blob: 92fde8d1aa821c38b59ba467fe386eb3306c2dfe [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
Joe Perches91df42b2012-05-15 14:11:54 +000037#define pr_fmt(fmt) "TCP: " fmt
38
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <net/tcp.h>
40
41#include <linux/compiler.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090042#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
45/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070046int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070047
Jianjun Kong09cb1052008-11-03 00:27:11 -080048/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080049 * interpret the window field as a signed quantity.
50 */
Brian Haleyab32ea52006-09-22 14:15:41 -070051int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080052
Eric Dumazet46d3cea2012-07-11 05:50:31 +000053/* Default TSQ limit of two TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
55
Linus Torvalds1da177e2005-04-16 15:20:36 -070056/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
Brian Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070061
Brian Haleyab32ea52006-09-22 14:15:41 -070062int sysctl_tcp_mtu_probing __read_mostly = 0;
Shan Wei97b1ce252010-12-01 18:04:50 +000063int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
John Heffner5d424d52006-03-20 17:53:41 -080064
David S. Miller35089bb2006-06-13 22:33:04 -070065/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070066int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070067
Eric Dumazet46d3cea2012-07-11 05:50:31 +000068static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
William Allen Simpson519855c2009-12-02 18:14:19 +000070
Andi Kleen67edfef2009-07-21 23:00:40 +000071/* Account for new data that has been sent to the network. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -040072static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070073{
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000074 struct inet_connection_sock *icsk = inet_csk(sk);
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070075 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080076 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070077
David S. Millerfe067e82007-03-07 12:12:44 -080078 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070079 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080080
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080081 tp->packets_out += tcp_skb_pcount(skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +000082 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000083 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
Yuchung Cheng750ea2b2012-05-02 13:30:04 +000084 tcp_rearm_rto(sk);
Eric Dumazet6a5dc9e2013-04-29 08:39:56 +000085 }
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".
137 * This is the first part of cwnd validation mechanism. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400138static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700140 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 s32 delta = tcp_time_stamp - tp->lsndtime;
142 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
143 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;
Cong Wangbcefe172013-06-15 09:39:18 +0800163 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164
David S. Miller35089bb2006-06-13 22:33:04 -0700165 if (sysctl_tcp_slow_start_after_idle &&
166 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700167 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168
169 tp->lsndtime = now;
170
171 /* If it is a reply for ato after last received
172 * packet, enter pingpong mode.
173 */
Cong Wangbcefe172013-06-15 09:39:18 +0800174 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
175 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
176 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177}
178
Andi Kleen67edfef2009-07-21 23:00:40 +0000179/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800180static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700182 tcp_dec_quickack_mode(sk, pkts);
183 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184}
185
Yuchung Cheng85f16522013-06-11 15:35:32 -0700186
187u32 tcp_default_init_rwnd(u32 mss)
188{
189 /* Initial receive window should be twice of TCP_INIT_CWND to
Weiping Pan9ef71e02013-06-18 21:00:31 +0800190 * enable proper sending of new unsent data during fast recovery
Yuchung Cheng85f16522013-06-11 15:35:32 -0700191 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
192 * limit when mss is larger than 1460.
193 */
194 u32 init_rwnd = TCP_INIT_CWND * 2;
195
196 if (mss > 1460)
197 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
198 return init_rwnd;
199}
200
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201/* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
207 */
208void tcp_select_initial_window(int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000210 int wscale_ok, __u8 *rcv_wscale,
211 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212{
213 unsigned int space = (__space < 0 ? 0 : __space);
214
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (65535 << 14);
218 space = min(*window_clamp, space);
219
220 /* Quantize space offering to a multiple of mss if possible. */
221 if (space > mss)
222 space = (space / mss) * mss;
223
224 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 */
Rick Jones15d99e02006-03-20 22:40:29 -0800232 if (sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
234 else
235 (*rcv_wnd) = space;
236
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 (*rcv_wscale) = 0;
238 if (wscale_ok) {
239 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900240 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 */
242 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700243 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244 while (space > 65535 && (*rcv_wscale) < 14) {
245 space >>= 1;
246 (*rcv_wscale)++;
247 }
248 }
249
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800250 if (mss > (1 << *rcv_wscale)) {
Yuchung Cheng85f16522013-06-11 15:35:32 -0700251 if (!init_rcv_wnd) /* Use default unless specified otherwise */
252 init_rcv_wnd = tcp_default_init_rwnd(mss);
253 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 }
255
256 /* Set the clamp no higher than max representable value */
257 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
258}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000259EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260
261/* Chose a new window to advertise, update state in tcp_sock for the
262 * socket, and return result with RFC1323 scaling applied. The return
263 * value can be stuffed directly into th->window for an outgoing
264 * frame.
265 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800266static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267{
268 struct tcp_sock *tp = tcp_sk(sk);
269 u32 cur_win = tcp_receive_window(tp);
270 u32 new_win = __tcp_select_window(sk);
271
272 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800273 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 /* Danger Will Robinson!
275 * Don't update rcv_wup/rcv_wnd here or else
276 * we will not be able to advertise a zero
277 * window in time. --DaveM
278 *
279 * Relax Will Robinson.
280 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700281 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 }
283 tp->rcv_wnd = new_win;
284 tp->rcv_wup = tp->rcv_nxt;
285
286 /* Make sure we do not exceed the maximum possible
287 * scaled window.
288 */
Rick Jones15d99e02006-03-20 22:40:29 -0800289 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 new_win = min(new_win, MAX_TCP_WINDOW);
291 else
292 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
293
294 /* RFC1323 scaling applied */
295 new_win >>= tp->rx_opt.rcv_wscale;
296
297 /* If we advertise zero window, disable fast path. */
298 if (new_win == 0)
299 tp->pred_flags = 0;
300
301 return new_win;
302}
303
Andi Kleen67edfef2009-07-21 23:00:40 +0000304/* Packet ECN state for a SYN-ACK */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400305static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700306{
Eric Dumazet4de075e2011-09-27 13:25:05 -0400307 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800308 if (!(tp->ecn_flags & TCP_ECN_OK))
Eric Dumazet4de075e2011-09-27 13:25:05 -0400309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700310}
311
Andi Kleen67edfef2009-07-21 23:00:40 +0000312/* Packet ECN state for a SYN. */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700313static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
314{
315 struct tcp_sock *tp = tcp_sk(sk);
316
317 tp->ecn_flags = 0;
Hannes Frederic Sowa5d134f12013-01-05 16:10:48 +0000318 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
Eric Dumazet4de075e2011-09-27 13:25:05 -0400319 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700320 tp->ecn_flags = TCP_ECN_OK;
321 }
322}
323
324static __inline__ void
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400325TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700326{
327 if (inet_rsk(req)->ecn_ok)
328 th->ece = 1;
329}
330
Andi Kleen67edfef2009-07-21 23:00:40 +0000331/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
332 * be sent.
333 */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700334static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
335 int tcp_header_len)
336{
337 struct tcp_sock *tp = tcp_sk(sk);
338
339 if (tp->ecn_flags & TCP_ECN_OK) {
340 /* Not-retransmitted data segment: set ECT and inject CWR. */
341 if (skb->len != tcp_header_len &&
342 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
343 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800344 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700345 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
346 tcp_hdr(skb)->cwr = 1;
347 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
348 }
349 } else {
350 /* ACK or retransmitted segment: clear ECT|CE */
351 INET_ECN_dontxmit(sk);
352 }
353 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
354 tcp_hdr(skb)->ece = 1;
355 }
356}
357
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800358/* Constructs common control bits of non-data skb. If SYN/FIN is present,
359 * auto increment end seqno.
360 */
361static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
362{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700363 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800364 skb->csum = 0;
365
Eric Dumazet4de075e2011-09-27 13:25:05 -0400366 TCP_SKB_CB(skb)->tcp_flags = flags;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800367 TCP_SKB_CB(skb)->sacked = 0;
368
369 skb_shinfo(skb)->gso_segs = 1;
370 skb_shinfo(skb)->gso_size = 0;
371 skb_shinfo(skb)->gso_type = 0;
372
373 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000374 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800375 seq++;
376 TCP_SKB_CB(skb)->end_seq = seq;
377}
378
Eric Dumazeta2a385d2012-05-16 23:15:34 +0000379static inline bool tcp_urg_mode(const struct tcp_sock *tp)
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700380{
381 return tp->snd_una != tp->snd_up;
382}
383
Adam Langley33ad7982008-07-19 00:04:31 -0700384#define OPTION_SACK_ADVERTISE (1 << 0)
385#define OPTION_TS (1 << 1)
386#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000387#define OPTION_WSCALE (1 << 3)
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000388#define OPTION_FAST_OPEN_COOKIE (1 << 8)
Adam Langley33ad7982008-07-19 00:04:31 -0700389
390struct tcp_out_options {
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000391 u16 options; /* bit field of OPTION_* */
392 u16 mss; /* 0 to disable */
Adam Langley33ad7982008-07-19 00:04:31 -0700393 u8 ws; /* window scale, 0 to disable */
394 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000395 u8 hash_size; /* bytes in hash_location */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000396 __u8 *hash_location; /* temporary pointer, overloaded */
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000397 __u32 tsval, tsecr; /* need to include OPTION_TS */
398 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
Adam Langley33ad7982008-07-19 00:04:31 -0700399};
400
Andi Kleen67edfef2009-07-21 23:00:40 +0000401/* Write previously computed TCP options to the packet.
402 *
403 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700404 * TCP options, we learned this through the hard way, so be careful here.
405 * Luckily we can at least blame others for their non-compliance but from
406 * inter-operatibility perspective it seems that we're somewhat stuck with
407 * the ordering which we have been using if we want to keep working with
408 * those broken things (not that it currently hurts anybody as there isn't
409 * particular reason why the ordering would need to be changed).
410 *
411 * At least SACK_PERM as the first option is known to lead to a disaster
412 * (but it may well be that other scenarios fail similarly).
413 */
Adam Langley33ad7982008-07-19 00:04:31 -0700414static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000415 struct tcp_out_options *opts)
416{
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000417 u16 options = opts->options; /* mungable copy */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000418
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000419 if (unlikely(OPTION_MD5 & options)) {
Christoph Paasch1a2c6182013-03-17 08:23:34 +0000420 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
421 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000422 /* overload cookie hash location */
423 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700424 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800425 }
Adam Langley33ad7982008-07-19 00:04:31 -0700426
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700427 if (unlikely(opts->mss)) {
428 *ptr++ = htonl((TCPOPT_MSS << 24) |
429 (TCPOLEN_MSS << 16) |
430 opts->mss);
431 }
432
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000433 if (likely(OPTION_TS & options)) {
434 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700435 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
436 (TCPOLEN_SACK_PERM << 16) |
437 (TCPOPT_TIMESTAMP << 8) |
438 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000439 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700440 } else {
441 *ptr++ = htonl((TCPOPT_NOP << 24) |
442 (TCPOPT_NOP << 16) |
443 (TCPOPT_TIMESTAMP << 8) |
444 TCPOLEN_TIMESTAMP);
445 }
446 *ptr++ = htonl(opts->tsval);
447 *ptr++ = htonl(opts->tsecr);
448 }
449
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000450 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700451 *ptr++ = htonl((TCPOPT_NOP << 24) |
452 (TCPOPT_NOP << 16) |
453 (TCPOPT_SACK_PERM << 8) |
454 TCPOLEN_SACK_PERM);
455 }
456
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000457 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700458 *ptr++ = htonl((TCPOPT_NOP << 24) |
459 (TCPOPT_WINDOW << 16) |
460 (TCPOLEN_WINDOW << 8) |
461 opts->ws);
462 }
463
464 if (unlikely(opts->num_sack_blocks)) {
465 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
466 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800467 int this_sack;
468
469 *ptr++ = htonl((TCPOPT_NOP << 24) |
470 (TCPOPT_NOP << 16) |
471 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700472 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800473 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800474
Adam Langley33ad7982008-07-19 00:04:31 -0700475 for (this_sack = 0; this_sack < opts->num_sack_blocks;
476 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800477 *ptr++ = htonl(sp[this_sack].start_seq);
478 *ptr++ = htonl(sp[this_sack].end_seq);
479 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800480
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000481 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800482 }
Yuchung Cheng2100c8d2012-07-19 06:43:05 +0000483
484 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
485 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
486
487 *ptr++ = htonl((TCPOPT_EXP << 24) |
488 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
489 TCPOPT_FASTOPEN_MAGIC);
490
491 memcpy(ptr, foc->val, foc->len);
492 if ((foc->len & 3) == 2) {
493 u8 *align = ((u8 *)ptr) + foc->len;
494 align[0] = align[1] = TCPOPT_NOP;
495 }
496 ptr += (foc->len + 3) >> 2;
497 }
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800498}
499
Andi Kleen67edfef2009-07-21 23:00:40 +0000500/* Compute TCP options for SYN packets. This is not the final
501 * network wire format yet.
502 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000503static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700504 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400505 struct tcp_md5sig_key **md5)
506{
Adam Langley33ad7982008-07-19 00:04:31 -0700507 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000508 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Yuchung Cheng783237e2012-07-19 06:43:07 +0000509 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
Adam Langley33ad7982008-07-19 00:04:31 -0700510
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800511#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700512 *md5 = tp->af_specific->md5_lookup(sk, sk);
513 if (*md5) {
514 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000515 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800516 }
Adam Langley33ad7982008-07-19 00:04:31 -0700517#else
518 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800519#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700520
521 /* We always get an MSS option. The option bytes which will be seen in
522 * normal data packets should timestamps be used, must be in the MSS
523 * advertised. But we subtract them from tp->mss_cache so that
524 * calculations in tcp_sendmsg are simpler etc. So account for this
525 * fact here if necessary. If we don't do this correctly, as a
526 * receiver we won't recognize data packets as being full sized when we
527 * should, and thus we won't abide by the delayed ACK rules correctly.
528 * SACKs don't matter, we never delay an ACK when we have any of those
529 * going out. */
530 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000531 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700532
David S. Millerbb5b7c12009-12-15 20:56:42 -0800533 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700534 opts->options |= OPTION_TS;
Andrey Vaginee684b62013-02-11 05:50:19 +0000535 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
Adam Langley33ad7982008-07-19 00:04:31 -0700536 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000537 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700538 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800539 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700540 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000541 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000542 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700543 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800544 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700545 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700546 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000547 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700548 }
549
Yuchung Cheng783237e2012-07-19 06:43:07 +0000550 if (fastopen && fastopen->cookie.len >= 0) {
551 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
552 need = (need + 3) & ~3U; /* Align to 32 bits */
553 if (remaining >= need) {
554 opts->options |= OPTION_FAST_OPEN_COOKIE;
555 opts->fastopen_cookie = &fastopen->cookie;
556 remaining -= need;
557 tp->syn_fastopen = 1;
558 }
559 }
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000560
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000561 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700562}
563
Andi Kleen67edfef2009-07-21 23:00:40 +0000564/* Set up TCP options for SYN-ACKs. */
Eric Dumazet95c96172012-04-15 05:58:06 +0000565static unsigned int tcp_synack_options(struct sock *sk,
Adam Langley33ad7982008-07-19 00:04:31 -0700566 struct request_sock *req,
Eric Dumazet95c96172012-04-15 05:58:06 +0000567 unsigned int mss, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700568 struct tcp_out_options *opts,
William Allen Simpson4957faade2009-12-02 18:25:27 +0000569 struct tcp_md5sig_key **md5,
Jerry Chu83368862012-08-31 12:29:12 +0000570 struct tcp_fastopen_cookie *foc)
William Allen Simpson4957faade2009-12-02 18:25:27 +0000571{
Adam Langley33ad7982008-07-19 00:04:31 -0700572 struct inet_request_sock *ireq = inet_rsk(req);
Eric Dumazet95c96172012-04-15 05:58:06 +0000573 unsigned int remaining = MAX_TCP_OPTION_SPACE;
Adam Langley33ad7982008-07-19 00:04:31 -0700574
575#ifdef CONFIG_TCP_MD5SIG
576 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
577 if (*md5) {
578 opts->options |= OPTION_MD5;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000579 remaining -= TCPOLEN_MD5SIG_ALIGNED;
580
581 /* We can't fit any SACK blocks in a packet with MD5 + TS
582 * options. There was discussion about disabling SACK
583 * rather than TS in order to fit in better with old,
584 * buggy kernels, but that was deemed to be unnecessary.
585 */
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700586 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700587 }
588#else
589 *md5 = NULL;
590#endif
591
William Allen Simpson4957faade2009-12-02 18:25:27 +0000592 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700593 opts->mss = mss;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000594 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700595
596 if (likely(ireq->wscale_ok)) {
597 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000598 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000599 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700600 }
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700601 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700602 opts->options |= OPTION_TS;
603 opts->tsval = TCP_SKB_CB(skb)->when;
604 opts->tsecr = req->ts_recent;
William Allen Simpson4957faade2009-12-02 18:25:27 +0000605 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700606 }
607 if (likely(ireq->sack_ok)) {
608 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e5e2010-05-17 22:35:36 -0700609 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faade2009-12-02 18:25:27 +0000610 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700611 }
Jerry Chu83368862012-08-31 12:29:12 +0000612 if (foc != NULL) {
613 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
614 need = (need + 3) & ~3U; /* Align to 32 bits */
615 if (remaining >= need) {
616 opts->options |= OPTION_FAST_OPEN_COOKIE;
617 opts->fastopen_cookie = foc;
618 remaining -= need;
619 }
620 }
William Allen Simpson4957faade2009-12-02 18:25:27 +0000621
William Allen Simpson4957faade2009-12-02 18:25:27 +0000622 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700623}
624
Andi Kleen67edfef2009-07-21 23:00:40 +0000625/* Compute TCP options for ESTABLISHED sockets. This is not the
626 * final wire format yet.
627 */
Eric Dumazet95c96172012-04-15 05:58:06 +0000628static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
Adam Langley33ad7982008-07-19 00:04:31 -0700629 struct tcp_out_options *opts,
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400630 struct tcp_md5sig_key **md5)
631{
Adam Langley33ad7982008-07-19 00:04:31 -0700632 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
633 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet95c96172012-04-15 05:58:06 +0000634 unsigned int size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000635 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700636
637#ifdef CONFIG_TCP_MD5SIG
638 *md5 = tp->af_specific->md5_lookup(sk, sk);
639 if (unlikely(*md5)) {
640 opts->options |= OPTION_MD5;
641 size += TCPOLEN_MD5SIG_ALIGNED;
642 }
643#else
644 *md5 = NULL;
645#endif
646
647 if (likely(tp->rx_opt.tstamp_ok)) {
648 opts->options |= OPTION_TS;
Andrey Vaginee684b62013-02-11 05:50:19 +0000649 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700650 opts->tsecr = tp->rx_opt.ts_recent;
651 size += TCPOLEN_TSTAMP_ALIGNED;
652 }
653
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000654 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
655 if (unlikely(eff_sacks)) {
Eric Dumazet95c96172012-04-15 05:58:06 +0000656 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
Adam Langley33ad7982008-07-19 00:04:31 -0700657 opts->num_sack_blocks =
Eric Dumazet95c96172012-04-15 05:58:06 +0000658 min_t(unsigned int, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700659 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
660 TCPOLEN_SACK_PERBLOCK);
661 size += TCPOLEN_SACK_BASE_ALIGNED +
662 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
663 }
664
665 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800666}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000668
669/* TCP SMALL QUEUES (TSQ)
670 *
671 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
672 * to reduce RTT and bufferbloat.
673 * We do this using a special skb destructor (tcp_wfree).
674 *
675 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
676 * needs to be reallocated in a driver.
677 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
678 *
679 * Since transmit from skb destructor is forbidden, we use a tasklet
680 * to process all sockets that eventually need to send more skbs.
681 * We use one tasklet per cpu, with its own queue of sockets.
682 */
683struct tsq_tasklet {
684 struct tasklet_struct tasklet;
685 struct list_head head; /* queue of tcp sockets */
686};
687static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
688
Eric Dumazet6f458df2012-07-20 05:45:50 +0000689static void tcp_tsq_handler(struct sock *sk)
690{
691 if ((1 << sk->sk_state) &
692 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
693 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
694 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
695}
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000696/*
697 * One tasklest per cpu tries to send more skbs.
698 * We run in tasklet context but need to disable irqs when
699 * transfering tsq->head because tcp_wfree() might
700 * interrupt us (non NAPI drivers)
701 */
702static void tcp_tasklet_func(unsigned long data)
703{
704 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
705 LIST_HEAD(list);
706 unsigned long flags;
707 struct list_head *q, *n;
708 struct tcp_sock *tp;
709 struct sock *sk;
710
711 local_irq_save(flags);
712 list_splice_init(&tsq->head, &list);
713 local_irq_restore(flags);
714
715 list_for_each_safe(q, n, &list) {
716 tp = list_entry(q, struct tcp_sock, tsq_node);
717 list_del(&tp->tsq_node);
718
719 sk = (struct sock *)tp;
720 bh_lock_sock(sk);
721
722 if (!sock_owned_by_user(sk)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000723 tcp_tsq_handler(sk);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000724 } else {
725 /* defer the work to tcp_release_cb() */
Eric Dumazet6f458df2012-07-20 05:45:50 +0000726 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000727 }
728 bh_unlock_sock(sk);
729
730 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
731 sk_free(sk);
732 }
733}
734
Eric Dumazet6f458df2012-07-20 05:45:50 +0000735#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
736 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
Eric Dumazet563d34d2012-07-23 09:48:52 +0200737 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
738 (1UL << TCP_MTU_REDUCED_DEFERRED))
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000739/**
740 * tcp_release_cb - tcp release_sock() callback
741 * @sk: socket
742 *
743 * called from release_sock() to perform protocol dependent
744 * actions before socket release.
745 */
746void tcp_release_cb(struct sock *sk)
747{
748 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet6f458df2012-07-20 05:45:50 +0000749 unsigned long flags, nflags;
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000750
Eric Dumazet6f458df2012-07-20 05:45:50 +0000751 /* perform an atomic operation only if at least one flag is set */
752 do {
753 flags = tp->tsq_flags;
754 if (!(flags & TCP_DEFERRED_ALL))
755 return;
756 nflags = flags & ~TCP_DEFERRED_ALL;
757 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
758
759 if (flags & (1UL << TCP_TSQ_DEFERRED))
760 tcp_tsq_handler(sk);
761
Eric Dumazet144d56e2012-08-20 00:22:46 +0000762 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000763 tcp_write_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000764 __sock_put(sk);
765 }
766 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
Eric Dumazet6f458df2012-07-20 05:45:50 +0000767 tcp_delack_timer_handler(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000768 __sock_put(sk);
769 }
770 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
Eric Dumazet563d34d2012-07-23 09:48:52 +0200771 sk->sk_prot->mtu_reduced(sk);
Eric Dumazet144d56e2012-08-20 00:22:46 +0000772 __sock_put(sk);
773 }
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000774}
775EXPORT_SYMBOL(tcp_release_cb);
776
777void __init tcp_tasklet_init(void)
778{
779 int i;
780
781 for_each_possible_cpu(i) {
782 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
783
784 INIT_LIST_HEAD(&tsq->head);
785 tasklet_init(&tsq->tasklet,
786 tcp_tasklet_func,
787 (unsigned long)tsq);
788 }
789}
790
791/*
792 * Write buffer destructor automatically called from kfree_skb.
793 * We cant xmit new skbs from this context, as we might already
794 * hold qdisc lock.
795 */
Eric Dumazetd6a4a102013-04-12 11:31:52 +0000796void tcp_wfree(struct sk_buff *skb)
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000797{
798 struct sock *sk = skb->sk;
799 struct tcp_sock *tp = tcp_sk(sk);
800
801 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
802 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
803 unsigned long flags;
804 struct tsq_tasklet *tsq;
805
806 /* Keep a ref on socket.
807 * This last ref will be released in tcp_tasklet_func()
808 */
809 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
810
811 /* queue this socket to tasklet queue */
812 local_irq_save(flags);
813 tsq = &__get_cpu_var(tsq_tasklet);
814 list_add(&tp->tsq_node, &tsq->head);
815 tasklet_schedule(&tsq->tasklet);
816 local_irq_restore(flags);
817 } else {
818 sock_wfree(skb);
819 }
820}
821
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822/* This routine actually transmits TCP packets queued in by
823 * tcp_do_sendmsg(). This is used by both the initial
824 * transmission and possible later retransmissions.
825 * All SKB's seen here are completely headerless. It is our
826 * job to build the TCP header, and pass the packet down to
827 * IP so it can do the same plus pass the packet off to the
828 * device.
829 *
830 * We are working here with either a clone of the original
831 * SKB, or a fresh unique copy made by the retransmit engine.
832 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800833static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
834 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700835{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800836 const struct inet_connection_sock *icsk = inet_csk(sk);
837 struct inet_sock *inet;
838 struct tcp_sock *tp;
839 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700840 struct tcp_out_options opts;
Eric Dumazet95c96172012-04-15 05:58:06 +0000841 unsigned int tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800842 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800843 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800844 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800846 BUG_ON(!skb || !tcp_skb_pcount(skb));
847
848 /* If congestion control is doing timestamping, we must
849 * take such a timestamp before we potentially clone/copy.
850 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700851 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800852 __net_timestamp(skb);
853
854 if (likely(clone_it)) {
Eric Dumazet0e280af2013-04-18 06:52:51 +0000855 const struct sk_buff *fclone = skb + 1;
856
857 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
858 fclone->fclone == SKB_FCLONE_CLONE))
859 NET_INC_STATS_BH(sock_net(sk),
860 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
861
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800862 if (unlikely(skb_cloned(skb)))
863 skb = pskb_copy(skb, gfp_mask);
864 else
865 skb = skb_clone(skb, gfp_mask);
866 if (unlikely(!skb))
867 return -ENOBUFS;
868 }
869
870 inet = inet_sk(sk);
871 tp = tcp_sk(sk);
872 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700873 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
Eric Dumazet4de075e2011-09-27 13:25:05 -0400875 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700876 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
877 else
878 tcp_options_size = tcp_established_options(sk, skb, &opts,
879 &md5);
880 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900881
Eric Dumazet547669d2013-05-23 07:44:20 +0000882 if (tcp_packets_in_flight(tp) == 0)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800883 tcp_ca_event(sk, CA_EVENT_TX_START);
Eric Dumazet547669d2013-05-23 07:44:20 +0000884
885 /* if no packet is in qdisc/device queue, then allow XPS to select
886 * another queue.
887 */
888 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800889
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700890 skb_push(skb, tcp_header_size);
891 skb_reset_transport_header(skb);
Eric Dumazet46d3cea2012-07-11 05:50:31 +0000892
893 skb_orphan(skb);
894 skb->sk = sk;
895 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
896 tcp_wfree : sock_wfree;
897 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800898
899 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700900 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000901 th->source = inet->inet_sport;
902 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800903 th->seq = htonl(tcb->seq);
904 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700905 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
Eric Dumazet4de075e2011-09-27 13:25:05 -0400906 tcb->tcp_flags);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800907
Eric Dumazet4de075e2011-09-27 13:25:05 -0400908 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800909 /* RFC1323: The window in SYN & SYN/ACK segments
910 * is never scaled.
911 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800912 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800913 } else {
914 th->window = htons(tcp_select_window(sk));
915 }
916 th->check = 0;
917 th->urg_ptr = 0;
918
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700919 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800920 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
921 if (before(tp->snd_up, tcb->seq + 0x10000)) {
922 th->urg_ptr = htons(tp->snd_up - tcb->seq);
923 th->urg = 1;
924 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700925 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800926 th->urg = 1;
927 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800928 }
929
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000930 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Eric Dumazet4de075e2011-09-27 13:25:05 -0400931 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700932 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800933
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800934#ifdef CONFIG_TCP_MD5SIG
935 /* Calculate the MD5 hash, as we have all we need now */
936 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -0700937 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000938 tp->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700939 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800940 }
941#endif
942
Herbert Xubb296242010-04-11 02:15:55 +0000943 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800944
Eric Dumazet4de075e2011-09-27 13:25:05 -0400945 if (likely(tcb->tcp_flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800946 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
947
948 if (skb->len != tcp_header_size)
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400949 tcp_event_data_sent(tp, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800950
Wei Yongjunbd37a082006-08-07 21:04:15 -0700951 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +0000952 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
953 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800954
David S. Millerd9d8da82011-05-06 22:23:20 -0700955 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
Hua Zhong83de47c2006-04-28 15:26:50 -0700956 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800957 return err;
958
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800959 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800960
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200961 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962}
963
Andi Kleen67edfef2009-07-21 23:00:40 +0000964/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 *
966 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
967 * otherwise socket can stall.
968 */
969static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
970{
971 struct tcp_sock *tp = tcp_sk(sk);
972
973 /* Advance write_seq and place onto the write_queue. */
974 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
975 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800976 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800977 sk->sk_wmem_queued += skb->truesize;
978 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979}
980
Andi Kleen67edfef2009-07-21 23:00:40 +0000981/* Initialize TSO segments for a packet. */
Eric Dumazetcf533ea2011-10-21 05:22:42 -0400982static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800983 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700984{
Herbert Xu8e5b9dd2009-06-28 18:03:30 +0000985 if (skb->len <= mss_now || !sk_can_gso(sk) ||
986 skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -0700987 /* Avoid the costly divide in the normal
988 * non-TSO case.
989 */
Herbert Xu79671682006-06-22 02:40:14 -0700990 skb_shinfo(skb)->gso_segs = 1;
991 skb_shinfo(skb)->gso_size = 0;
Pravin B Shelarc9af6db2013-02-11 09:27:41 +0000992 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700993 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700994 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700995 skb_shinfo(skb)->gso_size = mss_now;
Pravin B Shelarc9af6db2013-02-11 09:27:41 +0000996 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 }
998}
999
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001000/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -08001001 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001002 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001003static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001004 int decr)
1005{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -08001006 struct tcp_sock *tp = tcp_sk(sk);
1007
Ilpo Järvinendc869672007-10-01 15:27:19 -07001008 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001009 return;
1010
Ilpo Järvinen6859d492007-12-02 00:48:06 +02001011 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001012 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -07001013}
1014
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001015/* Pcount in the middle of the write queue got changed, we need to do various
1016 * tweaks to fix counters
1017 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001018static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001019{
1020 struct tcp_sock *tp = tcp_sk(sk);
1021
1022 tp->packets_out -= decr;
1023
1024 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1025 tp->sacked_out -= decr;
1026 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1027 tp->retrans_out -= decr;
1028 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1029 tp->lost_out -= decr;
1030
1031 /* Reno case is special. Sigh... */
1032 if (tcp_is_reno(tp) && decr > 0)
1033 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1034
1035 tcp_adjust_fackets_out(sk, skb, decr);
1036
1037 if (tp->lost_skb_hint &&
1038 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +00001039 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001040 tp->lost_cnt_hint -= decr;
1041
1042 tcp_verify_left_out(tp);
1043}
1044
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045/* Function to create two new TCP segments. Shrinks the given segment
1046 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001047 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 * Remember, these are still headerless SKBs at this point.
1049 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001050int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1051 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052{
1053 struct tcp_sock *tp = tcp_sk(sk);
1054 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001055 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001056 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001057 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058
Ilpo Järvinen2fceec12011-04-01 21:47:41 -07001059 if (WARN_ON(len > skb->len))
1060 return -EINVAL;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001061
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 nsize = skb_headlen(skb) - len;
1063 if (nsize < 0)
1064 nsize = 0;
1065
1066 if (skb_cloned(skb) &&
1067 skb_is_nonlinear(skb) &&
1068 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1069 return -ENOMEM;
1070
1071 /* Get a new skb... force flag on. */
1072 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1073 if (buff == NULL)
1074 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001075
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001076 sk->sk_wmem_queued += buff->truesize;
1077 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001078 nlen = skb->len - len - nsize;
1079 buff->truesize += nlen;
1080 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081
1082 /* Correct the sequence numbers. */
1083 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1084 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1085 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1086
1087 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001088 flags = TCP_SKB_CB(skb)->tcp_flags;
1089 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1090 TCP_SKB_CB(buff)->tcp_flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001091 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092
Patrick McHardy84fa7932006-08-29 16:44:56 -07001093 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001095 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1096 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 nsize, 0);
1098
1099 skb_trim(skb, len);
1100
1101 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1102 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001103 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104 skb_split(skb, buff, len);
1105 }
1106
1107 buff->ip_summed = skb->ip_summed;
1108
1109 /* Looks stupid, but our code really uses when of
1110 * skbs, which it never sent before. --ANK
1111 */
1112 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001113 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
David S. Miller6475be12005-09-01 22:47:01 -07001115 old_factor = tcp_skb_pcount(skb);
1116
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001118 tcp_set_skb_tso_segs(sk, skb, mss_now);
1119 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
David S. Miller6475be12005-09-01 22:47:01 -07001121 /* If this packet has been sent out already, we must
1122 * adjust the various packet counters.
1123 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001124 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001125 int diff = old_factor - tcp_skb_pcount(skb) -
1126 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001128 if (diff)
1129 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 }
1131
1132 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -07001133 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001134 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135
1136 return 0;
1137}
1138
1139/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1140 * eventually). The difference is that pulled data not copied, but
1141 * immediately discarded.
1142 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001143static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144{
1145 int i, k, eat;
1146
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001147 eat = min_t(int, len, skb_headlen(skb));
1148 if (eat) {
1149 __skb_pull(skb, eat);
1150 len -= eat;
1151 if (!len)
1152 return;
1153 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 eat = len;
1155 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001156 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Eric Dumazet9e903e02011-10-18 21:00:24 +00001157 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1158
1159 if (size <= eat) {
Ian Campbellaff65da2011-08-22 23:44:59 +00001160 skb_frag_unref(skb, i);
Eric Dumazet9e903e02011-10-18 21:00:24 +00001161 eat -= size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 } else {
1163 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1164 if (eat) {
1165 skb_shinfo(skb)->frags[k].page_offset += eat;
Eric Dumazet9e903e02011-10-18 21:00:24 +00001166 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 eat = 0;
1168 }
1169 k++;
1170 }
1171 }
1172 skb_shinfo(skb)->nr_frags = k;
1173
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001174 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001175 skb->data_len -= len;
1176 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177}
1178
Andi Kleen67edfef2009-07-21 23:00:40 +00001179/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1181{
Pravin B Shelar14bbd6a2013-02-14 09:44:49 +00001182 if (skb_unclone(skb, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 return -ENOMEM;
1184
Eric Dumazet4fa48bf2011-12-04 08:51:08 +00001185 __pskb_trim_head(skb, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186
1187 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001188 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189
1190 skb->truesize -= len;
1191 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001192 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1194
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001195 /* Any change of skb->len requires recalculation of tso factor. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 if (tcp_skb_pcount(skb) > 1)
Neal Cardwell5b35e1e2012-01-28 17:29:46 +00001197 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198
1199 return 0;
1200}
1201
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001202/* Calculate MSS not accounting any TCP options. */
1203static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
John Heffner5d424d52006-03-20 17:53:41 -08001204{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001205 const struct tcp_sock *tp = tcp_sk(sk);
1206 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001207 int mss_now;
1208
1209 /* Calculate base mss without TCP options:
1210 It is MMS_S - sizeof(tcphdr) of rfc1122
1211 */
1212 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1213
Eric Dumazet67469602012-04-24 07:37:38 +00001214 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1215 if (icsk->icsk_af_ops->net_frag_header_len) {
1216 const struct dst_entry *dst = __sk_dst_get(sk);
1217
1218 if (dst && dst_allfrag(dst))
1219 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1220 }
1221
John Heffner5d424d52006-03-20 17:53:41 -08001222 /* Clamp it (mss_clamp does not include tcp options) */
1223 if (mss_now > tp->rx_opt.mss_clamp)
1224 mss_now = tp->rx_opt.mss_clamp;
1225
1226 /* Now subtract optional transport overhead */
1227 mss_now -= icsk->icsk_ext_hdr_len;
1228
1229 /* Then reserve room for full set of TCP options and 8 bytes of data */
1230 if (mss_now < 48)
1231 mss_now = 48;
John Heffner5d424d52006-03-20 17:53:41 -08001232 return mss_now;
1233}
1234
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00001235/* Calculate MSS. Not accounting for SACKs here. */
1236int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1237{
1238 /* Subtract TCP options size, not including SACKs */
1239 return __tcp_mtu_to_mss(sk, pmtu) -
1240 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1241}
1242
John Heffner5d424d52006-03-20 17:53:41 -08001243/* Inverse of above */
Eric Dumazet67469602012-04-24 07:37:38 +00001244int tcp_mss_to_mtu(struct sock *sk, int mss)
John Heffner5d424d52006-03-20 17:53:41 -08001245{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001246 const struct tcp_sock *tp = tcp_sk(sk);
1247 const struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001248 int mtu;
1249
1250 mtu = mss +
1251 tp->tcp_header_len +
1252 icsk->icsk_ext_hdr_len +
1253 icsk->icsk_af_ops->net_header_len;
1254
Eric Dumazet67469602012-04-24 07:37:38 +00001255 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1256 if (icsk->icsk_af_ops->net_frag_header_len) {
1257 const struct dst_entry *dst = __sk_dst_get(sk);
1258
1259 if (dst && dst_allfrag(dst))
1260 mtu += icsk->icsk_af_ops->net_frag_header_len;
1261 }
John Heffner5d424d52006-03-20 17:53:41 -08001262 return mtu;
1263}
1264
Andi Kleen67edfef2009-07-21 23:00:40 +00001265/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001266void tcp_mtup_init(struct sock *sk)
1267{
1268 struct tcp_sock *tp = tcp_sk(sk);
1269 struct inet_connection_sock *icsk = inet_csk(sk);
1270
1271 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1272 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001273 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -08001274 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1275 icsk->icsk_mtup.probe_size = 0;
1276}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001277EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001278
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279/* This function synchronize snd mss to current pmtu/exthdr set.
1280
1281 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1282 for TCP options, but includes only bare TCP header.
1283
1284 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001285 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 It also does not include TCP options.
1287
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001288 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289
1290 tp->mss_cache is current effective sending mss, including
1291 all tcp options except for SACKs. It is evaluated,
1292 taking into account current pmtu, but never exceeds
1293 tp->rx_opt.mss_clamp.
1294
1295 NOTE1. rfc1122 clearly states that advertised MSS
1296 DOES NOT include either tcp or ip options.
1297
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001298 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1299 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1302{
1303 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001304 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001305 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306
John Heffner5d424d52006-03-20 17:53:41 -08001307 if (icsk->icsk_mtup.search_high > pmtu)
1308 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309
John Heffner5d424d52006-03-20 17:53:41 -08001310 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001311 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312
1313 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001314 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001315 if (icsk->icsk_mtup.enabled)
1316 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001317 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318
1319 return mss_now;
1320}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001321EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
1323/* Compute the current effective MSS, taking SACKs and IP options,
1324 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001326unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001328 const struct tcp_sock *tp = tcp_sk(sk);
1329 const struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001330 u32 mss_now;
Eric Dumazet95c96172012-04-15 05:58:06 +00001331 unsigned int header_len;
Adam Langley33ad7982008-07-19 00:04:31 -07001332 struct tcp_out_options opts;
1333 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001335 mss_now = tp->mss_cache;
1336
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 if (dst) {
1338 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001339 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 mss_now = tcp_sync_mss(sk, mtu);
1341 }
1342
Adam Langley33ad7982008-07-19 00:04:31 -07001343 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1344 sizeof(struct tcphdr);
1345 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1346 * some common options. If this is an odd packet (because we have SACK
1347 * blocks etc) then our calculated header_len will be different, and
1348 * we have to adjust mss_now correspondingly */
1349 if (header_len != tp->tcp_header_len) {
1350 int delta = (int) header_len - tp->tcp_header_len;
1351 mss_now -= delta;
1352 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001353
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 return mss_now;
1355}
1356
David S. Millera762a982005-07-05 15:18:51 -07001357/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001358static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001359{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001360 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001361
Ilpo Järvinend436d682007-12-31 14:58:00 -08001362 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001363 /* Network is feed fully. */
1364 tp->snd_cwnd_used = 0;
1365 tp->snd_cwnd_stamp = tcp_time_stamp;
1366 } else {
1367 /* Network starves. */
1368 if (tp->packets_out > tp->snd_cwnd_used)
1369 tp->snd_cwnd_used = tp->packets_out;
1370
David S. Miller15d33c02007-04-09 13:23:14 -07001371 if (sysctl_tcp_slow_start_after_idle &&
1372 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001373 tcp_cwnd_application_limited(sk);
1374 }
1375}
1376
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001377/* Returns the portion of skb which can be sent right away without
1378 * introducing MSS oddities to segment boundaries. In rare cases where
1379 * mss_now != mss_cache, we will request caller to create a small skb
1380 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001381 *
1382 * We explicitly want to create a request for splitting write queue tail
1383 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1384 * thus all the complexity (cwnd_len is always MSS multiple which we
1385 * return whenever allowed by the other factors). Basically we need the
1386 * modulo only when the receiver window alone is the limiting factor or
1387 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001388 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001389static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
Ben Hutchings14853482012-07-30 16:11:42 +00001390 unsigned int mss_now, unsigned int max_segs)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001391{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001392 const struct tcp_sock *tp = tcp_sk(sk);
Ben Hutchings14853482012-07-30 16:11:42 +00001393 u32 needed, window, max_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001394
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001395 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Ben Hutchings14853482012-07-30 16:11:42 +00001396 max_len = mss_now * max_segs;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001397
Ben Hutchings14853482012-07-30 16:11:42 +00001398 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1399 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001400
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001401 needed = min(skb->len, window);
1402
Ben Hutchings14853482012-07-30 16:11:42 +00001403 if (max_len <= needed)
1404 return max_len;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001405
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001406 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001407}
1408
1409/* Can at least one segment of SKB be sent right now, according to the
1410 * congestion window rules? If so, return how many segments are allowed.
1411 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001412static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1413 const struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001414{
1415 u32 in_flight, cwnd;
1416
1417 /* Don't be strict about the congestion window for the final FIN. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001418 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1419 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001420 return 1;
1421
1422 in_flight = tcp_packets_in_flight(tp);
1423 cwnd = tp->snd_cwnd;
1424 if (in_flight < cwnd)
1425 return (cwnd - in_flight);
1426
1427 return 0;
1428}
1429
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001430/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001431 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001432 * SKB onto the wire.
1433 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001434static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001435 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001436{
1437 int tso_segs = tcp_skb_pcount(skb);
1438
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001439 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001440 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001441 tso_segs = tcp_skb_pcount(skb);
1442 }
1443 return tso_segs;
1444}
1445
Andi Kleen67edfef2009-07-21 23:00:40 +00001446/* Minshall's variant of the Nagle send check. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001447static inline bool tcp_minshall_check(const struct tcp_sock *tp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001448{
Jianjun Kong09cb1052008-11-03 00:27:11 -08001449 return after(tp->snd_sml, tp->snd_una) &&
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001450 !after(tp->snd_sml, tp->snd_nxt);
1451}
1452
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001453/* Return false, if packet can be sent now without violation Nagle's rules:
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001454 * 1. It is full sized.
1455 * 2. Or it contains FIN. (already checked by caller)
Feng King6d67e9b2011-11-05 04:23:23 +00001456 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001457 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1458 * With Minshall's modification: all sent small packets are ACKed.
1459 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001460static inline bool tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001461 const struct sk_buff *skb,
Eric Dumazet95c96172012-04-15 05:58:06 +00001462 unsigned int mss_now, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001463{
Eric Dumazeta02cec22010-09-22 20:43:57 +00001464 return skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001465 ((nonagle & TCP_NAGLE_CORK) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00001466 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001467}
1468
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001469/* Return true if the Nagle test allows this packet to be
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001470 * sent now.
1471 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001472static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1473 unsigned int cur_mss, int nonagle)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001474{
1475 /* Nagle rule does not apply to frames, which sit in the middle of the
1476 * write_queue (they have no chances to get new data).
1477 *
1478 * This is implemented in the callers, where they modify the 'nonagle'
1479 * argument based upon the location of SKB in the send queue.
1480 */
1481 if (nonagle & TCP_NAGLE_PUSH)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001482 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001483
Yuchung Cheng9b441902013-03-20 13:32:58 +00001484 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1485 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001486 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001487
1488 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001489 return true;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001490
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001491 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001492}
1493
1494/* Does at least the first segment of SKB fit into the send window? */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001495static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1496 const struct sk_buff *skb,
1497 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001498{
1499 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1500
1501 if (skb->len > cur_mss)
1502 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1503
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001504 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001505}
1506
David S. Millerfe067e82007-03-07 12:12:44 -08001507/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001508 * should be put on the wire right now. If so, it returns the number of
1509 * packets allowed by the congestion window.
1510 */
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001511static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001512 unsigned int cur_mss, int nonagle)
1513{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001514 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001515 unsigned int cwnd_quota;
1516
David S. Miller846998a2005-08-04 19:52:01 -07001517 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001518
1519 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1520 return 0;
1521
1522 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001523 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001524 cwnd_quota = 0;
1525
1526 return cwnd_quota;
1527}
1528
Andi Kleen67edfef2009-07-21 23:00:40 +00001529/* Test if sending is allowed right now. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001530bool tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001531{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04001532 const struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001533 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001534
Eric Dumazeta02cec22010-09-22 20:43:57 +00001535 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001536 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001537 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001538 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001539}
1540
1541/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1542 * which is put after SKB on the list. It is very much like
1543 * tcp_fragment() except that it may make several kinds of assumptions
1544 * in order to speed up the splitting operation. In particular, we
1545 * know that all the data is in scatter-gather pages, and that the
1546 * packet has never been sent out before (and thus is not cloned).
1547 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001548static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001549 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001550{
1551 struct sk_buff *buff;
1552 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001553 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001554
1555 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001556 if (skb->len != skb->data_len)
1557 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001558
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001559 buff = sk_stream_alloc_skb(sk, 0, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001560 if (unlikely(buff == NULL))
1561 return -ENOMEM;
1562
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001563 sk->sk_wmem_queued += buff->truesize;
1564 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001565 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001566 skb->truesize -= nlen;
1567
1568 /* Correct the sequence numbers. */
1569 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1570 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1571 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1572
1573 /* PSH and FIN should only be set in the second packet. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001574 flags = TCP_SKB_CB(skb)->tcp_flags;
1575 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1576 TCP_SKB_CB(buff)->tcp_flags = flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001577
1578 /* This packet was never sent out yet, so no SACK bits. */
1579 TCP_SKB_CB(buff)->sacked = 0;
1580
Patrick McHardy84fa7932006-08-29 16:44:56 -07001581 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001582 skb_split(skb, buff, len);
1583
1584 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001585 tcp_set_skb_tso_segs(sk, skb, mss_now);
1586 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001587
1588 /* Link BUFF into the send queue. */
1589 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001590 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001591
1592 return 0;
1593}
1594
1595/* Try to defer sending, if possible, in order to minimize the amount
1596 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1597 *
1598 * This algorithm is from John Heffner.
1599 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001600static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001601{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001602 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001603 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001604 u32 send_win, cong_win, limit, in_flight;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001605 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001606
Eric Dumazet4de075e2011-09-27 13:25:05 -04001607 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001608 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001609
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001610 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001611 goto send_now;
1612
1613 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001614 if (tp->tso_deferred &&
Ilpo Järvinena2acde02008-12-05 22:49:37 -08001615 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001616 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001617
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001618 in_flight = tcp_packets_in_flight(tp);
1619
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001620 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001621
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001622 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001623
1624 /* From in_flight test above, we know that cwnd > in_flight. */
1625 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1626
1627 limit = min(send_win, cong_win);
1628
David S. Millerba244fe2006-03-11 18:51:49 -08001629 /* If a full-sized TSO skb can be sent, do it. */
Ben Hutchings14853482012-07-30 16:11:42 +00001630 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1631 sk->sk_gso_max_segs * tp->mss_cache))
John Heffnerae8064a2006-10-18 20:36:48 -07001632 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001633
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001634 /* Middle in queue won't get any more data, full sendable already? */
1635 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1636 goto send_now;
1637
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001638 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1639 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001640 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1641
1642 /* If at least some fraction of a window is available,
1643 * just use it.
1644 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001645 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001646 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001647 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001648 } else {
1649 /* Different approach, try not to defer past a single
1650 * ACK. Receiver should ACK every other full sized
1651 * frame, so if we have space for more than 3 frames
1652 * then send now.
1653 */
Neal Cardwell6b5a5c02011-11-21 17:15:14 +00001654 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001655 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001656 }
1657
Eric Dumazetf4541d62013-03-21 17:36:09 +00001658 /* Ok, it looks like it is advisable to defer.
1659 * Do not rearm the timer if already set to not break TCP ACK clocking.
1660 */
1661 if (!tp->tso_deferred)
1662 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001663
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001664 return true;
John Heffnerae8064a2006-10-18 20:36:48 -07001665
1666send_now:
1667 tp->tso_deferred = 0;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001668 return false;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001669}
1670
John Heffner5d424d52006-03-20 17:53:41 -08001671/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001672 * MTU probe is regularly attempting to increase the path MTU by
1673 * deliberately sending larger packets. This discovers routing
1674 * changes resulting in larger path MTUs.
1675 *
John Heffner5d424d52006-03-20 17:53:41 -08001676 * Returns 0 if we should wait to probe (no cwnd available),
1677 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001678 * -1 otherwise
1679 */
John Heffner5d424d52006-03-20 17:53:41 -08001680static int tcp_mtu_probe(struct sock *sk)
1681{
1682 struct tcp_sock *tp = tcp_sk(sk);
1683 struct inet_connection_sock *icsk = inet_csk(sk);
1684 struct sk_buff *skb, *nskb, *next;
1685 int len;
1686 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001687 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001688 int copy;
1689 int mss_now;
1690
1691 /* Not currently probing/verifying,
1692 * not in recovery,
1693 * have enough cwnd, and
1694 * not SACKing (the variable headers throw things off) */
1695 if (!icsk->icsk_mtup.enabled ||
1696 icsk->icsk_mtup.probe_size ||
1697 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1698 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001699 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001700 return -1;
1701
1702 /* Very simple search strategy: just double the MSS. */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001703 mss_now = tcp_current_mss(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001704 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001705 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001706 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1707 /* TODO: set timer for probe_converge_event */
1708 return -1;
1709 }
1710
1711 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001712 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001713 return -1;
1714
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001715 if (tp->snd_wnd < size_needed)
1716 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001717 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001718 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001719
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001720 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1721 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1722 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001723 return -1;
1724 else
1725 return 0;
1726 }
1727
1728 /* We're allowed to probe. Build it now. */
1729 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1730 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001731 sk->sk_wmem_queued += nskb->truesize;
1732 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001733
David S. Millerfe067e82007-03-07 12:12:44 -08001734 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001735
1736 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1737 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Eric Dumazet4de075e2011-09-27 13:25:05 -04001738 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001739 TCP_SKB_CB(nskb)->sacked = 0;
1740 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001741 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001742
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001743 tcp_insert_write_queue_before(nskb, skb, sk);
1744
John Heffner5d424d52006-03-20 17:53:41 -08001745 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001746 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001747 copy = min_t(int, skb->len, probe_size - len);
1748 if (nskb->ip_summed)
1749 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1750 else
1751 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001752 skb_put(nskb, copy),
1753 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001754
1755 if (skb->len <= copy) {
1756 /* We've eaten all the data from this skb.
1757 * Throw it away. */
Eric Dumazet4de075e2011-09-27 13:25:05 -04001758 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001759 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001760 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001761 } else {
Eric Dumazet4de075e2011-09-27 13:25:05 -04001762 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001763 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001764 if (!skb_shinfo(skb)->nr_frags) {
1765 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001766 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001767 skb->csum = csum_partial(skb->data,
1768 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001769 } else {
1770 __pskb_trim_head(skb, copy);
1771 tcp_set_skb_tso_segs(sk, skb, mss_now);
1772 }
1773 TCP_SKB_CB(skb)->seq += copy;
1774 }
1775
1776 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001777
1778 if (len >= probe_size)
1779 break;
John Heffner5d424d52006-03-20 17:53:41 -08001780 }
1781 tcp_init_tso_segs(sk, nskb, nskb->len);
1782
1783 /* We're ready to send. If this fails, the probe will
1784 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1785 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1786 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1787 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001788 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001789 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001790 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001791
1792 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001793 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1794 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001795
1796 return 1;
1797 }
1798
1799 return -1;
1800}
1801
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802/* This routine writes packets to the network. It advances the
1803 * send_head. This happens as incoming acks open up the remote
1804 * window for us.
1805 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001806 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1807 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1808 * account rare use of URG, this is not a big flaw.
1809 *
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001810 * Send at most one packet when push_one > 0. Temporarily ignore
1811 * cwnd limit to force at most one packet out when push_one == 2.
1812
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001813 * Returns true, if no segments are in flight and we have queued segments,
1814 * but cannot send anything now because of SWS or another problem.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001816static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1817 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818{
1819 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001820 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001821 unsigned int tso_segs, sent_pkts;
1822 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001823 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824
David S. Miller92df7b52005-07-05 15:19:06 -07001825 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001826
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001827 if (!push_one) {
1828 /* Do MTU probing. */
1829 result = tcp_mtu_probe(sk);
1830 if (!result) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001831 return false;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001832 } else if (result > 0) {
1833 sent_pkts = 1;
1834 }
John Heffner5d424d52006-03-20 17:53:41 -08001835 }
1836
David S. Millerfe067e82007-03-07 12:12:44 -08001837 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001838 unsigned int limit;
1839
Eric Dumazet46d3cea2012-07-11 05:50:31 +00001840
Herbert Xub68e9f82005-08-04 19:52:02 -07001841 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001842 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001843
Andrew Vaginec342322012-11-15 04:03:17 +00001844 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1845 goto repair; /* Skip network transmission */
1846
Herbert Xub68e9f82005-08-04 19:52:02 -07001847 cwnd_quota = tcp_cwnd_test(tp, skb);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001848 if (!cwnd_quota) {
1849 if (push_one == 2)
1850 /* Force out a loss probe pkt. */
1851 cwnd_quota = 1;
1852 else
1853 break;
1854 }
Herbert Xub68e9f82005-08-04 19:52:02 -07001855
1856 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1857 break;
1858
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001859 if (tso_segs == 1) {
1860 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1861 (tcp_skb_is_last(sk, skb) ?
1862 nonagle : TCP_NAGLE_PUSH))))
1863 break;
1864 } else {
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001865 if (!push_one && tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001866 break;
1867 }
David S. Milleraa934662005-07-05 15:20:09 -07001868
Eric Dumazet46d3cea2012-07-11 05:50:31 +00001869 /* TSQ : sk_wmem_alloc accounts skb truesize,
1870 * including skb overhead. But thats OK.
1871 */
1872 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1873 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1874 break;
1875 }
Herbert Xuc8ac3772005-08-16 20:43:40 -07001876 limit = mss_now;
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001877 if (tso_segs > 1 && !tcp_urg_mode(tp))
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001878 limit = tcp_mss_split_point(sk, skb, mss_now,
Ben Hutchings14853482012-07-30 16:11:42 +00001879 min_t(unsigned int,
1880 cwnd_quota,
1881 sk->sk_gso_max_segs));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882
Herbert Xuc8ac3772005-08-16 20:43:40 -07001883 if (skb->len > limit &&
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001884 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
Herbert Xuc8ac3772005-08-16 20:43:40 -07001885 break;
1886
David S. Miller92df7b52005-07-05 15:19:06 -07001887 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001888
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001889 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07001890 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891
Andrew Vaginec342322012-11-15 04:03:17 +00001892repair:
David S. Miller92df7b52005-07-05 15:19:06 -07001893 /* Advance the send_head. This one is sent out.
1894 * This call will increment packets_out.
1895 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001896 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001897
1898 tcp_minshall_update(tp, mss_now, skb);
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00001899 sent_pkts += tcp_skb_pcount(skb);
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001900
1901 if (push_one)
1902 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 }
David S. Miller92df7b52005-07-05 15:19:06 -07001904
David S. Milleraa934662005-07-05 15:20:09 -07001905 if (likely(sent_pkts)) {
Yuchung Cheng684bad12012-09-02 17:38:04 +00001906 if (tcp_in_cwnd_reduction(sk))
1907 tp->prr_out += sent_pkts;
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001908
1909 /* Send one loss probe per tail loss episode. */
1910 if (push_one != 2)
1911 tcp_schedule_loss_probe(sk);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001912 tcp_cwnd_validate(sk);
Eric Dumazeta2a385d2012-05-16 23:15:34 +00001913 return false;
David S. Miller92df7b52005-07-05 15:19:06 -07001914 }
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001915 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1916}
1917
1918bool tcp_schedule_loss_probe(struct sock *sk)
1919{
1920 struct inet_connection_sock *icsk = inet_csk(sk);
1921 struct tcp_sock *tp = tcp_sk(sk);
1922 u32 timeout, tlp_time_stamp, rto_time_stamp;
1923 u32 rtt = tp->srtt >> 3;
1924
1925 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1926 return false;
1927 /* No consecutive loss probes. */
1928 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1929 tcp_rearm_rto(sk);
1930 return false;
1931 }
1932 /* Don't do any loss probe on a Fast Open connection before 3WHS
1933 * finishes.
1934 */
1935 if (sk->sk_state == TCP_SYN_RECV)
1936 return false;
1937
1938 /* TLP is only scheduled when next timer event is RTO. */
1939 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1940 return false;
1941
1942 /* Schedule a loss probe in 2*RTT for SACK capable connections
1943 * in Open state, that are either limited by cwnd or application.
1944 */
1945 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1946 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1947 return false;
1948
1949 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1950 tcp_send_head(sk))
1951 return false;
1952
1953 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1954 * for delayed ack when there's one outstanding packet.
1955 */
1956 timeout = rtt << 1;
1957 if (tp->packets_out == 1)
1958 timeout = max_t(u32, timeout,
1959 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1960 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1961
1962 /* If RTO is shorter, just schedule TLP in its place. */
1963 tlp_time_stamp = tcp_time_stamp + timeout;
1964 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1965 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1966 s32 delta = rto_time_stamp - tcp_time_stamp;
1967 if (delta > 0)
1968 timeout = delta;
1969 }
1970
1971 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1972 TCP_RTO_MAX);
1973 return true;
1974}
1975
1976/* When probe timeout (PTO) fires, send a new segment if one exists, else
1977 * retransmit the last segment.
1978 */
1979void tcp_send_loss_probe(struct sock *sk)
1980{
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00001981 struct tcp_sock *tp = tcp_sk(sk);
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001982 struct sk_buff *skb;
1983 int pcount;
1984 int mss = tcp_current_mss(sk);
1985 int err = -1;
1986
1987 if (tcp_send_head(sk) != NULL) {
1988 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1989 goto rearm_timer;
1990 }
1991
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00001992 /* At most one outstanding TLP retransmission. */
1993 if (tp->tlp_high_seq)
1994 goto rearm_timer;
1995
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00001996 /* Retransmit last segment. */
1997 skb = tcp_write_queue_tail(sk);
1998 if (WARN_ON(!skb))
1999 goto rearm_timer;
2000
2001 pcount = tcp_skb_pcount(skb);
2002 if (WARN_ON(!pcount))
2003 goto rearm_timer;
2004
2005 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2006 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2007 goto rearm_timer;
2008 skb = tcp_write_queue_tail(sk);
2009 }
2010
2011 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2012 goto rearm_timer;
2013
2014 /* Probe with zero data doesn't trigger fast recovery. */
2015 if (skb->len > 0)
2016 err = __tcp_retransmit_skb(sk, skb);
2017
Nandita Dukkipati9b717a82013-03-11 10:00:44 +00002018 /* Record snd_nxt for loss detection. */
2019 if (likely(!err))
2020 tp->tlp_high_seq = tp->snd_nxt;
2021
Nandita Dukkipati6ba8a3b2013-03-11 10:00:43 +00002022rearm_timer:
2023 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2024 inet_csk(sk)->icsk_rto,
2025 TCP_RTO_MAX);
2026
2027 if (likely(!err))
2028 NET_INC_STATS_BH(sock_net(sk),
2029 LINUX_MIB_TCPLOSSPROBES);
2030 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031}
2032
David S. Millera762a982005-07-05 15:18:51 -07002033/* Push out any pending frames which were held back due to
2034 * TCP_CORK or attempt at coalescing tiny packets.
2035 * The socket must be locked by the caller.
2036 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002037void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2038 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07002039{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002040 /* If we are closed, the bytes will have to remain here.
2041 * In time closedown will finish, we empty the write queue and
2042 * all will be happy.
2043 */
2044 if (unlikely(sk->sk_state == TCP_CLOSE))
2045 return;
2046
Mel Gorman99a1dec2012-07-31 16:44:14 -07002047 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2048 sk_gfp_atomic(sk, GFP_ATOMIC)))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08002049 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07002050}
2051
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002052/* Send _single_ skb sitting at the send head. This function requires
2053 * true push pending frames to setup probe timer etc.
2054 */
2055void tcp_push_one(struct sock *sk, unsigned int mss_now)
2056{
David S. Millerfe067e82007-03-07 12:12:44 -08002057 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002058
2059 BUG_ON(!skb || skb->len < mss_now);
2060
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08002061 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07002062}
2063
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064/* This function returns the amount that we can raise the
2065 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002066 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 * 1. The window can never be shrunk once it is offered (RFC 793)
2068 * 2. We limit memory per socket
2069 *
2070 * RFC 1122:
2071 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2072 * RECV.NEXT + RCV.WIN fixed until:
2073 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2074 *
2075 * i.e. don't raise the right edge of the window until you can raise
2076 * it at least MSS bytes.
2077 *
2078 * Unfortunately, the recommended algorithm breaks header prediction,
2079 * since header prediction assumes th->window stays fixed.
2080 *
2081 * Strictly speaking, keeping th->window fixed violates the receiver
2082 * side SWS prevention criteria. The problem is that under this rule
2083 * a stream of single byte packets will cause the right side of the
2084 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002085 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 * Of course, if the sender implements sender side SWS prevention
2087 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002088 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002090 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 * If the free space is less than the 1/4 of the maximum
2092 * space available and the free space is less than 1/2 mss,
2093 * then set the window to 0.
2094 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2095 * Otherwise, just prevent the window from shrinking
2096 * and from being larger than the largest representable value.
2097 *
2098 * This prevents incremental opening of the window in the regime
2099 * where TCP is limited by the speed of the reader side taking
2100 * data out of the TCP receive queue. It does nothing about
2101 * those cases where the window is constrained on the sender side
2102 * because the pipeline is full.
2103 *
2104 * BSD also seems to "accidentally" limit itself to windows that are a
2105 * multiple of MSS, at least until the free space gets quite small.
2106 * This would appear to be a side effect of the mbuf implementation.
2107 * Combining these two algorithms results in the observed behavior
2108 * of having a fixed window size at almost all times.
2109 *
2110 * Below we obtain similar behavior by forcing the offered window to
2111 * a multiple of the mss when it is feasible to do so.
2112 *
2113 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2114 * Regular options like TIMESTAMP are taken into account.
2115 */
2116u32 __tcp_select_window(struct sock *sk)
2117{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002118 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002120 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 * here. I don't know if the value based on our guesses
2122 * of peer's MSS is better for the performance. It's more correct
2123 * but may be worse for the performance because of rcv_mss
2124 * fluctuations. --SAW 1998/11/1
2125 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002126 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 int free_space = tcp_space(sk);
2128 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2129 int window;
2130
2131 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002132 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002134 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002135 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136
Glauber Costa180d8cd2011-12-11 21:47:02 +00002137 if (sk_under_memory_pressure(sk))
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002138 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2139 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140
2141 if (free_space < mss)
2142 return 0;
2143 }
2144
2145 if (free_space > tp->rcv_ssthresh)
2146 free_space = tp->rcv_ssthresh;
2147
2148 /* Don't do rounding if we are using window scaling, since the
2149 * scaled window will not line up with the MSS boundary anyway.
2150 */
2151 window = tp->rcv_wnd;
2152 if (tp->rx_opt.rcv_wscale) {
2153 window = free_space;
2154
2155 /* Advertise enough space so that it won't get scaled away.
2156 * Import case: prevent zero window announcement if
2157 * 1<<rcv_wscale > mss.
2158 */
2159 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2160 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2161 << tp->rx_opt.rcv_wscale);
2162 } else {
2163 /* Get the largest window that is a nice multiple of mss.
2164 * Window clamp already applied above.
2165 * If our current window offering is within 1 mss of the
2166 * free space we just keep it. This prevents the divide
2167 * and multiply from happening most of the time.
2168 * We also don't do any window rounding when the free space
2169 * is too small.
2170 */
2171 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002172 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07002173 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08002174 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07002175 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176 }
2177
2178 return window;
2179}
2180
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002181/* Collapses two adjacent SKB's during retransmission. */
2182static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183{
2184 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002185 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002186 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002188 skb_size = skb->len;
2189 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002191 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07002192
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002193 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002195 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002197 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2198 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002200 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2201 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002203 if (skb->ip_summed != CHECKSUM_PARTIAL)
2204 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002206 /* Update sequence range on original skb. */
2207 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00002209 /* Merge over control information. This moves PSH/FIN etc. over */
Eric Dumazet4de075e2011-09-27 13:25:05 -04002210 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002212 /* All done, get rid of second SKB and account for it so
2213 * packet counting does not break.
2214 */
2215 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07002216
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002217 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07002218 tcp_clear_retrans_hints_partial(tp);
2219 if (next_skb == tp->retransmit_skb_hint)
2220 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002221
Ilpo Järvinen797108d2009-04-01 23:15:17 +00002222 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2223
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002224 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225}
2226
Andi Kleen67edfef2009-07-21 23:00:40 +00002227/* Check if coalescing SKBs is legal. */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002228static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002229{
2230 if (tcp_skb_pcount(skb) > 1)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002231 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002232 /* TODO: SACK collapsing could be used to remove this condition */
2233 if (skb_shinfo(skb)->nr_frags != 0)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002234 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002235 if (skb_cloned(skb))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002236 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002237 if (skb == tcp_send_head(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002238 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002239 /* Some heurestics for collapsing over SACK'd could be invented */
2240 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002241 return false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002242
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002243 return true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002244}
2245
Andi Kleen67edfef2009-07-21 23:00:40 +00002246/* Collapse packets in the retransmit queue to make to create
2247 * less packets on the wire. This is only done on retransmission.
2248 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002249static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2250 int space)
2251{
2252 struct tcp_sock *tp = tcp_sk(sk);
2253 struct sk_buff *skb = to, *tmp;
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002254 bool first = true;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002255
2256 if (!sysctl_tcp_retrans_collapse)
2257 return;
Eric Dumazet4de075e2011-09-27 13:25:05 -04002258 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002259 return;
2260
2261 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2262 if (!tcp_can_collapse(sk, skb))
2263 break;
2264
2265 space -= skb->len;
2266
2267 if (first) {
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002268 first = false;
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002269 continue;
2270 }
2271
2272 if (space < 0)
2273 break;
2274 /* Punt if not enough space exists in the first SKB for
2275 * the data in the second
2276 */
Eric Dumazeta21d4572012-04-10 20:30:48 +00002277 if (skb->len > skb_availroom(to))
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002278 break;
2279
2280 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2281 break;
2282
2283 tcp_collapse_retrans(sk, to);
2284 }
2285}
2286
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287/* This retransmits one SKB. Policy decisions and retransmit queue
2288 * state updates are done by the caller. Returns non-zero if an
2289 * error occurred which prevented the send.
2290 */
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002291int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292{
2293 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002294 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002295 unsigned int cur_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296
John Heffner5d424d52006-03-20 17:53:41 -08002297 /* Inconslusive MTU probe */
2298 if (icsk->icsk_mtup.probe_size) {
2299 icsk->icsk_mtup.probe_size = 0;
2300 }
2301
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002303 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 */
2305 if (atomic_read(&sk->sk_wmem_alloc) >
2306 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2307 return -EAGAIN;
2308
2309 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2310 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2311 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2313 return -ENOMEM;
2314 }
2315
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002316 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2317 return -EHOSTUNREACH; /* Routing failure or similar. */
2318
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002319 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002320
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 /* If receiver has shrunk his window, and skb is out of
2322 * new window, do not retransmit it. The exception is the
2323 * case, when window is shrunk to zero. In this case
2324 * our retransmit serves as a zero window probe.
2325 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002326 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2327 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 return -EAGAIN;
2329
2330 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07002331 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002333 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002334 int oldpcount = tcp_skb_pcount(skb);
2335
2336 if (unlikely(oldpcount > 1)) {
2337 tcp_init_tso_segs(sk, skb, cur_mss);
2338 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2339 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 }
2341
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002342 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 /* Some Solaris stacks overoptimize and ignore the FIN on a
2345 * retransmit when old data is attached. So strip it off
2346 * since it is cheap to do so and saves bytes on the network.
2347 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002348 if (skb->len > 0 &&
Eric Dumazet4de075e2011-09-27 13:25:05 -04002349 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002350 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002352 /* Reuse, even though it does some unnecessary work */
2353 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
Eric Dumazet4de075e2011-09-27 13:25:05 -04002354 TCP_SKB_CB(skb)->tcp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 }
2357 }
2358
2359 /* Make a copy, if the first transmission SKB clone we made
2360 * is still in somebody's hands, else make a clone.
2361 */
2362 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363
Thomas Graf50bceae2013-04-11 10:57:18 +00002364 /* make sure skb->data is aligned on arches that require it
2365 * and check if ack-trimming & collapsing extended the headroom
2366 * beyond what csum_start can cover.
2367 */
2368 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2369 skb_headroom(skb) >= 0xFFFF)) {
Eric Dumazet117632e2011-12-03 21:39:53 +00002370 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2371 GFP_ATOMIC);
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002372 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2373 -ENOBUFS;
Eric Dumazet117632e2011-12-03 21:39:53 +00002374 } else {
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002375 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Eric Dumazet117632e2011-12-03 21:39:53 +00002376 }
Yuchung Cheng93b174a2012-12-06 08:45:32 +00002377}
2378
2379int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2380{
2381 struct tcp_sock *tp = tcp_sk(sk);
2382 int err = __tcp_retransmit_skb(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383
2384 if (err == 0) {
2385 /* Update global TCP statistics. */
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002386 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
2388 tp->total_retrans++;
2389
2390#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002391 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Joe Perchese87cc472012-05-13 21:56:26 +00002392 net_dbg_ratelimited("retrans_out leaked\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 }
2394#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002395 if (!tp->retrans_out)
2396 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2398 tp->retrans_out += tcp_skb_pcount(skb);
2399
2400 /* Save stamp of the first retransmit. */
2401 if (!tp->retrans_stamp)
2402 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2403
Yuchung Chengc24f6912011-02-07 12:57:04 +00002404 tp->undo_retrans += tcp_skb_pcount(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405
2406 /* snd_nxt is stored to detect loss of retransmitted segment,
2407 * see tcp_input.c tcp_sacktag_write_queue().
2408 */
2409 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002410 } else {
2411 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 }
2413 return err;
2414}
2415
Andi Kleen67edfef2009-07-21 23:00:40 +00002416/* Check if we forward retransmits are possible in the current
2417 * window/congestion state.
2418 */
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002419static bool tcp_can_forward_retransmit(struct sock *sk)
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002420{
2421 const struct inet_connection_sock *icsk = inet_csk(sk);
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002422 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002423
2424 /* Forward retransmissions are possible only during Recovery. */
2425 if (icsk->icsk_ca_state != TCP_CA_Recovery)
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002426 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002427
2428 /* No forward retransmissions in Reno are possible. */
2429 if (tcp_is_reno(tp))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002430 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002431
2432 /* Yeah, we have to make difficult choice between forward transmission
2433 * and retransmission... Both ways have their merits...
2434 *
2435 * For now we do not retransmit anything, while we have some new
2436 * segments to send. In the other cases, follow rule 3 for
2437 * NextSeg() specified in RFC3517.
2438 */
2439
2440 if (tcp_may_send_now(sk))
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002441 return false;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002442
Eric Dumazeta2a385d2012-05-16 23:15:34 +00002443 return true;
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002444}
2445
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446/* This gets called after a retransmit timeout, and the initially
2447 * retransmitted data is acknowledged. It tries to continue
2448 * resending the rest of the retransmit queue, until either
2449 * we've sent it all or the congestion window limit is reached.
2450 * If doing SACK, the first ACK which comes back for a timeout
2451 * based retransmit packet might feed us FACK information again.
2452 * If so, we use it to avoid unnecessarily retransmissions.
2453 */
2454void tcp_xmit_retransmit_queue(struct sock *sk)
2455{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002456 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 struct tcp_sock *tp = tcp_sk(sk);
2458 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002459 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002460 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002461 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002462 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002463
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002464 if (!tp->packets_out)
2465 return;
2466
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002467 if (!tp->lost_out)
2468 tp->retransmit_high = tp->snd_una;
2469
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002470 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002471 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002472 last_lost = TCP_SKB_CB(skb)->end_seq;
2473 if (after(last_lost, tp->retransmit_high))
2474 last_lost = tp->retransmit_high;
2475 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002476 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002477 last_lost = tp->snd_una;
2478 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002480 tcp_for_write_queue_from(skb, sk) {
2481 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002483 if (skb == tcp_send_head(sk))
2484 break;
2485 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002486 if (hole == NULL)
2487 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002488
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002489 /* Assume this retransmit will generate
2490 * only one packet for congestion window
2491 * calculation purposes. This works because
2492 * tcp_retransmit_skb() will chop up the
2493 * packet to be MSS sized and all the
2494 * packet counting works out.
2495 */
2496 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2497 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002498
2499 if (fwd_rexmitting) {
2500begin_fwd:
2501 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2502 break;
2503 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2504
2505 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002506 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002507 if (!tcp_can_forward_retransmit(sk))
2508 break;
2509 /* Backtrack if necessary to non-L'ed skb */
2510 if (hole != NULL) {
2511 skb = hole;
2512 hole = NULL;
2513 }
2514 fwd_rexmitting = 1;
2515 goto begin_fwd;
2516
2517 } else if (!(sacked & TCPCB_LOST)) {
Ilpo Järvinenac11ba72009-02-28 04:44:27 +00002518 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002519 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002520 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002522 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002523 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002524 if (icsk->icsk_ca_state != TCP_CA_Loss)
2525 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2526 else
2527 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2528 }
2529
2530 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002531 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002532
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002533 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002534 return;
Yuchung Cheng24ab6be2013-07-12 11:33:04 -07002535
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002536 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2537
Yuchung Cheng684bad12012-09-02 17:38:04 +00002538 if (tcp_in_cwnd_reduction(sk))
Nandita Dukkipatia262f0c2011-08-21 20:21:57 +00002539 tp->prr_out += tcp_skb_pcount(skb);
2540
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002541 if (skb == tcp_write_queue_head(sk))
2542 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2543 inet_csk(sk)->icsk_rto,
2544 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546}
2547
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548/* Send a fin. The caller locks the socket for us. This cannot be
2549 * allowed to fail queueing a FIN frame under any circumstances.
2550 */
2551void tcp_send_fin(struct sock *sk)
2552{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002553 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002554 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002556
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 /* Optimization, tack on the FIN if we have a queue of
2558 * unsent frames. But be careful about outgoing SACKS
2559 * and IP options.
2560 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002561 mss_now = tcp_current_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562
David S. Millerfe067e82007-03-07 12:12:44 -08002563 if (tcp_send_head(sk) != NULL) {
Eric Dumazet4de075e2011-09-27 13:25:05 -04002564 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 TCP_SKB_CB(skb)->end_seq++;
2566 tp->write_seq++;
2567 } else {
2568 /* Socket is locked, keep trying until memory is available. */
2569 for (;;) {
Wu Fengguangaa133072009-09-02 23:45:45 -07002570 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2571 sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 if (skb)
2573 break;
2574 yield();
2575 }
2576
2577 /* Reserve space for headers and prepare control bits. */
2578 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002580 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002581 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 tcp_queue_skb(sk, skb);
2583 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002584 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585}
2586
2587/* We get here when a process closes a file descriptor (either due to
2588 * an explicit close() or as a byproduct of exit()'ing) and there
2589 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002590 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 */
Al Virodd0fc662005-10-07 07:46:04 +01002592void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 struct sk_buff *skb;
2595
2596 /* NOTE: No TCP options attached and we never retransmit this. */
2597 skb = alloc_skb(MAX_TCP_HEADER, priority);
2598 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002599 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 return;
2601 }
2602
2603 /* Reserve space for headers and prepare control bits. */
2604 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002605 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002606 TCPHDR_ACK | TCPHDR_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002609 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002610 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002611
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002612 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613}
2614
Andi Kleen67edfef2009-07-21 23:00:40 +00002615/* Send a crossed SYN-ACK during socket establishment.
2616 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 * a SYN packet that crossed the incoming SYN that caused this routine
2618 * to get called. If this assumption fails then the initial rcv_wnd
2619 * and rcv_wscale values will not be correct.
2620 */
2621int tcp_send_synack(struct sock *sk)
2622{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002623 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
David S. Millerfe067e82007-03-07 12:12:44 -08002625 skb = tcp_write_queue_head(sk);
Eric Dumazet4de075e2011-09-27 13:25:05 -04002626 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
Joe Perches91df42b2012-05-15 14:11:54 +00002627 pr_debug("%s: wrong queue state\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 return -EFAULT;
2629 }
Eric Dumazet4de075e2011-09-27 13:25:05 -04002630 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 if (skb_cloned(skb)) {
2632 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2633 if (nskb == NULL)
2634 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002635 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002637 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002638 sk_wmem_free_skb(sk, skb);
2639 sk->sk_wmem_queued += nskb->truesize;
2640 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 skb = nskb;
2642 }
2643
Eric Dumazet4de075e2011-09-27 13:25:05 -04002644 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 TCP_ECN_send_synack(tcp_sk(sk), skb);
2646 }
2647 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002648 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649}
2650
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002651/**
2652 * tcp_make_synack - Prepare a SYN-ACK.
2653 * sk: listener socket
2654 * dst: dst entry attached to the SYNACK
2655 * req: request_sock pointer
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002656 *
2657 * Allocate one skb and build a SYNACK packet.
2658 * @dst is consumed : Caller should not use it again.
2659 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002660struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002661 struct request_sock *req,
Jerry Chu83368862012-08-31 12:29:12 +00002662 struct tcp_fastopen_cookie *foc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002664 struct tcp_out_options opts;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002665 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 struct tcp_sock *tp = tcp_sk(sk);
2667 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002669 struct tcp_md5sig_key *md5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002670 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002671 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672
Christoph Paasch1a2c6182013-03-17 08:23:34 +00002673 skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC));
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002674 if (unlikely(!skb)) {
2675 dst_release(dst);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 return NULL;
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002677 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 /* Reserve space for headers. */
2679 skb_reserve(skb, MAX_TCP_HEADER);
2680
Eric Dumazet4aea39c2012-06-03 20:33:21 +00002681 skb_dst_set(skb, dst);
Eric Dumazetca10b9e2013-04-08 17:58:11 +00002682 security_skb_owned_by(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683
David S. Miller0dbaee32010-12-13 12:52:14 -08002684 mss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002685 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2686 mss = tp->rx_opt.user_mss;
2687
Adam Langley33ad7982008-07-19 00:04:31 -07002688 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2689 __u8 rcv_wscale;
2690 /* Set this up on the first call only */
2691 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00002692
2693 /* limit the window selection if the user enforce a smaller rx buffer */
2694 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2695 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2696 req->window_clamp = tcp_full_space(sk);
2697
Adam Langley33ad7982008-07-19 00:04:31 -07002698 /* tcp_full_space because it is guaranteed to be the first packet */
2699 tcp_select_initial_window(tcp_full_space(sk),
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002700 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Adam Langley33ad7982008-07-19 00:04:31 -07002701 &req->rcv_wnd,
2702 &req->window_clamp,
2703 ireq->wscale_ok,
laurent chavey31d12922009-12-15 11:15:28 +00002704 &rcv_wscale,
2705 dst_metric(dst, RTAX_INITRWND));
Adam Langley33ad7982008-07-19 00:04:31 -07002706 ireq->rcv_wscale = rcv_wscale;
2707 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002708
Adam Langley33ad7982008-07-19 00:04:31 -07002709 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002710#ifdef CONFIG_SYN_COOKIES
2711 if (unlikely(req->cookie_ts))
2712 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2713 else
2714#endif
Adam Langley33ad7982008-07-19 00:04:31 -07002715 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Christoph Paasch1a2c6182013-03-17 08:23:34 +00002716 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2717 foc) + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002718
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002719 skb_push(skb, tcp_header_size);
2720 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002721
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002722 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 memset(th, 0, sizeof(struct tcphdr));
2724 th->syn = 1;
2725 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 TCP_ECN_make_synack(req, th);
KOVACS Krisztiana3116ac52008-10-01 07:46:49 -07002727 th->source = ireq->loc_port;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002728 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002729 /* Setting of flags are superfluous here for callers (and ECE is
2730 * not even correctly set)
2731 */
2732 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00002733 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faade2009-12-02 18:25:27 +00002734
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Jerry Chu83368862012-08-31 12:29:12 +00002736 /* XXX data is queued and acked as is. No buffer/window check */
2737 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738
2739 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002740 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002741 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 th->doff = (tcp_header_size >> 2);
Tom Herbertaa2ea052010-04-22 07:00:24 +00002743 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002744
2745#ifdef CONFIG_TCP_MD5SIG
2746 /* Okay, we have all we need - do the md5 hash if needed */
2747 if (md5) {
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002748 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002749 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002750 }
2751#endif
2752
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 return skb;
2754}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002755EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
Andi Kleen67edfef2009-07-21 23:00:40 +00002757/* Do all connect socket setups that can be done AF independent. */
Pavel Emelyanov370816a2012-04-19 03:40:01 +00002758void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759{
Eric Dumazetcf533ea2011-10-21 05:22:42 -04002760 const struct dst_entry *dst = __sk_dst_get(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761 struct tcp_sock *tp = tcp_sk(sk);
2762 __u8 rcv_wscale;
2763
2764 /* We'll fix this up when we get a response from the other end.
2765 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2766 */
2767 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08002768 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002770#ifdef CONFIG_TCP_MD5SIG
2771 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2772 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2773#endif
2774
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 /* If user gave his TCP_MAXSEG, record it to clamp */
2776 if (tp->rx_opt.user_mss)
2777 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2778 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002779 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 tcp_sync_mss(sk, dst_mtu(dst));
2781
2782 if (!tp->window_clamp)
2783 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08002784 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002785 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2786 tp->advmss = tp->rx_opt.user_mss;
2787
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00002790 /* limit the window selection if the user enforce a smaller rx buffer */
2791 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2792 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2793 tp->window_clamp = tcp_full_space(sk);
2794
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 tcp_select_initial_window(tcp_full_space(sk),
2796 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2797 &tp->rcv_wnd,
2798 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08002799 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00002800 &rcv_wscale,
2801 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802
2803 tp->rx_opt.rcv_wscale = rcv_wscale;
2804 tp->rcv_ssthresh = tp->rcv_wnd;
2805
2806 sk->sk_err = 0;
2807 sock_reset_flag(sk, SOCK_DONE);
2808 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08002809 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 tp->snd_una = tp->write_seq;
2811 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002812 tp->snd_up = tp->write_seq;
Pavel Emelyanov370816a2012-04-19 03:40:01 +00002813 tp->snd_nxt = tp->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00002814
2815 if (likely(!tp->repair))
2816 tp->rcv_nxt = 0;
2817 tp->rcv_wup = tp->rcv_nxt;
2818 tp->copied_seq = tp->rcv_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002820 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2821 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 tcp_clear_retrans(tp);
2823}
2824
Yuchung Cheng783237e2012-07-19 06:43:07 +00002825static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2826{
2827 struct tcp_sock *tp = tcp_sk(sk);
2828 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2829
2830 tcb->end_seq += skb->len;
2831 skb_header_release(skb);
2832 __tcp_add_write_queue_tail(sk, skb);
2833 sk->sk_wmem_queued += skb->truesize;
2834 sk_mem_charge(sk, skb->truesize);
2835 tp->write_seq = tcb->end_seq;
2836 tp->packets_out += tcp_skb_pcount(skb);
2837}
2838
2839/* Build and send a SYN with data and (cached) Fast Open cookie. However,
2840 * queue a data-only packet after the regular SYN, such that regular SYNs
2841 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2842 * only the SYN sequence, the data are retransmitted in the first ACK.
2843 * If cookie is not cached or other error occurs, falls back to send a
2844 * regular SYN with Fast Open cookie request option.
2845 */
2846static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2847{
2848 struct tcp_sock *tp = tcp_sk(sk);
2849 struct tcp_fastopen_request *fo = tp->fastopen_req;
Yuchung Chengaab48742012-07-19 06:43:10 +00002850 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
Yuchung Cheng783237e2012-07-19 06:43:07 +00002851 struct sk_buff *syn_data = NULL, *data;
Yuchung Chengaab48742012-07-19 06:43:10 +00002852 unsigned long last_syn_loss = 0;
Yuchung Cheng783237e2012-07-19 06:43:07 +00002853
Yuchung Cheng67da22d2012-07-19 06:43:11 +00002854 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
Yuchung Chengaab48742012-07-19 06:43:10 +00002855 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2856 &syn_loss, &last_syn_loss);
2857 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2858 if (syn_loss > 1 &&
2859 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2860 fo->cookie.len = -1;
2861 goto fallback;
2862 }
2863
Yuchung Cheng67da22d2012-07-19 06:43:11 +00002864 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2865 fo->cookie.len = -1;
2866 else if (fo->cookie.len <= 0)
Yuchung Cheng783237e2012-07-19 06:43:07 +00002867 goto fallback;
2868
2869 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2870 * user-MSS. Reserve maximum option space for middleboxes that add
2871 * private TCP options. The cost is reduced data space in SYN :(
2872 */
2873 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2874 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
Yuchung Cheng1b63edd2013-02-22 08:59:06 +00002875 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
Yuchung Cheng783237e2012-07-19 06:43:07 +00002876 MAX_TCP_OPTION_SPACE;
2877
2878 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2879 sk->sk_allocation);
2880 if (syn_data == NULL)
2881 goto fallback;
2882
2883 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2884 struct iovec *iov = &fo->data->msg_iov[i];
2885 unsigned char __user *from = iov->iov_base;
2886 int len = iov->iov_len;
2887
2888 if (syn_data->len + len > space)
2889 len = space - syn_data->len;
2890 else if (i + 1 == iovlen)
2891 /* No more data pending in inet_wait_for_connect() */
2892 fo->data = NULL;
2893
2894 if (skb_add_data(syn_data, from, len))
2895 goto fallback;
2896 }
2897
2898 /* Queue a data-only packet after the regular SYN for retransmission */
2899 data = pskb_copy(syn_data, sk->sk_allocation);
2900 if (data == NULL)
2901 goto fallback;
2902 TCP_SKB_CB(data)->seq++;
2903 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2904 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2905 tcp_connect_queue_skb(sk, data);
2906 fo->copied = data->len;
2907
2908 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
Yuchung Cheng67da22d2012-07-19 06:43:11 +00002909 tp->syn_data = (fo->copied > 0);
Yuchung Cheng783237e2012-07-19 06:43:07 +00002910 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2911 goto done;
2912 }
2913 syn_data = NULL;
2914
2915fallback:
2916 /* Send a regular SYN with Fast Open cookie request option */
2917 if (fo->cookie.len > 0)
2918 fo->cookie.len = 0;
2919 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2920 if (err)
2921 tp->syn_fastopen = 0;
2922 kfree_skb(syn_data);
2923done:
2924 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2925 return err;
2926}
2927
Andi Kleen67edfef2009-07-21 23:00:40 +00002928/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929int tcp_connect(struct sock *sk)
2930{
2931 struct tcp_sock *tp = tcp_sk(sk);
2932 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00002933 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934
2935 tcp_connect_init(sk);
2936
Andrey Vagin2b916472012-11-22 01:13:58 +00002937 if (unlikely(tp->repair)) {
2938 tcp_finish_connect(sk, NULL);
2939 return 0;
2940 }
2941
David S. Millerd179cd12005-08-17 14:57:30 -07002942 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 if (unlikely(buff == NULL))
2944 return -ENOBUFS;
2945
2946 /* Reserve space for headers. */
2947 skb_reserve(buff, MAX_TCP_HEADER);
2948
Changli Gaoa3433f32010-06-12 14:01:43 +00002949 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Yuchung Cheng783237e2012-07-19 06:43:07 +00002950 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2951 tcp_connect_queue_skb(sk, buff);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002952 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953
Yuchung Cheng783237e2012-07-19 06:43:07 +00002954 /* Send off SYN; include data in Fast Open. */
2955 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2956 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Eric Parisee586812010-11-16 11:52:49 +00002957 if (err == -ECONNREFUSED)
2958 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002959
2960 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2961 * in order to make this packet get counted in tcpOutSegs.
2962 */
2963 tp->snd_nxt = tp->write_seq;
2964 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002965 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966
2967 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002968 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2969 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 return 0;
2971}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002972EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973
2974/* Send out a delayed ack, the caller does the policy checking
2975 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2976 * for details.
2977 */
2978void tcp_send_delayed_ack(struct sock *sk)
2979{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002980 struct inet_connection_sock *icsk = inet_csk(sk);
2981 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 unsigned long timeout;
2983
2984 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002985 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002986 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002988 if (icsk->icsk_ack.pingpong ||
2989 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 max_ato = TCP_DELACK_MAX;
2991
2992 /* Slow path, intersegment interval is "high". */
2993
2994 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002995 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 * directly.
2997 */
2998 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002999 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000
3001 if (rtt < max_ato)
3002 max_ato = rtt;
3003 }
3004
3005 ato = min(ato, max_ato);
3006 }
3007
3008 /* Stay within the limit we were given */
3009 timeout = jiffies + ato;
3010
3011 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003012 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 /* If delack timer was blocked or is about to expire,
3014 * send ACK now.
3015 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003016 if (icsk->icsk_ack.blocked ||
3017 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 tcp_send_ack(sk);
3019 return;
3020 }
3021
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003022 if (!time_before(timeout, icsk->icsk_ack.timeout))
3023 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003025 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3026 icsk->icsk_ack.timeout = timeout;
3027 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003028}
3029
3030/* This routine sends an ack and also updates the window. */
3031void tcp_send_ack(struct sock *sk)
3032{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003033 struct sk_buff *buff;
3034
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003036 if (sk->sk_state == TCP_CLOSE)
3037 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003039 /* We are not putting this on the write queue, so
3040 * tcp_transmit_skb() will set the ownership to this
3041 * sock.
3042 */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003043 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003044 if (buff == NULL) {
3045 inet_csk_schedule_ack(sk);
3046 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3047 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3048 TCP_DELACK_MAX, TCP_RTO_MAX);
3049 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003051
3052 /* Reserve space for headers and prepare control bits. */
3053 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00003054 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003055
3056 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003057 TCP_SKB_CB(buff)->when = tcp_time_stamp;
Mel Gorman99a1dec2012-07-31 16:44:14 -07003058 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059}
3060
3061/* This routine sends a packet with an out of date sequence
3062 * number. It assumes the other end will try to ack it.
3063 *
3064 * Question: what should we make while urgent mode?
3065 * 4.4BSD forces sending single byte of data. We cannot send
3066 * out of window data, because we have SND.NXT==SND.MAX...
3067 *
3068 * Current solution: to send TWO zero-length segments in urgent mode:
3069 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3070 * out-of-date with SND.UNA-1 to probe window.
3071 */
3072static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3073{
3074 struct tcp_sock *tp = tcp_sk(sk);
3075 struct sk_buff *skb;
3076
3077 /* We don't queue it, tcp_transmit_skb() sets ownership. */
Mel Gorman99a1dec2012-07-31 16:44:14 -07003078 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003079 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 return -1;
3081
3082 /* Reserve space for headers and set control bits. */
3083 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 /* Use a previous sequence. This should cause the other
3085 * end to send an ack. Don't queue or clone SKB, just
3086 * send it.
3087 */
Changli Gaoa3433f32010-06-12 14:01:43 +00003088 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08003090 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091}
3092
Pavel Emelyanovee995282012-04-19 03:40:39 +00003093void tcp_send_window_probe(struct sock *sk)
3094{
3095 if (sk->sk_state == TCP_ESTABLISHED) {
3096 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
Pavel Emelyanovc0e88ff2012-04-19 03:41:01 +00003097 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
Pavel Emelyanovee995282012-04-19 03:40:39 +00003098 tcp_xmit_probe_skb(sk, 0);
3099 }
3100}
3101
Andi Kleen67edfef2009-07-21 23:00:40 +00003102/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103int tcp_write_wakeup(struct sock *sk)
3104{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003105 struct tcp_sock *tp = tcp_sk(sk);
3106 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003108 if (sk->sk_state == TCP_CLOSE)
3109 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003111 if ((skb = tcp_send_head(sk)) != NULL &&
3112 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3113 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00003114 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003115 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003117 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3118 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003120 /* We are probing the opening of a window
3121 * but the window size is != 0
3122 * must have been a result SWS avoidance ( sender )
3123 */
3124 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3125 skb->len > mss) {
3126 seg_size = min(seg_size, mss);
Eric Dumazet4de075e2011-09-27 13:25:05 -04003127 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003128 if (tcp_fragment(sk, skb, seg_size, mss))
3129 return -1;
3130 } else if (!tcp_skb_pcount(skb))
3131 tcp_set_skb_tso_segs(sk, skb, mss);
3132
Eric Dumazet4de075e2011-09-27 13:25:05 -04003133 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003134 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3135 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3136 if (!err)
3137 tcp_event_new_data_sent(sk, skb);
3138 return err;
3139 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07003140 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08003141 tcp_xmit_probe_skb(sk, 1);
3142 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144}
3145
3146/* A window probe timeout has occurred. If window is not closed send
3147 * a partial packet else a zero probe.
3148 */
3149void tcp_send_probe0(struct sock *sk)
3150{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003151 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 struct tcp_sock *tp = tcp_sk(sk);
3153 int err;
3154
3155 err = tcp_write_wakeup(sk);
3156
David S. Millerfe067e82007-03-07 12:12:44 -08003157 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003159 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003160 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 return;
3162 }
3163
3164 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003165 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3166 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003167 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003168 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003169 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3170 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171 } else {
3172 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003173 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 * Let local senders to fight for local resources.
3175 *
3176 * Use accumulated backoff yet.
3177 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03003178 if (!icsk->icsk_probes_out)
3179 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09003180 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07003181 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07003182 TCP_RESOURCE_PROBE_INTERVAL),
3183 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 }
3185}