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.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
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>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
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
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse = 1;
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.
52 int sysctl_tcp_tso_win_divisor = 8;
54 static inline void update_send_head(struct sock *sk, struct tcp_sock *tp,
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);
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:
70 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
72 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
75 return tp->snd_una+tp->snd_wnd;
78 /* Calculate mss to advertise in SYN segment.
79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
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
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".
92 static __u16 tcp_advertise_mss(struct sock *sk)
94 struct tcp_sock *tp = tcp_sk(sk);
95 struct dst_entry *dst = __sk_dst_get(sk);
98 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
99 mss = dst_metric(dst, RTAX_ADVMSS);
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)
110 s32 delta = tcp_time_stamp - tp->lsndtime;
111 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
112 u32 cwnd = tp->snd_cwnd;
114 if (tcp_is_vegas(tp))
115 tcp_vegas_enable(tp);
117 tp->snd_ssthresh = tcp_current_ssthresh(tp);
118 restart_cwnd = min(restart_cwnd, cwnd);
120 while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd)
122 tp->snd_cwnd = max(cwnd, restart_cwnd);
123 tp->snd_cwnd_stamp = tcp_time_stamp;
124 tp->snd_cwnd_used = 0;
127 static inline void tcp_event_data_sent(struct tcp_sock *tp,
128 struct sk_buff *skb, struct sock *sk)
130 u32 now = tcp_time_stamp;
132 if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto)
133 tcp_cwnd_restart(tp, __sk_dst_get(sk));
137 /* If it is a reply for ato after last received
138 * packet, enter pingpong mode.
140 if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato)
141 tp->ack.pingpong = 1;
144 static __inline__ void tcp_event_ack_sent(struct sock *sk)
146 struct tcp_sock *tp = tcp_sk(sk);
148 tcp_dec_quickack_mode(tp);
149 tcp_clear_xmit_timer(sk, TCP_TIME_DACK);
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.
159 void tcp_select_initial_window(int __space, __u32 mss,
160 __u32 *rcv_wnd, __u32 *window_clamp,
161 int wscale_ok, __u8 *rcv_wscale)
163 unsigned int space = (__space < 0 ? 0 : __space);
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);
170 /* Quantize space offering to a multiple of mss if possible. */
172 space = (space / mss) * mss;
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.
180 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
183 /* Set window scaling on max possible window
184 * See RFC1323 for an explanation of the limit to 14
186 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
187 while (space > 65535 && (*rcv_wscale) < 14) {
193 /* Set initial window to value enough for senders,
194 * following RFC1414. Senders, not following this RFC,
195 * will be satisfied with 2.
197 if (mss > (1<<*rcv_wscale)) {
203 if (*rcv_wnd > init_cwnd*mss)
204 *rcv_wnd = init_cwnd*mss;
207 /* Set the clamp no higher than max representable value */
208 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
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
216 static __inline__ u16 tcp_select_window(struct sock *sk)
218 struct tcp_sock *tp = tcp_sk(sk);
219 u32 cur_win = tcp_receive_window(tp);
220 u32 new_win = __tcp_select_window(sk);
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
229 * Relax Will Robinson.
233 tp->rcv_wnd = new_win;
234 tp->rcv_wup = tp->rcv_nxt;
236 /* Make sure we do not exceed the maximum possible
239 if (!tp->rx_opt.rcv_wscale)
240 new_win = min(new_win, MAX_TCP_WINDOW);
242 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
244 /* RFC1323 scaling applied */
245 new_win >>= tp->rx_opt.rcv_wscale;
247 /* If we advertise zero window, disable fast path. */
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
263 * We are working here with either a clone of the original
264 * SKB, or a fresh unique copy made by the retransmit engine.
266 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
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;
277 BUG_ON(!tcp_skb_pcount(skb));
279 #define SYSCTL_FLAG_TSTAMPS 0x1
280 #define SYSCTL_FLAG_WSCALE 0x2
281 #define SYSCTL_FLAG_SACK 0x4
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;
290 if(sysctl_tcp_window_scaling) {
291 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
292 sysctl_flags |= SYSCTL_FLAG_WSCALE;
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;
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.
303 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
304 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
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
316 if (tcp_is_vegas(tp) && tcp_packets_in_flight(tp) == 0)
317 tcp_vegas_enable(tp);
319 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
321 skb_set_owner_w(skb, sk);
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
333 th->window = htons(tp->rcv_wnd);
335 th->window = htons(tcp_select_window(sk));
341 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
342 th->urg_ptr = htons(tp->snd_up-tcb->seq);
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,
354 tp->rx_opt.ts_recent);
356 tcp_build_and_update_options((__u32 *)(th + 1),
359 TCP_ECN_send(sk, tp, skb, tcp_header_size);
361 tp->af_specific->send_check(sk, th, skb->len, skb);
363 if (tcb->flags & TCPCB_FLAG_ACK)
364 tcp_event_ack_sent(sk);
366 if (skb->len != tcp_header_size)
367 tcp_event_data_sent(tp, skb, sk);
369 TCP_INC_STATS(TCP_MIB_OUTSEGS);
371 err = tp->af_specific->queue_xmit(skb, 0);
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.
383 return err == NET_XMIT_CN ? 0 : err;
386 #undef SYSCTL_FLAG_TSTAMPS
387 #undef SYSCTL_FLAG_WSCALE
388 #undef SYSCTL_FLAG_SACK
392 /* This routine just queue's the buffer
394 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
395 * otherwise socket can stall.
397 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
399 struct tcp_sock *tp = tcp_sk(sk);
401 /* Advance write_seq and place onto the write_queue. */
402 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
403 skb_header_release(skb);
404 __skb_queue_tail(&sk->sk_write_queue, skb);
405 sk_charge_skb(sk, skb);
407 /* Queue it, remembering where we must start sending. */
408 if (sk->sk_send_head == NULL)
409 sk->sk_send_head = skb;
412 static inline void tcp_tso_set_push(struct sk_buff *skb)
414 /* Force push to be on for any TSO frames to workaround
415 * problems with busted implementations like Mac OS-X that
416 * hold off socket receive wakeups until push is seen.
418 if (tcp_skb_pcount(skb) > 1)
419 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
422 /* Send _single_ skb sitting at the send head. This function requires
423 * true push pending frames to setup probe timer etc.
425 void tcp_push_one(struct sock *sk, unsigned cur_mss)
427 struct tcp_sock *tp = tcp_sk(sk);
428 struct sk_buff *skb = sk->sk_send_head;
430 if (tcp_snd_test(sk, skb, cur_mss, TCP_NAGLE_PUSH)) {
431 /* Send it out now. */
432 TCP_SKB_CB(skb)->when = tcp_time_stamp;
433 tcp_tso_set_push(skb);
434 if (!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation))) {
435 sk->sk_send_head = NULL;
436 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
437 tcp_packets_out_inc(sk, tp, skb);
443 void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb)
445 struct tcp_sock *tp = tcp_sk(sk);
447 if (skb->len <= tp->mss_cache_std ||
448 !(sk->sk_route_caps & NETIF_F_TSO)) {
449 /* Avoid the costly divide in the normal
452 skb_shinfo(skb)->tso_segs = 1;
453 skb_shinfo(skb)->tso_size = 0;
457 factor = skb->len + (tp->mss_cache_std - 1);
458 factor /= tp->mss_cache_std;
459 skb_shinfo(skb)->tso_segs = factor;
460 skb_shinfo(skb)->tso_size = tp->mss_cache_std;
464 EXPORT_SYMBOL_GPL(tcp_set_skb_tso_segs);
466 /* Function to create two new TCP segments. Shrinks the given segment
467 * to the specified size and appends a new segment with the rest of the
468 * packet to the list. This won't be called frequently, I hope.
469 * Remember, these are still headerless SKBs at this point.
471 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len)
473 struct tcp_sock *tp = tcp_sk(sk);
474 struct sk_buff *buff;
478 nsize = skb_headlen(skb) - len;
482 if (skb_cloned(skb) &&
483 skb_is_nonlinear(skb) &&
484 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
487 /* Get a new skb... force flag on. */
488 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
490 return -ENOMEM; /* We'll just try again later. */
491 sk_charge_skb(sk, buff);
493 /* Correct the sequence numbers. */
494 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
495 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
496 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
498 /* PSH and FIN should only be set in the second packet. */
499 flags = TCP_SKB_CB(skb)->flags;
500 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
501 TCP_SKB_CB(buff)->flags = flags;
502 TCP_SKB_CB(buff)->sacked =
503 (TCP_SKB_CB(skb)->sacked &
504 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
505 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
507 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
508 /* Copy and checksum data tail into the new buffer. */
509 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
514 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
516 skb->ip_summed = CHECKSUM_HW;
517 skb_split(skb, buff, len);
520 buff->ip_summed = skb->ip_summed;
522 /* Looks stupid, but our code really uses when of
523 * skbs, which it never sent before. --ANK
525 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
527 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
528 tp->lost_out -= tcp_skb_pcount(skb);
529 tp->left_out -= tcp_skb_pcount(skb);
532 /* Fix up tso_factor for both original and new SKB. */
533 tcp_set_skb_tso_segs(sk, skb);
534 tcp_set_skb_tso_segs(sk, buff);
536 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
537 tp->lost_out += tcp_skb_pcount(skb);
538 tp->left_out += tcp_skb_pcount(skb);
541 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) {
542 tp->lost_out += tcp_skb_pcount(buff);
543 tp->left_out += tcp_skb_pcount(buff);
546 /* Link BUFF into the send queue. */
547 __skb_append(skb, buff);
552 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
553 * eventually). The difference is that pulled data not copied, but
554 * immediately discarded.
556 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
562 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
563 if (skb_shinfo(skb)->frags[i].size <= eat) {
564 put_page(skb_shinfo(skb)->frags[i].page);
565 eat -= skb_shinfo(skb)->frags[i].size;
567 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
569 skb_shinfo(skb)->frags[k].page_offset += eat;
570 skb_shinfo(skb)->frags[k].size -= eat;
576 skb_shinfo(skb)->nr_frags = k;
578 skb->tail = skb->data;
579 skb->data_len -= len;
580 skb->len = skb->data_len;
584 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
586 if (skb_cloned(skb) &&
587 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
590 if (len <= skb_headlen(skb)) {
591 __skb_pull(skb, len);
593 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
597 TCP_SKB_CB(skb)->seq += len;
598 skb->ip_summed = CHECKSUM_HW;
600 skb->truesize -= len;
601 sk->sk_wmem_queued -= len;
602 sk->sk_forward_alloc += len;
603 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
605 /* Any change of skb->len requires recalculation of tso
608 if (tcp_skb_pcount(skb) > 1)
609 tcp_set_skb_tso_segs(sk, skb);
614 /* This function synchronize snd mss to current pmtu/exthdr set.
616 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
617 for TCP options, but includes only bare TCP header.
619 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
620 It is minumum of user_mss and mss received with SYN.
621 It also does not include TCP options.
623 tp->pmtu_cookie is last pmtu, seen by this function.
625 tp->mss_cache is current effective sending mss, including
626 all tcp options except for SACKs. It is evaluated,
627 taking into account current pmtu, but never exceeds
628 tp->rx_opt.mss_clamp.
630 NOTE1. rfc1122 clearly states that advertised MSS
631 DOES NOT include either tcp or ip options.
633 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
634 this function. --ANK (980731)
637 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
639 struct tcp_sock *tp = tcp_sk(sk);
642 /* Calculate base mss without TCP options:
643 It is MMS_S - sizeof(tcphdr) of rfc1122
645 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
647 /* Clamp it (mss_clamp does not include tcp options) */
648 if (mss_now > tp->rx_opt.mss_clamp)
649 mss_now = tp->rx_opt.mss_clamp;
651 /* Now subtract optional transport overhead */
652 mss_now -= tp->ext_header_len;
654 /* Then reserve room for full set of TCP options and 8 bytes of data */
658 /* Now subtract TCP options size, not including SACKs */
659 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
661 /* Bound mss with half of window */
662 if (tp->max_window && mss_now > (tp->max_window>>1))
663 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
665 /* And store cached results */
666 tp->pmtu_cookie = pmtu;
667 tp->mss_cache = tp->mss_cache_std = mss_now;
672 /* Compute the current effective MSS, taking SACKs and IP options,
673 * and even PMTU discovery events into account.
675 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
676 * cannot be large. However, taking into account rare use of URG, this
680 unsigned int tcp_current_mss(struct sock *sk, int large)
682 struct tcp_sock *tp = tcp_sk(sk);
683 struct dst_entry *dst = __sk_dst_get(sk);
684 unsigned int do_large, mss_now;
686 mss_now = tp->mss_cache_std;
688 u32 mtu = dst_mtu(dst);
689 if (mtu != tp->pmtu_cookie)
690 mss_now = tcp_sync_mss(sk, mtu);
694 (sk->sk_route_caps & NETIF_F_TSO) &&
698 unsigned int large_mss, factor, limit;
700 large_mss = 65535 - tp->af_specific->net_header_len -
701 tp->ext_header_len - tp->tcp_header_len;
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);
707 factor = large_mss / mss_now;
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
714 limit = tp->snd_cwnd;
715 if (sysctl_tcp_tso_win_divisor)
716 limit /= sysctl_tcp_tso_win_divisor;
717 limit = max(1U, limit);
721 tp->mss_cache = mss_now * factor;
723 mss_now = tp->mss_cache;
726 if (tp->rx_opt.eff_sacks)
727 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
728 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
731 EXPORT_SYMBOL_GPL(tcp_current_mss);
733 /* This routine writes packets to the network. It advances the
734 * send_head. This happens as incoming acks open up the remote
737 * Returns 1, if no segments are in flight and we have queued segments, but
738 * cannot send anything now because of SWS or another problem.
740 int tcp_write_xmit(struct sock *sk, int nonagle)
742 struct tcp_sock *tp = tcp_sk(sk);
743 unsigned int mss_now;
745 /* If we are closed, the bytes will have to remain here.
746 * In time closedown will finish, we empty the write queue and all
749 if (sk->sk_state != TCP_CLOSE) {
753 /* Account for SACKS, we may need to fragment due to this.
754 * It is just like the real MSS changing on us midstream.
755 * We also handle things correctly when the user adds some
756 * IP options mid-stream. Silly to do, but cover it.
758 mss_now = tcp_current_mss(sk, 1);
760 while ((skb = sk->sk_send_head) &&
761 tcp_snd_test(sk, skb, mss_now,
762 tcp_skb_is_last(sk, skb) ? nonagle :
764 if (skb->len > mss_now) {
765 if (tcp_fragment(sk, skb, mss_now))
769 TCP_SKB_CB(skb)->when = tcp_time_stamp;
770 tcp_tso_set_push(skb);
771 if (tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)))
774 /* Advance the send_head. This one is sent out.
775 * This call will increment packets_out.
777 update_send_head(sk, tp, skb);
779 tcp_minshall_update(tp, mss_now, skb);
784 tcp_cwnd_validate(sk, tp);
788 return !tp->packets_out && sk->sk_send_head;
793 EXPORT_SYMBOL_GPL(tcp_write_xmit);
795 /* This function returns the amount that we can raise the
796 * usable window based on the following constraints
798 * 1. The window can never be shrunk once it is offered (RFC 793)
799 * 2. We limit memory per socket
802 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
803 * RECV.NEXT + RCV.WIN fixed until:
804 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
806 * i.e. don't raise the right edge of the window until you can raise
807 * it at least MSS bytes.
809 * Unfortunately, the recommended algorithm breaks header prediction,
810 * since header prediction assumes th->window stays fixed.
812 * Strictly speaking, keeping th->window fixed violates the receiver
813 * side SWS prevention criteria. The problem is that under this rule
814 * a stream of single byte packets will cause the right side of the
815 * window to always advance by a single byte.
817 * Of course, if the sender implements sender side SWS prevention
818 * then this will not be a problem.
820 * BSD seems to make the following compromise:
822 * If the free space is less than the 1/4 of the maximum
823 * space available and the free space is less than 1/2 mss,
824 * then set the window to 0.
825 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
826 * Otherwise, just prevent the window from shrinking
827 * and from being larger than the largest representable value.
829 * This prevents incremental opening of the window in the regime
830 * where TCP is limited by the speed of the reader side taking
831 * data out of the TCP receive queue. It does nothing about
832 * those cases where the window is constrained on the sender side
833 * because the pipeline is full.
835 * BSD also seems to "accidentally" limit itself to windows that are a
836 * multiple of MSS, at least until the free space gets quite small.
837 * This would appear to be a side effect of the mbuf implementation.
838 * Combining these two algorithms results in the observed behavior
839 * of having a fixed window size at almost all times.
841 * Below we obtain similar behavior by forcing the offered window to
842 * a multiple of the mss when it is feasible to do so.
844 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
845 * Regular options like TIMESTAMP are taken into account.
847 u32 __tcp_select_window(struct sock *sk)
849 struct tcp_sock *tp = tcp_sk(sk);
850 /* MSS for the peer's data. Previous verions used mss_clamp
851 * here. I don't know if the value based on our guesses
852 * of peer's MSS is better for the performance. It's more correct
853 * but may be worse for the performance because of rcv_mss
854 * fluctuations. --SAW 1998/11/1
856 int mss = tp->ack.rcv_mss;
857 int free_space = tcp_space(sk);
858 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
861 if (mss > full_space)
864 if (free_space < full_space/2) {
867 if (tcp_memory_pressure)
868 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
870 if (free_space < mss)
874 if (free_space > tp->rcv_ssthresh)
875 free_space = tp->rcv_ssthresh;
877 /* Don't do rounding if we are using window scaling, since the
878 * scaled window will not line up with the MSS boundary anyway.
880 window = tp->rcv_wnd;
881 if (tp->rx_opt.rcv_wscale) {
884 /* Advertise enough space so that it won't get scaled away.
885 * Import case: prevent zero window announcement if
886 * 1<<rcv_wscale > mss.
888 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
889 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
890 << tp->rx_opt.rcv_wscale);
892 /* Get the largest window that is a nice multiple of mss.
893 * Window clamp already applied above.
894 * If our current window offering is within 1 mss of the
895 * free space we just keep it. This prevents the divide
896 * and multiply from happening most of the time.
897 * We also don't do any window rounding when the free space
900 if (window <= free_space - mss || window > free_space)
901 window = (free_space/mss)*mss;
907 /* Attempt to collapse two adjacent SKB's during retransmission. */
908 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
910 struct tcp_sock *tp = tcp_sk(sk);
911 struct sk_buff *next_skb = skb->next;
913 /* The first test we must make is that neither of these two
914 * SKB's are still referenced by someone else.
916 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
917 int skb_size = skb->len, next_skb_size = next_skb->len;
918 u16 flags = TCP_SKB_CB(skb)->flags;
920 /* Also punt if next skb has been SACK'd. */
921 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
924 /* Next skb is out of window. */
925 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
928 /* Punt if not enough space exists in the first SKB for
929 * the data in the second, or the total combined payload
930 * would exceed the MSS.
932 if ((next_skb_size > skb_tailroom(skb)) ||
933 ((skb_size + next_skb_size) > mss_now))
936 BUG_ON(tcp_skb_pcount(skb) != 1 ||
937 tcp_skb_pcount(next_skb) != 1);
939 /* Ok. We will be able to collapse the packet. */
940 __skb_unlink(next_skb, next_skb->list);
942 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
944 if (next_skb->ip_summed == CHECKSUM_HW)
945 skb->ip_summed = CHECKSUM_HW;
947 if (skb->ip_summed != CHECKSUM_HW)
948 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
950 /* Update sequence range on original skb. */
951 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
953 /* Merge over control information. */
954 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
955 TCP_SKB_CB(skb)->flags = flags;
957 /* All done, get rid of second SKB and account for it so
958 * packet counting does not break.
960 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
961 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
962 tp->retrans_out -= tcp_skb_pcount(next_skb);
963 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
964 tp->lost_out -= tcp_skb_pcount(next_skb);
965 tp->left_out -= tcp_skb_pcount(next_skb);
967 /* Reno case is special. Sigh... */
968 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
969 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
970 tp->left_out -= tcp_skb_pcount(next_skb);
973 /* Not quite right: it can be > snd.fack, but
974 * it is better to underestimate fackets.
976 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
977 tcp_packets_out_dec(tp, next_skb);
978 sk_stream_free_skb(sk, next_skb);
982 /* Do a simple retransmit without using the backoff mechanisms in
983 * tcp_timer. This is used for path mtu discovery.
984 * The socket is already locked here.
986 void tcp_simple_retransmit(struct sock *sk)
988 struct tcp_sock *tp = tcp_sk(sk);
990 unsigned int mss = tcp_current_mss(sk, 0);
993 sk_stream_for_retrans_queue(skb, sk) {
994 if (skb->len > mss &&
995 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
996 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
997 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
998 tp->retrans_out -= tcp_skb_pcount(skb);
1000 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1001 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1002 tp->lost_out += tcp_skb_pcount(skb);
1011 tcp_sync_left_out(tp);
1013 /* Don't muck with the congestion window here.
1014 * Reason is that we do not increase amount of _data_
1015 * in network, but units changed and effective
1016 * cwnd/ssthresh really reduced now.
1018 if (tp->ca_state != TCP_CA_Loss) {
1019 tp->high_seq = tp->snd_nxt;
1020 tp->snd_ssthresh = tcp_current_ssthresh(tp);
1021 tp->prior_ssthresh = 0;
1022 tp->undo_marker = 0;
1023 tcp_set_ca_state(tp, TCP_CA_Loss);
1025 tcp_xmit_retransmit_queue(sk);
1028 /* This retransmits one SKB. Policy decisions and retransmit queue
1029 * state updates are done by the caller. Returns non-zero if an
1030 * error occurred which prevented the send.
1032 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1034 struct tcp_sock *tp = tcp_sk(sk);
1035 unsigned int cur_mss = tcp_current_mss(sk, 0);
1038 /* Do not sent more than we queued. 1/4 is reserved for possible
1039 * copying overhead: frgagmentation, tunneling, mangling etc.
1041 if (atomic_read(&sk->sk_wmem_alloc) >
1042 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1045 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1046 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1049 if (sk->sk_route_caps & NETIF_F_TSO) {
1050 sk->sk_route_caps &= ~NETIF_F_TSO;
1051 sock_set_flag(sk, SOCK_NO_LARGESEND);
1052 tp->mss_cache = tp->mss_cache_std;
1055 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1059 /* If receiver has shrunk his window, and skb is out of
1060 * new window, do not retransmit it. The exception is the
1061 * case, when window is shrunk to zero. In this case
1062 * our retransmit serves as a zero window probe.
1064 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1065 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1068 if (skb->len > cur_mss) {
1069 int old_factor = tcp_skb_pcount(skb);
1072 if (tcp_fragment(sk, skb, cur_mss))
1073 return -ENOMEM; /* We'll try again later. */
1075 /* New SKB created, account for it. */
1076 new_factor = tcp_skb_pcount(skb);
1077 tp->packets_out -= old_factor - new_factor;
1078 tp->packets_out += tcp_skb_pcount(skb->next);
1081 /* Collapse two adjacent packets if worthwhile and we can. */
1082 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1083 (skb->len < (cur_mss >> 1)) &&
1084 (skb->next != sk->sk_send_head) &&
1085 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1086 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1087 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1088 (sysctl_tcp_retrans_collapse != 0))
1089 tcp_retrans_try_collapse(sk, skb, cur_mss);
1091 if(tp->af_specific->rebuild_header(sk))
1092 return -EHOSTUNREACH; /* Routing failure or similar. */
1094 /* Some Solaris stacks overoptimize and ignore the FIN on a
1095 * retransmit when old data is attached. So strip it off
1096 * since it is cheap to do so and saves bytes on the network.
1099 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1100 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1101 if (!pskb_trim(skb, 0)) {
1102 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1103 skb_shinfo(skb)->tso_segs = 1;
1104 skb_shinfo(skb)->tso_size = 0;
1105 skb->ip_summed = CHECKSUM_NONE;
1110 /* Make a copy, if the first transmission SKB clone we made
1111 * is still in somebody's hands, else make a clone.
1113 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1114 tcp_tso_set_push(skb);
1116 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1117 pskb_copy(skb, GFP_ATOMIC):
1118 skb_clone(skb, GFP_ATOMIC)));
1121 /* Update global TCP statistics. */
1122 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1124 tp->total_retrans++;
1126 #if FASTRETRANS_DEBUG > 0
1127 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1128 if (net_ratelimit())
1129 printk(KERN_DEBUG "retrans_out leaked.\n");
1132 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1133 tp->retrans_out += tcp_skb_pcount(skb);
1135 /* Save stamp of the first retransmit. */
1136 if (!tp->retrans_stamp)
1137 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1141 /* snd_nxt is stored to detect loss of retransmitted segment,
1142 * see tcp_input.c tcp_sacktag_write_queue().
1144 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1149 /* This gets called after a retransmit timeout, and the initially
1150 * retransmitted data is acknowledged. It tries to continue
1151 * resending the rest of the retransmit queue, until either
1152 * we've sent it all or the congestion window limit is reached.
1153 * If doing SACK, the first ACK which comes back for a timeout
1154 * based retransmit packet might feed us FACK information again.
1155 * If so, we use it to avoid unnecessarily retransmissions.
1157 void tcp_xmit_retransmit_queue(struct sock *sk)
1159 struct tcp_sock *tp = tcp_sk(sk);
1160 struct sk_buff *skb;
1161 int packet_cnt = tp->lost_out;
1163 /* First pass: retransmit lost packets. */
1165 sk_stream_for_retrans_queue(skb, sk) {
1166 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1168 /* Assume this retransmit will generate
1169 * only one packet for congestion window
1170 * calculation purposes. This works because
1171 * tcp_retransmit_skb() will chop up the
1172 * packet to be MSS sized and all the
1173 * packet counting works out.
1175 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1178 if (sacked&TCPCB_LOST) {
1179 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1180 if (tcp_retransmit_skb(sk, skb))
1182 if (tp->ca_state != TCP_CA_Loss)
1183 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1185 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1188 skb_peek(&sk->sk_write_queue))
1189 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1192 packet_cnt -= tcp_skb_pcount(skb);
1193 if (packet_cnt <= 0)
1199 /* OK, demanded retransmission is finished. */
1201 /* Forward retransmissions are possible only during Recovery. */
1202 if (tp->ca_state != TCP_CA_Recovery)
1205 /* No forward retransmissions in Reno are possible. */
1206 if (!tp->rx_opt.sack_ok)
1209 /* Yeah, we have to make difficult choice between forward transmission
1210 * and retransmission... Both ways have their merits...
1212 * For now we do not retransmit anything, while we have some new
1216 if (tcp_may_send_now(sk, tp))
1221 sk_stream_for_retrans_queue(skb, sk) {
1222 /* Similar to the retransmit loop above we
1223 * can pretend that the retransmitted SKB
1224 * we send out here will be composed of one
1225 * real MSS sized packet because tcp_retransmit_skb()
1226 * will fragment it if necessary.
1228 if (++packet_cnt > tp->fackets_out)
1231 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1234 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1237 /* Ok, retransmit it. */
1238 if (tcp_retransmit_skb(sk, skb))
1241 if (skb == skb_peek(&sk->sk_write_queue))
1242 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1244 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1249 /* Send a fin. The caller locks the socket for us. This cannot be
1250 * allowed to fail queueing a FIN frame under any circumstances.
1252 void tcp_send_fin(struct sock *sk)
1254 struct tcp_sock *tp = tcp_sk(sk);
1255 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1258 /* Optimization, tack on the FIN if we have a queue of
1259 * unsent frames. But be careful about outgoing SACKS
1262 mss_now = tcp_current_mss(sk, 1);
1264 if (sk->sk_send_head != NULL) {
1265 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1266 TCP_SKB_CB(skb)->end_seq++;
1269 /* Socket is locked, keep trying until memory is available. */
1271 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL);
1277 /* Reserve space for headers and prepare control bits. */
1278 skb_reserve(skb, MAX_TCP_HEADER);
1280 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1281 TCP_SKB_CB(skb)->sacked = 0;
1282 skb_shinfo(skb)->tso_segs = 1;
1283 skb_shinfo(skb)->tso_size = 0;
1285 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1286 TCP_SKB_CB(skb)->seq = tp->write_seq;
1287 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1288 tcp_queue_skb(sk, skb);
1290 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1293 /* We get here when a process closes a file descriptor (either due to
1294 * an explicit close() or as a byproduct of exit()'ing) and there
1295 * was unread data in the receive queue. This behavior is recommended
1296 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1298 void tcp_send_active_reset(struct sock *sk, int priority)
1300 struct tcp_sock *tp = tcp_sk(sk);
1301 struct sk_buff *skb;
1303 /* NOTE: No TCP options attached and we never retransmit this. */
1304 skb = alloc_skb(MAX_TCP_HEADER, priority);
1306 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1310 /* Reserve space for headers and prepare control bits. */
1311 skb_reserve(skb, MAX_TCP_HEADER);
1313 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1314 TCP_SKB_CB(skb)->sacked = 0;
1315 skb_shinfo(skb)->tso_segs = 1;
1316 skb_shinfo(skb)->tso_size = 0;
1319 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1320 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1321 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1322 if (tcp_transmit_skb(sk, skb))
1323 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1326 /* WARNING: This routine must only be called when we have already sent
1327 * a SYN packet that crossed the incoming SYN that caused this routine
1328 * to get called. If this assumption fails then the initial rcv_wnd
1329 * and rcv_wscale values will not be correct.
1331 int tcp_send_synack(struct sock *sk)
1333 struct sk_buff* skb;
1335 skb = skb_peek(&sk->sk_write_queue);
1336 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1337 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1340 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1341 if (skb_cloned(skb)) {
1342 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1345 __skb_unlink(skb, &sk->sk_write_queue);
1346 skb_header_release(nskb);
1347 __skb_queue_head(&sk->sk_write_queue, nskb);
1348 sk_stream_free_skb(sk, skb);
1349 sk_charge_skb(sk, nskb);
1353 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1354 TCP_ECN_send_synack(tcp_sk(sk), skb);
1356 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1357 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1361 * Prepare a SYN-ACK.
1363 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1364 struct open_request *req)
1366 struct tcp_sock *tp = tcp_sk(sk);
1368 int tcp_header_size;
1369 struct sk_buff *skb;
1371 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1375 /* Reserve space for headers. */
1376 skb_reserve(skb, MAX_TCP_HEADER);
1378 skb->dst = dst_clone(dst);
1380 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1381 (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1382 (req->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1383 /* SACK_PERM is in the place of NOP NOP of TS */
1384 ((req->sack_ok && !req->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1385 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1387 memset(th, 0, sizeof(struct tcphdr));
1390 if (dst->dev->features&NETIF_F_TSO)
1392 TCP_ECN_make_synack(req, th);
1393 th->source = inet_sk(sk)->sport;
1394 th->dest = req->rmt_port;
1395 TCP_SKB_CB(skb)->seq = req->snt_isn;
1396 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1397 TCP_SKB_CB(skb)->sacked = 0;
1398 skb_shinfo(skb)->tso_segs = 1;
1399 skb_shinfo(skb)->tso_size = 0;
1400 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1401 th->ack_seq = htonl(req->rcv_isn + 1);
1402 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1404 /* Set this up on the first call only */
1405 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1406 /* tcp_full_space because it is guaranteed to be the first packet */
1407 tcp_select_initial_window(tcp_full_space(sk),
1408 dst_metric(dst, RTAX_ADVMSS) - (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1413 req->rcv_wscale = rcv_wscale;
1416 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1417 th->window = htons(req->rcv_wnd);
1419 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1420 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), req->tstamp_ok,
1421 req->sack_ok, req->wscale_ok, req->rcv_wscale,
1422 TCP_SKB_CB(skb)->when,
1426 th->doff = (tcp_header_size >> 2);
1427 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1432 * Do all connect socket setups that can be done AF independent.
1434 static inline void tcp_connect_init(struct sock *sk)
1436 struct dst_entry *dst = __sk_dst_get(sk);
1437 struct tcp_sock *tp = tcp_sk(sk);
1440 /* We'll fix this up when we get a response from the other end.
1441 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1443 tp->tcp_header_len = sizeof(struct tcphdr) +
1444 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1446 /* If user gave his TCP_MAXSEG, record it to clamp */
1447 if (tp->rx_opt.user_mss)
1448 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1450 tcp_sync_mss(sk, dst_mtu(dst));
1452 if (!tp->window_clamp)
1453 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1454 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1455 tcp_initialize_rcv_mss(sk);
1458 tcp_select_initial_window(tcp_full_space(sk),
1459 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1462 sysctl_tcp_window_scaling,
1465 tp->rx_opt.rcv_wscale = rcv_wscale;
1466 tp->rcv_ssthresh = tp->rcv_wnd;
1469 sock_reset_flag(sk, SOCK_DONE);
1471 tcp_init_wl(tp, tp->write_seq, 0);
1472 tp->snd_una = tp->write_seq;
1473 tp->snd_sml = tp->write_seq;
1478 tp->rto = TCP_TIMEOUT_INIT;
1479 tp->retransmits = 0;
1480 tcp_clear_retrans(tp);
1484 * Build a SYN and send it off.
1486 int tcp_connect(struct sock *sk)
1488 struct tcp_sock *tp = tcp_sk(sk);
1489 struct sk_buff *buff;
1491 tcp_connect_init(sk);
1493 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation);
1494 if (unlikely(buff == NULL))
1497 /* Reserve space for headers. */
1498 skb_reserve(buff, MAX_TCP_HEADER);
1500 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1501 TCP_ECN_send_syn(sk, tp, buff);
1502 TCP_SKB_CB(buff)->sacked = 0;
1503 skb_shinfo(buff)->tso_segs = 1;
1504 skb_shinfo(buff)->tso_size = 0;
1506 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1507 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1508 tp->snd_nxt = tp->write_seq;
1509 tp->pushed_seq = tp->write_seq;
1513 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1514 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1515 skb_header_release(buff);
1516 __skb_queue_tail(&sk->sk_write_queue, buff);
1517 sk_charge_skb(sk, buff);
1518 tp->packets_out += tcp_skb_pcount(buff);
1519 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1520 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1522 /* Timer for repeating the SYN until an answer. */
1523 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1527 /* Send out a delayed ack, the caller does the policy checking
1528 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1531 void tcp_send_delayed_ack(struct sock *sk)
1533 struct tcp_sock *tp = tcp_sk(sk);
1534 int ato = tp->ack.ato;
1535 unsigned long timeout;
1537 if (ato > TCP_DELACK_MIN) {
1540 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED))
1541 max_ato = TCP_DELACK_MAX;
1543 /* Slow path, intersegment interval is "high". */
1545 /* If some rtt estimate is known, use it to bound delayed ack.
1546 * Do not use tp->rto here, use results of rtt measurements
1550 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1556 ato = min(ato, max_ato);
1559 /* Stay within the limit we were given */
1560 timeout = jiffies + ato;
1562 /* Use new timeout only if there wasn't a older one earlier. */
1563 if (tp->ack.pending&TCP_ACK_TIMER) {
1564 /* If delack timer was blocked or is about to expire,
1567 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) {
1572 if (!time_before(timeout, tp->ack.timeout))
1573 timeout = tp->ack.timeout;
1575 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER;
1576 tp->ack.timeout = timeout;
1577 sk_reset_timer(sk, &tp->delack_timer, timeout);
1580 /* This routine sends an ack and also updates the window. */
1581 void tcp_send_ack(struct sock *sk)
1583 /* If we have been reset, we may not send again. */
1584 if (sk->sk_state != TCP_CLOSE) {
1585 struct tcp_sock *tp = tcp_sk(sk);
1586 struct sk_buff *buff;
1588 /* We are not putting this on the write queue, so
1589 * tcp_transmit_skb() will set the ownership to this
1592 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1594 tcp_schedule_ack(tp);
1595 tp->ack.ato = TCP_ATO_MIN;
1596 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
1600 /* Reserve space for headers and prepare control bits. */
1601 skb_reserve(buff, MAX_TCP_HEADER);
1603 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1604 TCP_SKB_CB(buff)->sacked = 0;
1605 skb_shinfo(buff)->tso_segs = 1;
1606 skb_shinfo(buff)->tso_size = 0;
1608 /* Send it off, this clears delayed acks for us. */
1609 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1610 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1611 tcp_transmit_skb(sk, buff);
1615 /* This routine sends a packet with an out of date sequence
1616 * number. It assumes the other end will try to ack it.
1618 * Question: what should we make while urgent mode?
1619 * 4.4BSD forces sending single byte of data. We cannot send
1620 * out of window data, because we have SND.NXT==SND.MAX...
1622 * Current solution: to send TWO zero-length segments in urgent mode:
1623 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1624 * out-of-date with SND.UNA-1 to probe window.
1626 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1628 struct tcp_sock *tp = tcp_sk(sk);
1629 struct sk_buff *skb;
1631 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1632 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1636 /* Reserve space for headers and set control bits. */
1637 skb_reserve(skb, MAX_TCP_HEADER);
1639 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1640 TCP_SKB_CB(skb)->sacked = urgent;
1641 skb_shinfo(skb)->tso_segs = 1;
1642 skb_shinfo(skb)->tso_size = 0;
1644 /* Use a previous sequence. This should cause the other
1645 * end to send an ack. Don't queue or clone SKB, just
1648 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1649 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1650 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1651 return tcp_transmit_skb(sk, skb);
1654 int tcp_write_wakeup(struct sock *sk)
1656 if (sk->sk_state != TCP_CLOSE) {
1657 struct tcp_sock *tp = tcp_sk(sk);
1658 struct sk_buff *skb;
1660 if ((skb = sk->sk_send_head) != NULL &&
1661 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1663 unsigned int mss = tcp_current_mss(sk, 0);
1664 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1666 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1667 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1669 /* We are probing the opening of a window
1670 * but the window size is != 0
1671 * must have been a result SWS avoidance ( sender )
1673 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1675 seg_size = min(seg_size, mss);
1676 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1677 if (tcp_fragment(sk, skb, seg_size))
1679 /* SWS override triggered forced fragmentation.
1680 * Disable TSO, the connection is too sick. */
1681 if (sk->sk_route_caps & NETIF_F_TSO) {
1682 sock_set_flag(sk, SOCK_NO_LARGESEND);
1683 sk->sk_route_caps &= ~NETIF_F_TSO;
1684 tp->mss_cache = tp->mss_cache_std;
1686 } else if (!tcp_skb_pcount(skb))
1687 tcp_set_skb_tso_segs(sk, skb);
1689 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1690 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1691 tcp_tso_set_push(skb);
1692 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1694 update_send_head(sk, tp, skb);
1699 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
1700 tcp_xmit_probe_skb(sk, TCPCB_URG);
1701 return tcp_xmit_probe_skb(sk, 0);
1707 /* A window probe timeout has occurred. If window is not closed send
1708 * a partial packet else a zero probe.
1710 void tcp_send_probe0(struct sock *sk)
1712 struct tcp_sock *tp = tcp_sk(sk);
1715 err = tcp_write_wakeup(sk);
1717 if (tp->packets_out || !sk->sk_send_head) {
1718 /* Cancel probe timer, if it is not required. */
1725 if (tp->backoff < sysctl_tcp_retries2)
1728 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
1729 min(tp->rto << tp->backoff, TCP_RTO_MAX));
1731 /* If packet was not sent due to local congestion,
1732 * do not backoff and do not remember probes_out.
1733 * Let local senders to fight for local resources.
1735 * Use accumulated backoff yet.
1737 if (!tp->probes_out)
1739 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
1740 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL));
1744 EXPORT_SYMBOL(tcp_connect);
1745 EXPORT_SYMBOL(tcp_make_synack);
1746 EXPORT_SYMBOL(tcp_simple_retransmit);
1747 EXPORT_SYMBOL(tcp_sync_mss);