blob: 85b07eba1879e8ef8a88c70a394a1b8b22193462 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
Jesper Juhl02c30a82005-05-05 16:16:16 -07008 * Authors: Ross Biro
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#include <net/tcp.h>
38
39#include <linux/compiler.h>
40#include <linux/module.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42/* People can turn this off for buggy TCP's found in printers etc. */
Brian Haleyab32ea52006-09-22 14:15:41 -070043int sysctl_tcp_retrans_collapse __read_mostly = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -070044
Rick Jones15d99e02006-03-20 22:40:29 -080045/* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
47 */
Brian Haleyab32ea52006-09-22 14:15:41 -070048int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
Rick Jones15d99e02006-03-20 22:40:29 -080049
Linus Torvalds1da177e2005-04-16 15:20:36 -070050/* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
53 */
Brian Haleyab32ea52006-09-22 14:15:41 -070054int sysctl_tcp_tso_win_divisor __read_mostly = 3;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
Brian Haleyab32ea52006-09-22 14:15:41 -070056int sysctl_tcp_mtu_probing __read_mostly = 0;
57int sysctl_tcp_base_mss __read_mostly = 512;
John Heffner5d424d52006-03-20 17:53:41 -080058
David S. Miller35089bb2006-06-13 22:33:04 -070059/* By default, RFC2861 behavior. */
Brian Haleyab32ea52006-09-22 14:15:41 -070060int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
David S. Miller35089bb2006-06-13 22:33:04 -070061
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080062static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
Ilpo Järvinen6ff03ac2007-08-24 22:44:06 -070063{
64 struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080065 unsigned int prior_packets = tp->packets_out;
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070066
David S. Millerfe067e82007-03-07 12:12:44 -080067 tcp_advance_send_head(sk, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -070068 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen85124302007-11-26 20:17:38 +080069
70 /* Don't override Nagle indefinately with F-RTO */
71 if (tp->frto_counter == 2)
72 tp->frto_counter = 3;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -080073
74 tp->packets_out += tcp_skb_pcount(skb);
75 if (!prior_packets)
76 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
77 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -070078}
79
80/* SND.NXT, if window was not shrunk.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
85 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070086static inline __u32 tcp_acceptable_seq(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -070087{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -070088 struct tcp_sock *tp = tcp_sk(sk);
89
Ilpo Järvinen90840de2007-12-31 04:48:41 -080090 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 return tp->snd_nxt;
92 else
Ilpo Järvinen90840de2007-12-31 04:48:41 -080093 return tcp_wnd_end(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -070094}
95
96/* Calculate mss to advertise in SYN segment.
97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
98 *
99 * 1. It is independent of path mtu.
100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102 * attached devices, because some buggy hosts are confused by
103 * large MSS.
104 * 4. We do not make 3, we advertise MSS, calculated from first
105 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
106 * This may be overridden via information stored in routing table.
107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108 * probably even Jumbo".
109 */
110static __u16 tcp_advertise_mss(struct sock *sk)
111{
112 struct tcp_sock *tp = tcp_sk(sk);
113 struct dst_entry *dst = __sk_dst_get(sk);
114 int mss = tp->advmss;
115
116 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
117 mss = dst_metric(dst, RTAX_ADVMSS);
118 tp->advmss = mss;
119 }
120
121 return (__u16)mss;
122}
123
124/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125 * This is the first part of cwnd validation mechanism. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700126static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700128 struct tcp_sock *tp = tcp_sk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129 s32 delta = tcp_time_stamp - tp->lsndtime;
130 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
131 u32 cwnd = tp->snd_cwnd;
132
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300133 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -0300135 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 restart_cwnd = min(restart_cwnd, cwnd);
137
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700138 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 cwnd >>= 1;
140 tp->snd_cwnd = max(cwnd, restart_cwnd);
141 tp->snd_cwnd_stamp = tcp_time_stamp;
142 tp->snd_cwnd_used = 0;
143}
144
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800145static void tcp_event_data_sent(struct tcp_sock *tp,
146 struct sk_buff *skb, struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700148 struct inet_connection_sock *icsk = inet_csk(sk);
149 const u32 now = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150
David S. Miller35089bb2006-06-13 22:33:04 -0700151 if (sysctl_tcp_slow_start_after_idle &&
152 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700153 tcp_cwnd_restart(sk, __sk_dst_get(sk));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154
155 tp->lsndtime = now;
156
157 /* If it is a reply for ato after last received
158 * packet, enter pingpong mode.
159 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700160 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
161 icsk->icsk_ack.pingpong = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162}
163
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800164static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -0700166 tcp_dec_quickack_mode(sk, pkts);
167 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168}
169
170/* Determine a window scaling and initial window to offer.
171 * Based on the assumption that the given amount of space
172 * will be offered. Store the results in the tp structure.
173 * NOTE: for smooth operation initial space offering should
174 * be a multiple of mss if possible. We assume here that mss >= 1.
175 * This MUST be enforced by all callers.
176 */
177void tcp_select_initial_window(int __space, __u32 mss,
178 __u32 *rcv_wnd, __u32 *window_clamp,
179 int wscale_ok, __u8 *rcv_wscale)
180{
181 unsigned int space = (__space < 0 ? 0 : __space);
182
183 /* If no clamp set the clamp to the max possible scaled window */
184 if (*window_clamp == 0)
185 (*window_clamp) = (65535 << 14);
186 space = min(*window_clamp, space);
187
188 /* Quantize space offering to a multiple of mss if possible. */
189 if (space > mss)
190 space = (space / mss) * mss;
191
192 /* NOTE: offering an initial window larger than 32767
Rick Jones15d99e02006-03-20 22:40:29 -0800193 * will break some buggy TCP stacks. If the admin tells us
194 * it is likely we could be speaking with such a buggy stack
195 * we will truncate our initial window offering to 32K-1
196 * unless the remote has sent us a window scaling option,
197 * which we interpret as a sign the remote TCP is not
198 * misinterpreting the window field as a signed quantity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199 */
Rick Jones15d99e02006-03-20 22:40:29 -0800200 if (sysctl_tcp_workaround_signed_windows)
201 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
202 else
203 (*rcv_wnd) = space;
204
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205 (*rcv_wscale) = 0;
206 if (wscale_ok) {
207 /* Set window scaling on max possible window
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900208 * See RFC1323 for an explanation of the limit to 14
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 */
210 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
Stephen Hemminger316c1592006-08-22 00:06:11 -0700211 space = min_t(u32, space, *window_clamp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212 while (space > 65535 && (*rcv_wscale) < 14) {
213 space >>= 1;
214 (*rcv_wscale)++;
215 }
216 }
217
218 /* Set initial window to value enough for senders,
David S. Miller6b251852005-09-28 16:31:48 -0700219 * following RFC2414. Senders, not following this RFC,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 * will be satisfied with 2.
221 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800222 if (mss > (1 << *rcv_wscale)) {
David S. Miller01ff3672005-09-29 17:07:20 -0700223 int init_cwnd = 4;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800224 if (mss > 1460 * 3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 init_cwnd = 2;
David S. Miller01ff3672005-09-29 17:07:20 -0700226 else if (mss > 1460)
227 init_cwnd = 3;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800228 if (*rcv_wnd > init_cwnd * mss)
229 *rcv_wnd = init_cwnd * mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 }
231
232 /* Set the clamp no higher than max representable value */
233 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
234}
235
236/* Chose a new window to advertise, update state in tcp_sock for the
237 * socket, and return result with RFC1323 scaling applied. The return
238 * value can be stuffed directly into th->window for an outgoing
239 * frame.
240 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800241static u16 tcp_select_window(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242{
243 struct tcp_sock *tp = tcp_sk(sk);
244 u32 cur_win = tcp_receive_window(tp);
245 u32 new_win = __tcp_select_window(sk);
246
247 /* Never shrink the offered window */
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800248 if (new_win < cur_win) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 /* Danger Will Robinson!
250 * Don't update rcv_wup/rcv_wnd here or else
251 * we will not be able to advertise a zero
252 * window in time. --DaveM
253 *
254 * Relax Will Robinson.
255 */
Patrick McHardy607bfbf2008-03-20 16:11:27 -0700256 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 }
258 tp->rcv_wnd = new_win;
259 tp->rcv_wup = tp->rcv_nxt;
260
261 /* Make sure we do not exceed the maximum possible
262 * scaled window.
263 */
Rick Jones15d99e02006-03-20 22:40:29 -0800264 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265 new_win = min(new_win, MAX_TCP_WINDOW);
266 else
267 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
268
269 /* RFC1323 scaling applied */
270 new_win >>= tp->rx_opt.rcv_wscale;
271
272 /* If we advertise zero window, disable fast path. */
273 if (new_win == 0)
274 tp->pred_flags = 0;
275
276 return new_win;
277}
278
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800279static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700280{
281 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800282 if (!(tp->ecn_flags & TCP_ECN_OK))
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700283 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
284}
285
286static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
287{
288 struct tcp_sock *tp = tcp_sk(sk);
289
290 tp->ecn_flags = 0;
291 if (sysctl_tcp_ecn) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800292 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700293 tp->ecn_flags = TCP_ECN_OK;
294 }
295}
296
297static __inline__ void
298TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
299{
300 if (inet_rsk(req)->ecn_ok)
301 th->ece = 1;
302}
303
304static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
305 int tcp_header_len)
306{
307 struct tcp_sock *tp = tcp_sk(sk);
308
309 if (tp->ecn_flags & TCP_ECN_OK) {
310 /* Not-retransmitted data segment: set ECT and inject CWR. */
311 if (skb->len != tcp_header_len &&
312 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
313 INET_ECN_xmit(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800314 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
Ilpo Järvinenbdf1ee52007-05-27 02:04:16 -0700315 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
316 tcp_hdr(skb)->cwr = 1;
317 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
318 }
319 } else {
320 /* ACK or retransmitted segment: clear ECT|CE */
321 INET_ECN_dontxmit(sk);
322 }
323 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
324 tcp_hdr(skb)->ece = 1;
325 }
326}
327
Ilpo Järvinene870a8e2008-01-03 20:39:01 -0800328/* Constructs common control bits of non-data skb. If SYN/FIN is present,
329 * auto increment end seqno.
330 */
331static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
332{
333 skb->csum = 0;
334
335 TCP_SKB_CB(skb)->flags = flags;
336 TCP_SKB_CB(skb)->sacked = 0;
337
338 skb_shinfo(skb)->gso_segs = 1;
339 skb_shinfo(skb)->gso_size = 0;
340 skb_shinfo(skb)->gso_type = 0;
341
342 TCP_SKB_CB(skb)->seq = seq;
343 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
344 seq++;
345 TCP_SKB_CB(skb)->end_seq = seq;
346}
347
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700348static inline int tcp_urg_mode(const struct tcp_sock *tp)
349{
350 return tp->snd_una != tp->snd_up;
351}
352
Adam Langley33ad7982008-07-19 00:04:31 -0700353#define OPTION_SACK_ADVERTISE (1 << 0)
354#define OPTION_TS (1 << 1)
355#define OPTION_MD5 (1 << 2)
356
357struct tcp_out_options {
358 u8 options; /* bit field of OPTION_* */
359 u8 ws; /* window scale, 0 to disable */
360 u8 num_sack_blocks; /* number of SACK blocks to include */
361 u16 mss; /* 0 to disable */
362 __u32 tsval, tsecr; /* need to include OPTION_TS */
363};
364
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700365/* Beware: Something in the Internet is very sensitive to the ordering of
366 * TCP options, we learned this through the hard way, so be careful here.
367 * Luckily we can at least blame others for their non-compliance but from
368 * inter-operatibility perspective it seems that we're somewhat stuck with
369 * the ordering which we have been using if we want to keep working with
370 * those broken things (not that it currently hurts anybody as there isn't
371 * particular reason why the ordering would need to be changed).
372 *
373 * At least SACK_PERM as the first option is known to lead to a disaster
374 * (but it may well be that other scenarios fail similarly).
375 */
Adam Langley33ad7982008-07-19 00:04:31 -0700376static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
377 const struct tcp_out_options *opts,
378 __u8 **md5_hash) {
379 if (unlikely(OPTION_MD5 & opts->options)) {
YOSHIFUJI Hideaki496c98d2006-10-10 19:41:21 -0700380 *ptr++ = htonl((TCPOPT_NOP << 24) |
381 (TCPOPT_NOP << 16) |
Adam Langley33ad7982008-07-19 00:04:31 -0700382 (TCPOPT_MD5SIG << 8) |
383 TCPOLEN_MD5SIG);
384 *md5_hash = (__u8 *)ptr;
385 ptr += 4;
386 } else {
387 *md5_hash = NULL;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800388 }
Adam Langley33ad7982008-07-19 00:04:31 -0700389
Ilpo Järvinenfd6149d2008-10-23 14:06:35 -0700390 if (unlikely(opts->mss)) {
391 *ptr++ = htonl((TCPOPT_MSS << 24) |
392 (TCPOLEN_MSS << 16) |
393 opts->mss);
394 }
395
Adam Langley33ad7982008-07-19 00:04:31 -0700396 if (likely(OPTION_TS & opts->options)) {
397 if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
398 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
399 (TCPOLEN_SACK_PERM << 16) |
400 (TCPOPT_TIMESTAMP << 8) |
401 TCPOLEN_TIMESTAMP);
402 } else {
403 *ptr++ = htonl((TCPOPT_NOP << 24) |
404 (TCPOPT_NOP << 16) |
405 (TCPOPT_TIMESTAMP << 8) |
406 TCPOLEN_TIMESTAMP);
407 }
408 *ptr++ = htonl(opts->tsval);
409 *ptr++ = htonl(opts->tsecr);
410 }
411
Adam Langley33ad7982008-07-19 00:04:31 -0700412 if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
413 !(OPTION_TS & opts->options))) {
414 *ptr++ = htonl((TCPOPT_NOP << 24) |
415 (TCPOPT_NOP << 16) |
416 (TCPOPT_SACK_PERM << 8) |
417 TCPOLEN_SACK_PERM);
418 }
419
420 if (unlikely(opts->ws)) {
421 *ptr++ = htonl((TCPOPT_NOP << 24) |
422 (TCPOPT_WINDOW << 16) |
423 (TCPOLEN_WINDOW << 8) |
424 opts->ws);
425 }
426
427 if (unlikely(opts->num_sack_blocks)) {
428 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
429 tp->duplicate_sack : tp->selective_acks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800430 int this_sack;
431
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
433 (TCPOPT_NOP << 16) |
434 (TCPOPT_SACK << 8) |
Adam Langley33ad7982008-07-19 00:04:31 -0700435 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800436 TCPOLEN_SACK_PERBLOCK)));
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800437
Adam Langley33ad7982008-07-19 00:04:31 -0700438 for (this_sack = 0; this_sack < opts->num_sack_blocks;
439 ++this_sack) {
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800440 *ptr++ = htonl(sp[this_sack].start_seq);
441 *ptr++ = htonl(sp[this_sack].end_seq);
442 }
Stephen Hemminger2de979b2007-03-08 20:45:19 -0800443
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800444 if (tp->rx_opt.dsack) {
445 tp->rx_opt.dsack = 0;
Ilpo Järvinen75e3d8d2008-10-21 16:28:36 -0700446 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800447 }
448 }
449}
450
Adam Langley33ad7982008-07-19 00:04:31 -0700451static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
452 struct tcp_out_options *opts,
453 struct tcp_md5sig_key **md5) {
454 struct tcp_sock *tp = tcp_sk(sk);
455 unsigned size = 0;
456
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800457#ifdef CONFIG_TCP_MD5SIG
Adam Langley33ad7982008-07-19 00:04:31 -0700458 *md5 = tp->af_specific->md5_lookup(sk, sk);
459 if (*md5) {
460 opts->options |= OPTION_MD5;
461 size += TCPOLEN_MD5SIG_ALIGNED;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800462 }
Adam Langley33ad7982008-07-19 00:04:31 -0700463#else
464 *md5 = NULL;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800465#endif
Adam Langley33ad7982008-07-19 00:04:31 -0700466
467 /* We always get an MSS option. The option bytes which will be seen in
468 * normal data packets should timestamps be used, must be in the MSS
469 * advertised. But we subtract them from tp->mss_cache so that
470 * calculations in tcp_sendmsg are simpler etc. So account for this
471 * fact here if necessary. If we don't do this correctly, as a
472 * receiver we won't recognize data packets as being full sized when we
473 * should, and thus we won't abide by the delayed ACK rules correctly.
474 * SACKs don't matter, we never delay an ACK when we have any of those
475 * going out. */
476 opts->mss = tcp_advertise_mss(sk);
477 size += TCPOLEN_MSS_ALIGNED;
478
479 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
480 opts->options |= OPTION_TS;
481 opts->tsval = TCP_SKB_CB(skb)->when;
482 opts->tsecr = tp->rx_opt.ts_recent;
483 size += TCPOLEN_TSTAMP_ALIGNED;
484 }
485 if (likely(sysctl_tcp_window_scaling)) {
486 opts->ws = tp->rx_opt.rcv_wscale;
Philip Love7982d5e2008-08-27 02:33:50 -0700487 if(likely(opts->ws))
488 size += TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700489 }
490 if (likely(sysctl_tcp_sack)) {
491 opts->options |= OPTION_SACK_ADVERTISE;
David S. Millerb32d1312008-07-21 18:45:34 -0700492 if (unlikely(!(OPTION_TS & opts->options)))
Adam Langley33ad7982008-07-19 00:04:31 -0700493 size += TCPOLEN_SACKPERM_ALIGNED;
494 }
495
496 return size;
497}
498
499static unsigned tcp_synack_options(struct sock *sk,
500 struct request_sock *req,
501 unsigned mss, struct sk_buff *skb,
502 struct tcp_out_options *opts,
503 struct tcp_md5sig_key **md5) {
504 unsigned size = 0;
505 struct inet_request_sock *ireq = inet_rsk(req);
506 char doing_ts;
507
508#ifdef CONFIG_TCP_MD5SIG
509 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
510 if (*md5) {
511 opts->options |= OPTION_MD5;
512 size += TCPOLEN_MD5SIG_ALIGNED;
513 }
514#else
515 *md5 = NULL;
516#endif
517
518 /* we can't fit any SACK blocks in a packet with MD5 + TS
519 options. There was discussion about disabling SACK rather than TS in
520 order to fit in better with old, buggy kernels, but that was deemed
521 to be unnecessary. */
522 doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
523
524 opts->mss = mss;
525 size += TCPOLEN_MSS_ALIGNED;
526
527 if (likely(ireq->wscale_ok)) {
528 opts->ws = ireq->rcv_wscale;
Philip Love7982d5e2008-08-27 02:33:50 -0700529 if(likely(opts->ws))
530 size += TCPOLEN_WSCALE_ALIGNED;
Adam Langley33ad7982008-07-19 00:04:31 -0700531 }
532 if (likely(doing_ts)) {
533 opts->options |= OPTION_TS;
534 opts->tsval = TCP_SKB_CB(skb)->when;
535 opts->tsecr = req->ts_recent;
536 size += TCPOLEN_TSTAMP_ALIGNED;
537 }
538 if (likely(ireq->sack_ok)) {
539 opts->options |= OPTION_SACK_ADVERTISE;
540 if (unlikely(!doing_ts))
541 size += TCPOLEN_SACKPERM_ALIGNED;
542 }
543
544 return size;
545}
546
547static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
548 struct tcp_out_options *opts,
549 struct tcp_md5sig_key **md5) {
550 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
551 struct tcp_sock *tp = tcp_sk(sk);
552 unsigned size = 0;
553
554#ifdef CONFIG_TCP_MD5SIG
555 *md5 = tp->af_specific->md5_lookup(sk, sk);
556 if (unlikely(*md5)) {
557 opts->options |= OPTION_MD5;
558 size += TCPOLEN_MD5SIG_ALIGNED;
559 }
560#else
561 *md5 = NULL;
562#endif
563
564 if (likely(tp->rx_opt.tstamp_ok)) {
565 opts->options |= OPTION_TS;
566 opts->tsval = tcb ? tcb->when : 0;
567 opts->tsecr = tp->rx_opt.ts_recent;
568 size += TCPOLEN_TSTAMP_ALIGNED;
569 }
570
571 if (unlikely(tp->rx_opt.eff_sacks)) {
572 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
573 opts->num_sack_blocks =
574 min_t(unsigned, tp->rx_opt.eff_sacks,
575 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
576 TCPOLEN_SACK_PERBLOCK);
577 size += TCPOLEN_SACK_BASE_ALIGNED +
578 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
579 }
580
581 return size;
Stephen Hemminger40efc6f2006-01-03 16:03:49 -0800582}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583
584/* This routine actually transmits TCP packets queued in by
585 * tcp_do_sendmsg(). This is used by both the initial
586 * transmission and possible later retransmissions.
587 * All SKB's seen here are completely headerless. It is our
588 * job to build the TCP header, and pass the packet down to
589 * IP so it can do the same plus pass the packet off to the
590 * device.
591 *
592 * We are working here with either a clone of the original
593 * SKB, or a fresh unique copy made by the retransmit engine.
594 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800595static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
596 gfp_t gfp_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597{
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800598 const struct inet_connection_sock *icsk = inet_csk(sk);
599 struct inet_sock *inet;
600 struct tcp_sock *tp;
601 struct tcp_skb_cb *tcb;
Adam Langley33ad7982008-07-19 00:04:31 -0700602 struct tcp_out_options opts;
603 unsigned tcp_options_size, tcp_header_size;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800604 struct tcp_md5sig_key *md5;
605 __u8 *md5_hash_location;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800606 struct tcphdr *th;
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800607 int err;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800609 BUG_ON(!skb || !tcp_skb_pcount(skb));
610
611 /* If congestion control is doing timestamping, we must
612 * take such a timestamp before we potentially clone/copy.
613 */
Stephen Hemminger164891a2007-04-23 22:26:16 -0700614 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800615 __net_timestamp(skb);
616
617 if (likely(clone_it)) {
618 if (unlikely(skb_cloned(skb)))
619 skb = pskb_copy(skb, gfp_mask);
620 else
621 skb = skb_clone(skb, gfp_mask);
622 if (unlikely(!skb))
623 return -ENOBUFS;
624 }
625
626 inet = inet_sk(sk);
627 tp = tcp_sk(sk);
628 tcb = TCP_SKB_CB(skb);
Adam Langley33ad7982008-07-19 00:04:31 -0700629 memset(&opts, 0, sizeof(opts));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
Adam Langley33ad7982008-07-19 00:04:31 -0700631 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
632 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
633 else
634 tcp_options_size = tcp_established_options(sk, skb, &opts,
635 &md5);
636 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900637
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800638 if (tcp_packets_in_flight(tp) == 0)
639 tcp_ca_event(sk, CA_EVENT_TX_START);
640
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700641 skb_push(skb, tcp_header_size);
642 skb_reset_transport_header(skb);
David S. Millere89862f2007-01-26 01:04:55 -0800643 skb_set_owner_w(skb, sk);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800644
645 /* Build TCP header and checksum it. */
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -0700646 th = tcp_hdr(skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800647 th->source = inet->sport;
648 th->dest = inet->dport;
649 th->seq = htonl(tcb->seq);
650 th->ack_seq = htonl(tp->rcv_nxt);
Al Virodf7a3b02006-09-27 18:38:52 -0700651 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800652 tcb->flags);
653
654 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
655 /* RFC1323: The window in SYN & SYN/ACK segments
656 * is never scaled.
657 */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -0800658 th->window = htons(min(tp->rcv_wnd, 65535U));
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800659 } else {
660 th->window = htons(tcp_select_window(sk));
661 }
662 th->check = 0;
663 th->urg_ptr = 0;
664
Ilpo Järvinen33f5f572008-10-07 14:43:06 -0700665 /* The urg_mode check is necessary during a below snd_una win probe */
666 if (unlikely(tcp_urg_mode(tp) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800667 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
668 th->urg_ptr = htons(tp->snd_up - tcb->seq);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800669 th->urg = 1;
670 }
671
Adam Langley33ad7982008-07-19 00:04:31 -0700672 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
673 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -0700674 TCP_ECN_send(sk, skb, tcp_header_size);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800675
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800676#ifdef CONFIG_TCP_MD5SIG
677 /* Calculate the MD5 hash, as we have all we need now */
678 if (md5) {
Adam Langley33ad7982008-07-19 00:04:31 -0700679 sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800680 tp->af_specific->calc_md5_hash(md5_hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -0700681 md5, sk, NULL, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -0800682 }
683#endif
684
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -0800685 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800686
687 if (likely(tcb->flags & TCPCB_FLAG_ACK))
688 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
689
690 if (skb->len != tcp_header_size)
691 tcp_event_data_sent(tp, skb, sk);
692
Wei Yongjunbd37a082006-08-07 21:04:15 -0700693 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -0700694 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800695
David S. Millere89862f2007-01-26 01:04:55 -0800696 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
Hua Zhong83de47c2006-04-28 15:26:50 -0700697 if (likely(err <= 0))
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800698 return err;
699
Ilpo Järvinen3cfe3ba2007-02-27 10:09:49 -0800700 tcp_enter_cwr(sk, 1);
David S. Millerdfb4b9d2005-12-06 16:24:52 -0800701
Gerrit Renkerb9df3cb2006-11-14 11:21:36 -0200702 return net_xmit_eval(err);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703}
704
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900705/* This routine just queue's the buffer
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 *
707 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
708 * otherwise socket can stall.
709 */
710static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
711{
712 struct tcp_sock *tp = tcp_sk(sk);
713
714 /* Advance write_seq and place onto the write_queue. */
715 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
716 skb_header_release(skb);
David S. Millerfe067e82007-03-07 12:12:44 -0800717 tcp_add_write_queue_tail(sk, skb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800718 sk->sk_wmem_queued += skb->truesize;
719 sk_mem_charge(sk, skb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720}
721
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800722static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
723 unsigned int mss_now)
David S. Millerf6302d12005-07-05 15:18:03 -0700724{
Petr Tesarik33cf71c2008-11-21 16:42:58 -0800725 if (skb->len <= mss_now || !sk_can_gso(sk) ||
726 tcp_urg_mode(tcp_sk(sk))) {
David S. Millerf6302d12005-07-05 15:18:03 -0700727 /* Avoid the costly divide in the normal
728 * non-TSO case.
729 */
Herbert Xu79671682006-06-22 02:40:14 -0700730 skb_shinfo(skb)->gso_segs = 1;
731 skb_shinfo(skb)->gso_size = 0;
732 skb_shinfo(skb)->gso_type = 0;
David S. Millerf6302d12005-07-05 15:18:03 -0700733 } else {
Ilpo Järvinen356f89e2007-08-24 23:00:31 -0700734 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
Herbert Xu79671682006-06-22 02:40:14 -0700735 skb_shinfo(skb)->gso_size = mss_now;
Herbert Xubcd76112006-06-30 13:36:35 -0700736 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 }
738}
739
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700740/* When a modification to fackets out becomes necessary, we need to check
Ilpo Järvinen68f83532007-11-15 19:50:37 -0800741 * skb is counted to fackets_out or not.
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700742 */
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800743static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700744 int decr)
745{
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800746 struct tcp_sock *tp = tcp_sk(sk);
747
Ilpo Järvinendc869672007-10-01 15:27:19 -0700748 if (!tp->sacked_out || tcp_is_reno(tp))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700749 return;
750
Ilpo Järvinen6859d492007-12-02 00:48:06 +0200751 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700752 tp->fackets_out -= decr;
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700753}
754
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755/* Function to create two new TCP segments. Shrinks the given segment
756 * to the specified size and appends a new segment with the rest of the
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900757 * packet to the list. This won't be called frequently, I hope.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 * Remember, these are still headerless SKBs at this point.
759 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800760int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
761 unsigned int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762{
763 struct tcp_sock *tp = tcp_sk(sk);
764 struct sk_buff *buff;
David S. Miller6475be12005-09-01 22:47:01 -0700765 int nsize, old_factor;
Herbert Xub60b49e2006-04-19 21:35:00 -0700766 int nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767 u16 flags;
768
Herbert Xub2cc99f02005-10-20 17:13:13 -0200769 BUG_ON(len > skb->len);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -0800770
Ilpo Järvinenef9da472008-09-20 21:25:15 -0700771 tcp_clear_retrans_hints_partial(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 nsize = skb_headlen(skb) - len;
773 if (nsize < 0)
774 nsize = 0;
775
776 if (skb_cloned(skb) &&
777 skb_is_nonlinear(skb) &&
778 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
779 return -ENOMEM;
780
781 /* Get a new skb... force flag on. */
782 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
783 if (buff == NULL)
784 return -ENOMEM; /* We'll just try again later. */
Herbert Xuef5cb972006-04-18 13:24:14 -0700785
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800786 sk->sk_wmem_queued += buff->truesize;
787 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -0700788 nlen = skb->len - len - nsize;
789 buff->truesize += nlen;
790 skb->truesize -= nlen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791
792 /* Correct the sequence numbers. */
793 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
794 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
795 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
796
797 /* PSH and FIN should only be set in the second packet. */
798 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800799 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 TCP_SKB_CB(buff)->flags = flags;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700801 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Patrick McHardy84fa7932006-08-29 16:44:56 -0700803 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 /* Copy and checksum data tail into the new buffer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800805 buff->csum = csum_partial_copy_nocheck(skb->data + len,
806 skb_put(buff, nsize),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 nsize, 0);
808
809 skb_trim(skb, len);
810
811 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
812 } else {
Patrick McHardy84fa7932006-08-29 16:44:56 -0700813 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 skb_split(skb, buff, len);
815 }
816
817 buff->ip_summed = skb->ip_summed;
818
819 /* Looks stupid, but our code really uses when of
820 * skbs, which it never sent before. --ANK
821 */
822 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
Patrick McHardya61bbcf2005-08-14 17:24:31 -0700823 buff->tstamp = skb->tstamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700824
David S. Miller6475be12005-09-01 22:47:01 -0700825 old_factor = tcp_skb_pcount(skb);
826
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -0700828 tcp_set_skb_tso_segs(sk, skb, mss_now);
829 tcp_set_skb_tso_segs(sk, buff, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830
David S. Miller6475be12005-09-01 22:47:01 -0700831 /* If this packet has been sent out already, we must
832 * adjust the various packet counters.
833 */
Herbert Xucf0b4502005-09-08 15:10:52 -0700834 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
David S. Miller6475be12005-09-01 22:47:01 -0700835 int diff = old_factor - tcp_skb_pcount(skb) -
836 tcp_skb_pcount(buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
David S. Miller6475be12005-09-01 22:47:01 -0700838 tp->packets_out -= diff;
Herbert Xue14c3ca2005-09-19 18:18:38 -0700839
840 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
841 tp->sacked_out -= diff;
842 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
843 tp->retrans_out -= diff;
844
Ilpo Järvinenb5860bb2007-08-09 14:33:18 +0300845 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
David S. Miller6475be12005-09-01 22:47:01 -0700846 tp->lost_out -= diff;
Herbert Xu83ca28b2005-09-22 23:32:56 -0700847
Ilpo Järvinen91fed7a2007-10-09 01:24:15 -0700848 /* Adjust Reno SACK estimate. */
849 if (tcp_is_reno(tp) && diff > 0) {
850 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
851 tcp_verify_left_out(tp);
David S. Miller6475be12005-09-01 22:47:01 -0700852 }
Ilpo Järvinena47e5a92007-11-15 19:41:46 -0800853 tcp_adjust_fackets_out(sk, skb, diff);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 }
855
856 /* Link BUFF into the send queue. */
David S. Millerf44b5272005-07-05 15:18:34 -0700857 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -0800858 tcp_insert_write_queue_after(skb, buff, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859
860 return 0;
861}
862
863/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
864 * eventually). The difference is that pulled data not copied, but
865 * immediately discarded.
866 */
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700867static void __pskb_trim_head(struct sk_buff *skb, int len)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700868{
869 int i, k, eat;
870
871 eat = len;
872 k = 0;
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800873 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 if (skb_shinfo(skb)->frags[i].size <= eat) {
875 put_page(skb_shinfo(skb)->frags[i].page);
876 eat -= skb_shinfo(skb)->frags[i].size;
877 } else {
878 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
879 if (eat) {
880 skb_shinfo(skb)->frags[k].page_offset += eat;
881 skb_shinfo(skb)->frags[k].size -= eat;
882 eat = 0;
883 }
884 k++;
885 }
886 }
887 skb_shinfo(skb)->nr_frags = k;
888
Arnaldo Carvalho de Melo27a884d2007-04-19 20:29:13 -0700889 skb_reset_tail_pointer(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 skb->data_len -= len;
891 skb->len = skb->data_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892}
893
894int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
895{
Ilpo Järvinen056834d2007-12-31 14:57:14 -0800896 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 return -ENOMEM;
898
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700899 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
900 if (unlikely(len < skb_headlen(skb)))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 __skb_pull(skb, len);
Herbert Xu ~{PmVHI~}f2911962006-06-05 15:03:37 -0700902 else
903 __pskb_trim_head(skb, len - skb_headlen(skb));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700904
905 TCP_SKB_CB(skb)->seq += len;
Patrick McHardy84fa7932006-08-29 16:44:56 -0700906 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700907
908 skb->truesize -= len;
909 sk->sk_wmem_queued -= len;
Hideo Aoki3ab224b2007-12-31 00:11:19 -0800910 sk_mem_uncharge(sk, len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
912
913 /* Any change of skb->len requires recalculation of tso
914 * factor and mss.
915 */
916 if (tcp_skb_pcount(skb) > 1)
David S. Miller846998a2005-08-04 19:52:01 -0700917 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
919 return 0;
920}
921
John Heffner5d424d52006-03-20 17:53:41 -0800922/* Not accounting for SACKs here. */
923int tcp_mtu_to_mss(struct sock *sk, int pmtu)
924{
925 struct tcp_sock *tp = tcp_sk(sk);
926 struct inet_connection_sock *icsk = inet_csk(sk);
927 int mss_now;
928
929 /* Calculate base mss without TCP options:
930 It is MMS_S - sizeof(tcphdr) of rfc1122
931 */
932 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
933
934 /* Clamp it (mss_clamp does not include tcp options) */
935 if (mss_now > tp->rx_opt.mss_clamp)
936 mss_now = tp->rx_opt.mss_clamp;
937
938 /* Now subtract optional transport overhead */
939 mss_now -= icsk->icsk_ext_hdr_len;
940
941 /* Then reserve room for full set of TCP options and 8 bytes of data */
942 if (mss_now < 48)
943 mss_now = 48;
944
945 /* Now subtract TCP options size, not including SACKs */
946 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
947
948 return mss_now;
949}
950
951/* Inverse of above */
952int tcp_mss_to_mtu(struct sock *sk, int mss)
953{
954 struct tcp_sock *tp = tcp_sk(sk);
955 struct inet_connection_sock *icsk = inet_csk(sk);
956 int mtu;
957
958 mtu = mss +
959 tp->tcp_header_len +
960 icsk->icsk_ext_hdr_len +
961 icsk->icsk_af_ops->net_header_len;
962
963 return mtu;
964}
965
966void tcp_mtup_init(struct sock *sk)
967{
968 struct tcp_sock *tp = tcp_sk(sk);
969 struct inet_connection_sock *icsk = inet_csk(sk);
970
971 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
972 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +0900973 icsk->icsk_af_ops->net_header_len;
John Heffner5d424d52006-03-20 17:53:41 -0800974 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
975 icsk->icsk_mtup.probe_size = 0;
976}
977
Ilpo Järvinen409d22b2007-12-31 14:57:40 -0800978/* Bound MSS / TSO packet size with the half of the window */
979static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
980{
981 if (tp->max_window && pktsize > (tp->max_window >> 1))
982 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
983 else
984 return pktsize;
985}
986
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987/* This function synchronize snd mss to current pmtu/exthdr set.
988
989 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
990 for TCP options, but includes only bare TCP header.
991
992 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -0800993 It is minimum of user_mss and mss received with SYN.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 It also does not include TCP options.
995
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -0800996 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
998 tp->mss_cache is current effective sending mss, including
999 all tcp options except for SACKs. It is evaluated,
1000 taking into account current pmtu, but never exceeds
1001 tp->rx_opt.mss_clamp.
1002
1003 NOTE1. rfc1122 clearly states that advertised MSS
1004 DOES NOT include either tcp or ip options.
1005
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001006 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1007 are READ ONLY outside this function. --ANK (980731)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1010{
1011 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001012 struct inet_connection_sock *icsk = inet_csk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001013 int mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014
John Heffner5d424d52006-03-20 17:53:41 -08001015 if (icsk->icsk_mtup.search_high > pmtu)
1016 icsk->icsk_mtup.search_high = pmtu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017
John Heffner5d424d52006-03-20 17:53:41 -08001018 mss_now = tcp_mtu_to_mss(sk, pmtu);
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001019 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020
1021 /* And store cached results */
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001022 icsk->icsk_pmtu_cookie = pmtu;
John Heffner5d424d52006-03-20 17:53:41 -08001023 if (icsk->icsk_mtup.enabled)
1024 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001025 tp->mss_cache = mss_now;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
1027 return mss_now;
1028}
1029
1030/* Compute the current effective MSS, taking SACKs and IP options,
1031 * and even PMTU discovery events into account.
1032 *
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001033 * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 * cannot be large. However, taking into account rare use of URG, this
1035 * is not a big flaw.
1036 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001037unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038{
1039 struct tcp_sock *tp = tcp_sk(sk);
1040 struct dst_entry *dst = __sk_dst_get(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001041 u32 mss_now;
1042 u16 xmit_size_goal;
1043 int doing_tso = 0;
Adam Langley33ad7982008-07-19 00:04:31 -07001044 unsigned header_len;
1045 struct tcp_out_options opts;
1046 struct tcp_md5sig_key *md5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001048 mss_now = tp->mss_cache;
1049
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001050 if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001051 doing_tso = 1;
1052
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 if (dst) {
1054 u32 mtu = dst_mtu(dst);
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001055 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056 mss_now = tcp_sync_mss(sk, mtu);
1057 }
1058
Adam Langley33ad7982008-07-19 00:04:31 -07001059 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1060 sizeof(struct tcphdr);
1061 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1062 * some common options. If this is an odd packet (because we have SACK
1063 * blocks etc) then our calculated header_len will be different, and
1064 * we have to adjust mss_now correspondingly */
1065 if (header_len != tp->tcp_header_len) {
1066 int delta = (int) header_len - tp->tcp_header_len;
1067 mss_now -= delta;
1068 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08001069
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001070 xmit_size_goal = mss_now;
1071
1072 if (doing_tso) {
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001073 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
Arnaldo Carvalho de Melo8292a172005-12-13 23:15:52 -08001074 inet_csk(sk)->icsk_af_ops->net_header_len -
Arnaldo Carvalho de Melod83d8462005-12-13 23:26:10 -08001075 inet_csk(sk)->icsk_ext_hdr_len -
1076 tp->tcp_header_len);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001077
Ilpo Järvinen409d22b2007-12-31 14:57:40 -08001078 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001079 xmit_size_goal -= (xmit_size_goal % mss_now);
1080 }
1081 tp->xmit_size_goal = xmit_size_goal;
1082
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 return mss_now;
1084}
1085
David S. Millera762a982005-07-05 15:18:51 -07001086/* Congestion window validation. (RFC2861) */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001087static void tcp_cwnd_validate(struct sock *sk)
David S. Millera762a982005-07-05 15:18:51 -07001088{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001089 struct tcp_sock *tp = tcp_sk(sk);
David S. Millera762a982005-07-05 15:18:51 -07001090
Ilpo Järvinend436d682007-12-31 14:58:00 -08001091 if (tp->packets_out >= tp->snd_cwnd) {
David S. Millera762a982005-07-05 15:18:51 -07001092 /* Network is feed fully. */
1093 tp->snd_cwnd_used = 0;
1094 tp->snd_cwnd_stamp = tcp_time_stamp;
1095 } else {
1096 /* Network starves. */
1097 if (tp->packets_out > tp->snd_cwnd_used)
1098 tp->snd_cwnd_used = tp->packets_out;
1099
David S. Miller15d33c02007-04-09 13:23:14 -07001100 if (sysctl_tcp_slow_start_after_idle &&
1101 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
David S. Millera762a982005-07-05 15:18:51 -07001102 tcp_cwnd_application_limited(sk);
1103 }
1104}
1105
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001106/* Returns the portion of skb which can be sent right away without
1107 * introducing MSS oddities to segment boundaries. In rare cases where
1108 * mss_now != mss_cache, we will request caller to create a small skb
1109 * per input skb which could be mostly avoided here (if desired).
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001110 *
1111 * We explicitly want to create a request for splitting write queue tail
1112 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1113 * thus all the complexity (cwnd_len is always MSS multiple which we
1114 * return whenever allowed by the other factors). Basically we need the
1115 * modulo only when the receiver window alone is the limiting factor or
1116 * when we would be allowed to send the split-due-to-Nagle skb fully.
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001117 */
1118static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001119 unsigned int mss_now, unsigned int cwnd)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001120{
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001121 struct tcp_sock *tp = tcp_sk(sk);
1122 u32 needed, window, cwnd_len;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001123
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001124 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001125 cwnd_len = mss_now * cwnd;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001126
1127 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1128 return cwnd_len;
1129
Ilpo Järvinen5ea3a742008-03-11 17:55:27 -07001130 needed = min(skb->len, window);
1131
Ilpo Järvinen17515402008-04-15 20:36:55 -07001132 if (cwnd_len <= needed)
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001133 return cwnd_len;
1134
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001135 return needed - needed % mss_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001136}
1137
1138/* Can at least one segment of SKB be sent right now, according to the
1139 * congestion window rules? If so, return how many segments are allowed.
1140 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001141static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1142 struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001143{
1144 u32 in_flight, cwnd;
1145
1146 /* Don't be strict about the congestion window for the final FIN. */
John Heffner104439a2007-02-05 17:53:11 -08001147 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1148 tcp_skb_pcount(skb) == 1)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001149 return 1;
1150
1151 in_flight = tcp_packets_in_flight(tp);
1152 cwnd = tp->snd_cwnd;
1153 if (in_flight < cwnd)
1154 return (cwnd - in_flight);
1155
1156 return 0;
1157}
1158
1159/* This must be invoked the first time we consider transmitting
1160 * SKB onto the wire.
1161 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001162static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1163 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001164{
1165 int tso_segs = tcp_skb_pcount(skb);
1166
Petr Tesarik33cf71c2008-11-21 16:42:58 -08001167 if (!tso_segs ||
1168 (tso_segs > 1 && (tcp_skb_mss(skb) != mss_now ||
1169 tcp_urg_mode(tcp_sk(sk))))) {
David S. Miller846998a2005-08-04 19:52:01 -07001170 tcp_set_skb_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001171 tso_segs = tcp_skb_pcount(skb);
1172 }
1173 return tso_segs;
1174}
1175
1176static inline int tcp_minshall_check(const struct tcp_sock *tp)
1177{
1178 return after(tp->snd_sml,tp->snd_una) &&
1179 !after(tp->snd_sml, tp->snd_nxt);
1180}
1181
1182/* Return 0, if packet can be sent now without violation Nagle's rules:
1183 * 1. It is full sized.
1184 * 2. Or it contains FIN. (already checked by caller)
1185 * 3. Or TCP_NODELAY was set.
1186 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1187 * With Minshall's modification: all sent small packets are ACKed.
1188 */
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001189static inline int tcp_nagle_check(const struct tcp_sock *tp,
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001190 const struct sk_buff *skb,
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001191 unsigned mss_now, int nonagle)
1192{
1193 return (skb->len < mss_now &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001194 ((nonagle & TCP_NAGLE_CORK) ||
1195 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001196}
1197
1198/* Return non-zero if the Nagle test allows this packet to be
1199 * sent now.
1200 */
1201static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1202 unsigned int cur_mss, int nonagle)
1203{
1204 /* Nagle rule does not apply to frames, which sit in the middle of the
1205 * write_queue (they have no chances to get new data).
1206 *
1207 * This is implemented in the callers, where they modify the 'nonagle'
1208 * argument based upon the location of SKB in the send queue.
1209 */
1210 if (nonagle & TCP_NAGLE_PUSH)
1211 return 1;
1212
Ilpo Järvinend551e452007-04-30 00:42:20 -07001213 /* Don't use the nagle rule for urgent data (or for the final FIN).
1214 * Nagle can be ignored during F-RTO too (see RFC4138).
1215 */
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07001216 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001217 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1218 return 1;
1219
1220 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1221 return 1;
1222
1223 return 0;
1224}
1225
1226/* Does at least the first segment of SKB fit into the send window? */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001227static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1228 unsigned int cur_mss)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001229{
1230 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1231
1232 if (skb->len > cur_mss)
1233 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1234
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001235 return !after(end_seq, tcp_wnd_end(tp));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001236}
1237
David S. Millerfe067e82007-03-07 12:12:44 -08001238/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001239 * should be put on the wire right now. If so, it returns the number of
1240 * packets allowed by the congestion window.
1241 */
1242static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1243 unsigned int cur_mss, int nonagle)
1244{
1245 struct tcp_sock *tp = tcp_sk(sk);
1246 unsigned int cwnd_quota;
1247
David S. Miller846998a2005-08-04 19:52:01 -07001248 tcp_init_tso_segs(sk, skb, cur_mss);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001249
1250 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1251 return 0;
1252
1253 cwnd_quota = tcp_cwnd_test(tp, skb);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001254 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001255 cwnd_quota = 0;
1256
1257 return cwnd_quota;
1258}
1259
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001260int tcp_may_send_now(struct sock *sk)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001261{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001262 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001263 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001264
1265 return (skb &&
1266 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1267 (tcp_skb_is_last(sk, skb) ?
Ilpo Järvinen4e67d872007-12-05 02:25:32 -08001268 tp->nonagle : TCP_NAGLE_PUSH)));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001269}
1270
1271/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1272 * which is put after SKB on the list. It is very much like
1273 * tcp_fragment() except that it may make several kinds of assumptions
1274 * in order to speed up the splitting operation. In particular, we
1275 * know that all the data is in scatter-gather pages, and that the
1276 * packet has never been sent out before (and thus is not cloned).
1277 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001278static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1279 unsigned int mss_now)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001280{
1281 struct sk_buff *buff;
1282 int nlen = skb->len - len;
1283 u16 flags;
1284
1285 /* All of a TSO frame must be composed of paged data. */
Herbert Xuc8ac3772005-08-16 20:43:40 -07001286 if (skb->len != skb->data_len)
1287 return tcp_fragment(sk, skb, len, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001288
Pavel Emelyanovdf97c702007-11-29 21:22:33 +11001289 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001290 if (unlikely(buff == NULL))
1291 return -ENOMEM;
1292
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001293 sk->sk_wmem_queued += buff->truesize;
1294 sk_mem_charge(sk, buff->truesize);
Herbert Xub60b49e2006-04-19 21:35:00 -07001295 buff->truesize += nlen;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001296 skb->truesize -= nlen;
1297
1298 /* Correct the sequence numbers. */
1299 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1300 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1301 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1302
1303 /* PSH and FIN should only be set in the second packet. */
1304 flags = TCP_SKB_CB(skb)->flags;
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001305 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001306 TCP_SKB_CB(buff)->flags = flags;
1307
1308 /* This packet was never sent out yet, so no SACK bits. */
1309 TCP_SKB_CB(buff)->sacked = 0;
1310
Patrick McHardy84fa7932006-08-29 16:44:56 -07001311 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001312 skb_split(skb, buff, len);
1313
1314 /* Fix up tso_factor for both original and new SKB. */
David S. Miller846998a2005-08-04 19:52:01 -07001315 tcp_set_skb_tso_segs(sk, skb, mss_now);
1316 tcp_set_skb_tso_segs(sk, buff, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001317
1318 /* Link BUFF into the send queue. */
1319 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08001320 tcp_insert_write_queue_after(skb, buff, sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001321
1322 return 0;
1323}
1324
1325/* Try to defer sending, if possible, in order to minimize the amount
1326 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1327 *
1328 * This algorithm is from John Heffner.
1329 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001330static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001331{
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001332 struct tcp_sock *tp = tcp_sk(sk);
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001333 const struct inet_connection_sock *icsk = inet_csk(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001334 u32 send_win, cong_win, limit, in_flight;
1335
1336 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
John Heffnerae8064a2006-10-18 20:36:48 -07001337 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001338
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001339 if (icsk->icsk_ca_state != TCP_CA_Open)
John Heffnerae8064a2006-10-18 20:36:48 -07001340 goto send_now;
1341
1342 /* Defer for less than two clock ticks. */
Ilpo Järvinenbd515c32007-12-20 20:36:03 -08001343 if (tp->tso_deferred &&
1344 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
John Heffnerae8064a2006-10-18 20:36:48 -07001345 goto send_now;
David S. Miller908a75c2005-07-05 15:43:58 -07001346
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001347 in_flight = tcp_packets_in_flight(tp);
1348
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001349 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001350
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001351 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001352
1353 /* From in_flight test above, we know that cwnd > in_flight. */
1354 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1355
1356 limit = min(send_win, cong_win);
1357
David S. Millerba244fe2006-03-11 18:51:49 -08001358 /* If a full-sized TSO skb can be sent, do it. */
Peter P Waskiewicz Jr82cc1a72008-03-21 03:43:19 -07001359 if (limit >= sk->sk_gso_max_size)
John Heffnerae8064a2006-10-18 20:36:48 -07001360 goto send_now;
David S. Millerba244fe2006-03-11 18:51:49 -08001361
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001362 if (sysctl_tcp_tso_win_divisor) {
1363 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1364
1365 /* If at least some fraction of a window is available,
1366 * just use it.
1367 */
1368 chunk /= sysctl_tcp_tso_win_divisor;
1369 if (limit >= chunk)
John Heffnerae8064a2006-10-18 20:36:48 -07001370 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001371 } else {
1372 /* Different approach, try not to defer past a single
1373 * ACK. Receiver should ACK every other full sized
1374 * frame, so if we have space for more than 3 frames
1375 * then send now.
1376 */
1377 if (limit > tcp_max_burst(tp) * tp->mss_cache)
John Heffnerae8064a2006-10-18 20:36:48 -07001378 goto send_now;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001379 }
1380
1381 /* Ok, it looks like it is advisable to defer. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001382 tp->tso_deferred = 1 | (jiffies << 1);
John Heffnerae8064a2006-10-18 20:36:48 -07001383
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001384 return 1;
John Heffnerae8064a2006-10-18 20:36:48 -07001385
1386send_now:
1387 tp->tso_deferred = 0;
1388 return 0;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001389}
1390
John Heffner5d424d52006-03-20 17:53:41 -08001391/* Create a new MTU probe if we are ready.
1392 * Returns 0 if we should wait to probe (no cwnd available),
1393 * 1 if a probe was sent,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001394 * -1 otherwise
1395 */
John Heffner5d424d52006-03-20 17:53:41 -08001396static int tcp_mtu_probe(struct sock *sk)
1397{
1398 struct tcp_sock *tp = tcp_sk(sk);
1399 struct inet_connection_sock *icsk = inet_csk(sk);
1400 struct sk_buff *skb, *nskb, *next;
1401 int len;
1402 int probe_size;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001403 int size_needed;
John Heffner5d424d52006-03-20 17:53:41 -08001404 int copy;
1405 int mss_now;
1406
1407 /* Not currently probing/verifying,
1408 * not in recovery,
1409 * have enough cwnd, and
1410 * not SACKing (the variable headers throw things off) */
1411 if (!icsk->icsk_mtup.enabled ||
1412 icsk->icsk_mtup.probe_size ||
1413 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1414 tp->snd_cwnd < 11 ||
1415 tp->rx_opt.eff_sacks)
1416 return -1;
1417
1418 /* Very simple search strategy: just double the MSS. */
1419 mss_now = tcp_current_mss(sk, 0);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001420 probe_size = 2 * tp->mss_cache;
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001421 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
John Heffner5d424d52006-03-20 17:53:41 -08001422 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1423 /* TODO: set timer for probe_converge_event */
1424 return -1;
1425 }
1426
1427 /* Have enough data in the send queue to probe? */
Ilpo Järvinen7f9c33e2007-11-23 19:10:56 +08001428 if (tp->write_seq - tp->snd_nxt < size_needed)
John Heffner5d424d52006-03-20 17:53:41 -08001429 return -1;
1430
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001431 if (tp->snd_wnd < size_needed)
1432 return -1;
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001433 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
Ilpo Järvinen91cc17c2007-11-23 19:08:16 +08001434 return 0;
John Heffner5d424d52006-03-20 17:53:41 -08001435
Ilpo Järvinend67c58e2007-12-02 00:48:01 +02001436 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1437 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1438 if (!tcp_packets_in_flight(tp))
John Heffner5d424d52006-03-20 17:53:41 -08001439 return -1;
1440 else
1441 return 0;
1442 }
1443
1444 /* We're allowed to probe. Build it now. */
1445 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1446 return -1;
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001447 sk->sk_wmem_queued += nskb->truesize;
1448 sk_mem_charge(sk, nskb->truesize);
John Heffner5d424d52006-03-20 17:53:41 -08001449
David S. Millerfe067e82007-03-07 12:12:44 -08001450 skb = tcp_send_head(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001451
1452 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1453 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1454 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1455 TCP_SKB_CB(nskb)->sacked = 0;
1456 nskb->csum = 0;
Patrick McHardy84fa7932006-08-29 16:44:56 -07001457 nskb->ip_summed = skb->ip_summed;
John Heffner5d424d52006-03-20 17:53:41 -08001458
Ilpo Järvinen50c48172007-12-02 00:48:00 +02001459 tcp_insert_write_queue_before(nskb, skb, sk);
1460
John Heffner5d424d52006-03-20 17:53:41 -08001461 len = 0;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001462 tcp_for_write_queue_from_safe(skb, next, sk) {
John Heffner5d424d52006-03-20 17:53:41 -08001463 copy = min_t(int, skb->len, probe_size - len);
1464 if (nskb->ip_summed)
1465 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1466 else
1467 nskb->csum = skb_copy_and_csum_bits(skb, 0,
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001468 skb_put(nskb, copy),
1469 copy, nskb->csum);
John Heffner5d424d52006-03-20 17:53:41 -08001470
1471 if (skb->len <= copy) {
1472 /* We've eaten all the data from this skb.
1473 * Throw it away. */
1474 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
David S. Millerfe067e82007-03-07 12:12:44 -08001475 tcp_unlink_write_queue(skb, sk);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08001476 sk_wmem_free_skb(sk, skb);
John Heffner5d424d52006-03-20 17:53:41 -08001477 } else {
1478 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001479 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
John Heffner5d424d52006-03-20 17:53:41 -08001480 if (!skb_shinfo(skb)->nr_frags) {
1481 skb_pull(skb, copy);
Patrick McHardy84fa7932006-08-29 16:44:56 -07001482 if (skb->ip_summed != CHECKSUM_PARTIAL)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001483 skb->csum = csum_partial(skb->data,
1484 skb->len, 0);
John Heffner5d424d52006-03-20 17:53:41 -08001485 } else {
1486 __pskb_trim_head(skb, copy);
1487 tcp_set_skb_tso_segs(sk, skb, mss_now);
1488 }
1489 TCP_SKB_CB(skb)->seq += copy;
1490 }
1491
1492 len += copy;
Ilpo Järvinen234b6862007-12-02 00:48:02 +02001493
1494 if (len >= probe_size)
1495 break;
John Heffner5d424d52006-03-20 17:53:41 -08001496 }
1497 tcp_init_tso_segs(sk, nskb, nskb->len);
1498
1499 /* We're ready to send. If this fails, the probe will
1500 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1501 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1502 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1503 /* Decrement cwnd here because we are sending
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001504 * effectively two packets. */
John Heffner5d424d52006-03-20 17:53:41 -08001505 tp->snd_cwnd--;
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001506 tcp_event_new_data_sent(sk, nskb);
John Heffner5d424d52006-03-20 17:53:41 -08001507
1508 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
John Heffner0e7b1362006-03-20 21:32:58 -08001509 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1510 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
John Heffner5d424d52006-03-20 17:53:41 -08001511
1512 return 1;
1513 }
1514
1515 return -1;
1516}
1517
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518/* This routine writes packets to the network. It advances the
1519 * send_head. This happens as incoming acks open up the remote
1520 * window for us.
1521 *
1522 * Returns 1, if no segments are in flight and we have queued segments, but
1523 * cannot send anything now because of SWS or another problem.
1524 */
David S. Millera2e2a592005-07-05 15:19:23 -07001525static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526{
1527 struct tcp_sock *tp = tcp_sk(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001528 struct sk_buff *skb;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001529 unsigned int tso_segs, sent_pkts;
1530 int cwnd_quota;
John Heffner5d424d52006-03-20 17:53:41 -08001531 int result;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
1533 /* If we are closed, the bytes will have to remain here.
1534 * In time closedown will finish, we empty the write queue and all
1535 * will be happy.
1536 */
David S. Miller92df7b52005-07-05 15:19:06 -07001537 if (unlikely(sk->sk_state == TCP_CLOSE))
1538 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539
David S. Miller92df7b52005-07-05 15:19:06 -07001540 sent_pkts = 0;
John Heffner5d424d52006-03-20 17:53:41 -08001541
1542 /* Do MTU probing. */
1543 if ((result = tcp_mtu_probe(sk)) == 0) {
1544 return 0;
1545 } else if (result > 0) {
1546 sent_pkts = 1;
1547 }
1548
David S. Millerfe067e82007-03-07 12:12:44 -08001549 while ((skb = tcp_send_head(sk))) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001550 unsigned int limit;
1551
Herbert Xub68e9f82005-08-04 19:52:02 -07001552 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001553 BUG_ON(!tso_segs);
David S. Milleraa934662005-07-05 15:20:09 -07001554
Herbert Xub68e9f82005-08-04 19:52:02 -07001555 cwnd_quota = tcp_cwnd_test(tp, skb);
1556 if (!cwnd_quota)
1557 break;
1558
1559 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1560 break;
1561
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001562 if (tso_segs == 1) {
1563 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1564 (tcp_skb_is_last(sk, skb) ?
1565 nonagle : TCP_NAGLE_PUSH))))
1566 break;
1567 } else {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001568 if (tcp_tso_should_defer(sk, skb))
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001569 break;
1570 }
David S. Milleraa934662005-07-05 15:20:09 -07001571
Herbert Xuc8ac3772005-08-16 20:43:40 -07001572 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001573 if (tso_segs > 1)
1574 limit = tcp_mss_split_point(sk, skb, mss_now,
1575 cwnd_quota);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576
Herbert Xuc8ac3772005-08-16 20:43:40 -07001577 if (skb->len > limit &&
1578 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1579 break;
1580
David S. Miller92df7b52005-07-05 15:19:06 -07001581 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001582
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001583 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
David S. Miller92df7b52005-07-05 15:19:06 -07001584 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585
David S. Miller92df7b52005-07-05 15:19:06 -07001586 /* Advance the send_head. This one is sent out.
1587 * This call will increment packets_out.
1588 */
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001589 tcp_event_new_data_sent(sk, skb);
David S. Miller92df7b52005-07-05 15:19:06 -07001590
1591 tcp_minshall_update(tp, mss_now, skb);
David S. Milleraa934662005-07-05 15:20:09 -07001592 sent_pkts++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 }
David S. Miller92df7b52005-07-05 15:19:06 -07001594
David S. Milleraa934662005-07-05 15:20:09 -07001595 if (likely(sent_pkts)) {
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001596 tcp_cwnd_validate(sk);
David S. Miller92df7b52005-07-05 15:19:06 -07001597 return 0;
1598 }
David S. Millerfe067e82007-03-07 12:12:44 -08001599 return !tp->packets_out && tcp_send_head(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600}
1601
David S. Millera762a982005-07-05 15:18:51 -07001602/* Push out any pending frames which were held back due to
1603 * TCP_CORK or attempt at coalescing tiny packets.
1604 * The socket must be locked by the caller.
1605 */
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001606void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1607 int nonagle)
David S. Millera762a982005-07-05 15:18:51 -07001608{
David S. Millerfe067e82007-03-07 12:12:44 -08001609 struct sk_buff *skb = tcp_send_head(sk);
David S. Millera762a982005-07-05 15:18:51 -07001610
1611 if (skb) {
David S. Miller55c97f32005-07-05 15:19:38 -07001612 if (tcp_write_xmit(sk, cur_mss, nonagle))
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001613 tcp_check_probe_timer(sk);
David S. Millera762a982005-07-05 15:18:51 -07001614 }
1615}
1616
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001617/* Send _single_ skb sitting at the send head. This function requires
1618 * true push pending frames to setup probe timer etc.
1619 */
1620void tcp_push_one(struct sock *sk, unsigned int mss_now)
1621{
David S. Millerfe067e82007-03-07 12:12:44 -08001622 struct sk_buff *skb = tcp_send_head(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001623 unsigned int tso_segs, cwnd_quota;
1624
1625 BUG_ON(!skb || skb->len < mss_now);
1626
David S. Miller846998a2005-08-04 19:52:01 -07001627 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001628 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1629
1630 if (likely(cwnd_quota)) {
Herbert Xuc8ac3772005-08-16 20:43:40 -07001631 unsigned int limit;
1632
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001633 BUG_ON(!tso_segs);
1634
Herbert Xuc8ac3772005-08-16 20:43:40 -07001635 limit = mss_now;
Ilpo Järvinen0e3a4802007-12-24 21:33:45 -08001636 if (tso_segs > 1)
1637 limit = tcp_mss_split_point(sk, skb, mss_now,
1638 cwnd_quota);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001639
Herbert Xuc8ac3772005-08-16 20:43:40 -07001640 if (skb->len > limit &&
1641 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1642 return;
1643
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001644 /* Send it out now. */
1645 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1646
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001647 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
Ilpo Järvinen66f5fe62007-12-31 04:43:57 -08001648 tcp_event_new_data_sent(sk, skb);
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07001649 tcp_cwnd_validate(sk);
David S. Millerc1b4a7e2005-07-05 15:24:38 -07001650 return;
1651 }
1652 }
1653}
1654
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655/* This function returns the amount that we can raise the
1656 * usable window based on the following constraints
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001657 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 * 1. The window can never be shrunk once it is offered (RFC 793)
1659 * 2. We limit memory per socket
1660 *
1661 * RFC 1122:
1662 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1663 * RECV.NEXT + RCV.WIN fixed until:
1664 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1665 *
1666 * i.e. don't raise the right edge of the window until you can raise
1667 * it at least MSS bytes.
1668 *
1669 * Unfortunately, the recommended algorithm breaks header prediction,
1670 * since header prediction assumes th->window stays fixed.
1671 *
1672 * Strictly speaking, keeping th->window fixed violates the receiver
1673 * side SWS prevention criteria. The problem is that under this rule
1674 * a stream of single byte packets will cause the right side of the
1675 * window to always advance by a single byte.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001676 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677 * Of course, if the sender implements sender side SWS prevention
1678 * then this will not be a problem.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001679 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 * BSD seems to make the following compromise:
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001681 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 * If the free space is less than the 1/4 of the maximum
1683 * space available and the free space is less than 1/2 mss,
1684 * then set the window to 0.
1685 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1686 * Otherwise, just prevent the window from shrinking
1687 * and from being larger than the largest representable value.
1688 *
1689 * This prevents incremental opening of the window in the regime
1690 * where TCP is limited by the speed of the reader side taking
1691 * data out of the TCP receive queue. It does nothing about
1692 * those cases where the window is constrained on the sender side
1693 * because the pipeline is full.
1694 *
1695 * BSD also seems to "accidentally" limit itself to windows that are a
1696 * multiple of MSS, at least until the free space gets quite small.
1697 * This would appear to be a side effect of the mbuf implementation.
1698 * Combining these two algorithms results in the observed behavior
1699 * of having a fixed window size at almost all times.
1700 *
1701 * Below we obtain similar behavior by forcing the offered window to
1702 * a multiple of the mss when it is feasible to do so.
1703 *
1704 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1705 * Regular options like TIMESTAMP are taken into account.
1706 */
1707u32 __tcp_select_window(struct sock *sk)
1708{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001709 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 struct tcp_sock *tp = tcp_sk(sk);
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001711 /* MSS for the peer's data. Previous versions used mss_clamp
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 * here. I don't know if the value based on our guesses
1713 * of peer's MSS is better for the performance. It's more correct
1714 * but may be worse for the performance because of rcv_mss
1715 * fluctuations. --SAW 1998/11/1
1716 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001717 int mss = icsk->icsk_ack.rcv_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 int free_space = tcp_space(sk);
1719 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1720 int window;
1721
1722 if (mss > full_space)
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001723 mss = full_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001725 if (free_space < (full_space >> 1)) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07001726 icsk->icsk_ack.quick = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727
1728 if (tcp_memory_pressure)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001729 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1730 4U * tp->advmss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731
1732 if (free_space < mss)
1733 return 0;
1734 }
1735
1736 if (free_space > tp->rcv_ssthresh)
1737 free_space = tp->rcv_ssthresh;
1738
1739 /* Don't do rounding if we are using window scaling, since the
1740 * scaled window will not line up with the MSS boundary anyway.
1741 */
1742 window = tp->rcv_wnd;
1743 if (tp->rx_opt.rcv_wscale) {
1744 window = free_space;
1745
1746 /* Advertise enough space so that it won't get scaled away.
1747 * Import case: prevent zero window announcement if
1748 * 1<<rcv_wscale > mss.
1749 */
1750 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1751 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1752 << tp->rx_opt.rcv_wscale);
1753 } else {
1754 /* Get the largest window that is a nice multiple of mss.
1755 * Window clamp already applied above.
1756 * If our current window offering is within 1 mss of the
1757 * free space we just keep it. This prevents the divide
1758 * and multiply from happening most of the time.
1759 * We also don't do any window rounding when the free space
1760 * is too small.
1761 */
1762 if (window <= free_space - mss || window > free_space)
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001763 window = (free_space / mss) * mss;
John Heffner84565072007-04-02 13:56:32 -07001764 else if (mss == full_space &&
Eric Dumazetb92edbe2007-12-20 21:48:32 -08001765 free_space > window + (full_space >> 1))
John Heffner84565072007-04-02 13:56:32 -07001766 window = free_space;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767 }
1768
1769 return window;
1770}
1771
1772/* Attempt to collapse two adjacent SKB's during retransmission. */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001773static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1774 int mss_now)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775{
1776 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08001777 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001778 int skb_size, next_skb_size;
1779 u16 flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780
1781 /* The first test we must make is that neither of these two
1782 * SKB's are still referenced by someone else.
1783 */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001784 if (skb_cloned(skb) || skb_cloned(next_skb))
1785 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001787 skb_size = skb->len;
1788 next_skb_size = next_skb->len;
1789 flags = TCP_SKB_CB(skb)->flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001791 /* Also punt if next skb has been SACK'd. */
1792 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1793 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001795 /* Next skb is out of window. */
1796 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1797 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001799 /* Punt if not enough space exists in the first SKB for
1800 * the data in the second, or the total combined payload
1801 * would exceed the MSS.
1802 */
1803 if ((next_skb_size > skb_tailroom(skb)) ||
1804 ((skb_size + next_skb_size) > mss_now))
1805 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001807 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
Ilpo Järvinena6963a62007-09-25 22:44:14 -07001808
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001809 tcp_highest_sack_combine(sk, next_skb, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001811 /* Ok. We will be able to collapse the packet. */
1812 tcp_unlink_write_queue(next_skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001814 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1815 next_skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001817 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1818 skb->ip_summed = CHECKSUM_PARTIAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001820 if (skb->ip_summed != CHECKSUM_PARTIAL)
1821 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001823 /* Update sequence range on original skb. */
1824 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001826 /* Merge over control information. */
1827 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1828 TCP_SKB_CB(skb)->flags = flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001830 /* All done, get rid of second SKB and account for it so
1831 * packet counting does not break.
1832 */
1833 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1834 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1835 tp->retrans_out -= tcp_skb_pcount(next_skb);
1836 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1837 tp->lost_out -= tcp_skb_pcount(next_skb);
1838 /* Reno case is special. Sigh... */
1839 if (tcp_is_reno(tp) && tp->sacked_out)
1840 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001841
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001842 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1843 tp->packets_out -= tcp_skb_pcount(next_skb);
Ilpo Järvinenb7689202007-09-20 11:37:19 -07001844
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001845 /* changed transmit queue under us so clear hints */
Ilpo Järvinenef9da472008-09-20 21:25:15 -07001846 tcp_clear_retrans_hints_partial(tp);
1847 if (next_skb == tp->retransmit_skb_hint)
1848 tp->retransmit_skb_hint = skb;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08001849
1850 sk_wmem_free_skb(sk, next_skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851}
1852
1853/* Do a simple retransmit without using the backoff mechanisms in
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001854 * tcp_timer. This is used for path mtu discovery.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 * The socket is already locked here.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001856 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857void tcp_simple_retransmit(struct sock *sk)
1858{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001859 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 struct tcp_sock *tp = tcp_sk(sk);
1861 struct sk_buff *skb;
1862 unsigned int mss = tcp_current_mss(sk, 0);
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001863 u32 prior_lost = tp->lost_out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864
David S. Millerfe067e82007-03-07 12:12:44 -08001865 tcp_for_write_queue(skb, sk) {
1866 if (skb == tcp_send_head(sk))
1867 break;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001868 if (skb->len > mss &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001869 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1870 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1872 tp->retrans_out -= tcp_skb_pcount(skb);
1873 }
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001874 tcp_skb_mark_lost_uncond_verify(tp, skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 }
1876 }
1877
Ilpo Järvinenef9da472008-09-20 21:25:15 -07001878 tcp_clear_retrans_hints_partial(tp);
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08001879
Ilpo Järvinen006f5822008-09-20 21:20:20 -07001880 if (prior_lost == tp->lost_out)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 return;
1882
Ilpo Järvinen882beba2008-04-07 22:33:07 -07001883 if (tcp_is_reno(tp))
1884 tcp_limit_reno_sacked(tp);
1885
Ilpo Järvinen005903b2007-08-09 14:44:16 +03001886 tcp_verify_left_out(tp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09001888 /* Don't muck with the congestion window here.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 * Reason is that we do not increase amount of _data_
1890 * in network, but units changed and effective
1891 * cwnd/ssthresh really reduced now.
1892 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001893 if (icsk->icsk_ca_state != TCP_CA_Loss) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 tp->high_seq = tp->snd_nxt;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001895 tp->snd_ssthresh = tcp_current_ssthresh(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 tp->prior_ssthresh = 0;
1897 tp->undo_marker = 0;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03001898 tcp_set_ca_state(sk, TCP_CA_Loss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 }
1900 tcp_xmit_retransmit_queue(sk);
1901}
1902
1903/* This retransmits one SKB. Policy decisions and retransmit queue
1904 * state updates are done by the caller. Returns non-zero if an
1905 * error occurred which prevented the send.
1906 */
1907int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1908{
1909 struct tcp_sock *tp = tcp_sk(sk);
John Heffner5d424d52006-03-20 17:53:41 -08001910 struct inet_connection_sock *icsk = inet_csk(sk);
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07001911 unsigned int cur_mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 int err;
1913
John Heffner5d424d52006-03-20 17:53:41 -08001914 /* Inconslusive MTU probe */
1915 if (icsk->icsk_mtup.probe_size) {
1916 icsk->icsk_mtup.probe_size = 0;
1917 }
1918
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 /* Do not sent more than we queued. 1/4 is reserved for possible
Stephen Hemmingercaa20d9a2005-11-10 17:13:47 -08001920 * copying overhead: fragmentation, tunneling, mangling etc.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 */
1922 if (atomic_read(&sk->sk_wmem_alloc) >
1923 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1924 return -EAGAIN;
1925
1926 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1927 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1928 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1930 return -ENOMEM;
1931 }
1932
Sridhar Samudrala7d227cd22008-05-21 16:42:20 -07001933 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1934 return -EHOSTUNREACH; /* Routing failure or similar. */
1935
1936 cur_mss = tcp_current_mss(sk, 0);
1937
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 /* If receiver has shrunk his window, and skb is out of
1939 * new window, do not retransmit it. The exception is the
1940 * case, when window is shrunk to zero. In this case
1941 * our retransmit serves as a zero window probe.
1942 */
Ilpo Järvinen90840de2007-12-31 04:48:41 -08001943 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1945 return -EAGAIN;
1946
1947 if (skb->len > cur_mss) {
David S. Miller846998a2005-08-04 19:52:01 -07001948 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 return -ENOMEM; /* We'll try again later. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 }
1951
1952 /* Collapse two adjacent packets if worthwhile and we can. */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001953 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1954 (skb->len < (cur_mss >> 1)) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001955 (!tcp_skb_is_last(sk, skb)) &&
David S. Miller77d40a02008-09-23 01:29:23 -07001956 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001957 (skb_shinfo(skb)->nr_frags == 0 &&
1958 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1959 (tcp_skb_pcount(skb) == 1 &&
1960 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001961 (sysctl_tcp_retrans_collapse != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 tcp_retrans_try_collapse(sk, skb, cur_mss);
1963
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 /* Some Solaris stacks overoptimize and ignore the FIN on a
1965 * retransmit when old data is attached. So strip it off
1966 * since it is cheap to do so and saves bytes on the network.
1967 */
Stephen Hemminger2de979b2007-03-08 20:45:19 -08001968 if (skb->len > 0 &&
1969 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1970 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 if (!pskb_trim(skb, 0)) {
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08001972 /* Reuse, even though it does some unnecessary work */
1973 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1974 TCP_SKB_CB(skb)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 skb->ip_summed = CHECKSUM_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 }
1977 }
1978
1979 /* Make a copy, if the first transmission SKB clone we made
1980 * is still in somebody's hands, else make a clone.
1981 */
1982 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983
David S. Millerdfb4b9d2005-12-06 16:24:52 -08001984 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985
1986 if (err == 0) {
1987 /* Update global TCP statistics. */
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07001988 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989
1990 tp->total_retrans++;
1991
1992#if FASTRETRANS_DEBUG > 0
Ilpo Järvinen056834d2007-12-31 14:57:14 -08001993 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 if (net_ratelimit())
1995 printk(KERN_DEBUG "retrans_out leaked.\n");
1996 }
1997#endif
Ilpo Järvinenb08d6cb2007-10-11 17:36:13 -07001998 if (!tp->retrans_out)
1999 tp->lost_retrans_low = tp->snd_nxt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2001 tp->retrans_out += tcp_skb_pcount(skb);
2002
2003 /* Save stamp of the first retransmit. */
2004 if (!tp->retrans_stamp)
2005 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2006
2007 tp->undo_retrans++;
2008
2009 /* snd_nxt is stored to detect loss of retransmitted segment,
2010 * see tcp_input.c tcp_sacktag_write_queue().
2011 */
2012 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2013 }
2014 return err;
2015}
2016
Ilpo Järvinenb5afe7b2008-09-20 21:21:54 -07002017static int tcp_can_forward_retransmit(struct sock *sk)
2018{
2019 const struct inet_connection_sock *icsk = inet_csk(sk);
2020 struct tcp_sock *tp = tcp_sk(sk);
2021
2022 /* Forward retransmissions are possible only during Recovery. */
2023 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2024 return 0;
2025
2026 /* No forward retransmissions in Reno are possible. */
2027 if (tcp_is_reno(tp))
2028 return 0;
2029
2030 /* Yeah, we have to make difficult choice between forward transmission
2031 * and retransmission... Both ways have their merits...
2032 *
2033 * For now we do not retransmit anything, while we have some new
2034 * segments to send. In the other cases, follow rule 3 for
2035 * NextSeg() specified in RFC3517.
2036 */
2037
2038 if (tcp_may_send_now(sk))
2039 return 0;
2040
2041 return 1;
2042}
2043
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044/* This gets called after a retransmit timeout, and the initially
2045 * retransmitted data is acknowledged. It tries to continue
2046 * resending the rest of the retransmit queue, until either
2047 * we've sent it all or the congestion window limit is reached.
2048 * If doing SACK, the first ACK which comes back for a timeout
2049 * based retransmit packet might feed us FACK information again.
2050 * If so, we use it to avoid unnecessarily retransmissions.
2051 */
2052void tcp_xmit_retransmit_queue(struct sock *sk)
2053{
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002054 const struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct sk_buff *skb;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002057 struct sk_buff *hole = NULL;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002058 u32 last_lost;
Ilpo Järvinen61eb55f2008-09-20 21:22:59 -07002059 int mib_idx;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002060 int fwd_rexmitting = 0;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002061
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002062 if (!tp->lost_out)
2063 tp->retransmit_high = tp->snd_una;
2064
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002065 if (tp->retransmit_skb_hint) {
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002066 skb = tp->retransmit_skb_hint;
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002067 last_lost = TCP_SKB_CB(skb)->end_seq;
2068 if (after(last_lost, tp->retransmit_high))
2069 last_lost = tp->retransmit_high;
2070 } else {
David S. Millerfe067e82007-03-07 12:12:44 -08002071 skb = tcp_write_queue_head(sk);
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002072 last_lost = tp->snd_una;
2073 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
2075 /* First pass: retransmit lost packets. */
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002076 tcp_for_write_queue_from(skb, sk) {
2077 __u8 sacked = TCP_SKB_CB(skb)->sacked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002079 if (skb == tcp_send_head(sk))
2080 break;
2081 /* we could do better than to assign each time */
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002082 if (hole == NULL)
2083 tp->retransmit_skb_hint = skb;
Stephen Hemminger6a438bb2005-11-10 17:14:59 -08002084
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002085 /* Assume this retransmit will generate
2086 * only one packet for congestion window
2087 * calculation purposes. This works because
2088 * tcp_retransmit_skb() will chop up the
2089 * packet to be MSS sized and all the
2090 * packet counting works out.
2091 */
2092 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2093 return;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002094
2095 if (fwd_rexmitting) {
2096begin_fwd:
2097 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2098 break;
2099 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2100
2101 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002102 tp->retransmit_high = last_lost;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002103 if (!tcp_can_forward_retransmit(sk))
2104 break;
2105 /* Backtrack if necessary to non-L'ed skb */
2106 if (hole != NULL) {
2107 skb = hole;
2108 hole = NULL;
2109 }
2110 fwd_rexmitting = 1;
2111 goto begin_fwd;
2112
2113 } else if (!(sacked & TCPCB_LOST)) {
2114 if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS))
2115 hole = skb;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002116 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002118 } else {
Ilpo Järvinen618d9f22008-09-20 21:26:22 -07002119 last_lost = TCP_SKB_CB(skb)->end_seq;
Ilpo Järvinen0e1c54c2008-09-20 21:24:21 -07002120 if (icsk->icsk_ca_state != TCP_CA_Loss)
2121 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2122 else
2123 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2124 }
2125
2126 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002127 continue;
Pavel Emelyanov40b215e2008-07-03 01:05:41 -07002128
Ilpo Järvinenf0ceb0e2008-09-20 21:24:49 -07002129 if (tcp_retransmit_skb(sk, skb))
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002130 return;
Ilpo Järvinen08ebd172008-09-20 21:23:49 -07002131 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2132
2133 if (skb == tcp_write_queue_head(sk))
2134 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2135 inet_csk(sk)->icsk_rto,
2136 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138}
2139
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140/* Send a fin. The caller locks the socket for us. This cannot be
2141 * allowed to fail queueing a FIN frame under any circumstances.
2142 */
2143void tcp_send_fin(struct sock *sk)
2144{
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002145 struct tcp_sock *tp = tcp_sk(sk);
David S. Millerfe067e82007-03-07 12:12:44 -08002146 struct sk_buff *skb = tcp_write_queue_tail(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147 int mss_now;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002148
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 /* Optimization, tack on the FIN if we have a queue of
2150 * unsent frames. But be careful about outgoing SACKS
2151 * and IP options.
2152 */
2153 mss_now = tcp_current_mss(sk, 1);
2154
David S. Millerfe067e82007-03-07 12:12:44 -08002155 if (tcp_send_head(sk) != NULL) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2157 TCP_SKB_CB(skb)->end_seq++;
2158 tp->write_seq++;
2159 } else {
2160 /* Socket is locked, keep trying until memory is available. */
2161 for (;;) {
David S. Millerd179cd12005-08-17 14:57:30 -07002162 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 if (skb)
2164 break;
2165 yield();
2166 }
2167
2168 /* Reserve space for headers and prepare control bits. */
2169 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002171 tcp_init_nondata_skb(skb, tp->write_seq,
2172 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 tcp_queue_skb(sk, skb);
2174 }
Ilpo Järvinen9e412ba2007-04-20 22:18:02 -07002175 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176}
2177
2178/* We get here when a process closes a file descriptor (either due to
2179 * an explicit close() or as a byproduct of exit()'ing) and there
2180 * was unread data in the receive queue. This behavior is recommended
Gerrit Renker65bb7232007-04-28 21:21:46 -07002181 * by RFC 2525, section 2.17. -DaveM
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182 */
Al Virodd0fc662005-10-07 07:46:04 +01002183void tcp_send_active_reset(struct sock *sk, gfp_t priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 struct sk_buff *skb;
2186
2187 /* NOTE: No TCP options attached and we never retransmit this. */
2188 skb = alloc_skb(MAX_TCP_HEADER, priority);
2189 if (!skb) {
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002190 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 return;
2192 }
2193
2194 /* Reserve space for headers and prepare control bits. */
2195 skb_reserve(skb, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002196 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2197 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 /* Send it off. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002200 if (tcp_transmit_skb(sk, skb, 0, priority))
Pavel Emelyanov4e673442008-07-16 20:30:14 -07002201 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
Sridhar Samudrala26af65c2008-06-04 15:19:35 -07002202
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002203 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204}
2205
2206/* WARNING: This routine must only be called when we have already sent
2207 * a SYN packet that crossed the incoming SYN that caused this routine
2208 * to get called. If this assumption fails then the initial rcv_wnd
2209 * and rcv_wscale values will not be correct.
2210 */
2211int tcp_send_synack(struct sock *sk)
2212{
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002213 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214
David S. Millerfe067e82007-03-07 12:12:44 -08002215 skb = tcp_write_queue_head(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002216 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2218 return -EFAULT;
2219 }
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002220 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 if (skb_cloned(skb)) {
2222 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2223 if (nskb == NULL)
2224 return -ENOMEM;
David S. Millerfe067e82007-03-07 12:12:44 -08002225 tcp_unlink_write_queue(skb, sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 skb_header_release(nskb);
David S. Millerfe067e82007-03-07 12:12:44 -08002227 __tcp_add_write_queue_head(sk, nskb);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002228 sk_wmem_free_skb(sk, skb);
2229 sk->sk_wmem_queued += nskb->truesize;
2230 sk_mem_charge(sk, nskb->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 skb = nskb;
2232 }
2233
2234 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2235 TCP_ECN_send_synack(tcp_sk(sk), skb);
2236 }
2237 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002238 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
2240
2241/*
2242 * Prepare a SYN-ACK.
2243 */
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002244struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2245 struct request_sock *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246{
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002247 struct inet_request_sock *ireq = inet_rsk(req);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 struct tcp_sock *tp = tcp_sk(sk);
2249 struct tcphdr *th;
2250 int tcp_header_size;
Adam Langley33ad7982008-07-19 00:04:31 -07002251 struct tcp_out_options opts;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 struct sk_buff *skb;
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002253 struct tcp_md5sig_key *md5;
2254 __u8 *md5_hash_location;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002255 int mss;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256
2257 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2258 if (skb == NULL)
2259 return NULL;
2260
2261 /* Reserve space for headers. */
2262 skb_reserve(skb, MAX_TCP_HEADER);
2263
2264 skb->dst = dst_clone(dst);
2265
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002266 mss = dst_metric(dst, RTAX_ADVMSS);
2267 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2268 mss = tp->rx_opt.user_mss;
2269
Adam Langley33ad7982008-07-19 00:04:31 -07002270 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2271 __u8 rcv_wscale;
2272 /* Set this up on the first call only */
2273 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2274 /* tcp_full_space because it is guaranteed to be the first packet */
2275 tcp_select_initial_window(tcp_full_space(sk),
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002276 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
Adam Langley33ad7982008-07-19 00:04:31 -07002277 &req->rcv_wnd,
2278 &req->window_clamp,
2279 ireq->wscale_ok,
2280 &rcv_wscale);
2281 ireq->rcv_wscale = rcv_wscale;
2282 }
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002283
Adam Langley33ad7982008-07-19 00:04:31 -07002284 memset(&opts, 0, sizeof(opts));
Florian Westphal8b5f12d2008-10-26 23:10:12 -07002285#ifdef CONFIG_SYN_COOKIES
2286 if (unlikely(req->cookie_ts))
2287 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2288 else
2289#endif
Adam Langley33ad7982008-07-19 00:04:31 -07002290 TCP_SKB_CB(skb)->when = tcp_time_stamp;
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002291 tcp_header_size = tcp_synack_options(sk, req, mss,
Adam Langley33ad7982008-07-19 00:04:31 -07002292 skb, &opts, &md5) +
2293 sizeof(struct tcphdr);
2294
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002295 skb_push(skb, tcp_header_size);
2296 skb_reset_transport_header(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297
Arnaldo Carvalho de Meloaa8223c2007-04-10 21:04:22 -07002298 th = tcp_hdr(skb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 memset(th, 0, sizeof(struct tcphdr));
2300 th->syn = 1;
2301 th->ack = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 TCP_ECN_make_synack(req, th);
KOVACS Krisztiana3116ac52008-10-01 07:46:49 -07002303 th->source = ireq->loc_port;
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002304 th->dest = ireq->rmt_port;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002305 /* Setting of flags are superfluous here for callers (and ECE is
2306 * not even correctly set)
2307 */
2308 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2309 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 th->seq = htonl(TCP_SKB_CB(skb)->seq);
Arnaldo Carvalho de Melo2e6599c2005-06-18 22:46:52 -07002311 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312
2313 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
Ilpo Järvinen600ff0c2007-02-13 12:42:11 -08002314 th->window = htons(min(req->rcv_wnd, 65535U));
Adam Langley33ad7982008-07-19 00:04:31 -07002315 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 th->doff = (tcp_header_size >> 2);
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002317 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002318
2319#ifdef CONFIG_TCP_MD5SIG
2320 /* Okay, we have all we need - do the md5 hash if needed */
2321 if (md5) {
2322 tp->af_specific->calc_md5_hash(md5_hash_location,
Adam Langley49a72df2008-07-19 00:01:42 -07002323 md5, NULL, req, skb);
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002324 }
2325#endif
2326
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 return skb;
2328}
2329
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002330/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 * Do all connect socket setups that can be done AF independent.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002332 */
Stephen Hemminger40efc6f2006-01-03 16:03:49 -08002333static void tcp_connect_init(struct sock *sk)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334{
2335 struct dst_entry *dst = __sk_dst_get(sk);
2336 struct tcp_sock *tp = tcp_sk(sk);
2337 __u8 rcv_wscale;
2338
2339 /* We'll fix this up when we get a response from the other end.
2340 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2341 */
2342 tp->tcp_header_len = sizeof(struct tcphdr) +
2343 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2344
YOSHIFUJI Hideakicfb6eeb2006-11-14 19:07:45 -08002345#ifdef CONFIG_TCP_MD5SIG
2346 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2347 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2348#endif
2349
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 /* If user gave his TCP_MAXSEG, record it to clamp */
2351 if (tp->rx_opt.user_mss)
2352 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2353 tp->max_window = 0;
John Heffner5d424d52006-03-20 17:53:41 -08002354 tcp_mtup_init(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 tcp_sync_mss(sk, dst_mtu(dst));
2356
2357 if (!tp->window_clamp)
2358 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2359 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
Tom Quetchenbachf5fff5d2008-09-21 00:21:51 -07002360 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2361 tp->advmss = tp->rx_opt.user_mss;
2362
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 tcp_initialize_rcv_mss(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
2365 tcp_select_initial_window(tcp_full_space(sk),
2366 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2367 &tp->rcv_wnd,
2368 &tp->window_clamp,
2369 sysctl_tcp_window_scaling,
2370 &rcv_wscale);
2371
2372 tp->rx_opt.rcv_wscale = rcv_wscale;
2373 tp->rcv_ssthresh = tp->rcv_wnd;
2374
2375 sk->sk_err = 0;
2376 sock_reset_flag(sk, SOCK_DONE);
2377 tp->snd_wnd = 0;
2378 tcp_init_wl(tp, tp->write_seq, 0);
2379 tp->snd_una = tp->write_seq;
2380 tp->snd_sml = tp->write_seq;
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002381 tp->snd_up = tp->write_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 tp->rcv_nxt = 0;
2383 tp->rcv_wup = 0;
2384 tp->copied_seq = 0;
2385
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002386 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2387 inet_csk(sk)->icsk_retransmits = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388 tcp_clear_retrans(tp);
2389}
2390
2391/*
2392 * Build a SYN and send it off.
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002393 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394int tcp_connect(struct sock *sk)
2395{
2396 struct tcp_sock *tp = tcp_sk(sk);
2397 struct sk_buff *buff;
2398
2399 tcp_connect_init(sk);
2400
David S. Millerd179cd12005-08-17 14:57:30 -07002401 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 if (unlikely(buff == NULL))
2403 return -ENOBUFS;
2404
2405 /* Reserve space for headers. */
2406 skb_reserve(buff, MAX_TCP_HEADER);
2407
Wei Yongjunbd37a082006-08-07 21:04:15 -07002408 tp->snd_nxt = tp->write_seq;
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002409 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2410 TCP_ECN_send_syn(sk, buff);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
2412 /* Send it off. */
2413 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2414 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2415 skb_header_release(buff);
David S. Millerfe067e82007-03-07 12:12:44 -08002416 __tcp_add_write_queue_tail(sk, buff);
Hideo Aoki3ab224b2007-12-31 00:11:19 -08002417 sk->sk_wmem_queued += buff->truesize;
2418 sk_mem_charge(sk, buff->truesize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 tp->packets_out += tcp_skb_pcount(buff);
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002420 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
Wei Yongjunbd37a082006-08-07 21:04:15 -07002421
2422 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2423 * in order to make this packet get counted in tcpOutSegs.
2424 */
2425 tp->snd_nxt = tp->write_seq;
2426 tp->pushed_seq = tp->write_seq;
Pavel Emelyanov81cc8a72008-07-16 20:22:04 -07002427 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428
2429 /* Timer for repeating the SYN until an answer. */
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002430 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2431 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 return 0;
2433}
2434
2435/* Send out a delayed ack, the caller does the policy checking
2436 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2437 * for details.
2438 */
2439void tcp_send_delayed_ack(struct sock *sk)
2440{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002441 struct inet_connection_sock *icsk = inet_csk(sk);
2442 int ato = icsk->icsk_ack.ato;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 unsigned long timeout;
2444
2445 if (ato > TCP_DELACK_MIN) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002446 const struct tcp_sock *tp = tcp_sk(sk);
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002447 int max_ato = HZ / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002449 if (icsk->icsk_ack.pingpong ||
2450 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 max_ato = TCP_DELACK_MAX;
2452
2453 /* Slow path, intersegment interval is "high". */
2454
2455 /* If some rtt estimate is known, use it to bound delayed ack.
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002456 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 * directly.
2458 */
2459 if (tp->srtt) {
Ilpo Järvinen056834d2007-12-31 14:57:14 -08002460 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461
2462 if (rtt < max_ato)
2463 max_ato = rtt;
2464 }
2465
2466 ato = min(ato, max_ato);
2467 }
2468
2469 /* Stay within the limit we were given */
2470 timeout = jiffies + ato;
2471
2472 /* Use new timeout only if there wasn't a older one earlier. */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002473 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 /* If delack timer was blocked or is about to expire,
2475 * send ACK now.
2476 */
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002477 if (icsk->icsk_ack.blocked ||
2478 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 tcp_send_ack(sk);
2480 return;
2481 }
2482
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002483 if (!time_before(timeout, icsk->icsk_ack.timeout))
2484 timeout = icsk->icsk_ack.timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 }
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002486 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2487 icsk->icsk_ack.timeout = timeout;
2488 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489}
2490
2491/* This routine sends an ack and also updates the window. */
2492void tcp_send_ack(struct sock *sk)
2493{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002494 struct sk_buff *buff;
2495
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 /* If we have been reset, we may not send again. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002497 if (sk->sk_state == TCP_CLOSE)
2498 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002500 /* We are not putting this on the write queue, so
2501 * tcp_transmit_skb() will set the ownership to this
2502 * sock.
2503 */
2504 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2505 if (buff == NULL) {
2506 inet_csk_schedule_ack(sk);
2507 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2508 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2509 TCP_DELACK_MAX, TCP_RTO_MAX);
2510 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 }
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002512
2513 /* Reserve space for headers and prepare control bits. */
2514 skb_reserve(buff, MAX_TCP_HEADER);
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002515 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002516
2517 /* Send it off, this clears delayed acks for us. */
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002518 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2519 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520}
2521
2522/* This routine sends a packet with an out of date sequence
2523 * number. It assumes the other end will try to ack it.
2524 *
2525 * Question: what should we make while urgent mode?
2526 * 4.4BSD forces sending single byte of data. We cannot send
2527 * out of window data, because we have SND.NXT==SND.MAX...
2528 *
2529 * Current solution: to send TWO zero-length segments in urgent mode:
2530 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2531 * out-of-date with SND.UNA-1 to probe window.
2532 */
2533static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2534{
2535 struct tcp_sock *tp = tcp_sk(sk);
2536 struct sk_buff *skb;
2537
2538 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2539 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002540 if (skb == NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 return -1;
2542
2543 /* Reserve space for headers and set control bits. */
2544 skb_reserve(skb, MAX_TCP_HEADER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 /* Use a previous sequence. This should cause the other
2546 * end to send an ack. Don't queue or clone SKB, just
2547 * send it.
2548 */
Ilpo Järvinene870a8e2008-01-03 20:39:01 -08002549 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 TCP_SKB_CB(skb)->when = tcp_time_stamp;
David S. Millerdfb4b9d2005-12-06 16:24:52 -08002551 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552}
2553
2554int tcp_write_wakeup(struct sock *sk)
2555{
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002556 struct tcp_sock *tp = tcp_sk(sk);
2557 struct sk_buff *skb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002558
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002559 if (sk->sk_state == TCP_CLOSE)
2560 return -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002562 if ((skb = tcp_send_head(sk)) != NULL &&
2563 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2564 int err;
2565 unsigned int mss = tcp_current_mss(sk, 0);
2566 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002568 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2569 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002571 /* We are probing the opening of a window
2572 * but the window size is != 0
2573 * must have been a result SWS avoidance ( sender )
2574 */
2575 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2576 skb->len > mss) {
2577 seg_size = min(seg_size, mss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002579 if (tcp_fragment(sk, skb, seg_size, mss))
2580 return -1;
2581 } else if (!tcp_skb_pcount(skb))
2582 tcp_set_skb_tso_segs(sk, skb, mss);
2583
2584 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2585 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2586 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2587 if (!err)
2588 tcp_event_new_data_sent(sk, skb);
2589 return err;
2590 } else {
Ilpo Järvinen33f5f572008-10-07 14:43:06 -07002591 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
Ilpo Järvinen058dc332007-12-31 04:51:11 -08002592 tcp_xmit_probe_skb(sk, 1);
2593 return tcp_xmit_probe_skb(sk, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595}
2596
2597/* A window probe timeout has occurred. If window is not closed send
2598 * a partial packet else a zero probe.
2599 */
2600void tcp_send_probe0(struct sock *sk)
2601{
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002602 struct inet_connection_sock *icsk = inet_csk(sk);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 struct tcp_sock *tp = tcp_sk(sk);
2604 int err;
2605
2606 err = tcp_write_wakeup(sk);
2607
David S. Millerfe067e82007-03-07 12:12:44 -08002608 if (tp->packets_out || !tcp_send_head(sk)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609 /* Cancel probe timer, if it is not required. */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002610 icsk->icsk_probes_out = 0;
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002611 icsk->icsk_backoff = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 return;
2613 }
2614
2615 if (err <= 0) {
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002616 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2617 icsk->icsk_backoff++;
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002618 icsk->icsk_probes_out++;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002619 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002620 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2621 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 } else {
2623 /* If packet was not sent due to local congestion,
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002624 * do not backoff and do not remember icsk_probes_out.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 * Let local senders to fight for local resources.
2626 *
2627 * Use accumulated backoff yet.
2628 */
Arnaldo Carvalho de Melo6687e982005-08-10 04:03:31 -03002629 if (!icsk->icsk_probes_out)
2630 icsk->icsk_probes_out = 1;
YOSHIFUJI Hideakie905a9e2007-02-09 23:24:47 +09002631 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
Arnaldo Carvalho de Melo463c84b2005-08-09 20:10:42 -07002632 min(icsk->icsk_rto << icsk->icsk_backoff,
Arnaldo Carvalho de Melo3f421ba2005-08-09 20:11:08 -07002633 TCP_RESOURCE_PROBE_INTERVAL),
2634 TCP_RTO_MAX);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 }
2636}
2637
Glenn Griffinc6aefaf2008-02-07 21:49:26 -08002638EXPORT_SYMBOL(tcp_select_initial_window);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639EXPORT_SYMBOL(tcp_connect);
2640EXPORT_SYMBOL(tcp_make_synack);
2641EXPORT_SYMBOL(tcp_simple_retransmit);
2642EXPORT_SYMBOL(tcp_sync_mss);
John Heffner5d424d52006-03-20 17:53:41 -08002643EXPORT_SYMBOL(tcp_mtup_init);