]>
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> | |
43 | #include <linux/smp_lock.h> | |
44 | ||
45 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
46 | int sysctl_tcp_retrans_collapse = 1; | |
47 | ||
48 | /* This limits the percentage of the congestion window which we | |
49 | * will allow a single TSO frame to consume. Building TSO frames | |
50 | * which are too large can cause TCP streams to be bursty. | |
51 | */ | |
52 | int sysctl_tcp_tso_win_divisor = 8; | |
53 | ||
54 | static inline void update_send_head(struct sock *sk, struct tcp_sock *tp, | |
55 | struct sk_buff *skb) | |
56 | { | |
57 | sk->sk_send_head = skb->next; | |
58 | if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue) | |
59 | sk->sk_send_head = NULL; | |
60 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; | |
61 | tcp_packets_out_inc(sk, tp, skb); | |
62 | } | |
63 | ||
64 | /* SND.NXT, if window was not shrunk. | |
65 | * If window has been shrunk, what should we make? It is not clear at all. | |
66 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
67 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
68 | * invalid. OK, let's make this for now: | |
69 | */ | |
70 | static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp) | |
71 | { | |
72 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) | |
73 | return tp->snd_nxt; | |
74 | else | |
75 | return tp->snd_una+tp->snd_wnd; | |
76 | } | |
77 | ||
78 | /* Calculate mss to advertise in SYN segment. | |
79 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
80 | * | |
81 | * 1. It is independent of path mtu. | |
82 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
83 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
84 | * attached devices, because some buggy hosts are confused by | |
85 | * large MSS. | |
86 | * 4. We do not make 3, we advertise MSS, calculated from first | |
87 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
88 | * This may be overridden via information stored in routing table. | |
89 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
90 | * probably even Jumbo". | |
91 | */ | |
92 | static __u16 tcp_advertise_mss(struct sock *sk) | |
93 | { | |
94 | struct tcp_sock *tp = tcp_sk(sk); | |
95 | struct dst_entry *dst = __sk_dst_get(sk); | |
96 | int mss = tp->advmss; | |
97 | ||
98 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
99 | mss = dst_metric(dst, RTAX_ADVMSS); | |
100 | tp->advmss = mss; | |
101 | } | |
102 | ||
103 | return (__u16)mss; | |
104 | } | |
105 | ||
106 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
107 | * This is the first part of cwnd validation mechanism. */ | |
108 | static void tcp_cwnd_restart(struct tcp_sock *tp, struct dst_entry *dst) | |
109 | { | |
110 | s32 delta = tcp_time_stamp - tp->lsndtime; | |
111 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
112 | u32 cwnd = tp->snd_cwnd; | |
113 | ||
317a76f9 | 114 | tcp_ca_event(tp, CA_EVENT_CWND_RESTART); |
1da177e4 LT |
115 | |
116 | tp->snd_ssthresh = tcp_current_ssthresh(tp); | |
117 | restart_cwnd = min(restart_cwnd, cwnd); | |
118 | ||
119 | while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd) | |
120 | cwnd >>= 1; | |
121 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
122 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
123 | tp->snd_cwnd_used = 0; | |
124 | } | |
125 | ||
126 | static inline void tcp_event_data_sent(struct tcp_sock *tp, | |
127 | struct sk_buff *skb, struct sock *sk) | |
128 | { | |
129 | u32 now = tcp_time_stamp; | |
130 | ||
131 | if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto) | |
132 | tcp_cwnd_restart(tp, __sk_dst_get(sk)); | |
133 | ||
134 | tp->lsndtime = now; | |
135 | ||
136 | /* If it is a reply for ato after last received | |
137 | * packet, enter pingpong mode. | |
138 | */ | |
139 | if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato) | |
140 | tp->ack.pingpong = 1; | |
141 | } | |
142 | ||
fc6415bc | 143 | static __inline__ void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 LT |
144 | { |
145 | struct tcp_sock *tp = tcp_sk(sk); | |
146 | ||
fc6415bc | 147 | tcp_dec_quickack_mode(tp, pkts); |
1da177e4 LT |
148 | tcp_clear_xmit_timer(sk, TCP_TIME_DACK); |
149 | } | |
150 | ||
151 | /* Determine a window scaling and initial window to offer. | |
152 | * Based on the assumption that the given amount of space | |
153 | * will be offered. Store the results in the tp structure. | |
154 | * NOTE: for smooth operation initial space offering should | |
155 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
156 | * This MUST be enforced by all callers. | |
157 | */ | |
158 | void tcp_select_initial_window(int __space, __u32 mss, | |
159 | __u32 *rcv_wnd, __u32 *window_clamp, | |
160 | int wscale_ok, __u8 *rcv_wscale) | |
161 | { | |
162 | unsigned int space = (__space < 0 ? 0 : __space); | |
163 | ||
164 | /* If no clamp set the clamp to the max possible scaled window */ | |
165 | if (*window_clamp == 0) | |
166 | (*window_clamp) = (65535 << 14); | |
167 | space = min(*window_clamp, space); | |
168 | ||
169 | /* Quantize space offering to a multiple of mss if possible. */ | |
170 | if (space > mss) | |
171 | space = (space / mss) * mss; | |
172 | ||
173 | /* NOTE: offering an initial window larger than 32767 | |
174 | * will break some buggy TCP stacks. We try to be nice. | |
175 | * If we are not window scaling, then this truncates | |
176 | * our initial window offering to 32k. There should also | |
177 | * be a sysctl option to stop being nice. | |
178 | */ | |
179 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
180 | (*rcv_wscale) = 0; | |
181 | if (wscale_ok) { | |
182 | /* Set window scaling on max possible window | |
183 | * See RFC1323 for an explanation of the limit to 14 | |
184 | */ | |
185 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
186 | while (space > 65535 && (*rcv_wscale) < 14) { | |
187 | space >>= 1; | |
188 | (*rcv_wscale)++; | |
189 | } | |
190 | } | |
191 | ||
192 | /* Set initial window to value enough for senders, | |
193 | * following RFC1414. Senders, not following this RFC, | |
194 | * will be satisfied with 2. | |
195 | */ | |
196 | if (mss > (1<<*rcv_wscale)) { | |
197 | int init_cwnd = 4; | |
198 | if (mss > 1460*3) | |
199 | init_cwnd = 2; | |
200 | else if (mss > 1460) | |
201 | init_cwnd = 3; | |
202 | if (*rcv_wnd > init_cwnd*mss) | |
203 | *rcv_wnd = init_cwnd*mss; | |
204 | } | |
205 | ||
206 | /* Set the clamp no higher than max representable value */ | |
207 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
208 | } | |
209 | ||
210 | /* Chose a new window to advertise, update state in tcp_sock for the | |
211 | * socket, and return result with RFC1323 scaling applied. The return | |
212 | * value can be stuffed directly into th->window for an outgoing | |
213 | * frame. | |
214 | */ | |
215 | static __inline__ u16 tcp_select_window(struct sock *sk) | |
216 | { | |
217 | struct tcp_sock *tp = tcp_sk(sk); | |
218 | u32 cur_win = tcp_receive_window(tp); | |
219 | u32 new_win = __tcp_select_window(sk); | |
220 | ||
221 | /* Never shrink the offered window */ | |
222 | if(new_win < cur_win) { | |
223 | /* Danger Will Robinson! | |
224 | * Don't update rcv_wup/rcv_wnd here or else | |
225 | * we will not be able to advertise a zero | |
226 | * window in time. --DaveM | |
227 | * | |
228 | * Relax Will Robinson. | |
229 | */ | |
230 | new_win = cur_win; | |
231 | } | |
232 | tp->rcv_wnd = new_win; | |
233 | tp->rcv_wup = tp->rcv_nxt; | |
234 | ||
235 | /* Make sure we do not exceed the maximum possible | |
236 | * scaled window. | |
237 | */ | |
238 | if (!tp->rx_opt.rcv_wscale) | |
239 | new_win = min(new_win, MAX_TCP_WINDOW); | |
240 | else | |
241 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
242 | ||
243 | /* RFC1323 scaling applied */ | |
244 | new_win >>= tp->rx_opt.rcv_wscale; | |
245 | ||
246 | /* If we advertise zero window, disable fast path. */ | |
247 | if (new_win == 0) | |
248 | tp->pred_flags = 0; | |
249 | ||
250 | return new_win; | |
251 | } | |
252 | ||
253 | ||
254 | /* This routine actually transmits TCP packets queued in by | |
255 | * tcp_do_sendmsg(). This is used by both the initial | |
256 | * transmission and possible later retransmissions. | |
257 | * All SKB's seen here are completely headerless. It is our | |
258 | * job to build the TCP header, and pass the packet down to | |
259 | * IP so it can do the same plus pass the packet off to the | |
260 | * device. | |
261 | * | |
262 | * We are working here with either a clone of the original | |
263 | * SKB, or a fresh unique copy made by the retransmit engine. | |
264 | */ | |
265 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb) | |
266 | { | |
267 | if (skb != NULL) { | |
268 | struct inet_sock *inet = inet_sk(sk); | |
269 | struct tcp_sock *tp = tcp_sk(sk); | |
270 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
271 | int tcp_header_size = tp->tcp_header_len; | |
272 | struct tcphdr *th; | |
273 | int sysctl_flags; | |
274 | int err; | |
275 | ||
276 | BUG_ON(!tcp_skb_pcount(skb)); | |
277 | ||
278 | #define SYSCTL_FLAG_TSTAMPS 0x1 | |
279 | #define SYSCTL_FLAG_WSCALE 0x2 | |
280 | #define SYSCTL_FLAG_SACK 0x4 | |
281 | ||
317a76f9 SH |
282 | /* If congestion control is doing timestamping */ |
283 | if (tp->ca_ops->rtt_sample) | |
284 | do_gettimeofday(&skb->stamp); | |
285 | ||
1da177e4 LT |
286 | sysctl_flags = 0; |
287 | if (tcb->flags & TCPCB_FLAG_SYN) { | |
288 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; | |
289 | if(sysctl_tcp_timestamps) { | |
290 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; | |
291 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; | |
292 | } | |
293 | if(sysctl_tcp_window_scaling) { | |
294 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; | |
295 | sysctl_flags |= SYSCTL_FLAG_WSCALE; | |
296 | } | |
297 | if(sysctl_tcp_sack) { | |
298 | sysctl_flags |= SYSCTL_FLAG_SACK; | |
299 | if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) | |
300 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; | |
301 | } | |
302 | } else if (tp->rx_opt.eff_sacks) { | |
303 | /* A SACK is 2 pad bytes, a 2 byte header, plus | |
304 | * 2 32-bit sequence numbers for each SACK block. | |
305 | */ | |
306 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + | |
307 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
308 | } | |
309 | ||
317a76f9 SH |
310 | if (tcp_packets_in_flight(tp) == 0) |
311 | tcp_ca_event(tp, CA_EVENT_TX_START); | |
1da177e4 LT |
312 | |
313 | th = (struct tcphdr *) skb_push(skb, tcp_header_size); | |
314 | skb->h.th = th; | |
315 | skb_set_owner_w(skb, sk); | |
316 | ||
317 | /* Build TCP header and checksum it. */ | |
318 | th->source = inet->sport; | |
319 | th->dest = inet->dport; | |
320 | th->seq = htonl(tcb->seq); | |
321 | th->ack_seq = htonl(tp->rcv_nxt); | |
322 | *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags); | |
323 | if (tcb->flags & TCPCB_FLAG_SYN) { | |
324 | /* RFC1323: The window in SYN & SYN/ACK segments | |
325 | * is never scaled. | |
326 | */ | |
327 | th->window = htons(tp->rcv_wnd); | |
328 | } else { | |
329 | th->window = htons(tcp_select_window(sk)); | |
330 | } | |
331 | th->check = 0; | |
332 | th->urg_ptr = 0; | |
333 | ||
334 | if (tp->urg_mode && | |
335 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) { | |
336 | th->urg_ptr = htons(tp->snd_up-tcb->seq); | |
337 | th->urg = 1; | |
338 | } | |
339 | ||
340 | if (tcb->flags & TCPCB_FLAG_SYN) { | |
341 | tcp_syn_build_options((__u32 *)(th + 1), | |
342 | tcp_advertise_mss(sk), | |
343 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), | |
344 | (sysctl_flags & SYSCTL_FLAG_SACK), | |
345 | (sysctl_flags & SYSCTL_FLAG_WSCALE), | |
346 | tp->rx_opt.rcv_wscale, | |
347 | tcb->when, | |
348 | tp->rx_opt.ts_recent); | |
349 | } else { | |
350 | tcp_build_and_update_options((__u32 *)(th + 1), | |
351 | tp, tcb->when); | |
352 | ||
353 | TCP_ECN_send(sk, tp, skb, tcp_header_size); | |
354 | } | |
355 | tp->af_specific->send_check(sk, th, skb->len, skb); | |
356 | ||
357 | if (tcb->flags & TCPCB_FLAG_ACK) | |
fc6415bc | 358 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 LT |
359 | |
360 | if (skb->len != tcp_header_size) | |
361 | tcp_event_data_sent(tp, skb, sk); | |
362 | ||
363 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
364 | ||
365 | err = tp->af_specific->queue_xmit(skb, 0); | |
366 | if (err <= 0) | |
367 | return err; | |
368 | ||
369 | tcp_enter_cwr(tp); | |
370 | ||
371 | /* NET_XMIT_CN is special. It does not guarantee, | |
372 | * that this packet is lost. It tells that device | |
373 | * is about to start to drop packets or already | |
374 | * drops some packets of the same priority and | |
375 | * invokes us to send less aggressively. | |
376 | */ | |
377 | return err == NET_XMIT_CN ? 0 : err; | |
378 | } | |
379 | return -ENOBUFS; | |
380 | #undef SYSCTL_FLAG_TSTAMPS | |
381 | #undef SYSCTL_FLAG_WSCALE | |
382 | #undef SYSCTL_FLAG_SACK | |
383 | } | |
384 | ||
385 | ||
386 | /* This routine just queue's the buffer | |
387 | * | |
388 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
389 | * otherwise socket can stall. | |
390 | */ | |
391 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
392 | { | |
393 | struct tcp_sock *tp = tcp_sk(sk); | |
394 | ||
395 | /* Advance write_seq and place onto the write_queue. */ | |
396 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
397 | skb_header_release(skb); | |
398 | __skb_queue_tail(&sk->sk_write_queue, skb); | |
399 | sk_charge_skb(sk, skb); | |
400 | ||
401 | /* Queue it, remembering where we must start sending. */ | |
402 | if (sk->sk_send_head == NULL) | |
403 | sk->sk_send_head = skb; | |
404 | } | |
405 | ||
406 | static inline void tcp_tso_set_push(struct sk_buff *skb) | |
407 | { | |
408 | /* Force push to be on for any TSO frames to workaround | |
409 | * problems with busted implementations like Mac OS-X that | |
410 | * hold off socket receive wakeups until push is seen. | |
411 | */ | |
412 | if (tcp_skb_pcount(skb) > 1) | |
413 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
414 | } | |
415 | ||
f6302d1d DM |
416 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb) |
417 | { | |
418 | struct tcp_sock *tp = tcp_sk(sk); | |
419 | ||
420 | if (skb->len <= tp->mss_cache_std || | |
421 | !(sk->sk_route_caps & NETIF_F_TSO)) { | |
422 | /* Avoid the costly divide in the normal | |
423 | * non-TSO case. | |
424 | */ | |
425 | skb_shinfo(skb)->tso_segs = 1; | |
426 | skb_shinfo(skb)->tso_size = 0; | |
427 | } else { | |
428 | unsigned int factor; | |
429 | ||
430 | factor = skb->len + (tp->mss_cache_std - 1); | |
431 | factor /= tp->mss_cache_std; | |
432 | skb_shinfo(skb)->tso_segs = factor; | |
433 | skb_shinfo(skb)->tso_size = tp->mss_cache_std; | |
434 | } | |
435 | } | |
436 | ||
7f4dd0a9 DM |
437 | /* Does SKB fit into the send window? */ |
438 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) | |
439 | { | |
440 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
441 | ||
442 | return !after(end_seq, tp->snd_una + tp->snd_wnd); | |
443 | } | |
444 | ||
445 | /* Can at least one segment of SKB be sent right now, according to the | |
446 | * congestion window rules? If so, return how many segments are allowed. | |
447 | */ | |
448 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) | |
449 | { | |
450 | u32 in_flight, cwnd; | |
451 | ||
452 | /* Don't be strict about the congestion window for the final FIN. */ | |
453 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) | |
454 | return 1; | |
455 | ||
456 | in_flight = tcp_packets_in_flight(tp); | |
457 | cwnd = tp->snd_cwnd; | |
458 | if (in_flight < cwnd) | |
459 | return (cwnd - in_flight); | |
460 | ||
461 | return 0; | |
462 | } | |
463 | ||
f6302d1d DM |
464 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
465 | { | |
466 | return after(tp->snd_sml,tp->snd_una) && | |
467 | !after(tp->snd_sml, tp->snd_nxt); | |
468 | } | |
469 | ||
470 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
471 | * 1. It is full sized. | |
7f4dd0a9 | 472 | * 2. Or it contains FIN. (already checked by caller) |
f6302d1d DM |
473 | * 3. Or TCP_NODELAY was set. |
474 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
475 | * With Minshall's modification: all sent small packets are ACKed. | |
476 | */ | |
477 | ||
478 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
479 | const struct sk_buff *skb, | |
480 | unsigned mss_now, int nonagle) | |
481 | { | |
482 | return (skb->len < mss_now && | |
f6302d1d DM |
483 | ((nonagle&TCP_NAGLE_CORK) || |
484 | (!nonagle && | |
485 | tp->packets_out && | |
486 | tcp_minshall_check(tp)))); | |
487 | } | |
488 | ||
7f4dd0a9 DM |
489 | /* Return non-zero if the Nagle test allows this packet to be |
490 | * sent now. | |
f6302d1d | 491 | */ |
7f4dd0a9 DM |
492 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, |
493 | unsigned int cur_mss, int nonagle) | |
f6302d1d | 494 | { |
7f4dd0a9 DM |
495 | /* Nagle rule does not apply to frames, which sit in the middle of the |
496 | * write_queue (they have no chances to get new data). | |
497 | * | |
498 | * This is implemented in the callers, where they modify the 'nonagle' | |
499 | * argument based upon the location of SKB in the send queue. | |
500 | */ | |
501 | if (nonagle & TCP_NAGLE_PUSH) | |
502 | return 1; | |
503 | ||
504 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ | |
505 | if (tp->urg_mode || | |
506 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) | |
507 | return 1; | |
508 | ||
509 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
510 | return 1; | |
511 | ||
512 | return 0; | |
513 | } | |
f6302d1d | 514 | |
7f4dd0a9 DM |
515 | /* This must be invoked the first time we consider transmitting |
516 | * SKB onto the wire. | |
517 | */ | |
518 | static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb) | |
519 | { | |
520 | int tso_segs = tcp_skb_pcount(skb); | |
521 | ||
522 | if (!tso_segs) { | |
f6302d1d | 523 | tcp_set_skb_tso_segs(sk, skb); |
7f4dd0a9 | 524 | tso_segs = tcp_skb_pcount(skb); |
f6302d1d | 525 | } |
7f4dd0a9 DM |
526 | return tso_segs; |
527 | } | |
f6302d1d | 528 | |
7f4dd0a9 DM |
529 | /* This checks if the data bearing packet SKB (usually sk->sk_send_head) |
530 | * should be put on the wire right now. If so, it returns the number of | |
531 | * packets allowed by the congestion window. | |
532 | */ | |
533 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
534 | unsigned int cur_mss, int nonagle) | |
535 | { | |
536 | struct tcp_sock *tp = tcp_sk(sk); | |
537 | unsigned int cwnd_quota; | |
f6302d1d | 538 | |
7f4dd0a9 DM |
539 | tcp_init_tso_segs(sk, skb); |
540 | ||
541 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
542 | return 0; | |
543 | ||
544 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
545 | if (cwnd_quota && | |
546 | !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
547 | cwnd_quota = 0; | |
548 | ||
549 | return cwnd_quota; | |
f6302d1d DM |
550 | } |
551 | ||
552 | static inline int tcp_skb_is_last(const struct sock *sk, | |
553 | const struct sk_buff *skb) | |
554 | { | |
555 | return skb->next == (struct sk_buff *)&sk->sk_write_queue; | |
556 | } | |
557 | ||
f6302d1d DM |
558 | int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp) |
559 | { | |
560 | struct sk_buff *skb = sk->sk_send_head; | |
561 | ||
562 | return (skb && | |
563 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
564 | (tcp_skb_is_last(sk, skb) ? | |
565 | TCP_NAGLE_PUSH : | |
566 | tp->nonagle))); | |
567 | } | |
568 | ||
569 | ||
1da177e4 LT |
570 | /* Send _single_ skb sitting at the send head. This function requires |
571 | * true push pending frames to setup probe timer etc. | |
572 | */ | |
573 | void tcp_push_one(struct sock *sk, unsigned cur_mss) | |
574 | { | |
575 | struct tcp_sock *tp = tcp_sk(sk); | |
576 | struct sk_buff *skb = sk->sk_send_head; | |
577 | ||
d5ac99a6 | 578 | if (tcp_snd_test(sk, skb, cur_mss, TCP_NAGLE_PUSH)) { |
1da177e4 LT |
579 | /* Send it out now. */ |
580 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
581 | tcp_tso_set_push(skb); | |
582 | if (!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation))) { | |
583 | sk->sk_send_head = NULL; | |
584 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; | |
585 | tcp_packets_out_inc(sk, tp, skb); | |
586 | return; | |
587 | } | |
588 | } | |
589 | } | |
590 | ||
1da177e4 LT |
591 | /* Function to create two new TCP segments. Shrinks the given segment |
592 | * to the specified size and appends a new segment with the rest of the | |
593 | * packet to the list. This won't be called frequently, I hope. | |
594 | * Remember, these are still headerless SKBs at this point. | |
595 | */ | |
596 | static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len) | |
597 | { | |
598 | struct tcp_sock *tp = tcp_sk(sk); | |
599 | struct sk_buff *buff; | |
600 | int nsize; | |
601 | u16 flags; | |
602 | ||
603 | nsize = skb_headlen(skb) - len; | |
604 | if (nsize < 0) | |
605 | nsize = 0; | |
606 | ||
607 | if (skb_cloned(skb) && | |
608 | skb_is_nonlinear(skb) && | |
609 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
610 | return -ENOMEM; | |
611 | ||
612 | /* Get a new skb... force flag on. */ | |
613 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
614 | if (buff == NULL) | |
615 | return -ENOMEM; /* We'll just try again later. */ | |
616 | sk_charge_skb(sk, buff); | |
617 | ||
618 | /* Correct the sequence numbers. */ | |
619 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
620 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
621 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
622 | ||
623 | /* PSH and FIN should only be set in the second packet. */ | |
624 | flags = TCP_SKB_CB(skb)->flags; | |
625 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
626 | TCP_SKB_CB(buff)->flags = flags; | |
627 | TCP_SKB_CB(buff)->sacked = | |
628 | (TCP_SKB_CB(skb)->sacked & | |
629 | (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL)); | |
630 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; | |
631 | ||
632 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) { | |
633 | /* Copy and checksum data tail into the new buffer. */ | |
634 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), | |
635 | nsize, 0); | |
636 | ||
637 | skb_trim(skb, len); | |
638 | ||
639 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
640 | } else { | |
641 | skb->ip_summed = CHECKSUM_HW; | |
642 | skb_split(skb, buff, len); | |
643 | } | |
644 | ||
645 | buff->ip_summed = skb->ip_summed; | |
646 | ||
647 | /* Looks stupid, but our code really uses when of | |
648 | * skbs, which it never sent before. --ANK | |
649 | */ | |
650 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
317a76f9 | 651 | buff->stamp = skb->stamp; |
1da177e4 LT |
652 | |
653 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { | |
654 | tp->lost_out -= tcp_skb_pcount(skb); | |
655 | tp->left_out -= tcp_skb_pcount(skb); | |
656 | } | |
657 | ||
658 | /* Fix up tso_factor for both original and new SKB. */ | |
d5ac99a6 DM |
659 | tcp_set_skb_tso_segs(sk, skb); |
660 | tcp_set_skb_tso_segs(sk, buff); | |
1da177e4 LT |
661 | |
662 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { | |
663 | tp->lost_out += tcp_skb_pcount(skb); | |
664 | tp->left_out += tcp_skb_pcount(skb); | |
665 | } | |
666 | ||
667 | if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) { | |
668 | tp->lost_out += tcp_skb_pcount(buff); | |
669 | tp->left_out += tcp_skb_pcount(buff); | |
670 | } | |
671 | ||
672 | /* Link BUFF into the send queue. */ | |
f44b5271 | 673 | skb_header_release(buff); |
1da177e4 LT |
674 | __skb_append(skb, buff); |
675 | ||
676 | return 0; | |
677 | } | |
678 | ||
679 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
680 | * eventually). The difference is that pulled data not copied, but | |
681 | * immediately discarded. | |
682 | */ | |
683 | static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len) | |
684 | { | |
685 | int i, k, eat; | |
686 | ||
687 | eat = len; | |
688 | k = 0; | |
689 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { | |
690 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
691 | put_page(skb_shinfo(skb)->frags[i].page); | |
692 | eat -= skb_shinfo(skb)->frags[i].size; | |
693 | } else { | |
694 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
695 | if (eat) { | |
696 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
697 | skb_shinfo(skb)->frags[k].size -= eat; | |
698 | eat = 0; | |
699 | } | |
700 | k++; | |
701 | } | |
702 | } | |
703 | skb_shinfo(skb)->nr_frags = k; | |
704 | ||
705 | skb->tail = skb->data; | |
706 | skb->data_len -= len; | |
707 | skb->len = skb->data_len; | |
708 | return skb->tail; | |
709 | } | |
710 | ||
711 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
712 | { | |
713 | if (skb_cloned(skb) && | |
714 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
715 | return -ENOMEM; | |
716 | ||
717 | if (len <= skb_headlen(skb)) { | |
718 | __skb_pull(skb, len); | |
719 | } else { | |
720 | if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL) | |
721 | return -ENOMEM; | |
722 | } | |
723 | ||
724 | TCP_SKB_CB(skb)->seq += len; | |
725 | skb->ip_summed = CHECKSUM_HW; | |
726 | ||
727 | skb->truesize -= len; | |
728 | sk->sk_wmem_queued -= len; | |
729 | sk->sk_forward_alloc += len; | |
730 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
731 | ||
732 | /* Any change of skb->len requires recalculation of tso | |
733 | * factor and mss. | |
734 | */ | |
735 | if (tcp_skb_pcount(skb) > 1) | |
d5ac99a6 | 736 | tcp_set_skb_tso_segs(sk, skb); |
1da177e4 LT |
737 | |
738 | return 0; | |
739 | } | |
740 | ||
741 | /* This function synchronize snd mss to current pmtu/exthdr set. | |
742 | ||
743 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
744 | for TCP options, but includes only bare TCP header. | |
745 | ||
746 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
747 | It is minumum of user_mss and mss received with SYN. | |
748 | It also does not include TCP options. | |
749 | ||
750 | tp->pmtu_cookie is last pmtu, seen by this function. | |
751 | ||
752 | tp->mss_cache is current effective sending mss, including | |
753 | all tcp options except for SACKs. It is evaluated, | |
754 | taking into account current pmtu, but never exceeds | |
755 | tp->rx_opt.mss_clamp. | |
756 | ||
757 | NOTE1. rfc1122 clearly states that advertised MSS | |
758 | DOES NOT include either tcp or ip options. | |
759 | ||
760 | NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside | |
761 | this function. --ANK (980731) | |
762 | */ | |
763 | ||
764 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
765 | { | |
766 | struct tcp_sock *tp = tcp_sk(sk); | |
767 | int mss_now; | |
768 | ||
769 | /* Calculate base mss without TCP options: | |
770 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
771 | */ | |
772 | mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr); | |
773 | ||
774 | /* Clamp it (mss_clamp does not include tcp options) */ | |
775 | if (mss_now > tp->rx_opt.mss_clamp) | |
776 | mss_now = tp->rx_opt.mss_clamp; | |
777 | ||
778 | /* Now subtract optional transport overhead */ | |
779 | mss_now -= tp->ext_header_len; | |
780 | ||
781 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
782 | if (mss_now < 48) | |
783 | mss_now = 48; | |
784 | ||
785 | /* Now subtract TCP options size, not including SACKs */ | |
786 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
787 | ||
788 | /* Bound mss with half of window */ | |
789 | if (tp->max_window && mss_now > (tp->max_window>>1)) | |
790 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); | |
791 | ||
792 | /* And store cached results */ | |
793 | tp->pmtu_cookie = pmtu; | |
794 | tp->mss_cache = tp->mss_cache_std = mss_now; | |
795 | ||
796 | return mss_now; | |
797 | } | |
798 | ||
799 | /* Compute the current effective MSS, taking SACKs and IP options, | |
800 | * and even PMTU discovery events into account. | |
801 | * | |
802 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
803 | * cannot be large. However, taking into account rare use of URG, this | |
804 | * is not a big flaw. | |
805 | */ | |
806 | ||
807 | unsigned int tcp_current_mss(struct sock *sk, int large) | |
808 | { | |
809 | struct tcp_sock *tp = tcp_sk(sk); | |
810 | struct dst_entry *dst = __sk_dst_get(sk); | |
811 | unsigned int do_large, mss_now; | |
812 | ||
813 | mss_now = tp->mss_cache_std; | |
814 | if (dst) { | |
815 | u32 mtu = dst_mtu(dst); | |
816 | if (mtu != tp->pmtu_cookie) | |
817 | mss_now = tcp_sync_mss(sk, mtu); | |
818 | } | |
819 | ||
820 | do_large = (large && | |
821 | (sk->sk_route_caps & NETIF_F_TSO) && | |
822 | !tp->urg_mode); | |
823 | ||
824 | if (do_large) { | |
825 | unsigned int large_mss, factor, limit; | |
826 | ||
827 | large_mss = 65535 - tp->af_specific->net_header_len - | |
828 | tp->ext_header_len - tp->tcp_header_len; | |
829 | ||
830 | if (tp->max_window && large_mss > (tp->max_window>>1)) | |
831 | large_mss = max((tp->max_window>>1), | |
832 | 68U - tp->tcp_header_len); | |
833 | ||
834 | factor = large_mss / mss_now; | |
835 | ||
836 | /* Always keep large mss multiple of real mss, but | |
837 | * do not exceed 1/tso_win_divisor of the congestion window | |
838 | * so we can keep the ACK clock ticking and minimize | |
839 | * bursting. | |
840 | */ | |
841 | limit = tp->snd_cwnd; | |
842 | if (sysctl_tcp_tso_win_divisor) | |
843 | limit /= sysctl_tcp_tso_win_divisor; | |
844 | limit = max(1U, limit); | |
845 | if (factor > limit) | |
846 | factor = limit; | |
847 | ||
848 | tp->mss_cache = mss_now * factor; | |
849 | ||
850 | mss_now = tp->mss_cache; | |
851 | } | |
852 | ||
853 | if (tp->rx_opt.eff_sacks) | |
854 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + | |
855 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
856 | return mss_now; | |
857 | } | |
858 | ||
a762a980 DM |
859 | /* Congestion window validation. (RFC2861) */ |
860 | ||
861 | static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp) | |
862 | { | |
863 | __u32 packets_out = tp->packets_out; | |
864 | ||
865 | if (packets_out >= tp->snd_cwnd) { | |
866 | /* Network is feed fully. */ | |
867 | tp->snd_cwnd_used = 0; | |
868 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
869 | } else { | |
870 | /* Network starves. */ | |
871 | if (tp->packets_out > tp->snd_cwnd_used) | |
872 | tp->snd_cwnd_used = tp->packets_out; | |
873 | ||
874 | if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= tp->rto) | |
875 | tcp_cwnd_application_limited(sk); | |
876 | } | |
877 | } | |
878 | ||
1da177e4 LT |
879 | /* This routine writes packets to the network. It advances the |
880 | * send_head. This happens as incoming acks open up the remote | |
881 | * window for us. | |
882 | * | |
883 | * Returns 1, if no segments are in flight and we have queued segments, but | |
884 | * cannot send anything now because of SWS or another problem. | |
885 | */ | |
a2e2a59c | 886 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) |
1da177e4 LT |
887 | { |
888 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 889 | struct sk_buff *skb; |
92df7b51 | 890 | int sent_pkts; |
1da177e4 LT |
891 | |
892 | /* If we are closed, the bytes will have to remain here. | |
893 | * In time closedown will finish, we empty the write queue and all | |
894 | * will be happy. | |
895 | */ | |
92df7b51 DM |
896 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
897 | return 0; | |
1da177e4 | 898 | |
92df7b51 DM |
899 | sent_pkts = 0; |
900 | while ((skb = sk->sk_send_head) && | |
901 | tcp_snd_test(sk, skb, mss_now, | |
902 | tcp_skb_is_last(sk, skb) ? nonagle : | |
903 | TCP_NAGLE_PUSH)) { | |
904 | if (skb->len > mss_now) { | |
905 | if (tcp_fragment(sk, skb, mss_now)) | |
1da177e4 | 906 | break; |
92df7b51 | 907 | } |
1da177e4 | 908 | |
92df7b51 DM |
909 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
910 | tcp_tso_set_push(skb); | |
911 | if (tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))) | |
912 | break; | |
1da177e4 | 913 | |
92df7b51 DM |
914 | /* Advance the send_head. This one is sent out. |
915 | * This call will increment packets_out. | |
916 | */ | |
917 | update_send_head(sk, tp, skb); | |
1da177e4 | 918 | |
92df7b51 DM |
919 | tcp_minshall_update(tp, mss_now, skb); |
920 | sent_pkts = 1; | |
921 | } | |
1da177e4 | 922 | |
92df7b51 DM |
923 | if (sent_pkts) { |
924 | tcp_cwnd_validate(sk, tp); | |
925 | return 0; | |
1da177e4 | 926 | } |
92df7b51 DM |
927 | |
928 | return !tp->packets_out && sk->sk_send_head; | |
1da177e4 LT |
929 | } |
930 | ||
a762a980 DM |
931 | /* Push out any pending frames which were held back due to |
932 | * TCP_CORK or attempt at coalescing tiny packets. | |
933 | * The socket must be locked by the caller. | |
934 | */ | |
935 | void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp, | |
a2e2a59c | 936 | unsigned int cur_mss, int nonagle) |
a762a980 DM |
937 | { |
938 | struct sk_buff *skb = sk->sk_send_head; | |
939 | ||
940 | if (skb) { | |
55c97f3e | 941 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
a762a980 DM |
942 | tcp_check_probe_timer(sk, tp); |
943 | } | |
944 | } | |
945 | ||
1da177e4 LT |
946 | /* This function returns the amount that we can raise the |
947 | * usable window based on the following constraints | |
948 | * | |
949 | * 1. The window can never be shrunk once it is offered (RFC 793) | |
950 | * 2. We limit memory per socket | |
951 | * | |
952 | * RFC 1122: | |
953 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
954 | * RECV.NEXT + RCV.WIN fixed until: | |
955 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
956 | * | |
957 | * i.e. don't raise the right edge of the window until you can raise | |
958 | * it at least MSS bytes. | |
959 | * | |
960 | * Unfortunately, the recommended algorithm breaks header prediction, | |
961 | * since header prediction assumes th->window stays fixed. | |
962 | * | |
963 | * Strictly speaking, keeping th->window fixed violates the receiver | |
964 | * side SWS prevention criteria. The problem is that under this rule | |
965 | * a stream of single byte packets will cause the right side of the | |
966 | * window to always advance by a single byte. | |
967 | * | |
968 | * Of course, if the sender implements sender side SWS prevention | |
969 | * then this will not be a problem. | |
970 | * | |
971 | * BSD seems to make the following compromise: | |
972 | * | |
973 | * If the free space is less than the 1/4 of the maximum | |
974 | * space available and the free space is less than 1/2 mss, | |
975 | * then set the window to 0. | |
976 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
977 | * Otherwise, just prevent the window from shrinking | |
978 | * and from being larger than the largest representable value. | |
979 | * | |
980 | * This prevents incremental opening of the window in the regime | |
981 | * where TCP is limited by the speed of the reader side taking | |
982 | * data out of the TCP receive queue. It does nothing about | |
983 | * those cases where the window is constrained on the sender side | |
984 | * because the pipeline is full. | |
985 | * | |
986 | * BSD also seems to "accidentally" limit itself to windows that are a | |
987 | * multiple of MSS, at least until the free space gets quite small. | |
988 | * This would appear to be a side effect of the mbuf implementation. | |
989 | * Combining these two algorithms results in the observed behavior | |
990 | * of having a fixed window size at almost all times. | |
991 | * | |
992 | * Below we obtain similar behavior by forcing the offered window to | |
993 | * a multiple of the mss when it is feasible to do so. | |
994 | * | |
995 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
996 | * Regular options like TIMESTAMP are taken into account. | |
997 | */ | |
998 | u32 __tcp_select_window(struct sock *sk) | |
999 | { | |
1000 | struct tcp_sock *tp = tcp_sk(sk); | |
1001 | /* MSS for the peer's data. Previous verions used mss_clamp | |
1002 | * here. I don't know if the value based on our guesses | |
1003 | * of peer's MSS is better for the performance. It's more correct | |
1004 | * but may be worse for the performance because of rcv_mss | |
1005 | * fluctuations. --SAW 1998/11/1 | |
1006 | */ | |
1007 | int mss = tp->ack.rcv_mss; | |
1008 | int free_space = tcp_space(sk); | |
1009 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1010 | int window; | |
1011 | ||
1012 | if (mss > full_space) | |
1013 | mss = full_space; | |
1014 | ||
1015 | if (free_space < full_space/2) { | |
1016 | tp->ack.quick = 0; | |
1017 | ||
1018 | if (tcp_memory_pressure) | |
1019 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); | |
1020 | ||
1021 | if (free_space < mss) | |
1022 | return 0; | |
1023 | } | |
1024 | ||
1025 | if (free_space > tp->rcv_ssthresh) | |
1026 | free_space = tp->rcv_ssthresh; | |
1027 | ||
1028 | /* Don't do rounding if we are using window scaling, since the | |
1029 | * scaled window will not line up with the MSS boundary anyway. | |
1030 | */ | |
1031 | window = tp->rcv_wnd; | |
1032 | if (tp->rx_opt.rcv_wscale) { | |
1033 | window = free_space; | |
1034 | ||
1035 | /* Advertise enough space so that it won't get scaled away. | |
1036 | * Import case: prevent zero window announcement if | |
1037 | * 1<<rcv_wscale > mss. | |
1038 | */ | |
1039 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1040 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1041 | << tp->rx_opt.rcv_wscale); | |
1042 | } else { | |
1043 | /* Get the largest window that is a nice multiple of mss. | |
1044 | * Window clamp already applied above. | |
1045 | * If our current window offering is within 1 mss of the | |
1046 | * free space we just keep it. This prevents the divide | |
1047 | * and multiply from happening most of the time. | |
1048 | * We also don't do any window rounding when the free space | |
1049 | * is too small. | |
1050 | */ | |
1051 | if (window <= free_space - mss || window > free_space) | |
1052 | window = (free_space/mss)*mss; | |
1053 | } | |
1054 | ||
1055 | return window; | |
1056 | } | |
1057 | ||
1058 | /* Attempt to collapse two adjacent SKB's during retransmission. */ | |
1059 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) | |
1060 | { | |
1061 | struct tcp_sock *tp = tcp_sk(sk); | |
1062 | struct sk_buff *next_skb = skb->next; | |
1063 | ||
1064 | /* The first test we must make is that neither of these two | |
1065 | * SKB's are still referenced by someone else. | |
1066 | */ | |
1067 | if (!skb_cloned(skb) && !skb_cloned(next_skb)) { | |
1068 | int skb_size = skb->len, next_skb_size = next_skb->len; | |
1069 | u16 flags = TCP_SKB_CB(skb)->flags; | |
1070 | ||
1071 | /* Also punt if next skb has been SACK'd. */ | |
1072 | if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) | |
1073 | return; | |
1074 | ||
1075 | /* Next skb is out of window. */ | |
1076 | if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) | |
1077 | return; | |
1078 | ||
1079 | /* Punt if not enough space exists in the first SKB for | |
1080 | * the data in the second, or the total combined payload | |
1081 | * would exceed the MSS. | |
1082 | */ | |
1083 | if ((next_skb_size > skb_tailroom(skb)) || | |
1084 | ((skb_size + next_skb_size) > mss_now)) | |
1085 | return; | |
1086 | ||
1087 | BUG_ON(tcp_skb_pcount(skb) != 1 || | |
1088 | tcp_skb_pcount(next_skb) != 1); | |
1089 | ||
1090 | /* Ok. We will be able to collapse the packet. */ | |
1091 | __skb_unlink(next_skb, next_skb->list); | |
1092 | ||
1093 | memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size); | |
1094 | ||
1095 | if (next_skb->ip_summed == CHECKSUM_HW) | |
1096 | skb->ip_summed = CHECKSUM_HW; | |
1097 | ||
1098 | if (skb->ip_summed != CHECKSUM_HW) | |
1099 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1100 | ||
1101 | /* Update sequence range on original skb. */ | |
1102 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1103 | ||
1104 | /* Merge over control information. */ | |
1105 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ | |
1106 | TCP_SKB_CB(skb)->flags = flags; | |
1107 | ||
1108 | /* All done, get rid of second SKB and account for it so | |
1109 | * packet counting does not break. | |
1110 | */ | |
1111 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); | |
1112 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) | |
1113 | tp->retrans_out -= tcp_skb_pcount(next_skb); | |
1114 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { | |
1115 | tp->lost_out -= tcp_skb_pcount(next_skb); | |
1116 | tp->left_out -= tcp_skb_pcount(next_skb); | |
1117 | } | |
1118 | /* Reno case is special. Sigh... */ | |
1119 | if (!tp->rx_opt.sack_ok && tp->sacked_out) { | |
1120 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); | |
1121 | tp->left_out -= tcp_skb_pcount(next_skb); | |
1122 | } | |
1123 | ||
1124 | /* Not quite right: it can be > snd.fack, but | |
1125 | * it is better to underestimate fackets. | |
1126 | */ | |
1127 | tcp_dec_pcount_approx(&tp->fackets_out, next_skb); | |
1128 | tcp_packets_out_dec(tp, next_skb); | |
1129 | sk_stream_free_skb(sk, next_skb); | |
1130 | } | |
1131 | } | |
1132 | ||
1133 | /* Do a simple retransmit without using the backoff mechanisms in | |
1134 | * tcp_timer. This is used for path mtu discovery. | |
1135 | * The socket is already locked here. | |
1136 | */ | |
1137 | void tcp_simple_retransmit(struct sock *sk) | |
1138 | { | |
1139 | struct tcp_sock *tp = tcp_sk(sk); | |
1140 | struct sk_buff *skb; | |
1141 | unsigned int mss = tcp_current_mss(sk, 0); | |
1142 | int lost = 0; | |
1143 | ||
1144 | sk_stream_for_retrans_queue(skb, sk) { | |
1145 | if (skb->len > mss && | |
1146 | !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { | |
1147 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1148 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1149 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1150 | } | |
1151 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { | |
1152 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1153 | tp->lost_out += tcp_skb_pcount(skb); | |
1154 | lost = 1; | |
1155 | } | |
1156 | } | |
1157 | } | |
1158 | ||
1159 | if (!lost) | |
1160 | return; | |
1161 | ||
1162 | tcp_sync_left_out(tp); | |
1163 | ||
1164 | /* Don't muck with the congestion window here. | |
1165 | * Reason is that we do not increase amount of _data_ | |
1166 | * in network, but units changed and effective | |
1167 | * cwnd/ssthresh really reduced now. | |
1168 | */ | |
1169 | if (tp->ca_state != TCP_CA_Loss) { | |
1170 | tp->high_seq = tp->snd_nxt; | |
1171 | tp->snd_ssthresh = tcp_current_ssthresh(tp); | |
1172 | tp->prior_ssthresh = 0; | |
1173 | tp->undo_marker = 0; | |
1174 | tcp_set_ca_state(tp, TCP_CA_Loss); | |
1175 | } | |
1176 | tcp_xmit_retransmit_queue(sk); | |
1177 | } | |
1178 | ||
1179 | /* This retransmits one SKB. Policy decisions and retransmit queue | |
1180 | * state updates are done by the caller. Returns non-zero if an | |
1181 | * error occurred which prevented the send. | |
1182 | */ | |
1183 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
1184 | { | |
1185 | struct tcp_sock *tp = tcp_sk(sk); | |
1186 | unsigned int cur_mss = tcp_current_mss(sk, 0); | |
1187 | int err; | |
1188 | ||
1189 | /* Do not sent more than we queued. 1/4 is reserved for possible | |
1190 | * copying overhead: frgagmentation, tunneling, mangling etc. | |
1191 | */ | |
1192 | if (atomic_read(&sk->sk_wmem_alloc) > | |
1193 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
1194 | return -EAGAIN; | |
1195 | ||
1196 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
1197 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1198 | BUG(); | |
1199 | ||
1200 | if (sk->sk_route_caps & NETIF_F_TSO) { | |
1201 | sk->sk_route_caps &= ~NETIF_F_TSO; | |
1202 | sock_set_flag(sk, SOCK_NO_LARGESEND); | |
1203 | tp->mss_cache = tp->mss_cache_std; | |
1204 | } | |
1205 | ||
1206 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) | |
1207 | return -ENOMEM; | |
1208 | } | |
1209 | ||
1210 | /* If receiver has shrunk his window, and skb is out of | |
1211 | * new window, do not retransmit it. The exception is the | |
1212 | * case, when window is shrunk to zero. In this case | |
1213 | * our retransmit serves as a zero window probe. | |
1214 | */ | |
1215 | if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) | |
1216 | && TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1217 | return -EAGAIN; | |
1218 | ||
1219 | if (skb->len > cur_mss) { | |
1220 | int old_factor = tcp_skb_pcount(skb); | |
1221 | int new_factor; | |
1222 | ||
1223 | if (tcp_fragment(sk, skb, cur_mss)) | |
1224 | return -ENOMEM; /* We'll try again later. */ | |
1225 | ||
1226 | /* New SKB created, account for it. */ | |
1227 | new_factor = tcp_skb_pcount(skb); | |
1228 | tp->packets_out -= old_factor - new_factor; | |
1229 | tp->packets_out += tcp_skb_pcount(skb->next); | |
1230 | } | |
1231 | ||
1232 | /* Collapse two adjacent packets if worthwhile and we can. */ | |
1233 | if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && | |
1234 | (skb->len < (cur_mss >> 1)) && | |
1235 | (skb->next != sk->sk_send_head) && | |
1236 | (skb->next != (struct sk_buff *)&sk->sk_write_queue) && | |
1237 | (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) && | |
1238 | (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) && | |
1239 | (sysctl_tcp_retrans_collapse != 0)) | |
1240 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1241 | ||
1242 | if(tp->af_specific->rebuild_header(sk)) | |
1243 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
1244 | ||
1245 | /* Some Solaris stacks overoptimize and ignore the FIN on a | |
1246 | * retransmit when old data is attached. So strip it off | |
1247 | * since it is cheap to do so and saves bytes on the network. | |
1248 | */ | |
1249 | if(skb->len > 0 && | |
1250 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && | |
1251 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { | |
1252 | if (!pskb_trim(skb, 0)) { | |
1253 | TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; | |
1254 | skb_shinfo(skb)->tso_segs = 1; | |
1255 | skb_shinfo(skb)->tso_size = 0; | |
1256 | skb->ip_summed = CHECKSUM_NONE; | |
1257 | skb->csum = 0; | |
1258 | } | |
1259 | } | |
1260 | ||
1261 | /* Make a copy, if the first transmission SKB clone we made | |
1262 | * is still in somebody's hands, else make a clone. | |
1263 | */ | |
1264 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1265 | tcp_tso_set_push(skb); | |
1266 | ||
1267 | err = tcp_transmit_skb(sk, (skb_cloned(skb) ? | |
1268 | pskb_copy(skb, GFP_ATOMIC): | |
1269 | skb_clone(skb, GFP_ATOMIC))); | |
1270 | ||
1271 | if (err == 0) { | |
1272 | /* Update global TCP statistics. */ | |
1273 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); | |
1274 | ||
1275 | tp->total_retrans++; | |
1276 | ||
1277 | #if FASTRETRANS_DEBUG > 0 | |
1278 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1279 | if (net_ratelimit()) | |
1280 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
1281 | } | |
1282 | #endif | |
1283 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
1284 | tp->retrans_out += tcp_skb_pcount(skb); | |
1285 | ||
1286 | /* Save stamp of the first retransmit. */ | |
1287 | if (!tp->retrans_stamp) | |
1288 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
1289 | ||
1290 | tp->undo_retrans++; | |
1291 | ||
1292 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
1293 | * see tcp_input.c tcp_sacktag_write_queue(). | |
1294 | */ | |
1295 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
1296 | } | |
1297 | return err; | |
1298 | } | |
1299 | ||
1300 | /* This gets called after a retransmit timeout, and the initially | |
1301 | * retransmitted data is acknowledged. It tries to continue | |
1302 | * resending the rest of the retransmit queue, until either | |
1303 | * we've sent it all or the congestion window limit is reached. | |
1304 | * If doing SACK, the first ACK which comes back for a timeout | |
1305 | * based retransmit packet might feed us FACK information again. | |
1306 | * If so, we use it to avoid unnecessarily retransmissions. | |
1307 | */ | |
1308 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
1309 | { | |
1310 | struct tcp_sock *tp = tcp_sk(sk); | |
1311 | struct sk_buff *skb; | |
1312 | int packet_cnt = tp->lost_out; | |
1313 | ||
1314 | /* First pass: retransmit lost packets. */ | |
1315 | if (packet_cnt) { | |
1316 | sk_stream_for_retrans_queue(skb, sk) { | |
1317 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1318 | ||
1319 | /* Assume this retransmit will generate | |
1320 | * only one packet for congestion window | |
1321 | * calculation purposes. This works because | |
1322 | * tcp_retransmit_skb() will chop up the | |
1323 | * packet to be MSS sized and all the | |
1324 | * packet counting works out. | |
1325 | */ | |
1326 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
1327 | return; | |
1328 | ||
1329 | if (sacked&TCPCB_LOST) { | |
1330 | if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { | |
1331 | if (tcp_retransmit_skb(sk, skb)) | |
1332 | return; | |
1333 | if (tp->ca_state != TCP_CA_Loss) | |
1334 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); | |
1335 | else | |
1336 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); | |
1337 | ||
1338 | if (skb == | |
1339 | skb_peek(&sk->sk_write_queue)) | |
1340 | tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); | |
1341 | } | |
1342 | ||
1343 | packet_cnt -= tcp_skb_pcount(skb); | |
1344 | if (packet_cnt <= 0) | |
1345 | break; | |
1346 | } | |
1347 | } | |
1348 | } | |
1349 | ||
1350 | /* OK, demanded retransmission is finished. */ | |
1351 | ||
1352 | /* Forward retransmissions are possible only during Recovery. */ | |
1353 | if (tp->ca_state != TCP_CA_Recovery) | |
1354 | return; | |
1355 | ||
1356 | /* No forward retransmissions in Reno are possible. */ | |
1357 | if (!tp->rx_opt.sack_ok) | |
1358 | return; | |
1359 | ||
1360 | /* Yeah, we have to make difficult choice between forward transmission | |
1361 | * and retransmission... Both ways have their merits... | |
1362 | * | |
1363 | * For now we do not retransmit anything, while we have some new | |
1364 | * segments to send. | |
1365 | */ | |
1366 | ||
1367 | if (tcp_may_send_now(sk, tp)) | |
1368 | return; | |
1369 | ||
1370 | packet_cnt = 0; | |
1371 | ||
1372 | sk_stream_for_retrans_queue(skb, sk) { | |
1373 | /* Similar to the retransmit loop above we | |
1374 | * can pretend that the retransmitted SKB | |
1375 | * we send out here will be composed of one | |
1376 | * real MSS sized packet because tcp_retransmit_skb() | |
1377 | * will fragment it if necessary. | |
1378 | */ | |
1379 | if (++packet_cnt > tp->fackets_out) | |
1380 | break; | |
1381 | ||
1382 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
1383 | break; | |
1384 | ||
1385 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) | |
1386 | continue; | |
1387 | ||
1388 | /* Ok, retransmit it. */ | |
1389 | if (tcp_retransmit_skb(sk, skb)) | |
1390 | break; | |
1391 | ||
1392 | if (skb == skb_peek(&sk->sk_write_queue)) | |
1393 | tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); | |
1394 | ||
1395 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); | |
1396 | } | |
1397 | } | |
1398 | ||
1399 | ||
1400 | /* Send a fin. The caller locks the socket for us. This cannot be | |
1401 | * allowed to fail queueing a FIN frame under any circumstances. | |
1402 | */ | |
1403 | void tcp_send_fin(struct sock *sk) | |
1404 | { | |
1405 | struct tcp_sock *tp = tcp_sk(sk); | |
1406 | struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue); | |
1407 | int mss_now; | |
1408 | ||
1409 | /* Optimization, tack on the FIN if we have a queue of | |
1410 | * unsent frames. But be careful about outgoing SACKS | |
1411 | * and IP options. | |
1412 | */ | |
1413 | mss_now = tcp_current_mss(sk, 1); | |
1414 | ||
1415 | if (sk->sk_send_head != NULL) { | |
1416 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; | |
1417 | TCP_SKB_CB(skb)->end_seq++; | |
1418 | tp->write_seq++; | |
1419 | } else { | |
1420 | /* Socket is locked, keep trying until memory is available. */ | |
1421 | for (;;) { | |
1422 | skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL); | |
1423 | if (skb) | |
1424 | break; | |
1425 | yield(); | |
1426 | } | |
1427 | ||
1428 | /* Reserve space for headers and prepare control bits. */ | |
1429 | skb_reserve(skb, MAX_TCP_HEADER); | |
1430 | skb->csum = 0; | |
1431 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); | |
1432 | TCP_SKB_CB(skb)->sacked = 0; | |
1433 | skb_shinfo(skb)->tso_segs = 1; | |
1434 | skb_shinfo(skb)->tso_size = 0; | |
1435 | ||
1436 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ | |
1437 | TCP_SKB_CB(skb)->seq = tp->write_seq; | |
1438 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; | |
1439 | tcp_queue_skb(sk, skb); | |
1440 | } | |
1441 | __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF); | |
1442 | } | |
1443 | ||
1444 | /* We get here when a process closes a file descriptor (either due to | |
1445 | * an explicit close() or as a byproduct of exit()'ing) and there | |
1446 | * was unread data in the receive queue. This behavior is recommended | |
1447 | * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM | |
1448 | */ | |
1449 | void tcp_send_active_reset(struct sock *sk, int priority) | |
1450 | { | |
1451 | struct tcp_sock *tp = tcp_sk(sk); | |
1452 | struct sk_buff *skb; | |
1453 | ||
1454 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
1455 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
1456 | if (!skb) { | |
1457 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); | |
1458 | return; | |
1459 | } | |
1460 | ||
1461 | /* Reserve space for headers and prepare control bits. */ | |
1462 | skb_reserve(skb, MAX_TCP_HEADER); | |
1463 | skb->csum = 0; | |
1464 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); | |
1465 | TCP_SKB_CB(skb)->sacked = 0; | |
1466 | skb_shinfo(skb)->tso_segs = 1; | |
1467 | skb_shinfo(skb)->tso_size = 0; | |
1468 | ||
1469 | /* Send it off. */ | |
1470 | TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp); | |
1471 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; | |
1472 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1473 | if (tcp_transmit_skb(sk, skb)) | |
1474 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); | |
1475 | } | |
1476 | ||
1477 | /* WARNING: This routine must only be called when we have already sent | |
1478 | * a SYN packet that crossed the incoming SYN that caused this routine | |
1479 | * to get called. If this assumption fails then the initial rcv_wnd | |
1480 | * and rcv_wscale values will not be correct. | |
1481 | */ | |
1482 | int tcp_send_synack(struct sock *sk) | |
1483 | { | |
1484 | struct sk_buff* skb; | |
1485 | ||
1486 | skb = skb_peek(&sk->sk_write_queue); | |
1487 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { | |
1488 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); | |
1489 | return -EFAULT; | |
1490 | } | |
1491 | if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { | |
1492 | if (skb_cloned(skb)) { | |
1493 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
1494 | if (nskb == NULL) | |
1495 | return -ENOMEM; | |
1496 | __skb_unlink(skb, &sk->sk_write_queue); | |
1497 | skb_header_release(nskb); | |
1498 | __skb_queue_head(&sk->sk_write_queue, nskb); | |
1499 | sk_stream_free_skb(sk, skb); | |
1500 | sk_charge_skb(sk, nskb); | |
1501 | skb = nskb; | |
1502 | } | |
1503 | ||
1504 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; | |
1505 | TCP_ECN_send_synack(tcp_sk(sk), skb); | |
1506 | } | |
1507 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1508 | return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); | |
1509 | } | |
1510 | ||
1511 | /* | |
1512 | * Prepare a SYN-ACK. | |
1513 | */ | |
1514 | struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, | |
60236fdd | 1515 | struct request_sock *req) |
1da177e4 | 1516 | { |
2e6599cb | 1517 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
1518 | struct tcp_sock *tp = tcp_sk(sk); |
1519 | struct tcphdr *th; | |
1520 | int tcp_header_size; | |
1521 | struct sk_buff *skb; | |
1522 | ||
1523 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); | |
1524 | if (skb == NULL) | |
1525 | return NULL; | |
1526 | ||
1527 | /* Reserve space for headers. */ | |
1528 | skb_reserve(skb, MAX_TCP_HEADER); | |
1529 | ||
1530 | skb->dst = dst_clone(dst); | |
1531 | ||
1532 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + | |
2e6599cb ACM |
1533 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
1534 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + | |
1da177e4 | 1535 | /* SACK_PERM is in the place of NOP NOP of TS */ |
2e6599cb | 1536 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
1da177e4 LT |
1537 | skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size); |
1538 | ||
1539 | memset(th, 0, sizeof(struct tcphdr)); | |
1540 | th->syn = 1; | |
1541 | th->ack = 1; | |
1542 | if (dst->dev->features&NETIF_F_TSO) | |
2e6599cb | 1543 | ireq->ecn_ok = 0; |
1da177e4 LT |
1544 | TCP_ECN_make_synack(req, th); |
1545 | th->source = inet_sk(sk)->sport; | |
2e6599cb ACM |
1546 | th->dest = ireq->rmt_port; |
1547 | TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; | |
1da177e4 LT |
1548 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; |
1549 | TCP_SKB_CB(skb)->sacked = 0; | |
1550 | skb_shinfo(skb)->tso_segs = 1; | |
1551 | skb_shinfo(skb)->tso_size = 0; | |
1552 | th->seq = htonl(TCP_SKB_CB(skb)->seq); | |
2e6599cb | 1553 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
1554 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
1555 | __u8 rcv_wscale; | |
1556 | /* Set this up on the first call only */ | |
1557 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
1558 | /* tcp_full_space because it is guaranteed to be the first packet */ | |
1559 | tcp_select_initial_window(tcp_full_space(sk), | |
2e6599cb | 1560 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
1da177e4 LT |
1561 | &req->rcv_wnd, |
1562 | &req->window_clamp, | |
2e6599cb | 1563 | ireq->wscale_ok, |
1da177e4 | 1564 | &rcv_wscale); |
2e6599cb | 1565 | ireq->rcv_wscale = rcv_wscale; |
1da177e4 LT |
1566 | } |
1567 | ||
1568 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
1569 | th->window = htons(req->rcv_wnd); | |
1570 | ||
1571 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
2e6599cb ACM |
1572 | tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, |
1573 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, | |
1da177e4 LT |
1574 | TCP_SKB_CB(skb)->when, |
1575 | req->ts_recent); | |
1576 | ||
1577 | skb->csum = 0; | |
1578 | th->doff = (tcp_header_size >> 2); | |
1579 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
1580 | return skb; | |
1581 | } | |
1582 | ||
1583 | /* | |
1584 | * Do all connect socket setups that can be done AF independent. | |
1585 | */ | |
1586 | static inline void tcp_connect_init(struct sock *sk) | |
1587 | { | |
1588 | struct dst_entry *dst = __sk_dst_get(sk); | |
1589 | struct tcp_sock *tp = tcp_sk(sk); | |
1590 | __u8 rcv_wscale; | |
1591 | ||
1592 | /* We'll fix this up when we get a response from the other end. | |
1593 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
1594 | */ | |
1595 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
1596 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); | |
1597 | ||
1598 | /* If user gave his TCP_MAXSEG, record it to clamp */ | |
1599 | if (tp->rx_opt.user_mss) | |
1600 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1601 | tp->max_window = 0; | |
1602 | tcp_sync_mss(sk, dst_mtu(dst)); | |
1603 | ||
1604 | if (!tp->window_clamp) | |
1605 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
1606 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
1607 | tcp_initialize_rcv_mss(sk); | |
1da177e4 LT |
1608 | |
1609 | tcp_select_initial_window(tcp_full_space(sk), | |
1610 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
1611 | &tp->rcv_wnd, | |
1612 | &tp->window_clamp, | |
1613 | sysctl_tcp_window_scaling, | |
1614 | &rcv_wscale); | |
1615 | ||
1616 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
1617 | tp->rcv_ssthresh = tp->rcv_wnd; | |
1618 | ||
1619 | sk->sk_err = 0; | |
1620 | sock_reset_flag(sk, SOCK_DONE); | |
1621 | tp->snd_wnd = 0; | |
1622 | tcp_init_wl(tp, tp->write_seq, 0); | |
1623 | tp->snd_una = tp->write_seq; | |
1624 | tp->snd_sml = tp->write_seq; | |
1625 | tp->rcv_nxt = 0; | |
1626 | tp->rcv_wup = 0; | |
1627 | tp->copied_seq = 0; | |
1628 | ||
1629 | tp->rto = TCP_TIMEOUT_INIT; | |
1630 | tp->retransmits = 0; | |
1631 | tcp_clear_retrans(tp); | |
1632 | } | |
1633 | ||
1634 | /* | |
1635 | * Build a SYN and send it off. | |
1636 | */ | |
1637 | int tcp_connect(struct sock *sk) | |
1638 | { | |
1639 | struct tcp_sock *tp = tcp_sk(sk); | |
1640 | struct sk_buff *buff; | |
1641 | ||
1642 | tcp_connect_init(sk); | |
1643 | ||
1644 | buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation); | |
1645 | if (unlikely(buff == NULL)) | |
1646 | return -ENOBUFS; | |
1647 | ||
1648 | /* Reserve space for headers. */ | |
1649 | skb_reserve(buff, MAX_TCP_HEADER); | |
1650 | ||
1651 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; | |
1652 | TCP_ECN_send_syn(sk, tp, buff); | |
1653 | TCP_SKB_CB(buff)->sacked = 0; | |
1654 | skb_shinfo(buff)->tso_segs = 1; | |
1655 | skb_shinfo(buff)->tso_size = 0; | |
1656 | buff->csum = 0; | |
1657 | TCP_SKB_CB(buff)->seq = tp->write_seq++; | |
1658 | TCP_SKB_CB(buff)->end_seq = tp->write_seq; | |
1659 | tp->snd_nxt = tp->write_seq; | |
1660 | tp->pushed_seq = tp->write_seq; | |
1da177e4 LT |
1661 | |
1662 | /* Send it off. */ | |
1663 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
1664 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
1665 | skb_header_release(buff); | |
1666 | __skb_queue_tail(&sk->sk_write_queue, buff); | |
1667 | sk_charge_skb(sk, buff); | |
1668 | tp->packets_out += tcp_skb_pcount(buff); | |
1669 | tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL)); | |
1670 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); | |
1671 | ||
1672 | /* Timer for repeating the SYN until an answer. */ | |
1673 | tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); | |
1674 | return 0; | |
1675 | } | |
1676 | ||
1677 | /* Send out a delayed ack, the caller does the policy checking | |
1678 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
1679 | * for details. | |
1680 | */ | |
1681 | void tcp_send_delayed_ack(struct sock *sk) | |
1682 | { | |
1683 | struct tcp_sock *tp = tcp_sk(sk); | |
1684 | int ato = tp->ack.ato; | |
1685 | unsigned long timeout; | |
1686 | ||
1687 | if (ato > TCP_DELACK_MIN) { | |
1688 | int max_ato = HZ/2; | |
1689 | ||
1690 | if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED)) | |
1691 | max_ato = TCP_DELACK_MAX; | |
1692 | ||
1693 | /* Slow path, intersegment interval is "high". */ | |
1694 | ||
1695 | /* If some rtt estimate is known, use it to bound delayed ack. | |
1696 | * Do not use tp->rto here, use results of rtt measurements | |
1697 | * directly. | |
1698 | */ | |
1699 | if (tp->srtt) { | |
1700 | int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); | |
1701 | ||
1702 | if (rtt < max_ato) | |
1703 | max_ato = rtt; | |
1704 | } | |
1705 | ||
1706 | ato = min(ato, max_ato); | |
1707 | } | |
1708 | ||
1709 | /* Stay within the limit we were given */ | |
1710 | timeout = jiffies + ato; | |
1711 | ||
1712 | /* Use new timeout only if there wasn't a older one earlier. */ | |
1713 | if (tp->ack.pending&TCP_ACK_TIMER) { | |
1714 | /* If delack timer was blocked or is about to expire, | |
1715 | * send ACK now. | |
1716 | */ | |
1717 | if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) { | |
1718 | tcp_send_ack(sk); | |
1719 | return; | |
1720 | } | |
1721 | ||
1722 | if (!time_before(timeout, tp->ack.timeout)) | |
1723 | timeout = tp->ack.timeout; | |
1724 | } | |
1725 | tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER; | |
1726 | tp->ack.timeout = timeout; | |
1727 | sk_reset_timer(sk, &tp->delack_timer, timeout); | |
1728 | } | |
1729 | ||
1730 | /* This routine sends an ack and also updates the window. */ | |
1731 | void tcp_send_ack(struct sock *sk) | |
1732 | { | |
1733 | /* If we have been reset, we may not send again. */ | |
1734 | if (sk->sk_state != TCP_CLOSE) { | |
1735 | struct tcp_sock *tp = tcp_sk(sk); | |
1736 | struct sk_buff *buff; | |
1737 | ||
1738 | /* We are not putting this on the write queue, so | |
1739 | * tcp_transmit_skb() will set the ownership to this | |
1740 | * sock. | |
1741 | */ | |
1742 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
1743 | if (buff == NULL) { | |
1744 | tcp_schedule_ack(tp); | |
1745 | tp->ack.ato = TCP_ATO_MIN; | |
1746 | tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX); | |
1747 | return; | |
1748 | } | |
1749 | ||
1750 | /* Reserve space for headers and prepare control bits. */ | |
1751 | skb_reserve(buff, MAX_TCP_HEADER); | |
1752 | buff->csum = 0; | |
1753 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; | |
1754 | TCP_SKB_CB(buff)->sacked = 0; | |
1755 | skb_shinfo(buff)->tso_segs = 1; | |
1756 | skb_shinfo(buff)->tso_size = 0; | |
1757 | ||
1758 | /* Send it off, this clears delayed acks for us. */ | |
1759 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp); | |
1760 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
1761 | tcp_transmit_skb(sk, buff); | |
1762 | } | |
1763 | } | |
1764 | ||
1765 | /* This routine sends a packet with an out of date sequence | |
1766 | * number. It assumes the other end will try to ack it. | |
1767 | * | |
1768 | * Question: what should we make while urgent mode? | |
1769 | * 4.4BSD forces sending single byte of data. We cannot send | |
1770 | * out of window data, because we have SND.NXT==SND.MAX... | |
1771 | * | |
1772 | * Current solution: to send TWO zero-length segments in urgent mode: | |
1773 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
1774 | * out-of-date with SND.UNA-1 to probe window. | |
1775 | */ | |
1776 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
1777 | { | |
1778 | struct tcp_sock *tp = tcp_sk(sk); | |
1779 | struct sk_buff *skb; | |
1780 | ||
1781 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
1782 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
1783 | if (skb == NULL) | |
1784 | return -1; | |
1785 | ||
1786 | /* Reserve space for headers and set control bits. */ | |
1787 | skb_reserve(skb, MAX_TCP_HEADER); | |
1788 | skb->csum = 0; | |
1789 | TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; | |
1790 | TCP_SKB_CB(skb)->sacked = urgent; | |
1791 | skb_shinfo(skb)->tso_segs = 1; | |
1792 | skb_shinfo(skb)->tso_size = 0; | |
1793 | ||
1794 | /* Use a previous sequence. This should cause the other | |
1795 | * end to send an ack. Don't queue or clone SKB, just | |
1796 | * send it. | |
1797 | */ | |
1798 | TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; | |
1799 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; | |
1800 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1801 | return tcp_transmit_skb(sk, skb); | |
1802 | } | |
1803 | ||
1804 | int tcp_write_wakeup(struct sock *sk) | |
1805 | { | |
1806 | if (sk->sk_state != TCP_CLOSE) { | |
1807 | struct tcp_sock *tp = tcp_sk(sk); | |
1808 | struct sk_buff *skb; | |
1809 | ||
1810 | if ((skb = sk->sk_send_head) != NULL && | |
1811 | before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { | |
1812 | int err; | |
1813 | unsigned int mss = tcp_current_mss(sk, 0); | |
1814 | unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; | |
1815 | ||
1816 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
1817 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
1818 | ||
1819 | /* We are probing the opening of a window | |
1820 | * but the window size is != 0 | |
1821 | * must have been a result SWS avoidance ( sender ) | |
1822 | */ | |
1823 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
1824 | skb->len > mss) { | |
1825 | seg_size = min(seg_size, mss); | |
1826 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
1827 | if (tcp_fragment(sk, skb, seg_size)) | |
1828 | return -1; | |
1829 | /* SWS override triggered forced fragmentation. | |
1830 | * Disable TSO, the connection is too sick. */ | |
1831 | if (sk->sk_route_caps & NETIF_F_TSO) { | |
1832 | sock_set_flag(sk, SOCK_NO_LARGESEND); | |
1833 | sk->sk_route_caps &= ~NETIF_F_TSO; | |
1834 | tp->mss_cache = tp->mss_cache_std; | |
1835 | } | |
1836 | } else if (!tcp_skb_pcount(skb)) | |
d5ac99a6 | 1837 | tcp_set_skb_tso_segs(sk, skb); |
1da177e4 LT |
1838 | |
1839 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
1840 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1841 | tcp_tso_set_push(skb); | |
1842 | err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); | |
1843 | if (!err) { | |
1844 | update_send_head(sk, tp, skb); | |
1845 | } | |
1846 | return err; | |
1847 | } else { | |
1848 | if (tp->urg_mode && | |
1849 | between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) | |
1850 | tcp_xmit_probe_skb(sk, TCPCB_URG); | |
1851 | return tcp_xmit_probe_skb(sk, 0); | |
1852 | } | |
1853 | } | |
1854 | return -1; | |
1855 | } | |
1856 | ||
1857 | /* A window probe timeout has occurred. If window is not closed send | |
1858 | * a partial packet else a zero probe. | |
1859 | */ | |
1860 | void tcp_send_probe0(struct sock *sk) | |
1861 | { | |
1862 | struct tcp_sock *tp = tcp_sk(sk); | |
1863 | int err; | |
1864 | ||
1865 | err = tcp_write_wakeup(sk); | |
1866 | ||
1867 | if (tp->packets_out || !sk->sk_send_head) { | |
1868 | /* Cancel probe timer, if it is not required. */ | |
1869 | tp->probes_out = 0; | |
1870 | tp->backoff = 0; | |
1871 | return; | |
1872 | } | |
1873 | ||
1874 | if (err <= 0) { | |
1875 | if (tp->backoff < sysctl_tcp_retries2) | |
1876 | tp->backoff++; | |
1877 | tp->probes_out++; | |
1878 | tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, | |
1879 | min(tp->rto << tp->backoff, TCP_RTO_MAX)); | |
1880 | } else { | |
1881 | /* If packet was not sent due to local congestion, | |
1882 | * do not backoff and do not remember probes_out. | |
1883 | * Let local senders to fight for local resources. | |
1884 | * | |
1885 | * Use accumulated backoff yet. | |
1886 | */ | |
1887 | if (!tp->probes_out) | |
1888 | tp->probes_out=1; | |
1889 | tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, | |
1890 | min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL)); | |
1891 | } | |
1892 | } | |
1893 | ||
1894 | EXPORT_SYMBOL(tcp_connect); | |
1895 | EXPORT_SYMBOL(tcp_make_synack); | |
1896 | EXPORT_SYMBOL(tcp_simple_retransmit); | |
1897 | EXPORT_SYMBOL(tcp_sync_mss); |