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 *
10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu>
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
114 if (tcp_is_vegas(tp))
115 tcp_vegas_enable(tp);
116
117 tp->snd_ssthresh = tcp_current_ssthresh(tp);
118 restart_cwnd = min(restart_cwnd, cwnd);
119
120 while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd)
121 cwnd >>= 1;
122 tp->snd_cwnd = max(cwnd, restart_cwnd);
123 tp->snd_cwnd_stamp = tcp_time_stamp;
124 tp->snd_cwnd_used = 0;
125 }
126
127 static inline void tcp_event_data_sent(struct tcp_sock *tp,
128 struct sk_buff *skb, struct sock *sk)
129 {
130 u32 now = tcp_time_stamp;
131
132 if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto)
133 tcp_cwnd_restart(tp, __sk_dst_get(sk));
134
135 tp->lsndtime = now;
136
137 /* If it is a reply for ato after last received
138 * packet, enter pingpong mode.
139 */
140 if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato)
141 tp->ack.pingpong = 1;
142 }
143
144 static __inline__ void tcp_event_ack_sent(struct sock *sk)
145 {
146 struct tcp_sock *tp = tcp_sk(sk);
147
148 tcp_dec_quickack_mode(tp);
149 tcp_clear_xmit_timer(sk, TCP_TIME_DACK);
150 }
151
152 /* Determine a window scaling and initial window to offer.
153 * Based on the assumption that the given amount of space
154 * will be offered. Store the results in the tp structure.
155 * NOTE: for smooth operation initial space offering should
156 * be a multiple of mss if possible. We assume here that mss >= 1.
157 * This MUST be enforced by all callers.
158 */
159 void tcp_select_initial_window(int __space, __u32 mss,
160 __u32 *rcv_wnd, __u32 *window_clamp,
161 int wscale_ok, __u8 *rcv_wscale)
162 {
163 unsigned int space = (__space < 0 ? 0 : __space);
164
165 /* If no clamp set the clamp to the max possible scaled window */
166 if (*window_clamp == 0)
167 (*window_clamp) = (65535 << 14);
168 space = min(*window_clamp, space);
169
170 /* Quantize space offering to a multiple of mss if possible. */
171 if (space > mss)
172 space = (space / mss) * mss;
173
174 /* NOTE: offering an initial window larger than 32767
175 * will break some buggy TCP stacks. We try to be nice.
176 * If we are not window scaling, then this truncates
177 * our initial window offering to 32k. There should also
178 * be a sysctl option to stop being nice.
179 */
180 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
181 (*rcv_wscale) = 0;
182 if (wscale_ok) {
183 /* Set window scaling on max possible window
184 * See RFC1323 for an explanation of the limit to 14
185 */
186 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
187 while (space > 65535 && (*rcv_wscale) < 14) {
188 space >>= 1;
189 (*rcv_wscale)++;
190 }
191 }
192
193 /* Set initial window to value enough for senders,
194 * following RFC1414. Senders, not following this RFC,
195 * will be satisfied with 2.
196 */
197 if (mss > (1<<*rcv_wscale)) {
198 int init_cwnd = 4;
199 if (mss > 1460*3)
200 init_cwnd = 2;
201 else if (mss > 1460)
202 init_cwnd = 3;
203 if (*rcv_wnd > init_cwnd*mss)
204 *rcv_wnd = init_cwnd*mss;
205 }
206
207 /* Set the clamp no higher than max representable value */
208 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
209 }
210
211 /* Chose a new window to advertise, update state in tcp_sock for the
212 * socket, and return result with RFC1323 scaling applied. The return
213 * value can be stuffed directly into th->window for an outgoing
214 * frame.
215 */
216 static __inline__ u16 tcp_select_window(struct sock *sk)
217 {
218 struct tcp_sock *tp = tcp_sk(sk);
219 u32 cur_win = tcp_receive_window(tp);
220 u32 new_win = __tcp_select_window(sk);
221
222 /* Never shrink the offered window */
223 if(new_win < cur_win) {
224 /* Danger Will Robinson!
225 * Don't update rcv_wup/rcv_wnd here or else
226 * we will not be able to advertise a zero
227 * window in time. --DaveM
228 *
229 * Relax Will Robinson.
230 */
231 new_win = cur_win;
232 }
233 tp->rcv_wnd = new_win;
234 tp->rcv_wup = tp->rcv_nxt;
235
236 /* Make sure we do not exceed the maximum possible
237 * scaled window.
238 */
239 if (!tp->rx_opt.rcv_wscale)
240 new_win = min(new_win, MAX_TCP_WINDOW);
241 else
242 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
243
244 /* RFC1323 scaling applied */
245 new_win >>= tp->rx_opt.rcv_wscale;
246
247 /* If we advertise zero window, disable fast path. */
248 if (new_win == 0)
249 tp->pred_flags = 0;
250
251 return new_win;
252 }
253
254
255 /* This routine actually transmits TCP packets queued in by
256 * tcp_do_sendmsg(). This is used by both the initial
257 * transmission and possible later retransmissions.
258 * All SKB's seen here are completely headerless. It is our
259 * job to build the TCP header, and pass the packet down to
260 * IP so it can do the same plus pass the packet off to the
261 * device.
262 *
263 * We are working here with either a clone of the original
264 * SKB, or a fresh unique copy made by the retransmit engine.
265 */
266 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
267 {
268 if (skb != NULL) {
269 struct inet_sock *inet = inet_sk(sk);
270 struct tcp_sock *tp = tcp_sk(sk);
271 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
272 int tcp_header_size = tp->tcp_header_len;
273 struct tcphdr *th;
274 int sysctl_flags;
275 int err;
276
277 BUG_ON(!tcp_skb_pcount(skb));
278
279 #define SYSCTL_FLAG_TSTAMPS 0x1
280 #define SYSCTL_FLAG_WSCALE 0x2
281 #define SYSCTL_FLAG_SACK 0x4
282
283 sysctl_flags = 0;
284 if (tcb->flags & TCPCB_FLAG_SYN) {
285 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
286 if(sysctl_tcp_timestamps) {
287 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
288 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
289 }
290 if(sysctl_tcp_window_scaling) {
291 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
292 sysctl_flags |= SYSCTL_FLAG_WSCALE;
293 }
294 if(sysctl_tcp_sack) {
295 sysctl_flags |= SYSCTL_FLAG_SACK;
296 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
297 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
298 }
299 } else if (tp->rx_opt.eff_sacks) {
300 /* A SACK is 2 pad bytes, a 2 byte header, plus
301 * 2 32-bit sequence numbers for each SACK block.
302 */
303 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
304 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
305 }
306
307 /*
308 * If the connection is idle and we are restarting,
309 * then we don't want to do any Vegas calculations
310 * until we get fresh RTT samples. So when we
311 * restart, we reset our Vegas state to a clean
312 * slate. After we get acks for this flight of
313 * packets, _then_ we can make Vegas calculations
314 * again.
315 */
316 if (tcp_is_vegas(tp) && tcp_packets_in_flight(tp) == 0)
317 tcp_vegas_enable(tp);
318
319 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
320 skb->h.th = th;
321 skb_set_owner_w(skb, sk);
322
323 /* Build TCP header and checksum it. */
324 th->source = inet->sport;
325 th->dest = inet->dport;
326 th->seq = htonl(tcb->seq);
327 th->ack_seq = htonl(tp->rcv_nxt);
328 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
329 if (tcb->flags & TCPCB_FLAG_SYN) {
330 /* RFC1323: The window in SYN & SYN/ACK segments
331 * is never scaled.
332 */
333 th->window = htons(tp->rcv_wnd);
334 } else {
335 th->window = htons(tcp_select_window(sk));
336 }
337 th->check = 0;
338 th->urg_ptr = 0;
339
340 if (tp->urg_mode &&
341 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
342 th->urg_ptr = htons(tp->snd_up-tcb->seq);
343 th->urg = 1;
344 }
345
346 if (tcb->flags & TCPCB_FLAG_SYN) {
347 tcp_syn_build_options((__u32 *)(th + 1),
348 tcp_advertise_mss(sk),
349 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
350 (sysctl_flags & SYSCTL_FLAG_SACK),
351 (sysctl_flags & SYSCTL_FLAG_WSCALE),
352 tp->rx_opt.rcv_wscale,
353 tcb->when,
354 tp->rx_opt.ts_recent);
355 } else {
356 tcp_build_and_update_options((__u32 *)(th + 1),
357 tp, tcb->when);
358
359 TCP_ECN_send(sk, tp, skb, tcp_header_size);
360 }
361 tp->af_specific->send_check(sk, th, skb->len, skb);
362
363 if (tcb->flags & TCPCB_FLAG_ACK)
364 tcp_event_ack_sent(sk);
365
366 if (skb->len != tcp_header_size)
367 tcp_event_data_sent(tp, skb, sk);
368
369 TCP_INC_STATS(TCP_MIB_OUTSEGS);
370
371 err = tp->af_specific->queue_xmit(skb, 0);
372 if (err <= 0)
373 return err;
374
375 tcp_enter_cwr(tp);
376
377 /* NET_XMIT_CN is special. It does not guarantee,
378 * that this packet is lost. It tells that device
379 * is about to start to drop packets or already
380 * drops some packets of the same priority and
381 * invokes us to send less aggressively.
382 */
383 return err == NET_XMIT_CN ? 0 : err;
384 }
385 return -ENOBUFS;
386 #undef SYSCTL_FLAG_TSTAMPS
387 #undef SYSCTL_FLAG_WSCALE
388 #undef SYSCTL_FLAG_SACK
389 }
390
391
392 /* This routine just queue's the buffer
393 *
394 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
395 * otherwise socket can stall.
396 */
397 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
398 {
399 struct tcp_sock *tp = tcp_sk(sk);
400
401 /* Advance write_seq and place onto the write_queue. */
402 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
403 __skb_queue_tail(&sk->sk_write_queue, skb);
404 sk_charge_skb(sk, skb);
405
406 /* Queue it, remembering where we must start sending. */
407 if (sk->sk_send_head == NULL)
408 sk->sk_send_head = skb;
409 }
410
411 static inline void tcp_tso_set_push(struct sk_buff *skb)
412 {
413 /* Force push to be on for any TSO frames to workaround
414 * problems with busted implementations like Mac OS-X that
415 * hold off socket receive wakeups until push is seen.
416 */
417 if (tcp_skb_pcount(skb) > 1)
418 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
419 }
420
421 /* Send _single_ skb sitting at the send head. This function requires
422 * true push pending frames to setup probe timer etc.
423 */
424 void tcp_push_one(struct sock *sk, unsigned cur_mss)
425 {
426 struct tcp_sock *tp = tcp_sk(sk);
427 struct sk_buff *skb = sk->sk_send_head;
428
429 if (tcp_snd_test(tp, skb, cur_mss, TCP_NAGLE_PUSH)) {
430 /* Send it out now. */
431 TCP_SKB_CB(skb)->when = tcp_time_stamp;
432 tcp_tso_set_push(skb);
433 if (!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation))) {
434 sk->sk_send_head = NULL;
435 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
436 tcp_packets_out_inc(sk, tp, skb);
437 return;
438 }
439 }
440 }
441
442 void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_std)
443 {
444 if (skb->len <= mss_std) {
445 /* Avoid the costly divide in the normal
446 * non-TSO case.
447 */
448 skb_shinfo(skb)->tso_segs = 1;
449 skb_shinfo(skb)->tso_size = 0;
450 } else {
451 unsigned int factor;
452
453 factor = skb->len + (mss_std - 1);
454 factor /= mss_std;
455 skb_shinfo(skb)->tso_segs = factor;
456 skb_shinfo(skb)->tso_size = mss_std;
457 }
458 }
459
460 /* Function to create two new TCP segments. Shrinks the given segment
461 * to the specified size and appends a new segment with the rest of the
462 * packet to the list. This won't be called frequently, I hope.
463 * Remember, these are still headerless SKBs at this point.
464 */
465 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len)
466 {
467 struct tcp_sock *tp = tcp_sk(sk);
468 struct sk_buff *buff;
469 int nsize;
470 u16 flags;
471
472 nsize = skb_headlen(skb) - len;
473 if (nsize < 0)
474 nsize = 0;
475
476 if (skb_cloned(skb) &&
477 skb_is_nonlinear(skb) &&
478 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
479 return -ENOMEM;
480
481 /* Get a new skb... force flag on. */
482 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
483 if (buff == NULL)
484 return -ENOMEM; /* We'll just try again later. */
485 sk_charge_skb(sk, buff);
486
487 /* Correct the sequence numbers. */
488 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
489 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
490 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
491
492 /* PSH and FIN should only be set in the second packet. */
493 flags = TCP_SKB_CB(skb)->flags;
494 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
495 TCP_SKB_CB(buff)->flags = flags;
496 TCP_SKB_CB(buff)->sacked =
497 (TCP_SKB_CB(skb)->sacked &
498 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
499 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
500
501 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
502 /* Copy and checksum data tail into the new buffer. */
503 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
504 nsize, 0);
505
506 skb_trim(skb, len);
507
508 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
509 } else {
510 skb->ip_summed = CHECKSUM_HW;
511 skb_split(skb, buff, len);
512 }
513
514 buff->ip_summed = skb->ip_summed;
515
516 /* Looks stupid, but our code really uses when of
517 * skbs, which it never sent before. --ANK
518 */
519 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
520
521 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
522 tp->lost_out -= tcp_skb_pcount(skb);
523 tp->left_out -= tcp_skb_pcount(skb);
524 }
525
526 /* Fix up tso_factor for both original and new SKB. */
527 tcp_set_skb_tso_segs(skb, tp->mss_cache_std);
528 tcp_set_skb_tso_segs(buff, tp->mss_cache_std);
529
530 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
531 tp->lost_out += tcp_skb_pcount(skb);
532 tp->left_out += tcp_skb_pcount(skb);
533 }
534
535 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) {
536 tp->lost_out += tcp_skb_pcount(buff);
537 tp->left_out += tcp_skb_pcount(buff);
538 }
539
540 /* Link BUFF into the send queue. */
541 __skb_append(skb, buff);
542
543 return 0;
544 }
545
546 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
547 * eventually). The difference is that pulled data not copied, but
548 * immediately discarded.
549 */
550 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
551 {
552 int i, k, eat;
553
554 eat = len;
555 k = 0;
556 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
557 if (skb_shinfo(skb)->frags[i].size <= eat) {
558 put_page(skb_shinfo(skb)->frags[i].page);
559 eat -= skb_shinfo(skb)->frags[i].size;
560 } else {
561 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
562 if (eat) {
563 skb_shinfo(skb)->frags[k].page_offset += eat;
564 skb_shinfo(skb)->frags[k].size -= eat;
565 eat = 0;
566 }
567 k++;
568 }
569 }
570 skb_shinfo(skb)->nr_frags = k;
571
572 skb->tail = skb->data;
573 skb->data_len -= len;
574 skb->len = skb->data_len;
575 return skb->tail;
576 }
577
578 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
579 {
580 if (skb_cloned(skb) &&
581 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
582 return -ENOMEM;
583
584 if (len <= skb_headlen(skb)) {
585 __skb_pull(skb, len);
586 } else {
587 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
588 return -ENOMEM;
589 }
590
591 TCP_SKB_CB(skb)->seq += len;
592 skb->ip_summed = CHECKSUM_HW;
593
594 skb->truesize -= len;
595 sk->sk_queue_shrunk = 1;
596 sk->sk_wmem_queued -= len;
597 sk->sk_forward_alloc += len;
598
599 /* Any change of skb->len requires recalculation of tso
600 * factor and mss.
601 */
602 if (tcp_skb_pcount(skb) > 1)
603 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
604
605 return 0;
606 }
607
608 /* This function synchronize snd mss to current pmtu/exthdr set.
609
610 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
611 for TCP options, but includes only bare TCP header.
612
613 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
614 It is minumum of user_mss and mss received with SYN.
615 It also does not include TCP options.
616
617 tp->pmtu_cookie is last pmtu, seen by this function.
618
619 tp->mss_cache is current effective sending mss, including
620 all tcp options except for SACKs. It is evaluated,
621 taking into account current pmtu, but never exceeds
622 tp->rx_opt.mss_clamp.
623
624 NOTE1. rfc1122 clearly states that advertised MSS
625 DOES NOT include either tcp or ip options.
626
627 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
628 this function. --ANK (980731)
629 */
630
631 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
632 {
633 struct tcp_sock *tp = tcp_sk(sk);
634 struct dst_entry *dst = __sk_dst_get(sk);
635 int mss_now;
636
637 if (dst && dst->ops->get_mss)
638 pmtu = dst->ops->get_mss(dst, pmtu);
639
640 /* Calculate base mss without TCP options:
641 It is MMS_S - sizeof(tcphdr) of rfc1122
642 */
643 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
644
645 /* Clamp it (mss_clamp does not include tcp options) */
646 if (mss_now > tp->rx_opt.mss_clamp)
647 mss_now = tp->rx_opt.mss_clamp;
648
649 /* Now subtract optional transport overhead */
650 mss_now -= tp->ext_header_len + tp->ext2_header_len;
651
652 /* Then reserve room for full set of TCP options and 8 bytes of data */
653 if (mss_now < 48)
654 mss_now = 48;
655
656 /* Now subtract TCP options size, not including SACKs */
657 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
658
659 /* Bound mss with half of window */
660 if (tp->max_window && mss_now > (tp->max_window>>1))
661 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
662
663 /* And store cached results */
664 tp->pmtu_cookie = pmtu;
665 tp->mss_cache = tp->mss_cache_std = mss_now;
666
667 return mss_now;
668 }
669
670 /* Compute the current effective MSS, taking SACKs and IP options,
671 * and even PMTU discovery events into account.
672 *
673 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
674 * cannot be large. However, taking into account rare use of URG, this
675 * is not a big flaw.
676 */
677
678 unsigned int tcp_current_mss(struct sock *sk, int large)
679 {
680 struct tcp_sock *tp = tcp_sk(sk);
681 struct dst_entry *dst = __sk_dst_get(sk);
682 unsigned int do_large, mss_now;
683
684 mss_now = tp->mss_cache_std;
685 if (dst) {
686 u32 mtu = dst_pmtu(dst);
687 if (mtu != tp->pmtu_cookie ||
688 tp->ext2_header_len != dst->header_len)
689 mss_now = tcp_sync_mss(sk, mtu);
690 }
691
692 do_large = (large &&
693 (sk->sk_route_caps & NETIF_F_TSO) &&
694 !tp->urg_mode);
695
696 if (do_large) {
697 unsigned int large_mss, factor, limit;
698
699 large_mss = 65535 - tp->af_specific->net_header_len -
700 tp->ext_header_len - tp->ext2_header_len -
701 tp->tcp_header_len;
702
703 if (tp->max_window && large_mss > (tp->max_window>>1))
704 large_mss = max((tp->max_window>>1),
705 68U - tp->tcp_header_len);
706
707 factor = large_mss / mss_now;
708
709 /* Always keep large mss multiple of real mss, but
710 * do not exceed 1/tso_win_divisor of the congestion window
711 * so we can keep the ACK clock ticking and minimize
712 * bursting.
713 */
714 limit = tp->snd_cwnd;
715 if (sysctl_tcp_tso_win_divisor)
716 limit /= sysctl_tcp_tso_win_divisor;
717 limit = max(1U, limit);
718 if (factor > limit)
719 factor = limit;
720
721 tp->mss_cache = mss_now * factor;
722
723 mss_now = tp->mss_cache;
724 }
725
726 if (tp->rx_opt.eff_sacks)
727 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
728 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
729 return mss_now;
730 }
731
732 /* This routine writes packets to the network. It advances the
733 * send_head. This happens as incoming acks open up the remote
734 * window for us.
735 *
736 * Returns 1, if no segments are in flight and we have queued segments, but
737 * cannot send anything now because of SWS or another problem.
738 */
739 int tcp_write_xmit(struct sock *sk, int nonagle)
740 {
741 struct tcp_sock *tp = tcp_sk(sk);
742 unsigned int mss_now;
743
744 /* If we are closed, the bytes will have to remain here.
745 * In time closedown will finish, we empty the write queue and all
746 * will be happy.
747 */
748 if (sk->sk_state != TCP_CLOSE) {
749 struct sk_buff *skb;
750 int sent_pkts = 0;
751
752 /* Account for SACKS, we may need to fragment due to this.
753 * It is just like the real MSS changing on us midstream.
754 * We also handle things correctly when the user adds some
755 * IP options mid-stream. Silly to do, but cover it.
756 */
757 mss_now = tcp_current_mss(sk, 1);
758
759 while ((skb = sk->sk_send_head) &&
760 tcp_snd_test(tp, skb, mss_now,
761 tcp_skb_is_last(sk, skb) ? nonagle :
762 TCP_NAGLE_PUSH)) {
763 if (skb->len > mss_now) {
764 if (tcp_fragment(sk, skb, mss_now))
765 break;
766 }
767
768 TCP_SKB_CB(skb)->when = tcp_time_stamp;
769 tcp_tso_set_push(skb);
770 if (tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)))
771 break;
772
773 /* Advance the send_head. This one is sent out.
774 * This call will increment packets_out.
775 */
776 update_send_head(sk, tp, skb);
777
778 tcp_minshall_update(tp, mss_now, skb);
779 sent_pkts = 1;
780 }
781
782 if (sent_pkts) {
783 tcp_cwnd_validate(sk, tp);
784 return 0;
785 }
786
787 return !tp->packets_out && sk->sk_send_head;
788 }
789 return 0;
790 }
791
792 /* This function returns the amount that we can raise the
793 * usable window based on the following constraints
794 *
795 * 1. The window can never be shrunk once it is offered (RFC 793)
796 * 2. We limit memory per socket
797 *
798 * RFC 1122:
799 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
800 * RECV.NEXT + RCV.WIN fixed until:
801 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
802 *
803 * i.e. don't raise the right edge of the window until you can raise
804 * it at least MSS bytes.
805 *
806 * Unfortunately, the recommended algorithm breaks header prediction,
807 * since header prediction assumes th->window stays fixed.
808 *
809 * Strictly speaking, keeping th->window fixed violates the receiver
810 * side SWS prevention criteria. The problem is that under this rule
811 * a stream of single byte packets will cause the right side of the
812 * window to always advance by a single byte.
813 *
814 * Of course, if the sender implements sender side SWS prevention
815 * then this will not be a problem.
816 *
817 * BSD seems to make the following compromise:
818 *
819 * If the free space is less than the 1/4 of the maximum
820 * space available and the free space is less than 1/2 mss,
821 * then set the window to 0.
822 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
823 * Otherwise, just prevent the window from shrinking
824 * and from being larger than the largest representable value.
825 *
826 * This prevents incremental opening of the window in the regime
827 * where TCP is limited by the speed of the reader side taking
828 * data out of the TCP receive queue. It does nothing about
829 * those cases where the window is constrained on the sender side
830 * because the pipeline is full.
831 *
832 * BSD also seems to "accidentally" limit itself to windows that are a
833 * multiple of MSS, at least until the free space gets quite small.
834 * This would appear to be a side effect of the mbuf implementation.
835 * Combining these two algorithms results in the observed behavior
836 * of having a fixed window size at almost all times.
837 *
838 * Below we obtain similar behavior by forcing the offered window to
839 * a multiple of the mss when it is feasible to do so.
840 *
841 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
842 * Regular options like TIMESTAMP are taken into account.
843 */
844 u32 __tcp_select_window(struct sock *sk)
845 {
846 struct tcp_sock *tp = tcp_sk(sk);
847 /* MSS for the peer's data. Previous verions used mss_clamp
848 * here. I don't know if the value based on our guesses
849 * of peer's MSS is better for the performance. It's more correct
850 * but may be worse for the performance because of rcv_mss
851 * fluctuations. --SAW 1998/11/1
852 */
853 int mss = tp->ack.rcv_mss;
854 int free_space = tcp_space(sk);
855 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
856 int window;
857
858 if (mss > full_space)
859 mss = full_space;
860
861 if (free_space < full_space/2) {
862 tp->ack.quick = 0;
863
864 if (tcp_memory_pressure)
865 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
866
867 if (free_space < mss)
868 return 0;
869 }
870
871 if (free_space > tp->rcv_ssthresh)
872 free_space = tp->rcv_ssthresh;
873
874 /* Don't do rounding if we are using window scaling, since the
875 * scaled window will not line up with the MSS boundary anyway.
876 */
877 window = tp->rcv_wnd;
878 if (tp->rx_opt.rcv_wscale) {
879 window = free_space;
880
881 /* Advertise enough space so that it won't get scaled away.
882 * Import case: prevent zero window announcement if
883 * 1<<rcv_wscale > mss.
884 */
885 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
886 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
887 << tp->rx_opt.rcv_wscale);
888 } else {
889 /* Get the largest window that is a nice multiple of mss.
890 * Window clamp already applied above.
891 * If our current window offering is within 1 mss of the
892 * free space we just keep it. This prevents the divide
893 * and multiply from happening most of the time.
894 * We also don't do any window rounding when the free space
895 * is too small.
896 */
897 if (window <= free_space - mss || window > free_space)
898 window = (free_space/mss)*mss;
899 }
900
901 return window;
902 }
903
904 /* Attempt to collapse two adjacent SKB's during retransmission. */
905 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
906 {
907 struct tcp_sock *tp = tcp_sk(sk);
908 struct sk_buff *next_skb = skb->next;
909
910 /* The first test we must make is that neither of these two
911 * SKB's are still referenced by someone else.
912 */
913 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
914 int skb_size = skb->len, next_skb_size = next_skb->len;
915 u16 flags = TCP_SKB_CB(skb)->flags;
916
917 /* Also punt if next skb has been SACK'd. */
918 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
919 return;
920
921 /* Next skb is out of window. */
922 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
923 return;
924
925 /* Punt if not enough space exists in the first SKB for
926 * the data in the second, or the total combined payload
927 * would exceed the MSS.
928 */
929 if ((next_skb_size > skb_tailroom(skb)) ||
930 ((skb_size + next_skb_size) > mss_now))
931 return;
932
933 BUG_ON(tcp_skb_pcount(skb) != 1 ||
934 tcp_skb_pcount(next_skb) != 1);
935
936 /* Ok. We will be able to collapse the packet. */
937 __skb_unlink(next_skb, next_skb->list);
938
939 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
940
941 if (next_skb->ip_summed == CHECKSUM_HW)
942 skb->ip_summed = CHECKSUM_HW;
943
944 if (skb->ip_summed != CHECKSUM_HW)
945 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
946
947 /* Update sequence range on original skb. */
948 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
949
950 /* Merge over control information. */
951 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
952 TCP_SKB_CB(skb)->flags = flags;
953
954 /* All done, get rid of second SKB and account for it so
955 * packet counting does not break.
956 */
957 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
958 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
959 tp->retrans_out -= tcp_skb_pcount(next_skb);
960 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
961 tp->lost_out -= tcp_skb_pcount(next_skb);
962 tp->left_out -= tcp_skb_pcount(next_skb);
963 }
964 /* Reno case is special. Sigh... */
965 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
966 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
967 tp->left_out -= tcp_skb_pcount(next_skb);
968 }
969
970 /* Not quite right: it can be > snd.fack, but
971 * it is better to underestimate fackets.
972 */
973 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
974 tcp_packets_out_dec(tp, next_skb);
975 sk_stream_free_skb(sk, next_skb);
976 }
977 }
978
979 /* Do a simple retransmit without using the backoff mechanisms in
980 * tcp_timer. This is used for path mtu discovery.
981 * The socket is already locked here.
982 */
983 void tcp_simple_retransmit(struct sock *sk)
984 {
985 struct tcp_sock *tp = tcp_sk(sk);
986 struct sk_buff *skb;
987 unsigned int mss = tcp_current_mss(sk, 0);
988 int lost = 0;
989
990 sk_stream_for_retrans_queue(skb, sk) {
991 if (skb->len > mss &&
992 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
993 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
994 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
995 tp->retrans_out -= tcp_skb_pcount(skb);
996 }
997 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
998 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
999 tp->lost_out += tcp_skb_pcount(skb);
1000 lost = 1;
1001 }
1002 }
1003 }
1004
1005 if (!lost)
1006 return;
1007
1008 tcp_sync_left_out(tp);
1009
1010 /* Don't muck with the congestion window here.
1011 * Reason is that we do not increase amount of _data_
1012 * in network, but units changed and effective
1013 * cwnd/ssthresh really reduced now.
1014 */
1015 if (tp->ca_state != TCP_CA_Loss) {
1016 tp->high_seq = tp->snd_nxt;
1017 tp->snd_ssthresh = tcp_current_ssthresh(tp);
1018 tp->prior_ssthresh = 0;
1019 tp->undo_marker = 0;
1020 tcp_set_ca_state(tp, TCP_CA_Loss);
1021 }
1022 tcp_xmit_retransmit_queue(sk);
1023 }
1024
1025 /* This retransmits one SKB. Policy decisions and retransmit queue
1026 * state updates are done by the caller. Returns non-zero if an
1027 * error occurred which prevented the send.
1028 */
1029 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1030 {
1031 struct tcp_sock *tp = tcp_sk(sk);
1032 unsigned int cur_mss = tcp_current_mss(sk, 0);
1033 int err;
1034
1035 /* Do not sent more than we queued. 1/4 is reserved for possible
1036 * copying overhead: frgagmentation, tunneling, mangling etc.
1037 */
1038 if (atomic_read(&sk->sk_wmem_alloc) >
1039 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1040 return -EAGAIN;
1041
1042 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1043 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1044 BUG();
1045
1046 if (sk->sk_route_caps & NETIF_F_TSO) {
1047 sk->sk_route_caps &= ~NETIF_F_TSO;
1048 sk->sk_no_largesend = 1;
1049 tp->mss_cache = tp->mss_cache_std;
1050 }
1051
1052 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1053 return -ENOMEM;
1054 }
1055
1056 /* If receiver has shrunk his window, and skb is out of
1057 * new window, do not retransmit it. The exception is the
1058 * case, when window is shrunk to zero. In this case
1059 * our retransmit serves as a zero window probe.
1060 */
1061 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1062 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1063 return -EAGAIN;
1064
1065 if (skb->len > cur_mss) {
1066 int old_factor = tcp_skb_pcount(skb);
1067 int new_factor;
1068
1069 if (tcp_fragment(sk, skb, cur_mss))
1070 return -ENOMEM; /* We'll try again later. */
1071
1072 /* New SKB created, account for it. */
1073 new_factor = tcp_skb_pcount(skb);
1074 tp->packets_out -= old_factor - new_factor;
1075 tp->packets_out += tcp_skb_pcount(skb->next);
1076 }
1077
1078 /* Collapse two adjacent packets if worthwhile and we can. */
1079 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1080 (skb->len < (cur_mss >> 1)) &&
1081 (skb->next != sk->sk_send_head) &&
1082 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1083 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1084 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1085 (sysctl_tcp_retrans_collapse != 0))
1086 tcp_retrans_try_collapse(sk, skb, cur_mss);
1087
1088 if(tp->af_specific->rebuild_header(sk))
1089 return -EHOSTUNREACH; /* Routing failure or similar. */
1090
1091 /* Some Solaris stacks overoptimize and ignore the FIN on a
1092 * retransmit when old data is attached. So strip it off
1093 * since it is cheap to do so and saves bytes on the network.
1094 */
1095 if(skb->len > 0 &&
1096 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1097 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1098 if (!pskb_trim(skb, 0)) {
1099 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1100 skb_shinfo(skb)->tso_segs = 1;
1101 skb_shinfo(skb)->tso_size = 0;
1102 skb->ip_summed = CHECKSUM_NONE;
1103 skb->csum = 0;
1104 }
1105 }
1106
1107 /* Make a copy, if the first transmission SKB clone we made
1108 * is still in somebody's hands, else make a clone.
1109 */
1110 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1111 tcp_tso_set_push(skb);
1112
1113 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1114 pskb_copy(skb, GFP_ATOMIC):
1115 skb_clone(skb, GFP_ATOMIC)));
1116
1117 if (err == 0) {
1118 /* Update global TCP statistics. */
1119 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1120
1121 tp->total_retrans++;
1122
1123 #if FASTRETRANS_DEBUG > 0
1124 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1125 if (net_ratelimit())
1126 printk(KERN_DEBUG "retrans_out leaked.\n");
1127 }
1128 #endif
1129 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1130 tp->retrans_out += tcp_skb_pcount(skb);
1131
1132 /* Save stamp of the first retransmit. */
1133 if (!tp->retrans_stamp)
1134 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1135
1136 tp->undo_retrans++;
1137
1138 /* snd_nxt is stored to detect loss of retransmitted segment,
1139 * see tcp_input.c tcp_sacktag_write_queue().
1140 */
1141 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1142 }
1143 return err;
1144 }
1145
1146 /* This gets called after a retransmit timeout, and the initially
1147 * retransmitted data is acknowledged. It tries to continue
1148 * resending the rest of the retransmit queue, until either
1149 * we've sent it all or the congestion window limit is reached.
1150 * If doing SACK, the first ACK which comes back for a timeout
1151 * based retransmit packet might feed us FACK information again.
1152 * If so, we use it to avoid unnecessarily retransmissions.
1153 */
1154 void tcp_xmit_retransmit_queue(struct sock *sk)
1155 {
1156 struct tcp_sock *tp = tcp_sk(sk);
1157 struct sk_buff *skb;
1158 int packet_cnt = tp->lost_out;
1159
1160 /* First pass: retransmit lost packets. */
1161 if (packet_cnt) {
1162 sk_stream_for_retrans_queue(skb, sk) {
1163 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1164
1165 /* Assume this retransmit will generate
1166 * only one packet for congestion window
1167 * calculation purposes. This works because
1168 * tcp_retransmit_skb() will chop up the
1169 * packet to be MSS sized and all the
1170 * packet counting works out.
1171 */
1172 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1173 return;
1174
1175 if (sacked&TCPCB_LOST) {
1176 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1177 if (tcp_retransmit_skb(sk, skb))
1178 return;
1179 if (tp->ca_state != TCP_CA_Loss)
1180 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1181 else
1182 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1183
1184 if (skb ==
1185 skb_peek(&sk->sk_write_queue))
1186 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1187 }
1188
1189 packet_cnt -= tcp_skb_pcount(skb);
1190 if (packet_cnt <= 0)
1191 break;
1192 }
1193 }
1194 }
1195
1196 /* OK, demanded retransmission is finished. */
1197
1198 /* Forward retransmissions are possible only during Recovery. */
1199 if (tp->ca_state != TCP_CA_Recovery)
1200 return;
1201
1202 /* No forward retransmissions in Reno are possible. */
1203 if (!tp->rx_opt.sack_ok)
1204 return;
1205
1206 /* Yeah, we have to make difficult choice between forward transmission
1207 * and retransmission... Both ways have their merits...
1208 *
1209 * For now we do not retransmit anything, while we have some new
1210 * segments to send.
1211 */
1212
1213 if (tcp_may_send_now(sk, tp))
1214 return;
1215
1216 packet_cnt = 0;
1217
1218 sk_stream_for_retrans_queue(skb, sk) {
1219 /* Similar to the retransmit loop above we
1220 * can pretend that the retransmitted SKB
1221 * we send out here will be composed of one
1222 * real MSS sized packet because tcp_retransmit_skb()
1223 * will fragment it if necessary.
1224 */
1225 if (++packet_cnt > tp->fackets_out)
1226 break;
1227
1228 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1229 break;
1230
1231 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1232 continue;
1233
1234 /* Ok, retransmit it. */
1235 if (tcp_retransmit_skb(sk, skb))
1236 break;
1237
1238 if (skb == skb_peek(&sk->sk_write_queue))
1239 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1240
1241 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1242 }
1243 }
1244
1245
1246 /* Send a fin. The caller locks the socket for us. This cannot be
1247 * allowed to fail queueing a FIN frame under any circumstances.
1248 */
1249 void tcp_send_fin(struct sock *sk)
1250 {
1251 struct tcp_sock *tp = tcp_sk(sk);
1252 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1253 int mss_now;
1254
1255 /* Optimization, tack on the FIN if we have a queue of
1256 * unsent frames. But be careful about outgoing SACKS
1257 * and IP options.
1258 */
1259 mss_now = tcp_current_mss(sk, 1);
1260
1261 if (sk->sk_send_head != NULL) {
1262 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1263 TCP_SKB_CB(skb)->end_seq++;
1264 tp->write_seq++;
1265 } else {
1266 /* Socket is locked, keep trying until memory is available. */
1267 for (;;) {
1268 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL);
1269 if (skb)
1270 break;
1271 yield();
1272 }
1273
1274 /* Reserve space for headers and prepare control bits. */
1275 skb_reserve(skb, MAX_TCP_HEADER);
1276 skb->csum = 0;
1277 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1278 TCP_SKB_CB(skb)->sacked = 0;
1279 skb_shinfo(skb)->tso_segs = 1;
1280 skb_shinfo(skb)->tso_size = 0;
1281
1282 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1283 TCP_SKB_CB(skb)->seq = tp->write_seq;
1284 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1285 tcp_queue_skb(sk, skb);
1286 }
1287 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1288 }
1289
1290 /* We get here when a process closes a file descriptor (either due to
1291 * an explicit close() or as a byproduct of exit()'ing) and there
1292 * was unread data in the receive queue. This behavior is recommended
1293 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1294 */
1295 void tcp_send_active_reset(struct sock *sk, int priority)
1296 {
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct sk_buff *skb;
1299
1300 /* NOTE: No TCP options attached and we never retransmit this. */
1301 skb = alloc_skb(MAX_TCP_HEADER, priority);
1302 if (!skb) {
1303 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1304 return;
1305 }
1306
1307 /* Reserve space for headers and prepare control bits. */
1308 skb_reserve(skb, MAX_TCP_HEADER);
1309 skb->csum = 0;
1310 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1311 TCP_SKB_CB(skb)->sacked = 0;
1312 skb_shinfo(skb)->tso_segs = 1;
1313 skb_shinfo(skb)->tso_size = 0;
1314
1315 /* Send it off. */
1316 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1317 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1318 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1319 if (tcp_transmit_skb(sk, skb))
1320 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1321 }
1322
1323 /* WARNING: This routine must only be called when we have already sent
1324 * a SYN packet that crossed the incoming SYN that caused this routine
1325 * to get called. If this assumption fails then the initial rcv_wnd
1326 * and rcv_wscale values will not be correct.
1327 */
1328 int tcp_send_synack(struct sock *sk)
1329 {
1330 struct sk_buff* skb;
1331
1332 skb = skb_peek(&sk->sk_write_queue);
1333 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1334 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1335 return -EFAULT;
1336 }
1337 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1338 if (skb_cloned(skb)) {
1339 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1340 if (nskb == NULL)
1341 return -ENOMEM;
1342 __skb_unlink(skb, &sk->sk_write_queue);
1343 __skb_queue_head(&sk->sk_write_queue, nskb);
1344 sk_stream_free_skb(sk, skb);
1345 sk_charge_skb(sk, nskb);
1346 skb = nskb;
1347 }
1348
1349 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1350 TCP_ECN_send_synack(tcp_sk(sk), skb);
1351 }
1352 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1353 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1354 }
1355
1356 /*
1357 * Prepare a SYN-ACK.
1358 */
1359 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1360 struct open_request *req)
1361 {
1362 struct tcp_sock *tp = tcp_sk(sk);
1363 struct tcphdr *th;
1364 int tcp_header_size;
1365 struct sk_buff *skb;
1366
1367 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1368 if (skb == NULL)
1369 return NULL;
1370
1371 /* Reserve space for headers. */
1372 skb_reserve(skb, MAX_TCP_HEADER);
1373
1374 skb->dst = dst_clone(dst);
1375
1376 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1377 (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1378 (req->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1379 /* SACK_PERM is in the place of NOP NOP of TS */
1380 ((req->sack_ok && !req->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1381 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1382
1383 memset(th, 0, sizeof(struct tcphdr));
1384 th->syn = 1;
1385 th->ack = 1;
1386 if (dst->dev->features&NETIF_F_TSO)
1387 req->ecn_ok = 0;
1388 TCP_ECN_make_synack(req, th);
1389 th->source = inet_sk(sk)->sport;
1390 th->dest = req->rmt_port;
1391 TCP_SKB_CB(skb)->seq = req->snt_isn;
1392 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1393 TCP_SKB_CB(skb)->sacked = 0;
1394 skb_shinfo(skb)->tso_segs = 1;
1395 skb_shinfo(skb)->tso_size = 0;
1396 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1397 th->ack_seq = htonl(req->rcv_isn + 1);
1398 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1399 __u8 rcv_wscale;
1400 /* Set this up on the first call only */
1401 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1402 /* tcp_full_space because it is guaranteed to be the first packet */
1403 tcp_select_initial_window(tcp_full_space(sk),
1404 dst_metric(dst, RTAX_ADVMSS) - (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1405 &req->rcv_wnd,
1406 &req->window_clamp,
1407 req->wscale_ok,
1408 &rcv_wscale);
1409 req->rcv_wscale = rcv_wscale;
1410 }
1411
1412 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1413 th->window = htons(req->rcv_wnd);
1414
1415 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1416 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), req->tstamp_ok,
1417 req->sack_ok, req->wscale_ok, req->rcv_wscale,
1418 TCP_SKB_CB(skb)->when,
1419 req->ts_recent);
1420
1421 skb->csum = 0;
1422 th->doff = (tcp_header_size >> 2);
1423 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1424 return skb;
1425 }
1426
1427 /*
1428 * Do all connect socket setups that can be done AF independent.
1429 */
1430 static inline void tcp_connect_init(struct sock *sk)
1431 {
1432 struct dst_entry *dst = __sk_dst_get(sk);
1433 struct tcp_sock *tp = tcp_sk(sk);
1434
1435 /* We'll fix this up when we get a response from the other end.
1436 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1437 */
1438 tp->tcp_header_len = sizeof(struct tcphdr) +
1439 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1440
1441 /* If user gave his TCP_MAXSEG, record it to clamp */
1442 if (tp->rx_opt.user_mss)
1443 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1444 tp->max_window = 0;
1445 tcp_sync_mss(sk, dst_pmtu(dst));
1446
1447 if (!tp->window_clamp)
1448 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1449 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1450 tcp_initialize_rcv_mss(sk);
1451 tcp_ca_init(tp);
1452
1453 tcp_select_initial_window(tcp_full_space(sk),
1454 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1455 &tp->rcv_wnd,
1456 &tp->window_clamp,
1457 sysctl_tcp_window_scaling,
1458 &tp->rx_opt.rcv_wscale);
1459
1460 tp->rcv_ssthresh = tp->rcv_wnd;
1461
1462 sk->sk_err = 0;
1463 sock_reset_flag(sk, SOCK_DONE);
1464 tp->snd_wnd = 0;
1465 tcp_init_wl(tp, tp->write_seq, 0);
1466 tp->snd_una = tp->write_seq;
1467 tp->snd_sml = tp->write_seq;
1468 tp->rcv_nxt = 0;
1469 tp->rcv_wup = 0;
1470 tp->copied_seq = 0;
1471
1472 tp->rto = TCP_TIMEOUT_INIT;
1473 tp->retransmits = 0;
1474 tcp_clear_retrans(tp);
1475 }
1476
1477 /*
1478 * Build a SYN and send it off.
1479 */
1480 int tcp_connect(struct sock *sk)
1481 {
1482 struct tcp_sock *tp = tcp_sk(sk);
1483 struct sk_buff *buff;
1484
1485 tcp_connect_init(sk);
1486
1487 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation);
1488 if (unlikely(buff == NULL))
1489 return -ENOBUFS;
1490
1491 /* Reserve space for headers. */
1492 skb_reserve(buff, MAX_TCP_HEADER);
1493
1494 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1495 TCP_ECN_send_syn(sk, tp, buff);
1496 TCP_SKB_CB(buff)->sacked = 0;
1497 skb_shinfo(buff)->tso_segs = 1;
1498 skb_shinfo(buff)->tso_size = 0;
1499 buff->csum = 0;
1500 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1501 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1502 tp->snd_nxt = tp->write_seq;
1503 tp->pushed_seq = tp->write_seq;
1504 tcp_ca_init(tp);
1505
1506 /* Send it off. */
1507 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1508 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1509 __skb_queue_tail(&sk->sk_write_queue, buff);
1510 sk_charge_skb(sk, buff);
1511 tp->packets_out += tcp_skb_pcount(buff);
1512 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1513 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1514
1515 /* Timer for repeating the SYN until an answer. */
1516 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1517 return 0;
1518 }
1519
1520 /* Send out a delayed ack, the caller does the policy checking
1521 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1522 * for details.
1523 */
1524 void tcp_send_delayed_ack(struct sock *sk)
1525 {
1526 struct tcp_sock *tp = tcp_sk(sk);
1527 int ato = tp->ack.ato;
1528 unsigned long timeout;
1529
1530 if (ato > TCP_DELACK_MIN) {
1531 int max_ato = HZ/2;
1532
1533 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED))
1534 max_ato = TCP_DELACK_MAX;
1535
1536 /* Slow path, intersegment interval is "high". */
1537
1538 /* If some rtt estimate is known, use it to bound delayed ack.
1539 * Do not use tp->rto here, use results of rtt measurements
1540 * directly.
1541 */
1542 if (tp->srtt) {
1543 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1544
1545 if (rtt < max_ato)
1546 max_ato = rtt;
1547 }
1548
1549 ato = min(ato, max_ato);
1550 }
1551
1552 /* Stay within the limit we were given */
1553 timeout = jiffies + ato;
1554
1555 /* Use new timeout only if there wasn't a older one earlier. */
1556 if (tp->ack.pending&TCP_ACK_TIMER) {
1557 /* If delack timer was blocked or is about to expire,
1558 * send ACK now.
1559 */
1560 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) {
1561 tcp_send_ack(sk);
1562 return;
1563 }
1564
1565 if (!time_before(timeout, tp->ack.timeout))
1566 timeout = tp->ack.timeout;
1567 }
1568 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER;
1569 tp->ack.timeout = timeout;
1570 sk_reset_timer(sk, &tp->delack_timer, timeout);
1571 }
1572
1573 /* This routine sends an ack and also updates the window. */
1574 void tcp_send_ack(struct sock *sk)
1575 {
1576 /* If we have been reset, we may not send again. */
1577 if (sk->sk_state != TCP_CLOSE) {
1578 struct tcp_sock *tp = tcp_sk(sk);
1579 struct sk_buff *buff;
1580
1581 /* We are not putting this on the write queue, so
1582 * tcp_transmit_skb() will set the ownership to this
1583 * sock.
1584 */
1585 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1586 if (buff == NULL) {
1587 tcp_schedule_ack(tp);
1588 tp->ack.ato = TCP_ATO_MIN;
1589 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
1590 return;
1591 }
1592
1593 /* Reserve space for headers and prepare control bits. */
1594 skb_reserve(buff, MAX_TCP_HEADER);
1595 buff->csum = 0;
1596 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1597 TCP_SKB_CB(buff)->sacked = 0;
1598 skb_shinfo(buff)->tso_segs = 1;
1599 skb_shinfo(buff)->tso_size = 0;
1600
1601 /* Send it off, this clears delayed acks for us. */
1602 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1603 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1604 tcp_transmit_skb(sk, buff);
1605 }
1606 }
1607
1608 /* This routine sends a packet with an out of date sequence
1609 * number. It assumes the other end will try to ack it.
1610 *
1611 * Question: what should we make while urgent mode?
1612 * 4.4BSD forces sending single byte of data. We cannot send
1613 * out of window data, because we have SND.NXT==SND.MAX...
1614 *
1615 * Current solution: to send TWO zero-length segments in urgent mode:
1616 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1617 * out-of-date with SND.UNA-1 to probe window.
1618 */
1619 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1620 {
1621 struct tcp_sock *tp = tcp_sk(sk);
1622 struct sk_buff *skb;
1623
1624 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1625 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1626 if (skb == NULL)
1627 return -1;
1628
1629 /* Reserve space for headers and set control bits. */
1630 skb_reserve(skb, MAX_TCP_HEADER);
1631 skb->csum = 0;
1632 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1633 TCP_SKB_CB(skb)->sacked = urgent;
1634 skb_shinfo(skb)->tso_segs = 1;
1635 skb_shinfo(skb)->tso_size = 0;
1636
1637 /* Use a previous sequence. This should cause the other
1638 * end to send an ack. Don't queue or clone SKB, just
1639 * send it.
1640 */
1641 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1642 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1643 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1644 return tcp_transmit_skb(sk, skb);
1645 }
1646
1647 int tcp_write_wakeup(struct sock *sk)
1648 {
1649 if (sk->sk_state != TCP_CLOSE) {
1650 struct tcp_sock *tp = tcp_sk(sk);
1651 struct sk_buff *skb;
1652
1653 if ((skb = sk->sk_send_head) != NULL &&
1654 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1655 int err;
1656 unsigned int mss = tcp_current_mss(sk, 0);
1657 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1658
1659 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1660 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1661
1662 /* We are probing the opening of a window
1663 * but the window size is != 0
1664 * must have been a result SWS avoidance ( sender )
1665 */
1666 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1667 skb->len > mss) {
1668 seg_size = min(seg_size, mss);
1669 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1670 if (tcp_fragment(sk, skb, seg_size))
1671 return -1;
1672 /* SWS override triggered forced fragmentation.
1673 * Disable TSO, the connection is too sick. */
1674 if (sk->sk_route_caps & NETIF_F_TSO) {
1675 sk->sk_no_largesend = 1;
1676 sk->sk_route_caps &= ~NETIF_F_TSO;
1677 tp->mss_cache = tp->mss_cache_std;
1678 }
1679 } else if (!tcp_skb_pcount(skb))
1680 tcp_set_skb_tso_segs(skb, tp->mss_cache_std);
1681
1682 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1683 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1684 tcp_tso_set_push(skb);
1685 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1686 if (!err) {
1687 update_send_head(sk, tp, skb);
1688 }
1689 return err;
1690 } else {
1691 if (tp->urg_mode &&
1692 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
1693 tcp_xmit_probe_skb(sk, TCPCB_URG);
1694 return tcp_xmit_probe_skb(sk, 0);
1695 }
1696 }
1697 return -1;
1698 }
1699
1700 /* A window probe timeout has occurred. If window is not closed send
1701 * a partial packet else a zero probe.
1702 */
1703 void tcp_send_probe0(struct sock *sk)
1704 {
1705 struct tcp_sock *tp = tcp_sk(sk);
1706 int err;
1707
1708 err = tcp_write_wakeup(sk);
1709
1710 if (tp->packets_out || !sk->sk_send_head) {
1711 /* Cancel probe timer, if it is not required. */
1712 tp->probes_out = 0;
1713 tp->backoff = 0;
1714 return;
1715 }
1716
1717 if (err <= 0) {
1718 if (tp->backoff < sysctl_tcp_retries2)
1719 tp->backoff++;
1720 tp->probes_out++;
1721 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
1722 min(tp->rto << tp->backoff, TCP_RTO_MAX));
1723 } else {
1724 /* If packet was not sent due to local congestion,
1725 * do not backoff and do not remember probes_out.
1726 * Let local senders to fight for local resources.
1727 *
1728 * Use accumulated backoff yet.
1729 */
1730 if (!tp->probes_out)
1731 tp->probes_out=1;
1732 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
1733 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL));
1734 }
1735 }
1736
1737 EXPORT_SYMBOL(tcp_connect);
1738 EXPORT_SYMBOL(tcp_make_synack);
1739 EXPORT_SYMBOL(tcp_simple_retransmit);
1740 EXPORT_SYMBOL(tcp_sync_mss);
1741
|
This page was automatically generated by the
LXR engine.
|