blob: 6f35cdd5f2f09ae771168d62b5c3dd90dc626e45 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
Joe Perches91df42b2012-05-15 14:11:54 +000037#define pr_fmt(fmt) "TCP: " fmt
38
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <net/tcp.h>
40
41#include <linux/compiler.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090042#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070046int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Jianjun Kong09cb1052008-11-03 00:27:11 -080048/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080049 * interpret the window field as a signed quantity.
50 */
Brian Haleyab32ea52006-09-22 14:15:41 -070051int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080052
Wei Liuc39c4c62015-06-03 11:10:42 +010053/* Default TSQ limit of four TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
Eric Dumazet46d3cea2012-07-11 05:50:31 +000055
Linus Torvalds1da177e2005-04-16 15:20:36 -070056/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
Brian Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
David S. Miller35089bb2006-06-13 22:33:04 -070062/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070063int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070064
Eric Dumazet46d3cea2012-07-11 05:50:31 +000065static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000067
Andi Kleen67edfef2009-07-21 23:00:40 +000068/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040069static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070070{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000071 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070072 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080073 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070074
David S. Millerfe067e82007-03-07 12:12:44 -080075 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070076 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080077
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080078 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000079 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000080 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000081 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000082 }
Yuchung Chengf19c29e2014-03-03 12:31:36 -080083
David S. Millerf7324ac2014-03-06 15:03:17 -050084 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 tcp_skb_pcount(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -070086}
87
88/* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040094static inline __u32 tcp_acceptable_seq(const struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070095{
Eric Dumazetcf533ea2011-10-21 05:22:42 -040096 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070097
Ilpo Järvinen90840de2007-12-31 04:48:41 -080098 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070099 return tp->snd_nxt;
100 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -0800101 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102}
103
104/* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 *
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
111 * large MSS.
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
117 */
118static __u16 tcp_advertise_mss(struct sock *sk)
119{
120 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400121 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 int mss = tp->advmss;
123
David S. Miller0dbaee32010-12-13 12:52:14 -0800124 if (dst) {
125 unsigned int metric = dst_metric_advmss(dst);
126
127 if (metric < mss) {
128 mss = metric;
129 tp->advmss = mss;
130 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 }
132
133 return (__u16)mss;
134}
135
136/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
Eric Dumazet6f021c62015-08-21 12:30:00 -0700137 * This is the first part of cwnd validation mechanism.
138 */
139void tcp_cwnd_restart(struct sock *sk, s32 delta)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700141 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f021c62015-08-21 12:30:00 -0700142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 u32 cwnd = tp->snd_cwnd;
144
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 restart_cwnd = min(restart_cwnd, cwnd);
149
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 cwnd >>= 1;
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
155}
156
Andi Kleen67edfef2009-07-21 23:00:40 +0000157/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800158static void tcp_event_data_sent(struct tcp_sock *tp,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400159 struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Neal Cardwell05c5a462015-09-09 21:54:37 -0700164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
166
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
Jon Maxwell2251ae42015-07-08 10:12:28 +1000172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174}
175
Andi Kleen67edfef2009-07-21 23:00:40 +0000176/* Account for an ACK we sent. */
Yuchung Cheng57ec8822018-07-18 13:56:35 -0700177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
178 u32 rcv_nxt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179{
Yuchung Cheng57ec8822018-07-18 13:56:35 -0700180 struct tcp_sock *tp = tcp_sk(sk);
181
182 if (unlikely(rcv_nxt != tp->rcv_nxt))
183 return; /* Special ACK sent by DCTCP to reflect ECN */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700184 tcp_dec_quickack_mode(sk, pkts);
185 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186}
187
Yuchung Cheng85f16522013-06-11 15:35:32 -0700188
189u32 tcp_default_init_rwnd(u32 mss)
190{
191 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800192 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700193 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
194 * limit when mss is larger than 1460.
195 */
196 u32 init_rwnd = TCP_INIT_CWND * 2;
197
198 if (mss > 1460)
199 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
200 return init_rwnd;
201}
202
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203/* Determine a window scaling and initial window to offer.
204 * Based on the assumption that the given amount of space
205 * will be offered. Store the results in the tp structure.
206 * NOTE: for smooth operation initial space offering should
207 * be a multiple of mss if possible. We assume here that mss >= 1.
208 * This MUST be enforced by all callers.
209 */
210void tcp_select_initial_window(int __space, __u32 mss,
211 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000212 int wscale_ok, __u8 *rcv_wscale,
213 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214{
215 unsigned int space = (__space < 0 ? 0 : __space);
216
217 /* If no clamp set the clamp to the max possible scaled window */
218 if (*window_clamp == 0)
219 (*window_clamp) = (65535 << 14);
220 space = min(*window_clamp, space);
221
222 /* Quantize space offering to a multiple of mss if possible. */
223 if (space > mss)
224 space = (space / mss) * mss;
225
226 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800227 * will break some buggy TCP stacks. If the admin tells us
228 * it is likely we could be speaking with such a buggy stack
229 * we will truncate our initial window offering to 32K-1
230 * unless the remote has sent us a window scaling option,
231 * which we interpret as a sign the remote TCP is not
232 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233 */
Rick Jones15d99e02006-03-20 22:40:29 -0800234 if (sysctl_tcp_workaround_signed_windows)
235 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
236 else
237 (*rcv_wnd) = space;
238
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239 (*rcv_wscale) = 0;
240 if (wscale_ok) {
241 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900242 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700243 */
Soheil Hassas Yeganehf6263002016-07-29 09:34:02 -0400244 space = max_t(u32, space, sysctl_tcp_rmem[2]);
245 space = max_t(u32, space, 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
Daniel Borkmann49213552015-05-19 21:04:22 +0200353static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
354{
355 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
356 /* tp->ecn_flags are cleared at a later point in time when
357 * SYN ACK is ultimatively being received.
358 */
359 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
360}
361
Florian Westphal735d3832014-09-29 13:08:30 +0200362static void
Eric Dumazet6ac705b2015-09-25 07:39:18 -0700363tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700364{
Eric Dumazet6ac705b2015-09-25 07:39:18 -0700365 if (inet_rsk(req)->ecn_ok)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700366 th->ece = 1;
367}
368
Andi Kleen67edfef2009-07-21 23:00:40 +0000369/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
370 * be sent.
371 */
Florian Westphal735d3832014-09-29 13:08:30 +0200372static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
Eric Dumazetea1627c2016-05-13 09:16:40 -0700373 struct tcphdr *th, int tcp_header_len)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700374{
375 struct tcp_sock *tp = tcp_sk(sk);
376
377 if (tp->ecn_flags & TCP_ECN_OK) {
378 /* Not-retransmitted data segment: set ECT and inject CWR. */
379 if (skb->len != tcp_header_len &&
380 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
381 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800382 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700383 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
Eric Dumazetea1627c2016-05-13 09:16:40 -0700384 th->cwr = 1;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700385 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
386 }
Daniel Borkmann30e502a2014-09-26 22:37:33 +0200387 } else if (!tcp_ca_needs_ecn(sk)) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700388 /* ACK or retransmitted segment: clear ECT|CE */
389 INET_ECN_dontxmit(sk);
390 }
391 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
Eric Dumazetea1627c2016-05-13 09:16:40 -0700392 th->ece = 1;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700393 }
394}
395
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800396/* Constructs common control bits of non-data skb. If SYN/FIN is present,
397 * auto increment end seqno.
398 */
399static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
400{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700401 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800402 skb->csum = 0;
403
Eric Dumazet4de075e2011-09-27 13:25:05 -0400404 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800405 TCP_SKB_CB(skb)->sacked = 0;
406
Eric Dumazetcd7d8492014-09-24 04:11:22 -0700407 tcp_skb_pcount_set(skb, 1);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800408
409 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000410 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800411 seq++;
412 TCP_SKB_CB(skb)->end_seq = seq;
413}
414
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000415static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700416{
417 return tp->snd_una != tp->snd_up;
418}
419
Adam Langley33ad7982008-07-19 00:04:31 -0700420#define OPTION_SACK_ADVERTISE (1 << 0)
421#define OPTION_TS (1 << 1)
422#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000423#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000424#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700425
426struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000427 u16 options; /* bit field of OPTION_* */
428 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700429 u8 ws; /* window scale, 0 to disable */
430 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000431 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000432 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000433 __u32 tsval, tsecr; /* need to include OPTION_TS */
434 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700435};
436
Andi Kleen67edfef2009-07-21 23:00:40 +0000437/* Write previously computed TCP options to the packet.
438 *
439 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700440 * TCP options, we learned this through the hard way, so be careful here.
441 * Luckily we can at least blame others for their non-compliance but from
stephen hemminger8e3bff92013-12-08 12:15:44 -0800442 * inter-operability perspective it seems that we're somewhat stuck with
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700443 * the ordering which we have been using if we want to keep working with
444 * those broken things (not that it currently hurts anybody as there isn't
445 * particular reason why the ordering would need to be changed).
446 *
447 * At least SACK_PERM as the first option is known to lead to a disaster
448 * (but it may well be that other scenarios fail similarly).
449 */
Adam Langley33ad7982008-07-19 00:04:31 -0700450static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000451 struct tcp_out_options *opts)
452{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000453 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000454
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000455 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000456 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
457 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000458 /* overload cookie hash location */
459 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700460 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800461 }
Adam Langley33ad7982008-07-19 00:04:31 -0700462
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700463 if (unlikely(opts->mss)) {
464 *ptr++ = htonl((TCPOPT_MSS << 24) |
465 (TCPOLEN_MSS << 16) |
466 opts->mss);
467 }
468
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000469 if (likely(OPTION_TS & options)) {
470 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700471 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
472 (TCPOLEN_SACK_PERM << 16) |
473 (TCPOPT_TIMESTAMP << 8) |
474 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000475 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700476 } else {
477 *ptr++ = htonl((TCPOPT_NOP << 24) |
478 (TCPOPT_NOP << 16) |
479 (TCPOPT_TIMESTAMP << 8) |
480 TCPOLEN_TIMESTAMP);
481 }
482 *ptr++ = htonl(opts->tsval);
483 *ptr++ = htonl(opts->tsecr);
484 }
485
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000486 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700487 *ptr++ = htonl((TCPOPT_NOP << 24) |
488 (TCPOPT_NOP << 16) |
489 (TCPOPT_SACK_PERM << 8) |
490 TCPOLEN_SACK_PERM);
491 }
492
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000493 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700494 *ptr++ = htonl((TCPOPT_NOP << 24) |
495 (TCPOPT_WINDOW << 16) |
496 (TCPOLEN_WINDOW << 8) |
497 opts->ws);
498 }
499
500 if (unlikely(opts->num_sack_blocks)) {
501 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
502 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800503 int this_sack;
504
505 *ptr++ = htonl((TCPOPT_NOP << 24) |
506 (TCPOPT_NOP << 16) |
507 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700508 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800509 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800510
Adam Langley33ad7982008-07-19 00:04:31 -0700511 for (this_sack = 0; this_sack < opts->num_sack_blocks;
512 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800513 *ptr++ = htonl(sp[this_sack].start_seq);
514 *ptr++ = htonl(sp[this_sack].end_seq);
515 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800516
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000517 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800518 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000519
520 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
521 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
Daniel Lee7f9b8382015-04-06 14:37:26 -0700522 u8 *p = (u8 *)ptr;
523 u32 len; /* Fast Open option length */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000524
Daniel Lee7f9b8382015-04-06 14:37:26 -0700525 if (foc->exp) {
526 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
527 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
528 TCPOPT_FASTOPEN_MAGIC);
529 p += TCPOLEN_EXP_FASTOPEN_BASE;
530 } else {
531 len = TCPOLEN_FASTOPEN_BASE + foc->len;
532 *p++ = TCPOPT_FASTOPEN;
533 *p++ = len;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000534 }
Daniel Lee7f9b8382015-04-06 14:37:26 -0700535
536 memcpy(p, foc->val, foc->len);
537 if ((len & 3) == 2) {
538 p[foc->len] = TCPOPT_NOP;
539 p[foc->len + 1] = TCPOPT_NOP;
540 }
541 ptr += (len + 3) >> 2;
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000542 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800543}
544
Andi Kleen67edfef2009-07-21 23:00:40 +0000545/* Compute TCP options for SYN packets. This is not the final
546 * network wire format yet.
547 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000548static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700549 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400550 struct tcp_md5sig_key **md5)
551{
Adam Langley33ad7982008-07-19 00:04:31 -0700552 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000553 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000554 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700555
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800556#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700557 *md5 = tp->af_specific->md5_lookup(sk, sk);
558 if (*md5) {
559 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000560 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800561 }
Adam Langley33ad7982008-07-19 00:04:31 -0700562#else
563 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800564#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700565
566 /* We always get an MSS option. The option bytes which will be seen in
567 * normal data packets should timestamps be used, must be in the MSS
568 * advertised. But we subtract them from tp->mss_cache so that
569 * calculations in tcp_sendmsg are simpler etc. So account for this
570 * fact here if necessary. If we don't do this correctly, as a
571 * receiver we won't recognize data packets as being full sized when we
572 * should, and thus we won't abide by the delayed ACK rules correctly.
573 * SACKs don't matter, we never delay an ACK when we have any of those
574 * going out. */
575 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000576 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700577
Ian Morris51456b22015-04-03 09:17:26 +0100578 if (likely(sysctl_tcp_timestamps && !*md5)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700579 opts->options |= OPTION_TS;
Eric Dumazet7faee5c2014-09-05 15:33:33 -0700580 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700581 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000582 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700583 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800584 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700585 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000586 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000587 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700588 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800589 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700590 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700591 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000592 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700593 }
594
Yuchung Cheng783237e2012-07-19 06:43:07 +0000595 if (fastopen && fastopen->cookie.len >= 0) {
Daniel Lee2646c832015-04-06 14:37:27 -0700596 u32 need = fastopen->cookie.len;
597
598 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
599 TCPOLEN_FASTOPEN_BASE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000600 need = (need + 3) & ~3U; /* Align to 32 bits */
601 if (remaining >= need) {
602 opts->options |= OPTION_FAST_OPEN_COOKIE;
603 opts->fastopen_cookie = &fastopen->cookie;
604 remaining -= need;
605 tp->syn_fastopen = 1;
Daniel Lee2646c832015-04-06 14:37:27 -0700606 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000607 }
608 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000609
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000610 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700611}
612
Andi Kleen67edfef2009-07-21 23:00:40 +0000613/* Set up TCP options for SYN-ACKs. */
Eric Dumazet37bfbdd2015-09-25 07:39:17 -0700614static unsigned int tcp_synack_options(struct request_sock *req,
615 unsigned int mss, struct sk_buff *skb,
616 struct tcp_out_options *opts,
617 const struct tcp_md5sig_key *md5,
618 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 |
Eric Dumazetf9616c32016-09-20 22:45:58 -0700742 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
743 struct tcp_sock *tp = tcp_sk(sk);
744
745 if (tp->lost_out > tp->retrans_out &&
746 tp->snd_cwnd > tcp_packets_in_flight(tp))
747 tcp_xmit_retransmit_queue(sk);
748
749 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
John Ognessbf062002014-02-09 18:40:11 -0800750 0, GFP_ATOMIC);
Eric Dumazetf9616c32016-09-20 22:45:58 -0700751 }
Eric Dumazet6f458df2012-07-20 05:45:50 +0000752}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000753/*
stephen hemminger8e3bff92013-12-08 12:15:44 -0800754 * One tasklet per cpu tries to send more skbs.
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000755 * We run in tasklet context but need to disable irqs when
stephen hemminger8e3bff92013-12-08 12:15:44 -0800756 * transferring tsq->head because tcp_wfree() might
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000757 * interrupt us (non NAPI drivers)
758 */
759static void tcp_tasklet_func(unsigned long data)
760{
761 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
762 LIST_HEAD(list);
763 unsigned long flags;
764 struct list_head *q, *n;
765 struct tcp_sock *tp;
766 struct sock *sk;
767
768 local_irq_save(flags);
769 list_splice_init(&tsq->head, &list);
770 local_irq_restore(flags);
771
772 list_for_each_safe(q, n, &list) {
773 tp = list_entry(q, struct tcp_sock, tsq_node);
774 list_del(&tp->tsq_node);
775
776 sk = (struct sock *)tp;
777 bh_lock_sock(sk);
778
779 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000780 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000781 } else {
782 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000783 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000784 }
785 bh_unlock_sock(sk);
786
787 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
788 sk_free(sk);
789 }
790}
791
Eric Dumazet6f458df2012-07-20 05:45:50 +0000792#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
793 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200794 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
795 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000796/**
797 * tcp_release_cb - tcp release_sock() callback
798 * @sk: socket
799 *
800 * called from release_sock() to perform protocol dependent
801 * actions before socket release.
802 */
803void tcp_release_cb(struct sock *sk)
804{
805 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000806 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000807
Eric Dumazet6f458df2012-07-20 05:45:50 +0000808 /* perform an atomic operation only if at least one flag is set */
809 do {
810 flags = tp->tsq_flags;
811 if (!(flags & TCP_DEFERRED_ALL))
812 return;
813 nflags = flags & ~TCP_DEFERRED_ALL;
814 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
815
816 if (flags & (1UL << TCP_TSQ_DEFERRED))
817 tcp_tsq_handler(sk);
818
Eric Dumazetc3f9b012014-03-10 09:50:11 -0700819 /* Here begins the tricky part :
820 * We are called from release_sock() with :
821 * 1) BH disabled
822 * 2) sk_lock.slock spinlock held
823 * 3) socket owned by us (sk->sk_lock.owned == 1)
824 *
825 * But following code is meant to be called from BH handlers,
826 * so we should keep BH disabled, but early release socket ownership
827 */
828 sock_release_ownership(sk);
829
Eric Dumazet144d56e2012-08-20 00:22:46 +0000830 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000831 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000832 __sock_put(sk);
833 }
834 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000835 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000836 __sock_put(sk);
837 }
838 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Neal Cardwell4fab9072014-08-14 12:40:05 -0400839 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000840 __sock_put(sk);
841 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000842}
843EXPORT_SYMBOL(tcp_release_cb);
844
845void __init tcp_tasklet_init(void)
846{
847 int i;
848
849 for_each_possible_cpu(i) {
850 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
851
852 INIT_LIST_HEAD(&tsq->head);
853 tasklet_init(&tsq->tasklet,
854 tcp_tasklet_func,
855 (unsigned long)tsq);
856 }
857}
858
859/*
860 * Write buffer destructor automatically called from kfree_skb.
stephen hemminger8e3bff92013-12-08 12:15:44 -0800861 * We can't xmit new skbs from this context, as we might already
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000862 * hold qdisc lock.
863 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000864void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000865{
866 struct sock *sk = skb->sk;
867 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700868 int wmem;
869
870 /* Keep one reference on sk_wmem_alloc.
871 * Will be released by sk_free() from here or tcp_tasklet_func()
872 */
873 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
874
875 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
876 * Wait until our queues (qdisc + devices) are drained.
877 * This gives :
878 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
879 * - chance for incoming ACK (processed by another cpu maybe)
880 * to migrate this flow (skb->ooo_okay will be eventually set)
881 */
882 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
883 goto out;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000884
885 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
886 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
887 unsigned long flags;
888 struct tsq_tasklet *tsq;
889
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000890 /* queue this socket to tasklet queue */
891 local_irq_save(flags);
Christoph Lameter903ceff2014-08-17 12:30:35 -0500892 tsq = this_cpu_ptr(&tsq_tasklet);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000893 list_add(&tp->tsq_node, &tsq->head);
894 tasklet_schedule(&tsq->tasklet);
895 local_irq_restore(flags);
Eric Dumazet9b462d02014-10-13 06:27:47 -0700896 return;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000897 }
Eric Dumazet9b462d02014-10-13 06:27:47 -0700898out:
899 sk_free(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000900}
901
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902/* This routine actually transmits TCP packets queued in by
903 * tcp_do_sendmsg(). This is used by both the initial
904 * transmission and possible later retransmissions.
905 * All SKB's seen here are completely headerless. It is our
906 * job to build the TCP header, and pass the packet down to
907 * IP so it can do the same plus pass the packet off to the
908 * device.
909 *
910 * We are working here with either a clone of the original
911 * SKB, or a fresh unique copy made by the retransmit engine.
912 */
Yuchung Cheng1fcccc52018-07-18 13:56:34 -0700913static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
914 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800916 const struct inet_connection_sock *icsk = inet_csk(sk);
917 struct inet_sock *inet;
918 struct tcp_sock *tp;
919 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700920 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000921 unsigned int tcp_options_size, tcp_header_size;
Eric Dumazet186a9c52017-09-13 20:30:39 -0700922 struct sk_buff *oskb = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800923 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800924 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800925 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800927 BUG_ON(!skb || !tcp_skb_pcount(skb));
Lawrence Brakmo6f094b92016-06-08 21:16:44 -0700928 tp = tcp_sk(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800929
Eric Dumazetccdbb6e2013-10-10 08:43:00 -0700930 if (clone_it) {
Lawrence Brakmo6f094b92016-06-08 21:16:44 -0700931 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
932 - tp->snd_una;
Eric Dumazet186a9c52017-09-13 20:30:39 -0700933 oskb = skb;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800934 if (unlikely(skb_cloned(skb)))
935 skb = pskb_copy(skb, gfp_mask);
936 else
937 skb = skb_clone(skb, gfp_mask);
938 if (unlikely(!skb))
939 return -ENOBUFS;
940 }
Eric Dumazet186a9c52017-09-13 20:30:39 -0700941 skb_mstamp_get(&skb->skb_mstamp);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800942
943 inet = inet_sk(sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800944 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700945 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
Eric Dumazet4de075e2011-09-27 13:25:05 -0400947 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700948 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
949 else
950 tcp_options_size = tcp_established_options(sk, skb, &opts,
951 &md5);
952 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900953
Eric Dumazet547669d2013-05-23 07:44:20 +0000954 /* if no packet is in qdisc/device queue, then allow XPS to select
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700955 * another queue. We can be called from tcp_tsq_handler()
956 * which holds one reference to sk_wmem_alloc.
957 *
958 * TODO: Ideally, in-flight pure ACK packets should not matter here.
959 * One way to get this would be to set skb->truesize = 2 on them.
Eric Dumazet547669d2013-05-23 07:44:20 +0000960 */
Eric Dumazetb2532eb2014-10-10 18:06:35 -0700961 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800962
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700963 skb_push(skb, tcp_header_size);
964 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000965
966 skb_orphan(skb);
967 skb->sk = sk;
Eric Dumazet1d2077a2016-05-02 10:56:27 -0700968 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
Tom Herbertb73c3d02014-07-01 21:32:17 -0700969 skb_set_hash_from_sk(skb, sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000970 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800971
972 /* Build TCP header and checksum it. */
Eric Dumazetea1627c2016-05-13 09:16:40 -0700973 th = (struct tcphdr *)skb->data;
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000974 th->source = inet->inet_sport;
975 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800976 th->seq = htonl(tcb->seq);
Yuchung Cheng1fcccc52018-07-18 13:56:34 -0700977 th->ack_seq = htonl(rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700978 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400979 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800980
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800981 th->check = 0;
982 th->urg_ptr = 0;
983
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700984 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800985 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
986 if (before(tp->snd_up, tcb->seq + 0x10000)) {
987 th->urg_ptr = htons(tp->snd_up - tcb->seq);
988 th->urg = 1;
989 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700990 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800991 th->urg = 1;
992 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800993 }
994
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000995 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet51466a72015-06-11 09:15:16 -0700996 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Eric Dumazetea1627c2016-05-13 09:16:40 -0700997 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
998 th->window = htons(tcp_select_window(sk));
999 tcp_ecn_send(sk, skb, th, tcp_header_size);
1000 } else {
1001 /* RFC1323: The window in SYN & SYN/ACK segments
1002 * is never scaled.
1003 */
1004 th->window = htons(min(tp->rcv_wnd, 65535U));
1005 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001006#ifdef CONFIG_TCP_MD5SIG
1007 /* Calculate the MD5 hash, as we have all we need now */
1008 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -07001009 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00001010 tp->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07001011 md5, sk, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001012 }
1013#endif
1014
Herbert Xubb296242010-04-11 02:15:55 +00001015 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001016
Eric Dumazet4de075e2011-09-27 13:25:05 -04001017 if (likely(tcb->tcp_flags & TCPHDR_ACK))
Yuchung Cheng57ec8822018-07-18 13:56:35 -07001018 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001019
Martin KaFai Laua44d6ea2016-03-14 10:52:15 -07001020 if (skb->len != tcp_header_size) {
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001021 tcp_event_data_sent(tp, sk);
Martin KaFai Laua44d6ea2016-03-14 10:52:15 -07001022 tp->data_segs_out += tcp_skb_pcount(skb);
1023 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001024
Wei Yongjunbd37a082006-08-07 21:04:15 -07001025 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +00001026 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1027 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001028
Marcelo Ricardo Leitner2efd0552015-05-20 16:35:41 -07001029 tp->segs_out += tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001030 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001031 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001032 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001033
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001034 /* Our usage of tstamp should remain private */
1035 skb->tstamp.tv64 = 0;
Eric Dumazet971f10e2014-09-27 09:50:57 -07001036
1037 /* Cleanup our debris for IP stacks */
1038 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1039 sizeof(struct inet6_skb_parm)));
1040
Eric Dumazetb0270e92014-04-15 12:58:34 -04001041 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07001042
Eric Dumazet186a9c52017-09-13 20:30:39 -07001043 if (unlikely(err > 0)) {
1044 tcp_enter_cwr(sk);
1045 err = net_xmit_eval(err);
1046 }
Eric Dumazet85908cc2017-09-15 16:47:42 -07001047 if (!err && oskb) {
Eric Dumazet186a9c52017-09-13 20:30:39 -07001048 skb_mstamp_get(&oskb->skb_mstamp);
Eric Dumazet85908cc2017-09-15 16:47:42 -07001049 tcp_rate_skb_sent(sk, oskb);
1050 }
Eric Dumazet186a9c52017-09-13 20:30:39 -07001051 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052}
1053
Yuchung Cheng1fcccc52018-07-18 13:56:34 -07001054static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1055 gfp_t gfp_mask)
1056{
1057 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1058 tcp_sk(sk)->rcv_nxt);
1059}
1060
Andi Kleen67edfef2009-07-21 23:00:40 +00001061/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 *
1063 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1064 * otherwise socket can stall.
1065 */
1066static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1067{
1068 struct tcp_sock *tp = tcp_sk(sk);
1069
1070 /* Advance write_seq and place onto the write_queue. */
1071 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001072 __skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -08001073 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001074 sk->sk_wmem_queued += skb->truesize;
1075 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076}
1077
Andi Kleen67edfef2009-07-21 23:00:40 +00001078/* Initialize TSO segments for a packet. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001079static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -07001080{
Eric Dumazet8f26fb12013-10-15 12:24:54 -07001081 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -07001082 /* Avoid the costly divide in the normal
1083 * non-TSO case.
1084 */
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001085 tcp_skb_pcount_set(skb, 1);
Eric Dumazetf69ad292015-06-11 09:15:18 -07001086 TCP_SKB_CB(skb)->tcp_gso_size = 0;
David S. Millerf6302d12005-07-05 15:18:03 -07001087 } else {
Eric Dumazetcd7d8492014-09-24 04:11:22 -07001088 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
Eric Dumazetf69ad292015-06-11 09:15:18 -07001089 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 }
1091}
1092
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001093/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001094 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001095 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001096static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001097 int decr)
1098{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001099 struct tcp_sock *tp = tcp_sk(sk);
1100
Ilpo Järvinendc869672007-10-01 15:27:19 -07001101 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001102 return;
1103
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001104 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001105 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001106}
1107
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001108/* Pcount in the middle of the write queue got changed, we need to do various
1109 * tweaks to fix counters
1110 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001111static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001112{
1113 struct tcp_sock *tp = tcp_sk(sk);
1114
1115 tp->packets_out -= decr;
1116
1117 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1118 tp->sacked_out -= decr;
1119 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1120 tp->retrans_out -= decr;
1121 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1122 tp->lost_out -= decr;
1123
1124 /* Reno case is special. Sigh... */
1125 if (tcp_is_reno(tp) && decr > 0)
1126 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1127
1128 tcp_adjust_fackets_out(sk, skb, decr);
1129
1130 if (tp->lost_skb_hint &&
1131 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001132 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001133 tp->lost_cnt_hint -= decr;
1134
1135 tcp_verify_left_out(tp);
1136}
1137
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04001138static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1139{
1140 return TCP_SKB_CB(skb)->txstamp_ack ||
1141 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1142}
1143
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001144static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1145{
1146 struct skb_shared_info *shinfo = skb_shinfo(skb);
1147
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04001148 if (unlikely(tcp_has_tx_tstamp(skb)) &&
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001149 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1150 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1151 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1152
1153 shinfo->tx_flags &= ~tsflags;
1154 shinfo2->tx_flags |= tsflags;
1155 swap(shinfo->tskey, shinfo2->tskey);
Martin KaFai Laub51e13f2016-04-19 22:50:47 -07001156 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1157 TCP_SKB_CB(skb)->txstamp_ack = 0;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001158 }
1159}
1160
Martin KaFai Laua1661402016-04-25 14:44:50 -07001161static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1162{
1163 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1164 TCP_SKB_CB(skb)->eor = 0;
1165}
1166
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167/* Function to create two new TCP segments. Shrinks the given segment
1168 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001169 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170 * Remember, these are still headerless SKBs at this point.
1171 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001172int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001173 unsigned int mss_now, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174{
1175 struct tcp_sock *tp = tcp_sk(sk);
1176 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001177 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001178 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001179 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001181 if (WARN_ON(len > skb->len))
1182 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001183
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 nsize = skb_headlen(skb) - len;
1185 if (nsize < 0)
1186 nsize = 0;
1187
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001188 if (skb_unclone(skb, gfp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 return -ENOMEM;
1190
1191 /* Get a new skb... force flag on. */
Eric Dumazeteb934472015-05-19 13:26:55 -07001192 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001193 if (!buff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001195
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001196 sk->sk_wmem_queued += buff->truesize;
1197 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001198 nlen = skb->len - len - nsize;
1199 buff->truesize += nlen;
1200 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201
1202 /* Correct the sequence numbers. */
1203 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1204 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1205 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1206
1207 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001208 flags = TCP_SKB_CB(skb)->tcp_flags;
1209 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1210 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001211 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Martin KaFai Laua1661402016-04-25 14:44:50 -07001212 tcp_skb_fragment_eor(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213
Patrick McHardy84fa7932006-08-29 16:44:56 -07001214 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001216 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1217 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 nsize, 0);
1219
1220 skb_trim(skb, len);
1221
1222 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1223 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001224 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 skb_split(skb, buff, len);
1226 }
1227
1228 buff->ip_summed = skb->ip_summed;
1229
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001230 buff->tstamp = skb->tstamp;
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001231 tcp_fragment_tstamp(skb, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
David S. Miller6475be12005-09-01 22:47:01 -07001233 old_factor = tcp_skb_pcount(skb);
1234
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001236 tcp_set_skb_tso_segs(skb, mss_now);
1237 tcp_set_skb_tso_segs(buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238
Yuchung Chengb9f64822016-09-19 23:39:14 -04001239 /* Update delivered info for the new segment */
1240 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1241
David S. Miller6475be12005-09-01 22:47:01 -07001242 /* If this packet has been sent out already, we must
1243 * adjust the various packet counters.
1244 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001245 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001246 int diff = old_factor - tcp_skb_pcount(skb) -
1247 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001249 if (diff)
1250 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 }
1252
1253 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001254 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001255 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
1257 return 0;
1258}
1259
1260/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1261 * eventually). The difference is that pulled data not copied, but
1262 * immediately discarded.
1263 */
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001264static int __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265{
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001266 struct skb_shared_info *shinfo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 int i, k, eat;
1268
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001269 eat = min_t(int, len, skb_headlen(skb));
1270 if (eat) {
1271 __skb_pull(skb, eat);
1272 len -= eat;
1273 if (!len)
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001274 return 0;
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001275 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 eat = len;
1277 k = 0;
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001278 shinfo = skb_shinfo(skb);
1279 for (i = 0; i < shinfo->nr_frags; i++) {
1280 int size = skb_frag_size(&shinfo->frags[i]);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001281
1282 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001283 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001284 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 } else {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001286 shinfo->frags[k] = shinfo->frags[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 if (eat) {
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001288 shinfo->frags[k].page_offset += eat;
1289 skb_frag_size_sub(&shinfo->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 eat = 0;
1291 }
1292 k++;
1293 }
1294 }
Eric Dumazet7b7fc972013-12-05 22:31:30 -08001295 shinfo->nr_frags = k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001297 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 skb->data_len -= len;
1299 skb->len = skb->data_len;
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001300 return len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301}
1302
Andi Kleen67edfef2009-07-21 23:00:40 +00001303/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1305{
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001306 u32 delta_truesize;
1307
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001308 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 return -ENOMEM;
1310
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001311 delta_truesize = __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312
1313 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001314 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315
Eric Dumazetf3235cb2017-04-26 17:15:40 -07001316 if (delta_truesize) {
1317 skb->truesize -= delta_truesize;
1318 sk->sk_wmem_queued -= delta_truesize;
1319 sk_mem_uncharge(sk, delta_truesize);
1320 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1321 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001323 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 if (tcp_skb_pcount(skb) > 1)
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001325 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
1327 return 0;
1328}
1329
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001330/* Calculate MSS not accounting any TCP options. */
1331static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001332{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001333 const struct tcp_sock *tp = tcp_sk(sk);
1334 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001335 int mss_now;
1336
1337 /* Calculate base mss without TCP options:
1338 It is MMS_S - sizeof(tcphdr) of rfc1122
1339 */
1340 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1341
Eric Dumazet67469602012-04-24 07:37:38 +00001342 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1343 if (icsk->icsk_af_ops->net_frag_header_len) {
1344 const struct dst_entry *dst = __sk_dst_get(sk);
1345
1346 if (dst && dst_allfrag(dst))
1347 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1348 }
1349
John Heffner5d424d52006-03-20 17:53:41 -08001350 /* Clamp it (mss_clamp does not include tcp options) */
1351 if (mss_now > tp->rx_opt.mss_clamp)
1352 mss_now = tp->rx_opt.mss_clamp;
1353
1354 /* Now subtract optional transport overhead */
1355 mss_now -= icsk->icsk_ext_hdr_len;
1356
1357 /* Then reserve room for full set of TCP options and 8 bytes of data */
1358 if (mss_now < 48)
1359 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001360 return mss_now;
1361}
1362
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001363/* Calculate MSS. Not accounting for SACKs here. */
1364int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1365{
1366 /* Subtract TCP options size, not including SACKs */
1367 return __tcp_mtu_to_mss(sk, pmtu) -
1368 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1369}
1370
John Heffner5d424d52006-03-20 17:53:41 -08001371/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001372int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001373{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001374 const struct tcp_sock *tp = tcp_sk(sk);
1375 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001376 int mtu;
1377
1378 mtu = mss +
1379 tp->tcp_header_len +
1380 icsk->icsk_ext_hdr_len +
1381 icsk->icsk_af_ops->net_header_len;
1382
Eric Dumazet67469602012-04-24 07:37:38 +00001383 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1384 if (icsk->icsk_af_ops->net_frag_header_len) {
1385 const struct dst_entry *dst = __sk_dst_get(sk);
1386
1387 if (dst && dst_allfrag(dst))
1388 mtu += icsk->icsk_af_ops->net_frag_header_len;
1389 }
John Heffner5d424d52006-03-20 17:53:41 -08001390 return mtu;
1391}
Neal Cardwell556c6b42016-09-19 23:39:19 -04001392EXPORT_SYMBOL(tcp_mss_to_mtu);
John Heffner5d424d52006-03-20 17:53:41 -08001393
Andi Kleen67edfef2009-07-21 23:00:40 +00001394/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001395void tcp_mtup_init(struct sock *sk)
1396{
1397 struct tcp_sock *tp = tcp_sk(sk);
1398 struct inet_connection_sock *icsk = inet_csk(sk);
Fan Dub0f9ca52015-02-10 09:53:16 +08001399 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001400
Fan Dub0f9ca52015-02-10 09:53:16 +08001401 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
John Heffner5d424d52006-03-20 17:53:41 -08001402 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001403 icsk->icsk_af_ops->net_header_len;
Fan Dub0f9ca52015-02-10 09:53:16 +08001404 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
John Heffner5d424d52006-03-20 17:53:41 -08001405 icsk->icsk_mtup.probe_size = 0;
Fan Du05cbc0d2015-03-06 11:18:24 +08001406 if (icsk->icsk_mtup.enabled)
1407 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
John Heffner5d424d52006-03-20 17:53:41 -08001408}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001409EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001410
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411/* This function synchronize snd mss to current pmtu/exthdr set.
1412
1413 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1414 for TCP options, but includes only bare TCP header.
1415
1416 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001417 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 It also does not include TCP options.
1419
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001420 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421
1422 tp->mss_cache is current effective sending mss, including
1423 all tcp options except for SACKs. It is evaluated,
1424 taking into account current pmtu, but never exceeds
1425 tp->rx_opt.mss_clamp.
1426
1427 NOTE1. rfc1122 clearly states that advertised MSS
1428 DOES NOT include either tcp or ip options.
1429
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001430 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1431 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1434{
1435 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001436 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001437 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438
John Heffner5d424d52006-03-20 17:53:41 -08001439 if (icsk->icsk_mtup.search_high > pmtu)
1440 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441
John Heffner5d424d52006-03-20 17:53:41 -08001442 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001443 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444
1445 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001446 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001447 if (icsk->icsk_mtup.enabled)
1448 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001449 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450
1451 return mss_now;
1452}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001453EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454
1455/* Compute the current effective MSS, taking SACKs and IP options,
1456 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001458unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001460 const struct tcp_sock *tp = tcp_sk(sk);
1461 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001462 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001463 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001464 struct tcp_out_options opts;
1465 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001467 mss_now = tp->mss_cache;
1468
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 if (dst) {
1470 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001471 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 mss_now = tcp_sync_mss(sk, mtu);
1473 }
1474
Adam Langley33ad7982008-07-19 00:04:31 -07001475 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1476 sizeof(struct tcphdr);
1477 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1478 * some common options. If this is an odd packet (because we have SACK
1479 * blocks etc) then our calculated header_len will be different, and
1480 * we have to adjust mss_now correspondingly */
1481 if (header_len != tp->tcp_header_len) {
1482 int delta = (int) header_len - tp->tcp_header_len;
1483 mss_now -= delta;
1484 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001485
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 return mss_now;
1487}
1488
Weiping Pan86fd14a2014-04-18 12:27:46 +08001489/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1490 * As additional protections, we do not touch cwnd in retransmission phases,
1491 * and if application hit its sndbuf limit recently.
1492 */
1493static void tcp_cwnd_application_limited(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001494{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001495 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001496
Weiping Pan86fd14a2014-04-18 12:27:46 +08001497 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1498 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1499 /* Limited by application or receiver window. */
1500 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1501 u32 win_used = max(tp->snd_cwnd_used, init_win);
1502 if (win_used < tp->snd_cwnd) {
1503 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1504 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1505 }
1506 tp->snd_cwnd_used = 0;
1507 }
1508 tp->snd_cwnd_stamp = tcp_time_stamp;
1509}
1510
Neal Cardwellca8a2262014-05-22 10:41:08 -04001511static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
David S. Millera762a982005-07-05 15:18:51 -07001512{
1513 struct tcp_sock *tp = tcp_sk(sk);
1514
Neal Cardwellca8a2262014-05-22 10:41:08 -04001515 /* Track the maximum number of outstanding packets in each
1516 * window, and remember whether we were cwnd-limited then.
1517 */
1518 if (!before(tp->snd_una, tp->max_packets_seq) ||
1519 tp->packets_out > tp->max_packets_out) {
1520 tp->max_packets_out = tp->packets_out;
1521 tp->max_packets_seq = tp->snd_nxt;
1522 tp->is_cwnd_limited = is_cwnd_limited;
1523 }
Eric Dumazete114a712014-04-30 11:58:13 -07001524
Eric Dumazet24901552014-05-02 21:18:05 -07001525 if (tcp_is_cwnd_limited(sk)) {
David S. Millera762a982005-07-05 15:18:51 -07001526 /* Network is feed fully. */
1527 tp->snd_cwnd_used = 0;
1528 tp->snd_cwnd_stamp = tcp_time_stamp;
1529 } else {
1530 /* Network starves. */
1531 if (tp->packets_out > tp->snd_cwnd_used)
1532 tp->snd_cwnd_used = tp->packets_out;
1533
David S. Miller15d33c02007-04-09 13:23:14 -07001534 if (sysctl_tcp_slow_start_after_idle &&
1535 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001536 tcp_cwnd_application_limited(sk);
1537 }
1538}
1539
Eric Dumazetd4589922013-12-13 13:51:23 -08001540/* Minshall's variant of the Nagle send check. */
1541static bool tcp_minshall_check(const struct tcp_sock *tp)
1542{
1543 return after(tp->snd_sml, tp->snd_una) &&
1544 !after(tp->snd_sml, tp->snd_nxt);
1545}
1546
1547/* Update snd_sml if this skb is under mss
1548 * Note that a TSO packet might end with a sub-mss segment
1549 * The test is really :
1550 * if ((skb->len % mss) != 0)
1551 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1552 * But we can avoid doing the divide again given we already have
1553 * skb_pcount = skb->len / mss_now
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001554 */
Eric Dumazetd4589922013-12-13 13:51:23 -08001555static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1556 const struct sk_buff *skb)
1557{
1558 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1559 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1560}
1561
1562/* Return false, if packet can be sent now without violation Nagle's rules:
1563 * 1. It is full sized. (provided by caller in %partial bool)
1564 * 2. Or it contains FIN. (already checked by caller)
1565 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1566 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1567 * With Minshall's modification: all sent small packets are ACKed.
1568 */
1569static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001570 int nonagle)
Eric Dumazetd4589922013-12-13 13:51:23 -08001571{
1572 return partial &&
1573 ((nonagle & TCP_NAGLE_CORK) ||
1574 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1575}
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001576
1577/* Return how many segs we'd like on a TSO packet,
1578 * to send one TSO packet per ms
1579 */
Neal Cardwell1b3878c2016-09-19 23:39:18 -04001580u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1581 int min_tso_segs)
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001582{
1583 u32 bytes, segs;
1584
1585 bytes = min(sk->sk_pacing_rate >> 10,
1586 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1587
1588 /* Goal is to send at least one packet per ms,
1589 * not one big TSO packet every 100 ms.
1590 * This preserves ACK clocking and is consistent
1591 * with tcp_tso_should_defer() heuristic.
1592 */
Neal Cardwell1b3878c2016-09-19 23:39:18 -04001593 segs = max_t(u32, bytes / mss_now, min_tso_segs);
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001594
Eric Dumazet45fa6612018-02-21 06:43:03 -08001595 return segs;
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001596}
Neal Cardwell1b3878c2016-09-19 23:39:18 -04001597EXPORT_SYMBOL(tcp_tso_autosize);
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001598
Neal Cardwelled6e7262016-09-19 23:39:17 -04001599/* Return the number of segments we want in the skb we are transmitting.
1600 * See if congestion control module wants to decide; otherwise, autosize.
1601 */
1602static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1603{
1604 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1605 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1606
Eric Dumazet45fa6612018-02-21 06:43:03 -08001607 if (!tso_segs)
1608 tso_segs = tcp_tso_autosize(sk, mss_now,
1609 sysctl_tcp_min_tso_segs);
1610 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
Neal Cardwelled6e7262016-09-19 23:39:17 -04001611}
1612
Eric Dumazetd4589922013-12-13 13:51:23 -08001613/* Returns the portion of skb which can be sent right away */
1614static unsigned int tcp_mss_split_point(const struct sock *sk,
1615 const struct sk_buff *skb,
1616 unsigned int mss_now,
1617 unsigned int max_segs,
1618 int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001619{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001620 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazetd4589922013-12-13 13:51:23 -08001621 u32 partial, needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001622
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001623 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001624 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001625
Ben Hutchings14853482012-07-30 16:11:42 +00001626 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1627 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001628
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001629 needed = min(skb->len, window);
1630
Ben Hutchings14853482012-07-30 16:11:42 +00001631 if (max_len <= needed)
1632 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001633
Eric Dumazetd4589922013-12-13 13:51:23 -08001634 partial = needed % mss_now;
1635 /* If last segment is not a full MSS, check if Nagle rules allow us
1636 * to include this last segment in this skb.
1637 * Otherwise, we'll split the skb at last MSS boundary
1638 */
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001639 if (tcp_nagle_check(partial != 0, tp, nonagle))
Eric Dumazetd4589922013-12-13 13:51:23 -08001640 return needed - partial;
1641
1642 return needed;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001643}
1644
1645/* Can at least one segment of SKB be sent right now, according to the
1646 * congestion window rules? If so, return how many segments are allowed.
1647 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001648static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1649 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001650{
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001651 u32 in_flight, cwnd, halfcwnd;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001652
1653 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001654 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1655 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001656 return 1;
1657
1658 in_flight = tcp_packets_in_flight(tp);
1659 cwnd = tp->snd_cwnd;
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001660 if (in_flight >= cwnd)
1661 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001662
Eric Dumazetd649a7a2014-11-13 09:45:22 -08001663 /* For better scheduling, ensure we have at least
1664 * 2 GSO packets in flight.
1665 */
1666 halfcwnd = max(cwnd >> 1, 1U);
1667 return min(halfcwnd, cwnd - in_flight);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001668}
1669
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001670/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001671 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001672 * SKB onto the wire.
1673 */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001674static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001675{
1676 int tso_segs = tcp_skb_pcount(skb);
1677
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001678 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001679 tcp_set_skb_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001680 tso_segs = tcp_skb_pcount(skb);
1681 }
1682 return tso_segs;
1683}
1684
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001685
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001686/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001687 * sent now.
1688 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001689static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1690 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001691{
1692 /* Nagle rule does not apply to frames, which sit in the middle of the
1693 * write_queue (they have no chances to get new data).
1694 *
1695 * This is implemented in the callers, where they modify the 'nonagle'
1696 * argument based upon the location of SKB in the send queue.
1697 */
1698 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001699 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001700
Yuchung Cheng9b441902013-03-20 13:32:58 +00001701 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1702 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001703 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001704
Peter Pan(潘卫平)cc93fc52014-03-24 14:49:34 +08001705 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001706 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001707
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001708 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001709}
1710
1711/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001712static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1713 const struct sk_buff *skb,
1714 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001715{
1716 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1717
1718 if (skb->len > cur_mss)
1719 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1720
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001721 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001722}
1723
David S. Millerfe067e82007-03-07 12:12:44 -08001724/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001725 * should be put on the wire right now. If so, it returns the number of
1726 * packets allowed by the congestion window.
1727 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001728static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001729 unsigned int cur_mss, int nonagle)
1730{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001731 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001732 unsigned int cwnd_quota;
1733
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001734 tcp_init_tso_segs(skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001735
1736 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1737 return 0;
1738
1739 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001740 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001741 cwnd_quota = 0;
1742
1743 return cwnd_quota;
1744}
1745
Andi Kleen67edfef2009-07-21 23:00:40 +00001746/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001747bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001748{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001749 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001750 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001751
Eric Dumazeta02cec22010-09-22 20:43:57 +00001752 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001753 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001754 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001755 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001756}
1757
1758/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1759 * which is put after SKB on the list. It is very much like
1760 * tcp_fragment() except that it may make several kinds of assumptions
1761 * in order to speed up the splitting operation. In particular, we
1762 * know that all the data is in scatter-gather pages, and that the
1763 * packet has never been sent out before (and thus is not cloned).
1764 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001765static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001766 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001767{
1768 struct sk_buff *buff;
1769 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001770 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001771
1772 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001773 if (skb->len != skb->data_len)
Octavian Purdila6cc55e02014-06-06 17:32:37 +03001774 return tcp_fragment(sk, skb, len, mss_now, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001775
Eric Dumazeteb934472015-05-19 13:26:55 -07001776 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
Ian Morris51456b22015-04-03 09:17:26 +01001777 if (unlikely(!buff))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001778 return -ENOMEM;
1779
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001780 sk->sk_wmem_queued += buff->truesize;
1781 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001782 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001783 skb->truesize -= nlen;
1784
1785 /* Correct the sequence numbers. */
1786 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1787 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1788 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1789
1790 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001791 flags = TCP_SKB_CB(skb)->tcp_flags;
1792 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1793 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001794
1795 /* This packet was never sent out yet, so no SACK bits. */
1796 TCP_SKB_CB(buff)->sacked = 0;
1797
Martin KaFai Laua1661402016-04-25 14:44:50 -07001798 tcp_skb_fragment_eor(skb, buff);
1799
Patrick McHardy84fa7932006-08-29 16:44:56 -07001800 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001801 skb_split(skb, buff, len);
Willem de Bruijn490cc7d2014-08-12 15:08:12 -04001802 tcp_fragment_tstamp(skb, buff);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001803
1804 /* Fix up tso_factor for both original and new SKB. */
Eric Dumazet5bbb4322015-06-11 09:15:17 -07001805 tcp_set_skb_tso_segs(skb, mss_now);
1806 tcp_set_skb_tso_segs(buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001807
1808 /* Link BUFF into the send queue. */
Eric Dumazetf4a775d2014-09-22 16:29:32 -07001809 __skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001810 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001811
1812 return 0;
1813}
1814
1815/* Try to defer sending, if possible, in order to minimize the amount
1816 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1817 *
1818 * This algorithm is from John Heffner.
1819 */
Neal Cardwellca8a2262014-05-22 10:41:08 -04001820static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001821 bool *is_cwnd_limited, u32 max_segs)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001822{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001823 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet50c83392015-02-26 14:10:19 -08001824 u32 age, send_win, cong_win, limit, in_flight;
1825 struct tcp_sock *tp = tcp_sk(sk);
1826 struct skb_mstamp now;
1827 struct sk_buff *head;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001828 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001829
Eric Dumazet4de075e2011-09-27 13:25:05 -04001830 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001831 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001832
Eric Dumazet99d76622015-07-26 09:45:24 +02001833 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
John Heffnerae8064a2006-10-18 20:36:48 -07001834 goto send_now;
1835
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001836 /* Avoid bursty behavior by allowing defer
1837 * only if the last write was recent.
1838 */
1839 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
John Heffnerae8064a2006-10-18 20:36:48 -07001840 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001841
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001842 in_flight = tcp_packets_in_flight(tp);
1843
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001844 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001845
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001846 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001847
1848 /* From in_flight test above, we know that cwnd > in_flight. */
1849 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1850
1851 limit = min(send_win, cong_win);
1852
David S. Millerba244fe2006-03-11 18:51:49 -08001853 /* If a full-sized TSO skb can be sent, do it. */
Eric Dumazet605ad7f2014-12-07 12:22:18 -08001854 if (limit >= max_segs * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001855 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001856
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001857 /* Middle in queue won't get any more data, full sendable already? */
1858 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1859 goto send_now;
1860
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001861 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1862 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001863 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1864
1865 /* If at least some fraction of a window is available,
1866 * just use it.
1867 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001868 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001869 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001870 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001871 } else {
1872 /* Different approach, try not to defer past a single
1873 * ACK. Receiver should ACK every other full sized
1874 * frame, so if we have space for more than 3 frames
1875 * then send now.
1876 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001877 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001878 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001879 }
1880
Eric Dumazet50c83392015-02-26 14:10:19 -08001881 head = tcp_write_queue_head(sk);
1882 skb_mstamp_get(&now);
1883 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1884 /* If next ACK is likely to come too late (half srtt), do not defer */
1885 if (age < (tp->srtt_us >> 4))
1886 goto send_now;
1887
Eric Dumazet5f852eb2015-02-26 14:10:18 -08001888 /* Ok, it looks like it is advisable to defer. */
John Heffnerae8064a2006-10-18 20:36:48 -07001889
Bendik Rønning Opstadd2e13392015-09-23 18:49:53 +02001890 if (cong_win < send_win && cong_win <= skb->len)
Neal Cardwellca8a2262014-05-22 10:41:08 -04001891 *is_cwnd_limited = true;
1892
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001893 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001894
1895send_now:
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001896 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001897}
1898
Fan Du05cbc0d2015-03-06 11:18:24 +08001899static inline void tcp_mtu_check_reprobe(struct sock *sk)
1900{
1901 struct inet_connection_sock *icsk = inet_csk(sk);
1902 struct tcp_sock *tp = tcp_sk(sk);
1903 struct net *net = sock_net(sk);
1904 u32 interval;
1905 s32 delta;
1906
1907 interval = net->ipv4.sysctl_tcp_probe_interval;
1908 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1909 if (unlikely(delta >= interval * HZ)) {
1910 int mss = tcp_current_mss(sk);
1911
1912 /* Update current search range */
1913 icsk->icsk_mtup.probe_size = 0;
1914 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1915 sizeof(struct tcphdr) +
1916 icsk->icsk_af_ops->net_header_len;
1917 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1918
1919 /* Update probe time stamp */
1920 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1921 }
1922}
1923
Ilya Lesokhin3efb90a2018-02-12 12:57:04 +02001924static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
1925{
1926 struct sk_buff *skb, *next;
1927
1928 skb = tcp_send_head(sk);
1929 tcp_for_write_queue_from_safe(skb, next, sk) {
1930 if (len <= skb->len)
1931 break;
1932
1933 if (unlikely(TCP_SKB_CB(skb)->eor))
1934 return false;
1935
1936 len -= skb->len;
1937 }
1938
1939 return true;
1940}
1941
John Heffner5d424d52006-03-20 17:53:41 -08001942/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001943 * MTU probe is regularly attempting to increase the path MTU by
1944 * deliberately sending larger packets. This discovers routing
1945 * changes resulting in larger path MTUs.
1946 *
John Heffner5d424d52006-03-20 17:53:41 -08001947 * Returns 0 if we should wait to probe (no cwnd available),
1948 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001949 * -1 otherwise
1950 */
John Heffner5d424d52006-03-20 17:53:41 -08001951static int tcp_mtu_probe(struct sock *sk)
1952{
1953 struct tcp_sock *tp = tcp_sk(sk);
1954 struct inet_connection_sock *icsk = inet_csk(sk);
1955 struct sk_buff *skb, *nskb, *next;
Fan Du6b58e0a2015-03-06 11:18:23 +08001956 struct net *net = sock_net(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001957 int len;
1958 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001959 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001960 int copy;
1961 int mss_now;
Fan Du6b58e0a2015-03-06 11:18:23 +08001962 int interval;
John Heffner5d424d52006-03-20 17:53:41 -08001963
1964 /* Not currently probing/verifying,
1965 * not in recovery,
1966 * have enough cwnd, and
1967 * not SACKing (the variable headers throw things off) */
1968 if (!icsk->icsk_mtup.enabled ||
1969 icsk->icsk_mtup.probe_size ||
1970 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1971 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001972 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001973 return -1;
1974
Fan Du6b58e0a2015-03-06 11:18:23 +08001975 /* Use binary search for probe_size between tcp_mss_base,
1976 * and current mss_clamp. if (search_high - search_low)
1977 * smaller than a threshold, backoff from probing.
1978 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001979 mss_now = tcp_current_mss(sk);
Fan Du6b58e0a2015-03-06 11:18:23 +08001980 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1981 icsk->icsk_mtup.search_low) >> 1);
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001982 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
Fan Du6b58e0a2015-03-06 11:18:23 +08001983 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
Fan Du05cbc0d2015-03-06 11:18:24 +08001984 /* When misfortune happens, we are reprobing actively,
1985 * and then reprobe timer has expired. We stick with current
1986 * probing process by not resetting search range to its orignal.
1987 */
Fan Du6b58e0a2015-03-06 11:18:23 +08001988 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
Fan Du05cbc0d2015-03-06 11:18:24 +08001989 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1990 /* Check whether enough time has elaplased for
1991 * another round of probing.
1992 */
1993 tcp_mtu_check_reprobe(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001994 return -1;
1995 }
1996
1997 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001998 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001999 return -1;
2000
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08002001 if (tp->snd_wnd < size_needed)
2002 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08002003 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08002004 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08002005
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02002006 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2007 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2008 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08002009 return -1;
2010 else
2011 return 0;
2012 }
2013
Ilya Lesokhin3efb90a2018-02-12 12:57:04 +02002014 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2015 return -1;
2016
John Heffner5d424d52006-03-20 17:53:41 -08002017 /* We're allowed to probe. Build it now. */
Eric Dumazeteb934472015-05-19 13:26:55 -07002018 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
Ian Morris51456b22015-04-03 09:17:26 +01002019 if (!nskb)
John Heffner5d424d52006-03-20 17:53:41 -08002020 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002021 sk->sk_wmem_queued += nskb->truesize;
2022 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08002023
David S. Millerfe067e82007-03-07 12:12:44 -08002024 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002025
2026 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2027 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002028 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08002029 TCP_SKB_CB(nskb)->sacked = 0;
2030 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07002031 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08002032
Ilpo Järvinen50c48172007-12-02 00:48:00 +02002033 tcp_insert_write_queue_before(nskb, skb, sk);
Eric Dumazete12c42c2017-10-30 23:08:20 -07002034 tcp_highest_sack_replace(sk, skb, nskb);
Ilpo Järvinen50c48172007-12-02 00:48:00 +02002035
John Heffner5d424d52006-03-20 17:53:41 -08002036 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02002037 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08002038 copy = min_t(int, skb->len, probe_size - len);
Douglas Caetano dos Santos2fe664f2016-09-22 15:52:04 -03002039 if (nskb->ip_summed) {
John Heffner5d424d52006-03-20 17:53:41 -08002040 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
Douglas Caetano dos Santos2fe664f2016-09-22 15:52:04 -03002041 } else {
2042 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2043 skb_put(nskb, copy),
2044 copy, 0);
2045 nskb->csum = csum_block_add(nskb->csum, csum, len);
2046 }
John Heffner5d424d52006-03-20 17:53:41 -08002047
2048 if (skb->len <= copy) {
2049 /* We've eaten all the data from this skb.
2050 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002051 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
Ilya Lesokhin3efb90a2018-02-12 12:57:04 +02002052 /* If this is the last SKB we copy and eor is set
2053 * we need to propagate it to the new skb.
2054 */
2055 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
David S. Millerfe067e82007-03-07 12:12:44 -08002056 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002057 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08002058 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04002059 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00002060 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08002061 if (!skb_shinfo(skb)->nr_frags) {
2062 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07002063 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002064 skb->csum = csum_partial(skb->data,
2065 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08002066 } else {
2067 __pskb_trim_head(skb, copy);
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002068 tcp_set_skb_tso_segs(skb, mss_now);
John Heffner5d424d52006-03-20 17:53:41 -08002069 }
2070 TCP_SKB_CB(skb)->seq += copy;
2071 }
2072
2073 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02002074
2075 if (len >= probe_size)
2076 break;
John Heffner5d424d52006-03-20 17:53:41 -08002077 }
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002078 tcp_init_tso_segs(nskb, nskb->len);
John Heffner5d424d52006-03-20 17:53:41 -08002079
2080 /* We're ready to send. If this fails, the probe will
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002081 * be resegmented into mss-sized pieces by tcp_write_xmit().
2082 */
John Heffner5d424d52006-03-20 17:53:41 -08002083 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2084 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002085 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08002086 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002087 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08002088
2089 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08002090 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2091 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08002092
2093 return 1;
2094 }
2095
2096 return -1;
2097}
2098
Eric Dumazetf9616c32016-09-20 22:45:58 -07002099/* TCP Small Queues :
2100 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2101 * (These limits are doubled for retransmits)
2102 * This allows for :
2103 * - better RTT estimation and ACK scheduling
2104 * - faster recovery
2105 * - high rates
2106 * Alas, some drivers / subsystems require a fair amount
2107 * of queued bytes to ensure line rate.
2108 * One example is wifi aggregation (802.11 AMPDU)
2109 */
2110static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2111 unsigned int factor)
2112{
2113 unsigned int limit;
2114
2115 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2116 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2117 limit <<= factor;
2118
2119 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2120 set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
2121 /* It is possible TX completion already happened
2122 * before we set TSQ_THROTTLED, so we must
2123 * test again the condition.
2124 */
2125 smp_mb__after_atomic();
2126 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2127 return true;
2128 }
2129 return false;
2130}
2131
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132/* This routine writes packets to the network. It advances the
2133 * send_head. This happens as incoming acks open up the remote
2134 * window for us.
2135 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08002136 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2137 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2138 * account rare use of URG, this is not a big flaw.
2139 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002140 * Send at most one packet when push_one > 0. Temporarily ignore
2141 * cwnd limit to force at most one packet out when push_one == 2.
2142
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002143 * Returns true, if no segments are in flight and we have queued segments,
2144 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002146static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2147 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148{
2149 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07002150 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002151 unsigned int tso_segs, sent_pkts;
2152 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08002153 int result;
Neal Cardwellca8a2262014-05-22 10:41:08 -04002154 bool is_cwnd_limited = false;
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002155 u32 max_segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
David S. Miller92df7b52005-07-05 15:19:06 -07002157 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002158
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002159 if (!push_one) {
2160 /* Do MTU probing. */
2161 result = tcp_mtu_probe(sk);
2162 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002163 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002164 } else if (result > 0) {
2165 sent_pkts = 1;
2166 }
John Heffner5d424d52006-03-20 17:53:41 -08002167 }
2168
Neal Cardwelled6e7262016-09-19 23:39:17 -04002169 max_segs = tcp_tso_segs(sk, mss_now);
David S. Millerfe067e82007-03-07 12:12:44 -08002170 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07002171 unsigned int limit;
2172
Eric Dumazet5bbb4322015-06-11 09:15:17 -07002173 tso_segs = tcp_init_tso_segs(skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002174 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07002175
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002176 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002177 /* "skb_mstamp" is used as a start point for the retransmit timer */
2178 skb_mstamp_get(&skb->skb_mstamp);
Andrew Vaginec342322012-11-15 04:03:17 +00002179 goto repair; /* Skip network transmission */
Andrey Vagin9d186ca2014-08-13 16:03:10 +04002180 }
Andrew Vaginec342322012-11-15 04:03:17 +00002181
Herbert Xub68e9f82005-08-04 19:52:02 -07002182 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002183 if (!cwnd_quota) {
2184 if (push_one == 2)
2185 /* Force out a loss probe pkt. */
2186 cwnd_quota = 1;
2187 else
2188 break;
2189 }
Herbert Xub68e9f82005-08-04 19:52:02 -07002190
2191 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2192 break;
2193
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002194 if (tso_segs == 1) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002195 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2196 (tcp_skb_is_last(sk, skb) ?
2197 nonagle : TCP_NAGLE_PUSH))))
2198 break;
2199 } else {
Neal Cardwellca8a2262014-05-22 10:41:08 -04002200 if (!push_one &&
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002201 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2202 max_segs))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002203 break;
2204 }
David S. Milleraa934662005-07-05 15:20:09 -07002205
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002206 limit = mss_now;
Eric Dumazetd6a4e262015-05-26 08:55:28 -07002207 if (tso_segs > 1 && !tcp_urg_mode(tp))
Eric Dumazet605ad7f2014-12-07 12:22:18 -08002208 limit = tcp_mss_split_point(sk, skb, mss_now,
2209 min_t(unsigned int,
2210 cwnd_quota,
2211 max_segs),
2212 nonagle);
2213
2214 if (skb->len > limit &&
2215 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2216 break;
2217
Eric Dumazetf9616c32016-09-20 22:45:58 -07002218 if (tcp_small_queue_check(sk, skb, 0))
2219 break;
Eric Dumazetc9eeec22013-09-27 03:28:54 -07002220
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002221 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07002222 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
Andrew Vaginec342322012-11-15 04:03:17 +00002224repair:
David S. Miller92df7b52005-07-05 15:19:06 -07002225 /* Advance the send_head. This one is sent out.
2226 * This call will increment packets_out.
2227 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08002228 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07002229
2230 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002231 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002232
2233 if (push_one)
2234 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
David S. Miller92df7b52005-07-05 15:19:06 -07002236
David S. Milleraa934662005-07-05 15:20:09 -07002237 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00002238 if (tcp_in_cwnd_reduction(sk))
2239 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002240
2241 /* Send one loss probe per tail loss episode. */
2242 if (push_one != 2)
2243 tcp_schedule_loss_probe(sk);
Bendik Rønning Opstadd2e13392015-09-23 18:49:53 +02002244 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
Neal Cardwellca8a2262014-05-22 10:41:08 -04002245 tcp_cwnd_validate(sk, is_cwnd_limited);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002246 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07002247 }
Yuchung Chengb340b262015-08-12 11:18:19 -07002248 return !tp->packets_out && tcp_send_head(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002249}
2250
2251bool tcp_schedule_loss_probe(struct sock *sk)
2252{
2253 struct inet_connection_sock *icsk = inet_csk(sk);
2254 struct tcp_sock *tp = tcp_sk(sk);
2255 u32 timeout, tlp_time_stamp, rto_time_stamp;
Eric Dumazet740b0f12014-02-26 14:02:48 -08002256 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002257
2258 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2259 return false;
2260 /* No consecutive loss probes. */
2261 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2262 tcp_rearm_rto(sk);
2263 return false;
2264 }
2265 /* Don't do any loss probe on a Fast Open connection before 3WHS
2266 * finishes.
2267 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002268 if (tp->fastopen_rsk)
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002269 return false;
2270
2271 /* TLP is only scheduled when next timer event is RTO. */
2272 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2273 return false;
2274
2275 /* Schedule a loss probe in 2*RTT for SACK capable connections
2276 * in Open state, that are either limited by cwnd or application.
2277 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002278 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002279 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2280 return false;
2281
2282 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2283 tcp_send_head(sk))
2284 return false;
2285
2286 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
Yuchung Chengf9b99582015-09-18 11:40:33 -07002287 * for delayed ack when there's one outstanding packet. If no RTT
2288 * sample is available then probe after TCP_TIMEOUT_INIT.
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002289 */
Yuchung Chengf9b99582015-09-18 11:40:33 -07002290 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002291 if (tp->packets_out == 1)
2292 timeout = max_t(u32, timeout,
2293 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2294 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2295
2296 /* If RTO is shorter, just schedule TLP in its place. */
2297 tlp_time_stamp = tcp_time_stamp + timeout;
2298 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2299 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2300 s32 delta = rto_time_stamp - tcp_time_stamp;
2301 if (delta > 0)
2302 timeout = delta;
2303 }
2304
2305 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2306 TCP_RTO_MAX);
2307 return true;
2308}
2309
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002310/* Thanks to skb fast clones, we can detect if a prior transmit of
2311 * a packet is still in a qdisc or driver queue.
2312 * In this case, there is very little point doing a retransmit !
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002313 */
2314static bool skb_still_in_host_queue(const struct sock *sk,
2315 const struct sk_buff *skb)
2316{
Eric Dumazet39bb5e62014-10-30 10:32:34 -07002317 if (unlikely(skb_fclone_busy(sk, skb))) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07002318 NET_INC_STATS(sock_net(sk),
2319 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002320 return true;
2321 }
2322 return false;
2323}
2324
Yuchung Chengb340b262015-08-12 11:18:19 -07002325/* When probe timeout (PTO) fires, try send a new segment if possible, else
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002326 * retransmit the last segment.
2327 */
2328void tcp_send_loss_probe(struct sock *sk)
2329{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002330 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002331 struct sk_buff *skb;
2332 int pcount;
2333 int mss = tcp_current_mss(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002334
Yuchung Chengb340b262015-08-12 11:18:19 -07002335 skb = tcp_send_head(sk);
2336 if (skb) {
2337 if (tcp_snd_wnd_test(tp, skb, mss)) {
2338 pcount = tp->packets_out;
2339 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2340 if (tp->packets_out > pcount)
2341 goto probe_sent;
2342 goto rearm_timer;
2343 }
2344 skb = tcp_write_queue_prev(sk, skb);
2345 } else {
2346 skb = tcp_write_queue_tail(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002347 }
2348
Yuchung Chengbd3fa7e2018-12-05 14:38:38 -08002349 if (unlikely(!skb)) {
2350 WARN_ONCE(tp->packets_out,
2351 "invalid inflight: %u state %u cwnd %u mss %d\n",
2352 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2353 inet_csk(sk)->icsk_pending = 0;
2354 return;
2355 }
2356
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002357 /* At most one outstanding TLP retransmission. */
2358 if (tp->tlp_high_seq)
2359 goto rearm_timer;
2360
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002361 if (skb_still_in_host_queue(sk, skb))
2362 goto rearm_timer;
2363
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002364 pcount = tcp_skb_pcount(skb);
2365 if (WARN_ON(!pcount))
2366 goto rearm_timer;
2367
2368 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
Octavian Purdila6cc55e02014-06-06 17:32:37 +03002369 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2370 GFP_ATOMIC)))
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002371 goto rearm_timer;
Yuchung Chengb340b262015-08-12 11:18:19 -07002372 skb = tcp_write_queue_next(sk, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002373 }
2374
2375 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2376 goto rearm_timer;
2377
Eric Dumazet10d3be52016-04-21 10:55:23 -07002378 if (__tcp_retransmit_skb(sk, skb, 1))
Yuchung Chengb340b262015-08-12 11:18:19 -07002379 goto rearm_timer;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002380
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002381 /* Record snd_nxt for loss detection. */
Yuchung Chengb340b262015-08-12 11:18:19 -07002382 tp->tlp_high_seq = tp->snd_nxt;
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002383
Yuchung Chengb340b262015-08-12 11:18:19 -07002384probe_sent:
Eric Dumazetc10d9312016-04-29 14:16:47 -07002385 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
Yuchung Chengb340b262015-08-12 11:18:19 -07002386 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2387 inet_csk(sk)->icsk_pending = 0;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002388rearm_timer:
Yuchung Chengfcd16c02015-08-12 11:18:18 -07002389 tcp_rearm_rto(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390}
2391
David S. Millera762a982005-07-05 15:18:51 -07002392/* Push out any pending frames which were held back due to
2393 * TCP_CORK or attempt at coalescing tiny packets.
2394 * The socket must be locked by the caller.
2395 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002396void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2397 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002398{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002399 /* If we are closed, the bytes will have to remain here.
2400 * In time closedown will finish, we empty the write queue and
2401 * all will be happy.
2402 */
2403 if (unlikely(sk->sk_state == TCP_CLOSE))
2404 return;
2405
Mel Gorman99a1dec2012-07-31 16:44:14 -07002406 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
Eric Dumazet7450aaf2015-11-30 08:57:28 -08002407 sk_gfp_mask(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002408 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002409}
2410
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002411/* Send _single_ skb sitting at the send head. This function requires
2412 * true push pending frames to setup probe timer etc.
2413 */
2414void tcp_push_one(struct sock *sk, unsigned int mss_now)
2415{
David S. Millerfe067e82007-03-07 12:12:44 -08002416 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002417
2418 BUG_ON(!skb || skb->len < mss_now);
2419
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002420 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002421}
2422
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423/* This function returns the amount that we can raise the
2424 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002425 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 * 1. The window can never be shrunk once it is offered (RFC 793)
2427 * 2. We limit memory per socket
2428 *
2429 * RFC 1122:
2430 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2431 * RECV.NEXT + RCV.WIN fixed until:
2432 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2433 *
2434 * i.e. don't raise the right edge of the window until you can raise
2435 * it at least MSS bytes.
2436 *
2437 * Unfortunately, the recommended algorithm breaks header prediction,
2438 * since header prediction assumes th->window stays fixed.
2439 *
2440 * Strictly speaking, keeping th->window fixed violates the receiver
2441 * side SWS prevention criteria. The problem is that under this rule
2442 * a stream of single byte packets will cause the right side of the
2443 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002444 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 * Of course, if the sender implements sender side SWS prevention
2446 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002447 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002449 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 * If the free space is less than the 1/4 of the maximum
2451 * space available and the free space is less than 1/2 mss,
2452 * then set the window to 0.
2453 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2454 * Otherwise, just prevent the window from shrinking
2455 * and from being larger than the largest representable value.
2456 *
2457 * This prevents incremental opening of the window in the regime
2458 * where TCP is limited by the speed of the reader side taking
2459 * data out of the TCP receive queue. It does nothing about
2460 * those cases where the window is constrained on the sender side
2461 * because the pipeline is full.
2462 *
2463 * BSD also seems to "accidentally" limit itself to windows that are a
2464 * multiple of MSS, at least until the free space gets quite small.
2465 * This would appear to be a side effect of the mbuf implementation.
2466 * Combining these two algorithms results in the observed behavior
2467 * of having a fixed window size at almost all times.
2468 *
2469 * Below we obtain similar behavior by forcing the offered window to
2470 * a multiple of the mss when it is feasible to do so.
2471 *
2472 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2473 * Regular options like TIMESTAMP are taken into account.
2474 */
2475u32 __tcp_select_window(struct sock *sk)
2476{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002477 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002479 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 * here. I don't know if the value based on our guesses
2481 * of peer's MSS is better for the performance. It's more correct
2482 * but may be worse for the performance because of rcv_mss
2483 * fluctuations. --SAW 1998/11/1
2484 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002485 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 int free_space = tcp_space(sk);
Florian Westphal86c1a042014-02-19 12:51:10 +01002487 int allowed_space = tcp_full_space(sk);
2488 int full_space = min_t(int, tp->window_clamp, allowed_space);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 int window;
2490
Eric Dumazetca876df2017-02-01 08:33:53 -08002491 if (unlikely(mss > full_space)) {
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002492 mss = full_space;
Eric Dumazetca876df2017-02-01 08:33:53 -08002493 if (mss <= 0)
2494 return 0;
2495 }
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002496 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002497 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002499 if (tcp_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002500 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2501 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502
Florian Westphal86c1a042014-02-19 12:51:10 +01002503 /* free_space might become our new window, make sure we don't
2504 * increase it due to wscale.
2505 */
2506 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2507
2508 /* if free space is less than mss estimate, or is below 1/16th
2509 * of the maximum allowed, try to move to zero-window, else
2510 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2511 * new incoming data is dropped due to memory limits.
2512 * With large window, mss test triggers way too late in order
2513 * to announce zero window in time before rmem limit kicks in.
2514 */
2515 if (free_space < (allowed_space >> 4) || free_space < mss)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 return 0;
2517 }
2518
2519 if (free_space > tp->rcv_ssthresh)
2520 free_space = tp->rcv_ssthresh;
2521
2522 /* Don't do rounding if we are using window scaling, since the
2523 * scaled window will not line up with the MSS boundary anyway.
2524 */
2525 window = tp->rcv_wnd;
2526 if (tp->rx_opt.rcv_wscale) {
2527 window = free_space;
2528
2529 /* Advertise enough space so that it won't get scaled away.
2530 * Import case: prevent zero window announcement if
2531 * 1<<rcv_wscale > mss.
2532 */
2533 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2534 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2535 << tp->rx_opt.rcv_wscale);
2536 } else {
2537 /* Get the largest window that is a nice multiple of mss.
2538 * Window clamp already applied above.
2539 * If our current window offering is within 1 mss of the
2540 * free space we just keep it. This prevents the divide
2541 * and multiply from happening most of the time.
2542 * We also don't do any window rounding when the free space
2543 * is too small.
2544 */
2545 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002546 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002547 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002548 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002549 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 }
2551
2552 return window;
2553}
2554
Martin KaFai Laucfea5a62016-04-19 22:39:29 -07002555void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2556 const struct sk_buff *next_skb)
Martin KaFai Lau082ac2d2016-04-19 22:39:28 -07002557{
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04002558 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2559 const struct skb_shared_info *next_shinfo =
2560 skb_shinfo(next_skb);
Martin KaFai Lau082ac2d2016-04-19 22:39:28 -07002561 struct skb_shared_info *shinfo = skb_shinfo(skb);
2562
Soheil Hassas Yeganeh0a2cf202016-04-27 23:39:01 -04002563 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
Martin KaFai Lau082ac2d2016-04-19 22:39:28 -07002564 shinfo->tskey = next_shinfo->tskey;
Martin KaFai Lau2de80232016-04-19 22:50:48 -07002565 TCP_SKB_CB(skb)->txstamp_ack |=
2566 TCP_SKB_CB(next_skb)->txstamp_ack;
Martin KaFai Lau082ac2d2016-04-19 22:39:28 -07002567 }
2568}
2569
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002570/* Collapses two adjacent SKB's during retransmission. */
2571static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
2573 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002574 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002575 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002577 skb_size = skb->len;
2578 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002580 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002581
Eric Dumazete12c42c2017-10-30 23:08:20 -07002582 tcp_highest_sack_replace(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002584 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002586 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2587 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002589 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2590 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002592 if (skb->ip_summed != CHECKSUM_PARTIAL)
2593 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002595 /* Update sequence range on original skb. */
2596 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002598 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002599 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002601 /* All done, get rid of second SKB and account for it so
2602 * packet counting does not break.
2603 */
2604 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07002605 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002606
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002607 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002608 tcp_clear_retrans_hints_partial(tp);
2609 if (next_skb == tp->retransmit_skb_hint)
2610 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002611
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002612 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2613
Martin KaFai Lau082ac2d2016-04-19 22:39:28 -07002614 tcp_skb_collapse_tstamp(skb, next_skb);
2615
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002616 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617}
2618
Andi Kleen67edfef2009-07-21 23:00:40 +00002619/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002620static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002621{
2622 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002623 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002624 /* TODO: SACK collapsing could be used to remove this condition */
2625 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002626 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002627 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002628 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002629 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002630 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002631 /* Some heurestics for collapsing over SACK'd could be invented */
2632 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002633 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002634
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002635 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002636}
2637
Andi Kleen67edfef2009-07-21 23:00:40 +00002638/* Collapse packets in the retransmit queue to make to create
2639 * less packets on the wire. This is only done on retransmission.
2640 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002641static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2642 int space)
2643{
2644 struct tcp_sock *tp = tcp_sk(sk);
2645 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002646 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002647
2648 if (!sysctl_tcp_retrans_collapse)
2649 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002650 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002651 return;
2652
2653 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2654 if (!tcp_can_collapse(sk, skb))
2655 break;
2656
Martin KaFai Laua643b5d2016-04-25 14:44:49 -07002657 if (!tcp_skb_can_collapse_to(to))
2658 break;
2659
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002660 space -= skb->len;
2661
2662 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002663 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002664 continue;
2665 }
2666
2667 if (space < 0)
2668 break;
2669 /* Punt if not enough space exists in the first SKB for
2670 * the data in the second
2671 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002672 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002673 break;
2674
2675 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2676 break;
2677
2678 tcp_collapse_retrans(sk, to);
2679 }
2680}
2681
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682/* This retransmits one SKB. Policy decisions and retransmit queue
2683 * state updates are done by the caller. Returns non-zero if an
2684 * error occurred which prevented the send.
2685 */
Eric Dumazet10d3be52016-04-21 10:55:23 -07002686int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687{
John Heffner5d424d52006-03-20 17:53:41 -08002688 struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazet10d3be52016-04-21 10:55:23 -07002689 struct tcp_sock *tp = tcp_sk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002690 unsigned int cur_mss;
Eric Dumazet10d3be52016-04-21 10:55:23 -07002691 int diff, len, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692
Eric Dumazet10d3be52016-04-21 10:55:23 -07002693
2694 /* Inconclusive MTU probe */
2695 if (icsk->icsk_mtup.probe_size)
John Heffner5d424d52006-03-20 17:53:41 -08002696 icsk->icsk_mtup.probe_size = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002697
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002699 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 */
2701 if (atomic_read(&sk->sk_wmem_alloc) >
Eric Dumazetffb4d6c2016-09-15 08:12:33 -07002702 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2703 sk->sk_sndbuf))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704 return -EAGAIN;
2705
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002706 if (skb_still_in_host_queue(sk, skb))
2707 return -EBUSY;
2708
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
Eric Dumazet74a4c092018-05-14 21:14:26 -07002710 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2711 WARN_ON_ONCE(1);
2712 return -EINVAL;
2713 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2715 return -ENOMEM;
2716 }
2717
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002718 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2719 return -EHOSTUNREACH; /* Routing failure or similar. */
2720
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002721 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002722
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 /* If receiver has shrunk his window, and skb is out of
2724 * new window, do not retransmit it. The exception is the
2725 * case, when window is shrunk to zero. In this case
2726 * our retransmit serves as a zero window probe.
2727 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002728 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2729 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730 return -EAGAIN;
2731
Eric Dumazet10d3be52016-04-21 10:55:23 -07002732 len = cur_mss * segs;
2733 if (skb->len > len) {
2734 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002736 } else {
Eric Dumazet10d3be52016-04-21 10:55:23 -07002737 if (skb_unclone(skb, GFP_ATOMIC))
2738 return -ENOMEM;
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002739
Eric Dumazet10d3be52016-04-21 10:55:23 -07002740 diff = tcp_skb_pcount(skb);
2741 tcp_set_skb_tso_segs(skb, cur_mss);
2742 diff -= tcp_skb_pcount(skb);
2743 if (diff)
2744 tcp_adjust_pcount(sk, skb, diff);
2745 if (skb->len < cur_mss)
2746 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747 }
2748
Daniel Borkmann49213552015-05-19 21:04:22 +02002749 /* RFC3168, section 6.1.1.1. ECN fallback */
2750 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2751 tcp_ecn_clear_syn(sk, skb);
2752
Thomas Graf50bceae2013-04-11 10:57:18 +00002753 /* make sure skb->data is aligned on arches that require it
2754 * and check if ack-trimming & collapsing extended the headroom
2755 * beyond what csum_start can cover.
2756 */
2757 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2758 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet10a81982016-05-09 20:55:16 -07002759 struct sk_buff *nskb;
2760
Eric Dumazet10a81982016-05-09 20:55:16 -07002761 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
Yuchung Chengc84a5712014-02-28 16:42:26 -08002762 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2763 -ENOBUFS;
Eric Dumazet186a9c52017-09-13 20:30:39 -07002764 if (!err)
2765 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazet117632e2011-12-03 21:39:53 +00002766 } else {
Yuchung Chengc84a5712014-02-28 16:42:26 -08002767 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002768 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002769
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002770 if (likely(!err)) {
Eric Dumazet10d3be52016-04-21 10:55:23 -07002771 segs = tcp_skb_pcount(skb);
2772
Yuchung Chengc84a5712014-02-28 16:42:26 -08002773 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002774 /* Update global TCP statistics. */
Eric Dumazet10d3be52016-04-21 10:55:23 -07002775 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002776 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07002777 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
Eric Dumazet10d3be52016-04-21 10:55:23 -07002778 tp->total_retrans += segs;
Eric Dumazetfc9f3502014-04-28 22:00:29 -07002779 }
Yuchung Chengc84a5712014-02-28 16:42:26 -08002780 return err;
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002781}
2782
Eric Dumazet10d3be52016-04-21 10:55:23 -07002783int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002784{
2785 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet10d3be52016-04-21 10:55:23 -07002786 int err = __tcp_retransmit_skb(sk, skb, segs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787
2788 if (err == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002790 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002791 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 }
2793#endif
2794 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2795 tp->retrans_out += tcp_skb_pcount(skb);
2796
2797 /* Save stamp of the first retransmit. */
2798 if (!tp->retrans_stamp)
Eric Dumazet7faee5c2014-09-05 15:33:33 -07002799 tp->retrans_stamp = tcp_skb_timestamp(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800
Eric Dumazet1f3279a2014-04-20 17:58:17 -07002801 } else if (err != -EBUSY) {
Eric Dumazetc10d9312016-04-29 14:16:47 -07002802 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 }
Yuchung Cheng6e08d5e2014-07-02 12:07:16 -07002804
2805 if (tp->undo_retrans < 0)
2806 tp->undo_retrans = 0;
2807 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 return err;
2809}
2810
Andi Kleen67edfef2009-07-21 23:00:40 +00002811/* Check if we forward retransmits are possible in the current
2812 * window/congestion state.
2813 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002814static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002815{
2816 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002817 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002818
2819 /* Forward retransmissions are possible only during Recovery. */
2820 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002821 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002822
2823 /* No forward retransmissions in Reno are possible. */
2824 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002825 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002826
2827 /* Yeah, we have to make difficult choice between forward transmission
2828 * and retransmission... Both ways have their merits...
2829 *
2830 * For now we do not retransmit anything, while we have some new
2831 * segments to send. In the other cases, follow rule 3 for
2832 * NextSeg() specified in RFC3517.
2833 */
2834
2835 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002836 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002837
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002838 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002839}
2840
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841/* This gets called after a retransmit timeout, and the initially
2842 * retransmitted data is acknowledged. It tries to continue
2843 * resending the rest of the retransmit queue, until either
2844 * we've sent it all or the congestion window limit is reached.
2845 * If doing SACK, the first ACK which comes back for a timeout
2846 * based retransmit packet might feed us FACK information again.
2847 * If so, we use it to avoid unnecessarily retransmissions.
2848 */
2849void tcp_xmit_retransmit_queue(struct sock *sk)
2850{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002851 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 struct tcp_sock *tp = tcp_sk(sk);
2853 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002854 struct sk_buff *hole = NULL;
Eric Dumazeta3d2e9f2016-06-27 17:38:50 +02002855 u32 max_segs, last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002856 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002857 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002858
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002859 if (!tp->packets_out)
2860 return;
2861
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002862 if (!tp->lost_out)
2863 tp->retransmit_high = tp->snd_una;
2864
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002865 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002866 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002867 last_lost = TCP_SKB_CB(skb)->end_seq;
2868 if (after(last_lost, tp->retransmit_high))
2869 last_lost = tp->retransmit_high;
2870 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002871 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002872 last_lost = tp->snd_una;
2873 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874
Neal Cardwelled6e7262016-09-19 23:39:17 -04002875 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002876 tcp_for_write_queue_from(skb, sk) {
Eric Dumazetdca0aaf2016-08-17 07:48:36 -07002877 __u8 sacked;
Eric Dumazet10d3be52016-04-21 10:55:23 -07002878 int segs;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002880 if (skb == tcp_send_head(sk))
2881 break;
2882 /* we could do better than to assign each time */
Ian Morris51456b22015-04-03 09:17:26 +01002883 if (!hole)
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002884 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002885
Eric Dumazet10d3be52016-04-21 10:55:23 -07002886 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2887 if (segs <= 0)
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002888 return;
Eric Dumazetdca0aaf2016-08-17 07:48:36 -07002889 sacked = TCP_SKB_CB(skb)->sacked;
Eric Dumazeta3d2e9f2016-06-27 17:38:50 +02002890 /* In case tcp_shift_skb_data() have aggregated large skbs,
2891 * we need to make sure not sending too bigs TSO packets
2892 */
2893 segs = min_t(int, segs, max_segs);
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002894
2895 if (fwd_rexmitting) {
2896begin_fwd:
2897 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2898 break;
2899 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2900
2901 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002902 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002903 if (!tcp_can_forward_retransmit(sk))
2904 break;
2905 /* Backtrack if necessary to non-L'ed skb */
Ian Morris00db4122015-04-03 09:17:27 +01002906 if (hole) {
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002907 skb = hole;
2908 hole = NULL;
2909 }
2910 fwd_rexmitting = 1;
2911 goto begin_fwd;
2912
2913 } else if (!(sacked & TCPCB_LOST)) {
Ian Morris51456b22015-04-03 09:17:26 +01002914 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002915 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002916 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002918 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002919 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002920 if (icsk->icsk_ca_state != TCP_CA_Loss)
2921 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2922 else
2923 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2924 }
2925
2926 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002927 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002928
Eric Dumazetf9616c32016-09-20 22:45:58 -07002929 if (tcp_small_queue_check(sk, skb, 1))
2930 return;
2931
Eric Dumazet10d3be52016-04-21 10:55:23 -07002932 if (tcp_retransmit_skb(sk, skb, segs))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002933 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002934
Yuchung Chengde1d6572016-09-21 16:16:14 -07002935 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002936
Yuchung Cheng684bad12012-09-02 17:38:04 +00002937 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002938 tp->prr_out += tcp_skb_pcount(skb);
2939
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002940 if (skb == tcp_write_queue_head(sk))
2941 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2942 inet_csk(sk)->icsk_rto,
2943 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945}
2946
Eric Dumazetd83769a2015-04-21 18:32:24 -07002947/* We allow to exceed memory limits for FIN packets to expedite
2948 * connection tear down and (memory) recovery.
Eric Dumazet845704a2015-04-23 10:42:39 -07002949 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2950 * or even be forced to close flow without any FIN.
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002951 * In general, we want to allow one skb per socket to avoid hangs
2952 * with edge trigger epoll()
Eric Dumazetd83769a2015-04-21 18:32:24 -07002953 */
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07002954void sk_forced_mem_schedule(struct sock *sk, int size)
Eric Dumazetd83769a2015-04-21 18:32:24 -07002955{
Johannes Weinere8056052016-01-14 15:21:14 -08002956 int amt;
Eric Dumazetd83769a2015-04-21 18:32:24 -07002957
2958 if (size <= sk->sk_forward_alloc)
2959 return;
2960 amt = sk_mem_pages(size);
2961 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
Johannes Weinere8056052016-01-14 15:21:14 -08002962 sk_memory_allocated_add(sk, amt);
2963
Johannes Weinerbaac50b2016-01-14 15:21:17 -08002964 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2965 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
Eric Dumazetd83769a2015-04-21 18:32:24 -07002966}
2967
Eric Dumazet845704a2015-04-23 10:42:39 -07002968/* Send a FIN. The caller locks the socket for us.
2969 * We should try to send a FIN packet really hard, but eventually give up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 */
2971void tcp_send_fin(struct sock *sk)
2972{
Eric Dumazet845704a2015-04-23 10:42:39 -07002973 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002974 struct tcp_sock *tp = tcp_sk(sk);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002975
Eric Dumazet845704a2015-04-23 10:42:39 -07002976 /* Optimization, tack on the FIN if we have one skb in write queue and
2977 * this skb was not yet sent, or we are under memory pressure.
2978 * Note: in the latter case, FIN packet will be sent after a timeout,
2979 * as TCP stack thinks it has already been transmitted.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980 */
Eric Dumazetb8da51e2015-05-15 12:39:27 -07002981 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
Eric Dumazet845704a2015-04-23 10:42:39 -07002982coalesce:
2983 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2984 TCP_SKB_CB(tskb)->end_seq++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 tp->write_seq++;
Eric Dumazet845704a2015-04-23 10:42:39 -07002986 if (!tcp_send_head(sk)) {
2987 /* This means tskb was already sent.
2988 * Pretend we included the FIN on previous transmit.
2989 * We need to set tp->snd_nxt to the value it would have
2990 * if FIN had been sent. This is because retransmit path
2991 * does not change tp->snd_nxt.
2992 */
2993 tp->snd_nxt++;
2994 return;
2995 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 } else {
Eric Dumazet845704a2015-04-23 10:42:39 -07002997 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2998 if (unlikely(!skb)) {
2999 if (tskb)
3000 goto coalesce;
3001 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 }
Eric Dumazetd83769a2015-04-21 18:32:24 -07003003 skb_reserve(skb, MAX_TCP_HEADER);
Eric Dumazeta6c5ea42015-05-15 12:39:26 -07003004 sk_forced_mem_schedule(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08003006 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00003007 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008 tcp_queue_skb(sk, skb);
3009 }
Eric Dumazet845704a2015-04-23 10:42:39 -07003010 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011}
3012
3013/* We get here when a process closes a file descriptor (either due to
3014 * an explicit close() or as a byproduct of exit()'ing) and there
3015 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07003016 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 */
Al Virodd0fc662005-10-07 07:46:04 +01003018void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 struct sk_buff *skb;
3021
3022 /* NOTE: No TCP options attached and we never retransmit this. */
3023 skb = alloc_skb(MAX_TCP_HEADER, priority);
3024 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07003025 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 return;
3027 }
3028
3029 /* Reserve space for headers and prepare control bits. */
3030 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08003031 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00003032 TCPHDR_ACK | TCPHDR_RST);
Eric Dumazet675ee232015-09-23 14:00:21 -07003033 skb_mstamp_get(&skb->skb_mstamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 /* Send it off. */
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003035 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07003036 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07003037
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003038 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039}
3040
Andi Kleen67edfef2009-07-21 23:00:40 +00003041/* Send a crossed SYN-ACK during socket establishment.
3042 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 * a SYN packet that crossed the incoming SYN that caused this routine
3044 * to get called. If this assumption fails then the initial rcv_wnd
3045 * and rcv_wscale values will not be correct.
3046 */
3047int tcp_send_synack(struct sock *sk)
3048{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003049 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050
David S. Millerfe067e82007-03-07 12:12:44 -08003051 skb = tcp_write_queue_head(sk);
Ian Morris51456b22015-04-03 09:17:26 +01003052 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00003053 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 return -EFAULT;
3055 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04003056 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 if (skb_cloned(skb)) {
3058 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
Ian Morris51456b22015-04-03 09:17:26 +01003059 if (!nskb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08003061 tcp_unlink_write_queue(skb, sk);
Eric Dumazetf4a775d2014-09-22 16:29:32 -07003062 __skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08003063 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08003064 sk_wmem_free_skb(sk, skb);
3065 sk->sk_wmem_queued += nskb->truesize;
3066 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067 skb = nskb;
3068 }
3069
Eric Dumazet4de075e2011-09-27 13:25:05 -04003070 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Florian Westphal735d3832014-09-29 13:08:30 +02003071 tcp_ecn_send_synack(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003073 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003074}
3075
Eric Dumazet4aea39c2012-06-03 20:33:21 +00003076/**
3077 * tcp_make_synack - Prepare a SYN-ACK.
3078 * sk: listener socket
3079 * dst: dst entry attached to the SYNACK
3080 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00003081 *
3082 * Allocate one skb and build a SYNACK packet.
3083 * @dst is consumed : Caller should not use it again.
3084 */
Eric Dumazet5d062de2015-09-25 07:39:19 -07003085struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00003086 struct request_sock *req,
Eric Dumazetca6fb062015-10-02 11:43:35 -07003087 struct tcp_fastopen_cookie *foc,
Eric Dumazetb3d05142016-04-13 22:05:39 -07003088 enum tcp_synack_type synack_type)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07003090 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet5d062de2015-09-25 07:39:19 -07003091 const struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003092 struct tcp_md5sig_key *md5 = NULL;
Eric Dumazet5d062de2015-09-25 07:39:19 -07003093 struct tcp_out_options opts;
3094 struct sk_buff *skb;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003095 int tcp_header_size;
Eric Dumazet5d062de2015-09-25 07:39:19 -07003096 struct tcphdr *th;
3097 u16 user_mss;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003098 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099
Eric Dumazetca6fb062015-10-02 11:43:35 -07003100 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
Eric Dumazet4aea39c2012-06-03 20:33:21 +00003101 if (unlikely(!skb)) {
3102 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00003104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 /* Reserve space for headers. */
3106 skb_reserve(skb, MAX_TCP_HEADER);
3107
Eric Dumazetb3d05142016-04-13 22:05:39 -07003108 switch (synack_type) {
3109 case TCP_SYNACK_NORMAL:
Eric Dumazet9e17f8a2015-11-01 15:36:55 -08003110 skb_set_owner_w(skb, req_to_sk(req));
Eric Dumazetb3d05142016-04-13 22:05:39 -07003111 break;
3112 case TCP_SYNACK_COOKIE:
3113 /* Under synflood, we do not attach skb to a socket,
3114 * to avoid false sharing.
3115 */
3116 break;
3117 case TCP_SYNACK_FASTOPEN:
Eric Dumazetca6fb062015-10-02 11:43:35 -07003118 /* sk is a const pointer, because we want to express multiple
3119 * cpu might call us concurrently.
3120 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3121 */
3122 skb_set_owner_w(skb, (struct sock *)sk);
Eric Dumazetb3d05142016-04-13 22:05:39 -07003123 break;
Eric Dumazetca6fb062015-10-02 11:43:35 -07003124 }
Eric Dumazet4aea39c2012-06-03 20:33:21 +00003125 skb_dst_set(skb, dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126
David S. Miller0dbaee32010-12-13 12:52:14 -08003127 mss = dst_metric_advmss(dst);
Eric Dumazet5d062de2015-09-25 07:39:19 -07003128 user_mss = READ_ONCE(tp->rx_opt.user_mss);
3129 if (user_mss && user_mss < mss)
3130 mss = user_mss;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003131
Adam Langley33ad7982008-07-19 00:04:31 -07003132 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07003133#ifdef CONFIG_SYN_COOKIES
3134 if (unlikely(req->cookie_ts))
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003135 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
Florian Westphal8b5f12d2008-10-26 23:10:12 -07003136 else
3137#endif
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003138 skb_mstamp_get(&skb->skb_mstamp);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003139
3140#ifdef CONFIG_TCP_MD5SIG
3141 rcu_read_lock();
Eric Dumazetfd3a1542015-03-24 15:58:56 -07003142 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
Eric Dumazet80f03e22015-03-24 15:58:52 -07003143#endif
Eric Dumazet58d607d2015-09-15 15:24:20 -07003144 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
Eric Dumazet37bfbdd2015-09-25 07:39:17 -07003145 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3146 sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07003147
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07003148 skb_push(skb, tcp_header_size);
3149 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150
Eric Dumazetea1627c2016-05-13 09:16:40 -07003151 th = (struct tcphdr *)skb->data;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 memset(th, 0, sizeof(struct tcphdr));
3153 th->syn = 1;
3154 th->ack = 1;
Eric Dumazet6ac705b2015-09-25 07:39:18 -07003155 tcp_ecn_make_synack(req, th);
Eric Dumazetb44084c2013-10-10 00:04:37 -07003156 th->source = htons(ireq->ir_num);
Eric Dumazet634fb9792013-10-09 15:21:29 -07003157 th->dest = ireq->ir_rmt_port;
Eric Dumazet3920a5b2017-11-02 12:30:25 -07003158 skb->ip_summed = CHECKSUM_PARTIAL;
3159 th->seq = htonl(tcp_rsk(req)->snt_isn);
Jerry Chu83368862012-08-31 12:29:12 +00003160 /* XXX data is queued and acked as is. No buffer/window check */
3161 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162
3163 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Eric Dumazeted53d0a2015-10-08 19:33:23 -07003164 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
Eric Dumazet5d062de2015-09-25 07:39:19 -07003165 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 th->doff = (tcp_header_size >> 2);
Eric Dumazet90bbcc62016-04-27 16:44:32 -07003167 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003168
3169#ifdef CONFIG_TCP_MD5SIG
3170 /* Okay, we have all we need - do the md5 hash if needed */
Eric Dumazet80f03e22015-03-24 15:58:52 -07003171 if (md5)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00003172 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Eric Dumazet39f8e582015-03-24 15:58:55 -07003173 md5, req_to_sk(req), skb);
Eric Dumazet80f03e22015-03-24 15:58:52 -07003174 rcu_read_unlock();
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003175#endif
3176
Eric Dumazetb50edd72015-04-09 13:31:56 -07003177 /* Do not fool tcpdump (if any), clean our debris */
3178 skb->tstamp.tv64 = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179 return skb;
3180}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003181EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182
Daniel Borkmann81164412015-01-05 23:57:48 +01003183static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3184{
3185 struct inet_connection_sock *icsk = inet_csk(sk);
3186 const struct tcp_congestion_ops *ca;
3187 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3188
3189 if (ca_key == TCP_CA_UNSPEC)
3190 return;
3191
3192 rcu_read_lock();
3193 ca = tcp_ca_find_key(ca_key);
3194 if (likely(ca && try_module_get(ca->owner))) {
3195 module_put(icsk->icsk_ca_ops->owner);
3196 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3197 icsk->icsk_ca_ops = ca;
3198 }
3199 rcu_read_unlock();
3200}
3201
Andi Kleen67edfef2009-07-21 23:00:40 +00003202/* Do all connect socket setups that can be done AF independent. */
stephen hemmingerf7e56a72013-12-29 11:39:51 -08003203static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04003205 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 struct tcp_sock *tp = tcp_sk(sk);
3207 __u8 rcv_wscale;
3208
3209 /* We'll fix this up when we get a response from the other end.
3210 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3211 */
3212 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08003213 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003215#ifdef CONFIG_TCP_MD5SIG
Ian Morris00db4122015-04-03 09:17:27 +01003216 if (tp->af_specific->md5_lookup(sk, sk))
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08003217 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3218#endif
3219
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 /* If user gave his TCP_MAXSEG, record it to clamp */
3221 if (tp->rx_opt.user_mss)
3222 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3223 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08003224 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 tcp_sync_mss(sk, dst_mtu(dst));
3226
Daniel Borkmann81164412015-01-05 23:57:48 +01003227 tcp_ca_dst_init(sk, dst);
3228
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229 if (!tp->window_clamp)
3230 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08003231 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07003232 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3233 tp->advmss = tp->rx_opt.user_mss;
3234
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00003237 /* limit the window selection if the user enforce a smaller rx buffer */
3238 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3239 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3240 tp->window_clamp = tcp_full_space(sk);
3241
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 tcp_select_initial_window(tcp_full_space(sk),
3243 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3244 &tp->rcv_wnd,
3245 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08003246 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00003247 &rcv_wscale,
3248 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249
3250 tp->rx_opt.rcv_wscale = rcv_wscale;
3251 tp->rcv_ssthresh = tp->rcv_wnd;
3252
3253 sk->sk_err = 0;
3254 sock_reset_flag(sk, SOCK_DONE);
3255 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08003256 tcp_init_wl(tp, 0);
Eric Dumazet74a4c092018-05-14 21:14:26 -07003257 tcp_write_queue_purge(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258 tp->snd_una = tp->write_seq;
3259 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003260 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00003261 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003262
3263 if (likely(!tp->repair))
3264 tp->rcv_nxt = 0;
Andrew Vaginc7781a62013-08-27 12:20:40 +04003265 else
3266 tp->rcv_tstamp = tcp_time_stamp;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003267 tp->rcv_wup = tp->rcv_nxt;
3268 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003269
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003270 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3271 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003272 tcp_clear_retrans(tp);
3273}
3274
Yuchung Cheng783237e2012-07-19 06:43:07 +00003275static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3276{
3277 struct tcp_sock *tp = tcp_sk(sk);
3278 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3279
3280 tcb->end_seq += skb->len;
Eric Dumazetf4a775d2014-09-22 16:29:32 -07003281 __skb_header_release(skb);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003282 __tcp_add_write_queue_tail(sk, skb);
3283 sk->sk_wmem_queued += skb->truesize;
3284 sk_mem_charge(sk, skb->truesize);
3285 tp->write_seq = tcb->end_seq;
3286 tp->packets_out += tcp_skb_pcount(skb);
3287}
3288
3289/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3290 * queue a data-only packet after the regular SYN, such that regular SYNs
3291 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3292 * only the SYN sequence, the data are retransmitted in the first ACK.
3293 * If cookie is not cached or other error occurs, falls back to send a
3294 * regular SYN with Fast Open cookie request option.
3295 */
3296static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3297{
3298 struct tcp_sock *tp = tcp_sk(sk);
3299 struct tcp_fastopen_request *fo = tp->fastopen_req;
Eric Dumazet07e100f2015-12-16 13:53:10 -08003300 int syn_loss = 0, space, err = 0;
Yuchung Chengaab48742012-07-19 06:43:10 +00003301 unsigned long last_syn_loss = 0;
Eric Dumazet355a9012014-11-17 23:06:20 -08003302 struct sk_buff *syn_data;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003303
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003304 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00003305 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3306 &syn_loss, &last_syn_loss);
3307 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3308 if (syn_loss > 1 &&
3309 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3310 fo->cookie.len = -1;
3311 goto fallback;
3312 }
3313
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003314 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3315 fo->cookie.len = -1;
3316 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003317 goto fallback;
3318
3319 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3320 * user-MSS. Reserve maximum option space for middleboxes that add
3321 * private TCP options. The cost is reduced data space in SYN :(
3322 */
3323 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3324 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00003325 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00003326 MAX_TCP_OPTION_SPACE;
3327
Eric Dumazetf5ddcbb2014-02-20 10:09:18 -08003328 space = min_t(size_t, space, fo->size);
3329
3330 /* limit to order-0 allocations */
3331 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3332
Eric Dumazeteb934472015-05-19 13:26:55 -07003333 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
Eric Dumazet355a9012014-11-17 23:06:20 -08003334 if (!syn_data)
Yuchung Cheng783237e2012-07-19 06:43:07 +00003335 goto fallback;
Eric Dumazet355a9012014-11-17 23:06:20 -08003336 syn_data->ip_summed = CHECKSUM_PARTIAL;
3337 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
Eric Dumazet07e100f2015-12-16 13:53:10 -08003338 if (space) {
3339 int copied = copy_from_iter(skb_put(syn_data, space), space,
3340 &fo->data->msg_iter);
3341 if (unlikely(!copied)) {
3342 kfree_skb(syn_data);
3343 goto fallback;
3344 }
3345 if (copied != space) {
3346 skb_trim(syn_data, copied);
3347 space = copied;
3348 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003349 }
Eric Dumazet355a9012014-11-17 23:06:20 -08003350 /* No more data pending in inet_wait_for_connect() */
3351 if (space == fo->size)
3352 fo->data = NULL;
3353 fo->copied = space;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003354
Eric Dumazet355a9012014-11-17 23:06:20 -08003355 tcp_connect_queue_skb(sk, syn_data);
3356
3357 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3358
3359 syn->skb_mstamp = syn_data->skb_mstamp;
3360
3361 /* Now full SYN+DATA was cloned and sent (or not),
3362 * remove the SYN from the original skb (syn_data)
3363 * we keep in write queue in case of a retransmit, as we
3364 * also have the SYN packet (with no data) in the same queue.
Eric Dumazet431a9122014-03-09 17:36:02 -07003365 */
Eric Dumazet355a9012014-11-17 23:06:20 -08003366 TCP_SKB_CB(syn_data)->seq++;
3367 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3368 if (!err) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00003369 tp->syn_data = (fo->copied > 0);
Yuchung Chengf19c29e2014-03-03 12:31:36 -08003370 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
Yuchung Cheng783237e2012-07-19 06:43:07 +00003371 goto done;
3372 }
Yuchung Cheng783237e2012-07-19 06:43:07 +00003373
Eric Dumazetb4635212017-09-19 10:05:57 -07003374 /* data was not sent, this is our new send_head */
3375 sk->sk_send_head = syn_data;
3376 tp->packets_out -= tcp_skb_pcount(syn_data);
3377
Yuchung Cheng783237e2012-07-19 06:43:07 +00003378fallback:
3379 /* Send a regular SYN with Fast Open cookie request option */
3380 if (fo->cookie.len > 0)
3381 fo->cookie.len = 0;
3382 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3383 if (err)
3384 tp->syn_fastopen = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003385done:
3386 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3387 return err;
3388}
3389
Andi Kleen67edfef2009-07-21 23:00:40 +00003390/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391int tcp_connect(struct sock *sk)
3392{
3393 struct tcp_sock *tp = tcp_sk(sk);
3394 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00003395 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
Eric Dumazet87fdcfe2017-08-08 01:41:58 -07003397 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3398 return -EHOSTUNREACH; /* Routing failure or similar. */
3399
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 tcp_connect_init(sk);
3401
Andrey Vagin2b916472012-11-22 01:13:58 +00003402 if (unlikely(tp->repair)) {
3403 tcp_finish_connect(sk, NULL);
3404 return 0;
3405 }
3406
Eric Dumazeteb934472015-05-19 13:26:55 -07003407 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
Eric Dumazet355a9012014-11-17 23:06:20 -08003408 if (unlikely(!buff))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 return -ENOBUFS;
3410
Changli Gaoa3433f32010-06-12 14:01:43 +00003411 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003412 tp->retrans_stamp = tcp_time_stamp;
Yuchung Cheng783237e2012-07-19 06:43:07 +00003413 tcp_connect_queue_skb(sk, buff);
Florian Westphal735d3832014-09-29 13:08:30 +02003414 tcp_ecn_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415
Yuchung Cheng783237e2012-07-19 06:43:07 +00003416 /* Send off SYN; include data in Fast Open. */
3417 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3418 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00003419 if (err == -ECONNREFUSED)
3420 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07003421
3422 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3423 * in order to make this packet get counted in tcpOutSegs.
3424 */
3425 tp->snd_nxt = tp->write_seq;
3426 tp->pushed_seq = tp->write_seq;
Eric Dumazetb4635212017-09-19 10:05:57 -07003427 buff = tcp_send_head(sk);
3428 if (unlikely(buff)) {
3429 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3430 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3431 }
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07003432 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
3434 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003435 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3436 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437 return 0;
3438}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00003439EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440
3441/* Send out a delayed ack, the caller does the policy checking
3442 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3443 * for details.
3444 */
3445void tcp_send_delayed_ack(struct sock *sk)
3446{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003447 struct inet_connection_sock *icsk = inet_csk(sk);
3448 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 unsigned long timeout;
3450
3451 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003452 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003453 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454
Ilpo Järvinen056834d2007-12-31 14:57:14 -08003455 if (icsk->icsk_ack.pingpong ||
3456 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457 max_ato = TCP_DELACK_MAX;
3458
3459 /* Slow path, intersegment interval is "high". */
3460
3461 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003462 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463 * directly.
3464 */
Eric Dumazet740b0f12014-02-26 14:02:48 -08003465 if (tp->srtt_us) {
3466 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3467 TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468
3469 if (rtt < max_ato)
3470 max_ato = rtt;
3471 }
3472
3473 ato = min(ato, max_ato);
3474 }
3475
3476 /* Stay within the limit we were given */
3477 timeout = jiffies + ato;
3478
3479 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003480 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 /* If delack timer was blocked or is about to expire,
3482 * send ACK now.
3483 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003484 if (icsk->icsk_ack.blocked ||
3485 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486 tcp_send_ack(sk);
3487 return;
3488 }
3489
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003490 if (!time_before(timeout, icsk->icsk_ack.timeout))
3491 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003493 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3494 icsk->icsk_ack.timeout = timeout;
3495 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496}
3497
3498/* This routine sends an ack and also updates the window. */
Yuchung Cheng1fcccc52018-07-18 13:56:34 -07003499void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003501 struct sk_buff *buff;
3502
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003504 if (sk->sk_state == TCP_CLOSE)
3505 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003507 /* We are not putting this on the write queue, so
3508 * tcp_transmit_skb() will set the ownership to this
3509 * sock.
3510 */
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003511 buff = alloc_skb(MAX_TCP_HEADER,
3512 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3513 if (unlikely(!buff)) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003514 inet_csk_schedule_ack(sk);
3515 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3516 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3517 TCP_DELACK_MAX, TCP_RTO_MAX);
3518 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003520
3521 /* Reserve space for headers and prepare control bits. */
3522 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003523 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003524
Eric Dumazet98781962015-02-03 18:31:53 -08003525 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3526 * too much.
3527 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3528 * We also avoid tcp_wfree() overhead (cache line miss accessing
3529 * tp->tsq_flags) by using regular sock_wfree()
3530 */
3531 skb_set_tcp_pure_ack(buff);
3532
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003533 /* Send it off, this clears delayed acks for us. */
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003534 skb_mstamp_get(&buff->skb_mstamp);
Yuchung Cheng1fcccc52018-07-18 13:56:34 -07003535 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3536}
Yuchung Cheng57ec8822018-07-18 13:56:35 -07003537EXPORT_SYMBOL_GPL(__tcp_send_ack);
Yuchung Cheng1fcccc52018-07-18 13:56:34 -07003538
3539void tcp_send_ack(struct sock *sk)
3540{
3541 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542}
3543
3544/* This routine sends a packet with an out of date sequence
3545 * number. It assumes the other end will try to ack it.
3546 *
3547 * Question: what should we make while urgent mode?
3548 * 4.4BSD forces sending single byte of data. We cannot send
3549 * out of window data, because we have SND.NXT==SND.MAX...
3550 *
3551 * Current solution: to send TWO zero-length segments in urgent mode:
3552 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3553 * out-of-date with SND.UNA-1 to probe window.
3554 */
Eric Dumazete520af42015-05-06 14:26:25 -07003555static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556{
3557 struct tcp_sock *tp = tcp_sk(sk);
3558 struct sk_buff *skb;
3559
3560 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003561 skb = alloc_skb(MAX_TCP_HEADER,
3562 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
Ian Morris51456b22015-04-03 09:17:26 +01003563 if (!skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 return -1;
3565
3566 /* Reserve space for headers and set control bits. */
3567 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 /* Use a previous sequence. This should cause the other
3569 * end to send an ack. Don't queue or clone SKB, just
3570 * send it.
3571 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003572 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Eric Dumazet7faee5c2014-09-05 15:33:33 -07003573 skb_mstamp_get(&skb->skb_mstamp);
Renato Westphale2e80092015-10-19 18:51:34 -02003574 NET_INC_STATS(sock_net(sk), mib);
Eric Dumazet7450aaf2015-11-30 08:57:28 -08003575 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576}
3577
Pavel Emelyanovee995282012-04-19 03:40:39 +00003578void tcp_send_window_probe(struct sock *sk)
3579{
3580 if (sk->sk_state == TCP_ESTABLISHED) {
3581 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
Eric Dumazete520af42015-05-06 14:26:25 -07003582 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
Pavel Emelyanovee995282012-04-19 03:40:39 +00003583 }
3584}
3585
Andi Kleen67edfef2009-07-21 23:00:40 +00003586/* Initiate keepalive or window probe from timer. */
Eric Dumazete520af42015-05-06 14:26:25 -07003587int tcp_write_wakeup(struct sock *sk, int mib)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003589 struct tcp_sock *tp = tcp_sk(sk);
3590 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003592 if (sk->sk_state == TCP_CLOSE)
3593 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594
Ian Morris00db4122015-04-03 09:17:27 +01003595 skb = tcp_send_head(sk);
3596 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003597 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003598 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003599 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003601 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3602 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003604 /* We are probing the opening of a window
3605 * but the window size is != 0
3606 * must have been a result SWS avoidance ( sender )
3607 */
3608 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3609 skb->len > mss) {
3610 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003611 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Octavian Purdila6cc55e02014-06-06 17:32:37 +03003612 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003613 return -1;
3614 } else if (!tcp_skb_pcount(skb))
Eric Dumazet5bbb4322015-06-11 09:15:17 -07003615 tcp_set_skb_tso_segs(skb, mss);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003616
Eric Dumazet4de075e2011-09-27 13:25:05 -04003617 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003618 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3619 if (!err)
3620 tcp_event_new_data_sent(sk, skb);
3621 return err;
3622 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003623 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Eric Dumazete520af42015-05-06 14:26:25 -07003624 tcp_xmit_probe_skb(sk, 1, mib);
3625 return tcp_xmit_probe_skb(sk, 0, mib);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003626 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627}
3628
3629/* A window probe timeout has occurred. If window is not closed send
3630 * a partial packet else a zero probe.
3631 */
3632void tcp_send_probe0(struct sock *sk)
3633{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003634 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 struct tcp_sock *tp = tcp_sk(sk);
Nikolay Borisovc6214a92016-02-03 09:46:54 +02003636 struct net *net = sock_net(sk);
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003637 unsigned long probe_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 int err;
3639
Eric Dumazete520af42015-05-06 14:26:25 -07003640 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641
David S. Millerfe067e82007-03-07 12:12:44 -08003642 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003644 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003645 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 return;
3647 }
3648
3649 if (err <= 0) {
Nikolay Borisovc6214a92016-02-03 09:46:54 +02003650 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003651 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003652 icsk->icsk_probes_out++;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003653 probe_max = TCP_RTO_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654 } else {
3655 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003656 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657 * Let local senders to fight for local resources.
3658 *
3659 * Use accumulated backoff yet.
3660 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003661 if (!icsk->icsk_probes_out)
3662 icsk->icsk_probes_out = 1;
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003663 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 }
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003665 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Eric Dumazet21c8fe92015-05-06 14:26:24 -07003666 tcp_probe0_when(sk, probe_max),
Eric Dumazetfcdd1cf2014-09-22 13:19:44 -07003667 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668}
Octavian Purdila5db92c92014-06-25 17:09:59 +03003669
Eric Dumazetea3bea32015-09-25 07:39:23 -07003670int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
Octavian Purdila5db92c92014-06-25 17:09:59 +03003671{
3672 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3673 struct flowi fl;
3674 int res;
3675
Eric Dumazet58d607d2015-09-15 15:24:20 -07003676 tcp_rsk(req)->txhash = net_tx_rndhash();
Eric Dumazetb3d05142016-04-13 22:05:39 -07003677 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
Octavian Purdila5db92c92014-06-25 17:09:59 +03003678 if (!res) {
Eric Dumazet90bbcc62016-04-27 16:44:32 -07003679 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Eric Dumazet02a1d6e2016-04-27 16:44:39 -07003680 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
Yuchung Cheng7e32b442016-09-21 16:16:15 -07003681 if (unlikely(tcp_passive_fastopen(sk)))
3682 tcp_sk(sk)->total_retrans++;
Octavian Purdila5db92c92014-06-25 17:09:59 +03003683 }
3684 return res;
3685}
3686EXPORT_SYMBOL(tcp_rtx_synack);