blob: 93316a96d820f5afc16984af390b2460e1b37904 [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
37#include <net/tcp.h>
38
39#include <linux/compiler.h>
40#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070043int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
Jianjun Kong09cb1052008-11-03 00:27:11 -080045/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080046 * interpret the window field as a signed quantity.
47 */
Brian Haleyab32ea52006-09-22 14:15:41 -070048int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080049
Linus Torvalds1da177e2005-04-16 15:20:36 -070050/* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
53 */
Brian Haleyab32ea52006-09-22 14:15:41 -070054int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Brian Haleyab32ea52006-09-22 14:15:41 -070056int sysctl_tcp_mtu_probing __read_mostly = 0;
57int sysctl_tcp_base_mss __read_mostly = 512;
John Heffner5d424d52006-03-20 17:53:41 -080058
David S. Miller35089bb2006-06-13 22:33:04 -070059/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070061
William Allen Simpson519855c2009-12-02 18:14:19 +000062int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
David S. Millere6b09cc2009-12-02 22:18:58 -080063EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
William Allen Simpson519855c2009-12-02 18:14:19 +000064
65
Andi Kleen67edfef2009-07-21 23:00:40 +000066/* Account for new data that has been sent to the network. */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080067static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070068{
69 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080070 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070071
David S. Millerfe067e82007-03-07 12:12:44 -080072 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080074
75 /* Don't override Nagle indefinately with F-RTO */
76 if (tp->frto_counter == 2)
77 tp->frto_counter = 3;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080078
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets)
81 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
82 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -070083}
84
85/* SND.NXT, if window was not shrunk.
86 * If window has been shrunk, what should we make? It is not clear at all.
87 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
88 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
89 * invalid. OK, let's make this for now:
90 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070091static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070092{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070093 struct tcp_sock *tp = tcp_sk(sk);
94
Ilpo Järvinen90840de2007-12-31 04:48:41 -080095 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 return tp->snd_nxt;
97 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -080098 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -070099}
100
101/* Calculate mss to advertise in SYN segment.
102 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
103 *
104 * 1. It is independent of path mtu.
105 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
106 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
107 * attached devices, because some buggy hosts are confused by
108 * large MSS.
109 * 4. We do not make 3, we advertise MSS, calculated from first
110 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
111 * This may be overridden via information stored in routing table.
112 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
113 * probably even Jumbo".
114 */
115static __u16 tcp_advertise_mss(struct sock *sk)
116{
117 struct tcp_sock *tp = tcp_sk(sk);
118 struct dst_entry *dst = __sk_dst_get(sk);
119 int mss = tp->advmss;
120
121 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
122 mss = dst_metric(dst, RTAX_ADVMSS);
123 tp->advmss = mss;
124 }
125
126 return (__u16)mss;
127}
128
129/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
130 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700131static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700133 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 s32 delta = tcp_time_stamp - tp->lsndtime;
135 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
136 u32 cwnd = tp->snd_cwnd;
137
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300138 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300140 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 restart_cwnd = min(restart_cwnd, cwnd);
142
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700143 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 cwnd >>= 1;
145 tp->snd_cwnd = max(cwnd, restart_cwnd);
146 tp->snd_cwnd_stamp = tcp_time_stamp;
147 tp->snd_cwnd_used = 0;
148}
149
Andi Kleen67edfef2009-07-21 23:00:40 +0000150/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800151static void tcp_event_data_sent(struct tcp_sock *tp,
152 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700154 struct inet_connection_sock *icsk = inet_csk(sk);
155 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156
David S. Miller35089bb2006-06-13 22:33:04 -0700157 if (sysctl_tcp_slow_start_after_idle &&
158 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700159 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160
161 tp->lsndtime = now;
162
163 /* If it is a reply for ato after last received
164 * packet, enter pingpong mode.
165 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700166 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
167 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168}
169
Andi Kleen67edfef2009-07-21 23:00:40 +0000170/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800171static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700173 tcp_dec_quickack_mode(sk, pkts);
174 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175}
176
177/* Determine a window scaling and initial window to offer.
178 * Based on the assumption that the given amount of space
179 * will be offered. Store the results in the tp structure.
180 * NOTE: for smooth operation initial space offering should
181 * be a multiple of mss if possible. We assume here that mss >= 1.
182 * This MUST be enforced by all callers.
183 */
184void tcp_select_initial_window(int __space, __u32 mss,
185 __u32 *rcv_wnd, __u32 *window_clamp,
186 int wscale_ok, __u8 *rcv_wscale)
187{
188 unsigned int space = (__space < 0 ? 0 : __space);
189
190 /* If no clamp set the clamp to the max possible scaled window */
191 if (*window_clamp == 0)
192 (*window_clamp) = (65535 << 14);
193 space = min(*window_clamp, space);
194
195 /* Quantize space offering to a multiple of mss if possible. */
196 if (space > mss)
197 space = (space / mss) * mss;
198
199 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800200 * will break some buggy TCP stacks. If the admin tells us
201 * it is likely we could be speaking with such a buggy stack
202 * we will truncate our initial window offering to 32K-1
203 * unless the remote has sent us a window scaling option,
204 * which we interpret as a sign the remote TCP is not
205 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206 */
Rick Jones15d99e02006-03-20 22:40:29 -0800207 if (sysctl_tcp_workaround_signed_windows)
208 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
209 else
210 (*rcv_wnd) = space;
211
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 (*rcv_wscale) = 0;
213 if (wscale_ok) {
214 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900215 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 */
217 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700218 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 while (space > 65535 && (*rcv_wscale) < 14) {
220 space >>= 1;
221 (*rcv_wscale)++;
222 }
223 }
224
225 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700226 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227 * will be satisfied with 2.
228 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800229 if (mss > (1 << *rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700230 int init_cwnd = 4;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800231 if (mss > 1460 * 3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700233 else if (mss > 1460)
234 init_cwnd = 3;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800235 if (*rcv_wnd > init_cwnd * mss)
236 *rcv_wnd = init_cwnd * mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 }
238
239 /* Set the clamp no higher than max representable value */
240 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
241}
242
243/* Chose a new window to advertise, update state in tcp_sock for the
244 * socket, and return result with RFC1323 scaling applied. The return
245 * value can be stuffed directly into th->window for an outgoing
246 * frame.
247 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800248static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249{
250 struct tcp_sock *tp = tcp_sk(sk);
251 u32 cur_win = tcp_receive_window(tp);
252 u32 new_win = __tcp_select_window(sk);
253
254 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800255 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 /* Danger Will Robinson!
257 * Don't update rcv_wup/rcv_wnd here or else
258 * we will not be able to advertise a zero
259 * window in time. --DaveM
260 *
261 * Relax Will Robinson.
262 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700263 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 }
265 tp->rcv_wnd = new_win;
266 tp->rcv_wup = tp->rcv_nxt;
267
268 /* Make sure we do not exceed the maximum possible
269 * scaled window.
270 */
Rick Jones15d99e02006-03-20 22:40:29 -0800271 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 new_win = min(new_win, MAX_TCP_WINDOW);
273 else
274 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
275
276 /* RFC1323 scaling applied */
277 new_win >>= tp->rx_opt.rcv_wscale;
278
279 /* If we advertise zero window, disable fast path. */
280 if (new_win == 0)
281 tp->pred_flags = 0;
282
283 return new_win;
284}
285
Andi Kleen67edfef2009-07-21 23:00:40 +0000286/* Packet ECN state for a SYN-ACK */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800287static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700288{
289 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800290 if (!(tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700291 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
292}
293
Andi Kleen67edfef2009-07-21 23:00:40 +0000294/* Packet ECN state for a SYN. */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700295static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
296{
297 struct tcp_sock *tp = tcp_sk(sk);
298
299 tp->ecn_flags = 0;
Ilpo Järvinen255cac92009-05-04 11:07:36 -0700300 if (sysctl_tcp_ecn == 1) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800301 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700302 tp->ecn_flags = TCP_ECN_OK;
303 }
304}
305
306static __inline__ void
307TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
308{
309 if (inet_rsk(req)->ecn_ok)
310 th->ece = 1;
311}
312
Andi Kleen67edfef2009-07-21 23:00:40 +0000313/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
314 * be sent.
315 */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700316static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
317 int tcp_header_len)
318{
319 struct tcp_sock *tp = tcp_sk(sk);
320
321 if (tp->ecn_flags & TCP_ECN_OK) {
322 /* Not-retransmitted data segment: set ECT and inject CWR. */
323 if (skb->len != tcp_header_len &&
324 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
325 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800326 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700327 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
328 tcp_hdr(skb)->cwr = 1;
329 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
330 }
331 } else {
332 /* ACK or retransmitted segment: clear ECT|CE */
333 INET_ECN_dontxmit(sk);
334 }
335 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
336 tcp_hdr(skb)->ece = 1;
337 }
338}
339
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800340/* Constructs common control bits of non-data skb. If SYN/FIN is present,
341 * auto increment end seqno.
342 */
343static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
344{
345 skb->csum = 0;
346
347 TCP_SKB_CB(skb)->flags = flags;
348 TCP_SKB_CB(skb)->sacked = 0;
349
350 skb_shinfo(skb)->gso_segs = 1;
351 skb_shinfo(skb)->gso_size = 0;
352 skb_shinfo(skb)->gso_type = 0;
353
354 TCP_SKB_CB(skb)->seq = seq;
355 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
356 seq++;
357 TCP_SKB_CB(skb)->end_seq = seq;
358}
359
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700360static inline int tcp_urg_mode(const struct tcp_sock *tp)
361{
362 return tp->snd_una != tp->snd_up;
363}
364
Adam Langley33ad7982008-07-19 00:04:31 -0700365#define OPTION_SACK_ADVERTISE (1 << 0)
366#define OPTION_TS (1 << 1)
367#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000368#define OPTION_WSCALE (1 << 3)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000369#define OPTION_COOKIE_EXTENSION (1 << 4)
Adam Langley33ad7982008-07-19 00:04:31 -0700370
371struct tcp_out_options {
372 u8 options; /* bit field of OPTION_* */
373 u8 ws; /* window scale, 0 to disable */
374 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000375 u8 hash_size; /* bytes in hash_location */
Adam Langley33ad7982008-07-19 00:04:31 -0700376 u16 mss; /* 0 to disable */
377 __u32 tsval, tsecr; /* need to include OPTION_TS */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000378 __u8 *hash_location; /* temporary pointer, overloaded */
Adam Langley33ad7982008-07-19 00:04:31 -0700379};
380
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000381/* The sysctl int routines are generic, so check consistency here.
382 */
383static u8 tcp_cookie_size_check(u8 desired)
384{
385 if (desired > 0) {
386 /* previously specified */
387 return desired;
388 }
389 if (sysctl_tcp_cookie_size <= 0) {
390 /* no default specified */
391 return 0;
392 }
393 if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) {
394 /* value too small, specify minimum */
395 return TCP_COOKIE_MIN;
396 }
397 if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) {
398 /* value too large, specify maximum */
399 return TCP_COOKIE_MAX;
400 }
401 if (0x1 & sysctl_tcp_cookie_size) {
402 /* 8-bit multiple, illegal, fix it */
403 return (u8)(sysctl_tcp_cookie_size + 0x1);
404 }
405 return (u8)sysctl_tcp_cookie_size;
406}
407
Andi Kleen67edfef2009-07-21 23:00:40 +0000408/* Write previously computed TCP options to the packet.
409 *
410 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700411 * TCP options, we learned this through the hard way, so be careful here.
412 * Luckily we can at least blame others for their non-compliance but from
413 * inter-operatibility perspective it seems that we're somewhat stuck with
414 * the ordering which we have been using if we want to keep working with
415 * those broken things (not that it currently hurts anybody as there isn't
416 * particular reason why the ordering would need to be changed).
417 *
418 * At least SACK_PERM as the first option is known to lead to a disaster
419 * (but it may well be that other scenarios fail similarly).
420 */
Adam Langley33ad7982008-07-19 00:04:31 -0700421static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000422 struct tcp_out_options *opts)
423{
424 u8 options = opts->options; /* mungable copy */
425
426 /* Having both authentication and cookies for security is redundant,
427 * and there's certainly not enough room. Instead, the cookie-less
428 * extension variant is proposed.
429 *
430 * Consider the pessimal case with authentication. The options
431 * could look like:
432 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
433 */
434 if (unlikely(OPTION_MD5 & options)) {
435 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
436 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
437 (TCPOLEN_COOKIE_BASE << 16) |
438 (TCPOPT_MD5SIG << 8) |
439 TCPOLEN_MD5SIG);
440 } else {
441 *ptr++ = htonl((TCPOPT_NOP << 24) |
442 (TCPOPT_NOP << 16) |
443 (TCPOPT_MD5SIG << 8) |
444 TCPOLEN_MD5SIG);
445 }
446 options &= ~OPTION_COOKIE_EXTENSION;
447 /* overload cookie hash location */
448 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700449 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800450 }
Adam Langley33ad7982008-07-19 00:04:31 -0700451
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700452 if (unlikely(opts->mss)) {
453 *ptr++ = htonl((TCPOPT_MSS << 24) |
454 (TCPOLEN_MSS << 16) |
455 opts->mss);
456 }
457
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000458 if (likely(OPTION_TS & options)) {
459 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700460 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
461 (TCPOLEN_SACK_PERM << 16) |
462 (TCPOPT_TIMESTAMP << 8) |
463 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000464 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700465 } else {
466 *ptr++ = htonl((TCPOPT_NOP << 24) |
467 (TCPOPT_NOP << 16) |
468 (TCPOPT_TIMESTAMP << 8) |
469 TCPOLEN_TIMESTAMP);
470 }
471 *ptr++ = htonl(opts->tsval);
472 *ptr++ = htonl(opts->tsecr);
473 }
474
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000475 /* Specification requires after timestamp, so do it now.
476 *
477 * Consider the pessimal case without authentication. The options
478 * could look like:
479 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
480 */
481 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
482 __u8 *cookie_copy = opts->hash_location;
483 u8 cookie_size = opts->hash_size;
484
485 /* 8-bit multiple handled in tcp_cookie_size_check() above,
486 * and elsewhere.
487 */
488 if (0x2 & cookie_size) {
489 __u8 *p = (__u8 *)ptr;
490
491 /* 16-bit multiple */
492 *p++ = TCPOPT_COOKIE;
493 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
494 *p++ = *cookie_copy++;
495 *p++ = *cookie_copy++;
496 ptr++;
497 cookie_size -= 2;
498 } else {
499 /* 32-bit multiple */
500 *ptr++ = htonl(((TCPOPT_NOP << 24) |
501 (TCPOPT_NOP << 16) |
502 (TCPOPT_COOKIE << 8) |
503 TCPOLEN_COOKIE_BASE) +
504 cookie_size);
505 }
506
507 if (cookie_size > 0) {
508 memcpy(ptr, cookie_copy, cookie_size);
509 ptr += (cookie_size / 4);
510 }
511 }
512
513 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700514 *ptr++ = htonl((TCPOPT_NOP << 24) |
515 (TCPOPT_NOP << 16) |
516 (TCPOPT_SACK_PERM << 8) |
517 TCPOLEN_SACK_PERM);
518 }
519
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000520 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700521 *ptr++ = htonl((TCPOPT_NOP << 24) |
522 (TCPOPT_WINDOW << 16) |
523 (TCPOLEN_WINDOW << 8) |
524 opts->ws);
525 }
526
527 if (unlikely(opts->num_sack_blocks)) {
528 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
529 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800530 int this_sack;
531
532 *ptr++ = htonl((TCPOPT_NOP << 24) |
533 (TCPOPT_NOP << 16) |
534 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700535 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800536 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800537
Adam Langley33ad7982008-07-19 00:04:31 -0700538 for (this_sack = 0; this_sack < opts->num_sack_blocks;
539 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800540 *ptr++ = htonl(sp[this_sack].start_seq);
541 *ptr++ = htonl(sp[this_sack].end_seq);
542 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800543
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000544 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800545 }
546}
547
Andi Kleen67edfef2009-07-21 23:00:40 +0000548/* Compute TCP options for SYN packets. This is not the final
549 * network wire format yet.
550 */
Adam Langley33ad7982008-07-19 00:04:31 -0700551static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
552 struct tcp_out_options *opts,
553 struct tcp_md5sig_key **md5) {
554 struct tcp_sock *tp = tcp_sk(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000555 struct tcp_cookie_values *cvp = tp->cookie_values;
Gilad Ben-Yossef1aba7212009-10-28 04:15:24 +0000556 struct dst_entry *dst = __sk_dst_get(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000557 unsigned remaining = MAX_TCP_OPTION_SPACE;
558 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
559 tcp_cookie_size_check(cvp->cookie_desired) :
560 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700561
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800562#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700563 *md5 = tp->af_specific->md5_lookup(sk, sk);
564 if (*md5) {
565 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000566 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800567 }
Adam Langley33ad7982008-07-19 00:04:31 -0700568#else
569 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800570#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700571
572 /* We always get an MSS option. The option bytes which will be seen in
573 * normal data packets should timestamps be used, must be in the MSS
574 * advertised. But we subtract them from tp->mss_cache so that
575 * calculations in tcp_sendmsg are simpler etc. So account for this
576 * fact here if necessary. If we don't do this correctly, as a
577 * receiver we won't recognize data packets as being full sized when we
578 * should, and thus we won't abide by the delayed ACK rules correctly.
579 * SACKs don't matter, we never delay an ACK when we have any of those
580 * going out. */
581 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000582 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700583
Gilad Ben-Yossefcda42eb2009-10-28 04:15:25 +0000584 if (likely(sysctl_tcp_timestamps &&
585 !dst_feature(dst, RTAX_FEATURE_NO_TSTAMP) &&
586 *md5 == NULL)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700587 opts->options |= OPTION_TS;
588 opts->tsval = TCP_SKB_CB(skb)->when;
589 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000590 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700591 }
Gilad Ben-Yossef345cda22009-10-28 04:15:26 +0000592 if (likely(sysctl_tcp_window_scaling &&
593 !dst_feature(dst, RTAX_FEATURE_NO_WSCALE))) {
Adam Langley33ad7982008-07-19 00:04:31 -0700594 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000595 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000596 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700597 }
Gilad Ben-Yossef1aba7212009-10-28 04:15:24 +0000598 if (likely(sysctl_tcp_sack &&
599 !dst_feature(dst, RTAX_FEATURE_NO_SACK))) {
Adam Langley33ad7982008-07-19 00:04:31 -0700600 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700601 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000602 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700603 }
604
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000605 /* Note that timestamps are required by the specification.
606 *
607 * Odd numbers of bytes are prohibited by the specification, ensuring
608 * that the cookie is 16-bit aligned, and the resulting cookie pair is
609 * 32-bit aligned.
610 */
611 if (*md5 == NULL &&
612 (OPTION_TS & opts->options) &&
613 cookie_size > 0) {
614 int need = TCPOLEN_COOKIE_BASE + cookie_size;
615
616 if (0x2 & need) {
617 /* 32-bit multiple */
618 need += 2; /* NOPs */
619
620 if (need > remaining) {
621 /* try shrinking cookie to fit */
622 cookie_size -= 2;
623 need -= 4;
624 }
625 }
626 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
627 cookie_size -= 4;
628 need -= 4;
629 }
630 if (TCP_COOKIE_MIN <= cookie_size) {
631 opts->options |= OPTION_COOKIE_EXTENSION;
632 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
633 opts->hash_size = cookie_size;
634
635 /* Remember for future incarnations. */
636 cvp->cookie_desired = cookie_size;
637
638 if (cvp->cookie_desired != cvp->cookie_pair_size) {
639 /* Currently use random bytes as a nonce,
640 * assuming these are completely unpredictable
641 * by hostile users of the same system.
642 */
643 get_random_bytes(&cvp->cookie_pair[0],
644 cookie_size);
645 cvp->cookie_pair_size = cookie_size;
646 }
647
648 remaining -= need;
649 }
650 }
651 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700652}
653
Andi Kleen67edfef2009-07-21 23:00:40 +0000654/* Set up TCP options for SYN-ACKs. */
Adam Langley33ad7982008-07-19 00:04:31 -0700655static unsigned tcp_synack_options(struct sock *sk,
656 struct request_sock *req,
657 unsigned mss, struct sk_buff *skb,
658 struct tcp_out_options *opts,
William Allen Simpson4957faa2009-12-02 18:25:27 +0000659 struct tcp_md5sig_key **md5,
660 struct tcp_extend_values *xvp)
661{
Adam Langley33ad7982008-07-19 00:04:31 -0700662 struct inet_request_sock *ireq = inet_rsk(req);
William Allen Simpson4957faa2009-12-02 18:25:27 +0000663 unsigned remaining = MAX_TCP_OPTION_SPACE;
664 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
665 xvp->cookie_plus :
666 0;
667 bool doing_ts = ireq->tstamp_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700668
669#ifdef CONFIG_TCP_MD5SIG
670 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
671 if (*md5) {
672 opts->options |= OPTION_MD5;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000673 remaining -= TCPOLEN_MD5SIG_ALIGNED;
674
675 /* We can't fit any SACK blocks in a packet with MD5 + TS
676 * options. There was discussion about disabling SACK
677 * rather than TS in order to fit in better with old,
678 * buggy kernels, but that was deemed to be unnecessary.
679 */
680 doing_ts &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700681 }
682#else
683 *md5 = NULL;
684#endif
685
William Allen Simpson4957faa2009-12-02 18:25:27 +0000686 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700687 opts->mss = mss;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000688 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700689
690 if (likely(ireq->wscale_ok)) {
691 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000692 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000693 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700694 }
695 if (likely(doing_ts)) {
696 opts->options |= OPTION_TS;
697 opts->tsval = TCP_SKB_CB(skb)->when;
698 opts->tsecr = req->ts_recent;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000699 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700700 }
701 if (likely(ireq->sack_ok)) {
702 opts->options |= OPTION_SACK_ADVERTISE;
703 if (unlikely(!doing_ts))
William Allen Simpson4957faa2009-12-02 18:25:27 +0000704 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700705 }
706
William Allen Simpson4957faa2009-12-02 18:25:27 +0000707 /* Similar rationale to tcp_syn_options() applies here, too.
708 * If the <SYN> options fit, the same options should fit now!
709 */
710 if (*md5 == NULL &&
711 doing_ts &&
712 cookie_plus > TCPOLEN_COOKIE_BASE) {
713 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
714
715 if (0x2 & need) {
716 /* 32-bit multiple */
717 need += 2; /* NOPs */
718 }
719 if (need <= remaining) {
720 opts->options |= OPTION_COOKIE_EXTENSION;
721 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
722 remaining -= need;
723 } else {
724 /* There's no error return, so flag it. */
725 xvp->cookie_out_never = 1; /* true */
726 opts->hash_size = 0;
727 }
728 }
729 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700730}
731
Andi Kleen67edfef2009-07-21 23:00:40 +0000732/* Compute TCP options for ESTABLISHED sockets. This is not the
733 * final wire format yet.
734 */
Adam Langley33ad7982008-07-19 00:04:31 -0700735static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
736 struct tcp_out_options *opts,
737 struct tcp_md5sig_key **md5) {
738 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
739 struct tcp_sock *tp = tcp_sk(sk);
740 unsigned size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000741 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700742
743#ifdef CONFIG_TCP_MD5SIG
744 *md5 = tp->af_specific->md5_lookup(sk, sk);
745 if (unlikely(*md5)) {
746 opts->options |= OPTION_MD5;
747 size += TCPOLEN_MD5SIG_ALIGNED;
748 }
749#else
750 *md5 = NULL;
751#endif
752
753 if (likely(tp->rx_opt.tstamp_ok)) {
754 opts->options |= OPTION_TS;
755 opts->tsval = tcb ? tcb->when : 0;
756 opts->tsecr = tp->rx_opt.ts_recent;
757 size += TCPOLEN_TSTAMP_ALIGNED;
758 }
759
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000760 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
761 if (unlikely(eff_sacks)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700762 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
763 opts->num_sack_blocks =
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000764 min_t(unsigned, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700765 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
766 TCPOLEN_SACK_PERBLOCK);
767 size += TCPOLEN_SACK_BASE_ALIGNED +
768 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
769 }
770
771 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800772}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
774/* This routine actually transmits TCP packets queued in by
775 * tcp_do_sendmsg(). This is used by both the initial
776 * transmission and possible later retransmissions.
777 * All SKB's seen here are completely headerless. It is our
778 * job to build the TCP header, and pass the packet down to
779 * IP so it can do the same plus pass the packet off to the
780 * device.
781 *
782 * We are working here with either a clone of the original
783 * SKB, or a fresh unique copy made by the retransmit engine.
784 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800785static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
786 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800788 const struct inet_connection_sock *icsk = inet_csk(sk);
789 struct inet_sock *inet;
790 struct tcp_sock *tp;
791 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700792 struct tcp_out_options opts;
793 unsigned tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800794 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800795 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800796 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800798 BUG_ON(!skb || !tcp_skb_pcount(skb));
799
800 /* If congestion control is doing timestamping, we must
801 * take such a timestamp before we potentially clone/copy.
802 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700803 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800804 __net_timestamp(skb);
805
806 if (likely(clone_it)) {
807 if (unlikely(skb_cloned(skb)))
808 skb = pskb_copy(skb, gfp_mask);
809 else
810 skb = skb_clone(skb, gfp_mask);
811 if (unlikely(!skb))
812 return -ENOBUFS;
813 }
814
815 inet = inet_sk(sk);
816 tp = tcp_sk(sk);
817 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700818 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819
Adam Langley33ad7982008-07-19 00:04:31 -0700820 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
821 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
822 else
823 tcp_options_size = tcp_established_options(sk, skb, &opts,
824 &md5);
825 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900826
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800827 if (tcp_packets_in_flight(tp) == 0)
828 tcp_ca_event(sk, CA_EVENT_TX_START);
829
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700830 skb_push(skb, tcp_header_size);
831 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800832 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800833
834 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700835 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000836 th->source = inet->inet_sport;
837 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800838 th->seq = htonl(tcb->seq);
839 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700840 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800841 tcb->flags);
842
843 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
844 /* RFC1323: The window in SYN & SYN/ACK segments
845 * is never scaled.
846 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800847 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800848 } else {
849 th->window = htons(tcp_select_window(sk));
850 }
851 th->check = 0;
852 th->urg_ptr = 0;
853
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700854 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800855 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
856 if (before(tp->snd_up, tcb->seq + 0x10000)) {
857 th->urg_ptr = htons(tp->snd_up - tcb->seq);
858 th->urg = 1;
859 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
860 th->urg_ptr = 0xFFFF;
861 th->urg = 1;
862 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800863 }
864
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000865 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Adam Langley33ad7982008-07-19 00:04:31 -0700866 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700867 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800868
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800869#ifdef CONFIG_TCP_MD5SIG
870 /* Calculate the MD5 hash, as we have all we need now */
871 if (md5) {
Adam Langley33ad7982008-07-19 00:04:31 -0700872 sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000873 tp->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700874 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800875 }
876#endif
877
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800878 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800879
880 if (likely(tcb->flags & TCPCB_FLAG_ACK))
881 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
882
883 if (skb->len != tcp_header_size)
884 tcp_event_data_sent(tp, skb, sk);
885
Wei Yongjunbd37a082006-08-07 21:04:15 -0700886 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -0700887 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800888
David S. Millere89862f2007-01-26 01:04:55 -0800889 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700890 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800891 return err;
892
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800893 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800894
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200895 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896}
897
Andi Kleen67edfef2009-07-21 23:00:40 +0000898/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 *
900 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
901 * otherwise socket can stall.
902 */
903static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
904{
905 struct tcp_sock *tp = tcp_sk(sk);
906
907 /* Advance write_seq and place onto the write_queue. */
908 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
909 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800910 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800911 sk->sk_wmem_queued += skb->truesize;
912 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913}
914
Andi Kleen67edfef2009-07-21 23:00:40 +0000915/* Initialize TSO segments for a packet. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800916static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
917 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700918{
Herbert Xu8e5b9dd2009-06-28 18:03:30 +0000919 if (skb->len <= mss_now || !sk_can_gso(sk) ||
920 skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -0700921 /* Avoid the costly divide in the normal
922 * non-TSO case.
923 */
Herbert Xu79671682006-06-22 02:40:14 -0700924 skb_shinfo(skb)->gso_segs = 1;
925 skb_shinfo(skb)->gso_size = 0;
926 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700927 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700928 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700929 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700930 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 }
932}
933
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700934/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800935 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700936 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800937static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700938 int decr)
939{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800940 struct tcp_sock *tp = tcp_sk(sk);
941
Ilpo Järvinendc869672007-10-01 15:27:19 -0700942 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700943 return;
944
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200945 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700946 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700947}
948
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000949/* Pcount in the middle of the write queue got changed, we need to do various
950 * tweaks to fix counters
951 */
952static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
953{
954 struct tcp_sock *tp = tcp_sk(sk);
955
956 tp->packets_out -= decr;
957
958 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
959 tp->sacked_out -= decr;
960 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
961 tp->retrans_out -= decr;
962 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
963 tp->lost_out -= decr;
964
965 /* Reno case is special. Sigh... */
966 if (tcp_is_reno(tp) && decr > 0)
967 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
968
969 tcp_adjust_fackets_out(sk, skb, decr);
970
971 if (tp->lost_skb_hint &&
972 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +0000973 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000974 tp->lost_cnt_hint -= decr;
975
976 tcp_verify_left_out(tp);
977}
978
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979/* Function to create two new TCP segments. Shrinks the given segment
980 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900981 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 * Remember, these are still headerless SKBs at this point.
983 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800984int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
985 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986{
987 struct tcp_sock *tp = tcp_sk(sk);
988 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700989 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700990 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +0000991 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992
Herbert Xub2cc99f2005-10-20 17:13:13 -0200993 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800994
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 nsize = skb_headlen(skb) - len;
996 if (nsize < 0)
997 nsize = 0;
998
999 if (skb_cloned(skb) &&
1000 skb_is_nonlinear(skb) &&
1001 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1002 return -ENOMEM;
1003
1004 /* Get a new skb... force flag on. */
1005 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1006 if (buff == NULL)
1007 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001008
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001009 sk->sk_wmem_queued += buff->truesize;
1010 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001011 nlen = skb->len - len - nsize;
1012 buff->truesize += nlen;
1013 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014
1015 /* Correct the sequence numbers. */
1016 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1017 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1018 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1019
1020 /* PSH and FIN should only be set in the second packet. */
1021 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001022 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001024 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025
Patrick McHardy84fa7932006-08-29 16:44:56 -07001026 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001028 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1029 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 nsize, 0);
1031
1032 skb_trim(skb, len);
1033
1034 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1035 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001036 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 skb_split(skb, buff, len);
1038 }
1039
1040 buff->ip_summed = skb->ip_summed;
1041
1042 /* Looks stupid, but our code really uses when of
1043 * skbs, which it never sent before. --ANK
1044 */
1045 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001046 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047
David S. Miller6475be12005-09-01 22:47:01 -07001048 old_factor = tcp_skb_pcount(skb);
1049
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001051 tcp_set_skb_tso_segs(sk, skb, mss_now);
1052 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
David S. Miller6475be12005-09-01 22:47:01 -07001054 /* If this packet has been sent out already, we must
1055 * adjust the various packet counters.
1056 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001057 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001058 int diff = old_factor - tcp_skb_pcount(skb) -
1059 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001061 if (diff)
1062 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 }
1064
1065 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -07001066 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001067 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
1069 return 0;
1070}
1071
1072/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1073 * eventually). The difference is that pulled data not copied, but
1074 * immediately discarded.
1075 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001076static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077{
1078 int i, k, eat;
1079
1080 eat = len;
1081 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001082 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 if (skb_shinfo(skb)->frags[i].size <= eat) {
1084 put_page(skb_shinfo(skb)->frags[i].page);
1085 eat -= skb_shinfo(skb)->frags[i].size;
1086 } else {
1087 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1088 if (eat) {
1089 skb_shinfo(skb)->frags[k].page_offset += eat;
1090 skb_shinfo(skb)->frags[k].size -= eat;
1091 eat = 0;
1092 }
1093 k++;
1094 }
1095 }
1096 skb_shinfo(skb)->nr_frags = k;
1097
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001098 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 skb->data_len -= len;
1100 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101}
1102
Andi Kleen67edfef2009-07-21 23:00:40 +00001103/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001104int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1105{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001106 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 return -ENOMEM;
1108
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001109 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1110 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001112 else
1113 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114
1115 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001116 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117
1118 skb->truesize -= len;
1119 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001120 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1122
1123 /* Any change of skb->len requires recalculation of tso
1124 * factor and mss.
1125 */
1126 if (tcp_skb_pcount(skb) > 1)
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001127 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128
1129 return 0;
1130}
1131
Andi Kleen67edfef2009-07-21 23:00:40 +00001132/* Calculate MSS. Not accounting for SACKs here. */
John Heffner5d424d52006-03-20 17:53:41 -08001133int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1134{
1135 struct tcp_sock *tp = tcp_sk(sk);
1136 struct inet_connection_sock *icsk = inet_csk(sk);
1137 int mss_now;
1138
1139 /* Calculate base mss without TCP options:
1140 It is MMS_S - sizeof(tcphdr) of rfc1122
1141 */
1142 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1143
1144 /* Clamp it (mss_clamp does not include tcp options) */
1145 if (mss_now > tp->rx_opt.mss_clamp)
1146 mss_now = tp->rx_opt.mss_clamp;
1147
1148 /* Now subtract optional transport overhead */
1149 mss_now -= icsk->icsk_ext_hdr_len;
1150
1151 /* Then reserve room for full set of TCP options and 8 bytes of data */
1152 if (mss_now < 48)
1153 mss_now = 48;
1154
1155 /* Now subtract TCP options size, not including SACKs */
1156 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1157
1158 return mss_now;
1159}
1160
1161/* Inverse of above */
1162int tcp_mss_to_mtu(struct sock *sk, int mss)
1163{
1164 struct tcp_sock *tp = tcp_sk(sk);
1165 struct inet_connection_sock *icsk = inet_csk(sk);
1166 int mtu;
1167
1168 mtu = mss +
1169 tp->tcp_header_len +
1170 icsk->icsk_ext_hdr_len +
1171 icsk->icsk_af_ops->net_header_len;
1172
1173 return mtu;
1174}
1175
Andi Kleen67edfef2009-07-21 23:00:40 +00001176/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001177void tcp_mtup_init(struct sock *sk)
1178{
1179 struct tcp_sock *tp = tcp_sk(sk);
1180 struct inet_connection_sock *icsk = inet_csk(sk);
1181
1182 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1183 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001184 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -08001185 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1186 icsk->icsk_mtup.probe_size = 0;
1187}
1188
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189/* This function synchronize snd mss to current pmtu/exthdr set.
1190
1191 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1192 for TCP options, but includes only bare TCP header.
1193
1194 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001195 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 It also does not include TCP options.
1197
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001198 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
1200 tp->mss_cache is current effective sending mss, including
1201 all tcp options except for SACKs. It is evaluated,
1202 taking into account current pmtu, but never exceeds
1203 tp->rx_opt.mss_clamp.
1204
1205 NOTE1. rfc1122 clearly states that advertised MSS
1206 DOES NOT include either tcp or ip options.
1207
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001208 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1209 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1212{
1213 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001214 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001215 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216
John Heffner5d424d52006-03-20 17:53:41 -08001217 if (icsk->icsk_mtup.search_high > pmtu)
1218 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219
John Heffner5d424d52006-03-20 17:53:41 -08001220 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001221 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
1223 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001224 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001225 if (icsk->icsk_mtup.enabled)
1226 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001227 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228
1229 return mss_now;
1230}
1231
1232/* Compute the current effective MSS, taking SACKs and IP options,
1233 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001235unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236{
1237 struct tcp_sock *tp = tcp_sk(sk);
1238 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001239 u32 mss_now;
Adam Langley33ad7982008-07-19 00:04:31 -07001240 unsigned header_len;
1241 struct tcp_out_options opts;
1242 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001244 mss_now = tp->mss_cache;
1245
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 if (dst) {
1247 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001248 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 mss_now = tcp_sync_mss(sk, mtu);
1250 }
1251
Adam Langley33ad7982008-07-19 00:04:31 -07001252 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1253 sizeof(struct tcphdr);
1254 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1255 * some common options. If this is an odd packet (because we have SACK
1256 * blocks etc) then our calculated header_len will be different, and
1257 * we have to adjust mss_now correspondingly */
1258 if (header_len != tp->tcp_header_len) {
1259 int delta = (int) header_len - tp->tcp_header_len;
1260 mss_now -= delta;
1261 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001262
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 return mss_now;
1264}
1265
David S. Millera762a982005-07-05 15:18:51 -07001266/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001267static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001268{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001269 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001270
Ilpo Järvinend436d682007-12-31 14:58:00 -08001271 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001272 /* Network is feed fully. */
1273 tp->snd_cwnd_used = 0;
1274 tp->snd_cwnd_stamp = tcp_time_stamp;
1275 } else {
1276 /* Network starves. */
1277 if (tp->packets_out > tp->snd_cwnd_used)
1278 tp->snd_cwnd_used = tp->packets_out;
1279
David S. Miller15d33c02007-04-09 13:23:14 -07001280 if (sysctl_tcp_slow_start_after_idle &&
1281 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001282 tcp_cwnd_application_limited(sk);
1283 }
1284}
1285
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001286/* Returns the portion of skb which can be sent right away without
1287 * introducing MSS oddities to segment boundaries. In rare cases where
1288 * mss_now != mss_cache, we will request caller to create a small skb
1289 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001290 *
1291 * We explicitly want to create a request for splitting write queue tail
1292 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1293 * thus all the complexity (cwnd_len is always MSS multiple which we
1294 * return whenever allowed by the other factors). Basically we need the
1295 * modulo only when the receiver window alone is the limiting factor or
1296 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001297 */
1298static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001299 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001300{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001301 struct tcp_sock *tp = tcp_sk(sk);
1302 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001303
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001304 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001305 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001306
1307 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1308 return cwnd_len;
1309
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001310 needed = min(skb->len, window);
1311
Ilpo Järvinen17515402008-04-15 20:36:55 -07001312 if (cwnd_len <= needed)
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001313 return cwnd_len;
1314
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001315 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001316}
1317
1318/* Can at least one segment of SKB be sent right now, according to the
1319 * congestion window rules? If so, return how many segments are allowed.
1320 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001321static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1322 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001323{
1324 u32 in_flight, cwnd;
1325
1326 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001327 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1328 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001329 return 1;
1330
1331 in_flight = tcp_packets_in_flight(tp);
1332 cwnd = tp->snd_cwnd;
1333 if (in_flight < cwnd)
1334 return (cwnd - in_flight);
1335
1336 return 0;
1337}
1338
Andi Kleen67edfef2009-07-21 23:00:40 +00001339/* Intialize TSO state of a skb.
1340 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001341 * SKB onto the wire.
1342 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001343static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1344 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001345{
1346 int tso_segs = tcp_skb_pcount(skb);
1347
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001348 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001349 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001350 tso_segs = tcp_skb_pcount(skb);
1351 }
1352 return tso_segs;
1353}
1354
Andi Kleen67edfef2009-07-21 23:00:40 +00001355/* Minshall's variant of the Nagle send check. */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001356static inline int tcp_minshall_check(const struct tcp_sock *tp)
1357{
Jianjun Kong09cb1052008-11-03 00:27:11 -08001358 return after(tp->snd_sml, tp->snd_una) &&
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001359 !after(tp->snd_sml, tp->snd_nxt);
1360}
1361
1362/* Return 0, if packet can be sent now without violation Nagle's rules:
1363 * 1. It is full sized.
1364 * 2. Or it contains FIN. (already checked by caller)
1365 * 3. Or TCP_NODELAY was set.
1366 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1367 * With Minshall's modification: all sent small packets are ACKed.
1368 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001369static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001370 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001371 unsigned mss_now, int nonagle)
1372{
1373 return (skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001374 ((nonagle & TCP_NAGLE_CORK) ||
1375 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001376}
1377
1378/* Return non-zero if the Nagle test allows this packet to be
1379 * sent now.
1380 */
1381static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1382 unsigned int cur_mss, int nonagle)
1383{
1384 /* Nagle rule does not apply to frames, which sit in the middle of the
1385 * write_queue (they have no chances to get new data).
1386 *
1387 * This is implemented in the callers, where they modify the 'nonagle'
1388 * argument based upon the location of SKB in the send queue.
1389 */
1390 if (nonagle & TCP_NAGLE_PUSH)
1391 return 1;
1392
Ilpo Järvinend551e452007-04-30 00:42:20 -07001393 /* Don't use the nagle rule for urgent data (or for the final FIN).
1394 * Nagle can be ignored during F-RTO too (see RFC4138).
1395 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001396 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001397 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1398 return 1;
1399
1400 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1401 return 1;
1402
1403 return 0;
1404}
1405
1406/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001407static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1408 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001409{
1410 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1411
1412 if (skb->len > cur_mss)
1413 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1414
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001415 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001416}
1417
David S. Millerfe067e82007-03-07 12:12:44 -08001418/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001419 * should be put on the wire right now. If so, it returns the number of
1420 * packets allowed by the congestion window.
1421 */
1422static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1423 unsigned int cur_mss, int nonagle)
1424{
1425 struct tcp_sock *tp = tcp_sk(sk);
1426 unsigned int cwnd_quota;
1427
David S. Miller846998a2005-08-04 19:52:01 -07001428 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001429
1430 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1431 return 0;
1432
1433 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001434 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001435 cwnd_quota = 0;
1436
1437 return cwnd_quota;
1438}
1439
Andi Kleen67edfef2009-07-21 23:00:40 +00001440/* Test if sending is allowed right now. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001441int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001442{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001443 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001444 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001445
1446 return (skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001447 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001448 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001449 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001450}
1451
1452/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1453 * which is put after SKB on the list. It is very much like
1454 * tcp_fragment() except that it may make several kinds of assumptions
1455 * in order to speed up the splitting operation. In particular, we
1456 * know that all the data is in scatter-gather pages, and that the
1457 * packet has never been sent out before (and thus is not cloned).
1458 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001459static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1460 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001461{
1462 struct sk_buff *buff;
1463 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001464 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001465
1466 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001467 if (skb->len != skb->data_len)
1468 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001469
Pavel Emelyanovdf97c702007-11-29 21:22:33 +11001470 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001471 if (unlikely(buff == NULL))
1472 return -ENOMEM;
1473
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001474 sk->sk_wmem_queued += buff->truesize;
1475 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001476 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001477 skb->truesize -= nlen;
1478
1479 /* Correct the sequence numbers. */
1480 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1481 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1482 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1483
1484 /* PSH and FIN should only be set in the second packet. */
1485 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001486 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001487 TCP_SKB_CB(buff)->flags = flags;
1488
1489 /* This packet was never sent out yet, so no SACK bits. */
1490 TCP_SKB_CB(buff)->sacked = 0;
1491
Patrick McHardy84fa7932006-08-29 16:44:56 -07001492 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001493 skb_split(skb, buff, len);
1494
1495 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001496 tcp_set_skb_tso_segs(sk, skb, mss_now);
1497 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001498
1499 /* Link BUFF into the send queue. */
1500 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001501 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001502
1503 return 0;
1504}
1505
1506/* Try to defer sending, if possible, in order to minimize the amount
1507 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1508 *
1509 * This algorithm is from John Heffner.
1510 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001511static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001512{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001513 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001514 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001515 u32 send_win, cong_win, limit, in_flight;
1516
1517 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001518 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001519
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001520 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001521 goto send_now;
1522
1523 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001524 if (tp->tso_deferred &&
Ilpo Järvinena2acde02008-12-05 22:49:37 -08001525 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001526 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001527
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001528 in_flight = tcp_packets_in_flight(tp);
1529
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001530 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001531
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001532 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001533
1534 /* From in_flight test above, we know that cwnd > in_flight. */
1535 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1536
1537 limit = min(send_win, cong_win);
1538
David S. Millerba244fe2006-03-11 18:51:49 -08001539 /* If a full-sized TSO skb can be sent, do it. */
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001540 if (limit >= sk->sk_gso_max_size)
John Heffnerae8064a2006-10-18 20:36:48 -07001541 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001542
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001543 /* Middle in queue won't get any more data, full sendable already? */
1544 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1545 goto send_now;
1546
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001547 if (sysctl_tcp_tso_win_divisor) {
1548 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1549
1550 /* If at least some fraction of a window is available,
1551 * just use it.
1552 */
1553 chunk /= sysctl_tcp_tso_win_divisor;
1554 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001555 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001556 } else {
1557 /* Different approach, try not to defer past a single
1558 * ACK. Receiver should ACK every other full sized
1559 * frame, so if we have space for more than 3 frames
1560 * then send now.
1561 */
1562 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001563 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001564 }
1565
1566 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001567 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001568
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001569 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001570
1571send_now:
1572 tp->tso_deferred = 0;
1573 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001574}
1575
John Heffner5d424d52006-03-20 17:53:41 -08001576/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001577 * MTU probe is regularly attempting to increase the path MTU by
1578 * deliberately sending larger packets. This discovers routing
1579 * changes resulting in larger path MTUs.
1580 *
John Heffner5d424d52006-03-20 17:53:41 -08001581 * Returns 0 if we should wait to probe (no cwnd available),
1582 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001583 * -1 otherwise
1584 */
John Heffner5d424d52006-03-20 17:53:41 -08001585static int tcp_mtu_probe(struct sock *sk)
1586{
1587 struct tcp_sock *tp = tcp_sk(sk);
1588 struct inet_connection_sock *icsk = inet_csk(sk);
1589 struct sk_buff *skb, *nskb, *next;
1590 int len;
1591 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001592 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001593 int copy;
1594 int mss_now;
1595
1596 /* Not currently probing/verifying,
1597 * not in recovery,
1598 * have enough cwnd, and
1599 * not SACKing (the variable headers throw things off) */
1600 if (!icsk->icsk_mtup.enabled ||
1601 icsk->icsk_mtup.probe_size ||
1602 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1603 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001604 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001605 return -1;
1606
1607 /* Very simple search strategy: just double the MSS. */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001608 mss_now = tcp_current_mss(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001609 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001610 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001611 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1612 /* TODO: set timer for probe_converge_event */
1613 return -1;
1614 }
1615
1616 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001617 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001618 return -1;
1619
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001620 if (tp->snd_wnd < size_needed)
1621 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001622 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001623 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001624
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001625 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1626 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1627 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001628 return -1;
1629 else
1630 return 0;
1631 }
1632
1633 /* We're allowed to probe. Build it now. */
1634 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1635 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001636 sk->sk_wmem_queued += nskb->truesize;
1637 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001638
David S. Millerfe067e82007-03-07 12:12:44 -08001639 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001640
1641 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1642 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1643 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1644 TCP_SKB_CB(nskb)->sacked = 0;
1645 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001646 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001647
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001648 tcp_insert_write_queue_before(nskb, skb, sk);
1649
John Heffner5d424d52006-03-20 17:53:41 -08001650 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001651 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001652 copy = min_t(int, skb->len, probe_size - len);
1653 if (nskb->ip_summed)
1654 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1655 else
1656 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001657 skb_put(nskb, copy),
1658 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001659
1660 if (skb->len <= copy) {
1661 /* We've eaten all the data from this skb.
1662 * Throw it away. */
1663 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001664 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001665 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001666 } else {
1667 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001668 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001669 if (!skb_shinfo(skb)->nr_frags) {
1670 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001671 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001672 skb->csum = csum_partial(skb->data,
1673 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001674 } else {
1675 __pskb_trim_head(skb, copy);
1676 tcp_set_skb_tso_segs(sk, skb, mss_now);
1677 }
1678 TCP_SKB_CB(skb)->seq += copy;
1679 }
1680
1681 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001682
1683 if (len >= probe_size)
1684 break;
John Heffner5d424d52006-03-20 17:53:41 -08001685 }
1686 tcp_init_tso_segs(sk, nskb, nskb->len);
1687
1688 /* We're ready to send. If this fails, the probe will
1689 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1690 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1691 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1692 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001693 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001694 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001695 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001696
1697 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001698 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1699 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001700
1701 return 1;
1702 }
1703
1704 return -1;
1705}
1706
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707/* This routine writes packets to the network. It advances the
1708 * send_head. This happens as incoming acks open up the remote
1709 * window for us.
1710 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001711 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1712 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1713 * account rare use of URG, this is not a big flaw.
1714 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 * Returns 1, if no segments are in flight and we have queued segments, but
1716 * cannot send anything now because of SWS or another problem.
1717 */
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001718static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1719 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720{
1721 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001722 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001723 unsigned int tso_segs, sent_pkts;
1724 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001725 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726
David S. Miller92df7b52005-07-05 15:19:06 -07001727 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001728
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001729 if (!push_one) {
1730 /* Do MTU probing. */
1731 result = tcp_mtu_probe(sk);
1732 if (!result) {
1733 return 0;
1734 } else if (result > 0) {
1735 sent_pkts = 1;
1736 }
John Heffner5d424d52006-03-20 17:53:41 -08001737 }
1738
David S. Millerfe067e82007-03-07 12:12:44 -08001739 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001740 unsigned int limit;
1741
Herbert Xub68e9f82005-08-04 19:52:02 -07001742 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001743 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001744
Herbert Xub68e9f82005-08-04 19:52:02 -07001745 cwnd_quota = tcp_cwnd_test(tp, skb);
1746 if (!cwnd_quota)
1747 break;
1748
1749 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1750 break;
1751
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001752 if (tso_segs == 1) {
1753 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1754 (tcp_skb_is_last(sk, skb) ?
1755 nonagle : TCP_NAGLE_PUSH))))
1756 break;
1757 } else {
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001758 if (!push_one && tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001759 break;
1760 }
David S. Milleraa934662005-07-05 15:20:09 -07001761
Herbert Xuc8ac3772005-08-16 20:43:40 -07001762 limit = mss_now;
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001763 if (tso_segs > 1 && !tcp_urg_mode(tp))
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001764 limit = tcp_mss_split_point(sk, skb, mss_now,
1765 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766
Herbert Xuc8ac3772005-08-16 20:43:40 -07001767 if (skb->len > limit &&
1768 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1769 break;
1770
David S. Miller92df7b52005-07-05 15:19:06 -07001771 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001772
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001773 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07001774 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775
David S. Miller92df7b52005-07-05 15:19:06 -07001776 /* Advance the send_head. This one is sent out.
1777 * This call will increment packets_out.
1778 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001779 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001780
1781 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001782 sent_pkts++;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001783
1784 if (push_one)
1785 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 }
David S. Miller92df7b52005-07-05 15:19:06 -07001787
David S. Milleraa934662005-07-05 15:20:09 -07001788 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001789 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001790 return 0;
1791 }
David S. Millerfe067e82007-03-07 12:12:44 -08001792 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793}
1794
David S. Millera762a982005-07-05 15:18:51 -07001795/* Push out any pending frames which were held back due to
1796 * TCP_CORK or attempt at coalescing tiny packets.
1797 * The socket must be locked by the caller.
1798 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001799void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1800 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001801{
David S. Millerfe067e82007-03-07 12:12:44 -08001802 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001803
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001804 if (!skb)
1805 return;
1806
1807 /* If we are closed, the bytes will have to remain here.
1808 * In time closedown will finish, we empty the write queue and
1809 * all will be happy.
1810 */
1811 if (unlikely(sk->sk_state == TCP_CLOSE))
1812 return;
1813
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001814 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001815 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001816}
1817
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001818/* Send _single_ skb sitting at the send head. This function requires
1819 * true push pending frames to setup probe timer etc.
1820 */
1821void tcp_push_one(struct sock *sk, unsigned int mss_now)
1822{
David S. Millerfe067e82007-03-07 12:12:44 -08001823 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001824
1825 BUG_ON(!skb || skb->len < mss_now);
1826
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001827 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001828}
1829
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830/* This function returns the amount that we can raise the
1831 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001832 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 * 1. The window can never be shrunk once it is offered (RFC 793)
1834 * 2. We limit memory per socket
1835 *
1836 * RFC 1122:
1837 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1838 * RECV.NEXT + RCV.WIN fixed until:
1839 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1840 *
1841 * i.e. don't raise the right edge of the window until you can raise
1842 * it at least MSS bytes.
1843 *
1844 * Unfortunately, the recommended algorithm breaks header prediction,
1845 * since header prediction assumes th->window stays fixed.
1846 *
1847 * Strictly speaking, keeping th->window fixed violates the receiver
1848 * side SWS prevention criteria. The problem is that under this rule
1849 * a stream of single byte packets will cause the right side of the
1850 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001851 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 * Of course, if the sender implements sender side SWS prevention
1853 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001854 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001856 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 * If the free space is less than the 1/4 of the maximum
1858 * space available and the free space is less than 1/2 mss,
1859 * then set the window to 0.
1860 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1861 * Otherwise, just prevent the window from shrinking
1862 * and from being larger than the largest representable value.
1863 *
1864 * This prevents incremental opening of the window in the regime
1865 * where TCP is limited by the speed of the reader side taking
1866 * data out of the TCP receive queue. It does nothing about
1867 * those cases where the window is constrained on the sender side
1868 * because the pipeline is full.
1869 *
1870 * BSD also seems to "accidentally" limit itself to windows that are a
1871 * multiple of MSS, at least until the free space gets quite small.
1872 * This would appear to be a side effect of the mbuf implementation.
1873 * Combining these two algorithms results in the observed behavior
1874 * of having a fixed window size at almost all times.
1875 *
1876 * Below we obtain similar behavior by forcing the offered window to
1877 * a multiple of the mss when it is feasible to do so.
1878 *
1879 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1880 * Regular options like TIMESTAMP are taken into account.
1881 */
1882u32 __tcp_select_window(struct sock *sk)
1883{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001884 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001886 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 * here. I don't know if the value based on our guesses
1888 * of peer's MSS is better for the performance. It's more correct
1889 * but may be worse for the performance because of rcv_mss
1890 * fluctuations. --SAW 1998/11/1
1891 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001892 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 int free_space = tcp_space(sk);
1894 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1895 int window;
1896
1897 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001898 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001900 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001901 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902
1903 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001904 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1905 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906
1907 if (free_space < mss)
1908 return 0;
1909 }
1910
1911 if (free_space > tp->rcv_ssthresh)
1912 free_space = tp->rcv_ssthresh;
1913
1914 /* Don't do rounding if we are using window scaling, since the
1915 * scaled window will not line up with the MSS boundary anyway.
1916 */
1917 window = tp->rcv_wnd;
1918 if (tp->rx_opt.rcv_wscale) {
1919 window = free_space;
1920
1921 /* Advertise enough space so that it won't get scaled away.
1922 * Import case: prevent zero window announcement if
1923 * 1<<rcv_wscale > mss.
1924 */
1925 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1926 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1927 << tp->rx_opt.rcv_wscale);
1928 } else {
1929 /* Get the largest window that is a nice multiple of mss.
1930 * Window clamp already applied above.
1931 * If our current window offering is within 1 mss of the
1932 * free space we just keep it. This prevents the divide
1933 * and multiply from happening most of the time.
1934 * We also don't do any window rounding when the free space
1935 * is too small.
1936 */
1937 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001938 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001939 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001940 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001941 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 }
1943
1944 return window;
1945}
1946
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08001947/* Collapses two adjacent SKB's during retransmission. */
1948static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949{
1950 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001951 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001952 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001954 skb_size = skb->len;
1955 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001957 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001958
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001959 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001961 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001963 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1964 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001966 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1967 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001969 if (skb->ip_summed != CHECKSUM_PARTIAL)
1970 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001972 /* Update sequence range on original skb. */
1973 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00001975 /* Merge over control information. This moves PSH/FIN etc. over */
1976 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001978 /* All done, get rid of second SKB and account for it so
1979 * packet counting does not break.
1980 */
1981 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001982
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001983 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07001984 tcp_clear_retrans_hints_partial(tp);
1985 if (next_skb == tp->retransmit_skb_hint)
1986 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001987
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001988 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
1989
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001990 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991}
1992
Andi Kleen67edfef2009-07-21 23:00:40 +00001993/* Check if coalescing SKBs is legal. */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08001994static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1995{
1996 if (tcp_skb_pcount(skb) > 1)
1997 return 0;
1998 /* TODO: SACK collapsing could be used to remove this condition */
1999 if (skb_shinfo(skb)->nr_frags != 0)
2000 return 0;
2001 if (skb_cloned(skb))
2002 return 0;
2003 if (skb == tcp_send_head(sk))
2004 return 0;
2005 /* Some heurestics for collapsing over SACK'd could be invented */
2006 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2007 return 0;
2008
2009 return 1;
2010}
2011
Andi Kleen67edfef2009-07-21 23:00:40 +00002012/* Collapse packets in the retransmit queue to make to create
2013 * less packets on the wire. This is only done on retransmission.
2014 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002015static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2016 int space)
2017{
2018 struct tcp_sock *tp = tcp_sk(sk);
2019 struct sk_buff *skb = to, *tmp;
2020 int first = 1;
2021
2022 if (!sysctl_tcp_retrans_collapse)
2023 return;
2024 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
2025 return;
2026
2027 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2028 if (!tcp_can_collapse(sk, skb))
2029 break;
2030
2031 space -= skb->len;
2032
2033 if (first) {
2034 first = 0;
2035 continue;
2036 }
2037
2038 if (space < 0)
2039 break;
2040 /* Punt if not enough space exists in the first SKB for
2041 * the data in the second
2042 */
2043 if (skb->len > skb_tailroom(to))
2044 break;
2045
2046 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2047 break;
2048
2049 tcp_collapse_retrans(sk, to);
2050 }
2051}
2052
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053/* This retransmits one SKB. Policy decisions and retransmit queue
2054 * state updates are done by the caller. Returns non-zero if an
2055 * error occurred which prevented the send.
2056 */
2057int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2058{
2059 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002060 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002061 unsigned int cur_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 int err;
2063
John Heffner5d424d52006-03-20 17:53:41 -08002064 /* Inconslusive MTU probe */
2065 if (icsk->icsk_mtup.probe_size) {
2066 icsk->icsk_mtup.probe_size = 0;
2067 }
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002070 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 */
2072 if (atomic_read(&sk->sk_wmem_alloc) >
2073 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2074 return -EAGAIN;
2075
2076 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2077 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2078 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2080 return -ENOMEM;
2081 }
2082
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002083 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2084 return -EHOSTUNREACH; /* Routing failure or similar. */
2085
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002086 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 /* If receiver has shrunk his window, and skb is out of
2089 * new window, do not retransmit it. The exception is the
2090 * case, when window is shrunk to zero. In this case
2091 * our retransmit serves as a zero window probe.
2092 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002093 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2094 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 return -EAGAIN;
2096
2097 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07002098 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002100 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002101 int oldpcount = tcp_skb_pcount(skb);
2102
2103 if (unlikely(oldpcount > 1)) {
2104 tcp_init_tso_segs(sk, skb, cur_mss);
2105 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 }
2108
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002109 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111 /* Some Solaris stacks overoptimize and ignore the FIN on a
2112 * retransmit when old data is attached. So strip it off
2113 * since it is cheap to do so and saves bytes on the network.
2114 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002115 if (skb->len > 0 &&
2116 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
2117 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002119 /* Reuse, even though it does some unnecessary work */
2120 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2121 TCP_SKB_CB(skb)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 }
2124 }
2125
2126 /* Make a copy, if the first transmission SKB clone we made
2127 * is still in somebody's hands, else make a clone.
2128 */
2129 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002131 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
2133 if (err == 0) {
2134 /* Update global TCP statistics. */
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002135 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136
2137 tp->total_retrans++;
2138
2139#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002140 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 if (net_ratelimit())
2142 printk(KERN_DEBUG "retrans_out leaked.\n");
2143 }
2144#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002145 if (!tp->retrans_out)
2146 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2148 tp->retrans_out += tcp_skb_pcount(skb);
2149
2150 /* Save stamp of the first retransmit. */
2151 if (!tp->retrans_stamp)
2152 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2153
2154 tp->undo_retrans++;
2155
2156 /* snd_nxt is stored to detect loss of retransmitted segment,
2157 * see tcp_input.c tcp_sacktag_write_queue().
2158 */
2159 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2160 }
2161 return err;
2162}
2163
Andi Kleen67edfef2009-07-21 23:00:40 +00002164/* Check if we forward retransmits are possible in the current
2165 * window/congestion state.
2166 */
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002167static int tcp_can_forward_retransmit(struct sock *sk)
2168{
2169 const struct inet_connection_sock *icsk = inet_csk(sk);
2170 struct tcp_sock *tp = tcp_sk(sk);
2171
2172 /* Forward retransmissions are possible only during Recovery. */
2173 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2174 return 0;
2175
2176 /* No forward retransmissions in Reno are possible. */
2177 if (tcp_is_reno(tp))
2178 return 0;
2179
2180 /* Yeah, we have to make difficult choice between forward transmission
2181 * and retransmission... Both ways have their merits...
2182 *
2183 * For now we do not retransmit anything, while we have some new
2184 * segments to send. In the other cases, follow rule 3 for
2185 * NextSeg() specified in RFC3517.
2186 */
2187
2188 if (tcp_may_send_now(sk))
2189 return 0;
2190
2191 return 1;
2192}
2193
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194/* This gets called after a retransmit timeout, and the initially
2195 * retransmitted data is acknowledged. It tries to continue
2196 * resending the rest of the retransmit queue, until either
2197 * we've sent it all or the congestion window limit is reached.
2198 * If doing SACK, the first ACK which comes back for a timeout
2199 * based retransmit packet might feed us FACK information again.
2200 * If so, we use it to avoid unnecessarily retransmissions.
2201 */
2202void tcp_xmit_retransmit_queue(struct sock *sk)
2203{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002204 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 struct tcp_sock *tp = tcp_sk(sk);
2206 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002207 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002208 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002209 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002210 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002211
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002212 if (!tp->lost_out)
2213 tp->retransmit_high = tp->snd_una;
2214
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002215 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002216 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002217 last_lost = TCP_SKB_CB(skb)->end_seq;
2218 if (after(last_lost, tp->retransmit_high))
2219 last_lost = tp->retransmit_high;
2220 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002221 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002222 last_lost = tp->snd_una;
2223 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002225 tcp_for_write_queue_from(skb, sk) {
2226 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002228 if (skb == tcp_send_head(sk))
2229 break;
2230 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002231 if (hole == NULL)
2232 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002233
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002234 /* Assume this retransmit will generate
2235 * only one packet for congestion window
2236 * calculation purposes. This works because
2237 * tcp_retransmit_skb() will chop up the
2238 * packet to be MSS sized and all the
2239 * packet counting works out.
2240 */
2241 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2242 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002243
2244 if (fwd_rexmitting) {
2245begin_fwd:
2246 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2247 break;
2248 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2249
2250 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002251 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002252 if (!tcp_can_forward_retransmit(sk))
2253 break;
2254 /* Backtrack if necessary to non-L'ed skb */
2255 if (hole != NULL) {
2256 skb = hole;
2257 hole = NULL;
2258 }
2259 fwd_rexmitting = 1;
2260 goto begin_fwd;
2261
2262 } else if (!(sacked & TCPCB_LOST)) {
Ilpo Järvinenac11ba72009-02-28 04:44:27 +00002263 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002264 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002265 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002267 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002268 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002269 if (icsk->icsk_ca_state != TCP_CA_Loss)
2270 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2271 else
2272 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2273 }
2274
2275 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002276 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002277
Ilpo Järvinenf0ceb0e2008-09-20 21:24:49 -07002278 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002279 return;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002280 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2281
2282 if (skb == tcp_write_queue_head(sk))
2283 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2284 inet_csk(sk)->icsk_rto,
2285 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287}
2288
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289/* Send a fin. The caller locks the socket for us. This cannot be
2290 * allowed to fail queueing a FIN frame under any circumstances.
2291 */
2292void tcp_send_fin(struct sock *sk)
2293{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002294 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002295 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002297
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 /* Optimization, tack on the FIN if we have a queue of
2299 * unsent frames. But be careful about outgoing SACKS
2300 * and IP options.
2301 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002302 mss_now = tcp_current_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303
David S. Millerfe067e82007-03-07 12:12:44 -08002304 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2306 TCP_SKB_CB(skb)->end_seq++;
2307 tp->write_seq++;
2308 } else {
2309 /* Socket is locked, keep trying until memory is available. */
2310 for (;;) {
Wu Fengguangaa133072009-09-02 23:45:45 -07002311 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2312 sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 if (skb)
2314 break;
2315 yield();
2316 }
2317
2318 /* Reserve space for headers and prepare control bits. */
2319 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002321 tcp_init_nondata_skb(skb, tp->write_seq,
2322 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 tcp_queue_skb(sk, skb);
2324 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002325 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326}
2327
2328/* We get here when a process closes a file descriptor (either due to
2329 * an explicit close() or as a byproduct of exit()'ing) and there
2330 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002331 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 */
Al Virodd0fc662005-10-07 07:46:04 +01002333void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 struct sk_buff *skb;
2336
2337 /* NOTE: No TCP options attached and we never retransmit this. */
2338 skb = alloc_skb(MAX_TCP_HEADER, priority);
2339 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002340 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 return;
2342 }
2343
2344 /* Reserve space for headers and prepare control bits. */
2345 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002346 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2347 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002350 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002351 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002352
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002353 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354}
2355
Andi Kleen67edfef2009-07-21 23:00:40 +00002356/* Send a crossed SYN-ACK during socket establishment.
2357 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 * a SYN packet that crossed the incoming SYN that caused this routine
2359 * to get called. If this assumption fails then the initial rcv_wnd
2360 * and rcv_wscale values will not be correct.
2361 */
2362int tcp_send_synack(struct sock *sk)
2363{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002364 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365
David S. Millerfe067e82007-03-07 12:12:44 -08002366 skb = tcp_write_queue_head(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002367 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2369 return -EFAULT;
2370 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002371 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 if (skb_cloned(skb)) {
2373 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2374 if (nskb == NULL)
2375 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002376 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002378 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002379 sk_wmem_free_skb(sk, skb);
2380 sk->sk_wmem_queued += nskb->truesize;
2381 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 skb = nskb;
2383 }
2384
2385 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2386 TCP_ECN_send_synack(tcp_sk(sk), skb);
2387 }
2388 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002389 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390}
2391
Andi Kleen67edfef2009-07-21 23:00:40 +00002392/* Prepare a SYN-ACK. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002393struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002394 struct request_sock *req,
2395 struct request_values *rvp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002397 struct tcp_out_options opts;
William Allen Simpson4957faa2009-12-02 18:25:27 +00002398 struct tcp_extend_values *xvp = tcp_xv(rvp);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002399 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 struct tcp_sock *tp = tcp_sk(sk);
2401 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002403 struct tcp_md5sig_key *md5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002404 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002405 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
2407 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2408 if (skb == NULL)
2409 return NULL;
2410
2411 /* Reserve space for headers. */
2412 skb_reserve(skb, MAX_TCP_HEADER);
2413
Eric Dumazetadf30902009-06-02 05:19:30 +00002414 skb_dst_set(skb, dst_clone(dst));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002416 mss = dst_metric(dst, RTAX_ADVMSS);
2417 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2418 mss = tp->rx_opt.user_mss;
2419
Adam Langley33ad7982008-07-19 00:04:31 -07002420 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2421 __u8 rcv_wscale;
2422 /* Set this up on the first call only */
2423 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2424 /* tcp_full_space because it is guaranteed to be the first packet */
2425 tcp_select_initial_window(tcp_full_space(sk),
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002426 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Adam Langley33ad7982008-07-19 00:04:31 -07002427 &req->rcv_wnd,
2428 &req->window_clamp,
2429 ireq->wscale_ok,
2430 &rcv_wscale);
2431 ireq->rcv_wscale = rcv_wscale;
2432 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002433
Adam Langley33ad7982008-07-19 00:04:31 -07002434 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002435#ifdef CONFIG_SYN_COOKIES
2436 if (unlikely(req->cookie_ts))
2437 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2438 else
2439#endif
Adam Langley33ad7982008-07-19 00:04:31 -07002440 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002441 tcp_header_size = tcp_synack_options(sk, req, mss,
William Allen Simpson4957faa2009-12-02 18:25:27 +00002442 skb, &opts, &md5, xvp)
2443 + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002444
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002445 skb_push(skb, tcp_header_size);
2446 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002448 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 memset(th, 0, sizeof(struct tcphdr));
2450 th->syn = 1;
2451 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 TCP_ECN_make_synack(req, th);
KOVACS Krisztiana3116ac2008-10-01 07:46:49 -07002453 th->source = ireq->loc_port;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002454 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002455 /* Setting of flags are superfluous here for callers (and ECE is
2456 * not even correctly set)
2457 */
2458 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2459 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
William Allen Simpson4957faa2009-12-02 18:25:27 +00002460
2461 if (OPTION_COOKIE_EXTENSION & opts.options) {
2462 const struct tcp_cookie_values *cvp = tp->cookie_values;
2463
2464 if (cvp != NULL &&
2465 cvp->s_data_constant &&
2466 cvp->s_data_desired > 0) {
2467 u8 *buf = skb_put(skb, cvp->s_data_desired);
2468
2469 /* copy data directly from the listening socket. */
2470 memcpy(buf, cvp->s_data_payload, cvp->s_data_desired);
2471 TCP_SKB_CB(skb)->end_seq += cvp->s_data_desired;
2472 }
2473
2474 if (opts.hash_size > 0) {
2475 __u32 workspace[SHA_WORKSPACE_WORDS];
2476 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2477 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2478
2479 /* Secret recipe depends on the Timestamp, (future)
2480 * Sequence and Acknowledgment Numbers, Initiator
2481 * Cookie, and others handled by IP variant caller.
2482 */
2483 *tail-- ^= opts.tsval;
2484 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2485 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2486
2487 /* recommended */
2488 *tail-- ^= ((th->dest << 16) | th->source);
David S. Millerf9a2e692009-12-02 22:12:04 -08002489 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
William Allen Simpson4957faa2009-12-02 18:25:27 +00002490
2491 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2492 (char *)mess,
2493 &workspace[0]);
2494 opts.hash_location =
2495 (__u8 *)&xvp->cookie_bakery[0];
2496 }
2497 }
2498
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002500 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501
2502 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002503 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002504 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 th->doff = (tcp_header_size >> 2);
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002506 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002507
2508#ifdef CONFIG_TCP_MD5SIG
2509 /* Okay, we have all we need - do the md5 hash if needed */
2510 if (md5) {
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002511 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002512 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002513 }
2514#endif
2515
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 return skb;
2517}
2518
Andi Kleen67edfef2009-07-21 23:00:40 +00002519/* Do all connect socket setups that can be done AF independent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002520static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
2522 struct dst_entry *dst = __sk_dst_get(sk);
2523 struct tcp_sock *tp = tcp_sk(sk);
2524 __u8 rcv_wscale;
2525
2526 /* We'll fix this up when we get a response from the other end.
2527 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2528 */
2529 tp->tcp_header_len = sizeof(struct tcphdr) +
Gilad Ben-Yossefcda42eb2009-10-28 04:15:25 +00002530 (sysctl_tcp_timestamps &&
2531 (!dst_feature(dst, RTAX_FEATURE_NO_TSTAMP) ?
2532 TCPOLEN_TSTAMP_ALIGNED : 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002534#ifdef CONFIG_TCP_MD5SIG
2535 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2536 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2537#endif
2538
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 /* If user gave his TCP_MAXSEG, record it to clamp */
2540 if (tp->rx_opt.user_mss)
2541 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2542 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002543 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 tcp_sync_mss(sk, dst_mtu(dst));
2545
2546 if (!tp->window_clamp)
2547 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2548 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002549 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2550 tp->advmss = tp->rx_opt.user_mss;
2551
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553
2554 tcp_select_initial_window(tcp_full_space(sk),
2555 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2556 &tp->rcv_wnd,
2557 &tp->window_clamp,
Gilad Ben-Yossef345cda22009-10-28 04:15:26 +00002558 (sysctl_tcp_window_scaling &&
2559 !dst_feature(dst, RTAX_FEATURE_NO_WSCALE)),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 &rcv_wscale);
2561
2562 tp->rx_opt.rcv_wscale = rcv_wscale;
2563 tp->rcv_ssthresh = tp->rcv_wnd;
2564
2565 sk->sk_err = 0;
2566 sock_reset_flag(sk, SOCK_DONE);
2567 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08002568 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 tp->snd_una = tp->write_seq;
2570 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002571 tp->snd_up = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 tp->rcv_nxt = 0;
2573 tp->rcv_wup = 0;
2574 tp->copied_seq = 0;
2575
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002576 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2577 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 tcp_clear_retrans(tp);
2579}
2580
Andi Kleen67edfef2009-07-21 23:00:40 +00002581/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582int tcp_connect(struct sock *sk)
2583{
2584 struct tcp_sock *tp = tcp_sk(sk);
2585 struct sk_buff *buff;
2586
2587 tcp_connect_init(sk);
2588
David S. Millerd179cd12005-08-17 14:57:30 -07002589 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 if (unlikely(buff == NULL))
2591 return -ENOBUFS;
2592
2593 /* Reserve space for headers. */
2594 skb_reserve(buff, MAX_TCP_HEADER);
2595
Wei Yongjunbd37a082006-08-07 21:04:15 -07002596 tp->snd_nxt = tp->write_seq;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002597 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2598 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599
2600 /* Send it off. */
2601 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2602 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2603 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002604 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002605 sk->sk_wmem_queued += buff->truesize;
2606 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 tp->packets_out += tcp_skb_pcount(buff);
Wu Fengguangaa133072009-09-02 23:45:45 -07002608 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002609
2610 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2611 * in order to make this packet get counted in tcpOutSegs.
2612 */
2613 tp->snd_nxt = tp->write_seq;
2614 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002615 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616
2617 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002618 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2619 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 return 0;
2621}
2622
2623/* Send out a delayed ack, the caller does the policy checking
2624 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2625 * for details.
2626 */
2627void tcp_send_delayed_ack(struct sock *sk)
2628{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002629 struct inet_connection_sock *icsk = inet_csk(sk);
2630 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 unsigned long timeout;
2632
2633 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002634 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002635 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002637 if (icsk->icsk_ack.pingpong ||
2638 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639 max_ato = TCP_DELACK_MAX;
2640
2641 /* Slow path, intersegment interval is "high". */
2642
2643 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002644 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 * directly.
2646 */
2647 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002648 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649
2650 if (rtt < max_ato)
2651 max_ato = rtt;
2652 }
2653
2654 ato = min(ato, max_ato);
2655 }
2656
2657 /* Stay within the limit we were given */
2658 timeout = jiffies + ato;
2659
2660 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002661 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662 /* If delack timer was blocked or is about to expire,
2663 * send ACK now.
2664 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002665 if (icsk->icsk_ack.blocked ||
2666 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 tcp_send_ack(sk);
2668 return;
2669 }
2670
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002671 if (!time_before(timeout, icsk->icsk_ack.timeout))
2672 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002674 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2675 icsk->icsk_ack.timeout = timeout;
2676 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677}
2678
2679/* This routine sends an ack and also updates the window. */
2680void tcp_send_ack(struct sock *sk)
2681{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002682 struct sk_buff *buff;
2683
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002685 if (sk->sk_state == TCP_CLOSE)
2686 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002688 /* We are not putting this on the write queue, so
2689 * tcp_transmit_skb() will set the ownership to this
2690 * sock.
2691 */
2692 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2693 if (buff == NULL) {
2694 inet_csk_schedule_ack(sk);
2695 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2696 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2697 TCP_DELACK_MAX, TCP_RTO_MAX);
2698 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002700
2701 /* Reserve space for headers and prepare control bits. */
2702 skb_reserve(buff, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002703 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002704
2705 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002706 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2707 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708}
2709
2710/* This routine sends a packet with an out of date sequence
2711 * number. It assumes the other end will try to ack it.
2712 *
2713 * Question: what should we make while urgent mode?
2714 * 4.4BSD forces sending single byte of data. We cannot send
2715 * out of window data, because we have SND.NXT==SND.MAX...
2716 *
2717 * Current solution: to send TWO zero-length segments in urgent mode:
2718 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2719 * out-of-date with SND.UNA-1 to probe window.
2720 */
2721static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2722{
2723 struct tcp_sock *tp = tcp_sk(sk);
2724 struct sk_buff *skb;
2725
2726 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2727 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002728 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 return -1;
2730
2731 /* Reserve space for headers and set control bits. */
2732 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733 /* Use a previous sequence. This should cause the other
2734 * end to send an ack. Don't queue or clone SKB, just
2735 * send it.
2736 */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002737 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002739 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740}
2741
Andi Kleen67edfef2009-07-21 23:00:40 +00002742/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743int tcp_write_wakeup(struct sock *sk)
2744{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002745 struct tcp_sock *tp = tcp_sk(sk);
2746 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002748 if (sk->sk_state == TCP_CLOSE)
2749 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002751 if ((skb = tcp_send_head(sk)) != NULL &&
2752 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2753 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002754 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002755 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002757 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2758 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002760 /* We are probing the opening of a window
2761 * but the window size is != 0
2762 * must have been a result SWS avoidance ( sender )
2763 */
2764 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2765 skb->len > mss) {
2766 seg_size = min(seg_size, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002768 if (tcp_fragment(sk, skb, seg_size, mss))
2769 return -1;
2770 } else if (!tcp_skb_pcount(skb))
2771 tcp_set_skb_tso_segs(sk, skb, mss);
2772
2773 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2774 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2775 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2776 if (!err)
2777 tcp_event_new_data_sent(sk, skb);
2778 return err;
2779 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002780 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002781 tcp_xmit_probe_skb(sk, 1);
2782 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784}
2785
2786/* A window probe timeout has occurred. If window is not closed send
2787 * a partial packet else a zero probe.
2788 */
2789void tcp_send_probe0(struct sock *sk)
2790{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002791 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 struct tcp_sock *tp = tcp_sk(sk);
2793 int err;
2794
2795 err = tcp_write_wakeup(sk);
2796
David S. Millerfe067e82007-03-07 12:12:44 -08002797 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002799 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002800 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 return;
2802 }
2803
2804 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002805 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2806 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002807 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002808 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002809 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2810 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 } else {
2812 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002813 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 * Let local senders to fight for local resources.
2815 *
2816 * Use accumulated backoff yet.
2817 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002818 if (!icsk->icsk_probes_out)
2819 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002820 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002821 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002822 TCP_RESOURCE_PROBE_INTERVAL),
2823 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 }
2825}
2826
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08002827EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828EXPORT_SYMBOL(tcp_connect);
2829EXPORT_SYMBOL(tcp_make_synack);
2830EXPORT_SYMBOL(tcp_simple_retransmit);
2831EXPORT_SYMBOL(tcp_sync_mss);
John Heffner5d424d52006-03-20 17:53:41 -08002832EXPORT_SYMBOL(tcp_mtup_init);