, this clears delayed acks for us. */
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003369 skb_mstamp_get(&buff->skb_mstamp);
Mel Gorman99a1dec2012-07-31 16:44:14 -07003370 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371}
Daniel Borkmanne3118e82014-09-26 22:37:36 +02003372EXPORT_SYMBOL_GPL(tcp_send_ack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
3374/* This routine sends a packet with an out of date sequence
3375 * number. It assumes the other end will try to ack it.
3376 *
3377 * Question: what should we make while urgent mode?
3378 * 4.4BSD forces sending single byte of data. We cannot send
3379 * out of window data, because we have SND.NXT==SND.MAX...
3380 *
3381 * Current solution: to send TWO zero-length segments in urgent mode:
3382 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3383 * out-of-date with SND.UNA-1 to probe window.
3384 */
3385static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3386{
3387 struct tcp_sock *tp = tcp_sk(sk);
3388 struct sk_buff *skb;
3389
3390 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003391 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ian Morris51456b22015-04-03 09:17:26 +01003392 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 return -1;
3394
3395 /* Reserve space for headers and set control bits. */
3396 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397 /* Use a previous sequence. This should cause the other
3398 * end to send an ack. Don't queue or clone SKB, just
3399 * send it.
3400 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003401 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003402 skb_mstamp_get(&skb->skb_mstamp);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003403 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404}
3405
Pavel Emelyanovee995282012-04-19 03:40:39 +00003406void tcp_send_window_probe(struct sock *sk)
3407{
3408 if (sk->sk_state == TCP_ESTABLISHED) {
3409 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3410 tcp_xmit_probe_skb(sk, 0);
3411 }
3412}
3413
Andi Kleen67edfef2009-07-21 23:00:40 +00003414/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415int tcp_write_wakeup(struct sock *sk)
3416{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003417 struct tcp_sock *tp = tcp_sk(sk);
3418 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003419
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003420 if (sk->sk_state == TCP_CLOSE)
3421 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Ian Morris00db4122015-04-03 09:17:27 +01003423 skb = tcp_send_head(sk);
3424 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003425 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003426 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003427 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003429 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3430 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003432 /* We are probing the opening of a window
3433 * but the window size is != 0
3434 * must have been a result SWS avoidance ( sender )
3435 */
3436 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3437 skb->len > mss) {
3438 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003439 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Octavian Purdila6cc55e02014-06-06 17:32:37 +03003440 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003441 return -1;
3442 } else if (!tcp_skb_pcount(skb))
3443 tcp_set_skb_tso_segs(sk, skb, mss);
3444
Eric Dumazet4de075e2011-09-27 13:25:05 -04003445 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003446 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3447 if (!err)
3448 tcp_event_new_data_sent(sk, skb);
3449 return err;
3450 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003451 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003452 tcp_xmit_probe_skb(sk, 1);
3453 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455}
3456
3457/* A window probe timeout has occurred. If window is not closed send
3458 * a partial packet else a zero probe.
3459 */
3460void tcp_send_probe0(struct sock *sk)
3461{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003462 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003464 unsigned long probe_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465 int err;
3466
3467 err = tcp_write_wakeup(sk);
3468
David S. Millerfe067e82007-03-07 12:12:44 -08003469 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003471 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003472 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 return;
3474 }
3475
3476 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003477 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3478 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003479 icsk->icsk_probes_out++;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003480 probe_max = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 } else {
3482 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003483 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 * Let local senders to fight for local resources.
3485 *
3486 * Use accumulated backoff yet.
3487 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003488 if (!icsk->icsk_probes_out)
3489 icsk->icsk_probes_out = 1;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003490 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 }
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003492 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3493 inet_csk_rto_backoff(icsk, probe_max),
3494 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495}
Octavian Purdila5db92c92014-06-25 17:09:59 +03003496
3497int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3498{
3499 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3500 struct flowi fl;
3501 int res;
3502
3503 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3504 if (!res) {
3505 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3506 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3507 }
3508 return res;
3509}
3510EXPORT_SYMBOL(tcp_rtx_synack);