Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
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 | * | |
8 | * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ | |
9 | * | |
02c30a84 | 10 | * Authors: Ross Biro |
1da177e4 LT |
11 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
12 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
13 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
14 | * Florian La Roche, <flla@stud.uni-sb.de> | |
15 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
16 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
17 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
18 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
19 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
20 | * Jorge Cwik, <jorge@laser.satlink.net> | |
21 | */ | |
22 | ||
23 | /* | |
24 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
25 | * : Fragmentation on mtu decrease | |
26 | * : Segment collapse on retransmit | |
27 | * : AF independence | |
28 | * | |
29 | * Linus Torvalds : send_delayed_ack | |
30 | * David S. Miller : Charge memory using the right skb | |
31 | * during syn/ack processing. | |
32 | * David S. Miller : Output engine completely rewritten. | |
33 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
34 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
35 | * J Hadi Salim : ECN support | |
36 | * | |
37 | */ | |
38 | ||
39 | #include <net/tcp.h> | |
40 | ||
41 | #include <linux/compiler.h> | |
42 | #include <linux/module.h> | |
1da177e4 LT |
43 | |
44 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 45 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 46 | |
15d99e02 RJ |
47 | /* People can turn this on to work with those rare, broken TCPs that |
48 | * interpret the window field as a signed quantity. | |
49 | */ | |
ab32ea5d | 50 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 51 | |
1da177e4 LT |
52 | /* This limits the percentage of the congestion window which we |
53 | * will allow a single TSO frame to consume. Building TSO frames | |
54 | * which are too large can cause TCP streams to be bursty. | |
55 | */ | |
ab32ea5d | 56 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 57 | |
ab32ea5d BH |
58 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
59 | int sysctl_tcp_base_mss __read_mostly = 512; | |
5d424d5a | 60 | |
35089bb2 | 61 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 62 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 63 | |
6ff03ac3 IJ |
64 | static inline void tcp_packets_out_inc(struct sock *sk, |
65 | const struct sk_buff *skb) | |
66 | { | |
67 | struct tcp_sock *tp = tcp_sk(sk); | |
68 | int orig = tp->packets_out; | |
69 | ||
70 | tp->packets_out += tcp_skb_pcount(skb); | |
71 | if (!orig) | |
72 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
73 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
74 | } | |
75 | ||
9e412ba7 | 76 | static void update_send_head(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 77 | { |
9e412ba7 IJ |
78 | struct tcp_sock *tp = tcp_sk(sk); |
79 | ||
fe067e8a | 80 | tcp_advance_send_head(sk, skb); |
1da177e4 | 81 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
9e412ba7 | 82 | tcp_packets_out_inc(sk, skb); |
1da177e4 LT |
83 | } |
84 | ||
85 | /* SND.NXT, if window was not shrunk. | |
86 | * If window has been shrunk, what should we make? It is not clear at all. | |
87 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
88 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
89 | * invalid. OK, let's make this for now: | |
90 | */ | |
9e412ba7 | 91 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 92 | { |
9e412ba7 IJ |
93 | struct tcp_sock *tp = tcp_sk(sk); |
94 | ||
1da177e4 LT |
95 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) |
96 | return tp->snd_nxt; | |
97 | else | |
98 | return tp->snd_una+tp->snd_wnd; | |
99 | } | |
100 | ||
101 | /* Calculate mss to advertise in SYN segment. | |
102 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
103 | * | |
104 | * 1. It is independent of path mtu. | |
105 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
106 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
107 | * attached devices, because some buggy hosts are confused by | |
108 | * large MSS. | |
109 | * 4. We do not make 3, we advertise MSS, calculated from first | |
110 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
111 | * This may be overridden via information stored in routing table. | |
112 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
113 | * probably even Jumbo". | |
114 | */ | |
115 | static __u16 tcp_advertise_mss(struct sock *sk) | |
116 | { | |
117 | struct tcp_sock *tp = tcp_sk(sk); | |
118 | struct dst_entry *dst = __sk_dst_get(sk); | |
119 | int mss = tp->advmss; | |
120 | ||
121 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
122 | mss = dst_metric(dst, RTAX_ADVMSS); | |
123 | tp->advmss = mss; | |
124 | } | |
125 | ||
126 | return (__u16)mss; | |
127 | } | |
128 | ||
129 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
130 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 131 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 132 | { |
463c84b9 | 133 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
134 | s32 delta = tcp_time_stamp - tp->lsndtime; |
135 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
136 | u32 cwnd = tp->snd_cwnd; | |
137 | ||
6687e988 | 138 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 139 | |
6687e988 | 140 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
141 | restart_cwnd = min(restart_cwnd, cwnd); |
142 | ||
463c84b9 | 143 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
144 | cwnd >>= 1; |
145 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
146 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
147 | tp->snd_cwnd_used = 0; | |
148 | } | |
149 | ||
40efc6fa SH |
150 | static void tcp_event_data_sent(struct tcp_sock *tp, |
151 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 152 | { |
463c84b9 ACM |
153 | struct inet_connection_sock *icsk = inet_csk(sk); |
154 | const u32 now = tcp_time_stamp; | |
1da177e4 | 155 | |
35089bb2 DM |
156 | if (sysctl_tcp_slow_start_after_idle && |
157 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 158 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
159 | |
160 | tp->lsndtime = now; | |
161 | ||
162 | /* If it is a reply for ato after last received | |
163 | * packet, enter pingpong mode. | |
164 | */ | |
463c84b9 ACM |
165 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
166 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
167 | } |
168 | ||
40efc6fa | 169 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 170 | { |
463c84b9 ACM |
171 | tcp_dec_quickack_mode(sk, pkts); |
172 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
173 | } |
174 | ||
175 | /* Determine a window scaling and initial window to offer. | |
176 | * Based on the assumption that the given amount of space | |
177 | * will be offered. Store the results in the tp structure. | |
178 | * NOTE: for smooth operation initial space offering should | |
179 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
180 | * This MUST be enforced by all callers. | |
181 | */ | |
182 | void tcp_select_initial_window(int __space, __u32 mss, | |
183 | __u32 *rcv_wnd, __u32 *window_clamp, | |
184 | int wscale_ok, __u8 *rcv_wscale) | |
185 | { | |
186 | unsigned int space = (__space < 0 ? 0 : __space); | |
187 | ||
188 | /* If no clamp set the clamp to the max possible scaled window */ | |
189 | if (*window_clamp == 0) | |
190 | (*window_clamp) = (65535 << 14); | |
191 | space = min(*window_clamp, space); | |
192 | ||
193 | /* Quantize space offering to a multiple of mss if possible. */ | |
194 | if (space > mss) | |
195 | space = (space / mss) * mss; | |
196 | ||
197 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
198 | * will break some buggy TCP stacks. If the admin tells us |
199 | * it is likely we could be speaking with such a buggy stack | |
200 | * we will truncate our initial window offering to 32K-1 | |
201 | * unless the remote has sent us a window scaling option, | |
202 | * which we interpret as a sign the remote TCP is not | |
203 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 204 | */ |
15d99e02 RJ |
205 | if (sysctl_tcp_workaround_signed_windows) |
206 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
207 | else | |
208 | (*rcv_wnd) = space; | |
209 | ||
1da177e4 LT |
210 | (*rcv_wscale) = 0; |
211 | if (wscale_ok) { | |
212 | /* Set window scaling on max possible window | |
e905a9ed | 213 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
214 | */ |
215 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 216 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
217 | while (space > 65535 && (*rcv_wscale) < 14) { |
218 | space >>= 1; | |
219 | (*rcv_wscale)++; | |
220 | } | |
221 | } | |
222 | ||
223 | /* Set initial window to value enough for senders, | |
6b251858 | 224 | * following RFC2414. Senders, not following this RFC, |
1da177e4 LT |
225 | * will be satisfied with 2. |
226 | */ | |
227 | if (mss > (1<<*rcv_wscale)) { | |
01ff367e DM |
228 | int init_cwnd = 4; |
229 | if (mss > 1460*3) | |
1da177e4 | 230 | init_cwnd = 2; |
01ff367e DM |
231 | else if (mss > 1460) |
232 | init_cwnd = 3; | |
1da177e4 LT |
233 | if (*rcv_wnd > init_cwnd*mss) |
234 | *rcv_wnd = init_cwnd*mss; | |
235 | } | |
236 | ||
237 | /* Set the clamp no higher than max representable value */ | |
238 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
239 | } | |
240 | ||
241 | /* Chose a new window to advertise, update state in tcp_sock for the | |
242 | * socket, and return result with RFC1323 scaling applied. The return | |
243 | * value can be stuffed directly into th->window for an outgoing | |
244 | * frame. | |
245 | */ | |
40efc6fa | 246 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
247 | { |
248 | struct tcp_sock *tp = tcp_sk(sk); | |
249 | u32 cur_win = tcp_receive_window(tp); | |
250 | u32 new_win = __tcp_select_window(sk); | |
251 | ||
252 | /* Never shrink the offered window */ | |
2de979bd | 253 | if (new_win < cur_win) { |
1da177e4 LT |
254 | /* Danger Will Robinson! |
255 | * Don't update rcv_wup/rcv_wnd here or else | |
256 | * we will not be able to advertise a zero | |
257 | * window in time. --DaveM | |
258 | * | |
259 | * Relax Will Robinson. | |
260 | */ | |
261 | new_win = cur_win; | |
262 | } | |
263 | tp->rcv_wnd = new_win; | |
264 | tp->rcv_wup = tp->rcv_nxt; | |
265 | ||
266 | /* Make sure we do not exceed the maximum possible | |
267 | * scaled window. | |
268 | */ | |
15d99e02 | 269 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
270 | new_win = min(new_win, MAX_TCP_WINDOW); |
271 | else | |
272 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
273 | ||
274 | /* RFC1323 scaling applied */ | |
275 | new_win >>= tp->rx_opt.rcv_wscale; | |
276 | ||
277 | /* If we advertise zero window, disable fast path. */ | |
278 | if (new_win == 0) | |
279 | tp->pred_flags = 0; | |
280 | ||
281 | return new_win; | |
282 | } | |
283 | ||
bdf1ee5d IJ |
284 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, |
285 | struct sk_buff *skb) | |
286 | { | |
287 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; | |
288 | if (!(tp->ecn_flags&TCP_ECN_OK)) | |
289 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; | |
290 | } | |
291 | ||
292 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) | |
293 | { | |
294 | struct tcp_sock *tp = tcp_sk(sk); | |
295 | ||
296 | tp->ecn_flags = 0; | |
297 | if (sysctl_tcp_ecn) { | |
298 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR; | |
299 | tp->ecn_flags = TCP_ECN_OK; | |
300 | } | |
301 | } | |
302 | ||
303 | static __inline__ void | |
304 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
305 | { | |
306 | if (inet_rsk(req)->ecn_ok) | |
307 | th->ece = 1; | |
308 | } | |
309 | ||
310 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, | |
311 | int tcp_header_len) | |
312 | { | |
313 | struct tcp_sock *tp = tcp_sk(sk); | |
314 | ||
315 | if (tp->ecn_flags & TCP_ECN_OK) { | |
316 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
317 | if (skb->len != tcp_header_len && | |
318 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
319 | INET_ECN_xmit(sk); | |
320 | if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) { | |
321 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; | |
322 | tcp_hdr(skb)->cwr = 1; | |
323 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
324 | } | |
325 | } else { | |
326 | /* ACK or retransmitted segment: clear ECT|CE */ | |
327 | INET_ECN_dontxmit(sk); | |
328 | } | |
329 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
330 | tcp_hdr(skb)->ece = 1; | |
331 | } | |
332 | } | |
333 | ||
df7a3b07 | 334 | static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, |
cfb6eeb4 | 335 | __u32 tstamp, __u8 **md5_hash) |
40efc6fa SH |
336 | { |
337 | if (tp->rx_opt.tstamp_ok) { | |
496c98df YH |
338 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
339 | (TCPOPT_NOP << 16) | | |
340 | (TCPOPT_TIMESTAMP << 8) | | |
341 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
342 | *ptr++ = htonl(tstamp); |
343 | *ptr++ = htonl(tp->rx_opt.ts_recent); | |
344 | } | |
345 | if (tp->rx_opt.eff_sacks) { | |
346 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; | |
347 | int this_sack; | |
348 | ||
349 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
350 | (TCPOPT_NOP << 16) | | |
351 | (TCPOPT_SACK << 8) | | |
352 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * | |
353 | TCPOLEN_SACK_PERBLOCK))); | |
2de979bd SH |
354 | |
355 | for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { | |
40efc6fa SH |
356 | *ptr++ = htonl(sp[this_sack].start_seq); |
357 | *ptr++ = htonl(sp[this_sack].end_seq); | |
358 | } | |
2de979bd | 359 | |
40efc6fa SH |
360 | if (tp->rx_opt.dsack) { |
361 | tp->rx_opt.dsack = 0; | |
362 | tp->rx_opt.eff_sacks--; | |
363 | } | |
364 | } | |
cfb6eeb4 YH |
365 | #ifdef CONFIG_TCP_MD5SIG |
366 | if (md5_hash) { | |
367 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
368 | (TCPOPT_NOP << 16) | | |
369 | (TCPOPT_MD5SIG << 8) | | |
370 | TCPOLEN_MD5SIG); | |
371 | *md5_hash = (__u8 *)ptr; | |
372 | } | |
373 | #endif | |
40efc6fa SH |
374 | } |
375 | ||
376 | /* Construct a tcp options header for a SYN or SYN_ACK packet. | |
377 | * If this is every changed make sure to change the definition of | |
378 | * MAX_SYN_SIZE to match the new maximum number of options that you | |
379 | * can generate. | |
cfb6eeb4 YH |
380 | * |
381 | * Note - that with the RFC2385 TCP option, we make room for the | |
382 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the | |
383 | * location to be filled is passed back up. | |
40efc6fa | 384 | */ |
df7a3b07 | 385 | static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, |
40efc6fa | 386 | int offer_wscale, int wscale, __u32 tstamp, |
cfb6eeb4 | 387 | __u32 ts_recent, __u8 **md5_hash) |
40efc6fa SH |
388 | { |
389 | /* We always get an MSS option. | |
390 | * The option bytes which will be seen in normal data | |
391 | * packets should timestamps be used, must be in the MSS | |
392 | * advertised. But we subtract them from tp->mss_cache so | |
393 | * that calculations in tcp_sendmsg are simpler etc. | |
394 | * So account for this fact here if necessary. If we | |
395 | * don't do this correctly, as a receiver we won't | |
396 | * recognize data packets as being full sized when we | |
397 | * should, and thus we won't abide by the delayed ACK | |
398 | * rules correctly. | |
399 | * SACKs don't matter, we never delay an ACK when we | |
400 | * have any of those going out. | |
401 | */ | |
402 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); | |
403 | if (ts) { | |
2de979bd | 404 | if (sack) |
496c98df YH |
405 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
406 | (TCPOLEN_SACK_PERM << 16) | | |
407 | (TCPOPT_TIMESTAMP << 8) | | |
408 | TCPOLEN_TIMESTAMP); | |
40efc6fa | 409 | else |
496c98df YH |
410 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
411 | (TCPOPT_NOP << 16) | | |
412 | (TCPOPT_TIMESTAMP << 8) | | |
413 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
414 | *ptr++ = htonl(tstamp); /* TSVAL */ |
415 | *ptr++ = htonl(ts_recent); /* TSECR */ | |
2de979bd | 416 | } else if (sack) |
496c98df YH |
417 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
418 | (TCPOPT_NOP << 16) | | |
419 | (TCPOPT_SACK_PERM << 8) | | |
420 | TCPOLEN_SACK_PERM); | |
40efc6fa | 421 | if (offer_wscale) |
496c98df YH |
422 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
423 | (TCPOPT_WINDOW << 16) | | |
424 | (TCPOLEN_WINDOW << 8) | | |
425 | (wscale)); | |
cfb6eeb4 YH |
426 | #ifdef CONFIG_TCP_MD5SIG |
427 | /* | |
428 | * If MD5 is enabled, then we set the option, and include the size | |
429 | * (always 18). The actual MD5 hash is added just before the | |
430 | * packet is sent. | |
431 | */ | |
432 | if (md5_hash) { | |
433 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
434 | (TCPOPT_NOP << 16) | | |
435 | (TCPOPT_MD5SIG << 8) | | |
436 | TCPOLEN_MD5SIG); | |
437 | *md5_hash = (__u8 *) ptr; | |
438 | } | |
439 | #endif | |
40efc6fa | 440 | } |
1da177e4 LT |
441 | |
442 | /* This routine actually transmits TCP packets queued in by | |
443 | * tcp_do_sendmsg(). This is used by both the initial | |
444 | * transmission and possible later retransmissions. | |
445 | * All SKB's seen here are completely headerless. It is our | |
446 | * job to build the TCP header, and pass the packet down to | |
447 | * IP so it can do the same plus pass the packet off to the | |
448 | * device. | |
449 | * | |
450 | * We are working here with either a clone of the original | |
451 | * SKB, or a fresh unique copy made by the retransmit engine. | |
452 | */ | |
dfb4b9dc | 453 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) |
1da177e4 | 454 | { |
dfb4b9dc DM |
455 | const struct inet_connection_sock *icsk = inet_csk(sk); |
456 | struct inet_sock *inet; | |
457 | struct tcp_sock *tp; | |
458 | struct tcp_skb_cb *tcb; | |
459 | int tcp_header_size; | |
cfb6eeb4 YH |
460 | #ifdef CONFIG_TCP_MD5SIG |
461 | struct tcp_md5sig_key *md5; | |
462 | __u8 *md5_hash_location; | |
463 | #endif | |
dfb4b9dc DM |
464 | struct tcphdr *th; |
465 | int sysctl_flags; | |
466 | int err; | |
467 | ||
468 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
469 | ||
470 | /* If congestion control is doing timestamping, we must | |
471 | * take such a timestamp before we potentially clone/copy. | |
472 | */ | |
164891aa | 473 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
474 | __net_timestamp(skb); |
475 | ||
476 | if (likely(clone_it)) { | |
477 | if (unlikely(skb_cloned(skb))) | |
478 | skb = pskb_copy(skb, gfp_mask); | |
479 | else | |
480 | skb = skb_clone(skb, gfp_mask); | |
481 | if (unlikely(!skb)) | |
482 | return -ENOBUFS; | |
483 | } | |
1da177e4 | 484 | |
dfb4b9dc DM |
485 | inet = inet_sk(sk); |
486 | tp = tcp_sk(sk); | |
487 | tcb = TCP_SKB_CB(skb); | |
488 | tcp_header_size = tp->tcp_header_len; | |
1da177e4 LT |
489 | |
490 | #define SYSCTL_FLAG_TSTAMPS 0x1 | |
491 | #define SYSCTL_FLAG_WSCALE 0x2 | |
492 | #define SYSCTL_FLAG_SACK 0x4 | |
493 | ||
dfb4b9dc DM |
494 | sysctl_flags = 0; |
495 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
496 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; | |
2de979bd | 497 | if (sysctl_tcp_timestamps) { |
dfb4b9dc DM |
498 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
499 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; | |
1da177e4 | 500 | } |
dfb4b9dc DM |
501 | if (sysctl_tcp_window_scaling) { |
502 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; | |
503 | sysctl_flags |= SYSCTL_FLAG_WSCALE; | |
1da177e4 | 504 | } |
dfb4b9dc DM |
505 | if (sysctl_tcp_sack) { |
506 | sysctl_flags |= SYSCTL_FLAG_SACK; | |
507 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) | |
508 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; | |
1da177e4 | 509 | } |
dfb4b9dc DM |
510 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
511 | /* A SACK is 2 pad bytes, a 2 byte header, plus | |
512 | * 2 32-bit sequence numbers for each SACK block. | |
513 | */ | |
514 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + | |
515 | (tp->rx_opt.eff_sacks * | |
516 | TCPOLEN_SACK_PERBLOCK)); | |
517 | } | |
e905a9ed | 518 | |
dfb4b9dc DM |
519 | if (tcp_packets_in_flight(tp) == 0) |
520 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
521 | ||
cfb6eeb4 YH |
522 | #ifdef CONFIG_TCP_MD5SIG |
523 | /* | |
524 | * Are we doing MD5 on this segment? If so - make | |
525 | * room for it. | |
526 | */ | |
527 | md5 = tp->af_specific->md5_lookup(sk, sk); | |
528 | if (md5) | |
529 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
530 | #endif | |
531 | ||
aa8223c7 ACM |
532 | skb_push(skb, tcp_header_size); |
533 | skb_reset_transport_header(skb); | |
e89862f4 | 534 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
535 | |
536 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 537 | th = tcp_hdr(skb); |
dfb4b9dc DM |
538 | th->source = inet->sport; |
539 | th->dest = inet->dport; | |
540 | th->seq = htonl(tcb->seq); | |
541 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 542 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
543 | tcb->flags); |
544 | ||
545 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
546 | /* RFC1323: The window in SYN & SYN/ACK segments | |
547 | * is never scaled. | |
548 | */ | |
600ff0c2 | 549 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
550 | } else { |
551 | th->window = htons(tcp_select_window(sk)); | |
552 | } | |
553 | th->check = 0; | |
554 | th->urg_ptr = 0; | |
1da177e4 | 555 | |
dfb4b9dc DM |
556 | if (unlikely(tp->urg_mode && |
557 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { | |
558 | th->urg_ptr = htons(tp->snd_up-tcb->seq); | |
559 | th->urg = 1; | |
560 | } | |
1da177e4 | 561 | |
dfb4b9dc | 562 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
df7a3b07 | 563 | tcp_syn_build_options((__be32 *)(th + 1), |
dfb4b9dc DM |
564 | tcp_advertise_mss(sk), |
565 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), | |
566 | (sysctl_flags & SYSCTL_FLAG_SACK), | |
567 | (sysctl_flags & SYSCTL_FLAG_WSCALE), | |
568 | tp->rx_opt.rcv_wscale, | |
569 | tcb->when, | |
cfb6eeb4 YH |
570 | tp->rx_opt.ts_recent, |
571 | ||
572 | #ifdef CONFIG_TCP_MD5SIG | |
573 | md5 ? &md5_hash_location : | |
574 | #endif | |
575 | NULL); | |
dfb4b9dc | 576 | } else { |
df7a3b07 | 577 | tcp_build_and_update_options((__be32 *)(th + 1), |
cfb6eeb4 YH |
578 | tp, tcb->when, |
579 | #ifdef CONFIG_TCP_MD5SIG | |
580 | md5 ? &md5_hash_location : | |
581 | #endif | |
582 | NULL); | |
9e412ba7 | 583 | TCP_ECN_send(sk, skb, tcp_header_size); |
dfb4b9dc | 584 | } |
1da177e4 | 585 | |
cfb6eeb4 YH |
586 | #ifdef CONFIG_TCP_MD5SIG |
587 | /* Calculate the MD5 hash, as we have all we need now */ | |
588 | if (md5) { | |
589 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
590 | md5, | |
591 | sk, NULL, NULL, | |
aa8223c7 | 592 | tcp_hdr(skb), |
cfb6eeb4 YH |
593 | sk->sk_protocol, |
594 | skb->len); | |
595 | } | |
596 | #endif | |
597 | ||
8292a17a | 598 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
1da177e4 | 599 | |
dfb4b9dc DM |
600 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
601 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1da177e4 | 602 | |
dfb4b9dc DM |
603 | if (skb->len != tcp_header_size) |
604 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 605 | |
bd37a088 WY |
606 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
607 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
1da177e4 | 608 | |
e89862f4 | 609 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
83de47cd | 610 | if (likely(err <= 0)) |
dfb4b9dc DM |
611 | return err; |
612 | ||
3cfe3baa | 613 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 614 | |
b9df3cb8 | 615 | return net_xmit_eval(err); |
1da177e4 | 616 | |
1da177e4 LT |
617 | #undef SYSCTL_FLAG_TSTAMPS |
618 | #undef SYSCTL_FLAG_WSCALE | |
619 | #undef SYSCTL_FLAG_SACK | |
620 | } | |
621 | ||
622 | ||
e905a9ed | 623 | /* This routine just queue's the buffer |
1da177e4 LT |
624 | * |
625 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
626 | * otherwise socket can stall. | |
627 | */ | |
628 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
629 | { | |
630 | struct tcp_sock *tp = tcp_sk(sk); | |
631 | ||
632 | /* Advance write_seq and place onto the write_queue. */ | |
633 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
634 | skb_header_release(skb); | |
fe067e8a | 635 | tcp_add_write_queue_tail(sk, skb); |
1da177e4 | 636 | sk_charge_skb(sk, skb); |
1da177e4 LT |
637 | } |
638 | ||
846998ae | 639 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 640 | { |
bcd76111 | 641 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
f6302d1d DM |
642 | /* Avoid the costly divide in the normal |
643 | * non-TSO case. | |
644 | */ | |
7967168c HX |
645 | skb_shinfo(skb)->gso_segs = 1; |
646 | skb_shinfo(skb)->gso_size = 0; | |
647 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 648 | } else { |
356f89e1 | 649 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 650 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 651 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
652 | } |
653 | } | |
654 | ||
1da177e4 LT |
655 | /* Function to create two new TCP segments. Shrinks the given segment |
656 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 657 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
658 | * Remember, these are still headerless SKBs at this point. |
659 | */ | |
6475be16 | 660 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) |
1da177e4 LT |
661 | { |
662 | struct tcp_sock *tp = tcp_sk(sk); | |
663 | struct sk_buff *buff; | |
6475be16 | 664 | int nsize, old_factor; |
b60b49ea | 665 | int nlen; |
1da177e4 LT |
666 | u16 flags; |
667 | ||
b2cc99f0 | 668 | BUG_ON(len > skb->len); |
6a438bbe | 669 | |
e905a9ed | 670 | clear_all_retrans_hints(tp); |
1da177e4 LT |
671 | nsize = skb_headlen(skb) - len; |
672 | if (nsize < 0) | |
673 | nsize = 0; | |
674 | ||
675 | if (skb_cloned(skb) && | |
676 | skb_is_nonlinear(skb) && | |
677 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
678 | return -ENOMEM; | |
679 | ||
680 | /* Get a new skb... force flag on. */ | |
681 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
682 | if (buff == NULL) | |
683 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 684 | |
b60b49ea HX |
685 | sk_charge_skb(sk, buff); |
686 | nlen = skb->len - len - nsize; | |
687 | buff->truesize += nlen; | |
688 | skb->truesize -= nlen; | |
1da177e4 LT |
689 | |
690 | /* Correct the sequence numbers. */ | |
691 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
692 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
693 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
694 | ||
0dde7b54 IJ |
695 | if (tp->sacked_out && (TCP_SKB_CB(skb)->seq == tp->highest_sack)) |
696 | tp->highest_sack = TCP_SKB_CB(buff)->seq; | |
697 | ||
1da177e4 LT |
698 | /* PSH and FIN should only be set in the second packet. */ |
699 | flags = TCP_SKB_CB(skb)->flags; | |
700 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
701 | TCP_SKB_CB(buff)->flags = flags; | |
e14c3caf | 702 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 LT |
703 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; |
704 | ||
84fa7933 | 705 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 LT |
706 | /* Copy and checksum data tail into the new buffer. */ |
707 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), | |
708 | nsize, 0); | |
709 | ||
710 | skb_trim(skb, len); | |
711 | ||
712 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
713 | } else { | |
84fa7933 | 714 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
715 | skb_split(skb, buff, len); |
716 | } | |
717 | ||
718 | buff->ip_summed = skb->ip_summed; | |
719 | ||
720 | /* Looks stupid, but our code really uses when of | |
721 | * skbs, which it never sent before. --ANK | |
722 | */ | |
723 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 724 | buff->tstamp = skb->tstamp; |
1da177e4 | 725 | |
6475be16 DM |
726 | old_factor = tcp_skb_pcount(skb); |
727 | ||
1da177e4 | 728 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
729 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
730 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 731 | |
6475be16 DM |
732 | /* If this packet has been sent out already, we must |
733 | * adjust the various packet counters. | |
734 | */ | |
cf0b450c | 735 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
736 | int diff = old_factor - tcp_skb_pcount(skb) - |
737 | tcp_skb_pcount(buff); | |
1da177e4 | 738 | |
6475be16 | 739 | tp->packets_out -= diff; |
e14c3caf HX |
740 | |
741 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
742 | tp->sacked_out -= diff; | |
743 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
744 | tp->retrans_out -= diff; | |
745 | ||
b5860bba | 746 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
6475be16 | 747 | tp->lost_out -= diff; |
83ca28be | 748 | |
6475be16 | 749 | if (diff > 0) { |
83ca28be | 750 | /* Adjust Reno SACK estimate. */ |
e60402d0 | 751 | if (tcp_is_reno(tp)) { |
af610b4c | 752 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); |
005903bc | 753 | tcp_verify_left_out(tp); |
83ca28be HX |
754 | } |
755 | ||
af610b4c | 756 | tcp_dec_pcount_approx_int(&tp->fackets_out, diff); |
d0410051 IJ |
757 | /* SACK fastpath might overwrite it unless dealt with */ |
758 | if (tp->fastpath_skb_hint != NULL && | |
759 | after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, | |
760 | TCP_SKB_CB(skb)->seq)) { | |
af610b4c | 761 | tcp_dec_pcount_approx_int(&tp->fastpath_cnt_hint, diff); |
d0410051 | 762 | } |
6475be16 | 763 | } |
1da177e4 LT |
764 | } |
765 | ||
766 | /* Link BUFF into the send queue. */ | |
f44b5271 | 767 | skb_header_release(buff); |
fe067e8a | 768 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
769 | |
770 | return 0; | |
771 | } | |
772 | ||
773 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
774 | * eventually). The difference is that pulled data not copied, but | |
775 | * immediately discarded. | |
776 | */ | |
f2911969 | 777 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
778 | { |
779 | int i, k, eat; | |
780 | ||
781 | eat = len; | |
782 | k = 0; | |
783 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { | |
784 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
785 | put_page(skb_shinfo(skb)->frags[i].page); | |
786 | eat -= skb_shinfo(skb)->frags[i].size; | |
787 | } else { | |
788 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
789 | if (eat) { | |
790 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
791 | skb_shinfo(skb)->frags[k].size -= eat; | |
792 | eat = 0; | |
793 | } | |
794 | k++; | |
795 | } | |
796 | } | |
797 | skb_shinfo(skb)->nr_frags = k; | |
798 | ||
27a884dc | 799 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
800 | skb->data_len -= len; |
801 | skb->len = skb->data_len; | |
1da177e4 LT |
802 | } |
803 | ||
804 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
805 | { | |
806 | if (skb_cloned(skb) && | |
807 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
808 | return -ENOMEM; | |
809 | ||
f2911969 HXP |
810 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
811 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 812 | __skb_pull(skb, len); |
f2911969 HXP |
813 | else |
814 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
815 | |
816 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 817 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
818 | |
819 | skb->truesize -= len; | |
820 | sk->sk_wmem_queued -= len; | |
821 | sk->sk_forward_alloc += len; | |
822 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
823 | ||
824 | /* Any change of skb->len requires recalculation of tso | |
825 | * factor and mss. | |
826 | */ | |
827 | if (tcp_skb_pcount(skb) > 1) | |
846998ae | 828 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
1da177e4 LT |
829 | |
830 | return 0; | |
831 | } | |
832 | ||
5d424d5a JH |
833 | /* Not accounting for SACKs here. */ |
834 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
835 | { | |
836 | struct tcp_sock *tp = tcp_sk(sk); | |
837 | struct inet_connection_sock *icsk = inet_csk(sk); | |
838 | int mss_now; | |
839 | ||
840 | /* Calculate base mss without TCP options: | |
841 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
842 | */ | |
843 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
844 | ||
845 | /* Clamp it (mss_clamp does not include tcp options) */ | |
846 | if (mss_now > tp->rx_opt.mss_clamp) | |
847 | mss_now = tp->rx_opt.mss_clamp; | |
848 | ||
849 | /* Now subtract optional transport overhead */ | |
850 | mss_now -= icsk->icsk_ext_hdr_len; | |
851 | ||
852 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
853 | if (mss_now < 48) | |
854 | mss_now = 48; | |
855 | ||
856 | /* Now subtract TCP options size, not including SACKs */ | |
857 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
858 | ||
859 | return mss_now; | |
860 | } | |
861 | ||
862 | /* Inverse of above */ | |
863 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
864 | { | |
865 | struct tcp_sock *tp = tcp_sk(sk); | |
866 | struct inet_connection_sock *icsk = inet_csk(sk); | |
867 | int mtu; | |
868 | ||
869 | mtu = mss + | |
870 | tp->tcp_header_len + | |
871 | icsk->icsk_ext_hdr_len + | |
872 | icsk->icsk_af_ops->net_header_len; | |
873 | ||
874 | return mtu; | |
875 | } | |
876 | ||
877 | void tcp_mtup_init(struct sock *sk) | |
878 | { | |
879 | struct tcp_sock *tp = tcp_sk(sk); | |
880 | struct inet_connection_sock *icsk = inet_csk(sk); | |
881 | ||
882 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
883 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 884 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
885 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
886 | icsk->icsk_mtup.probe_size = 0; | |
887 | } | |
888 | ||
1da177e4 LT |
889 | /* This function synchronize snd mss to current pmtu/exthdr set. |
890 | ||
891 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
892 | for TCP options, but includes only bare TCP header. | |
893 | ||
894 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 895 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
896 | It also does not include TCP options. |
897 | ||
d83d8461 | 898 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
899 | |
900 | tp->mss_cache is current effective sending mss, including | |
901 | all tcp options except for SACKs. It is evaluated, | |
902 | taking into account current pmtu, but never exceeds | |
903 | tp->rx_opt.mss_clamp. | |
904 | ||
905 | NOTE1. rfc1122 clearly states that advertised MSS | |
906 | DOES NOT include either tcp or ip options. | |
907 | ||
d83d8461 ACM |
908 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
909 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 LT |
910 | */ |
911 | ||
912 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
913 | { | |
914 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 915 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 916 | int mss_now; |
1da177e4 | 917 | |
5d424d5a JH |
918 | if (icsk->icsk_mtup.search_high > pmtu) |
919 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 920 | |
5d424d5a | 921 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
1da177e4 LT |
922 | |
923 | /* Bound mss with half of window */ | |
924 | if (tp->max_window && mss_now > (tp->max_window>>1)) | |
925 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); | |
926 | ||
927 | /* And store cached results */ | |
d83d8461 | 928 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
929 | if (icsk->icsk_mtup.enabled) |
930 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 931 | tp->mss_cache = mss_now; |
1da177e4 LT |
932 | |
933 | return mss_now; | |
934 | } | |
935 | ||
936 | /* Compute the current effective MSS, taking SACKs and IP options, | |
937 | * and even PMTU discovery events into account. | |
938 | * | |
939 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
940 | * cannot be large. However, taking into account rare use of URG, this | |
941 | * is not a big flaw. | |
942 | */ | |
c1b4a7e6 | 943 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
1da177e4 LT |
944 | { |
945 | struct tcp_sock *tp = tcp_sk(sk); | |
946 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 DM |
947 | u32 mss_now; |
948 | u16 xmit_size_goal; | |
949 | int doing_tso = 0; | |
950 | ||
951 | mss_now = tp->mss_cache; | |
952 | ||
bcd76111 | 953 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
c1b4a7e6 | 954 | doing_tso = 1; |
1da177e4 | 955 | |
1da177e4 LT |
956 | if (dst) { |
957 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 958 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
959 | mss_now = tcp_sync_mss(sk, mtu); |
960 | } | |
961 | ||
c1b4a7e6 DM |
962 | if (tp->rx_opt.eff_sacks) |
963 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + | |
964 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
1da177e4 | 965 | |
cfb6eeb4 YH |
966 | #ifdef CONFIG_TCP_MD5SIG |
967 | if (tp->af_specific->md5_lookup(sk, sk)) | |
968 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; | |
969 | #endif | |
970 | ||
c1b4a7e6 | 971 | xmit_size_goal = mss_now; |
1da177e4 | 972 | |
c1b4a7e6 | 973 | if (doing_tso) { |
8292a17a ACM |
974 | xmit_size_goal = (65535 - |
975 | inet_csk(sk)->icsk_af_ops->net_header_len - | |
d83d8461 ACM |
976 | inet_csk(sk)->icsk_ext_hdr_len - |
977 | tp->tcp_header_len); | |
1da177e4 | 978 | |
c1b4a7e6 DM |
979 | if (tp->max_window && |
980 | (xmit_size_goal > (tp->max_window >> 1))) | |
981 | xmit_size_goal = max((tp->max_window >> 1), | |
982 | 68U - tp->tcp_header_len); | |
1da177e4 | 983 | |
c1b4a7e6 | 984 | xmit_size_goal -= (xmit_size_goal % mss_now); |
1da177e4 | 985 | } |
c1b4a7e6 | 986 | tp->xmit_size_goal = xmit_size_goal; |
1da177e4 | 987 | |
1da177e4 LT |
988 | return mss_now; |
989 | } | |
990 | ||
a762a980 DM |
991 | /* Congestion window validation. (RFC2861) */ |
992 | ||
9e412ba7 | 993 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 994 | { |
9e412ba7 | 995 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 DM |
996 | __u32 packets_out = tp->packets_out; |
997 | ||
998 | if (packets_out >= tp->snd_cwnd) { | |
999 | /* Network is feed fully. */ | |
1000 | tp->snd_cwnd_used = 0; | |
1001 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1002 | } else { | |
1003 | /* Network starves. */ | |
1004 | if (tp->packets_out > tp->snd_cwnd_used) | |
1005 | tp->snd_cwnd_used = tp->packets_out; | |
1006 | ||
15d33c07 DM |
1007 | if (sysctl_tcp_slow_start_after_idle && |
1008 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1009 | tcp_cwnd_application_limited(sk); |
1010 | } | |
1011 | } | |
1012 | ||
c1b4a7e6 DM |
1013 | static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) |
1014 | { | |
1015 | u32 window, cwnd_len; | |
1016 | ||
1017 | window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); | |
1018 | cwnd_len = mss_now * cwnd; | |
1019 | return min(window, cwnd_len); | |
1020 | } | |
1021 | ||
1022 | /* Can at least one segment of SKB be sent right now, according to the | |
1023 | * congestion window rules? If so, return how many segments are allowed. | |
1024 | */ | |
1025 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) | |
1026 | { | |
1027 | u32 in_flight, cwnd; | |
1028 | ||
1029 | /* Don't be strict about the congestion window for the final FIN. */ | |
104439a8 JH |
1030 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
1031 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1032 | return 1; |
1033 | ||
1034 | in_flight = tcp_packets_in_flight(tp); | |
1035 | cwnd = tp->snd_cwnd; | |
1036 | if (in_flight < cwnd) | |
1037 | return (cwnd - in_flight); | |
1038 | ||
1039 | return 0; | |
1040 | } | |
1041 | ||
1042 | /* This must be invoked the first time we consider transmitting | |
1043 | * SKB onto the wire. | |
1044 | */ | |
40efc6fa | 1045 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1046 | { |
1047 | int tso_segs = tcp_skb_pcount(skb); | |
1048 | ||
846998ae DM |
1049 | if (!tso_segs || |
1050 | (tso_segs > 1 && | |
7967168c | 1051 | tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1052 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1053 | tso_segs = tcp_skb_pcount(skb); |
1054 | } | |
1055 | return tso_segs; | |
1056 | } | |
1057 | ||
1058 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1059 | { | |
1060 | return after(tp->snd_sml,tp->snd_una) && | |
1061 | !after(tp->snd_sml, tp->snd_nxt); | |
1062 | } | |
1063 | ||
1064 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1065 | * 1. It is full sized. | |
1066 | * 2. Or it contains FIN. (already checked by caller) | |
1067 | * 3. Or TCP_NODELAY was set. | |
1068 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1069 | * With Minshall's modification: all sent small packets are ACKed. | |
1070 | */ | |
1071 | ||
1072 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
e905a9ed | 1073 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1074 | unsigned mss_now, int nonagle) |
1075 | { | |
1076 | return (skb->len < mss_now && | |
1077 | ((nonagle&TCP_NAGLE_CORK) || | |
1078 | (!nonagle && | |
1079 | tp->packets_out && | |
1080 | tcp_minshall_check(tp)))); | |
1081 | } | |
1082 | ||
1083 | /* Return non-zero if the Nagle test allows this packet to be | |
1084 | * sent now. | |
1085 | */ | |
1086 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1087 | unsigned int cur_mss, int nonagle) | |
1088 | { | |
1089 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1090 | * write_queue (they have no chances to get new data). | |
1091 | * | |
1092 | * This is implemented in the callers, where they modify the 'nonagle' | |
1093 | * argument based upon the location of SKB in the send queue. | |
1094 | */ | |
1095 | if (nonagle & TCP_NAGLE_PUSH) | |
1096 | return 1; | |
1097 | ||
d551e454 IJ |
1098 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1099 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1100 | */ | |
1101 | if (tp->urg_mode || (tp->frto_counter == 2) || | |
c1b4a7e6 DM |
1102 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
1103 | return 1; | |
1104 | ||
1105 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1106 | return 1; | |
1107 | ||
1108 | return 0; | |
1109 | } | |
1110 | ||
1111 | /* Does at least the first segment of SKB fit into the send window? */ | |
1112 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) | |
1113 | { | |
1114 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1115 | ||
1116 | if (skb->len > cur_mss) | |
1117 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1118 | ||
1119 | return !after(end_seq, tp->snd_una + tp->snd_wnd); | |
1120 | } | |
1121 | ||
fe067e8a | 1122 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1123 | * should be put on the wire right now. If so, it returns the number of |
1124 | * packets allowed by the congestion window. | |
1125 | */ | |
1126 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1127 | unsigned int cur_mss, int nonagle) | |
1128 | { | |
1129 | struct tcp_sock *tp = tcp_sk(sk); | |
1130 | unsigned int cwnd_quota; | |
1131 | ||
846998ae | 1132 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1133 | |
1134 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1135 | return 0; | |
1136 | ||
1137 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1138 | if (cwnd_quota && | |
1139 | !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1140 | cwnd_quota = 0; | |
1141 | ||
1142 | return cwnd_quota; | |
1143 | } | |
1144 | ||
9e412ba7 | 1145 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1146 | { |
9e412ba7 | 1147 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1148 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1149 | |
1150 | return (skb && | |
1151 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
1152 | (tcp_skb_is_last(sk, skb) ? | |
1153 | TCP_NAGLE_PUSH : | |
1154 | tp->nonagle))); | |
1155 | } | |
1156 | ||
1157 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1158 | * which is put after SKB on the list. It is very much like | |
1159 | * tcp_fragment() except that it may make several kinds of assumptions | |
1160 | * in order to speed up the splitting operation. In particular, we | |
1161 | * know that all the data is in scatter-gather pages, and that the | |
1162 | * packet has never been sent out before (and thus is not cloned). | |
1163 | */ | |
846998ae | 1164 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) |
c1b4a7e6 DM |
1165 | { |
1166 | struct sk_buff *buff; | |
1167 | int nlen = skb->len - len; | |
1168 | u16 flags; | |
1169 | ||
1170 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1171 | if (skb->len != skb->data_len) |
1172 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 DM |
1173 | |
1174 | buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC); | |
1175 | if (unlikely(buff == NULL)) | |
1176 | return -ENOMEM; | |
1177 | ||
b60b49ea HX |
1178 | sk_charge_skb(sk, buff); |
1179 | buff->truesize += nlen; | |
c1b4a7e6 DM |
1180 | skb->truesize -= nlen; |
1181 | ||
1182 | /* Correct the sequence numbers. */ | |
1183 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1184 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1185 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1186 | ||
1187 | /* PSH and FIN should only be set in the second packet. */ | |
1188 | flags = TCP_SKB_CB(skb)->flags; | |
1189 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
1190 | TCP_SKB_CB(buff)->flags = flags; | |
1191 | ||
1192 | /* This packet was never sent out yet, so no SACK bits. */ | |
1193 | TCP_SKB_CB(buff)->sacked = 0; | |
1194 | ||
84fa7933 | 1195 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1196 | skb_split(skb, buff, len); |
1197 | ||
1198 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1199 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1200 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1201 | |
1202 | /* Link BUFF into the send queue. */ | |
1203 | skb_header_release(buff); | |
fe067e8a | 1204 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1205 | |
1206 | return 0; | |
1207 | } | |
1208 | ||
1209 | /* Try to defer sending, if possible, in order to minimize the amount | |
1210 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1211 | * | |
1212 | * This algorithm is from John Heffner. | |
1213 | */ | |
9e412ba7 | 1214 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1215 | { |
9e412ba7 | 1216 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1217 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 DM |
1218 | u32 send_win, cong_win, limit, in_flight; |
1219 | ||
1220 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) | |
ae8064ac | 1221 | goto send_now; |
c1b4a7e6 | 1222 | |
6687e988 | 1223 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1224 | goto send_now; |
1225 | ||
1226 | /* Defer for less than two clock ticks. */ | |
1227 | if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1) | |
1228 | goto send_now; | |
908a75c1 | 1229 | |
c1b4a7e6 DM |
1230 | in_flight = tcp_packets_in_flight(tp); |
1231 | ||
1232 | BUG_ON(tcp_skb_pcount(skb) <= 1 || | |
1233 | (tp->snd_cwnd <= in_flight)); | |
1234 | ||
1235 | send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq; | |
1236 | ||
1237 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1238 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1239 | ||
1240 | limit = min(send_win, cong_win); | |
1241 | ||
ba244fe9 DM |
1242 | /* If a full-sized TSO skb can be sent, do it. */ |
1243 | if (limit >= 65536) | |
ae8064ac | 1244 | goto send_now; |
ba244fe9 | 1245 | |
c1b4a7e6 DM |
1246 | if (sysctl_tcp_tso_win_divisor) { |
1247 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1248 | ||
1249 | /* If at least some fraction of a window is available, | |
1250 | * just use it. | |
1251 | */ | |
1252 | chunk /= sysctl_tcp_tso_win_divisor; | |
1253 | if (limit >= chunk) | |
ae8064ac | 1254 | goto send_now; |
c1b4a7e6 DM |
1255 | } else { |
1256 | /* Different approach, try not to defer past a single | |
1257 | * ACK. Receiver should ACK every other full sized | |
1258 | * frame, so if we have space for more than 3 frames | |
1259 | * then send now. | |
1260 | */ | |
1261 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
ae8064ac | 1262 | goto send_now; |
c1b4a7e6 DM |
1263 | } |
1264 | ||
1265 | /* Ok, it looks like it is advisable to defer. */ | |
ae8064ac JH |
1266 | tp->tso_deferred = 1 | (jiffies<<1); |
1267 | ||
c1b4a7e6 | 1268 | return 1; |
ae8064ac JH |
1269 | |
1270 | send_now: | |
1271 | tp->tso_deferred = 0; | |
1272 | return 0; | |
c1b4a7e6 DM |
1273 | } |
1274 | ||
5d424d5a JH |
1275 | /* Create a new MTU probe if we are ready. |
1276 | * Returns 0 if we should wait to probe (no cwnd available), | |
1277 | * 1 if a probe was sent, | |
1278 | * -1 otherwise */ | |
1279 | static int tcp_mtu_probe(struct sock *sk) | |
1280 | { | |
1281 | struct tcp_sock *tp = tcp_sk(sk); | |
1282 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1283 | struct sk_buff *skb, *nskb, *next; | |
1284 | int len; | |
1285 | int probe_size; | |
1286 | unsigned int pif; | |
1287 | int copy; | |
1288 | int mss_now; | |
1289 | ||
1290 | /* Not currently probing/verifying, | |
1291 | * not in recovery, | |
1292 | * have enough cwnd, and | |
1293 | * not SACKing (the variable headers throw things off) */ | |
1294 | if (!icsk->icsk_mtup.enabled || | |
1295 | icsk->icsk_mtup.probe_size || | |
1296 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1297 | tp->snd_cwnd < 11 || | |
1298 | tp->rx_opt.eff_sacks) | |
1299 | return -1; | |
1300 | ||
1301 | /* Very simple search strategy: just double the MSS. */ | |
1302 | mss_now = tcp_current_mss(sk, 0); | |
1303 | probe_size = 2*tp->mss_cache; | |
1304 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { | |
1305 | /* TODO: set timer for probe_converge_event */ | |
1306 | return -1; | |
1307 | } | |
1308 | ||
1309 | /* Have enough data in the send queue to probe? */ | |
1310 | len = 0; | |
fe067e8a | 1311 | if ((skb = tcp_send_head(sk)) == NULL) |
5d424d5a JH |
1312 | return -1; |
1313 | while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb)) | |
fe067e8a | 1314 | skb = tcp_write_queue_next(sk, skb); |
5d424d5a JH |
1315 | if (len < probe_size) |
1316 | return -1; | |
1317 | ||
1318 | /* Receive window check. */ | |
1319 | if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) { | |
1320 | if (tp->snd_wnd < probe_size) | |
1321 | return -1; | |
1322 | else | |
1323 | return 0; | |
1324 | } | |
1325 | ||
1326 | /* Do we need to wait to drain cwnd? */ | |
1327 | pif = tcp_packets_in_flight(tp); | |
1328 | if (pif + 2 > tp->snd_cwnd) { | |
1329 | /* With no packets in flight, don't stall. */ | |
1330 | if (pif == 0) | |
1331 | return -1; | |
1332 | else | |
1333 | return 0; | |
1334 | } | |
1335 | ||
1336 | /* We're allowed to probe. Build it now. */ | |
1337 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1338 | return -1; | |
1339 | sk_charge_skb(sk, nskb); | |
1340 | ||
fe067e8a DM |
1341 | skb = tcp_send_head(sk); |
1342 | tcp_insert_write_queue_before(nskb, skb, sk); | |
1343 | tcp_advance_send_head(sk, skb); | |
5d424d5a JH |
1344 | |
1345 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1346 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
1347 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; | |
1348 | TCP_SKB_CB(nskb)->sacked = 0; | |
1349 | nskb->csum = 0; | |
84fa7933 | 1350 | nskb->ip_summed = skb->ip_summed; |
5d424d5a JH |
1351 | |
1352 | len = 0; | |
1353 | while (len < probe_size) { | |
fe067e8a | 1354 | next = tcp_write_queue_next(sk, skb); |
5d424d5a JH |
1355 | |
1356 | copy = min_t(int, skb->len, probe_size - len); | |
1357 | if (nskb->ip_summed) | |
1358 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1359 | else | |
1360 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
e905a9ed | 1361 | skb_put(nskb, copy), copy, nskb->csum); |
5d424d5a JH |
1362 | |
1363 | if (skb->len <= copy) { | |
1364 | /* We've eaten all the data from this skb. | |
1365 | * Throw it away. */ | |
1366 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
fe067e8a | 1367 | tcp_unlink_write_queue(skb, sk); |
5d424d5a JH |
1368 | sk_stream_free_skb(sk, skb); |
1369 | } else { | |
1370 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
e905a9ed | 1371 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
5d424d5a JH |
1372 | if (!skb_shinfo(skb)->nr_frags) { |
1373 | skb_pull(skb, copy); | |
84fa7933 | 1374 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
5d424d5a JH |
1375 | skb->csum = csum_partial(skb->data, skb->len, 0); |
1376 | } else { | |
1377 | __pskb_trim_head(skb, copy); | |
1378 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1379 | } | |
1380 | TCP_SKB_CB(skb)->seq += copy; | |
1381 | } | |
1382 | ||
1383 | len += copy; | |
1384 | skb = next; | |
1385 | } | |
1386 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1387 | ||
1388 | /* We're ready to send. If this fails, the probe will | |
1389 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1390 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1391 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1392 | /* Decrement cwnd here because we are sending | |
1393 | * effectively two packets. */ | |
1394 | tp->snd_cwnd--; | |
9e412ba7 | 1395 | update_send_head(sk, nskb); |
5d424d5a JH |
1396 | |
1397 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1398 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1399 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1400 | |
1401 | return 1; | |
1402 | } | |
1403 | ||
1404 | return -1; | |
1405 | } | |
1406 | ||
1407 | ||
1da177e4 LT |
1408 | /* This routine writes packets to the network. It advances the |
1409 | * send_head. This happens as incoming acks open up the remote | |
1410 | * window for us. | |
1411 | * | |
1412 | * Returns 1, if no segments are in flight and we have queued segments, but | |
1413 | * cannot send anything now because of SWS or another problem. | |
1414 | */ | |
a2e2a59c | 1415 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) |
1da177e4 LT |
1416 | { |
1417 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1418 | struct sk_buff *skb; |
c1b4a7e6 DM |
1419 | unsigned int tso_segs, sent_pkts; |
1420 | int cwnd_quota; | |
5d424d5a | 1421 | int result; |
1da177e4 LT |
1422 | |
1423 | /* If we are closed, the bytes will have to remain here. | |
1424 | * In time closedown will finish, we empty the write queue and all | |
1425 | * will be happy. | |
1426 | */ | |
92df7b51 DM |
1427 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
1428 | return 0; | |
1da177e4 | 1429 | |
92df7b51 | 1430 | sent_pkts = 0; |
5d424d5a JH |
1431 | |
1432 | /* Do MTU probing. */ | |
1433 | if ((result = tcp_mtu_probe(sk)) == 0) { | |
1434 | return 0; | |
1435 | } else if (result > 0) { | |
1436 | sent_pkts = 1; | |
1437 | } | |
1438 | ||
fe067e8a | 1439 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1440 | unsigned int limit; |
1441 | ||
b68e9f85 | 1442 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1443 | BUG_ON(!tso_segs); |
aa93466b | 1444 | |
b68e9f85 HX |
1445 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1446 | if (!cwnd_quota) | |
1447 | break; | |
1448 | ||
1449 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1450 | break; | |
1451 | ||
c1b4a7e6 DM |
1452 | if (tso_segs == 1) { |
1453 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1454 | (tcp_skb_is_last(sk, skb) ? | |
1455 | nonagle : TCP_NAGLE_PUSH)))) | |
1456 | break; | |
1457 | } else { | |
9e412ba7 | 1458 | if (tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1459 | break; |
1460 | } | |
aa93466b | 1461 | |
c8ac3774 | 1462 | limit = mss_now; |
c1b4a7e6 | 1463 | if (tso_segs > 1) { |
c8ac3774 HX |
1464 | limit = tcp_window_allows(tp, skb, |
1465 | mss_now, cwnd_quota); | |
c1b4a7e6 DM |
1466 | |
1467 | if (skb->len < limit) { | |
1468 | unsigned int trim = skb->len % mss_now; | |
aa93466b | 1469 | |
c1b4a7e6 DM |
1470 | if (trim) |
1471 | limit = skb->len - trim; | |
1472 | } | |
92df7b51 | 1473 | } |
1da177e4 | 1474 | |
c8ac3774 HX |
1475 | if (skb->len > limit && |
1476 | unlikely(tso_fragment(sk, skb, limit, mss_now))) | |
1477 | break; | |
1478 | ||
92df7b51 | 1479 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1480 | |
dfb4b9dc | 1481 | if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) |
92df7b51 | 1482 | break; |
1da177e4 | 1483 | |
92df7b51 DM |
1484 | /* Advance the send_head. This one is sent out. |
1485 | * This call will increment packets_out. | |
1486 | */ | |
9e412ba7 | 1487 | update_send_head(sk, skb); |
1da177e4 | 1488 | |
92df7b51 | 1489 | tcp_minshall_update(tp, mss_now, skb); |
aa93466b | 1490 | sent_pkts++; |
92df7b51 | 1491 | } |
1da177e4 | 1492 | |
aa93466b | 1493 | if (likely(sent_pkts)) { |
9e412ba7 | 1494 | tcp_cwnd_validate(sk); |
92df7b51 | 1495 | return 0; |
1da177e4 | 1496 | } |
fe067e8a | 1497 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1498 | } |
1499 | ||
a762a980 DM |
1500 | /* Push out any pending frames which were held back due to |
1501 | * TCP_CORK or attempt at coalescing tiny packets. | |
1502 | * The socket must be locked by the caller. | |
1503 | */ | |
9e412ba7 IJ |
1504 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1505 | int nonagle) | |
a762a980 | 1506 | { |
fe067e8a | 1507 | struct sk_buff *skb = tcp_send_head(sk); |
a762a980 DM |
1508 | |
1509 | if (skb) { | |
55c97f3e | 1510 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
9e412ba7 | 1511 | tcp_check_probe_timer(sk); |
a762a980 DM |
1512 | } |
1513 | } | |
1514 | ||
c1b4a7e6 DM |
1515 | /* Send _single_ skb sitting at the send head. This function requires |
1516 | * true push pending frames to setup probe timer etc. | |
1517 | */ | |
1518 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1519 | { | |
1520 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1521 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1522 | unsigned int tso_segs, cwnd_quota; |
1523 | ||
1524 | BUG_ON(!skb || skb->len < mss_now); | |
1525 | ||
846998ae | 1526 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1527 | cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); |
1528 | ||
1529 | if (likely(cwnd_quota)) { | |
c8ac3774 HX |
1530 | unsigned int limit; |
1531 | ||
c1b4a7e6 DM |
1532 | BUG_ON(!tso_segs); |
1533 | ||
c8ac3774 | 1534 | limit = mss_now; |
c1b4a7e6 | 1535 | if (tso_segs > 1) { |
c8ac3774 HX |
1536 | limit = tcp_window_allows(tp, skb, |
1537 | mss_now, cwnd_quota); | |
c1b4a7e6 DM |
1538 | |
1539 | if (skb->len < limit) { | |
1540 | unsigned int trim = skb->len % mss_now; | |
1541 | ||
1542 | if (trim) | |
1543 | limit = skb->len - trim; | |
1544 | } | |
c1b4a7e6 DM |
1545 | } |
1546 | ||
c8ac3774 HX |
1547 | if (skb->len > limit && |
1548 | unlikely(tso_fragment(sk, skb, limit, mss_now))) | |
1549 | return; | |
1550 | ||
c1b4a7e6 DM |
1551 | /* Send it out now. */ |
1552 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1553 | ||
dfb4b9dc | 1554 | if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { |
9e412ba7 IJ |
1555 | update_send_head(sk, skb); |
1556 | tcp_cwnd_validate(sk); | |
c1b4a7e6 DM |
1557 | return; |
1558 | } | |
1559 | } | |
1560 | } | |
1561 | ||
1da177e4 LT |
1562 | /* This function returns the amount that we can raise the |
1563 | * usable window based on the following constraints | |
e905a9ed | 1564 | * |
1da177e4 LT |
1565 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1566 | * 2. We limit memory per socket | |
1567 | * | |
1568 | * RFC 1122: | |
1569 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1570 | * RECV.NEXT + RCV.WIN fixed until: | |
1571 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1572 | * | |
1573 | * i.e. don't raise the right edge of the window until you can raise | |
1574 | * it at least MSS bytes. | |
1575 | * | |
1576 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1577 | * since header prediction assumes th->window stays fixed. | |
1578 | * | |
1579 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1580 | * side SWS prevention criteria. The problem is that under this rule | |
1581 | * a stream of single byte packets will cause the right side of the | |
1582 | * window to always advance by a single byte. | |
e905a9ed | 1583 | * |
1da177e4 LT |
1584 | * Of course, if the sender implements sender side SWS prevention |
1585 | * then this will not be a problem. | |
e905a9ed | 1586 | * |
1da177e4 | 1587 | * BSD seems to make the following compromise: |
e905a9ed | 1588 | * |
1da177e4 LT |
1589 | * If the free space is less than the 1/4 of the maximum |
1590 | * space available and the free space is less than 1/2 mss, | |
1591 | * then set the window to 0. | |
1592 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1593 | * Otherwise, just prevent the window from shrinking | |
1594 | * and from being larger than the largest representable value. | |
1595 | * | |
1596 | * This prevents incremental opening of the window in the regime | |
1597 | * where TCP is limited by the speed of the reader side taking | |
1598 | * data out of the TCP receive queue. It does nothing about | |
1599 | * those cases where the window is constrained on the sender side | |
1600 | * because the pipeline is full. | |
1601 | * | |
1602 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1603 | * multiple of MSS, at least until the free space gets quite small. | |
1604 | * This would appear to be a side effect of the mbuf implementation. | |
1605 | * Combining these two algorithms results in the observed behavior | |
1606 | * of having a fixed window size at almost all times. | |
1607 | * | |
1608 | * Below we obtain similar behavior by forcing the offered window to | |
1609 | * a multiple of the mss when it is feasible to do so. | |
1610 | * | |
1611 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1612 | * Regular options like TIMESTAMP are taken into account. | |
1613 | */ | |
1614 | u32 __tcp_select_window(struct sock *sk) | |
1615 | { | |
463c84b9 | 1616 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1617 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1618 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1619 | * here. I don't know if the value based on our guesses |
1620 | * of peer's MSS is better for the performance. It's more correct | |
1621 | * but may be worse for the performance because of rcv_mss | |
1622 | * fluctuations. --SAW 1998/11/1 | |
1623 | */ | |
463c84b9 | 1624 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1625 | int free_space = tcp_space(sk); |
1626 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1627 | int window; | |
1628 | ||
1629 | if (mss > full_space) | |
e905a9ed | 1630 | mss = full_space; |
1da177e4 LT |
1631 | |
1632 | if (free_space < full_space/2) { | |
463c84b9 | 1633 | icsk->icsk_ack.quick = 0; |
1da177e4 LT |
1634 | |
1635 | if (tcp_memory_pressure) | |
1636 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); | |
1637 | ||
1638 | if (free_space < mss) | |
1639 | return 0; | |
1640 | } | |
1641 | ||
1642 | if (free_space > tp->rcv_ssthresh) | |
1643 | free_space = tp->rcv_ssthresh; | |
1644 | ||
1645 | /* Don't do rounding if we are using window scaling, since the | |
1646 | * scaled window will not line up with the MSS boundary anyway. | |
1647 | */ | |
1648 | window = tp->rcv_wnd; | |
1649 | if (tp->rx_opt.rcv_wscale) { | |
1650 | window = free_space; | |
1651 | ||
1652 | /* Advertise enough space so that it won't get scaled away. | |
1653 | * Import case: prevent zero window announcement if | |
1654 | * 1<<rcv_wscale > mss. | |
1655 | */ | |
1656 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1657 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1658 | << tp->rx_opt.rcv_wscale); | |
1659 | } else { | |
1660 | /* Get the largest window that is a nice multiple of mss. | |
1661 | * Window clamp already applied above. | |
1662 | * If our current window offering is within 1 mss of the | |
1663 | * free space we just keep it. This prevents the divide | |
1664 | * and multiply from happening most of the time. | |
1665 | * We also don't do any window rounding when the free space | |
1666 | * is too small. | |
1667 | */ | |
1668 | if (window <= free_space - mss || window > free_space) | |
1669 | window = (free_space/mss)*mss; | |
84565070 | 1670 | else if (mss == full_space && |
9c681b43 | 1671 | free_space > window + full_space/2) |
84565070 | 1672 | window = free_space; |
1da177e4 LT |
1673 | } |
1674 | ||
1675 | return window; | |
1676 | } | |
1677 | ||
1678 | /* Attempt to collapse two adjacent SKB's during retransmission. */ | |
1679 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) | |
1680 | { | |
1681 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1682 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
1da177e4 LT |
1683 | |
1684 | /* The first test we must make is that neither of these two | |
1685 | * SKB's are still referenced by someone else. | |
1686 | */ | |
1687 | if (!skb_cloned(skb) && !skb_cloned(next_skb)) { | |
1688 | int skb_size = skb->len, next_skb_size = next_skb->len; | |
1689 | u16 flags = TCP_SKB_CB(skb)->flags; | |
1690 | ||
1691 | /* Also punt if next skb has been SACK'd. */ | |
2de979bd | 1692 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) |
1da177e4 LT |
1693 | return; |
1694 | ||
1695 | /* Next skb is out of window. */ | |
1696 | if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) | |
1697 | return; | |
1698 | ||
1699 | /* Punt if not enough space exists in the first SKB for | |
1700 | * the data in the second, or the total combined payload | |
1701 | * would exceed the MSS. | |
1702 | */ | |
1703 | if ((next_skb_size > skb_tailroom(skb)) || | |
1704 | ((skb_size + next_skb_size) > mss_now)) | |
1705 | return; | |
1706 | ||
1707 | BUG_ON(tcp_skb_pcount(skb) != 1 || | |
1708 | tcp_skb_pcount(next_skb) != 1); | |
1709 | ||
6a438bbe SH |
1710 | /* changing transmit queue under us so clear hints */ |
1711 | clear_all_retrans_hints(tp); | |
1712 | ||
1713 | /* Ok. We will be able to collapse the packet. */ | |
fe067e8a | 1714 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 1715 | |
1a4e2d09 ACM |
1716 | skb_copy_from_linear_data(next_skb, |
1717 | skb_put(skb, next_skb_size), | |
1718 | next_skb_size); | |
1da177e4 | 1719 | |
52d570aa JP |
1720 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
1721 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 1722 | |
84fa7933 | 1723 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
1da177e4 LT |
1724 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
1725 | ||
1726 | /* Update sequence range on original skb. */ | |
1727 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1728 | ||
0dde7b54 IJ |
1729 | if (WARN_ON(tp->sacked_out && |
1730 | (TCP_SKB_CB(next_skb)->seq == tp->highest_sack))) | |
1731 | return; | |
1732 | ||
1da177e4 LT |
1733 | /* Merge over control information. */ |
1734 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ | |
1735 | TCP_SKB_CB(skb)->flags = flags; | |
1736 | ||
1737 | /* All done, get rid of second SKB and account for it so | |
1738 | * packet counting does not break. | |
1739 | */ | |
1740 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); | |
1741 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) | |
1742 | tp->retrans_out -= tcp_skb_pcount(next_skb); | |
b5860bba | 1743 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) |
1da177e4 | 1744 | tp->lost_out -= tcp_skb_pcount(next_skb); |
1da177e4 | 1745 | /* Reno case is special. Sigh... */ |
e60402d0 | 1746 | if (tcp_is_reno(tp) && tp->sacked_out) |
1da177e4 | 1747 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); |
1da177e4 LT |
1748 | |
1749 | /* Not quite right: it can be > snd.fack, but | |
1750 | * it is better to underestimate fackets. | |
1751 | */ | |
1752 | tcp_dec_pcount_approx(&tp->fackets_out, next_skb); | |
e9144bd8 | 1753 | tp->packets_out -= tcp_skb_pcount(next_skb); |
1da177e4 LT |
1754 | sk_stream_free_skb(sk, next_skb); |
1755 | } | |
1756 | } | |
1757 | ||
1758 | /* Do a simple retransmit without using the backoff mechanisms in | |
e905a9ed | 1759 | * tcp_timer. This is used for path mtu discovery. |
1da177e4 | 1760 | * The socket is already locked here. |
e905a9ed | 1761 | */ |
1da177e4 LT |
1762 | void tcp_simple_retransmit(struct sock *sk) |
1763 | { | |
6687e988 | 1764 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1765 | struct tcp_sock *tp = tcp_sk(sk); |
1766 | struct sk_buff *skb; | |
1767 | unsigned int mss = tcp_current_mss(sk, 0); | |
1768 | int lost = 0; | |
1769 | ||
fe067e8a DM |
1770 | tcp_for_write_queue(skb, sk) { |
1771 | if (skb == tcp_send_head(sk)) | |
1772 | break; | |
e905a9ed | 1773 | if (skb->len > mss && |
1da177e4 LT |
1774 | !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { |
1775 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1776 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1777 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1778 | } | |
1779 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { | |
1780 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1781 | tp->lost_out += tcp_skb_pcount(skb); | |
1782 | lost = 1; | |
1783 | } | |
1784 | } | |
1785 | } | |
1786 | ||
6a438bbe SH |
1787 | clear_all_retrans_hints(tp); |
1788 | ||
1da177e4 LT |
1789 | if (!lost) |
1790 | return; | |
1791 | ||
005903bc | 1792 | tcp_verify_left_out(tp); |
1da177e4 | 1793 | |
e905a9ed | 1794 | /* Don't muck with the congestion window here. |
1da177e4 LT |
1795 | * Reason is that we do not increase amount of _data_ |
1796 | * in network, but units changed and effective | |
1797 | * cwnd/ssthresh really reduced now. | |
1798 | */ | |
6687e988 | 1799 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
1da177e4 | 1800 | tp->high_seq = tp->snd_nxt; |
6687e988 | 1801 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
1802 | tp->prior_ssthresh = 0; |
1803 | tp->undo_marker = 0; | |
6687e988 | 1804 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1805 | } |
1806 | tcp_xmit_retransmit_queue(sk); | |
1807 | } | |
1808 | ||
1809 | /* This retransmits one SKB. Policy decisions and retransmit queue | |
1810 | * state updates are done by the caller. Returns non-zero if an | |
1811 | * error occurred which prevented the send. | |
1812 | */ | |
1813 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
1814 | { | |
1815 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 1816 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 1817 | unsigned int cur_mss = tcp_current_mss(sk, 0); |
1da177e4 LT |
1818 | int err; |
1819 | ||
5d424d5a JH |
1820 | /* Inconslusive MTU probe */ |
1821 | if (icsk->icsk_mtup.probe_size) { | |
1822 | icsk->icsk_mtup.probe_size = 0; | |
1823 | } | |
1824 | ||
1da177e4 | 1825 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 1826 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
1827 | */ |
1828 | if (atomic_read(&sk->sk_wmem_alloc) > | |
1829 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
1830 | return -EAGAIN; | |
1831 | ||
1832 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
1833 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1834 | BUG(); | |
1da177e4 LT |
1835 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
1836 | return -ENOMEM; | |
1837 | } | |
1838 | ||
1839 | /* If receiver has shrunk his window, and skb is out of | |
1840 | * new window, do not retransmit it. The exception is the | |
1841 | * case, when window is shrunk to zero. In this case | |
1842 | * our retransmit serves as a zero window probe. | |
1843 | */ | |
1844 | if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) | |
1845 | && TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1846 | return -EAGAIN; | |
1847 | ||
1848 | if (skb->len > cur_mss) { | |
846998ae | 1849 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 1850 | return -ENOMEM; /* We'll try again later. */ |
1da177e4 LT |
1851 | } |
1852 | ||
1853 | /* Collapse two adjacent packets if worthwhile and we can. */ | |
2de979bd SH |
1854 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && |
1855 | (skb->len < (cur_mss >> 1)) && | |
1856 | (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && | |
1857 | (!tcp_skb_is_last(sk, skb)) && | |
1858 | (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && | |
1859 | (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && | |
1860 | (sysctl_tcp_retrans_collapse != 0)) | |
1da177e4 LT |
1861 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1862 | ||
8292a17a | 1863 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
1da177e4 LT |
1864 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
1865 | ||
1866 | /* Some Solaris stacks overoptimize and ignore the FIN on a | |
1867 | * retransmit when old data is attached. So strip it off | |
1868 | * since it is cheap to do so and saves bytes on the network. | |
1869 | */ | |
2de979bd SH |
1870 | if (skb->len > 0 && |
1871 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && | |
1872 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { | |
1da177e4 LT |
1873 | if (!pskb_trim(skb, 0)) { |
1874 | TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; | |
7967168c HX |
1875 | skb_shinfo(skb)->gso_segs = 1; |
1876 | skb_shinfo(skb)->gso_size = 0; | |
1877 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
1878 | skb->ip_summed = CHECKSUM_NONE; |
1879 | skb->csum = 0; | |
1880 | } | |
1881 | } | |
1882 | ||
1883 | /* Make a copy, if the first transmission SKB clone we made | |
1884 | * is still in somebody's hands, else make a clone. | |
1885 | */ | |
1886 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 1887 | |
dfb4b9dc | 1888 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
1889 | |
1890 | if (err == 0) { | |
1891 | /* Update global TCP statistics. */ | |
1892 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); | |
1893 | ||
1894 | tp->total_retrans++; | |
1895 | ||
1896 | #if FASTRETRANS_DEBUG > 0 | |
1897 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1898 | if (net_ratelimit()) | |
1899 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
1900 | } | |
1901 | #endif | |
1902 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
1903 | tp->retrans_out += tcp_skb_pcount(skb); | |
1904 | ||
1905 | /* Save stamp of the first retransmit. */ | |
1906 | if (!tp->retrans_stamp) | |
1907 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
1908 | ||
1909 | tp->undo_retrans++; | |
1910 | ||
1911 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
1912 | * see tcp_input.c tcp_sacktag_write_queue(). | |
1913 | */ | |
1914 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
1915 | } | |
1916 | return err; | |
1917 | } | |
1918 | ||
1919 | /* This gets called after a retransmit timeout, and the initially | |
1920 | * retransmitted data is acknowledged. It tries to continue | |
1921 | * resending the rest of the retransmit queue, until either | |
1922 | * we've sent it all or the congestion window limit is reached. | |
1923 | * If doing SACK, the first ACK which comes back for a timeout | |
1924 | * based retransmit packet might feed us FACK information again. | |
1925 | * If so, we use it to avoid unnecessarily retransmissions. | |
1926 | */ | |
1927 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
1928 | { | |
6687e988 | 1929 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1930 | struct tcp_sock *tp = tcp_sk(sk); |
1931 | struct sk_buff *skb; | |
6a438bbe SH |
1932 | int packet_cnt; |
1933 | ||
1934 | if (tp->retransmit_skb_hint) { | |
1935 | skb = tp->retransmit_skb_hint; | |
1936 | packet_cnt = tp->retransmit_cnt_hint; | |
1937 | }else{ | |
fe067e8a | 1938 | skb = tcp_write_queue_head(sk); |
6a438bbe SH |
1939 | packet_cnt = 0; |
1940 | } | |
1da177e4 LT |
1941 | |
1942 | /* First pass: retransmit lost packets. */ | |
6a438bbe | 1943 | if (tp->lost_out) { |
fe067e8a | 1944 | tcp_for_write_queue_from(skb, sk) { |
1da177e4 LT |
1945 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
1946 | ||
fe067e8a DM |
1947 | if (skb == tcp_send_head(sk)) |
1948 | break; | |
6a438bbe SH |
1949 | /* we could do better than to assign each time */ |
1950 | tp->retransmit_skb_hint = skb; | |
1951 | tp->retransmit_cnt_hint = packet_cnt; | |
1952 | ||
1da177e4 LT |
1953 | /* Assume this retransmit will generate |
1954 | * only one packet for congestion window | |
1955 | * calculation purposes. This works because | |
1956 | * tcp_retransmit_skb() will chop up the | |
1957 | * packet to be MSS sized and all the | |
1958 | * packet counting works out. | |
1959 | */ | |
1960 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
1961 | return; | |
1962 | ||
6a438bbe | 1963 | if (sacked & TCPCB_LOST) { |
1da177e4 | 1964 | if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { |
6a438bbe SH |
1965 | if (tcp_retransmit_skb(sk, skb)) { |
1966 | tp->retransmit_skb_hint = NULL; | |
1da177e4 | 1967 | return; |
6a438bbe | 1968 | } |
6687e988 | 1969 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
1da177e4 LT |
1970 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); |
1971 | else | |
1972 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); | |
1973 | ||
fe067e8a | 1974 | if (skb == tcp_write_queue_head(sk)) |
463c84b9 | 1975 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3f421baa ACM |
1976 | inet_csk(sk)->icsk_rto, |
1977 | TCP_RTO_MAX); | |
1da177e4 LT |
1978 | } |
1979 | ||
6a438bbe SH |
1980 | packet_cnt += tcp_skb_pcount(skb); |
1981 | if (packet_cnt >= tp->lost_out) | |
1da177e4 LT |
1982 | break; |
1983 | } | |
1984 | } | |
1985 | } | |
1986 | ||
1987 | /* OK, demanded retransmission is finished. */ | |
1988 | ||
1989 | /* Forward retransmissions are possible only during Recovery. */ | |
6687e988 | 1990 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
1da177e4 LT |
1991 | return; |
1992 | ||
1993 | /* No forward retransmissions in Reno are possible. */ | |
e60402d0 | 1994 | if (tcp_is_reno(tp)) |
1da177e4 LT |
1995 | return; |
1996 | ||
1997 | /* Yeah, we have to make difficult choice between forward transmission | |
1998 | * and retransmission... Both ways have their merits... | |
1999 | * | |
2000 | * For now we do not retransmit anything, while we have some new | |
539d243f IJ |
2001 | * segments to send. In the other cases, follow rule 3 for |
2002 | * NextSeg() specified in RFC3517. | |
1da177e4 LT |
2003 | */ |
2004 | ||
9e412ba7 | 2005 | if (tcp_may_send_now(sk)) |
1da177e4 LT |
2006 | return; |
2007 | ||
539d243f IJ |
2008 | /* If nothing is SACKed, highest_sack in the loop won't be valid */ |
2009 | if (!tp->sacked_out) | |
2010 | return; | |
2011 | ||
2012 | if (tp->forward_skb_hint) | |
6a438bbe | 2013 | skb = tp->forward_skb_hint; |
539d243f | 2014 | else |
fe067e8a | 2015 | skb = tcp_write_queue_head(sk); |
6a438bbe | 2016 | |
fe067e8a DM |
2017 | tcp_for_write_queue_from(skb, sk) { |
2018 | if (skb == tcp_send_head(sk)) | |
2019 | break; | |
6a438bbe | 2020 | tp->forward_skb_hint = skb; |
1da177e4 | 2021 | |
539d243f | 2022 | if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) |
1da177e4 LT |
2023 | break; |
2024 | ||
2025 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2026 | break; | |
2027 | ||
2028 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) | |
2029 | continue; | |
2030 | ||
2031 | /* Ok, retransmit it. */ | |
6a438bbe SH |
2032 | if (tcp_retransmit_skb(sk, skb)) { |
2033 | tp->forward_skb_hint = NULL; | |
1da177e4 | 2034 | break; |
6a438bbe | 2035 | } |
1da177e4 | 2036 | |
fe067e8a | 2037 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2038 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2039 | inet_csk(sk)->icsk_rto, | |
2040 | TCP_RTO_MAX); | |
1da177e4 LT |
2041 | |
2042 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); | |
2043 | } | |
2044 | } | |
2045 | ||
2046 | ||
2047 | /* Send a fin. The caller locks the socket for us. This cannot be | |
2048 | * allowed to fail queueing a FIN frame under any circumstances. | |
2049 | */ | |
2050 | void tcp_send_fin(struct sock *sk) | |
2051 | { | |
e905a9ed | 2052 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2053 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2054 | int mss_now; |
e905a9ed | 2055 | |
1da177e4 LT |
2056 | /* Optimization, tack on the FIN if we have a queue of |
2057 | * unsent frames. But be careful about outgoing SACKS | |
2058 | * and IP options. | |
2059 | */ | |
2060 | mss_now = tcp_current_mss(sk, 1); | |
2061 | ||
fe067e8a | 2062 | if (tcp_send_head(sk) != NULL) { |
1da177e4 LT |
2063 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; |
2064 | TCP_SKB_CB(skb)->end_seq++; | |
2065 | tp->write_seq++; | |
2066 | } else { | |
2067 | /* Socket is locked, keep trying until memory is available. */ | |
2068 | for (;;) { | |
d179cd12 | 2069 | skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); |
1da177e4 LT |
2070 | if (skb) |
2071 | break; | |
2072 | yield(); | |
2073 | } | |
2074 | ||
2075 | /* Reserve space for headers and prepare control bits. */ | |
2076 | skb_reserve(skb, MAX_TCP_HEADER); | |
2077 | skb->csum = 0; | |
2078 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); | |
2079 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2080 | skb_shinfo(skb)->gso_segs = 1; |
2081 | skb_shinfo(skb)->gso_size = 0; | |
2082 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2083 | |
2084 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ | |
2085 | TCP_SKB_CB(skb)->seq = tp->write_seq; | |
2086 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; | |
2087 | tcp_queue_skb(sk, skb); | |
2088 | } | |
9e412ba7 | 2089 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2090 | } |
2091 | ||
2092 | /* We get here when a process closes a file descriptor (either due to | |
2093 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2094 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2095 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2096 | */ |
dd0fc66f | 2097 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2098 | { |
1da177e4 LT |
2099 | struct sk_buff *skb; |
2100 | ||
2101 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2102 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2103 | if (!skb) { | |
2104 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); | |
2105 | return; | |
2106 | } | |
2107 | ||
2108 | /* Reserve space for headers and prepare control bits. */ | |
2109 | skb_reserve(skb, MAX_TCP_HEADER); | |
2110 | skb->csum = 0; | |
2111 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); | |
2112 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2113 | skb_shinfo(skb)->gso_segs = 1; |
2114 | skb_shinfo(skb)->gso_size = 0; | |
2115 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2116 | |
2117 | /* Send it off. */ | |
9e412ba7 | 2118 | TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk); |
1da177e4 LT |
2119 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; |
2120 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2121 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
1da177e4 LT |
2122 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
2123 | } | |
2124 | ||
2125 | /* WARNING: This routine must only be called when we have already sent | |
2126 | * a SYN packet that crossed the incoming SYN that caused this routine | |
2127 | * to get called. If this assumption fails then the initial rcv_wnd | |
2128 | * and rcv_wscale values will not be correct. | |
2129 | */ | |
2130 | int tcp_send_synack(struct sock *sk) | |
2131 | { | |
2132 | struct sk_buff* skb; | |
2133 | ||
fe067e8a | 2134 | skb = tcp_write_queue_head(sk); |
1da177e4 LT |
2135 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { |
2136 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); | |
2137 | return -EFAULT; | |
2138 | } | |
2139 | if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { | |
2140 | if (skb_cloned(skb)) { | |
2141 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2142 | if (nskb == NULL) | |
2143 | return -ENOMEM; | |
fe067e8a | 2144 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2145 | skb_header_release(nskb); |
fe067e8a | 2146 | __tcp_add_write_queue_head(sk, nskb); |
1da177e4 LT |
2147 | sk_stream_free_skb(sk, skb); |
2148 | sk_charge_skb(sk, nskb); | |
2149 | skb = nskb; | |
2150 | } | |
2151 | ||
2152 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; | |
2153 | TCP_ECN_send_synack(tcp_sk(sk), skb); | |
2154 | } | |
2155 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2156 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2157 | } |
2158 | ||
2159 | /* | |
2160 | * Prepare a SYN-ACK. | |
2161 | */ | |
2162 | struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, | |
60236fdd | 2163 | struct request_sock *req) |
1da177e4 | 2164 | { |
2e6599cb | 2165 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
2166 | struct tcp_sock *tp = tcp_sk(sk); |
2167 | struct tcphdr *th; | |
2168 | int tcp_header_size; | |
2169 | struct sk_buff *skb; | |
cfb6eeb4 YH |
2170 | #ifdef CONFIG_TCP_MD5SIG |
2171 | struct tcp_md5sig_key *md5; | |
2172 | __u8 *md5_hash_location; | |
2173 | #endif | |
1da177e4 LT |
2174 | |
2175 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); | |
2176 | if (skb == NULL) | |
2177 | return NULL; | |
2178 | ||
2179 | /* Reserve space for headers. */ | |
2180 | skb_reserve(skb, MAX_TCP_HEADER); | |
2181 | ||
2182 | skb->dst = dst_clone(dst); | |
2183 | ||
2184 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + | |
2e6599cb ACM |
2185 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
2186 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + | |
1da177e4 | 2187 | /* SACK_PERM is in the place of NOP NOP of TS */ |
2e6599cb | 2188 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
cfb6eeb4 YH |
2189 | |
2190 | #ifdef CONFIG_TCP_MD5SIG | |
2191 | /* Are we doing MD5 on this segment? If so - make room for it */ | |
2192 | md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); | |
2193 | if (md5) | |
2194 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
2195 | #endif | |
aa8223c7 ACM |
2196 | skb_push(skb, tcp_header_size); |
2197 | skb_reset_transport_header(skb); | |
1da177e4 | 2198 | |
aa8223c7 | 2199 | th = tcp_hdr(skb); |
1da177e4 LT |
2200 | memset(th, 0, sizeof(struct tcphdr)); |
2201 | th->syn = 1; | |
2202 | th->ack = 1; | |
1da177e4 LT |
2203 | TCP_ECN_make_synack(req, th); |
2204 | th->source = inet_sk(sk)->sport; | |
2e6599cb ACM |
2205 | th->dest = ireq->rmt_port; |
2206 | TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; | |
1da177e4 LT |
2207 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; |
2208 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2209 | skb_shinfo(skb)->gso_segs = 1; |
2210 | skb_shinfo(skb)->gso_size = 0; | |
2211 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 | 2212 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2213 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 | 2214 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
e905a9ed | 2215 | __u8 rcv_wscale; |
1da177e4 LT |
2216 | /* Set this up on the first call only */ |
2217 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
2218 | /* tcp_full_space because it is guaranteed to be the first packet */ | |
e905a9ed | 2219 | tcp_select_initial_window(tcp_full_space(sk), |
2e6599cb | 2220 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
1da177e4 LT |
2221 | &req->rcv_wnd, |
2222 | &req->window_clamp, | |
2e6599cb | 2223 | ireq->wscale_ok, |
1da177e4 | 2224 | &rcv_wscale); |
e905a9ed | 2225 | ireq->rcv_wscale = rcv_wscale; |
1da177e4 LT |
2226 | } |
2227 | ||
2228 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2229 | th->window = htons(min(req->rcv_wnd, 65535U)); |
1da177e4 LT |
2230 | |
2231 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
df7a3b07 | 2232 | tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, |
2e6599cb | 2233 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, |
1da177e4 | 2234 | TCP_SKB_CB(skb)->when, |
cfb6eeb4 YH |
2235 | req->ts_recent, |
2236 | ( | |
2237 | #ifdef CONFIG_TCP_MD5SIG | |
2238 | md5 ? &md5_hash_location : | |
2239 | #endif | |
2240 | NULL) | |
2241 | ); | |
1da177e4 LT |
2242 | |
2243 | skb->csum = 0; | |
2244 | th->doff = (tcp_header_size >> 2); | |
2245 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
cfb6eeb4 YH |
2246 | |
2247 | #ifdef CONFIG_TCP_MD5SIG | |
2248 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2249 | if (md5) { | |
2250 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
2251 | md5, | |
2252 | NULL, dst, req, | |
aa8223c7 | 2253 | tcp_hdr(skb), sk->sk_protocol, |
cfb6eeb4 YH |
2254 | skb->len); |
2255 | } | |
2256 | #endif | |
2257 | ||
1da177e4 LT |
2258 | return skb; |
2259 | } | |
2260 | ||
e905a9ed | 2261 | /* |
1da177e4 | 2262 | * Do all connect socket setups that can be done AF independent. |
e905a9ed | 2263 | */ |
40efc6fa | 2264 | static void tcp_connect_init(struct sock *sk) |
1da177e4 LT |
2265 | { |
2266 | struct dst_entry *dst = __sk_dst_get(sk); | |
2267 | struct tcp_sock *tp = tcp_sk(sk); | |
2268 | __u8 rcv_wscale; | |
2269 | ||
2270 | /* We'll fix this up when we get a response from the other end. | |
2271 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2272 | */ | |
2273 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
2274 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); | |
2275 | ||
cfb6eeb4 YH |
2276 | #ifdef CONFIG_TCP_MD5SIG |
2277 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2278 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2279 | #endif | |
2280 | ||
1da177e4 LT |
2281 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2282 | if (tp->rx_opt.user_mss) | |
2283 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2284 | tp->max_window = 0; | |
5d424d5a | 2285 | tcp_mtup_init(sk); |
1da177e4 LT |
2286 | tcp_sync_mss(sk, dst_mtu(dst)); |
2287 | ||
2288 | if (!tp->window_clamp) | |
2289 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2290 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
2291 | tcp_initialize_rcv_mss(sk); | |
1da177e4 LT |
2292 | |
2293 | tcp_select_initial_window(tcp_full_space(sk), | |
2294 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2295 | &tp->rcv_wnd, | |
2296 | &tp->window_clamp, | |
2297 | sysctl_tcp_window_scaling, | |
2298 | &rcv_wscale); | |
2299 | ||
2300 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2301 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2302 | ||
2303 | sk->sk_err = 0; | |
2304 | sock_reset_flag(sk, SOCK_DONE); | |
2305 | tp->snd_wnd = 0; | |
2306 | tcp_init_wl(tp, tp->write_seq, 0); | |
2307 | tp->snd_una = tp->write_seq; | |
2308 | tp->snd_sml = tp->write_seq; | |
2309 | tp->rcv_nxt = 0; | |
2310 | tp->rcv_wup = 0; | |
2311 | tp->copied_seq = 0; | |
2312 | ||
463c84b9 ACM |
2313 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2314 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2315 | tcp_clear_retrans(tp); |
2316 | } | |
2317 | ||
2318 | /* | |
2319 | * Build a SYN and send it off. | |
e905a9ed | 2320 | */ |
1da177e4 LT |
2321 | int tcp_connect(struct sock *sk) |
2322 | { | |
2323 | struct tcp_sock *tp = tcp_sk(sk); | |
2324 | struct sk_buff *buff; | |
2325 | ||
2326 | tcp_connect_init(sk); | |
2327 | ||
d179cd12 | 2328 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2329 | if (unlikely(buff == NULL)) |
2330 | return -ENOBUFS; | |
2331 | ||
2332 | /* Reserve space for headers. */ | |
2333 | skb_reserve(buff, MAX_TCP_HEADER); | |
2334 | ||
2335 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; | |
9e412ba7 | 2336 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 2337 | TCP_SKB_CB(buff)->sacked = 0; |
7967168c HX |
2338 | skb_shinfo(buff)->gso_segs = 1; |
2339 | skb_shinfo(buff)->gso_size = 0; | |
2340 | skb_shinfo(buff)->gso_type = 0; | |
1da177e4 | 2341 | buff->csum = 0; |
bd37a088 | 2342 | tp->snd_nxt = tp->write_seq; |
1da177e4 LT |
2343 | TCP_SKB_CB(buff)->seq = tp->write_seq++; |
2344 | TCP_SKB_CB(buff)->end_seq = tp->write_seq; | |
1da177e4 LT |
2345 | |
2346 | /* Send it off. */ | |
2347 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2348 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2349 | skb_header_release(buff); | |
fe067e8a | 2350 | __tcp_add_write_queue_tail(sk, buff); |
1da177e4 LT |
2351 | sk_charge_skb(sk, buff); |
2352 | tp->packets_out += tcp_skb_pcount(buff); | |
dfb4b9dc | 2353 | tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); |
bd37a088 WY |
2354 | |
2355 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2356 | * in order to make this packet get counted in tcpOutSegs. | |
2357 | */ | |
2358 | tp->snd_nxt = tp->write_seq; | |
2359 | tp->pushed_seq = tp->write_seq; | |
1da177e4 LT |
2360 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); |
2361 | ||
2362 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2363 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2364 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2365 | return 0; |
2366 | } | |
2367 | ||
2368 | /* Send out a delayed ack, the caller does the policy checking | |
2369 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2370 | * for details. | |
2371 | */ | |
2372 | void tcp_send_delayed_ack(struct sock *sk) | |
2373 | { | |
463c84b9 ACM |
2374 | struct inet_connection_sock *icsk = inet_csk(sk); |
2375 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2376 | unsigned long timeout; |
2377 | ||
2378 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2379 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2380 | int max_ato = HZ/2; |
2381 | ||
463c84b9 | 2382 | if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
1da177e4 LT |
2383 | max_ato = TCP_DELACK_MAX; |
2384 | ||
2385 | /* Slow path, intersegment interval is "high". */ | |
2386 | ||
2387 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2388 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2389 | * directly. |
2390 | */ | |
2391 | if (tp->srtt) { | |
2392 | int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); | |
2393 | ||
2394 | if (rtt < max_ato) | |
2395 | max_ato = rtt; | |
2396 | } | |
2397 | ||
2398 | ato = min(ato, max_ato); | |
2399 | } | |
2400 | ||
2401 | /* Stay within the limit we were given */ | |
2402 | timeout = jiffies + ato; | |
2403 | ||
2404 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2405 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2406 | /* If delack timer was blocked or is about to expire, |
2407 | * send ACK now. | |
2408 | */ | |
463c84b9 ACM |
2409 | if (icsk->icsk_ack.blocked || |
2410 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2411 | tcp_send_ack(sk); |
2412 | return; | |
2413 | } | |
2414 | ||
463c84b9 ACM |
2415 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2416 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2417 | } |
463c84b9 ACM |
2418 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2419 | icsk->icsk_ack.timeout = timeout; | |
2420 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2421 | } |
2422 | ||
2423 | /* This routine sends an ack and also updates the window. */ | |
2424 | void tcp_send_ack(struct sock *sk) | |
2425 | { | |
2426 | /* If we have been reset, we may not send again. */ | |
2427 | if (sk->sk_state != TCP_CLOSE) { | |
1da177e4 LT |
2428 | struct sk_buff *buff; |
2429 | ||
2430 | /* We are not putting this on the write queue, so | |
2431 | * tcp_transmit_skb() will set the ownership to this | |
2432 | * sock. | |
2433 | */ | |
2434 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2435 | if (buff == NULL) { | |
463c84b9 ACM |
2436 | inet_csk_schedule_ack(sk); |
2437 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3f421baa ACM |
2438 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
2439 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
2440 | return; |
2441 | } | |
2442 | ||
2443 | /* Reserve space for headers and prepare control bits. */ | |
2444 | skb_reserve(buff, MAX_TCP_HEADER); | |
2445 | buff->csum = 0; | |
2446 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; | |
2447 | TCP_SKB_CB(buff)->sacked = 0; | |
7967168c HX |
2448 | skb_shinfo(buff)->gso_segs = 1; |
2449 | skb_shinfo(buff)->gso_size = 0; | |
2450 | skb_shinfo(buff)->gso_type = 0; | |
1da177e4 LT |
2451 | |
2452 | /* Send it off, this clears delayed acks for us. */ | |
9e412ba7 | 2453 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk); |
1da177e4 | 2454 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
dfb4b9dc | 2455 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); |
1da177e4 LT |
2456 | } |
2457 | } | |
2458 | ||
2459 | /* This routine sends a packet with an out of date sequence | |
2460 | * number. It assumes the other end will try to ack it. | |
2461 | * | |
2462 | * Question: what should we make while urgent mode? | |
2463 | * 4.4BSD forces sending single byte of data. We cannot send | |
2464 | * out of window data, because we have SND.NXT==SND.MAX... | |
2465 | * | |
2466 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2467 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2468 | * out-of-date with SND.UNA-1 to probe window. | |
2469 | */ | |
2470 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2471 | { | |
2472 | struct tcp_sock *tp = tcp_sk(sk); | |
2473 | struct sk_buff *skb; | |
2474 | ||
2475 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2476 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2477 | if (skb == NULL) |
1da177e4 LT |
2478 | return -1; |
2479 | ||
2480 | /* Reserve space for headers and set control bits. */ | |
2481 | skb_reserve(skb, MAX_TCP_HEADER); | |
2482 | skb->csum = 0; | |
2483 | TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; | |
2484 | TCP_SKB_CB(skb)->sacked = urgent; | |
7967168c HX |
2485 | skb_shinfo(skb)->gso_segs = 1; |
2486 | skb_shinfo(skb)->gso_size = 0; | |
2487 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2488 | |
2489 | /* Use a previous sequence. This should cause the other | |
2490 | * end to send an ack. Don't queue or clone SKB, just | |
2491 | * send it. | |
2492 | */ | |
2493 | TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; | |
2494 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; | |
2495 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2496 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2497 | } |
2498 | ||
2499 | int tcp_write_wakeup(struct sock *sk) | |
2500 | { | |
2501 | if (sk->sk_state != TCP_CLOSE) { | |
2502 | struct tcp_sock *tp = tcp_sk(sk); | |
2503 | struct sk_buff *skb; | |
2504 | ||
fe067e8a | 2505 | if ((skb = tcp_send_head(sk)) != NULL && |
1da177e4 LT |
2506 | before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { |
2507 | int err; | |
2508 | unsigned int mss = tcp_current_mss(sk, 0); | |
2509 | unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; | |
2510 | ||
2511 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2512 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2513 | ||
2514 | /* We are probing the opening of a window | |
2515 | * but the window size is != 0 | |
2516 | * must have been a result SWS avoidance ( sender ) | |
2517 | */ | |
2518 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2519 | skb->len > mss) { | |
2520 | seg_size = min(seg_size, mss); | |
2521 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
846998ae | 2522 | if (tcp_fragment(sk, skb, seg_size, mss)) |
1da177e4 | 2523 | return -1; |
1da177e4 | 2524 | } else if (!tcp_skb_pcount(skb)) |
846998ae | 2525 | tcp_set_skb_tso_segs(sk, skb, mss); |
1da177e4 LT |
2526 | |
2527 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
2528 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2529 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 | 2530 | if (!err) { |
9e412ba7 | 2531 | update_send_head(sk, skb); |
1da177e4 LT |
2532 | } |
2533 | return err; | |
2534 | } else { | |
2535 | if (tp->urg_mode && | |
2536 | between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) | |
2537 | tcp_xmit_probe_skb(sk, TCPCB_URG); | |
2538 | return tcp_xmit_probe_skb(sk, 0); | |
2539 | } | |
2540 | } | |
2541 | return -1; | |
2542 | } | |
2543 | ||
2544 | /* A window probe timeout has occurred. If window is not closed send | |
2545 | * a partial packet else a zero probe. | |
2546 | */ | |
2547 | void tcp_send_probe0(struct sock *sk) | |
2548 | { | |
463c84b9 | 2549 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2550 | struct tcp_sock *tp = tcp_sk(sk); |
2551 | int err; | |
2552 | ||
2553 | err = tcp_write_wakeup(sk); | |
2554 | ||
fe067e8a | 2555 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2556 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2557 | icsk->icsk_probes_out = 0; |
463c84b9 | 2558 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2559 | return; |
2560 | } | |
2561 | ||
2562 | if (err <= 0) { | |
463c84b9 ACM |
2563 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2564 | icsk->icsk_backoff++; | |
6687e988 | 2565 | icsk->icsk_probes_out++; |
e905a9ed | 2566 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2567 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2568 | TCP_RTO_MAX); | |
1da177e4 LT |
2569 | } else { |
2570 | /* If packet was not sent due to local congestion, | |
6687e988 | 2571 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2572 | * Let local senders to fight for local resources. |
2573 | * | |
2574 | * Use accumulated backoff yet. | |
2575 | */ | |
6687e988 ACM |
2576 | if (!icsk->icsk_probes_out) |
2577 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2578 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2579 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2580 | TCP_RESOURCE_PROBE_INTERVAL), |
2581 | TCP_RTO_MAX); | |
1da177e4 LT |
2582 | } |
2583 | } | |
2584 | ||
2585 | EXPORT_SYMBOL(tcp_connect); | |
2586 | EXPORT_SYMBOL(tcp_make_synack); | |
2587 | EXPORT_SYMBOL(tcp_simple_retransmit); | |
2588 | EXPORT_SYMBOL(tcp_sync_mss); | |
f4805ede | 2589 | EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); |
5d424d5a | 2590 | EXPORT_SYMBOL(tcp_mtup_init); |