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 = 3;
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 sock *sk, struct dst_entry *dst)
110 struct tcp_sock *tp = tcp_sk(sk);
111 s32 delta = tcp_time_stamp - tp->lsndtime;
112 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
113 u32 cwnd = tp->snd_cwnd;
115 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
117 tp->snd_ssthresh = tcp_current_ssthresh(sk);
118 restart_cwnd = min(restart_cwnd, cwnd);
120 while ((delta -= inet_csk(sk)->icsk_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 struct inet_connection_sock *icsk = inet_csk(sk);
131 const u32 now = tcp_time_stamp;
133 if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)
134 tcp_cwnd_restart(sk, __sk_dst_get(sk));
138 /* If it is a reply for ato after last received
139 * packet, enter pingpong mode.
141 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
142 icsk->icsk_ack.pingpong = 1;
145 static __inline__ void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
147 tcp_dec_quickack_mode(sk, pkts);
148 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
151 /* Determine a window scaling and initial window to offer.
152 * Based on the assumption that the given amount of space
153 * will be offered. Store the results in the tp structure.
154 * NOTE: for smooth operation initial space offering should
155 * be a multiple of mss if possible. We assume here that mss >= 1.
156 * This MUST be enforced by all callers.
158 void tcp_select_initial_window(int __space, __u32 mss,
159 __u32 *rcv_wnd, __u32 *window_clamp,
160 int wscale_ok, __u8 *rcv_wscale)
162 unsigned int space = (__space < 0 ? 0 : __space);
164 /* If no clamp set the clamp to the max possible scaled window */
165 if (*window_clamp == 0)
166 (*window_clamp) = (65535 << 14);
167 space = min(*window_clamp, space);
169 /* Quantize space offering to a multiple of mss if possible. */
171 space = (space / mss) * mss;
173 /* NOTE: offering an initial window larger than 32767
174 * will break some buggy TCP stacks. We try to be nice.
175 * If we are not window scaling, then this truncates
176 * our initial window offering to 32k. There should also
177 * be a sysctl option to stop being nice.
179 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
182 /* Set window scaling on max possible window
183 * See RFC1323 for an explanation of the limit to 14
185 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
186 while (space > 65535 && (*rcv_wscale) < 14) {
192 /* Set initial window to value enough for senders,
193 * following RFC2414. Senders, not following this RFC,
194 * will be satisfied with 2.
196 if (mss > (1<<*rcv_wscale)) {
202 init_cwnd = (mss > 1095) ? 3 : 4;
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 const struct inet_connection_sock *icsk = inet_csk(sk);
270 struct inet_sock *inet = inet_sk(sk);
271 struct tcp_sock *tp = tcp_sk(sk);
272 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
273 int tcp_header_size = tp->tcp_header_len;
278 BUG_ON(!tcp_skb_pcount(skb));
280 #define SYSCTL_FLAG_TSTAMPS 0x1
281 #define SYSCTL_FLAG_WSCALE 0x2
282 #define SYSCTL_FLAG_SACK 0x4
284 /* If congestion control is doing timestamping */
285 if (icsk->icsk_ca_ops->rtt_sample)
286 __net_timestamp(skb);
289 if (tcb->flags & TCPCB_FLAG_SYN) {
290 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
291 if(sysctl_tcp_timestamps) {
292 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
293 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
295 if(sysctl_tcp_window_scaling) {
296 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
297 sysctl_flags |= SYSCTL_FLAG_WSCALE;
299 if(sysctl_tcp_sack) {
300 sysctl_flags |= SYSCTL_FLAG_SACK;
301 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
302 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
304 } else if (tp->rx_opt.eff_sacks) {
305 /* A SACK is 2 pad bytes, a 2 byte header, plus
306 * 2 32-bit sequence numbers for each SACK block.
308 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
309 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
312 if (tcp_packets_in_flight(tp) == 0)
313 tcp_ca_event(sk, CA_EVENT_TX_START);
315 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
317 skb_set_owner_w(skb, sk);
319 /* Build TCP header and checksum it. */
320 th->source = inet->sport;
321 th->dest = inet->dport;
322 th->seq = htonl(tcb->seq);
323 th->ack_seq = htonl(tp->rcv_nxt);
324 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
325 if (tcb->flags & TCPCB_FLAG_SYN) {
326 /* RFC1323: The window in SYN & SYN/ACK segments
329 th->window = htons(tp->rcv_wnd);
331 th->window = htons(tcp_select_window(sk));
337 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
338 th->urg_ptr = htons(tp->snd_up-tcb->seq);
342 if (tcb->flags & TCPCB_FLAG_SYN) {
343 tcp_syn_build_options((__u32 *)(th + 1),
344 tcp_advertise_mss(sk),
345 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
346 (sysctl_flags & SYSCTL_FLAG_SACK),
347 (sysctl_flags & SYSCTL_FLAG_WSCALE),
348 tp->rx_opt.rcv_wscale,
350 tp->rx_opt.ts_recent);
352 tcp_build_and_update_options((__u32 *)(th + 1),
355 TCP_ECN_send(sk, tp, skb, tcp_header_size);
357 tp->af_specific->send_check(sk, th, skb->len, skb);
359 if (tcb->flags & TCPCB_FLAG_ACK)
360 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
362 if (skb->len != tcp_header_size)
363 tcp_event_data_sent(tp, skb, sk);
365 TCP_INC_STATS(TCP_MIB_OUTSEGS);
367 err = tp->af_specific->queue_xmit(skb, 0);
373 /* NET_XMIT_CN is special. It does not guarantee,
374 * that this packet is lost. It tells that device
375 * is about to start to drop packets or already
376 * drops some packets of the same priority and
377 * invokes us to send less aggressively.
379 return err == NET_XMIT_CN ? 0 : err;
382 #undef SYSCTL_FLAG_TSTAMPS
383 #undef SYSCTL_FLAG_WSCALE
384 #undef SYSCTL_FLAG_SACK
388 /* This routine just queue's the buffer
390 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
391 * otherwise socket can stall.
393 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
395 struct tcp_sock *tp = tcp_sk(sk);
397 /* Advance write_seq and place onto the write_queue. */
398 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
399 skb_header_release(skb);
400 __skb_queue_tail(&sk->sk_write_queue, skb);
401 sk_charge_skb(sk, skb);
403 /* Queue it, remembering where we must start sending. */
404 if (sk->sk_send_head == NULL)
405 sk->sk_send_head = skb;
408 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
410 if (skb->len <= mss_now ||
411 !(sk->sk_route_caps & NETIF_F_TSO)) {
412 /* Avoid the costly divide in the normal
415 skb_shinfo(skb)->tso_segs = 1;
416 skb_shinfo(skb)->tso_size = 0;
420 factor = skb->len + (mss_now - 1);
422 skb_shinfo(skb)->tso_segs = factor;
423 skb_shinfo(skb)->tso_size = mss_now;
427 /* Function to create two new TCP segments. Shrinks the given segment
428 * to the specified size and appends a new segment with the rest of the
429 * packet to the list. This won't be called frequently, I hope.
430 * Remember, these are still headerless SKBs at this point.
432 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
434 struct tcp_sock *tp = tcp_sk(sk);
435 struct sk_buff *buff;
436 int nsize, old_factor;
439 BUG_ON(len >= skb->len);
441 nsize = skb_headlen(skb) - len;
445 if (skb_cloned(skb) &&
446 skb_is_nonlinear(skb) &&
447 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
450 /* Get a new skb... force flag on. */
451 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
453 return -ENOMEM; /* We'll just try again later. */
454 sk_charge_skb(sk, buff);
456 /* Correct the sequence numbers. */
457 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
458 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
459 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
461 /* PSH and FIN should only be set in the second packet. */
462 flags = TCP_SKB_CB(skb)->flags;
463 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
464 TCP_SKB_CB(buff)->flags = flags;
465 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
466 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
468 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
469 /* Copy and checksum data tail into the new buffer. */
470 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
475 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
477 skb->ip_summed = CHECKSUM_HW;
478 skb_split(skb, buff, len);
481 buff->ip_summed = skb->ip_summed;
483 /* Looks stupid, but our code really uses when of
484 * skbs, which it never sent before. --ANK
486 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
487 buff->tstamp = skb->tstamp;
489 old_factor = tcp_skb_pcount(skb);
491 /* Fix up tso_factor for both original and new SKB. */
492 tcp_set_skb_tso_segs(sk, skb, mss_now);
493 tcp_set_skb_tso_segs(sk, buff, mss_now);
495 /* If this packet has been sent out already, we must
496 * adjust the various packet counters.
498 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
499 int diff = old_factor - tcp_skb_pcount(skb) -
500 tcp_skb_pcount(buff);
502 tp->packets_out -= diff;
504 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
505 tp->sacked_out -= diff;
506 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
507 tp->retrans_out -= diff;
509 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
510 tp->lost_out -= diff;
511 tp->left_out -= diff;
515 /* Adjust Reno SACK estimate. */
516 if (!tp->rx_opt.sack_ok) {
517 tp->sacked_out -= diff;
518 if ((int)tp->sacked_out < 0)
520 tcp_sync_left_out(tp);
523 tp->fackets_out -= diff;
524 if ((int)tp->fackets_out < 0)
529 /* Link BUFF into the send queue. */
530 skb_header_release(buff);
531 __skb_append(skb, buff, &sk->sk_write_queue);
536 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
537 * eventually). The difference is that pulled data not copied, but
538 * immediately discarded.
540 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
546 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
547 if (skb_shinfo(skb)->frags[i].size <= eat) {
548 put_page(skb_shinfo(skb)->frags[i].page);
549 eat -= skb_shinfo(skb)->frags[i].size;
551 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
553 skb_shinfo(skb)->frags[k].page_offset += eat;
554 skb_shinfo(skb)->frags[k].size -= eat;
560 skb_shinfo(skb)->nr_frags = k;
562 skb->tail = skb->data;
563 skb->data_len -= len;
564 skb->len = skb->data_len;
568 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
570 if (skb_cloned(skb) &&
571 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
574 if (len <= skb_headlen(skb)) {
575 __skb_pull(skb, len);
577 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
581 TCP_SKB_CB(skb)->seq += len;
582 skb->ip_summed = CHECKSUM_HW;
584 skb->truesize -= len;
585 sk->sk_wmem_queued -= len;
586 sk->sk_forward_alloc += len;
587 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
589 /* Any change of skb->len requires recalculation of tso
592 if (tcp_skb_pcount(skb) > 1)
593 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
598 /* This function synchronize snd mss to current pmtu/exthdr set.
600 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
601 for TCP options, but includes only bare TCP header.
603 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
604 It is minumum of user_mss and mss received with SYN.
605 It also does not include TCP options.
607 tp->pmtu_cookie is last pmtu, seen by this function.
609 tp->mss_cache is current effective sending mss, including
610 all tcp options except for SACKs. It is evaluated,
611 taking into account current pmtu, but never exceeds
612 tp->rx_opt.mss_clamp.
614 NOTE1. rfc1122 clearly states that advertised MSS
615 DOES NOT include either tcp or ip options.
617 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
618 this function. --ANK (980731)
621 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
623 struct tcp_sock *tp = tcp_sk(sk);
626 /* Calculate base mss without TCP options:
627 It is MMS_S - sizeof(tcphdr) of rfc1122
629 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
631 /* Clamp it (mss_clamp does not include tcp options) */
632 if (mss_now > tp->rx_opt.mss_clamp)
633 mss_now = tp->rx_opt.mss_clamp;
635 /* Now subtract optional transport overhead */
636 mss_now -= tp->ext_header_len;
638 /* Then reserve room for full set of TCP options and 8 bytes of data */
642 /* Now subtract TCP options size, not including SACKs */
643 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
645 /* Bound mss with half of window */
646 if (tp->max_window && mss_now > (tp->max_window>>1))
647 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
649 /* And store cached results */
650 tp->pmtu_cookie = pmtu;
651 tp->mss_cache = mss_now;
656 /* Compute the current effective MSS, taking SACKs and IP options,
657 * and even PMTU discovery events into account.
659 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
660 * cannot be large. However, taking into account rare use of URG, this
663 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
665 struct tcp_sock *tp = tcp_sk(sk);
666 struct dst_entry *dst = __sk_dst_get(sk);
671 mss_now = tp->mss_cache;
674 (sk->sk_route_caps & NETIF_F_TSO) &&
679 u32 mtu = dst_mtu(dst);
680 if (mtu != tp->pmtu_cookie)
681 mss_now = tcp_sync_mss(sk, mtu);
684 if (tp->rx_opt.eff_sacks)
685 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
686 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
688 xmit_size_goal = mss_now;
691 xmit_size_goal = 65535 -
692 tp->af_specific->net_header_len -
693 tp->ext_header_len - tp->tcp_header_len;
695 if (tp->max_window &&
696 (xmit_size_goal > (tp->max_window >> 1)))
697 xmit_size_goal = max((tp->max_window >> 1),
698 68U - tp->tcp_header_len);
700 xmit_size_goal -= (xmit_size_goal % mss_now);
702 tp->xmit_size_goal = xmit_size_goal;
707 /* Congestion window validation. (RFC2861) */
709 static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
711 __u32 packets_out = tp->packets_out;
713 if (packets_out >= tp->snd_cwnd) {
714 /* Network is feed fully. */
715 tp->snd_cwnd_used = 0;
716 tp->snd_cwnd_stamp = tcp_time_stamp;
718 /* Network starves. */
719 if (tp->packets_out > tp->snd_cwnd_used)
720 tp->snd_cwnd_used = tp->packets_out;
722 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
723 tcp_cwnd_application_limited(sk);
727 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
729 u32 window, cwnd_len;
731 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
732 cwnd_len = mss_now * cwnd;
733 return min(window, cwnd_len);
736 /* Can at least one segment of SKB be sent right now, according to the
737 * congestion window rules? If so, return how many segments are allowed.
739 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
743 /* Don't be strict about the congestion window for the final FIN. */
744 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
747 in_flight = tcp_packets_in_flight(tp);
749 if (in_flight < cwnd)
750 return (cwnd - in_flight);
755 /* This must be invoked the first time we consider transmitting
758 static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
760 int tso_segs = tcp_skb_pcount(skb);
764 skb_shinfo(skb)->tso_size != mss_now)) {
765 tcp_set_skb_tso_segs(sk, skb, mss_now);
766 tso_segs = tcp_skb_pcount(skb);
771 static inline int tcp_minshall_check(const struct tcp_sock *tp)
773 return after(tp->snd_sml,tp->snd_una) &&
774 !after(tp->snd_sml, tp->snd_nxt);
777 /* Return 0, if packet can be sent now without violation Nagle's rules:
778 * 1. It is full sized.
779 * 2. Or it contains FIN. (already checked by caller)
780 * 3. Or TCP_NODELAY was set.
781 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
782 * With Minshall's modification: all sent small packets are ACKed.
785 static inline int tcp_nagle_check(const struct tcp_sock *tp,
786 const struct sk_buff *skb,
787 unsigned mss_now, int nonagle)
789 return (skb->len < mss_now &&
790 ((nonagle&TCP_NAGLE_CORK) ||
793 tcp_minshall_check(tp))));
796 /* Return non-zero if the Nagle test allows this packet to be
799 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
800 unsigned int cur_mss, int nonagle)
802 /* Nagle rule does not apply to frames, which sit in the middle of the
803 * write_queue (they have no chances to get new data).
805 * This is implemented in the callers, where they modify the 'nonagle'
806 * argument based upon the location of SKB in the send queue.
808 if (nonagle & TCP_NAGLE_PUSH)
811 /* Don't use the nagle rule for urgent data (or for the final FIN). */
813 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
816 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
822 /* Does at least the first segment of SKB fit into the send window? */
823 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
825 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
827 if (skb->len > cur_mss)
828 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
830 return !after(end_seq, tp->snd_una + tp->snd_wnd);
833 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
834 * should be put on the wire right now. If so, it returns the number of
835 * packets allowed by the congestion window.
837 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
838 unsigned int cur_mss, int nonagle)
840 struct tcp_sock *tp = tcp_sk(sk);
841 unsigned int cwnd_quota;
843 tcp_init_tso_segs(sk, skb, cur_mss);
845 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
848 cwnd_quota = tcp_cwnd_test(tp, skb);
850 !tcp_snd_wnd_test(tp, skb, cur_mss))
856 static inline int tcp_skb_is_last(const struct sock *sk,
857 const struct sk_buff *skb)
859 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
862 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
864 struct sk_buff *skb = sk->sk_send_head;
867 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
868 (tcp_skb_is_last(sk, skb) ?
873 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
874 * which is put after SKB on the list. It is very much like
875 * tcp_fragment() except that it may make several kinds of assumptions
876 * in order to speed up the splitting operation. In particular, we
877 * know that all the data is in scatter-gather pages, and that the
878 * packet has never been sent out before (and thus is not cloned).
880 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
882 struct sk_buff *buff;
883 int nlen = skb->len - len;
886 /* All of a TSO frame must be composed of paged data. */
887 if (skb->len != skb->data_len)
888 return tcp_fragment(sk, skb, len, mss_now);
890 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
891 if (unlikely(buff == NULL))
894 buff->truesize = nlen;
895 skb->truesize -= nlen;
897 /* Correct the sequence numbers. */
898 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
899 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
900 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
902 /* PSH and FIN should only be set in the second packet. */
903 flags = TCP_SKB_CB(skb)->flags;
904 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
905 TCP_SKB_CB(buff)->flags = flags;
907 /* This packet was never sent out yet, so no SACK bits. */
908 TCP_SKB_CB(buff)->sacked = 0;
910 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
911 skb_split(skb, buff, len);
913 /* Fix up tso_factor for both original and new SKB. */
914 tcp_set_skb_tso_segs(sk, skb, mss_now);
915 tcp_set_skb_tso_segs(sk, buff, mss_now);
917 /* Link BUFF into the send queue. */
918 skb_header_release(buff);
919 __skb_append(skb, buff, &sk->sk_write_queue);
924 /* Try to defer sending, if possible, in order to minimize the amount
925 * of TSO splitting we do. View it as a kind of TSO Nagle test.
927 * This algorithm is from John Heffner.
929 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
931 const struct inet_connection_sock *icsk = inet_csk(sk);
932 u32 send_win, cong_win, limit, in_flight;
934 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
937 if (icsk->icsk_ca_state != TCP_CA_Open)
940 in_flight = tcp_packets_in_flight(tp);
942 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
943 (tp->snd_cwnd <= in_flight));
945 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
947 /* From in_flight test above, we know that cwnd > in_flight. */
948 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
950 limit = min(send_win, cong_win);
952 if (sysctl_tcp_tso_win_divisor) {
953 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
955 /* If at least some fraction of a window is available,
958 chunk /= sysctl_tcp_tso_win_divisor;
962 /* Different approach, try not to defer past a single
963 * ACK. Receiver should ACK every other full sized
964 * frame, so if we have space for more than 3 frames
967 if (limit > tcp_max_burst(tp) * tp->mss_cache)
971 /* Ok, it looks like it is advisable to defer. */
975 /* This routine writes packets to the network. It advances the
976 * send_head. This happens as incoming acks open up the remote
979 * Returns 1, if no segments are in flight and we have queued segments, but
980 * cannot send anything now because of SWS or another problem.
982 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
984 struct tcp_sock *tp = tcp_sk(sk);
986 unsigned int tso_segs, sent_pkts;
989 /* If we are closed, the bytes will have to remain here.
990 * In time closedown will finish, we empty the write queue and all
993 if (unlikely(sk->sk_state == TCP_CLOSE))
997 while ((skb = sk->sk_send_head)) {
1000 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1003 cwnd_quota = tcp_cwnd_test(tp, skb);
1007 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1010 if (tso_segs == 1) {
1011 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1012 (tcp_skb_is_last(sk, skb) ?
1013 nonagle : TCP_NAGLE_PUSH))))
1016 if (tcp_tso_should_defer(sk, tp, skb))
1022 limit = tcp_window_allows(tp, skb,
1023 mss_now, cwnd_quota);
1025 if (skb->len < limit) {
1026 unsigned int trim = skb->len % mss_now;
1029 limit = skb->len - trim;
1033 if (skb->len > limit &&
1034 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1037 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1039 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
1042 /* Advance the send_head. This one is sent out.
1043 * This call will increment packets_out.
1045 update_send_head(sk, tp, skb);
1047 tcp_minshall_update(tp, mss_now, skb);
1051 if (likely(sent_pkts)) {
1052 tcp_cwnd_validate(sk, tp);
1055 return !tp->packets_out && sk->sk_send_head;
1058 /* Push out any pending frames which were held back due to
1059 * TCP_CORK or attempt at coalescing tiny packets.
1060 * The socket must be locked by the caller.
1062 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1063 unsigned int cur_mss, int nonagle)
1065 struct sk_buff *skb = sk->sk_send_head;
1068 if (tcp_write_xmit(sk, cur_mss, nonagle))
1069 tcp_check_probe_timer(sk, tp);
1073 /* Send _single_ skb sitting at the send head. This function requires
1074 * true push pending frames to setup probe timer etc.
1076 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1078 struct tcp_sock *tp = tcp_sk(sk);
1079 struct sk_buff *skb = sk->sk_send_head;
1080 unsigned int tso_segs, cwnd_quota;
1082 BUG_ON(!skb || skb->len < mss_now);
1084 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1085 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1087 if (likely(cwnd_quota)) {
1094 limit = tcp_window_allows(tp, skb,
1095 mss_now, cwnd_quota);
1097 if (skb->len < limit) {
1098 unsigned int trim = skb->len % mss_now;
1101 limit = skb->len - trim;
1105 if (skb->len > limit &&
1106 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1109 /* Send it out now. */
1110 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1112 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1113 update_send_head(sk, tp, skb);
1114 tcp_cwnd_validate(sk, tp);
1120 /* This function returns the amount that we can raise the
1121 * usable window based on the following constraints
1123 * 1. The window can never be shrunk once it is offered (RFC 793)
1124 * 2. We limit memory per socket
1127 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1128 * RECV.NEXT + RCV.WIN fixed until:
1129 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1131 * i.e. don't raise the right edge of the window until you can raise
1132 * it at least MSS bytes.
1134 * Unfortunately, the recommended algorithm breaks header prediction,
1135 * since header prediction assumes th->window stays fixed.
1137 * Strictly speaking, keeping th->window fixed violates the receiver
1138 * side SWS prevention criteria. The problem is that under this rule
1139 * a stream of single byte packets will cause the right side of the
1140 * window to always advance by a single byte.
1142 * Of course, if the sender implements sender side SWS prevention
1143 * then this will not be a problem.
1145 * BSD seems to make the following compromise:
1147 * If the free space is less than the 1/4 of the maximum
1148 * space available and the free space is less than 1/2 mss,
1149 * then set the window to 0.
1150 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1151 * Otherwise, just prevent the window from shrinking
1152 * and from being larger than the largest representable value.
1154 * This prevents incremental opening of the window in the regime
1155 * where TCP is limited by the speed of the reader side taking
1156 * data out of the TCP receive queue. It does nothing about
1157 * those cases where the window is constrained on the sender side
1158 * because the pipeline is full.
1160 * BSD also seems to "accidentally" limit itself to windows that are a
1161 * multiple of MSS, at least until the free space gets quite small.
1162 * This would appear to be a side effect of the mbuf implementation.
1163 * Combining these two algorithms results in the observed behavior
1164 * of having a fixed window size at almost all times.
1166 * Below we obtain similar behavior by forcing the offered window to
1167 * a multiple of the mss when it is feasible to do so.
1169 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1170 * Regular options like TIMESTAMP are taken into account.
1172 u32 __tcp_select_window(struct sock *sk)
1174 struct inet_connection_sock *icsk = inet_csk(sk);
1175 struct tcp_sock *tp = tcp_sk(sk);
1176 /* MSS for the peer's data. Previous verions used mss_clamp
1177 * here. I don't know if the value based on our guesses
1178 * of peer's MSS is better for the performance. It's more correct
1179 * but may be worse for the performance because of rcv_mss
1180 * fluctuations. --SAW 1998/11/1
1182 int mss = icsk->icsk_ack.rcv_mss;
1183 int free_space = tcp_space(sk);
1184 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1187 if (mss > full_space)
1190 if (free_space < full_space/2) {
1191 icsk->icsk_ack.quick = 0;
1193 if (tcp_memory_pressure)
1194 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1196 if (free_space < mss)
1200 if (free_space > tp->rcv_ssthresh)
1201 free_space = tp->rcv_ssthresh;
1203 /* Don't do rounding if we are using window scaling, since the
1204 * scaled window will not line up with the MSS boundary anyway.
1206 window = tp->rcv_wnd;
1207 if (tp->rx_opt.rcv_wscale) {
1208 window = free_space;
1210 /* Advertise enough space so that it won't get scaled away.
1211 * Import case: prevent zero window announcement if
1212 * 1<<rcv_wscale > mss.
1214 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1215 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1216 << tp->rx_opt.rcv_wscale);
1218 /* Get the largest window that is a nice multiple of mss.
1219 * Window clamp already applied above.
1220 * If our current window offering is within 1 mss of the
1221 * free space we just keep it. This prevents the divide
1222 * and multiply from happening most of the time.
1223 * We also don't do any window rounding when the free space
1226 if (window <= free_space - mss || window > free_space)
1227 window = (free_space/mss)*mss;
1233 /* Attempt to collapse two adjacent SKB's during retransmission. */
1234 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1236 struct tcp_sock *tp = tcp_sk(sk);
1237 struct sk_buff *next_skb = skb->next;
1239 /* The first test we must make is that neither of these two
1240 * SKB's are still referenced by someone else.
1242 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1243 int skb_size = skb->len, next_skb_size = next_skb->len;
1244 u16 flags = TCP_SKB_CB(skb)->flags;
1246 /* Also punt if next skb has been SACK'd. */
1247 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1250 /* Next skb is out of window. */
1251 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1254 /* Punt if not enough space exists in the first SKB for
1255 * the data in the second, or the total combined payload
1256 * would exceed the MSS.
1258 if ((next_skb_size > skb_tailroom(skb)) ||
1259 ((skb_size + next_skb_size) > mss_now))
1262 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1263 tcp_skb_pcount(next_skb) != 1);
1265 /* Ok. We will be able to collapse the packet. */
1266 __skb_unlink(next_skb, &sk->sk_write_queue);
1268 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1270 if (next_skb->ip_summed == CHECKSUM_HW)
1271 skb->ip_summed = CHECKSUM_HW;
1273 if (skb->ip_summed != CHECKSUM_HW)
1274 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1276 /* Update sequence range on original skb. */
1277 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1279 /* Merge over control information. */
1280 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1281 TCP_SKB_CB(skb)->flags = flags;
1283 /* All done, get rid of second SKB and account for it so
1284 * packet counting does not break.
1286 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1287 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1288 tp->retrans_out -= tcp_skb_pcount(next_skb);
1289 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1290 tp->lost_out -= tcp_skb_pcount(next_skb);
1291 tp->left_out -= tcp_skb_pcount(next_skb);
1293 /* Reno case is special. Sigh... */
1294 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1295 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1296 tp->left_out -= tcp_skb_pcount(next_skb);
1299 /* Not quite right: it can be > snd.fack, but
1300 * it is better to underestimate fackets.
1302 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1303 tcp_packets_out_dec(tp, next_skb);
1304 sk_stream_free_skb(sk, next_skb);
1308 /* Do a simple retransmit without using the backoff mechanisms in
1309 * tcp_timer. This is used for path mtu discovery.
1310 * The socket is already locked here.
1312 void tcp_simple_retransmit(struct sock *sk)
1314 const struct inet_connection_sock *icsk = inet_csk(sk);
1315 struct tcp_sock *tp = tcp_sk(sk);
1316 struct sk_buff *skb;
1317 unsigned int mss = tcp_current_mss(sk, 0);
1320 sk_stream_for_retrans_queue(skb, sk) {
1321 if (skb->len > mss &&
1322 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1323 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1324 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1325 tp->retrans_out -= tcp_skb_pcount(skb);
1327 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1328 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1329 tp->lost_out += tcp_skb_pcount(skb);
1338 tcp_sync_left_out(tp);
1340 /* Don't muck with the congestion window here.
1341 * Reason is that we do not increase amount of _data_
1342 * in network, but units changed and effective
1343 * cwnd/ssthresh really reduced now.
1345 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1346 tp->high_seq = tp->snd_nxt;
1347 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1348 tp->prior_ssthresh = 0;
1349 tp->undo_marker = 0;
1350 tcp_set_ca_state(sk, TCP_CA_Loss);
1352 tcp_xmit_retransmit_queue(sk);
1355 /* This retransmits one SKB. Policy decisions and retransmit queue
1356 * state updates are done by the caller. Returns non-zero if an
1357 * error occurred which prevented the send.
1359 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1361 struct tcp_sock *tp = tcp_sk(sk);
1362 unsigned int cur_mss = tcp_current_mss(sk, 0);
1365 /* Do not sent more than we queued. 1/4 is reserved for possible
1366 * copying overhead: frgagmentation, tunneling, mangling etc.
1368 if (atomic_read(&sk->sk_wmem_alloc) >
1369 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1372 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1373 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1375 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1379 /* If receiver has shrunk his window, and skb is out of
1380 * new window, do not retransmit it. The exception is the
1381 * case, when window is shrunk to zero. In this case
1382 * our retransmit serves as a zero window probe.
1384 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1385 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1388 if (skb->len > cur_mss) {
1389 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1390 return -ENOMEM; /* We'll try again later. */
1393 /* Collapse two adjacent packets if worthwhile and we can. */
1394 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1395 (skb->len < (cur_mss >> 1)) &&
1396 (skb->next != sk->sk_send_head) &&
1397 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1398 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1399 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1400 (sysctl_tcp_retrans_collapse != 0))
1401 tcp_retrans_try_collapse(sk, skb, cur_mss);
1403 if(tp->af_specific->rebuild_header(sk))
1404 return -EHOSTUNREACH; /* Routing failure or similar. */
1406 /* Some Solaris stacks overoptimize and ignore the FIN on a
1407 * retransmit when old data is attached. So strip it off
1408 * since it is cheap to do so and saves bytes on the network.
1411 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1412 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1413 if (!pskb_trim(skb, 0)) {
1414 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1415 skb_shinfo(skb)->tso_segs = 1;
1416 skb_shinfo(skb)->tso_size = 0;
1417 skb->ip_summed = CHECKSUM_NONE;
1422 /* Make a copy, if the first transmission SKB clone we made
1423 * is still in somebody's hands, else make a clone.
1425 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1427 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1428 pskb_copy(skb, GFP_ATOMIC):
1429 skb_clone(skb, GFP_ATOMIC)));
1432 /* Update global TCP statistics. */
1433 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1435 tp->total_retrans++;
1437 #if FASTRETRANS_DEBUG > 0
1438 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1439 if (net_ratelimit())
1440 printk(KERN_DEBUG "retrans_out leaked.\n");
1443 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1444 tp->retrans_out += tcp_skb_pcount(skb);
1446 /* Save stamp of the first retransmit. */
1447 if (!tp->retrans_stamp)
1448 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1452 /* snd_nxt is stored to detect loss of retransmitted segment,
1453 * see tcp_input.c tcp_sacktag_write_queue().
1455 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1460 /* This gets called after a retransmit timeout, and the initially
1461 * retransmitted data is acknowledged. It tries to continue
1462 * resending the rest of the retransmit queue, until either
1463 * we've sent it all or the congestion window limit is reached.
1464 * If doing SACK, the first ACK which comes back for a timeout
1465 * based retransmit packet might feed us FACK information again.
1466 * If so, we use it to avoid unnecessarily retransmissions.
1468 void tcp_xmit_retransmit_queue(struct sock *sk)
1470 const struct inet_connection_sock *icsk = inet_csk(sk);
1471 struct tcp_sock *tp = tcp_sk(sk);
1472 struct sk_buff *skb;
1473 int packet_cnt = tp->lost_out;
1475 /* First pass: retransmit lost packets. */
1477 sk_stream_for_retrans_queue(skb, sk) {
1478 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1480 /* Assume this retransmit will generate
1481 * only one packet for congestion window
1482 * calculation purposes. This works because
1483 * tcp_retransmit_skb() will chop up the
1484 * packet to be MSS sized and all the
1485 * packet counting works out.
1487 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1490 if (sacked&TCPCB_LOST) {
1491 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1492 if (tcp_retransmit_skb(sk, skb))
1494 if (icsk->icsk_ca_state != TCP_CA_Loss)
1495 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1497 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1500 skb_peek(&sk->sk_write_queue))
1501 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1502 inet_csk(sk)->icsk_rto,
1506 packet_cnt -= tcp_skb_pcount(skb);
1507 if (packet_cnt <= 0)
1513 /* OK, demanded retransmission is finished. */
1515 /* Forward retransmissions are possible only during Recovery. */
1516 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1519 /* No forward retransmissions in Reno are possible. */
1520 if (!tp->rx_opt.sack_ok)
1523 /* Yeah, we have to make difficult choice between forward transmission
1524 * and retransmission... Both ways have their merits...
1526 * For now we do not retransmit anything, while we have some new
1530 if (tcp_may_send_now(sk, tp))
1535 sk_stream_for_retrans_queue(skb, sk) {
1536 /* Similar to the retransmit loop above we
1537 * can pretend that the retransmitted SKB
1538 * we send out here will be composed of one
1539 * real MSS sized packet because tcp_retransmit_skb()
1540 * will fragment it if necessary.
1542 if (++packet_cnt > tp->fackets_out)
1545 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1548 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1551 /* Ok, retransmit it. */
1552 if (tcp_retransmit_skb(sk, skb))
1555 if (skb == skb_peek(&sk->sk_write_queue))
1556 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1557 inet_csk(sk)->icsk_rto,
1560 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1565 /* Send a fin. The caller locks the socket for us. This cannot be
1566 * allowed to fail queueing a FIN frame under any circumstances.
1568 void tcp_send_fin(struct sock *sk)
1570 struct tcp_sock *tp = tcp_sk(sk);
1571 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1574 /* Optimization, tack on the FIN if we have a queue of
1575 * unsent frames. But be careful about outgoing SACKS
1578 mss_now = tcp_current_mss(sk, 1);
1580 if (sk->sk_send_head != NULL) {
1581 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1582 TCP_SKB_CB(skb)->end_seq++;
1585 /* Socket is locked, keep trying until memory is available. */
1587 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1593 /* Reserve space for headers and prepare control bits. */
1594 skb_reserve(skb, MAX_TCP_HEADER);
1596 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1597 TCP_SKB_CB(skb)->sacked = 0;
1598 skb_shinfo(skb)->tso_segs = 1;
1599 skb_shinfo(skb)->tso_size = 0;
1601 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1602 TCP_SKB_CB(skb)->seq = tp->write_seq;
1603 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1604 tcp_queue_skb(sk, skb);
1606 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1609 /* We get here when a process closes a file descriptor (either due to
1610 * an explicit close() or as a byproduct of exit()'ing) and there
1611 * was unread data in the receive queue. This behavior is recommended
1612 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1614 void tcp_send_active_reset(struct sock *sk, unsigned int __nocast priority)
1616 struct tcp_sock *tp = tcp_sk(sk);
1617 struct sk_buff *skb;
1619 /* NOTE: No TCP options attached and we never retransmit this. */
1620 skb = alloc_skb(MAX_TCP_HEADER, priority);
1622 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1626 /* Reserve space for headers and prepare control bits. */
1627 skb_reserve(skb, MAX_TCP_HEADER);
1629 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1630 TCP_SKB_CB(skb)->sacked = 0;
1631 skb_shinfo(skb)->tso_segs = 1;
1632 skb_shinfo(skb)->tso_size = 0;
1635 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1636 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1637 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1638 if (tcp_transmit_skb(sk, skb))
1639 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1642 /* WARNING: This routine must only be called when we have already sent
1643 * a SYN packet that crossed the incoming SYN that caused this routine
1644 * to get called. If this assumption fails then the initial rcv_wnd
1645 * and rcv_wscale values will not be correct.
1647 int tcp_send_synack(struct sock *sk)
1649 struct sk_buff* skb;
1651 skb = skb_peek(&sk->sk_write_queue);
1652 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1653 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1656 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1657 if (skb_cloned(skb)) {
1658 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1661 __skb_unlink(skb, &sk->sk_write_queue);
1662 skb_header_release(nskb);
1663 __skb_queue_head(&sk->sk_write_queue, nskb);
1664 sk_stream_free_skb(sk, skb);
1665 sk_charge_skb(sk, nskb);
1669 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1670 TCP_ECN_send_synack(tcp_sk(sk), skb);
1672 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1673 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1677 * Prepare a SYN-ACK.
1679 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1680 struct request_sock *req)
1682 struct inet_request_sock *ireq = inet_rsk(req);
1683 struct tcp_sock *tp = tcp_sk(sk);
1685 int tcp_header_size;
1686 struct sk_buff *skb;
1688 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1692 /* Reserve space for headers. */
1693 skb_reserve(skb, MAX_TCP_HEADER);
1695 skb->dst = dst_clone(dst);
1697 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1698 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1699 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1700 /* SACK_PERM is in the place of NOP NOP of TS */
1701 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1702 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1704 memset(th, 0, sizeof(struct tcphdr));
1707 if (dst->dev->features&NETIF_F_TSO)
1709 TCP_ECN_make_synack(req, th);
1710 th->source = inet_sk(sk)->sport;
1711 th->dest = ireq->rmt_port;
1712 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
1713 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1714 TCP_SKB_CB(skb)->sacked = 0;
1715 skb_shinfo(skb)->tso_segs = 1;
1716 skb_shinfo(skb)->tso_size = 0;
1717 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1718 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
1719 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1721 /* Set this up on the first call only */
1722 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1723 /* tcp_full_space because it is guaranteed to be the first packet */
1724 tcp_select_initial_window(tcp_full_space(sk),
1725 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1730 ireq->rcv_wscale = rcv_wscale;
1733 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1734 th->window = htons(req->rcv_wnd);
1736 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1737 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1738 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
1739 TCP_SKB_CB(skb)->when,
1743 th->doff = (tcp_header_size >> 2);
1744 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1749 * Do all connect socket setups that can be done AF independent.
1751 static inline void tcp_connect_init(struct sock *sk)
1753 struct dst_entry *dst = __sk_dst_get(sk);
1754 struct tcp_sock *tp = tcp_sk(sk);
1757 /* We'll fix this up when we get a response from the other end.
1758 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1760 tp->tcp_header_len = sizeof(struct tcphdr) +
1761 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1763 /* If user gave his TCP_MAXSEG, record it to clamp */
1764 if (tp->rx_opt.user_mss)
1765 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1767 tcp_sync_mss(sk, dst_mtu(dst));
1769 if (!tp->window_clamp)
1770 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1771 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1772 tcp_initialize_rcv_mss(sk);
1774 tcp_select_initial_window(tcp_full_space(sk),
1775 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1778 sysctl_tcp_window_scaling,
1781 tp->rx_opt.rcv_wscale = rcv_wscale;
1782 tp->rcv_ssthresh = tp->rcv_wnd;
1785 sock_reset_flag(sk, SOCK_DONE);
1787 tcp_init_wl(tp, tp->write_seq, 0);
1788 tp->snd_una = tp->write_seq;
1789 tp->snd_sml = tp->write_seq;
1794 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
1795 inet_csk(sk)->icsk_retransmits = 0;
1796 tcp_clear_retrans(tp);
1800 * Build a SYN and send it off.
1802 int tcp_connect(struct sock *sk)
1804 struct tcp_sock *tp = tcp_sk(sk);
1805 struct sk_buff *buff;
1807 tcp_connect_init(sk);
1809 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
1810 if (unlikely(buff == NULL))
1813 /* Reserve space for headers. */
1814 skb_reserve(buff, MAX_TCP_HEADER);
1816 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1817 TCP_ECN_send_syn(sk, tp, buff);
1818 TCP_SKB_CB(buff)->sacked = 0;
1819 skb_shinfo(buff)->tso_segs = 1;
1820 skb_shinfo(buff)->tso_size = 0;
1822 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1823 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1824 tp->snd_nxt = tp->write_seq;
1825 tp->pushed_seq = tp->write_seq;
1828 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1829 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1830 skb_header_release(buff);
1831 __skb_queue_tail(&sk->sk_write_queue, buff);
1832 sk_charge_skb(sk, buff);
1833 tp->packets_out += tcp_skb_pcount(buff);
1834 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1835 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1837 /* Timer for repeating the SYN until an answer. */
1838 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1839 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1843 /* Send out a delayed ack, the caller does the policy checking
1844 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1847 void tcp_send_delayed_ack(struct sock *sk)
1849 struct inet_connection_sock *icsk = inet_csk(sk);
1850 int ato = icsk->icsk_ack.ato;
1851 unsigned long timeout;
1853 if (ato > TCP_DELACK_MIN) {
1854 const struct tcp_sock *tp = tcp_sk(sk);
1857 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
1858 max_ato = TCP_DELACK_MAX;
1860 /* Slow path, intersegment interval is "high". */
1862 /* If some rtt estimate is known, use it to bound delayed ack.
1863 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1867 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1873 ato = min(ato, max_ato);
1876 /* Stay within the limit we were given */
1877 timeout = jiffies + ato;
1879 /* Use new timeout only if there wasn't a older one earlier. */
1880 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
1881 /* If delack timer was blocked or is about to expire,
1884 if (icsk->icsk_ack.blocked ||
1885 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
1890 if (!time_before(timeout, icsk->icsk_ack.timeout))
1891 timeout = icsk->icsk_ack.timeout;
1893 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
1894 icsk->icsk_ack.timeout = timeout;
1895 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
1898 /* This routine sends an ack and also updates the window. */
1899 void tcp_send_ack(struct sock *sk)
1901 /* If we have been reset, we may not send again. */
1902 if (sk->sk_state != TCP_CLOSE) {
1903 struct tcp_sock *tp = tcp_sk(sk);
1904 struct sk_buff *buff;
1906 /* We are not putting this on the write queue, so
1907 * tcp_transmit_skb() will set the ownership to this
1910 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1912 inet_csk_schedule_ack(sk);
1913 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
1914 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
1915 TCP_DELACK_MAX, TCP_RTO_MAX);
1919 /* Reserve space for headers and prepare control bits. */
1920 skb_reserve(buff, MAX_TCP_HEADER);
1922 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1923 TCP_SKB_CB(buff)->sacked = 0;
1924 skb_shinfo(buff)->tso_segs = 1;
1925 skb_shinfo(buff)->tso_size = 0;
1927 /* Send it off, this clears delayed acks for us. */
1928 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1929 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1930 tcp_transmit_skb(sk, buff);
1934 /* This routine sends a packet with an out of date sequence
1935 * number. It assumes the other end will try to ack it.
1937 * Question: what should we make while urgent mode?
1938 * 4.4BSD forces sending single byte of data. We cannot send
1939 * out of window data, because we have SND.NXT==SND.MAX...
1941 * Current solution: to send TWO zero-length segments in urgent mode:
1942 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1943 * out-of-date with SND.UNA-1 to probe window.
1945 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1947 struct tcp_sock *tp = tcp_sk(sk);
1948 struct sk_buff *skb;
1950 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1951 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1955 /* Reserve space for headers and set control bits. */
1956 skb_reserve(skb, MAX_TCP_HEADER);
1958 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1959 TCP_SKB_CB(skb)->sacked = urgent;
1960 skb_shinfo(skb)->tso_segs = 1;
1961 skb_shinfo(skb)->tso_size = 0;
1963 /* Use a previous sequence. This should cause the other
1964 * end to send an ack. Don't queue or clone SKB, just
1967 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1968 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1969 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1970 return tcp_transmit_skb(sk, skb);
1973 int tcp_write_wakeup(struct sock *sk)
1975 if (sk->sk_state != TCP_CLOSE) {
1976 struct tcp_sock *tp = tcp_sk(sk);
1977 struct sk_buff *skb;
1979 if ((skb = sk->sk_send_head) != NULL &&
1980 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1982 unsigned int mss = tcp_current_mss(sk, 0);
1983 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1985 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1986 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1988 /* We are probing the opening of a window
1989 * but the window size is != 0
1990 * must have been a result SWS avoidance ( sender )
1992 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1994 seg_size = min(seg_size, mss);
1995 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1996 if (tcp_fragment(sk, skb, seg_size, mss))
1998 } else if (!tcp_skb_pcount(skb))
1999 tcp_set_skb_tso_segs(sk, skb, mss);
2001 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2002 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2003 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
2005 update_send_head(sk, tp, skb);
2010 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2011 tcp_xmit_probe_skb(sk, TCPCB_URG);
2012 return tcp_xmit_probe_skb(sk, 0);
2018 /* A window probe timeout has occurred. If window is not closed send
2019 * a partial packet else a zero probe.
2021 void tcp_send_probe0(struct sock *sk)
2023 struct inet_connection_sock *icsk = inet_csk(sk);
2024 struct tcp_sock *tp = tcp_sk(sk);
2027 err = tcp_write_wakeup(sk);
2029 if (tp->packets_out || !sk->sk_send_head) {
2030 /* Cancel probe timer, if it is not required. */
2031 icsk->icsk_probes_out = 0;
2032 icsk->icsk_backoff = 0;
2037 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2038 icsk->icsk_backoff++;
2039 icsk->icsk_probes_out++;
2040 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2041 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2044 /* If packet was not sent due to local congestion,
2045 * do not backoff and do not remember icsk_probes_out.
2046 * Let local senders to fight for local resources.
2048 * Use accumulated backoff yet.
2050 if (!icsk->icsk_probes_out)
2051 icsk->icsk_probes_out = 1;
2052 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2053 min(icsk->icsk_rto << icsk->icsk_backoff,
2054 TCP_RESOURCE_PROBE_INTERVAL),
2059 EXPORT_SYMBOL(tcp_connect);
2060 EXPORT_SYMBOL(tcp_make_synack);
2061 EXPORT_SYMBOL(tcp_simple_retransmit);
2062 EXPORT_SYMBOL(tcp_sync_mss);