blob: 406f320336e6591db59197fd4b75e47534b41ef7 [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090040#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
43/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070044int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
Jianjun Kong09cb1052008-11-03 00:27:11 -080046/* People can turn this on to work with those rare, broken TCPs that
Rick Jones15d99e02006-03-20 22:40:29 -080047 * interpret the window field as a signed quantity.
48 */
Brian Haleyab32ea52006-09-22 14:15:41 -070049int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080050
Linus Torvalds1da177e2005-04-16 15:20:36 -070051/* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
54 */
Brian Haleyab32ea52006-09-22 14:15:41 -070055int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
Brian Haleyab32ea52006-09-22 14:15:41 -070057int sysctl_tcp_mtu_probing __read_mostly = 0;
Shan Wei97b1ce22010-12-01 18:04:50 +000058int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
John Heffner5d424d52006-03-20 17:53:41 -080059
David S. Miller35089bb2006-06-13 22:33:04 -070060/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070061int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070062
William Allen Simpson519855c2009-12-02 18:14:19 +000063int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
David S. Millere6b09cc2009-12-02 22:18:58 -080064EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
William Allen Simpson519855c2009-12-02 18:14:19 +000065
66
Andi Kleen67edfef2009-07-21 23:00:40 +000067/* Account for new data that has been sent to the network. */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080068static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070069{
70 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080071 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070072
David S. Millerfe067e82007-03-07 12:12:44 -080073 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080075
76 /* Don't override Nagle indefinately with F-RTO */
77 if (tp->frto_counter == 2)
78 tp->frto_counter = 3;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080079
80 tp->packets_out += tcp_skb_pcount(skb);
81 if (!prior_packets)
82 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
83 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -070084}
85
86/* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
91 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070092static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070093{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070094 struct tcp_sock *tp = tcp_sk(sk);
95
Ilpo Järvinen90840de2007-12-31 04:48:41 -080096 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070097 return tp->snd_nxt;
98 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -080099 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100}
101
102/* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 *
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
109 * large MSS.
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
115 */
116static __u16 tcp_advertise_mss(struct sock *sk)
117{
118 struct tcp_sock *tp = tcp_sk(sk);
119 struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
121
David S. Miller0dbaee32010-12-13 12:52:14 -0800122 if (dst) {
123 unsigned int metric = dst_metric_advmss(dst);
124
125 if (metric < mss) {
126 mss = metric;
127 tp->advmss = mss;
128 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 }
130
131 return (__u16)mss;
132}
133
134/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700136static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700138 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 s32 delta = tcp_time_stamp - tp->lsndtime;
140 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
141 u32 cwnd = tp->snd_cwnd;
142
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 restart_cwnd = min(restart_cwnd, cwnd);
147
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149 cwnd >>= 1;
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_time_stamp;
152 tp->snd_cwnd_used = 0;
153}
154
Andi Kleen67edfef2009-07-21 23:00:40 +0000155/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800156static void tcp_event_data_sent(struct tcp_sock *tp,
157 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
David S. Miller35089bb2006-06-13 22:33:04 -0700162 if (sysctl_tcp_slow_start_after_idle &&
163 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700164 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165
166 tp->lsndtime = now;
167
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
170 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700171 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173}
174
Andi Kleen67edfef2009-07-21 23:00:40 +0000175/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800176static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180}
181
182/* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
188 */
189void tcp_select_initial_window(int __space, __u32 mss,
190 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000191 int wscale_ok, __u8 *rcv_wscale,
192 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193{
194 unsigned int space = (__space < 0 ? 0 : __space);
195
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp == 0)
198 (*window_clamp) = (65535 << 14);
199 space = min(*window_clamp, space);
200
201 /* Quantize space offering to a multiple of mss if possible. */
202 if (space > mss)
203 space = (space / mss) * mss;
204
205 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 */
Rick Jones15d99e02006-03-20 22:40:29 -0800213 if (sysctl_tcp_workaround_signed_windows)
214 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
215 else
216 (*rcv_wnd) = space;
217
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 (*rcv_wscale) = 0;
219 if (wscale_ok) {
220 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900221 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 */
223 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700224 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 while (space > 65535 && (*rcv_wscale) < 14) {
226 space >>= 1;
227 (*rcv_wscale)++;
228 }
229 }
230
Nandita Dukkipati356f0392010-12-20 14:15:56 +0000231 /* Set initial window to a value enough for senders starting with
232 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 * a limit on the initial window when mss is larger than 1460.
234 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800235 if (mss > (1 << *rcv_wscale)) {
Nandita Dukkipati356f0392010-12-20 14:15:56 +0000236 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
237 if (mss > 1460)
238 init_cwnd =
239 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
laurent chavey31d12922009-12-15 11:15:28 +0000240 /* when initializing use the value from init_rcv_wnd
241 * rather than the default from above
242 */
Nandita Dukkipatib1afde62010-12-03 13:33:44 +0000243 if (init_rcv_wnd)
244 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
245 else
246 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 }
248
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
251}
Eric Dumazet4bc2f182010-07-09 21:22:10 +0000252EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
254/* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
257 * frame.
258 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800259static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260{
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 cur_win = tcp_receive_window(tp);
263 u32 new_win = __tcp_select_window(sk);
264
265 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800266 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
271 *
272 * Relax Will Robinson.
273 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275 }
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
278
279 /* Make sure we do not exceed the maximum possible
280 * scaled window.
281 */
Rick Jones15d99e02006-03-20 22:40:29 -0800282 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 new_win = min(new_win, MAX_TCP_WINDOW);
284 else
285 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
286
287 /* RFC1323 scaling applied */
288 new_win >>= tp->rx_opt.rcv_wscale;
289
290 /* If we advertise zero window, disable fast path. */
291 if (new_win == 0)
292 tp->pred_flags = 0;
293
294 return new_win;
295}
296
Andi Kleen67edfef2009-07-21 23:00:40 +0000297/* Packet ECN state for a SYN-ACK */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800298static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700299{
Changli Gaoa3433f32010-06-12 14:01:43 +0000300 TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800301 if (!(tp->ecn_flags & TCP_ECN_OK))
Changli Gaoa3433f32010-06-12 14:01:43 +0000302 TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700303}
304
Andi Kleen67edfef2009-07-21 23:00:40 +0000305/* Packet ECN state for a SYN. */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700306static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
307{
308 struct tcp_sock *tp = tcp_sk(sk);
309
310 tp->ecn_flags = 0;
Ilpo Järvinen255cac92009-05-04 11:07:36 -0700311 if (sysctl_tcp_ecn == 1) {
Changli Gaoa3433f32010-06-12 14:01:43 +0000312 TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700313 tp->ecn_flags = TCP_ECN_OK;
314 }
315}
316
317static __inline__ void
318TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
319{
320 if (inet_rsk(req)->ecn_ok)
321 th->ece = 1;
322}
323
Andi Kleen67edfef2009-07-21 23:00:40 +0000324/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325 * be sent.
326 */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700327static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
328 int tcp_header_len)
329{
330 struct tcp_sock *tp = tcp_sk(sk);
331
332 if (tp->ecn_flags & TCP_ECN_OK) {
333 /* Not-retransmitted data segment: set ECT and inject CWR. */
334 if (skb->len != tcp_header_len &&
335 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
336 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800337 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700338 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
339 tcp_hdr(skb)->cwr = 1;
340 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
341 }
342 } else {
343 /* ACK or retransmitted segment: clear ECT|CE */
344 INET_ECN_dontxmit(sk);
345 }
346 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
347 tcp_hdr(skb)->ece = 1;
348 }
349}
350
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800351/* Constructs common control bits of non-data skb. If SYN/FIN is present,
352 * auto increment end seqno.
353 */
354static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
355{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700356 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800357 skb->csum = 0;
358
359 TCP_SKB_CB(skb)->flags = flags;
360 TCP_SKB_CB(skb)->sacked = 0;
361
362 skb_shinfo(skb)->gso_segs = 1;
363 skb_shinfo(skb)->gso_size = 0;
364 skb_shinfo(skb)->gso_type = 0;
365
366 TCP_SKB_CB(skb)->seq = seq;
Changli Gaoa3433f32010-06-12 14:01:43 +0000367 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800368 seq++;
369 TCP_SKB_CB(skb)->end_seq = seq;
370}
371
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700372static inline int tcp_urg_mode(const struct tcp_sock *tp)
373{
374 return tp->snd_una != tp->snd_up;
375}
376
Adam Langley33ad7982008-07-19 00:04:31 -0700377#define OPTION_SACK_ADVERTISE (1 << 0)
378#define OPTION_TS (1 << 1)
379#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000380#define OPTION_WSCALE (1 << 3)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000381#define OPTION_COOKIE_EXTENSION (1 << 4)
Adam Langley33ad7982008-07-19 00:04:31 -0700382
383struct tcp_out_options {
384 u8 options; /* bit field of OPTION_* */
385 u8 ws; /* window scale, 0 to disable */
386 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000387 u8 hash_size; /* bytes in hash_location */
Adam Langley33ad7982008-07-19 00:04:31 -0700388 u16 mss; /* 0 to disable */
389 __u32 tsval, tsecr; /* need to include OPTION_TS */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000390 __u8 *hash_location; /* temporary pointer, overloaded */
Adam Langley33ad7982008-07-19 00:04:31 -0700391};
392
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000393/* The sysctl int routines are generic, so check consistency here.
394 */
395static u8 tcp_cookie_size_check(u8 desired)
396{
Eric Dumazetf1987252010-12-07 12:20:47 +0000397 int cookie_size;
398
399 if (desired > 0)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000400 /* previously specified */
401 return desired;
Eric Dumazetf1987252010-12-07 12:20:47 +0000402
403 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
404 if (cookie_size <= 0)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000405 /* no default specified */
406 return 0;
Eric Dumazetf1987252010-12-07 12:20:47 +0000407
408 if (cookie_size <= TCP_COOKIE_MIN)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000409 /* value too small, specify minimum */
410 return TCP_COOKIE_MIN;
Eric Dumazetf1987252010-12-07 12:20:47 +0000411
412 if (cookie_size >= TCP_COOKIE_MAX)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000413 /* value too large, specify maximum */
414 return TCP_COOKIE_MAX;
Eric Dumazetf1987252010-12-07 12:20:47 +0000415
416 if (cookie_size & 1)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000417 /* 8-bit multiple, illegal, fix it */
Eric Dumazetf1987252010-12-07 12:20:47 +0000418 cookie_size++;
419
420 return (u8)cookie_size;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000421}
422
Andi Kleen67edfef2009-07-21 23:00:40 +0000423/* Write previously computed TCP options to the packet.
424 *
425 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700426 * TCP options, we learned this through the hard way, so be careful here.
427 * Luckily we can at least blame others for their non-compliance but from
428 * inter-operatibility perspective it seems that we're somewhat stuck with
429 * the ordering which we have been using if we want to keep working with
430 * those broken things (not that it currently hurts anybody as there isn't
431 * particular reason why the ordering would need to be changed).
432 *
433 * At least SACK_PERM as the first option is known to lead to a disaster
434 * (but it may well be that other scenarios fail similarly).
435 */
Adam Langley33ad7982008-07-19 00:04:31 -0700436static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000437 struct tcp_out_options *opts)
438{
439 u8 options = opts->options; /* mungable copy */
440
441 /* Having both authentication and cookies for security is redundant,
442 * and there's certainly not enough room. Instead, the cookie-less
443 * extension variant is proposed.
444 *
445 * Consider the pessimal case with authentication. The options
446 * could look like:
447 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
448 */
449 if (unlikely(OPTION_MD5 & options)) {
450 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
451 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
452 (TCPOLEN_COOKIE_BASE << 16) |
453 (TCPOPT_MD5SIG << 8) |
454 TCPOLEN_MD5SIG);
455 } else {
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
457 (TCPOPT_NOP << 16) |
458 (TCPOPT_MD5SIG << 8) |
459 TCPOLEN_MD5SIG);
460 }
461 options &= ~OPTION_COOKIE_EXTENSION;
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700464 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800465 }
Adam Langley33ad7982008-07-19 00:04:31 -0700466
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
470 opts->mss);
471 }
472
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
478 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000479 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700480 } else {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_TIMESTAMP << 8) |
484 TCPOLEN_TIMESTAMP);
485 }
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
488 }
489
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000490 /* Specification requires after timestamp, so do it now.
491 *
492 * Consider the pessimal case without authentication. The options
493 * could look like:
494 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
495 */
496 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
497 __u8 *cookie_copy = opts->hash_location;
498 u8 cookie_size = opts->hash_size;
499
500 /* 8-bit multiple handled in tcp_cookie_size_check() above,
501 * and elsewhere.
502 */
503 if (0x2 & cookie_size) {
504 __u8 *p = (__u8 *)ptr;
505
506 /* 16-bit multiple */
507 *p++ = TCPOPT_COOKIE;
508 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
509 *p++ = *cookie_copy++;
510 *p++ = *cookie_copy++;
511 ptr++;
512 cookie_size -= 2;
513 } else {
514 /* 32-bit multiple */
515 *ptr++ = htonl(((TCPOPT_NOP << 24) |
516 (TCPOPT_NOP << 16) |
517 (TCPOPT_COOKIE << 8) |
518 TCPOLEN_COOKIE_BASE) +
519 cookie_size);
520 }
521
522 if (cookie_size > 0) {
523 memcpy(ptr, cookie_copy, cookie_size);
524 ptr += (cookie_size / 4);
525 }
526 }
527
528 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700529 *ptr++ = htonl((TCPOPT_NOP << 24) |
530 (TCPOPT_NOP << 16) |
531 (TCPOPT_SACK_PERM << 8) |
532 TCPOLEN_SACK_PERM);
533 }
534
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000535 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700536 *ptr++ = htonl((TCPOPT_NOP << 24) |
537 (TCPOPT_WINDOW << 16) |
538 (TCPOLEN_WINDOW << 8) |
539 opts->ws);
540 }
541
542 if (unlikely(opts->num_sack_blocks)) {
543 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
544 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800545 int this_sack;
546
547 *ptr++ = htonl((TCPOPT_NOP << 24) |
548 (TCPOPT_NOP << 16) |
549 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700550 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800551 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800552
Adam Langley33ad7982008-07-19 00:04:31 -0700553 for (this_sack = 0; this_sack < opts->num_sack_blocks;
554 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800555 *ptr++ = htonl(sp[this_sack].start_seq);
556 *ptr++ = htonl(sp[this_sack].end_seq);
557 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800558
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000559 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800560 }
561}
562
Andi Kleen67edfef2009-07-21 23:00:40 +0000563/* Compute TCP options for SYN packets. This is not the final
564 * network wire format yet.
565 */
Adam Langley33ad7982008-07-19 00:04:31 -0700566static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
567 struct tcp_out_options *opts,
568 struct tcp_md5sig_key **md5) {
569 struct tcp_sock *tp = tcp_sk(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000570 struct tcp_cookie_values *cvp = tp->cookie_values;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000571 unsigned remaining = MAX_TCP_OPTION_SPACE;
572 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
573 tcp_cookie_size_check(cvp->cookie_desired) :
574 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700575
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800576#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700577 *md5 = tp->af_specific->md5_lookup(sk, sk);
578 if (*md5) {
579 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000580 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800581 }
Adam Langley33ad7982008-07-19 00:04:31 -0700582#else
583 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800584#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700585
586 /* We always get an MSS option. The option bytes which will be seen in
587 * normal data packets should timestamps be used, must be in the MSS
588 * advertised. But we subtract them from tp->mss_cache so that
589 * calculations in tcp_sendmsg are simpler etc. So account for this
590 * fact here if necessary. If we don't do this correctly, as a
591 * receiver we won't recognize data packets as being full sized when we
592 * should, and thus we won't abide by the delayed ACK rules correctly.
593 * SACKs don't matter, we never delay an ACK when we have any of those
594 * going out. */
595 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000596 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700597
David S. Millerbb5b7c12009-12-15 20:56:42 -0800598 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700599 opts->options |= OPTION_TS;
600 opts->tsval = TCP_SKB_CB(skb)->when;
601 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000602 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700603 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800604 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700605 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000606 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000607 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700608 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800609 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700610 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700611 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000612 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700613 }
614
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000615 /* Note that timestamps are required by the specification.
616 *
617 * Odd numbers of bytes are prohibited by the specification, ensuring
618 * that the cookie is 16-bit aligned, and the resulting cookie pair is
619 * 32-bit aligned.
620 */
621 if (*md5 == NULL &&
622 (OPTION_TS & opts->options) &&
623 cookie_size > 0) {
624 int need = TCPOLEN_COOKIE_BASE + cookie_size;
625
626 if (0x2 & need) {
627 /* 32-bit multiple */
628 need += 2; /* NOPs */
629
630 if (need > remaining) {
631 /* try shrinking cookie to fit */
632 cookie_size -= 2;
633 need -= 4;
634 }
635 }
636 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
637 cookie_size -= 4;
638 need -= 4;
639 }
640 if (TCP_COOKIE_MIN <= cookie_size) {
641 opts->options |= OPTION_COOKIE_EXTENSION;
642 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
643 opts->hash_size = cookie_size;
644
645 /* Remember for future incarnations. */
646 cvp->cookie_desired = cookie_size;
647
648 if (cvp->cookie_desired != cvp->cookie_pair_size) {
649 /* Currently use random bytes as a nonce,
650 * assuming these are completely unpredictable
651 * by hostile users of the same system.
652 */
653 get_random_bytes(&cvp->cookie_pair[0],
654 cookie_size);
655 cvp->cookie_pair_size = cookie_size;
656 }
657
658 remaining -= need;
659 }
660 }
661 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700662}
663
Andi Kleen67edfef2009-07-21 23:00:40 +0000664/* Set up TCP options for SYN-ACKs. */
Adam Langley33ad7982008-07-19 00:04:31 -0700665static unsigned tcp_synack_options(struct sock *sk,
666 struct request_sock *req,
667 unsigned mss, struct sk_buff *skb,
668 struct tcp_out_options *opts,
William Allen Simpson4957faa2009-12-02 18:25:27 +0000669 struct tcp_md5sig_key **md5,
670 struct tcp_extend_values *xvp)
671{
Adam Langley33ad7982008-07-19 00:04:31 -0700672 struct inet_request_sock *ireq = inet_rsk(req);
William Allen Simpson4957faa2009-12-02 18:25:27 +0000673 unsigned remaining = MAX_TCP_OPTION_SPACE;
674 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
675 xvp->cookie_plus :
676 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700677
678#ifdef CONFIG_TCP_MD5SIG
679 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
680 if (*md5) {
681 opts->options |= OPTION_MD5;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000682 remaining -= TCPOLEN_MD5SIG_ALIGNED;
683
684 /* We can't fit any SACK blocks in a packet with MD5 + TS
685 * options. There was discussion about disabling SACK
686 * rather than TS in order to fit in better with old,
687 * buggy kernels, but that was deemed to be unnecessary.
688 */
Eric Dumazetde213e52010-05-17 22:35:36 -0700689 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700690 }
691#else
692 *md5 = NULL;
693#endif
694
William Allen Simpson4957faa2009-12-02 18:25:27 +0000695 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700696 opts->mss = mss;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000697 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700698
699 if (likely(ireq->wscale_ok)) {
700 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000701 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000702 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700703 }
Eric Dumazetde213e52010-05-17 22:35:36 -0700704 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700705 opts->options |= OPTION_TS;
706 opts->tsval = TCP_SKB_CB(skb)->when;
707 opts->tsecr = req->ts_recent;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000708 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700709 }
710 if (likely(ireq->sack_ok)) {
711 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e52010-05-17 22:35:36 -0700712 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faa2009-12-02 18:25:27 +0000713 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700714 }
715
William Allen Simpson4957faa2009-12-02 18:25:27 +0000716 /* Similar rationale to tcp_syn_options() applies here, too.
717 * If the <SYN> options fit, the same options should fit now!
718 */
719 if (*md5 == NULL &&
Eric Dumazetde213e52010-05-17 22:35:36 -0700720 ireq->tstamp_ok &&
William Allen Simpson4957faa2009-12-02 18:25:27 +0000721 cookie_plus > TCPOLEN_COOKIE_BASE) {
722 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
723
724 if (0x2 & need) {
725 /* 32-bit multiple */
726 need += 2; /* NOPs */
727 }
728 if (need <= remaining) {
729 opts->options |= OPTION_COOKIE_EXTENSION;
730 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
731 remaining -= need;
732 } else {
733 /* There's no error return, so flag it. */
734 xvp->cookie_out_never = 1; /* true */
735 opts->hash_size = 0;
736 }
737 }
738 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700739}
740
Andi Kleen67edfef2009-07-21 23:00:40 +0000741/* Compute TCP options for ESTABLISHED sockets. This is not the
742 * final wire format yet.
743 */
Adam Langley33ad7982008-07-19 00:04:31 -0700744static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
745 struct tcp_out_options *opts,
746 struct tcp_md5sig_key **md5) {
747 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
748 struct tcp_sock *tp = tcp_sk(sk);
749 unsigned size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000750 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700751
752#ifdef CONFIG_TCP_MD5SIG
753 *md5 = tp->af_specific->md5_lookup(sk, sk);
754 if (unlikely(*md5)) {
755 opts->options |= OPTION_MD5;
756 size += TCPOLEN_MD5SIG_ALIGNED;
757 }
758#else
759 *md5 = NULL;
760#endif
761
762 if (likely(tp->rx_opt.tstamp_ok)) {
763 opts->options |= OPTION_TS;
764 opts->tsval = tcb ? tcb->when : 0;
765 opts->tsecr = tp->rx_opt.ts_recent;
766 size += TCPOLEN_TSTAMP_ALIGNED;
767 }
768
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000769 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 if (unlikely(eff_sacks)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700771 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
772 opts->num_sack_blocks =
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000773 min_t(unsigned, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700774 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 TCPOLEN_SACK_PERBLOCK);
776 size += TCPOLEN_SACK_BASE_ALIGNED +
777 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
778 }
779
780 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800781}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782
783/* This routine actually transmits TCP packets queued in by
784 * tcp_do_sendmsg(). This is used by both the initial
785 * transmission and possible later retransmissions.
786 * All SKB's seen here are completely headerless. It is our
787 * job to build the TCP header, and pass the packet down to
788 * IP so it can do the same plus pass the packet off to the
789 * device.
790 *
791 * We are working here with either a clone of the original
792 * SKB, or a fresh unique copy made by the retransmit engine.
793 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800794static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
795 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800797 const struct inet_connection_sock *icsk = inet_csk(sk);
798 struct inet_sock *inet;
799 struct tcp_sock *tp;
800 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700801 struct tcp_out_options opts;
802 unsigned tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800803 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800804 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800805 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800807 BUG_ON(!skb || !tcp_skb_pcount(skb));
808
809 /* If congestion control is doing timestamping, we must
810 * take such a timestamp before we potentially clone/copy.
811 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700812 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800813 __net_timestamp(skb);
814
815 if (likely(clone_it)) {
816 if (unlikely(skb_cloned(skb)))
817 skb = pskb_copy(skb, gfp_mask);
818 else
819 skb = skb_clone(skb, gfp_mask);
820 if (unlikely(!skb))
821 return -ENOBUFS;
822 }
823
824 inet = inet_sk(sk);
825 tp = tcp_sk(sk);
826 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700827 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828
Changli Gaoa3433f32010-06-12 14:01:43 +0000829 if (unlikely(tcb->flags & TCPHDR_SYN))
Adam Langley33ad7982008-07-19 00:04:31 -0700830 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
831 else
832 tcp_options_size = tcp_established_options(sk, skb, &opts,
833 &md5);
834 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900835
Tom Herbert3853b582010-11-21 13:17:29 +0000836 if (tcp_packets_in_flight(tp) == 0) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800837 tcp_ca_event(sk, CA_EVENT_TX_START);
Tom Herbert3853b582010-11-21 13:17:29 +0000838 skb->ooo_okay = 1;
839 } else
840 skb->ooo_okay = 0;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800841
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700842 skb_push(skb, tcp_header_size);
843 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800844 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800845
846 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700847 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000848 th->source = inet->inet_sport;
849 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800850 th->seq = htonl(tcb->seq);
851 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700852 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800853 tcb->flags);
854
Changli Gaoa3433f32010-06-12 14:01:43 +0000855 if (unlikely(tcb->flags & TCPHDR_SYN)) {
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800856 /* RFC1323: The window in SYN & SYN/ACK segments
857 * is never scaled.
858 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800859 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800860 } else {
861 th->window = htons(tcp_select_window(sk));
862 }
863 th->check = 0;
864 th->urg_ptr = 0;
865
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700866 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800867 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
868 if (before(tp->snd_up, tcb->seq + 0x10000)) {
869 th->urg_ptr = htons(tp->snd_up - tcb->seq);
870 th->urg = 1;
871 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700872 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800873 th->urg = 1;
874 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800875 }
876
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000877 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Changli Gaoa3433f32010-06-12 14:01:43 +0000878 if (likely((tcb->flags & TCPHDR_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700879 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800880
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800881#ifdef CONFIG_TCP_MD5SIG
882 /* Calculate the MD5 hash, as we have all we need now */
883 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -0700884 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000885 tp->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700886 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800887 }
888#endif
889
Herbert Xubb296242010-04-11 02:15:55 +0000890 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800891
Changli Gaoa3433f32010-06-12 14:01:43 +0000892 if (likely(tcb->flags & TCPHDR_ACK))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800893 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
894
895 if (skb->len != tcp_header_size)
896 tcp_event_data_sent(tp, skb, sk);
897
Wei Yongjunbd37a082006-08-07 21:04:15 -0700898 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +0000899 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
900 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800901
Shan Wei4e15ed42010-04-15 16:43:08 +0000902 err = icsk->icsk_af_ops->queue_xmit(skb);
Hua Zhong83de47c2006-04-28 15:26:50 -0700903 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800904 return err;
905
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800906 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800907
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200908 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909}
910
Andi Kleen67edfef2009-07-21 23:00:40 +0000911/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 *
913 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
914 * otherwise socket can stall.
915 */
916static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
917{
918 struct tcp_sock *tp = tcp_sk(sk);
919
920 /* Advance write_seq and place onto the write_queue. */
921 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
922 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800923 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800924 sk->sk_wmem_queued += skb->truesize;
925 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926}
927
Andi Kleen67edfef2009-07-21 23:00:40 +0000928/* Initialize TSO segments for a packet. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800929static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
930 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700931{
Herbert Xu8e5b9dd2009-06-28 18:03:30 +0000932 if (skb->len <= mss_now || !sk_can_gso(sk) ||
933 skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -0700934 /* Avoid the costly divide in the normal
935 * non-TSO case.
936 */
Herbert Xu79671682006-06-22 02:40:14 -0700937 skb_shinfo(skb)->gso_segs = 1;
938 skb_shinfo(skb)->gso_size = 0;
939 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700940 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700941 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700942 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700943 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 }
945}
946
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700947/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800948 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700949 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800950static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700951 int decr)
952{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800953 struct tcp_sock *tp = tcp_sk(sk);
954
Ilpo Järvinendc869672007-10-01 15:27:19 -0700955 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700956 return;
957
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200958 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700959 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700960}
961
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000962/* Pcount in the middle of the write queue got changed, we need to do various
963 * tweaks to fix counters
964 */
965static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
966{
967 struct tcp_sock *tp = tcp_sk(sk);
968
969 tp->packets_out -= decr;
970
971 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
972 tp->sacked_out -= decr;
973 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
974 tp->retrans_out -= decr;
975 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
976 tp->lost_out -= decr;
977
978 /* Reno case is special. Sigh... */
979 if (tcp_is_reno(tp) && decr > 0)
980 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
981
982 tcp_adjust_fackets_out(sk, skb, decr);
983
984 if (tp->lost_skb_hint &&
985 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +0000986 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000987 tp->lost_cnt_hint -= decr;
988
989 tcp_verify_left_out(tp);
990}
991
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992/* Function to create two new TCP segments. Shrinks the given segment
993 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900994 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 * Remember, these are still headerless SKBs at this point.
996 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800997int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
998 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999{
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -07001002 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -07001003 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001004 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
Herbert Xub2cc99f2005-10-20 17:13:13 -02001006 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001007
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 nsize = skb_headlen(skb) - len;
1009 if (nsize < 0)
1010 nsize = 0;
1011
1012 if (skb_cloned(skb) &&
1013 skb_is_nonlinear(skb) &&
1014 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1015 return -ENOMEM;
1016
1017 /* Get a new skb... force flag on. */
1018 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1019 if (buff == NULL)
1020 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001021
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001022 sk->sk_wmem_queued += buff->truesize;
1023 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001024 nlen = skb->len - len - nsize;
1025 buff->truesize += nlen;
1026 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027
1028 /* Correct the sequence numbers. */
1029 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1030 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1031 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1032
1033 /* PSH and FIN should only be set in the second packet. */
1034 flags = TCP_SKB_CB(skb)->flags;
Changli Gaoa3433f32010-06-12 14:01:43 +00001035 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001037 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038
Patrick McHardy84fa7932006-08-29 16:44:56 -07001039 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001041 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1042 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043 nsize, 0);
1044
1045 skb_trim(skb, len);
1046
1047 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1048 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001049 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050 skb_split(skb, buff, len);
1051 }
1052
1053 buff->ip_summed = skb->ip_summed;
1054
1055 /* Looks stupid, but our code really uses when of
1056 * skbs, which it never sent before. --ANK
1057 */
1058 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001059 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001060
David S. Miller6475be12005-09-01 22:47:01 -07001061 old_factor = tcp_skb_pcount(skb);
1062
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001064 tcp_set_skb_tso_segs(sk, skb, mss_now);
1065 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
David S. Miller6475be12005-09-01 22:47:01 -07001067 /* If this packet has been sent out already, we must
1068 * adjust the various packet counters.
1069 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001070 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001071 int diff = old_factor - tcp_skb_pcount(skb) -
1072 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001074 if (diff)
1075 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 }
1077
1078 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -07001079 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001080 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081
1082 return 0;
1083}
1084
1085/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1086 * eventually). The difference is that pulled data not copied, but
1087 * immediately discarded.
1088 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001089static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090{
1091 int i, k, eat;
1092
1093 eat = len;
1094 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001095 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 if (skb_shinfo(skb)->frags[i].size <= eat) {
1097 put_page(skb_shinfo(skb)->frags[i].page);
1098 eat -= skb_shinfo(skb)->frags[i].size;
1099 } else {
1100 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1101 if (eat) {
1102 skb_shinfo(skb)->frags[k].page_offset += eat;
1103 skb_shinfo(skb)->frags[k].size -= eat;
1104 eat = 0;
1105 }
1106 k++;
1107 }
1108 }
1109 skb_shinfo(skb)->nr_frags = k;
1110
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001111 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 skb->data_len -= len;
1113 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114}
1115
Andi Kleen67edfef2009-07-21 23:00:40 +00001116/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1118{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001119 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 return -ENOMEM;
1121
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001122 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1123 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001125 else
1126 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127
1128 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001129 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130
1131 skb->truesize -= len;
1132 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001133 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1135
1136 /* Any change of skb->len requires recalculation of tso
1137 * factor and mss.
1138 */
1139 if (tcp_skb_pcount(skb) > 1)
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001140 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141
1142 return 0;
1143}
1144
Andi Kleen67edfef2009-07-21 23:00:40 +00001145/* Calculate MSS. Not accounting for SACKs here. */
John Heffner5d424d52006-03-20 17:53:41 -08001146int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1147{
1148 struct tcp_sock *tp = tcp_sk(sk);
1149 struct inet_connection_sock *icsk = inet_csk(sk);
1150 int mss_now;
1151
1152 /* Calculate base mss without TCP options:
1153 It is MMS_S - sizeof(tcphdr) of rfc1122
1154 */
1155 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1156
1157 /* Clamp it (mss_clamp does not include tcp options) */
1158 if (mss_now > tp->rx_opt.mss_clamp)
1159 mss_now = tp->rx_opt.mss_clamp;
1160
1161 /* Now subtract optional transport overhead */
1162 mss_now -= icsk->icsk_ext_hdr_len;
1163
1164 /* Then reserve room for full set of TCP options and 8 bytes of data */
1165 if (mss_now < 48)
1166 mss_now = 48;
1167
1168 /* Now subtract TCP options size, not including SACKs */
1169 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1170
1171 return mss_now;
1172}
1173
1174/* Inverse of above */
1175int tcp_mss_to_mtu(struct sock *sk, int mss)
1176{
1177 struct tcp_sock *tp = tcp_sk(sk);
1178 struct inet_connection_sock *icsk = inet_csk(sk);
1179 int mtu;
1180
1181 mtu = mss +
1182 tp->tcp_header_len +
1183 icsk->icsk_ext_hdr_len +
1184 icsk->icsk_af_ops->net_header_len;
1185
1186 return mtu;
1187}
1188
Andi Kleen67edfef2009-07-21 23:00:40 +00001189/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001190void tcp_mtup_init(struct sock *sk)
1191{
1192 struct tcp_sock *tp = tcp_sk(sk);
1193 struct inet_connection_sock *icsk = inet_csk(sk);
1194
1195 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1196 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001197 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -08001198 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1199 icsk->icsk_mtup.probe_size = 0;
1200}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001201EXPORT_SYMBOL(tcp_mtup_init);
John Heffner5d424d52006-03-20 17:53:41 -08001202
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203/* This function synchronize snd mss to current pmtu/exthdr set.
1204
1205 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1206 for TCP options, but includes only bare TCP header.
1207
1208 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001209 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 It also does not include TCP options.
1211
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001212 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213
1214 tp->mss_cache is current effective sending mss, including
1215 all tcp options except for SACKs. It is evaluated,
1216 taking into account current pmtu, but never exceeds
1217 tp->rx_opt.mss_clamp.
1218
1219 NOTE1. rfc1122 clearly states that advertised MSS
1220 DOES NOT include either tcp or ip options.
1221
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001222 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1223 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1226{
1227 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001228 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001229 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
John Heffner5d424d52006-03-20 17:53:41 -08001231 if (icsk->icsk_mtup.search_high > pmtu)
1232 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233
John Heffner5d424d52006-03-20 17:53:41 -08001234 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001235 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236
1237 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001238 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001239 if (icsk->icsk_mtup.enabled)
1240 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001241 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242
1243 return mss_now;
1244}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00001245EXPORT_SYMBOL(tcp_sync_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246
1247/* Compute the current effective MSS, taking SACKs and IP options,
1248 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001250unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
1252 struct tcp_sock *tp = tcp_sk(sk);
1253 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001254 u32 mss_now;
Adam Langley33ad7982008-07-19 00:04:31 -07001255 unsigned header_len;
1256 struct tcp_out_options opts;
1257 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001259 mss_now = tp->mss_cache;
1260
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 if (dst) {
1262 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001263 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 mss_now = tcp_sync_mss(sk, mtu);
1265 }
1266
Adam Langley33ad7982008-07-19 00:04:31 -07001267 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1268 sizeof(struct tcphdr);
1269 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1270 * some common options. If this is an odd packet (because we have SACK
1271 * blocks etc) then our calculated header_len will be different, and
1272 * we have to adjust mss_now correspondingly */
1273 if (header_len != tp->tcp_header_len) {
1274 int delta = (int) header_len - tp->tcp_header_len;
1275 mss_now -= delta;
1276 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001277
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 return mss_now;
1279}
1280
David S. Millera762a982005-07-05 15:18:51 -07001281/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001282static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001283{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001284 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001285
Ilpo Järvinend436d682007-12-31 14:58:00 -08001286 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001287 /* Network is feed fully. */
1288 tp->snd_cwnd_used = 0;
1289 tp->snd_cwnd_stamp = tcp_time_stamp;
1290 } else {
1291 /* Network starves. */
1292 if (tp->packets_out > tp->snd_cwnd_used)
1293 tp->snd_cwnd_used = tp->packets_out;
1294
David S. Miller15d33c02007-04-09 13:23:14 -07001295 if (sysctl_tcp_slow_start_after_idle &&
1296 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001297 tcp_cwnd_application_limited(sk);
1298 }
1299}
1300
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001301/* Returns the portion of skb which can be sent right away without
1302 * introducing MSS oddities to segment boundaries. In rare cases where
1303 * mss_now != mss_cache, we will request caller to create a small skb
1304 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001305 *
1306 * We explicitly want to create a request for splitting write queue tail
1307 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1308 * thus all the complexity (cwnd_len is always MSS multiple which we
1309 * return whenever allowed by the other factors). Basically we need the
1310 * modulo only when the receiver window alone is the limiting factor or
1311 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001312 */
1313static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001314 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001315{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001316 struct tcp_sock *tp = tcp_sk(sk);
1317 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001318
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001319 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001320 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001321
1322 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1323 return cwnd_len;
1324
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001325 needed = min(skb->len, window);
1326
Ilpo Järvinen17515402008-04-15 20:36:55 -07001327 if (cwnd_len <= needed)
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001328 return cwnd_len;
1329
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001330 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001331}
1332
1333/* Can at least one segment of SKB be sent right now, according to the
1334 * congestion window rules? If so, return how many segments are allowed.
1335 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001336static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1337 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001338{
1339 u32 in_flight, cwnd;
1340
1341 /* Don't be strict about the congestion window for the final FIN. */
Changli Gaoa3433f32010-06-12 14:01:43 +00001342 if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001343 return 1;
1344
1345 in_flight = tcp_packets_in_flight(tp);
1346 cwnd = tp->snd_cwnd;
1347 if (in_flight < cwnd)
1348 return (cwnd - in_flight);
1349
1350 return 0;
1351}
1352
Uwe Kleine-Königb5950762010-11-01 15:38:34 -04001353/* Initialize TSO state of a skb.
Andi Kleen67edfef2009-07-21 23:00:40 +00001354 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001355 * SKB onto the wire.
1356 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001357static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1358 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001359{
1360 int tso_segs = tcp_skb_pcount(skb);
1361
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001362 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001363 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001364 tso_segs = tcp_skb_pcount(skb);
1365 }
1366 return tso_segs;
1367}
1368
Andi Kleen67edfef2009-07-21 23:00:40 +00001369/* Minshall's variant of the Nagle send check. */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001370static inline int tcp_minshall_check(const struct tcp_sock *tp)
1371{
Jianjun Kong09cb1052008-11-03 00:27:11 -08001372 return after(tp->snd_sml, tp->snd_una) &&
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001373 !after(tp->snd_sml, tp->snd_nxt);
1374}
1375
1376/* Return 0, if packet can be sent now without violation Nagle's rules:
1377 * 1. It is full sized.
1378 * 2. Or it contains FIN. (already checked by caller)
1379 * 3. Or TCP_NODELAY was set.
1380 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1381 * With Minshall's modification: all sent small packets are ACKed.
1382 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001383static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001384 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001385 unsigned mss_now, int nonagle)
1386{
Eric Dumazeta02cec22010-09-22 20:43:57 +00001387 return skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001388 ((nonagle & TCP_NAGLE_CORK) ||
Eric Dumazeta02cec22010-09-22 20:43:57 +00001389 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001390}
1391
1392/* Return non-zero if the Nagle test allows this packet to be
1393 * sent now.
1394 */
1395static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1396 unsigned int cur_mss, int nonagle)
1397{
1398 /* Nagle rule does not apply to frames, which sit in the middle of the
1399 * write_queue (they have no chances to get new data).
1400 *
1401 * This is implemented in the callers, where they modify the 'nonagle'
1402 * argument based upon the location of SKB in the send queue.
1403 */
1404 if (nonagle & TCP_NAGLE_PUSH)
1405 return 1;
1406
Ilpo Järvinend551e452007-04-30 00:42:20 -07001407 /* Don't use the nagle rule for urgent data (or for the final FIN).
1408 * Nagle can be ignored during F-RTO too (see RFC4138).
1409 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001410 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
Changli Gaoa3433f32010-06-12 14:01:43 +00001411 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001412 return 1;
1413
1414 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1415 return 1;
1416
1417 return 0;
1418}
1419
1420/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001421static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1422 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001423{
1424 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1425
1426 if (skb->len > cur_mss)
1427 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1428
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001429 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001430}
1431
David S. Millerfe067e82007-03-07 12:12:44 -08001432/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001433 * should be put on the wire right now. If so, it returns the number of
1434 * packets allowed by the congestion window.
1435 */
1436static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1437 unsigned int cur_mss, int nonagle)
1438{
1439 struct tcp_sock *tp = tcp_sk(sk);
1440 unsigned int cwnd_quota;
1441
David S. Miller846998a2005-08-04 19:52:01 -07001442 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001443
1444 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1445 return 0;
1446
1447 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001448 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001449 cwnd_quota = 0;
1450
1451 return cwnd_quota;
1452}
1453
Andi Kleen67edfef2009-07-21 23:00:40 +00001454/* Test if sending is allowed right now. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001455int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001456{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001457 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001458 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001459
Eric Dumazeta02cec22010-09-22 20:43:57 +00001460 return skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001461 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001462 (tcp_skb_is_last(sk, skb) ?
Eric Dumazeta02cec22010-09-22 20:43:57 +00001463 tp->nonagle : TCP_NAGLE_PUSH));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001464}
1465
1466/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1467 * which is put after SKB on the list. It is very much like
1468 * tcp_fragment() except that it may make several kinds of assumptions
1469 * in order to speed up the splitting operation. In particular, we
1470 * know that all the data is in scatter-gather pages, and that the
1471 * packet has never been sent out before (and thus is not cloned).
1472 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001473static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001474 unsigned int mss_now, gfp_t gfp)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001475{
1476 struct sk_buff *buff;
1477 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001478 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001479
1480 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001481 if (skb->len != skb->data_len)
1482 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001483
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001484 buff = sk_stream_alloc_skb(sk, 0, gfp);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001485 if (unlikely(buff == NULL))
1486 return -ENOMEM;
1487
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001488 sk->sk_wmem_queued += buff->truesize;
1489 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001490 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001491 skb->truesize -= nlen;
1492
1493 /* Correct the sequence numbers. */
1494 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1495 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1496 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1497
1498 /* PSH and FIN should only be set in the second packet. */
1499 flags = TCP_SKB_CB(skb)->flags;
Changli Gaoa3433f32010-06-12 14:01:43 +00001500 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001501 TCP_SKB_CB(buff)->flags = flags;
1502
1503 /* This packet was never sent out yet, so no SACK bits. */
1504 TCP_SKB_CB(buff)->sacked = 0;
1505
Patrick McHardy84fa7932006-08-29 16:44:56 -07001506 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001507 skb_split(skb, buff, len);
1508
1509 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001510 tcp_set_skb_tso_segs(sk, skb, mss_now);
1511 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001512
1513 /* Link BUFF into the send queue. */
1514 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001515 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001516
1517 return 0;
1518}
1519
1520/* Try to defer sending, if possible, in order to minimize the amount
1521 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1522 *
1523 * This algorithm is from John Heffner.
1524 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001525static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001526{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001527 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001528 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001529 u32 send_win, cong_win, limit, in_flight;
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001530 int win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001531
Changli Gaoa3433f32010-06-12 14:01:43 +00001532 if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001533 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001534
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001535 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001536 goto send_now;
1537
1538 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001539 if (tp->tso_deferred &&
Ilpo Järvinena2acde02008-12-05 22:49:37 -08001540 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001541 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001542
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001543 in_flight = tcp_packets_in_flight(tp);
1544
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001545 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001546
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001547 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001548
1549 /* From in_flight test above, we know that cwnd > in_flight. */
1550 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1551
1552 limit = min(send_win, cong_win);
1553
David S. Millerba244fe2006-03-11 18:51:49 -08001554 /* If a full-sized TSO skb can be sent, do it. */
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001555 if (limit >= sk->sk_gso_max_size)
John Heffnerae8064a2006-10-18 20:36:48 -07001556 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001557
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001558 /* Middle in queue won't get any more data, full sendable already? */
1559 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1560 goto send_now;
1561
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001562 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1563 if (win_divisor) {
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001564 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1565
1566 /* If at least some fraction of a window is available,
1567 * just use it.
1568 */
Eric Dumazetad9f4f52010-12-07 12:03:55 +00001569 chunk /= win_divisor;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001570 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001571 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001572 } else {
1573 /* Different approach, try not to defer past a single
1574 * ACK. Receiver should ACK every other full sized
1575 * frame, so if we have space for more than 3 frames
1576 * then send now.
1577 */
1578 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001579 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001580 }
1581
1582 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001583 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001584
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001585 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001586
1587send_now:
1588 tp->tso_deferred = 0;
1589 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001590}
1591
John Heffner5d424d52006-03-20 17:53:41 -08001592/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001593 * MTU probe is regularly attempting to increase the path MTU by
1594 * deliberately sending larger packets. This discovers routing
1595 * changes resulting in larger path MTUs.
1596 *
John Heffner5d424d52006-03-20 17:53:41 -08001597 * Returns 0 if we should wait to probe (no cwnd available),
1598 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001599 * -1 otherwise
1600 */
John Heffner5d424d52006-03-20 17:53:41 -08001601static int tcp_mtu_probe(struct sock *sk)
1602{
1603 struct tcp_sock *tp = tcp_sk(sk);
1604 struct inet_connection_sock *icsk = inet_csk(sk);
1605 struct sk_buff *skb, *nskb, *next;
1606 int len;
1607 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001608 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001609 int copy;
1610 int mss_now;
1611
1612 /* Not currently probing/verifying,
1613 * not in recovery,
1614 * have enough cwnd, and
1615 * not SACKing (the variable headers throw things off) */
1616 if (!icsk->icsk_mtup.enabled ||
1617 icsk->icsk_mtup.probe_size ||
1618 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1619 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001620 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001621 return -1;
1622
1623 /* Very simple search strategy: just double the MSS. */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001624 mss_now = tcp_current_mss(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001625 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001626 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001627 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1628 /* TODO: set timer for probe_converge_event */
1629 return -1;
1630 }
1631
1632 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001633 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001634 return -1;
1635
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001636 if (tp->snd_wnd < size_needed)
1637 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001638 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001639 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001640
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001641 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1642 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1643 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001644 return -1;
1645 else
1646 return 0;
1647 }
1648
1649 /* We're allowed to probe. Build it now. */
1650 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1651 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001652 sk->sk_wmem_queued += nskb->truesize;
1653 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001654
David S. Millerfe067e82007-03-07 12:12:44 -08001655 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001656
1657 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1658 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
Changli Gaoa3433f32010-06-12 14:01:43 +00001659 TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
John Heffner5d424d52006-03-20 17:53:41 -08001660 TCP_SKB_CB(nskb)->sacked = 0;
1661 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001662 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001663
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001664 tcp_insert_write_queue_before(nskb, skb, sk);
1665
John Heffner5d424d52006-03-20 17:53:41 -08001666 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001667 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001668 copy = min_t(int, skb->len, probe_size - len);
1669 if (nskb->ip_summed)
1670 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1671 else
1672 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001673 skb_put(nskb, copy),
1674 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001675
1676 if (skb->len <= copy) {
1677 /* We've eaten all the data from this skb.
1678 * Throw it away. */
1679 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001680 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001681 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001682 } else {
1683 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
Changli Gaoa3433f32010-06-12 14:01:43 +00001684 ~(TCPHDR_FIN|TCPHDR_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001685 if (!skb_shinfo(skb)->nr_frags) {
1686 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001687 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001688 skb->csum = csum_partial(skb->data,
1689 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001690 } else {
1691 __pskb_trim_head(skb, copy);
1692 tcp_set_skb_tso_segs(sk, skb, mss_now);
1693 }
1694 TCP_SKB_CB(skb)->seq += copy;
1695 }
1696
1697 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001698
1699 if (len >= probe_size)
1700 break;
John Heffner5d424d52006-03-20 17:53:41 -08001701 }
1702 tcp_init_tso_segs(sk, nskb, nskb->len);
1703
1704 /* We're ready to send. If this fails, the probe will
1705 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1706 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1707 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1708 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001709 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001710 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001711 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001712
1713 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001714 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1715 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001716
1717 return 1;
1718 }
1719
1720 return -1;
1721}
1722
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723/* This routine writes packets to the network. It advances the
1724 * send_head. This happens as incoming acks open up the remote
1725 * window for us.
1726 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001727 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1728 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1729 * account rare use of URG, this is not a big flaw.
1730 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 * Returns 1, if no segments are in flight and we have queued segments, but
1732 * cannot send anything now because of SWS or another problem.
1733 */
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001734static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1735 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736{
1737 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001738 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001739 unsigned int tso_segs, sent_pkts;
1740 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001741 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742
David S. Miller92df7b52005-07-05 15:19:06 -07001743 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001744
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001745 if (!push_one) {
1746 /* Do MTU probing. */
1747 result = tcp_mtu_probe(sk);
1748 if (!result) {
1749 return 0;
1750 } else if (result > 0) {
1751 sent_pkts = 1;
1752 }
John Heffner5d424d52006-03-20 17:53:41 -08001753 }
1754
David S. Millerfe067e82007-03-07 12:12:44 -08001755 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001756 unsigned int limit;
1757
Herbert Xub68e9f82005-08-04 19:52:02 -07001758 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001759 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001760
Herbert Xub68e9f82005-08-04 19:52:02 -07001761 cwnd_quota = tcp_cwnd_test(tp, skb);
1762 if (!cwnd_quota)
1763 break;
1764
1765 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1766 break;
1767
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001768 if (tso_segs == 1) {
1769 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1770 (tcp_skb_is_last(sk, skb) ?
1771 nonagle : TCP_NAGLE_PUSH))))
1772 break;
1773 } else {
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001774 if (!push_one && tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001775 break;
1776 }
David S. Milleraa934662005-07-05 15:20:09 -07001777
Herbert Xuc8ac3772005-08-16 20:43:40 -07001778 limit = mss_now;
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001779 if (tso_segs > 1 && !tcp_urg_mode(tp))
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001780 limit = tcp_mss_split_point(sk, skb, mss_now,
1781 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782
Herbert Xuc8ac3772005-08-16 20:43:40 -07001783 if (skb->len > limit &&
Eric Dumazetc4ead4c2010-06-24 01:00:22 +00001784 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
Herbert Xuc8ac3772005-08-16 20:43:40 -07001785 break;
1786
David S. Miller92df7b52005-07-05 15:19:06 -07001787 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001788
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001789 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07001790 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791
David S. Miller92df7b52005-07-05 15:19:06 -07001792 /* Advance the send_head. This one is sent out.
1793 * This call will increment packets_out.
1794 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001795 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001796
1797 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001798 sent_pkts++;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001799
1800 if (push_one)
1801 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 }
David S. Miller92df7b52005-07-05 15:19:06 -07001803
David S. Milleraa934662005-07-05 15:20:09 -07001804 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001805 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001806 return 0;
1807 }
David S. Millerfe067e82007-03-07 12:12:44 -08001808 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809}
1810
David S. Millera762a982005-07-05 15:18:51 -07001811/* Push out any pending frames which were held back due to
1812 * TCP_CORK or attempt at coalescing tiny packets.
1813 * The socket must be locked by the caller.
1814 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001815void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1816 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001817{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001818 /* If we are closed, the bytes will have to remain here.
1819 * In time closedown will finish, we empty the write queue and
1820 * all will be happy.
1821 */
1822 if (unlikely(sk->sk_state == TCP_CLOSE))
1823 return;
1824
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001825 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001826 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001827}
1828
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001829/* Send _single_ skb sitting at the send head. This function requires
1830 * true push pending frames to setup probe timer etc.
1831 */
1832void tcp_push_one(struct sock *sk, unsigned int mss_now)
1833{
David S. Millerfe067e82007-03-07 12:12:44 -08001834 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001835
1836 BUG_ON(!skb || skb->len < mss_now);
1837
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001838 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001839}
1840
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841/* This function returns the amount that we can raise the
1842 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001843 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 * 1. The window can never be shrunk once it is offered (RFC 793)
1845 * 2. We limit memory per socket
1846 *
1847 * RFC 1122:
1848 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1849 * RECV.NEXT + RCV.WIN fixed until:
1850 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1851 *
1852 * i.e. don't raise the right edge of the window until you can raise
1853 * it at least MSS bytes.
1854 *
1855 * Unfortunately, the recommended algorithm breaks header prediction,
1856 * since header prediction assumes th->window stays fixed.
1857 *
1858 * Strictly speaking, keeping th->window fixed violates the receiver
1859 * side SWS prevention criteria. The problem is that under this rule
1860 * a stream of single byte packets will cause the right side of the
1861 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001862 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 * Of course, if the sender implements sender side SWS prevention
1864 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001865 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001867 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 * If the free space is less than the 1/4 of the maximum
1869 * space available and the free space is less than 1/2 mss,
1870 * then set the window to 0.
1871 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1872 * Otherwise, just prevent the window from shrinking
1873 * and from being larger than the largest representable value.
1874 *
1875 * This prevents incremental opening of the window in the regime
1876 * where TCP is limited by the speed of the reader side taking
1877 * data out of the TCP receive queue. It does nothing about
1878 * those cases where the window is constrained on the sender side
1879 * because the pipeline is full.
1880 *
1881 * BSD also seems to "accidentally" limit itself to windows that are a
1882 * multiple of MSS, at least until the free space gets quite small.
1883 * This would appear to be a side effect of the mbuf implementation.
1884 * Combining these two algorithms results in the observed behavior
1885 * of having a fixed window size at almost all times.
1886 *
1887 * Below we obtain similar behavior by forcing the offered window to
1888 * a multiple of the mss when it is feasible to do so.
1889 *
1890 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1891 * Regular options like TIMESTAMP are taken into account.
1892 */
1893u32 __tcp_select_window(struct sock *sk)
1894{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001895 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001897 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 * here. I don't know if the value based on our guesses
1899 * of peer's MSS is better for the performance. It's more correct
1900 * but may be worse for the performance because of rcv_mss
1901 * fluctuations. --SAW 1998/11/1
1902 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001903 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904 int free_space = tcp_space(sk);
1905 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1906 int window;
1907
1908 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001909 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001911 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001912 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913
1914 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001915 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1916 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917
1918 if (free_space < mss)
1919 return 0;
1920 }
1921
1922 if (free_space > tp->rcv_ssthresh)
1923 free_space = tp->rcv_ssthresh;
1924
1925 /* Don't do rounding if we are using window scaling, since the
1926 * scaled window will not line up with the MSS boundary anyway.
1927 */
1928 window = tp->rcv_wnd;
1929 if (tp->rx_opt.rcv_wscale) {
1930 window = free_space;
1931
1932 /* Advertise enough space so that it won't get scaled away.
1933 * Import case: prevent zero window announcement if
1934 * 1<<rcv_wscale > mss.
1935 */
1936 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1937 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1938 << tp->rx_opt.rcv_wscale);
1939 } else {
1940 /* Get the largest window that is a nice multiple of mss.
1941 * Window clamp already applied above.
1942 * If our current window offering is within 1 mss of the
1943 * free space we just keep it. This prevents the divide
1944 * and multiply from happening most of the time.
1945 * We also don't do any window rounding when the free space
1946 * is too small.
1947 */
1948 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001949 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001950 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001951 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001952 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 }
1954
1955 return window;
1956}
1957
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08001958/* Collapses two adjacent SKB's during retransmission. */
1959static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960{
1961 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001962 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001963 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001965 skb_size = skb->len;
1966 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001968 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001969
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001970 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001972 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001974 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1975 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001977 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1978 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001980 if (skb->ip_summed != CHECKSUM_PARTIAL)
1981 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001983 /* Update sequence range on original skb. */
1984 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00001986 /* Merge over control information. This moves PSH/FIN etc. over */
1987 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001989 /* All done, get rid of second SKB and account for it so
1990 * packet counting does not break.
1991 */
1992 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001993
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001994 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07001995 tcp_clear_retrans_hints_partial(tp);
1996 if (next_skb == tp->retransmit_skb_hint)
1997 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001998
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001999 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2000
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002001 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002}
2003
Andi Kleen67edfef2009-07-21 23:00:40 +00002004/* Check if coalescing SKBs is legal. */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002005static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2006{
2007 if (tcp_skb_pcount(skb) > 1)
2008 return 0;
2009 /* TODO: SACK collapsing could be used to remove this condition */
2010 if (skb_shinfo(skb)->nr_frags != 0)
2011 return 0;
2012 if (skb_cloned(skb))
2013 return 0;
2014 if (skb == tcp_send_head(sk))
2015 return 0;
2016 /* Some heurestics for collapsing over SACK'd could be invented */
2017 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2018 return 0;
2019
2020 return 1;
2021}
2022
Andi Kleen67edfef2009-07-21 23:00:40 +00002023/* Collapse packets in the retransmit queue to make to create
2024 * less packets on the wire. This is only done on retransmission.
2025 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002026static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2027 int space)
2028{
2029 struct tcp_sock *tp = tcp_sk(sk);
2030 struct sk_buff *skb = to, *tmp;
2031 int first = 1;
2032
2033 if (!sysctl_tcp_retrans_collapse)
2034 return;
Changli Gaoa3433f32010-06-12 14:01:43 +00002035 if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002036 return;
2037
2038 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2039 if (!tcp_can_collapse(sk, skb))
2040 break;
2041
2042 space -= skb->len;
2043
2044 if (first) {
2045 first = 0;
2046 continue;
2047 }
2048
2049 if (space < 0)
2050 break;
2051 /* Punt if not enough space exists in the first SKB for
2052 * the data in the second
2053 */
2054 if (skb->len > skb_tailroom(to))
2055 break;
2056
2057 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2058 break;
2059
2060 tcp_collapse_retrans(sk, to);
2061 }
2062}
2063
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064/* This retransmits one SKB. Policy decisions and retransmit queue
2065 * state updates are done by the caller. Returns non-zero if an
2066 * error occurred which prevented the send.
2067 */
2068int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2069{
2070 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002071 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002072 unsigned int cur_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 int err;
2074
John Heffner5d424d52006-03-20 17:53:41 -08002075 /* Inconslusive MTU probe */
2076 if (icsk->icsk_mtup.probe_size) {
2077 icsk->icsk_mtup.probe_size = 0;
2078 }
2079
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002081 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 */
2083 if (atomic_read(&sk->sk_wmem_alloc) >
2084 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2085 return -EAGAIN;
2086
2087 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2088 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2089 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2091 return -ENOMEM;
2092 }
2093
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002094 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2095 return -EHOSTUNREACH; /* Routing failure or similar. */
2096
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002097 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002098
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 /* If receiver has shrunk his window, and skb is out of
2100 * new window, do not retransmit it. The exception is the
2101 * case, when window is shrunk to zero. In this case
2102 * our retransmit serves as a zero window probe.
2103 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002104 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2105 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 return -EAGAIN;
2107
2108 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07002109 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002111 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002112 int oldpcount = tcp_skb_pcount(skb);
2113
2114 if (unlikely(oldpcount > 1)) {
2115 tcp_init_tso_segs(sk, skb, cur_mss);
2116 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2117 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 }
2119
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002120 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 /* Some Solaris stacks overoptimize and ignore the FIN on a
2123 * retransmit when old data is attached. So strip it off
2124 * since it is cheap to do so and saves bytes on the network.
2125 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002126 if (skb->len > 0 &&
Changli Gaoa3433f32010-06-12 14:01:43 +00002127 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002128 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002130 /* Reuse, even though it does some unnecessary work */
2131 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2132 TCP_SKB_CB(skb)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 }
2135 }
2136
2137 /* Make a copy, if the first transmission SKB clone we made
2138 * is still in somebody's hands, else make a clone.
2139 */
2140 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002142 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143
2144 if (err == 0) {
2145 /* Update global TCP statistics. */
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002146 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147
2148 tp->total_retrans++;
2149
2150#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002151 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 if (net_ratelimit())
2153 printk(KERN_DEBUG "retrans_out leaked.\n");
2154 }
2155#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002156 if (!tp->retrans_out)
2157 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2159 tp->retrans_out += tcp_skb_pcount(skb);
2160
2161 /* Save stamp of the first retransmit. */
2162 if (!tp->retrans_stamp)
2163 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2164
2165 tp->undo_retrans++;
2166
2167 /* snd_nxt is stored to detect loss of retransmitted segment,
2168 * see tcp_input.c tcp_sacktag_write_queue().
2169 */
2170 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2171 }
2172 return err;
2173}
2174
Andi Kleen67edfef2009-07-21 23:00:40 +00002175/* Check if we forward retransmits are possible in the current
2176 * window/congestion state.
2177 */
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002178static int tcp_can_forward_retransmit(struct sock *sk)
2179{
2180 const struct inet_connection_sock *icsk = inet_csk(sk);
2181 struct tcp_sock *tp = tcp_sk(sk);
2182
2183 /* Forward retransmissions are possible only during Recovery. */
2184 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2185 return 0;
2186
2187 /* No forward retransmissions in Reno are possible. */
2188 if (tcp_is_reno(tp))
2189 return 0;
2190
2191 /* Yeah, we have to make difficult choice between forward transmission
2192 * and retransmission... Both ways have their merits...
2193 *
2194 * For now we do not retransmit anything, while we have some new
2195 * segments to send. In the other cases, follow rule 3 for
2196 * NextSeg() specified in RFC3517.
2197 */
2198
2199 if (tcp_may_send_now(sk))
2200 return 0;
2201
2202 return 1;
2203}
2204
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205/* This gets called after a retransmit timeout, and the initially
2206 * retransmitted data is acknowledged. It tries to continue
2207 * resending the rest of the retransmit queue, until either
2208 * we've sent it all or the congestion window limit is reached.
2209 * If doing SACK, the first ACK which comes back for a timeout
2210 * based retransmit packet might feed us FACK information again.
2211 * If so, we use it to avoid unnecessarily retransmissions.
2212 */
2213void tcp_xmit_retransmit_queue(struct sock *sk)
2214{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002215 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 struct tcp_sock *tp = tcp_sk(sk);
2217 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002218 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002219 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002220 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002221 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002222
Ilpo Järvinen45e77d32010-07-19 01:16:18 +00002223 if (!tp->packets_out)
2224 return;
2225
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002226 if (!tp->lost_out)
2227 tp->retransmit_high = tp->snd_una;
2228
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002229 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002230 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002231 last_lost = TCP_SKB_CB(skb)->end_seq;
2232 if (after(last_lost, tp->retransmit_high))
2233 last_lost = tp->retransmit_high;
2234 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002235 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002236 last_lost = tp->snd_una;
2237 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002239 tcp_for_write_queue_from(skb, sk) {
2240 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002242 if (skb == tcp_send_head(sk))
2243 break;
2244 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002245 if (hole == NULL)
2246 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002247
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002248 /* Assume this retransmit will generate
2249 * only one packet for congestion window
2250 * calculation purposes. This works because
2251 * tcp_retransmit_skb() will chop up the
2252 * packet to be MSS sized and all the
2253 * packet counting works out.
2254 */
2255 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2256 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002257
2258 if (fwd_rexmitting) {
2259begin_fwd:
2260 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2261 break;
2262 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2263
2264 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002265 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002266 if (!tcp_can_forward_retransmit(sk))
2267 break;
2268 /* Backtrack if necessary to non-L'ed skb */
2269 if (hole != NULL) {
2270 skb = hole;
2271 hole = NULL;
2272 }
2273 fwd_rexmitting = 1;
2274 goto begin_fwd;
2275
2276 } else if (!(sacked & TCPCB_LOST)) {
Ilpo Järvinenac11ba72009-02-28 04:44:27 +00002277 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002278 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002279 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002281 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002282 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002283 if (icsk->icsk_ca_state != TCP_CA_Loss)
2284 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2285 else
2286 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2287 }
2288
2289 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002290 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002291
Ilpo Järvinenf0ceb0e2008-09-20 21:24:49 -07002292 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002293 return;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002294 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2295
2296 if (skb == tcp_write_queue_head(sk))
2297 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2298 inet_csk(sk)->icsk_rto,
2299 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301}
2302
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303/* Send a fin. The caller locks the socket for us. This cannot be
2304 * allowed to fail queueing a FIN frame under any circumstances.
2305 */
2306void tcp_send_fin(struct sock *sk)
2307{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002308 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002309 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002311
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 /* Optimization, tack on the FIN if we have a queue of
2313 * unsent frames. But be careful about outgoing SACKS
2314 * and IP options.
2315 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002316 mss_now = tcp_current_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317
David S. Millerfe067e82007-03-07 12:12:44 -08002318 if (tcp_send_head(sk) != NULL) {
Changli Gaoa3433f32010-06-12 14:01:43 +00002319 TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 TCP_SKB_CB(skb)->end_seq++;
2321 tp->write_seq++;
2322 } else {
2323 /* Socket is locked, keep trying until memory is available. */
2324 for (;;) {
Wu Fengguangaa133072009-09-02 23:45:45 -07002325 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2326 sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 if (skb)
2328 break;
2329 yield();
2330 }
2331
2332 /* Reserve space for headers and prepare control bits. */
2333 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002335 tcp_init_nondata_skb(skb, tp->write_seq,
Changli Gaoa3433f32010-06-12 14:01:43 +00002336 TCPHDR_ACK | TCPHDR_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 tcp_queue_skb(sk, skb);
2338 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002339 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340}
2341
2342/* We get here when a process closes a file descriptor (either due to
2343 * an explicit close() or as a byproduct of exit()'ing) and there
2344 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002345 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 */
Al Virodd0fc662005-10-07 07:46:04 +01002347void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 struct sk_buff *skb;
2350
2351 /* NOTE: No TCP options attached and we never retransmit this. */
2352 skb = alloc_skb(MAX_TCP_HEADER, priority);
2353 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002354 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 return;
2356 }
2357
2358 /* Reserve space for headers and prepare control bits. */
2359 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002360 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
Changli Gaoa3433f32010-06-12 14:01:43 +00002361 TCPHDR_ACK | TCPHDR_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002364 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002365 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002366
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002367 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368}
2369
Andi Kleen67edfef2009-07-21 23:00:40 +00002370/* Send a crossed SYN-ACK during socket establishment.
2371 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 * a SYN packet that crossed the incoming SYN that caused this routine
2373 * to get called. If this assumption fails then the initial rcv_wnd
2374 * and rcv_wscale values will not be correct.
2375 */
2376int tcp_send_synack(struct sock *sk)
2377{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002378 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379
David S. Millerfe067e82007-03-07 12:12:44 -08002380 skb = tcp_write_queue_head(sk);
Changli Gaoa3433f32010-06-12 14:01:43 +00002381 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2383 return -EFAULT;
2384 }
Changli Gaoa3433f32010-06-12 14:01:43 +00002385 if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 if (skb_cloned(skb)) {
2387 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2388 if (nskb == NULL)
2389 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002390 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002392 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002393 sk_wmem_free_skb(sk, skb);
2394 sk->sk_wmem_queued += nskb->truesize;
2395 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 skb = nskb;
2397 }
2398
Changli Gaoa3433f32010-06-12 14:01:43 +00002399 TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 TCP_ECN_send_synack(tcp_sk(sk), skb);
2401 }
2402 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002403 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404}
2405
Andi Kleen67edfef2009-07-21 23:00:40 +00002406/* Prepare a SYN-ACK. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002407struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002408 struct request_sock *req,
2409 struct request_values *rvp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002411 struct tcp_out_options opts;
William Allen Simpson4957faa2009-12-02 18:25:27 +00002412 struct tcp_extend_values *xvp = tcp_xv(rvp);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002413 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet28b27742010-03-08 11:32:01 -08002415 const struct tcp_cookie_values *cvp = tp->cookie_values;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002418 struct tcp_md5sig_key *md5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002419 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002420 int mss;
Eric Dumazet28b27742010-03-08 11:32:01 -08002421 int s_data_desired = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
Eric Dumazet28b27742010-03-08 11:32:01 -08002423 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2424 s_data_desired = cvp->s_data_desired;
2425 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 if (skb == NULL)
2427 return NULL;
2428
2429 /* Reserve space for headers. */
2430 skb_reserve(skb, MAX_TCP_HEADER);
2431
Eric Dumazetadf30902009-06-02 05:19:30 +00002432 skb_dst_set(skb, dst_clone(dst));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433
David S. Miller0dbaee32010-12-13 12:52:14 -08002434 mss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002435 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2436 mss = tp->rx_opt.user_mss;
2437
Adam Langley33ad7982008-07-19 00:04:31 -07002438 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2439 __u8 rcv_wscale;
2440 /* Set this up on the first call only */
2441 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00002442
2443 /* limit the window selection if the user enforce a smaller rx buffer */
2444 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2445 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2446 req->window_clamp = tcp_full_space(sk);
2447
Adam Langley33ad7982008-07-19 00:04:31 -07002448 /* tcp_full_space because it is guaranteed to be the first packet */
2449 tcp_select_initial_window(tcp_full_space(sk),
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002450 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Adam Langley33ad7982008-07-19 00:04:31 -07002451 &req->rcv_wnd,
2452 &req->window_clamp,
2453 ireq->wscale_ok,
laurent chavey31d12922009-12-15 11:15:28 +00002454 &rcv_wscale,
2455 dst_metric(dst, RTAX_INITRWND));
Adam Langley33ad7982008-07-19 00:04:31 -07002456 ireq->rcv_wscale = rcv_wscale;
2457 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002458
Adam Langley33ad7982008-07-19 00:04:31 -07002459 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002460#ifdef CONFIG_SYN_COOKIES
2461 if (unlikely(req->cookie_ts))
2462 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2463 else
2464#endif
Adam Langley33ad7982008-07-19 00:04:31 -07002465 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002466 tcp_header_size = tcp_synack_options(sk, req, mss,
William Allen Simpson4957faa2009-12-02 18:25:27 +00002467 skb, &opts, &md5, xvp)
2468 + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002469
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002470 skb_push(skb, tcp_header_size);
2471 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002473 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 memset(th, 0, sizeof(struct tcphdr));
2475 th->syn = 1;
2476 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 TCP_ECN_make_synack(req, th);
KOVACS Krisztiana3116ac2008-10-01 07:46:49 -07002478 th->source = ireq->loc_port;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002479 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002480 /* Setting of flags are superfluous here for callers (and ECE is
2481 * not even correctly set)
2482 */
2483 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
Changli Gaoa3433f32010-06-12 14:01:43 +00002484 TCPHDR_SYN | TCPHDR_ACK);
William Allen Simpson4957faa2009-12-02 18:25:27 +00002485
2486 if (OPTION_COOKIE_EXTENSION & opts.options) {
Eric Dumazet28b27742010-03-08 11:32:01 -08002487 if (s_data_desired) {
2488 u8 *buf = skb_put(skb, s_data_desired);
William Allen Simpson4957faa2009-12-02 18:25:27 +00002489
2490 /* copy data directly from the listening socket. */
Eric Dumazet28b27742010-03-08 11:32:01 -08002491 memcpy(buf, cvp->s_data_payload, s_data_desired);
2492 TCP_SKB_CB(skb)->end_seq += s_data_desired;
William Allen Simpson4957faa2009-12-02 18:25:27 +00002493 }
2494
2495 if (opts.hash_size > 0) {
2496 __u32 workspace[SHA_WORKSPACE_WORDS];
2497 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2498 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2499
2500 /* Secret recipe depends on the Timestamp, (future)
2501 * Sequence and Acknowledgment Numbers, Initiator
2502 * Cookie, and others handled by IP variant caller.
2503 */
2504 *tail-- ^= opts.tsval;
2505 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2506 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2507
2508 /* recommended */
Eric Dumazet0eae88f2010-04-20 19:06:52 -07002509 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
David S. Millerf9a2e692009-12-02 22:12:04 -08002510 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
William Allen Simpson4957faa2009-12-02 18:25:27 +00002511
2512 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2513 (char *)mess,
2514 &workspace[0]);
2515 opts.hash_location =
2516 (__u8 *)&xvp->cookie_bakery[0];
2517 }
2518 }
2519
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002521 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
2523 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002524 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002525 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 th->doff = (tcp_header_size >> 2);
Tom Herbertaa2ea052010-04-22 07:00:24 +00002527 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002528
2529#ifdef CONFIG_TCP_MD5SIG
2530 /* Okay, we have all we need - do the md5 hash if needed */
2531 if (md5) {
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002532 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002533 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002534 }
2535#endif
2536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 return skb;
2538}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002539EXPORT_SYMBOL(tcp_make_synack);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540
Andi Kleen67edfef2009-07-21 23:00:40 +00002541/* Do all connect socket setups that can be done AF independent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002542static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543{
2544 struct dst_entry *dst = __sk_dst_get(sk);
2545 struct tcp_sock *tp = tcp_sk(sk);
2546 __u8 rcv_wscale;
2547
2548 /* We'll fix this up when we get a response from the other end.
2549 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2550 */
2551 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08002552 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002554#ifdef CONFIG_TCP_MD5SIG
2555 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2556 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2557#endif
2558
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 /* If user gave his TCP_MAXSEG, record it to clamp */
2560 if (tp->rx_opt.user_mss)
2561 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2562 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002563 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 tcp_sync_mss(sk, dst_mtu(dst));
2565
2566 if (!tp->window_clamp)
2567 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
David S. Miller0dbaee32010-12-13 12:52:14 -08002568 tp->advmss = dst_metric_advmss(dst);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002569 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2570 tp->advmss = tp->rx_opt.user_mss;
2571
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573
Hagen Paul Pfeifere88c64f2010-08-19 06:33:05 +00002574 /* limit the window selection if the user enforce a smaller rx buffer */
2575 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2576 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2577 tp->window_clamp = tcp_full_space(sk);
2578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 tcp_select_initial_window(tcp_full_space(sk),
2580 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2581 &tp->rcv_wnd,
2582 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08002583 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00002584 &rcv_wscale,
2585 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586
2587 tp->rx_opt.rcv_wscale = rcv_wscale;
2588 tp->rcv_ssthresh = tp->rcv_wnd;
2589
2590 sk->sk_err = 0;
2591 sock_reset_flag(sk, SOCK_DONE);
2592 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08002593 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 tp->snd_una = tp->write_seq;
2595 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002596 tp->snd_up = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 tp->rcv_nxt = 0;
2598 tp->rcv_wup = 0;
2599 tp->copied_seq = 0;
2600
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002601 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2602 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 tcp_clear_retrans(tp);
2604}
2605
Andi Kleen67edfef2009-07-21 23:00:40 +00002606/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607int tcp_connect(struct sock *sk)
2608{
2609 struct tcp_sock *tp = tcp_sk(sk);
2610 struct sk_buff *buff;
Eric Parisee586812010-11-16 11:52:49 +00002611 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612
2613 tcp_connect_init(sk);
2614
David S. Millerd179cd12005-08-17 14:57:30 -07002615 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 if (unlikely(buff == NULL))
2617 return -ENOBUFS;
2618
2619 /* Reserve space for headers. */
2620 skb_reserve(buff, MAX_TCP_HEADER);
2621
Wei Yongjunbd37a082006-08-07 21:04:15 -07002622 tp->snd_nxt = tp->write_seq;
Changli Gaoa3433f32010-06-12 14:01:43 +00002623 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002624 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625
2626 /* Send it off. */
2627 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2628 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2629 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002630 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002631 sk->sk_wmem_queued += buff->truesize;
2632 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 tp->packets_out += tcp_skb_pcount(buff);
Eric Parisee586812010-11-16 11:52:49 +00002634 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2635 if (err == -ECONNREFUSED)
2636 return err;
Wei Yongjunbd37a082006-08-07 21:04:15 -07002637
2638 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2639 * in order to make this packet get counted in tcpOutSegs.
2640 */
2641 tp->snd_nxt = tp->write_seq;
2642 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002643 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644
2645 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002646 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2647 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 return 0;
2649}
Eric Dumazet4bc2f182010-07-09 21:22:10 +00002650EXPORT_SYMBOL(tcp_connect);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651
2652/* Send out a delayed ack, the caller does the policy checking
2653 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2654 * for details.
2655 */
2656void tcp_send_delayed_ack(struct sock *sk)
2657{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002658 struct inet_connection_sock *icsk = inet_csk(sk);
2659 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 unsigned long timeout;
2661
2662 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002663 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002664 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002666 if (icsk->icsk_ack.pingpong ||
2667 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668 max_ato = TCP_DELACK_MAX;
2669
2670 /* Slow path, intersegment interval is "high". */
2671
2672 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002673 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 * directly.
2675 */
2676 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002677 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678
2679 if (rtt < max_ato)
2680 max_ato = rtt;
2681 }
2682
2683 ato = min(ato, max_ato);
2684 }
2685
2686 /* Stay within the limit we were given */
2687 timeout = jiffies + ato;
2688
2689 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002690 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 /* If delack timer was blocked or is about to expire,
2692 * send ACK now.
2693 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002694 if (icsk->icsk_ack.blocked ||
2695 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 tcp_send_ack(sk);
2697 return;
2698 }
2699
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002700 if (!time_before(timeout, icsk->icsk_ack.timeout))
2701 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002703 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2704 icsk->icsk_ack.timeout = timeout;
2705 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706}
2707
2708/* This routine sends an ack and also updates the window. */
2709void tcp_send_ack(struct sock *sk)
2710{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002711 struct sk_buff *buff;
2712
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002714 if (sk->sk_state == TCP_CLOSE)
2715 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002717 /* We are not putting this on the write queue, so
2718 * tcp_transmit_skb() will set the ownership to this
2719 * sock.
2720 */
2721 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2722 if (buff == NULL) {
2723 inet_csk_schedule_ack(sk);
2724 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2725 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2726 TCP_DELACK_MAX, TCP_RTO_MAX);
2727 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002729
2730 /* Reserve space for headers and prepare control bits. */
2731 skb_reserve(buff, MAX_TCP_HEADER);
Changli Gaoa3433f32010-06-12 14:01:43 +00002732 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002733
2734 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002735 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2736 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737}
2738
2739/* This routine sends a packet with an out of date sequence
2740 * number. It assumes the other end will try to ack it.
2741 *
2742 * Question: what should we make while urgent mode?
2743 * 4.4BSD forces sending single byte of data. We cannot send
2744 * out of window data, because we have SND.NXT==SND.MAX...
2745 *
2746 * Current solution: to send TWO zero-length segments in urgent mode:
2747 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2748 * out-of-date with SND.UNA-1 to probe window.
2749 */
2750static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2751{
2752 struct tcp_sock *tp = tcp_sk(sk);
2753 struct sk_buff *skb;
2754
2755 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2756 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002757 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 return -1;
2759
2760 /* Reserve space for headers and set control bits. */
2761 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 /* Use a previous sequence. This should cause the other
2763 * end to send an ack. Don't queue or clone SKB, just
2764 * send it.
2765 */
Changli Gaoa3433f32010-06-12 14:01:43 +00002766 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002768 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769}
2770
Andi Kleen67edfef2009-07-21 23:00:40 +00002771/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772int tcp_write_wakeup(struct sock *sk)
2773{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002774 struct tcp_sock *tp = tcp_sk(sk);
2775 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002777 if (sk->sk_state == TCP_CLOSE)
2778 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002780 if ((skb = tcp_send_head(sk)) != NULL &&
2781 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2782 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002783 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002784 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002786 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2787 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002789 /* We are probing the opening of a window
2790 * but the window size is != 0
2791 * must have been a result SWS avoidance ( sender )
2792 */
2793 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2794 skb->len > mss) {
2795 seg_size = min(seg_size, mss);
Changli Gaoa3433f32010-06-12 14:01:43 +00002796 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002797 if (tcp_fragment(sk, skb, seg_size, mss))
2798 return -1;
2799 } else if (!tcp_skb_pcount(skb))
2800 tcp_set_skb_tso_segs(sk, skb, mss);
2801
Changli Gaoa3433f32010-06-12 14:01:43 +00002802 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002803 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2804 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2805 if (!err)
2806 tcp_event_new_data_sent(sk, skb);
2807 return err;
2808 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002809 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002810 tcp_xmit_probe_skb(sk, 1);
2811 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813}
2814
2815/* A window probe timeout has occurred. If window is not closed send
2816 * a partial packet else a zero probe.
2817 */
2818void tcp_send_probe0(struct sock *sk)
2819{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002820 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821 struct tcp_sock *tp = tcp_sk(sk);
2822 int err;
2823
2824 err = tcp_write_wakeup(sk);
2825
David S. Millerfe067e82007-03-07 12:12:44 -08002826 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002828 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002829 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002830 return;
2831 }
2832
2833 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002834 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2835 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002836 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002837 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002838 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2839 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 } else {
2841 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002842 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 * Let local senders to fight for local resources.
2844 *
2845 * Use accumulated backoff yet.
2846 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002847 if (!icsk->icsk_probes_out)
2848 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002849 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002850 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002851 TCP_RESOURCE_PROBE_INTERVAL),
2852 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 }
2854}