blob: b4ed957f201a6f3f2ebf8d5564edec567b6435a6 [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;
58int sysctl_tcp_base_mss __read_mostly = 512;
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
122 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
123 mss = dst_metric(dst, RTAX_ADVMSS);
124 tp->advmss = mss;
125 }
126
127 return (__u16)mss;
128}
129
130/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
131 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700132static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700134 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135 s32 delta = tcp_time_stamp - tp->lsndtime;
136 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
137 u32 cwnd = tp->snd_cwnd;
138
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300139 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300141 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 restart_cwnd = min(restart_cwnd, cwnd);
143
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700144 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145 cwnd >>= 1;
146 tp->snd_cwnd = max(cwnd, restart_cwnd);
147 tp->snd_cwnd_stamp = tcp_time_stamp;
148 tp->snd_cwnd_used = 0;
149}
150
Andi Kleen67edfef2009-07-21 23:00:40 +0000151/* Congestion state accounting after a packet has been sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800152static void tcp_event_data_sent(struct tcp_sock *tp,
153 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700155 struct inet_connection_sock *icsk = inet_csk(sk);
156 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157
David S. Miller35089bb2006-06-13 22:33:04 -0700158 if (sysctl_tcp_slow_start_after_idle &&
159 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
162 tp->lsndtime = now;
163
164 /* If it is a reply for ato after last received
165 * packet, enter pingpong mode.
166 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700167 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
168 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
170
Andi Kleen67edfef2009-07-21 23:00:40 +0000171/* Account for an ACK we sent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800172static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700174 tcp_dec_quickack_mode(sk, pkts);
175 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176}
177
178/* Determine a window scaling and initial window to offer.
179 * Based on the assumption that the given amount of space
180 * will be offered. Store the results in the tp structure.
181 * NOTE: for smooth operation initial space offering should
182 * be a multiple of mss if possible. We assume here that mss >= 1.
183 * This MUST be enforced by all callers.
184 */
185void tcp_select_initial_window(int __space, __u32 mss,
186 __u32 *rcv_wnd, __u32 *window_clamp,
laurent chavey31d12922009-12-15 11:15:28 +0000187 int wscale_ok, __u8 *rcv_wscale,
188 __u32 init_rcv_wnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189{
190 unsigned int space = (__space < 0 ? 0 : __space);
191
192 /* If no clamp set the clamp to the max possible scaled window */
193 if (*window_clamp == 0)
194 (*window_clamp) = (65535 << 14);
195 space = min(*window_clamp, space);
196
197 /* Quantize space offering to a multiple of mss if possible. */
198 if (space > mss)
199 space = (space / mss) * mss;
200
201 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800202 * will break some buggy TCP stacks. If the admin tells us
203 * it is likely we could be speaking with such a buggy stack
204 * we will truncate our initial window offering to 32K-1
205 * unless the remote has sent us a window scaling option,
206 * which we interpret as a sign the remote TCP is not
207 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208 */
Rick Jones15d99e02006-03-20 22:40:29 -0800209 if (sysctl_tcp_workaround_signed_windows)
210 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
211 else
212 (*rcv_wnd) = space;
213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214 (*rcv_wscale) = 0;
215 if (wscale_ok) {
216 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900217 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 */
219 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700220 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 while (space > 65535 && (*rcv_wscale) < 14) {
222 space >>= 1;
223 (*rcv_wscale)++;
224 }
225 }
226
227 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700228 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 * will be satisfied with 2.
230 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800231 if (mss > (1 << *rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700232 int init_cwnd = 4;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800233 if (mss > 1460 * 3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700234 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700235 else if (mss > 1460)
236 init_cwnd = 3;
laurent chavey31d12922009-12-15 11:15:28 +0000237 /* when initializing use the value from init_rcv_wnd
238 * rather than the default from above
239 */
240 if (init_rcv_wnd &&
241 (*rcv_wnd > init_rcv_wnd * mss))
242 *rcv_wnd = init_rcv_wnd * mss;
243 else if (*rcv_wnd > init_cwnd * mss)
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800244 *rcv_wnd = init_cwnd * mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245 }
246
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
249}
250
251/* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
254 * frame.
255 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800256static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257{
258 struct tcp_sock *tp = tcp_sk(sk);
259 u32 cur_win = tcp_receive_window(tp);
260 u32 new_win = __tcp_select_window(sk);
261
262 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800263 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
268 *
269 * Relax Will Robinson.
270 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700271 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 }
273 tp->rcv_wnd = new_win;
274 tp->rcv_wup = tp->rcv_nxt;
275
276 /* Make sure we do not exceed the maximum possible
277 * scaled window.
278 */
Rick Jones15d99e02006-03-20 22:40:29 -0800279 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 new_win = min(new_win, MAX_TCP_WINDOW);
281 else
282 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
283
284 /* RFC1323 scaling applied */
285 new_win >>= tp->rx_opt.rcv_wscale;
286
287 /* If we advertise zero window, disable fast path. */
288 if (new_win == 0)
289 tp->pred_flags = 0;
290
291 return new_win;
292}
293
Andi Kleen67edfef2009-07-21 23:00:40 +0000294/* Packet ECN state for a SYN-ACK */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800295static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700296{
297 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800298 if (!(tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700299 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
300}
301
Andi Kleen67edfef2009-07-21 23:00:40 +0000302/* Packet ECN state for a SYN. */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700303static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
304{
305 struct tcp_sock *tp = tcp_sk(sk);
306
307 tp->ecn_flags = 0;
Ilpo Järvinen255cac92009-05-04 11:07:36 -0700308 if (sysctl_tcp_ecn == 1) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800309 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700310 tp->ecn_flags = TCP_ECN_OK;
311 }
312}
313
314static __inline__ void
315TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
316{
317 if (inet_rsk(req)->ecn_ok)
318 th->ece = 1;
319}
320
Andi Kleen67edfef2009-07-21 23:00:40 +0000321/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
322 * be sent.
323 */
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700324static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
325 int tcp_header_len)
326{
327 struct tcp_sock *tp = tcp_sk(sk);
328
329 if (tp->ecn_flags & TCP_ECN_OK) {
330 /* Not-retransmitted data segment: set ECT and inject CWR. */
331 if (skb->len != tcp_header_len &&
332 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
333 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800334 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700335 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
336 tcp_hdr(skb)->cwr = 1;
337 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
338 }
339 } else {
340 /* ACK or retransmitted segment: clear ECT|CE */
341 INET_ECN_dontxmit(sk);
342 }
343 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
344 tcp_hdr(skb)->ece = 1;
345 }
346}
347
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800348/* Constructs common control bits of non-data skb. If SYN/FIN is present,
349 * auto increment end seqno.
350 */
351static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
352{
David S. Miller2e8e18e2010-04-08 11:32:30 -0700353 skb->ip_summed = CHECKSUM_PARTIAL;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800354 skb->csum = 0;
355
356 TCP_SKB_CB(skb)->flags = flags;
357 TCP_SKB_CB(skb)->sacked = 0;
358
359 skb_shinfo(skb)->gso_segs = 1;
360 skb_shinfo(skb)->gso_size = 0;
361 skb_shinfo(skb)->gso_type = 0;
362
363 TCP_SKB_CB(skb)->seq = seq;
364 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
365 seq++;
366 TCP_SKB_CB(skb)->end_seq = seq;
367}
368
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700369static inline int tcp_urg_mode(const struct tcp_sock *tp)
370{
371 return tp->snd_una != tp->snd_up;
372}
373
Adam Langley33ad7982008-07-19 00:04:31 -0700374#define OPTION_SACK_ADVERTISE (1 << 0)
375#define OPTION_TS (1 << 1)
376#define OPTION_MD5 (1 << 2)
Ori Finkelman89e95a62009-10-01 06:41:59 +0000377#define OPTION_WSCALE (1 << 3)
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000378#define OPTION_COOKIE_EXTENSION (1 << 4)
Adam Langley33ad7982008-07-19 00:04:31 -0700379
380struct tcp_out_options {
381 u8 options; /* bit field of OPTION_* */
382 u8 ws; /* window scale, 0 to disable */
383 u8 num_sack_blocks; /* number of SACK blocks to include */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000384 u8 hash_size; /* bytes in hash_location */
Adam Langley33ad7982008-07-19 00:04:31 -0700385 u16 mss; /* 0 to disable */
386 __u32 tsval, tsecr; /* need to include OPTION_TS */
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000387 __u8 *hash_location; /* temporary pointer, overloaded */
Adam Langley33ad7982008-07-19 00:04:31 -0700388};
389
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000390/* The sysctl int routines are generic, so check consistency here.
391 */
392static u8 tcp_cookie_size_check(u8 desired)
393{
394 if (desired > 0) {
395 /* previously specified */
396 return desired;
397 }
398 if (sysctl_tcp_cookie_size <= 0) {
399 /* no default specified */
400 return 0;
401 }
402 if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) {
403 /* value too small, specify minimum */
404 return TCP_COOKIE_MIN;
405 }
406 if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) {
407 /* value too large, specify maximum */
408 return TCP_COOKIE_MAX;
409 }
410 if (0x1 & sysctl_tcp_cookie_size) {
411 /* 8-bit multiple, illegal, fix it */
412 return (u8)(sysctl_tcp_cookie_size + 0x1);
413 }
414 return (u8)sysctl_tcp_cookie_size;
415}
416
Andi Kleen67edfef2009-07-21 23:00:40 +0000417/* Write previously computed TCP options to the packet.
418 *
419 * Beware: Something in the Internet is very sensitive to the ordering of
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700420 * TCP options, we learned this through the hard way, so be careful here.
421 * Luckily we can at least blame others for their non-compliance but from
422 * inter-operatibility perspective it seems that we're somewhat stuck with
423 * the ordering which we have been using if we want to keep working with
424 * those broken things (not that it currently hurts anybody as there isn't
425 * particular reason why the ordering would need to be changed).
426 *
427 * At least SACK_PERM as the first option is known to lead to a disaster
428 * (but it may well be that other scenarios fail similarly).
429 */
Adam Langley33ad7982008-07-19 00:04:31 -0700430static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000431 struct tcp_out_options *opts)
432{
433 u8 options = opts->options; /* mungable copy */
434
435 /* Having both authentication and cookies for security is redundant,
436 * and there's certainly not enough room. Instead, the cookie-less
437 * extension variant is proposed.
438 *
439 * Consider the pessimal case with authentication. The options
440 * could look like:
441 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
442 */
443 if (unlikely(OPTION_MD5 & options)) {
444 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
445 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
446 (TCPOLEN_COOKIE_BASE << 16) |
447 (TCPOPT_MD5SIG << 8) |
448 TCPOLEN_MD5SIG);
449 } else {
450 *ptr++ = htonl((TCPOPT_NOP << 24) |
451 (TCPOPT_NOP << 16) |
452 (TCPOPT_MD5SIG << 8) |
453 TCPOLEN_MD5SIG);
454 }
455 options &= ~OPTION_COOKIE_EXTENSION;
456 /* overload cookie hash location */
457 opts->hash_location = (__u8 *)ptr;
Adam Langley33ad7982008-07-19 00:04:31 -0700458 ptr += 4;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800459 }
Adam Langley33ad7982008-07-19 00:04:31 -0700460
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700461 if (unlikely(opts->mss)) {
462 *ptr++ = htonl((TCPOPT_MSS << 24) |
463 (TCPOLEN_MSS << 16) |
464 opts->mss);
465 }
466
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000467 if (likely(OPTION_TS & options)) {
468 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700469 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
470 (TCPOLEN_SACK_PERM << 16) |
471 (TCPOPT_TIMESTAMP << 8) |
472 TCPOLEN_TIMESTAMP);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000473 options &= ~OPTION_SACK_ADVERTISE;
Adam Langley33ad7982008-07-19 00:04:31 -0700474 } else {
475 *ptr++ = htonl((TCPOPT_NOP << 24) |
476 (TCPOPT_NOP << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
478 TCPOLEN_TIMESTAMP);
479 }
480 *ptr++ = htonl(opts->tsval);
481 *ptr++ = htonl(opts->tsecr);
482 }
483
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000484 /* Specification requires after timestamp, so do it now.
485 *
486 * Consider the pessimal case without authentication. The options
487 * could look like:
488 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
489 */
490 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
491 __u8 *cookie_copy = opts->hash_location;
492 u8 cookie_size = opts->hash_size;
493
494 /* 8-bit multiple handled in tcp_cookie_size_check() above,
495 * and elsewhere.
496 */
497 if (0x2 & cookie_size) {
498 __u8 *p = (__u8 *)ptr;
499
500 /* 16-bit multiple */
501 *p++ = TCPOPT_COOKIE;
502 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
503 *p++ = *cookie_copy++;
504 *p++ = *cookie_copy++;
505 ptr++;
506 cookie_size -= 2;
507 } else {
508 /* 32-bit multiple */
509 *ptr++ = htonl(((TCPOPT_NOP << 24) |
510 (TCPOPT_NOP << 16) |
511 (TCPOPT_COOKIE << 8) |
512 TCPOLEN_COOKIE_BASE) +
513 cookie_size);
514 }
515
516 if (cookie_size > 0) {
517 memcpy(ptr, cookie_copy, cookie_size);
518 ptr += (cookie_size / 4);
519 }
520 }
521
522 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700523 *ptr++ = htonl((TCPOPT_NOP << 24) |
524 (TCPOPT_NOP << 16) |
525 (TCPOPT_SACK_PERM << 8) |
526 TCPOLEN_SACK_PERM);
527 }
528
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000529 if (unlikely(OPTION_WSCALE & options)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700530 *ptr++ = htonl((TCPOPT_NOP << 24) |
531 (TCPOPT_WINDOW << 16) |
532 (TCPOLEN_WINDOW << 8) |
533 opts->ws);
534 }
535
536 if (unlikely(opts->num_sack_blocks)) {
537 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
538 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800539 int this_sack;
540
541 *ptr++ = htonl((TCPOPT_NOP << 24) |
542 (TCPOPT_NOP << 16) |
543 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700544 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800545 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800546
Adam Langley33ad7982008-07-19 00:04:31 -0700547 for (this_sack = 0; this_sack < opts->num_sack_blocks;
548 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800549 *ptr++ = htonl(sp[this_sack].start_seq);
550 *ptr++ = htonl(sp[this_sack].end_seq);
551 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800552
Ilpo Järvinen5861f8e2009-03-14 14:23:01 +0000553 tp->rx_opt.dsack = 0;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800554 }
555}
556
Andi Kleen67edfef2009-07-21 23:00:40 +0000557/* Compute TCP options for SYN packets. This is not the final
558 * network wire format yet.
559 */
Adam Langley33ad7982008-07-19 00:04:31 -0700560static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
561 struct tcp_out_options *opts,
562 struct tcp_md5sig_key **md5) {
563 struct tcp_sock *tp = tcp_sk(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000564 struct tcp_cookie_values *cvp = tp->cookie_values;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000565 unsigned remaining = MAX_TCP_OPTION_SPACE;
566 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
567 tcp_cookie_size_check(cvp->cookie_desired) :
568 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700569
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800570#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700571 *md5 = tp->af_specific->md5_lookup(sk, sk);
572 if (*md5) {
573 opts->options |= OPTION_MD5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000574 remaining -= TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800575 }
Adam Langley33ad7982008-07-19 00:04:31 -0700576#else
577 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800578#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700579
580 /* We always get an MSS option. The option bytes which will be seen in
581 * normal data packets should timestamps be used, must be in the MSS
582 * advertised. But we subtract them from tp->mss_cache so that
583 * calculations in tcp_sendmsg are simpler etc. So account for this
584 * fact here if necessary. If we don't do this correctly, as a
585 * receiver we won't recognize data packets as being full sized when we
586 * should, and thus we won't abide by the delayed ACK rules correctly.
587 * SACKs don't matter, we never delay an ACK when we have any of those
588 * going out. */
589 opts->mss = tcp_advertise_mss(sk);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000590 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700591
David S. Millerbb5b7c12009-12-15 20:56:42 -0800592 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700593 opts->options |= OPTION_TS;
594 opts->tsval = TCP_SKB_CB(skb)->when;
595 opts->tsecr = tp->rx_opt.ts_recent;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000596 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700597 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800598 if (likely(sysctl_tcp_window_scaling)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700599 opts->ws = tp->rx_opt.rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000600 opts->options |= OPTION_WSCALE;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000601 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700602 }
David S. Millerbb5b7c12009-12-15 20:56:42 -0800603 if (likely(sysctl_tcp_sack)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700604 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700605 if (unlikely(!(OPTION_TS & opts->options)))
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000606 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700607 }
608
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000609 /* Note that timestamps are required by the specification.
610 *
611 * Odd numbers of bytes are prohibited by the specification, ensuring
612 * that the cookie is 16-bit aligned, and the resulting cookie pair is
613 * 32-bit aligned.
614 */
615 if (*md5 == NULL &&
616 (OPTION_TS & opts->options) &&
617 cookie_size > 0) {
618 int need = TCPOLEN_COOKIE_BASE + cookie_size;
619
620 if (0x2 & need) {
621 /* 32-bit multiple */
622 need += 2; /* NOPs */
623
624 if (need > remaining) {
625 /* try shrinking cookie to fit */
626 cookie_size -= 2;
627 need -= 4;
628 }
629 }
630 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
631 cookie_size -= 4;
632 need -= 4;
633 }
634 if (TCP_COOKIE_MIN <= cookie_size) {
635 opts->options |= OPTION_COOKIE_EXTENSION;
636 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
637 opts->hash_size = cookie_size;
638
639 /* Remember for future incarnations. */
640 cvp->cookie_desired = cookie_size;
641
642 if (cvp->cookie_desired != cvp->cookie_pair_size) {
643 /* Currently use random bytes as a nonce,
644 * assuming these are completely unpredictable
645 * by hostile users of the same system.
646 */
647 get_random_bytes(&cvp->cookie_pair[0],
648 cookie_size);
649 cvp->cookie_pair_size = cookie_size;
650 }
651
652 remaining -= need;
653 }
654 }
655 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700656}
657
Andi Kleen67edfef2009-07-21 23:00:40 +0000658/* Set up TCP options for SYN-ACKs. */
Adam Langley33ad7982008-07-19 00:04:31 -0700659static unsigned tcp_synack_options(struct sock *sk,
660 struct request_sock *req,
661 unsigned mss, struct sk_buff *skb,
662 struct tcp_out_options *opts,
William Allen Simpson4957faa2009-12-02 18:25:27 +0000663 struct tcp_md5sig_key **md5,
664 struct tcp_extend_values *xvp)
665{
Adam Langley33ad7982008-07-19 00:04:31 -0700666 struct inet_request_sock *ireq = inet_rsk(req);
William Allen Simpson4957faa2009-12-02 18:25:27 +0000667 unsigned remaining = MAX_TCP_OPTION_SPACE;
668 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
669 xvp->cookie_plus :
670 0;
Adam Langley33ad7982008-07-19 00:04:31 -0700671
672#ifdef CONFIG_TCP_MD5SIG
673 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
674 if (*md5) {
675 opts->options |= OPTION_MD5;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000676 remaining -= TCPOLEN_MD5SIG_ALIGNED;
677
678 /* We can't fit any SACK blocks in a packet with MD5 + TS
679 * options. There was discussion about disabling SACK
680 * rather than TS in order to fit in better with old,
681 * buggy kernels, but that was deemed to be unnecessary.
682 */
Eric Dumazetde213e52010-05-17 22:35:36 -0700683 ireq->tstamp_ok &= !ireq->sack_ok;
Adam Langley33ad7982008-07-19 00:04:31 -0700684 }
685#else
686 *md5 = NULL;
687#endif
688
William Allen Simpson4957faa2009-12-02 18:25:27 +0000689 /* We always send an MSS option. */
Adam Langley33ad7982008-07-19 00:04:31 -0700690 opts->mss = mss;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000691 remaining -= TCPOLEN_MSS_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700692
693 if (likely(ireq->wscale_ok)) {
694 opts->ws = ireq->rcv_wscale;
Ori Finkelman89e95a62009-10-01 06:41:59 +0000695 opts->options |= OPTION_WSCALE;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000696 remaining -= TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700697 }
Eric Dumazetde213e52010-05-17 22:35:36 -0700698 if (likely(ireq->tstamp_ok)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700699 opts->options |= OPTION_TS;
700 opts->tsval = TCP_SKB_CB(skb)->when;
701 opts->tsecr = req->ts_recent;
William Allen Simpson4957faa2009-12-02 18:25:27 +0000702 remaining -= TCPOLEN_TSTAMP_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700703 }
704 if (likely(ireq->sack_ok)) {
705 opts->options |= OPTION_SACK_ADVERTISE;
Eric Dumazetde213e52010-05-17 22:35:36 -0700706 if (unlikely(!ireq->tstamp_ok))
William Allen Simpson4957faa2009-12-02 18:25:27 +0000707 remaining -= TCPOLEN_SACKPERM_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700708 }
709
William Allen Simpson4957faa2009-12-02 18:25:27 +0000710 /* Similar rationale to tcp_syn_options() applies here, too.
711 * If the <SYN> options fit, the same options should fit now!
712 */
713 if (*md5 == NULL &&
Eric Dumazetde213e52010-05-17 22:35:36 -0700714 ireq->tstamp_ok &&
William Allen Simpson4957faa2009-12-02 18:25:27 +0000715 cookie_plus > TCPOLEN_COOKIE_BASE) {
716 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
717
718 if (0x2 & need) {
719 /* 32-bit multiple */
720 need += 2; /* NOPs */
721 }
722 if (need <= remaining) {
723 opts->options |= OPTION_COOKIE_EXTENSION;
724 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
725 remaining -= need;
726 } else {
727 /* There's no error return, so flag it. */
728 xvp->cookie_out_never = 1; /* true */
729 opts->hash_size = 0;
730 }
731 }
732 return MAX_TCP_OPTION_SPACE - remaining;
Adam Langley33ad7982008-07-19 00:04:31 -0700733}
734
Andi Kleen67edfef2009-07-21 23:00:40 +0000735/* Compute TCP options for ESTABLISHED sockets. This is not the
736 * final wire format yet.
737 */
Adam Langley33ad7982008-07-19 00:04:31 -0700738static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
739 struct tcp_out_options *opts,
740 struct tcp_md5sig_key **md5) {
741 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
742 struct tcp_sock *tp = tcp_sk(sk);
743 unsigned size = 0;
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000744 unsigned int eff_sacks;
Adam Langley33ad7982008-07-19 00:04:31 -0700745
746#ifdef CONFIG_TCP_MD5SIG
747 *md5 = tp->af_specific->md5_lookup(sk, sk);
748 if (unlikely(*md5)) {
749 opts->options |= OPTION_MD5;
750 size += TCPOLEN_MD5SIG_ALIGNED;
751 }
752#else
753 *md5 = NULL;
754#endif
755
756 if (likely(tp->rx_opt.tstamp_ok)) {
757 opts->options |= OPTION_TS;
758 opts->tsval = tcb ? tcb->when : 0;
759 opts->tsecr = tp->rx_opt.ts_recent;
760 size += TCPOLEN_TSTAMP_ALIGNED;
761 }
762
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000763 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
764 if (unlikely(eff_sacks)) {
Adam Langley33ad7982008-07-19 00:04:31 -0700765 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
766 opts->num_sack_blocks =
Ilpo Järvinencabeccb2009-02-28 04:44:38 +0000767 min_t(unsigned, eff_sacks,
Adam Langley33ad7982008-07-19 00:04:31 -0700768 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
769 TCPOLEN_SACK_PERBLOCK);
770 size += TCPOLEN_SACK_BASE_ALIGNED +
771 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
772 }
773
774 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800775}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776
777/* This routine actually transmits TCP packets queued in by
778 * tcp_do_sendmsg(). This is used by both the initial
779 * transmission and possible later retransmissions.
780 * All SKB's seen here are completely headerless. It is our
781 * job to build the TCP header, and pass the packet down to
782 * IP so it can do the same plus pass the packet off to the
783 * device.
784 *
785 * We are working here with either a clone of the original
786 * SKB, or a fresh unique copy made by the retransmit engine.
787 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800788static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
789 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800791 const struct inet_connection_sock *icsk = inet_csk(sk);
792 struct inet_sock *inet;
793 struct tcp_sock *tp;
794 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700795 struct tcp_out_options opts;
796 unsigned tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800797 struct tcp_md5sig_key *md5;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800798 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800799 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800801 BUG_ON(!skb || !tcp_skb_pcount(skb));
802
803 /* If congestion control is doing timestamping, we must
804 * take such a timestamp before we potentially clone/copy.
805 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700806 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800807 __net_timestamp(skb);
808
809 if (likely(clone_it)) {
810 if (unlikely(skb_cloned(skb)))
811 skb = pskb_copy(skb, gfp_mask);
812 else
813 skb = skb_clone(skb, gfp_mask);
814 if (unlikely(!skb))
815 return -ENOBUFS;
816 }
817
818 inet = inet_sk(sk);
819 tp = tcp_sk(sk);
820 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700821 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822
Adam Langley33ad7982008-07-19 00:04:31 -0700823 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
824 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
825 else
826 tcp_options_size = tcp_established_options(sk, skb, &opts,
827 &md5);
828 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900829
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800830 if (tcp_packets_in_flight(tp) == 0)
831 tcp_ca_event(sk, CA_EVENT_TX_START);
832
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700833 skb_push(skb, tcp_header_size);
834 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800835 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800836
837 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700838 th = tcp_hdr(skb);
Eric Dumazetc720c7e2009-10-15 06:30:45 +0000839 th->source = inet->inet_sport;
840 th->dest = inet->inet_dport;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800841 th->seq = htonl(tcb->seq);
842 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700843 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800844 tcb->flags);
845
846 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
847 /* RFC1323: The window in SYN & SYN/ACK segments
848 * is never scaled.
849 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800850 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800851 } else {
852 th->window = htons(tcp_select_window(sk));
853 }
854 th->check = 0;
855 th->urg_ptr = 0;
856
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700857 /* The urg_mode check is necessary during a below snd_una win probe */
Herbert Xu76913672009-02-21 23:52:29 -0800858 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
859 if (before(tp->snd_up, tcb->seq + 0x10000)) {
860 th->urg_ptr = htons(tp->snd_up - tcb->seq);
861 th->urg = 1;
862 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
Eric Dumazet0eae88f2010-04-20 19:06:52 -0700863 th->urg_ptr = htons(0xFFFF);
Herbert Xu76913672009-02-21 23:52:29 -0800864 th->urg = 1;
865 }
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800866 }
867
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000868 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Adam Langley33ad7982008-07-19 00:04:31 -0700869 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700870 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800871
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800872#ifdef CONFIG_TCP_MD5SIG
873 /* Calculate the MD5 hash, as we have all we need now */
874 if (md5) {
Eric Dumazeta4654192010-05-16 00:36:33 -0700875 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
William Allen Simpsonbd0388a2009-12-02 18:23:05 +0000876 tp->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700877 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800878 }
879#endif
880
Herbert Xubb296242010-04-11 02:15:55 +0000881 icsk->icsk_af_ops->send_check(sk, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800882
883 if (likely(tcb->flags & TCPCB_FLAG_ACK))
884 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
885
886 if (skb->len != tcp_header_size)
887 tcp_event_data_sent(tp, skb, sk);
888
Wei Yongjunbd37a082006-08-07 21:04:15 -0700889 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Tom Herbertaa2ea052010-04-22 07:00:24 +0000890 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
891 tcp_skb_pcount(skb));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800892
Shan Wei4e15ed42010-04-15 16:43:08 +0000893 err = icsk->icsk_af_ops->queue_xmit(skb);
Hua Zhong83de47c2006-04-28 15:26:50 -0700894 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800895 return err;
896
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800897 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800898
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200899 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900}
901
Andi Kleen67edfef2009-07-21 23:00:40 +0000902/* This routine just queues the buffer for sending.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903 *
904 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
905 * otherwise socket can stall.
906 */
907static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
908{
909 struct tcp_sock *tp = tcp_sk(sk);
910
911 /* Advance write_seq and place onto the write_queue. */
912 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
913 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800914 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800915 sk->sk_wmem_queued += skb->truesize;
916 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917}
918
Andi Kleen67edfef2009-07-21 23:00:40 +0000919/* Initialize TSO segments for a packet. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800920static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
921 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700922{
Herbert Xu8e5b9dd2009-06-28 18:03:30 +0000923 if (skb->len <= mss_now || !sk_can_gso(sk) ||
924 skb->ip_summed == CHECKSUM_NONE) {
David S. Millerf6302d12005-07-05 15:18:03 -0700925 /* Avoid the costly divide in the normal
926 * non-TSO case.
927 */
Herbert Xu79671682006-06-22 02:40:14 -0700928 skb_shinfo(skb)->gso_segs = 1;
929 skb_shinfo(skb)->gso_size = 0;
930 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700931 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700932 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700933 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700934 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 }
936}
937
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700938/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800939 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700940 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800941static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700942 int decr)
943{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800944 struct tcp_sock *tp = tcp_sk(sk);
945
Ilpo Järvinendc869672007-10-01 15:27:19 -0700946 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700947 return;
948
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200949 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700950 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700951}
952
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000953/* Pcount in the middle of the write queue got changed, we need to do various
954 * tweaks to fix counters
955 */
956static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
957{
958 struct tcp_sock *tp = tcp_sk(sk);
959
960 tp->packets_out -= decr;
961
962 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
963 tp->sacked_out -= decr;
964 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
965 tp->retrans_out -= decr;
966 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
967 tp->lost_out -= decr;
968
969 /* Reno case is special. Sigh... */
970 if (tcp_is_reno(tp) && decr > 0)
971 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
972
973 tcp_adjust_fackets_out(sk, skb, decr);
974
975 if (tp->lost_skb_hint &&
976 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
Ilpo Järvinen52cf3cc2009-04-18 05:48:48 +0000977 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
Ilpo Järvinen797108d2009-04-01 23:15:17 +0000978 tp->lost_cnt_hint -= decr;
979
980 tcp_verify_left_out(tp);
981}
982
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983/* Function to create two new TCP segments. Shrinks the given segment
984 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900985 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 * Remember, these are still headerless SKBs at this point.
987 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800988int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
989 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990{
991 struct tcp_sock *tp = tcp_sk(sk);
992 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700993 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700994 int nlen;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +0000995 u8 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996
Herbert Xub2cc99f2005-10-20 17:13:13 -0200997 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800998
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 nsize = skb_headlen(skb) - len;
1000 if (nsize < 0)
1001 nsize = 0;
1002
1003 if (skb_cloned(skb) &&
1004 skb_is_nonlinear(skb) &&
1005 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1006 return -ENOMEM;
1007
1008 /* Get a new skb... force flag on. */
1009 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1010 if (buff == NULL)
1011 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -07001012
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001013 sk->sk_wmem_queued += buff->truesize;
1014 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001015 nlen = skb->len - len - nsize;
1016 buff->truesize += nlen;
1017 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018
1019 /* Correct the sequence numbers. */
1020 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1021 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1022 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1023
1024 /* PSH and FIN should only be set in the second packet. */
1025 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001026 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -07001028 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029
Patrick McHardy84fa7932006-08-29 16:44:56 -07001030 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001032 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1033 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 nsize, 0);
1035
1036 skb_trim(skb, len);
1037
1038 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1039 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -07001040 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 skb_split(skb, buff, len);
1042 }
1043
1044 buff->ip_summed = skb->ip_summed;
1045
1046 /* Looks stupid, but our code really uses when of
1047 * skbs, which it never sent before. --ANK
1048 */
1049 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -07001050 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051
David S. Miller6475be12005-09-01 22:47:01 -07001052 old_factor = tcp_skb_pcount(skb);
1053
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001055 tcp_set_skb_tso_segs(sk, skb, mss_now);
1056 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
David S. Miller6475be12005-09-01 22:47:01 -07001058 /* If this packet has been sent out already, we must
1059 * adjust the various packet counters.
1060 */
Herbert Xucf0b4502005-09-08 15:10:52 -07001061 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -07001062 int diff = old_factor - tcp_skb_pcount(skb) -
1063 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001065 if (diff)
1066 tcp_adjust_pcount(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067 }
1068
1069 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -07001070 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001071 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072
1073 return 0;
1074}
1075
1076/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1077 * eventually). The difference is that pulled data not copied, but
1078 * immediately discarded.
1079 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001080static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001081{
1082 int i, k, eat;
1083
1084 eat = len;
1085 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001086 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 if (skb_shinfo(skb)->frags[i].size <= eat) {
1088 put_page(skb_shinfo(skb)->frags[i].page);
1089 eat -= skb_shinfo(skb)->frags[i].size;
1090 } else {
1091 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1092 if (eat) {
1093 skb_shinfo(skb)->frags[k].page_offset += eat;
1094 skb_shinfo(skb)->frags[k].size -= eat;
1095 eat = 0;
1096 }
1097 k++;
1098 }
1099 }
1100 skb_shinfo(skb)->nr_frags = k;
1101
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -07001102 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 skb->data_len -= len;
1104 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105}
1106
Andi Kleen67edfef2009-07-21 23:00:40 +00001107/* Remove acked data from a packet in the transmit queue. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1109{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001110 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 return -ENOMEM;
1112
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001113 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1114 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -07001116 else
1117 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
1119 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001120 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
1122 skb->truesize -= len;
1123 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001124 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1126
1127 /* Any change of skb->len requires recalculation of tso
1128 * factor and mss.
1129 */
1130 if (tcp_skb_pcount(skb) > 1)
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001131 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132
1133 return 0;
1134}
1135
Andi Kleen67edfef2009-07-21 23:00:40 +00001136/* Calculate MSS. Not accounting for SACKs here. */
John Heffner5d424d52006-03-20 17:53:41 -08001137int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1138{
1139 struct tcp_sock *tp = tcp_sk(sk);
1140 struct inet_connection_sock *icsk = inet_csk(sk);
1141 int mss_now;
1142
1143 /* Calculate base mss without TCP options:
1144 It is MMS_S - sizeof(tcphdr) of rfc1122
1145 */
1146 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1147
1148 /* Clamp it (mss_clamp does not include tcp options) */
1149 if (mss_now > tp->rx_opt.mss_clamp)
1150 mss_now = tp->rx_opt.mss_clamp;
1151
1152 /* Now subtract optional transport overhead */
1153 mss_now -= icsk->icsk_ext_hdr_len;
1154
1155 /* Then reserve room for full set of TCP options and 8 bytes of data */
1156 if (mss_now < 48)
1157 mss_now = 48;
1158
1159 /* Now subtract TCP options size, not including SACKs */
1160 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1161
1162 return mss_now;
1163}
1164
1165/* Inverse of above */
1166int tcp_mss_to_mtu(struct sock *sk, int mss)
1167{
1168 struct tcp_sock *tp = tcp_sk(sk);
1169 struct inet_connection_sock *icsk = inet_csk(sk);
1170 int mtu;
1171
1172 mtu = mss +
1173 tp->tcp_header_len +
1174 icsk->icsk_ext_hdr_len +
1175 icsk->icsk_af_ops->net_header_len;
1176
1177 return mtu;
1178}
1179
Andi Kleen67edfef2009-07-21 23:00:40 +00001180/* MTU probing init per socket */
John Heffner5d424d52006-03-20 17:53:41 -08001181void tcp_mtup_init(struct sock *sk)
1182{
1183 struct tcp_sock *tp = tcp_sk(sk);
1184 struct inet_connection_sock *icsk = inet_csk(sk);
1185
1186 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1187 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001188 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -08001189 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1190 icsk->icsk_mtup.probe_size = 0;
1191}
1192
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193/* This function synchronize snd mss to current pmtu/exthdr set.
1194
1195 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1196 for TCP options, but includes only bare TCP header.
1197
1198 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001199 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 It also does not include TCP options.
1201
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001202 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203
1204 tp->mss_cache is current effective sending mss, including
1205 all tcp options except for SACKs. It is evaluated,
1206 taking into account current pmtu, but never exceeds
1207 tp->rx_opt.mss_clamp.
1208
1209 NOTE1. rfc1122 clearly states that advertised MSS
1210 DOES NOT include either tcp or ip options.
1211
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001212 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1213 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1216{
1217 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001218 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001219 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220
John Heffner5d424d52006-03-20 17:53:41 -08001221 if (icsk->icsk_mtup.search_high > pmtu)
1222 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223
John Heffner5d424d52006-03-20 17:53:41 -08001224 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001225 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226
1227 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001228 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001229 if (icsk->icsk_mtup.enabled)
1230 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001231 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
1233 return mss_now;
1234}
1235
1236/* Compute the current effective MSS, taking SACKs and IP options,
1237 * and even PMTU discovery events into account.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001239unsigned int tcp_current_mss(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240{
1241 struct tcp_sock *tp = tcp_sk(sk);
1242 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001243 u32 mss_now;
Adam Langley33ad7982008-07-19 00:04:31 -07001244 unsigned header_len;
1245 struct tcp_out_options opts;
1246 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001248 mss_now = tp->mss_cache;
1249
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 if (dst) {
1251 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001252 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 mss_now = tcp_sync_mss(sk, mtu);
1254 }
1255
Adam Langley33ad7982008-07-19 00:04:31 -07001256 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1257 sizeof(struct tcphdr);
1258 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1259 * some common options. If this is an odd packet (because we have SACK
1260 * blocks etc) then our calculated header_len will be different, and
1261 * we have to adjust mss_now correspondingly */
1262 if (header_len != tp->tcp_header_len) {
1263 int delta = (int) header_len - tp->tcp_header_len;
1264 mss_now -= delta;
1265 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001266
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 return mss_now;
1268}
1269
David S. Millera762a982005-07-05 15:18:51 -07001270/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001271static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001272{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001273 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001274
Ilpo Järvinend436d682007-12-31 14:58:00 -08001275 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001276 /* Network is feed fully. */
1277 tp->snd_cwnd_used = 0;
1278 tp->snd_cwnd_stamp = tcp_time_stamp;
1279 } else {
1280 /* Network starves. */
1281 if (tp->packets_out > tp->snd_cwnd_used)
1282 tp->snd_cwnd_used = tp->packets_out;
1283
David S. Miller15d33c02007-04-09 13:23:14 -07001284 if (sysctl_tcp_slow_start_after_idle &&
1285 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001286 tcp_cwnd_application_limited(sk);
1287 }
1288}
1289
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001290/* Returns the portion of skb which can be sent right away without
1291 * introducing MSS oddities to segment boundaries. In rare cases where
1292 * mss_now != mss_cache, we will request caller to create a small skb
1293 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001294 *
1295 * We explicitly want to create a request for splitting write queue tail
1296 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1297 * thus all the complexity (cwnd_len is always MSS multiple which we
1298 * return whenever allowed by the other factors). Basically we need the
1299 * modulo only when the receiver window alone is the limiting factor or
1300 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001301 */
1302static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001303 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001304{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001305 struct tcp_sock *tp = tcp_sk(sk);
1306 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001307
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001308 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001309 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001310
1311 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1312 return cwnd_len;
1313
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001314 needed = min(skb->len, window);
1315
Ilpo Järvinen17515402008-04-15 20:36:55 -07001316 if (cwnd_len <= needed)
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001317 return cwnd_len;
1318
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001319 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001320}
1321
1322/* Can at least one segment of SKB be sent right now, according to the
1323 * congestion window rules? If so, return how many segments are allowed.
1324 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001325static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1326 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001327{
1328 u32 in_flight, cwnd;
1329
1330 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001331 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1332 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001333 return 1;
1334
1335 in_flight = tcp_packets_in_flight(tp);
1336 cwnd = tp->snd_cwnd;
1337 if (in_flight < cwnd)
1338 return (cwnd - in_flight);
1339
1340 return 0;
1341}
1342
Andi Kleen67edfef2009-07-21 23:00:40 +00001343/* Intialize TSO state of a skb.
1344 * This must be invoked the first time we consider transmitting
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001345 * SKB onto the wire.
1346 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001347static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1348 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001349{
1350 int tso_segs = tcp_skb_pcount(skb);
1351
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001352 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
David S. Miller846998a2005-08-04 19:52:01 -07001353 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001354 tso_segs = tcp_skb_pcount(skb);
1355 }
1356 return tso_segs;
1357}
1358
Andi Kleen67edfef2009-07-21 23:00:40 +00001359/* Minshall's variant of the Nagle send check. */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001360static inline int tcp_minshall_check(const struct tcp_sock *tp)
1361{
Jianjun Kong09cb1052008-11-03 00:27:11 -08001362 return after(tp->snd_sml, tp->snd_una) &&
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001363 !after(tp->snd_sml, tp->snd_nxt);
1364}
1365
1366/* Return 0, if packet can be sent now without violation Nagle's rules:
1367 * 1. It is full sized.
1368 * 2. Or it contains FIN. (already checked by caller)
1369 * 3. Or TCP_NODELAY was set.
1370 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1371 * With Minshall's modification: all sent small packets are ACKed.
1372 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001373static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001374 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001375 unsigned mss_now, int nonagle)
1376{
1377 return (skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001378 ((nonagle & TCP_NAGLE_CORK) ||
1379 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001380}
1381
1382/* Return non-zero if the Nagle test allows this packet to be
1383 * sent now.
1384 */
1385static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1386 unsigned int cur_mss, int nonagle)
1387{
1388 /* Nagle rule does not apply to frames, which sit in the middle of the
1389 * write_queue (they have no chances to get new data).
1390 *
1391 * This is implemented in the callers, where they modify the 'nonagle'
1392 * argument based upon the location of SKB in the send queue.
1393 */
1394 if (nonagle & TCP_NAGLE_PUSH)
1395 return 1;
1396
Ilpo Järvinend551e452007-04-30 00:42:20 -07001397 /* Don't use the nagle rule for urgent data (or for the final FIN).
1398 * Nagle can be ignored during F-RTO too (see RFC4138).
1399 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001400 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001401 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1402 return 1;
1403
1404 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1405 return 1;
1406
1407 return 0;
1408}
1409
1410/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001411static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1412 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001413{
1414 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1415
1416 if (skb->len > cur_mss)
1417 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1418
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001419 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001420}
1421
David S. Millerfe067e82007-03-07 12:12:44 -08001422/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001423 * should be put on the wire right now. If so, it returns the number of
1424 * packets allowed by the congestion window.
1425 */
1426static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1427 unsigned int cur_mss, int nonagle)
1428{
1429 struct tcp_sock *tp = tcp_sk(sk);
1430 unsigned int cwnd_quota;
1431
David S. Miller846998a2005-08-04 19:52:01 -07001432 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001433
1434 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1435 return 0;
1436
1437 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001438 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001439 cwnd_quota = 0;
1440
1441 return cwnd_quota;
1442}
1443
Andi Kleen67edfef2009-07-21 23:00:40 +00001444/* Test if sending is allowed right now. */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001445int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001446{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001447 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001448 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001449
1450 return (skb &&
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001451 tcp_snd_test(sk, skb, tcp_current_mss(sk),
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001452 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001453 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001454}
1455
1456/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1457 * which is put after SKB on the list. It is very much like
1458 * tcp_fragment() except that it may make several kinds of assumptions
1459 * in order to speed up the splitting operation. In particular, we
1460 * know that all the data is in scatter-gather pages, and that the
1461 * packet has never been sent out before (and thus is not cloned).
1462 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001463static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1464 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001465{
1466 struct sk_buff *buff;
1467 int nlen = skb->len - len;
Ilpo Järvinen9ce01462009-02-28 04:44:42 +00001468 u8 flags;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001469
1470 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001471 if (skb->len != skb->data_len)
1472 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001473
Pavel Emelyanovdf97c702007-11-29 21:22:33 +11001474 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001475 if (unlikely(buff == NULL))
1476 return -ENOMEM;
1477
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001478 sk->sk_wmem_queued += buff->truesize;
1479 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001480 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001481 skb->truesize -= nlen;
1482
1483 /* Correct the sequence numbers. */
1484 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1485 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1486 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1487
1488 /* PSH and FIN should only be set in the second packet. */
1489 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001490 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001491 TCP_SKB_CB(buff)->flags = flags;
1492
1493 /* This packet was never sent out yet, so no SACK bits. */
1494 TCP_SKB_CB(buff)->sacked = 0;
1495
Patrick McHardy84fa7932006-08-29 16:44:56 -07001496 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001497 skb_split(skb, buff, len);
1498
1499 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001500 tcp_set_skb_tso_segs(sk, skb, mss_now);
1501 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001502
1503 /* Link BUFF into the send queue. */
1504 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001505 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001506
1507 return 0;
1508}
1509
1510/* Try to defer sending, if possible, in order to minimize the amount
1511 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1512 *
1513 * This algorithm is from John Heffner.
1514 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001515static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001516{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001517 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001518 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001519 u32 send_win, cong_win, limit, in_flight;
1520
1521 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001522 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001523
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001524 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001525 goto send_now;
1526
1527 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001528 if (tp->tso_deferred &&
Ilpo Järvinena2acde02008-12-05 22:49:37 -08001529 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001530 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001531
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001532 in_flight = tcp_packets_in_flight(tp);
1533
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001534 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001535
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001536 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001537
1538 /* From in_flight test above, we know that cwnd > in_flight. */
1539 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1540
1541 limit = min(send_win, cong_win);
1542
David S. Millerba244fe2006-03-11 18:51:49 -08001543 /* If a full-sized TSO skb can be sent, do it. */
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001544 if (limit >= sk->sk_gso_max_size)
John Heffnerae8064a2006-10-18 20:36:48 -07001545 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001546
Ilpo Järvinen62ad2762009-02-28 04:44:29 +00001547 /* Middle in queue won't get any more data, full sendable already? */
1548 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1549 goto send_now;
1550
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001551 if (sysctl_tcp_tso_win_divisor) {
1552 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1553
1554 /* If at least some fraction of a window is available,
1555 * just use it.
1556 */
1557 chunk /= sysctl_tcp_tso_win_divisor;
1558 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001559 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001560 } else {
1561 /* Different approach, try not to defer past a single
1562 * ACK. Receiver should ACK every other full sized
1563 * frame, so if we have space for more than 3 frames
1564 * then send now.
1565 */
1566 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001567 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001568 }
1569
1570 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001571 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001572
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001573 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001574
1575send_now:
1576 tp->tso_deferred = 0;
1577 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001578}
1579
John Heffner5d424d52006-03-20 17:53:41 -08001580/* Create a new MTU probe if we are ready.
Andi Kleen67edfef2009-07-21 23:00:40 +00001581 * MTU probe is regularly attempting to increase the path MTU by
1582 * deliberately sending larger packets. This discovers routing
1583 * changes resulting in larger path MTUs.
1584 *
John Heffner5d424d52006-03-20 17:53:41 -08001585 * Returns 0 if we should wait to probe (no cwnd available),
1586 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001587 * -1 otherwise
1588 */
John Heffner5d424d52006-03-20 17:53:41 -08001589static int tcp_mtu_probe(struct sock *sk)
1590{
1591 struct tcp_sock *tp = tcp_sk(sk);
1592 struct inet_connection_sock *icsk = inet_csk(sk);
1593 struct sk_buff *skb, *nskb, *next;
1594 int len;
1595 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001596 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001597 int copy;
1598 int mss_now;
1599
1600 /* Not currently probing/verifying,
1601 * not in recovery,
1602 * have enough cwnd, and
1603 * not SACKing (the variable headers throw things off) */
1604 if (!icsk->icsk_mtup.enabled ||
1605 icsk->icsk_mtup.probe_size ||
1606 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1607 tp->snd_cwnd < 11 ||
Ilpo Järvinencabeccb2009-02-28 04:44:38 +00001608 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
John Heffner5d424d52006-03-20 17:53:41 -08001609 return -1;
1610
1611 /* Very simple search strategy: just double the MSS. */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00001612 mss_now = tcp_current_mss(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001613 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001614 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001615 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1616 /* TODO: set timer for probe_converge_event */
1617 return -1;
1618 }
1619
1620 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001621 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001622 return -1;
1623
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001624 if (tp->snd_wnd < size_needed)
1625 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001626 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001627 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001628
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001629 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1630 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1631 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001632 return -1;
1633 else
1634 return 0;
1635 }
1636
1637 /* We're allowed to probe. Build it now. */
1638 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1639 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001640 sk->sk_wmem_queued += nskb->truesize;
1641 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001642
David S. Millerfe067e82007-03-07 12:12:44 -08001643 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001644
1645 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1646 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1647 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1648 TCP_SKB_CB(nskb)->sacked = 0;
1649 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001650 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001651
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001652 tcp_insert_write_queue_before(nskb, skb, sk);
1653
John Heffner5d424d52006-03-20 17:53:41 -08001654 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001655 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001656 copy = min_t(int, skb->len, probe_size - len);
1657 if (nskb->ip_summed)
1658 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1659 else
1660 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001661 skb_put(nskb, copy),
1662 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001663
1664 if (skb->len <= copy) {
1665 /* We've eaten all the data from this skb.
1666 * Throw it away. */
1667 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001668 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001669 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001670 } else {
1671 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001672 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001673 if (!skb_shinfo(skb)->nr_frags) {
1674 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001675 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001676 skb->csum = csum_partial(skb->data,
1677 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001678 } else {
1679 __pskb_trim_head(skb, copy);
1680 tcp_set_skb_tso_segs(sk, skb, mss_now);
1681 }
1682 TCP_SKB_CB(skb)->seq += copy;
1683 }
1684
1685 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001686
1687 if (len >= probe_size)
1688 break;
John Heffner5d424d52006-03-20 17:53:41 -08001689 }
1690 tcp_init_tso_segs(sk, nskb, nskb->len);
1691
1692 /* We're ready to send. If this fails, the probe will
1693 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1694 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1695 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1696 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001697 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001698 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001699 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001700
1701 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001702 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1703 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001704
1705 return 1;
1706 }
1707
1708 return -1;
1709}
1710
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711/* This routine writes packets to the network. It advances the
1712 * send_head. This happens as incoming acks open up the remote
1713 * window for us.
1714 *
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001715 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1716 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1717 * account rare use of URG, this is not a big flaw.
1718 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 * Returns 1, if no segments are in flight and we have queued segments, but
1720 * cannot send anything now because of SWS or another problem.
1721 */
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001722static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1723 int push_one, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724{
1725 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001726 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001727 unsigned int tso_segs, sent_pkts;
1728 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001729 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730
David S. Miller92df7b52005-07-05 15:19:06 -07001731 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001732
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001733 if (!push_one) {
1734 /* Do MTU probing. */
1735 result = tcp_mtu_probe(sk);
1736 if (!result) {
1737 return 0;
1738 } else if (result > 0) {
1739 sent_pkts = 1;
1740 }
John Heffner5d424d52006-03-20 17:53:41 -08001741 }
1742
David S. Millerfe067e82007-03-07 12:12:44 -08001743 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001744 unsigned int limit;
1745
Herbert Xub68e9f82005-08-04 19:52:02 -07001746 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001747 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001748
Herbert Xub68e9f82005-08-04 19:52:02 -07001749 cwnd_quota = tcp_cwnd_test(tp, skb);
1750 if (!cwnd_quota)
1751 break;
1752
1753 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1754 break;
1755
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001756 if (tso_segs == 1) {
1757 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1758 (tcp_skb_is_last(sk, skb) ?
1759 nonagle : TCP_NAGLE_PUSH))))
1760 break;
1761 } else {
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001762 if (!push_one && tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001763 break;
1764 }
David S. Milleraa934662005-07-05 15:20:09 -07001765
Herbert Xuc8ac3772005-08-16 20:43:40 -07001766 limit = mss_now;
Ilpo Järvinenf8269a42008-12-03 21:24:48 -08001767 if (tso_segs > 1 && !tcp_urg_mode(tp))
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001768 limit = tcp_mss_split_point(sk, skb, mss_now,
1769 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770
Herbert Xuc8ac3772005-08-16 20:43:40 -07001771 if (skb->len > limit &&
1772 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1773 break;
1774
David S. Miller92df7b52005-07-05 15:19:06 -07001775 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001776
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001777 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
David S. Miller92df7b52005-07-05 15:19:06 -07001778 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779
David S. Miller92df7b52005-07-05 15:19:06 -07001780 /* Advance the send_head. This one is sent out.
1781 * This call will increment packets_out.
1782 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001783 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001784
1785 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001786 sent_pkts++;
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001787
1788 if (push_one)
1789 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 }
David S. Miller92df7b52005-07-05 15:19:06 -07001791
David S. Milleraa934662005-07-05 15:20:09 -07001792 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001793 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001794 return 0;
1795 }
David S. Millerfe067e82007-03-07 12:12:44 -08001796 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797}
1798
David S. Millera762a982005-07-05 15:18:51 -07001799/* Push out any pending frames which were held back due to
1800 * TCP_CORK or attempt at coalescing tiny packets.
1801 * The socket must be locked by the caller.
1802 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001803void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1804 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001805{
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001806 /* If we are closed, the bytes will have to remain here.
1807 * In time closedown will finish, we empty the write queue and
1808 * all will be happy.
1809 */
1810 if (unlikely(sk->sk_state == TCP_CLOSE))
1811 return;
1812
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001813 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
Ilpo Järvinen726e07a2008-12-05 22:43:56 -08001814 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001815}
1816
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001817/* Send _single_ skb sitting at the send head. This function requires
1818 * true push pending frames to setup probe timer etc.
1819 */
1820void tcp_push_one(struct sock *sk, unsigned int mss_now)
1821{
David S. Millerfe067e82007-03-07 12:12:44 -08001822 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001823
1824 BUG_ON(!skb || skb->len < mss_now);
1825
Ilpo Järvinend5dd9172008-12-05 22:48:55 -08001826 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001827}
1828
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829/* This function returns the amount that we can raise the
1830 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001831 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 * 1. The window can never be shrunk once it is offered (RFC 793)
1833 * 2. We limit memory per socket
1834 *
1835 * RFC 1122:
1836 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1837 * RECV.NEXT + RCV.WIN fixed until:
1838 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1839 *
1840 * i.e. don't raise the right edge of the window until you can raise
1841 * it at least MSS bytes.
1842 *
1843 * Unfortunately, the recommended algorithm breaks header prediction,
1844 * since header prediction assumes th->window stays fixed.
1845 *
1846 * Strictly speaking, keeping th->window fixed violates the receiver
1847 * side SWS prevention criteria. The problem is that under this rule
1848 * a stream of single byte packets will cause the right side of the
1849 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001850 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 * Of course, if the sender implements sender side SWS prevention
1852 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001853 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001855 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 * If the free space is less than the 1/4 of the maximum
1857 * space available and the free space is less than 1/2 mss,
1858 * then set the window to 0.
1859 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1860 * Otherwise, just prevent the window from shrinking
1861 * and from being larger than the largest representable value.
1862 *
1863 * This prevents incremental opening of the window in the regime
1864 * where TCP is limited by the speed of the reader side taking
1865 * data out of the TCP receive queue. It does nothing about
1866 * those cases where the window is constrained on the sender side
1867 * because the pipeline is full.
1868 *
1869 * BSD also seems to "accidentally" limit itself to windows that are a
1870 * multiple of MSS, at least until the free space gets quite small.
1871 * This would appear to be a side effect of the mbuf implementation.
1872 * Combining these two algorithms results in the observed behavior
1873 * of having a fixed window size at almost all times.
1874 *
1875 * Below we obtain similar behavior by forcing the offered window to
1876 * a multiple of the mss when it is feasible to do so.
1877 *
1878 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1879 * Regular options like TIMESTAMP are taken into account.
1880 */
1881u32 __tcp_select_window(struct sock *sk)
1882{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001883 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001885 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 * here. I don't know if the value based on our guesses
1887 * of peer's MSS is better for the performance. It's more correct
1888 * but may be worse for the performance because of rcv_mss
1889 * fluctuations. --SAW 1998/11/1
1890 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001891 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 int free_space = tcp_space(sk);
1893 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1894 int window;
1895
1896 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001897 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001899 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001900 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901
1902 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001903 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1904 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905
1906 if (free_space < mss)
1907 return 0;
1908 }
1909
1910 if (free_space > tp->rcv_ssthresh)
1911 free_space = tp->rcv_ssthresh;
1912
1913 /* Don't do rounding if we are using window scaling, since the
1914 * scaled window will not line up with the MSS boundary anyway.
1915 */
1916 window = tp->rcv_wnd;
1917 if (tp->rx_opt.rcv_wscale) {
1918 window = free_space;
1919
1920 /* Advertise enough space so that it won't get scaled away.
1921 * Import case: prevent zero window announcement if
1922 * 1<<rcv_wscale > mss.
1923 */
1924 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1925 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1926 << tp->rx_opt.rcv_wscale);
1927 } else {
1928 /* Get the largest window that is a nice multiple of mss.
1929 * Window clamp already applied above.
1930 * If our current window offering is within 1 mss of the
1931 * free space we just keep it. This prevents the divide
1932 * and multiply from happening most of the time.
1933 * We also don't do any window rounding when the free space
1934 * is too small.
1935 */
1936 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001937 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001938 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001939 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001940 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941 }
1942
1943 return window;
1944}
1945
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08001946/* Collapses two adjacent SKB's during retransmission. */
1947static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948{
1949 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001950 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001951 int skb_size, next_skb_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001953 skb_size = skb->len;
1954 next_skb_size = next_skb->len;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001956 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001957
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001958 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001960 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001962 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1963 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001965 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1966 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001968 if (skb->ip_summed != CHECKSUM_PARTIAL)
1969 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001971 /* Update sequence range on original skb. */
1972 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973
Ilpo Järvinene6c7d082009-02-28 04:44:35 +00001974 /* Merge over control information. This moves PSH/FIN etc. over */
1975 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001977 /* All done, get rid of second SKB and account for it so
1978 * packet counting does not break.
1979 */
1980 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001981
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001982 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07001983 tcp_clear_retrans_hints_partial(tp);
1984 if (next_skb == tp->retransmit_skb_hint)
1985 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001986
Ilpo Järvinen797108d2009-04-01 23:15:17 +00001987 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
1988
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001989 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990}
1991
Andi Kleen67edfef2009-07-21 23:00:40 +00001992/* Check if coalescing SKBs is legal. */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08001993static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1994{
1995 if (tcp_skb_pcount(skb) > 1)
1996 return 0;
1997 /* TODO: SACK collapsing could be used to remove this condition */
1998 if (skb_shinfo(skb)->nr_frags != 0)
1999 return 0;
2000 if (skb_cloned(skb))
2001 return 0;
2002 if (skb == tcp_send_head(sk))
2003 return 0;
2004 /* Some heurestics for collapsing over SACK'd could be invented */
2005 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2006 return 0;
2007
2008 return 1;
2009}
2010
Andi Kleen67edfef2009-07-21 23:00:40 +00002011/* Collapse packets in the retransmit queue to make to create
2012 * less packets on the wire. This is only done on retransmission.
2013 */
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002014static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2015 int space)
2016{
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 struct sk_buff *skb = to, *tmp;
2019 int first = 1;
2020
2021 if (!sysctl_tcp_retrans_collapse)
2022 return;
2023 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
2024 return;
2025
2026 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2027 if (!tcp_can_collapse(sk, skb))
2028 break;
2029
2030 space -= skb->len;
2031
2032 if (first) {
2033 first = 0;
2034 continue;
2035 }
2036
2037 if (space < 0)
2038 break;
2039 /* Punt if not enough space exists in the first SKB for
2040 * the data in the second
2041 */
2042 if (skb->len > skb_tailroom(to))
2043 break;
2044
2045 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2046 break;
2047
2048 tcp_collapse_retrans(sk, to);
2049 }
2050}
2051
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052/* This retransmits one SKB. Policy decisions and retransmit queue
2053 * state updates are done by the caller. Returns non-zero if an
2054 * error occurred which prevented the send.
2055 */
2056int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2057{
2058 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08002059 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002060 unsigned int cur_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 int err;
2062
John Heffner5d424d52006-03-20 17:53:41 -08002063 /* Inconslusive MTU probe */
2064 if (icsk->icsk_mtup.probe_size) {
2065 icsk->icsk_mtup.probe_size = 0;
2066 }
2067
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08002069 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 */
2071 if (atomic_read(&sk->sk_wmem_alloc) >
2072 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2073 return -EAGAIN;
2074
2075 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2076 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2077 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2079 return -ENOMEM;
2080 }
2081
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002082 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2083 return -EHOSTUNREACH; /* Routing failure or similar. */
2084
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002085 cur_mss = tcp_current_mss(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07002086
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 /* If receiver has shrunk his window, and skb is out of
2088 * new window, do not retransmit it. The exception is the
2089 * case, when window is shrunk to zero. In this case
2090 * our retransmit serves as a zero window probe.
2091 */
Joe Perches9d4fb272009-11-23 10:41:23 -08002092 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2093 TCP_SKB_CB(skb)->seq != tp->snd_una)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 return -EAGAIN;
2095
2096 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07002097 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 return -ENOMEM; /* We'll try again later. */
Ilpo Järvinen02276f32009-02-28 04:44:31 +00002099 } else {
Ilpo Järvinen9eb93622009-04-01 23:18:20 +00002100 int oldpcount = tcp_skb_pcount(skb);
2101
2102 if (unlikely(oldpcount > 1)) {
2103 tcp_init_tso_segs(sk, skb, cur_mss);
2104 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2105 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 }
2107
Ilpo Järvinen4a17fc32008-11-24 21:03:43 -08002108 tcp_retrans_try_collapse(sk, skb, cur_mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 /* Some Solaris stacks overoptimize and ignore the FIN on a
2111 * retransmit when old data is attached. So strip it off
2112 * since it is cheap to do so and saves bytes on the network.
2113 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08002114 if (skb->len > 0 &&
2115 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
2116 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002118 /* Reuse, even though it does some unnecessary work */
2119 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2120 TCP_SKB_CB(skb)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 }
2123 }
2124
2125 /* Make a copy, if the first transmission SKB clone we made
2126 * is still in somebody's hands, else make a clone.
2127 */
2128 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002130 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
2132 if (err == 0) {
2133 /* Update global TCP statistics. */
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002134 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135
2136 tp->total_retrans++;
2137
2138#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002139 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 if (net_ratelimit())
2141 printk(KERN_DEBUG "retrans_out leaked.\n");
2142 }
2143#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07002144 if (!tp->retrans_out)
2145 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2147 tp->retrans_out += tcp_skb_pcount(skb);
2148
2149 /* Save stamp of the first retransmit. */
2150 if (!tp->retrans_stamp)
2151 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2152
2153 tp->undo_retrans++;
2154
2155 /* snd_nxt is stored to detect loss of retransmitted segment,
2156 * see tcp_input.c tcp_sacktag_write_queue().
2157 */
2158 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2159 }
2160 return err;
2161}
2162
Andi Kleen67edfef2009-07-21 23:00:40 +00002163/* Check if we forward retransmits are possible in the current
2164 * window/congestion state.
2165 */
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002166static int tcp_can_forward_retransmit(struct sock *sk)
2167{
2168 const struct inet_connection_sock *icsk = inet_csk(sk);
2169 struct tcp_sock *tp = tcp_sk(sk);
2170
2171 /* Forward retransmissions are possible only during Recovery. */
2172 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2173 return 0;
2174
2175 /* No forward retransmissions in Reno are possible. */
2176 if (tcp_is_reno(tp))
2177 return 0;
2178
2179 /* Yeah, we have to make difficult choice between forward transmission
2180 * and retransmission... Both ways have their merits...
2181 *
2182 * For now we do not retransmit anything, while we have some new
2183 * segments to send. In the other cases, follow rule 3 for
2184 * NextSeg() specified in RFC3517.
2185 */
2186
2187 if (tcp_may_send_now(sk))
2188 return 0;
2189
2190 return 1;
2191}
2192
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193/* This gets called after a retransmit timeout, and the initially
2194 * retransmitted data is acknowledged. It tries to continue
2195 * resending the rest of the retransmit queue, until either
2196 * we've sent it all or the congestion window limit is reached.
2197 * If doing SACK, the first ACK which comes back for a timeout
2198 * based retransmit packet might feed us FACK information again.
2199 * If so, we use it to avoid unnecessarily retransmissions.
2200 */
2201void tcp_xmit_retransmit_queue(struct sock *sk)
2202{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002203 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 struct tcp_sock *tp = tcp_sk(sk);
2205 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002206 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002207 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002208 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002209 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002210
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002211 if (!tp->lost_out)
2212 tp->retransmit_high = tp->snd_una;
2213
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002214 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002215 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002216 last_lost = TCP_SKB_CB(skb)->end_seq;
2217 if (after(last_lost, tp->retransmit_high))
2218 last_lost = tp->retransmit_high;
2219 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002220 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002221 last_lost = tp->snd_una;
2222 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002224 tcp_for_write_queue_from(skb, sk) {
2225 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002227 if (skb == tcp_send_head(sk))
2228 break;
2229 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002230 if (hole == NULL)
2231 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002232
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002233 /* Assume this retransmit will generate
2234 * only one packet for congestion window
2235 * calculation purposes. This works because
2236 * tcp_retransmit_skb() will chop up the
2237 * packet to be MSS sized and all the
2238 * packet counting works out.
2239 */
2240 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2241 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002242
2243 if (fwd_rexmitting) {
2244begin_fwd:
2245 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2246 break;
2247 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2248
2249 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002250 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002251 if (!tcp_can_forward_retransmit(sk))
2252 break;
2253 /* Backtrack if necessary to non-L'ed skb */
2254 if (hole != NULL) {
2255 skb = hole;
2256 hole = NULL;
2257 }
2258 fwd_rexmitting = 1;
2259 goto begin_fwd;
2260
2261 } else if (!(sacked & TCPCB_LOST)) {
Ilpo Järvinenac11ba72009-02-28 04:44:27 +00002262 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002263 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002264 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002266 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002267 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002268 if (icsk->icsk_ca_state != TCP_CA_Loss)
2269 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2270 else
2271 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2272 }
2273
2274 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002275 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002276
Ilpo Järvinenf0ceb0e2008-09-20 21:24:49 -07002277 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002278 return;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002279 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2280
2281 if (skb == tcp_write_queue_head(sk))
2282 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2283 inet_csk(sk)->icsk_rto,
2284 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286}
2287
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288/* Send a fin. The caller locks the socket for us. This cannot be
2289 * allowed to fail queueing a FIN frame under any circumstances.
2290 */
2291void tcp_send_fin(struct sock *sk)
2292{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002293 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002294 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002296
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 /* Optimization, tack on the FIN if we have a queue of
2298 * unsent frames. But be careful about outgoing SACKS
2299 * and IP options.
2300 */
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002301 mss_now = tcp_current_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302
David S. Millerfe067e82007-03-07 12:12:44 -08002303 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2305 TCP_SKB_CB(skb)->end_seq++;
2306 tp->write_seq++;
2307 } else {
2308 /* Socket is locked, keep trying until memory is available. */
2309 for (;;) {
Wu Fengguangaa133072009-09-02 23:45:45 -07002310 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2311 sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 if (skb)
2313 break;
2314 yield();
2315 }
2316
2317 /* Reserve space for headers and prepare control bits. */
2318 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002320 tcp_init_nondata_skb(skb, tp->write_seq,
2321 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 tcp_queue_skb(sk, skb);
2323 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002324 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325}
2326
2327/* We get here when a process closes a file descriptor (either due to
2328 * an explicit close() or as a byproduct of exit()'ing) and there
2329 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002330 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 */
Al Virodd0fc662005-10-07 07:46:04 +01002332void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 struct sk_buff *skb;
2335
2336 /* NOTE: No TCP options attached and we never retransmit this. */
2337 skb = alloc_skb(MAX_TCP_HEADER, priority);
2338 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002339 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 return;
2341 }
2342
2343 /* Reserve space for headers and prepare control bits. */
2344 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002345 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2346 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002349 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002350 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002351
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002352 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353}
2354
Andi Kleen67edfef2009-07-21 23:00:40 +00002355/* Send a crossed SYN-ACK during socket establishment.
2356 * WARNING: This routine must only be called when we have already sent
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 * a SYN packet that crossed the incoming SYN that caused this routine
2358 * to get called. If this assumption fails then the initial rcv_wnd
2359 * and rcv_wscale values will not be correct.
2360 */
2361int tcp_send_synack(struct sock *sk)
2362{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002363 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
David S. Millerfe067e82007-03-07 12:12:44 -08002365 skb = tcp_write_queue_head(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002366 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2368 return -EFAULT;
2369 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002370 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 if (skb_cloned(skb)) {
2372 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2373 if (nskb == NULL)
2374 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002375 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002377 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002378 sk_wmem_free_skb(sk, skb);
2379 sk->sk_wmem_queued += nskb->truesize;
2380 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 skb = nskb;
2382 }
2383
2384 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2385 TCP_ECN_send_synack(tcp_sk(sk), skb);
2386 }
2387 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002388 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389}
2390
Andi Kleen67edfef2009-07-21 23:00:40 +00002391/* Prepare a SYN-ACK. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002392struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
William Allen Simpsone6b4d112009-12-02 18:07:39 +00002393 struct request_sock *req,
2394 struct request_values *rvp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395{
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002396 struct tcp_out_options opts;
William Allen Simpson4957faa2009-12-02 18:25:27 +00002397 struct tcp_extend_values *xvp = tcp_xv(rvp);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002398 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 struct tcp_sock *tp = tcp_sk(sk);
Eric Dumazet28b27742010-03-08 11:32:01 -08002400 const struct tcp_cookie_values *cvp = tp->cookie_values;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 struct tcphdr *th;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002403 struct tcp_md5sig_key *md5;
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002404 int tcp_header_size;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002405 int mss;
Eric Dumazet28b27742010-03-08 11:32:01 -08002406 int s_data_desired = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407
Eric Dumazet28b27742010-03-08 11:32:01 -08002408 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2409 s_data_desired = cvp->s_data_desired;
2410 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 if (skb == NULL)
2412 return NULL;
2413
2414 /* Reserve space for headers. */
2415 skb_reserve(skb, MAX_TCP_HEADER);
2416
Eric Dumazetadf30902009-06-02 05:19:30 +00002417 skb_dst_set(skb, dst_clone(dst));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002419 mss = dst_metric(dst, RTAX_ADVMSS);
2420 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2421 mss = tp->rx_opt.user_mss;
2422
Adam Langley33ad7982008-07-19 00:04:31 -07002423 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2424 __u8 rcv_wscale;
2425 /* Set this up on the first call only */
2426 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2427 /* tcp_full_space because it is guaranteed to be the first packet */
2428 tcp_select_initial_window(tcp_full_space(sk),
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002429 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Adam Langley33ad7982008-07-19 00:04:31 -07002430 &req->rcv_wnd,
2431 &req->window_clamp,
2432 ireq->wscale_ok,
laurent chavey31d12922009-12-15 11:15:28 +00002433 &rcv_wscale,
2434 dst_metric(dst, RTAX_INITRWND));
Adam Langley33ad7982008-07-19 00:04:31 -07002435 ireq->rcv_wscale = rcv_wscale;
2436 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002437
Adam Langley33ad7982008-07-19 00:04:31 -07002438 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002439#ifdef CONFIG_SYN_COOKIES
2440 if (unlikely(req->cookie_ts))
2441 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2442 else
2443#endif
Adam Langley33ad7982008-07-19 00:04:31 -07002444 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002445 tcp_header_size = tcp_synack_options(sk, req, mss,
William Allen Simpson4957faa2009-12-02 18:25:27 +00002446 skb, &opts, &md5, xvp)
2447 + sizeof(*th);
Adam Langley33ad7982008-07-19 00:04:31 -07002448
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002449 skb_push(skb, tcp_header_size);
2450 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002452 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 memset(th, 0, sizeof(struct tcphdr));
2454 th->syn = 1;
2455 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 TCP_ECN_make_synack(req, th);
KOVACS Krisztiana3116ac2008-10-01 07:46:49 -07002457 th->source = ireq->loc_port;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002458 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002459 /* Setting of flags are superfluous here for callers (and ECE is
2460 * not even correctly set)
2461 */
2462 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2463 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
William Allen Simpson4957faa2009-12-02 18:25:27 +00002464
2465 if (OPTION_COOKIE_EXTENSION & opts.options) {
Eric Dumazet28b27742010-03-08 11:32:01 -08002466 if (s_data_desired) {
2467 u8 *buf = skb_put(skb, s_data_desired);
William Allen Simpson4957faa2009-12-02 18:25:27 +00002468
2469 /* copy data directly from the listening socket. */
Eric Dumazet28b27742010-03-08 11:32:01 -08002470 memcpy(buf, cvp->s_data_payload, s_data_desired);
2471 TCP_SKB_CB(skb)->end_seq += s_data_desired;
William Allen Simpson4957faa2009-12-02 18:25:27 +00002472 }
2473
2474 if (opts.hash_size > 0) {
2475 __u32 workspace[SHA_WORKSPACE_WORDS];
2476 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2477 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2478
2479 /* Secret recipe depends on the Timestamp, (future)
2480 * Sequence and Acknowledgment Numbers, Initiator
2481 * Cookie, and others handled by IP variant caller.
2482 */
2483 *tail-- ^= opts.tsval;
2484 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2485 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2486
2487 /* recommended */
Eric Dumazet0eae88f2010-04-20 19:06:52 -07002488 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
David S. Millerf9a2e692009-12-02 22:12:04 -08002489 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
William Allen Simpson4957faa2009-12-02 18:25:27 +00002490
2491 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2492 (char *)mess,
2493 &workspace[0]);
2494 opts.hash_location =
2495 (__u8 *)&xvp->cookie_bakery[0];
2496 }
2497 }
2498
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002500 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501
2502 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002503 th->window = htons(min(req->rcv_wnd, 65535U));
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002504 tcp_options_write((__be32 *)(th + 1), tp, &opts);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505 th->doff = (tcp_header_size >> 2);
Tom Herbertaa2ea052010-04-22 07:00:24 +00002506 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002507
2508#ifdef CONFIG_TCP_MD5SIG
2509 /* Okay, we have all we need - do the md5 hash if needed */
2510 if (md5) {
William Allen Simpsonbd0388a2009-12-02 18:23:05 +00002511 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002512 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002513 }
2514#endif
2515
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 return skb;
2517}
2518
Andi Kleen67edfef2009-07-21 23:00:40 +00002519/* Do all connect socket setups that can be done AF independent. */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002520static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
2522 struct dst_entry *dst = __sk_dst_get(sk);
2523 struct tcp_sock *tp = tcp_sk(sk);
2524 __u8 rcv_wscale;
2525
2526 /* We'll fix this up when we get a response from the other end.
2527 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2528 */
2529 tp->tcp_header_len = sizeof(struct tcphdr) +
David S. Millerbb5b7c12009-12-15 20:56:42 -08002530 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002532#ifdef CONFIG_TCP_MD5SIG
2533 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2534 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2535#endif
2536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 /* If user gave his TCP_MAXSEG, record it to clamp */
2538 if (tp->rx_opt.user_mss)
2539 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2540 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002541 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 tcp_sync_mss(sk, dst_mtu(dst));
2543
2544 if (!tp->window_clamp)
2545 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2546 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002547 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2548 tp->advmss = tp->rx_opt.user_mss;
2549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551
2552 tcp_select_initial_window(tcp_full_space(sk),
2553 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2554 &tp->rcv_wnd,
2555 &tp->window_clamp,
David S. Millerbb5b7c12009-12-15 20:56:42 -08002556 sysctl_tcp_window_scaling,
laurent chavey31d12922009-12-15 11:15:28 +00002557 &rcv_wscale,
2558 dst_metric(dst, RTAX_INITRWND));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559
2560 tp->rx_opt.rcv_wscale = rcv_wscale;
2561 tp->rcv_ssthresh = tp->rcv_wnd;
2562
2563 sk->sk_err = 0;
2564 sock_reset_flag(sk, SOCK_DONE);
2565 tp->snd_wnd = 0;
Hantzis Fotisee7537b2009-03-02 22:42:02 -08002566 tcp_init_wl(tp, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 tp->snd_una = tp->write_seq;
2568 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002569 tp->snd_up = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 tp->rcv_nxt = 0;
2571 tp->rcv_wup = 0;
2572 tp->copied_seq = 0;
2573
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002574 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2575 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 tcp_clear_retrans(tp);
2577}
2578
Andi Kleen67edfef2009-07-21 23:00:40 +00002579/* Build a SYN and send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580int tcp_connect(struct sock *sk)
2581{
2582 struct tcp_sock *tp = tcp_sk(sk);
2583 struct sk_buff *buff;
2584
2585 tcp_connect_init(sk);
2586
David S. Millerd179cd12005-08-17 14:57:30 -07002587 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 if (unlikely(buff == NULL))
2589 return -ENOBUFS;
2590
2591 /* Reserve space for headers. */
2592 skb_reserve(buff, MAX_TCP_HEADER);
2593
Wei Yongjunbd37a082006-08-07 21:04:15 -07002594 tp->snd_nxt = tp->write_seq;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002595 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2596 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597
2598 /* Send it off. */
2599 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2600 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2601 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002602 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002603 sk->sk_wmem_queued += buff->truesize;
2604 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 tp->packets_out += tcp_skb_pcount(buff);
Wu Fengguangaa133072009-09-02 23:45:45 -07002606 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002607
2608 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2609 * in order to make this packet get counted in tcpOutSegs.
2610 */
2611 tp->snd_nxt = tp->write_seq;
2612 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002613 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614
2615 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002616 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2617 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 return 0;
2619}
2620
2621/* Send out a delayed ack, the caller does the policy checking
2622 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2623 * for details.
2624 */
2625void tcp_send_delayed_ack(struct sock *sk)
2626{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002627 struct inet_connection_sock *icsk = inet_csk(sk);
2628 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 unsigned long timeout;
2630
2631 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002632 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002633 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002635 if (icsk->icsk_ack.pingpong ||
2636 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 max_ato = TCP_DELACK_MAX;
2638
2639 /* Slow path, intersegment interval is "high". */
2640
2641 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002642 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 * directly.
2644 */
2645 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002646 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647
2648 if (rtt < max_ato)
2649 max_ato = rtt;
2650 }
2651
2652 ato = min(ato, max_ato);
2653 }
2654
2655 /* Stay within the limit we were given */
2656 timeout = jiffies + ato;
2657
2658 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002659 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 /* If delack timer was blocked or is about to expire,
2661 * send ACK now.
2662 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002663 if (icsk->icsk_ack.blocked ||
2664 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 tcp_send_ack(sk);
2666 return;
2667 }
2668
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002669 if (!time_before(timeout, icsk->icsk_ack.timeout))
2670 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002672 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2673 icsk->icsk_ack.timeout = timeout;
2674 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675}
2676
2677/* This routine sends an ack and also updates the window. */
2678void tcp_send_ack(struct sock *sk)
2679{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002680 struct sk_buff *buff;
2681
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002683 if (sk->sk_state == TCP_CLOSE)
2684 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002686 /* We are not putting this on the write queue, so
2687 * tcp_transmit_skb() will set the ownership to this
2688 * sock.
2689 */
2690 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2691 if (buff == NULL) {
2692 inet_csk_schedule_ack(sk);
2693 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2694 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2695 TCP_DELACK_MAX, TCP_RTO_MAX);
2696 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002698
2699 /* Reserve space for headers and prepare control bits. */
2700 skb_reserve(buff, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002701 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002702
2703 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002704 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2705 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706}
2707
2708/* This routine sends a packet with an out of date sequence
2709 * number. It assumes the other end will try to ack it.
2710 *
2711 * Question: what should we make while urgent mode?
2712 * 4.4BSD forces sending single byte of data. We cannot send
2713 * out of window data, because we have SND.NXT==SND.MAX...
2714 *
2715 * Current solution: to send TWO zero-length segments in urgent mode:
2716 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2717 * out-of-date with SND.UNA-1 to probe window.
2718 */
2719static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2720{
2721 struct tcp_sock *tp = tcp_sk(sk);
2722 struct sk_buff *skb;
2723
2724 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2725 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002726 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727 return -1;
2728
2729 /* Reserve space for headers and set control bits. */
2730 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731 /* Use a previous sequence. This should cause the other
2732 * end to send an ack. Don't queue or clone SKB, just
2733 * send it.
2734 */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002735 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002737 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738}
2739
Andi Kleen67edfef2009-07-21 23:00:40 +00002740/* Initiate keepalive or window probe from timer. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741int tcp_write_wakeup(struct sock *sk)
2742{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002743 struct tcp_sock *tp = tcp_sk(sk);
2744 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002746 if (sk->sk_state == TCP_CLOSE)
2747 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002749 if ((skb = tcp_send_head(sk)) != NULL &&
2750 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2751 int err;
Ilpo Järvinen0c54b852009-03-14 14:23:05 +00002752 unsigned int mss = tcp_current_mss(sk);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002753 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002755 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2756 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002758 /* We are probing the opening of a window
2759 * but the window size is != 0
2760 * must have been a result SWS avoidance ( sender )
2761 */
2762 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2763 skb->len > mss) {
2764 seg_size = min(seg_size, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002766 if (tcp_fragment(sk, skb, seg_size, mss))
2767 return -1;
2768 } else if (!tcp_skb_pcount(skb))
2769 tcp_set_skb_tso_segs(sk, skb, mss);
2770
2771 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2772 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2773 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2774 if (!err)
2775 tcp_event_new_data_sent(sk, skb);
2776 return err;
2777 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002778 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002779 tcp_xmit_probe_skb(sk, 1);
2780 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782}
2783
2784/* A window probe timeout has occurred. If window is not closed send
2785 * a partial packet else a zero probe.
2786 */
2787void tcp_send_probe0(struct sock *sk)
2788{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002789 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 struct tcp_sock *tp = tcp_sk(sk);
2791 int err;
2792
2793 err = tcp_write_wakeup(sk);
2794
David S. Millerfe067e82007-03-07 12:12:44 -08002795 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002797 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002798 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 return;
2800 }
2801
2802 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002803 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2804 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002805 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002806 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002807 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2808 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 } else {
2810 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002811 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 * Let local senders to fight for local resources.
2813 *
2814 * Use accumulated backoff yet.
2815 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002816 if (!icsk->icsk_probes_out)
2817 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002818 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002819 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002820 TCP_RESOURCE_PROBE_INTERVAL),
2821 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 }
2823}
2824
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08002825EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826EXPORT_SYMBOL(tcp_connect);
2827EXPORT_SYMBOL(tcp_make_synack);
2828EXPORT_SYMBOL(tcp_simple_retransmit);
2829EXPORT_SYMBOL(tcp_sync_mss);
John Heffner5d424d52006-03-20 17:53:41 -08002830EXPORT_SYMBOL(tcp_mtup_init);