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 RFC1414. Senders, not following this RFC,
194 * will be satisfied with 2.
196 if (mss > (1<<*rcv_wscale)) {
202 if (*rcv_wnd > init_cwnd*mss)
203 *rcv_wnd = init_cwnd*mss;
206 /* Set the clamp no higher than max representable value */
207 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
210 /* Chose a new window to advertise, update state in tcp_sock for the
211 * socket, and return result with RFC1323 scaling applied. The return
212 * value can be stuffed directly into th->window for an outgoing
215 static __inline__ u16 tcp_select_window(struct sock *sk)
217 struct tcp_sock *tp = tcp_sk(sk);
218 u32 cur_win = tcp_receive_window(tp);
219 u32 new_win = __tcp_select_window(sk);
221 /* Never shrink the offered window */
222 if(new_win < cur_win) {
223 /* Danger Will Robinson!
224 * Don't update rcv_wup/rcv_wnd here or else
225 * we will not be able to advertise a zero
226 * window in time. --DaveM
228 * Relax Will Robinson.
232 tp->rcv_wnd = new_win;
233 tp->rcv_wup = tp->rcv_nxt;
235 /* Make sure we do not exceed the maximum possible
238 if (!tp->rx_opt.rcv_wscale)
239 new_win = min(new_win, MAX_TCP_WINDOW);
241 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
243 /* RFC1323 scaling applied */
244 new_win >>= tp->rx_opt.rcv_wscale;
246 /* If we advertise zero window, disable fast path. */
254 /* This routine actually transmits TCP packets queued in by
255 * tcp_do_sendmsg(). This is used by both the initial
256 * transmission and possible later retransmissions.
257 * All SKB's seen here are completely headerless. It is our
258 * job to build the TCP header, and pass the packet down to
259 * IP so it can do the same plus pass the packet off to the
262 * We are working here with either a clone of the original
263 * SKB, or a fresh unique copy made by the retransmit engine.
265 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
268 const struct inet_connection_sock *icsk = inet_csk(sk);
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
283 /* If congestion control is doing timestamping */
284 if (icsk->icsk_ca_ops->rtt_sample)
285 __net_timestamp(skb);
288 if (tcb->flags & TCPCB_FLAG_SYN) {
289 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
290 if(sysctl_tcp_timestamps) {
291 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
292 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
294 if(sysctl_tcp_window_scaling) {
295 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
296 sysctl_flags |= SYSCTL_FLAG_WSCALE;
298 if(sysctl_tcp_sack) {
299 sysctl_flags |= SYSCTL_FLAG_SACK;
300 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
301 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
303 } else if (tp->rx_opt.eff_sacks) {
304 /* A SACK is 2 pad bytes, a 2 byte header, plus
305 * 2 32-bit sequence numbers for each SACK block.
307 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
308 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
311 if (tcp_packets_in_flight(tp) == 0)
312 tcp_ca_event(sk, CA_EVENT_TX_START);
314 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
316 skb_set_owner_w(skb, sk);
318 /* Build TCP header and checksum it. */
319 th->source = inet->sport;
320 th->dest = inet->dport;
321 th->seq = htonl(tcb->seq);
322 th->ack_seq = htonl(tp->rcv_nxt);
323 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
324 if (tcb->flags & TCPCB_FLAG_SYN) {
325 /* RFC1323: The window in SYN & SYN/ACK segments
328 th->window = htons(tp->rcv_wnd);
330 th->window = htons(tcp_select_window(sk));
336 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
337 th->urg_ptr = htons(tp->snd_up-tcb->seq);
341 if (tcb->flags & TCPCB_FLAG_SYN) {
342 tcp_syn_build_options((__u32 *)(th + 1),
343 tcp_advertise_mss(sk),
344 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
345 (sysctl_flags & SYSCTL_FLAG_SACK),
346 (sysctl_flags & SYSCTL_FLAG_WSCALE),
347 tp->rx_opt.rcv_wscale,
349 tp->rx_opt.ts_recent);
351 tcp_build_and_update_options((__u32 *)(th + 1),
354 TCP_ECN_send(sk, tp, skb, tcp_header_size);
356 tp->af_specific->send_check(sk, th, skb->len, skb);
358 if (tcb->flags & TCPCB_FLAG_ACK)
359 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
361 if (skb->len != tcp_header_size)
362 tcp_event_data_sent(tp, skb, sk);
364 TCP_INC_STATS(TCP_MIB_OUTSEGS);
366 err = tp->af_specific->queue_xmit(skb, 0);
372 /* NET_XMIT_CN is special. It does not guarantee,
373 * that this packet is lost. It tells that device
374 * is about to start to drop packets or already
375 * drops some packets of the same priority and
376 * invokes us to send less aggressively.
378 return err == NET_XMIT_CN ? 0 : err;
381 #undef SYSCTL_FLAG_TSTAMPS
382 #undef SYSCTL_FLAG_WSCALE
383 #undef SYSCTL_FLAG_SACK
387 /* This routine just queue's the buffer
389 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
390 * otherwise socket can stall.
392 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
394 struct tcp_sock *tp = tcp_sk(sk);
396 /* Advance write_seq and place onto the write_queue. */
397 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
398 skb_header_release(skb);
399 __skb_queue_tail(&sk->sk_write_queue, skb);
400 sk_charge_skb(sk, skb);
402 /* Queue it, remembering where we must start sending. */
403 if (sk->sk_send_head == NULL)
404 sk->sk_send_head = skb;
407 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
409 if (skb->len <= mss_now ||
410 !(sk->sk_route_caps & NETIF_F_TSO)) {
411 /* Avoid the costly divide in the normal
414 skb_shinfo(skb)->tso_segs = 1;
415 skb_shinfo(skb)->tso_size = 0;
419 factor = skb->len + (mss_now - 1);
421 skb_shinfo(skb)->tso_segs = factor;
422 skb_shinfo(skb)->tso_size = mss_now;
426 /* Function to create two new TCP segments. Shrinks the given segment
427 * to the specified size and appends a new segment with the rest of the
428 * packet to the list. This won't be called frequently, I hope.
429 * Remember, these are still headerless SKBs at this point.
431 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
433 struct tcp_sock *tp = tcp_sk(sk);
434 struct sk_buff *buff;
435 int nsize, old_factor;
438 nsize = skb_headlen(skb) - len;
442 if (skb_cloned(skb) &&
443 skb_is_nonlinear(skb) &&
444 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
447 /* Get a new skb... force flag on. */
448 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
450 return -ENOMEM; /* We'll just try again later. */
451 sk_charge_skb(sk, buff);
453 /* Correct the sequence numbers. */
454 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
455 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
456 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
458 /* PSH and FIN should only be set in the second packet. */
459 flags = TCP_SKB_CB(skb)->flags;
460 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
461 TCP_SKB_CB(buff)->flags = flags;
462 TCP_SKB_CB(buff)->sacked =
463 (TCP_SKB_CB(skb)->sacked &
464 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
465 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
467 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
468 /* Copy and checksum data tail into the new buffer. */
469 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
474 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
476 skb->ip_summed = CHECKSUM_HW;
477 skb_split(skb, buff, len);
480 buff->ip_summed = skb->ip_summed;
482 /* Looks stupid, but our code really uses when of
483 * skbs, which it never sent before. --ANK
485 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
486 buff->tstamp = skb->tstamp;
488 old_factor = tcp_skb_pcount(skb);
490 /* Fix up tso_factor for both original and new SKB. */
491 tcp_set_skb_tso_segs(sk, skb, mss_now);
492 tcp_set_skb_tso_segs(sk, buff, mss_now);
494 /* If this packet has been sent out already, we must
495 * adjust the various packet counters.
497 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
498 int diff = old_factor - tcp_skb_pcount(skb) -
499 tcp_skb_pcount(buff);
501 tp->packets_out -= diff;
502 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
503 tp->lost_out -= diff;
504 tp->left_out -= diff;
507 tp->fackets_out -= diff;
508 if ((int)tp->fackets_out < 0)
513 /* Link BUFF into the send queue. */
514 skb_header_release(buff);
515 __skb_append(skb, buff, &sk->sk_write_queue);
520 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
521 * eventually). The difference is that pulled data not copied, but
522 * immediately discarded.
524 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
530 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
531 if (skb_shinfo(skb)->frags[i].size <= eat) {
532 put_page(skb_shinfo(skb)->frags[i].page);
533 eat -= skb_shinfo(skb)->frags[i].size;
535 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
537 skb_shinfo(skb)->frags[k].page_offset += eat;
538 skb_shinfo(skb)->frags[k].size -= eat;
544 skb_shinfo(skb)->nr_frags = k;
546 skb->tail = skb->data;
547 skb->data_len -= len;
548 skb->len = skb->data_len;
552 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
554 if (skb_cloned(skb) &&
555 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
558 if (len <= skb_headlen(skb)) {
559 __skb_pull(skb, len);
561 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
565 TCP_SKB_CB(skb)->seq += len;
566 skb->ip_summed = CHECKSUM_HW;
568 skb->truesize -= len;
569 sk->sk_wmem_queued -= len;
570 sk->sk_forward_alloc += len;
571 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
573 /* Any change of skb->len requires recalculation of tso
576 if (tcp_skb_pcount(skb) > 1)
577 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
582 /* This function synchronize snd mss to current pmtu/exthdr set.
584 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
585 for TCP options, but includes only bare TCP header.
587 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
588 It is minumum of user_mss and mss received with SYN.
589 It also does not include TCP options.
591 tp->pmtu_cookie is last pmtu, seen by this function.
593 tp->mss_cache is current effective sending mss, including
594 all tcp options except for SACKs. It is evaluated,
595 taking into account current pmtu, but never exceeds
596 tp->rx_opt.mss_clamp.
598 NOTE1. rfc1122 clearly states that advertised MSS
599 DOES NOT include either tcp or ip options.
601 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
602 this function. --ANK (980731)
605 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
607 struct tcp_sock *tp = tcp_sk(sk);
610 /* Calculate base mss without TCP options:
611 It is MMS_S - sizeof(tcphdr) of rfc1122
613 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
615 /* Clamp it (mss_clamp does not include tcp options) */
616 if (mss_now > tp->rx_opt.mss_clamp)
617 mss_now = tp->rx_opt.mss_clamp;
619 /* Now subtract optional transport overhead */
620 mss_now -= tp->ext_header_len;
622 /* Then reserve room for full set of TCP options and 8 bytes of data */
626 /* Now subtract TCP options size, not including SACKs */
627 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
629 /* Bound mss with half of window */
630 if (tp->max_window && mss_now > (tp->max_window>>1))
631 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
633 /* And store cached results */
634 tp->pmtu_cookie = pmtu;
635 tp->mss_cache = mss_now;
640 /* Compute the current effective MSS, taking SACKs and IP options,
641 * and even PMTU discovery events into account.
643 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
644 * cannot be large. However, taking into account rare use of URG, this
647 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
649 struct tcp_sock *tp = tcp_sk(sk);
650 struct dst_entry *dst = __sk_dst_get(sk);
655 mss_now = tp->mss_cache;
658 (sk->sk_route_caps & NETIF_F_TSO) &&
663 u32 mtu = dst_mtu(dst);
664 if (mtu != tp->pmtu_cookie)
665 mss_now = tcp_sync_mss(sk, mtu);
668 if (tp->rx_opt.eff_sacks)
669 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
670 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
672 xmit_size_goal = mss_now;
675 xmit_size_goal = 65535 -
676 tp->af_specific->net_header_len -
677 tp->ext_header_len - tp->tcp_header_len;
679 if (tp->max_window &&
680 (xmit_size_goal > (tp->max_window >> 1)))
681 xmit_size_goal = max((tp->max_window >> 1),
682 68U - tp->tcp_header_len);
684 xmit_size_goal -= (xmit_size_goal % mss_now);
686 tp->xmit_size_goal = xmit_size_goal;
691 /* Congestion window validation. (RFC2861) */
693 static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
695 __u32 packets_out = tp->packets_out;
697 if (packets_out >= tp->snd_cwnd) {
698 /* Network is feed fully. */
699 tp->snd_cwnd_used = 0;
700 tp->snd_cwnd_stamp = tcp_time_stamp;
702 /* Network starves. */
703 if (tp->packets_out > tp->snd_cwnd_used)
704 tp->snd_cwnd_used = tp->packets_out;
706 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
707 tcp_cwnd_application_limited(sk);
711 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
713 u32 window, cwnd_len;
715 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
716 cwnd_len = mss_now * cwnd;
717 return min(window, cwnd_len);
720 /* Can at least one segment of SKB be sent right now, according to the
721 * congestion window rules? If so, return how many segments are allowed.
723 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
727 /* Don't be strict about the congestion window for the final FIN. */
728 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
731 in_flight = tcp_packets_in_flight(tp);
733 if (in_flight < cwnd)
734 return (cwnd - in_flight);
739 /* This must be invoked the first time we consider transmitting
742 static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
744 int tso_segs = tcp_skb_pcount(skb);
748 skb_shinfo(skb)->tso_size != mss_now)) {
749 tcp_set_skb_tso_segs(sk, skb, mss_now);
750 tso_segs = tcp_skb_pcount(skb);
755 static inline int tcp_minshall_check(const struct tcp_sock *tp)
757 return after(tp->snd_sml,tp->snd_una) &&
758 !after(tp->snd_sml, tp->snd_nxt);
761 /* Return 0, if packet can be sent now without violation Nagle's rules:
762 * 1. It is full sized.
763 * 2. Or it contains FIN. (already checked by caller)
764 * 3. Or TCP_NODELAY was set.
765 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
766 * With Minshall's modification: all sent small packets are ACKed.
769 static inline int tcp_nagle_check(const struct tcp_sock *tp,
770 const struct sk_buff *skb,
771 unsigned mss_now, int nonagle)
773 return (skb->len < mss_now &&
774 ((nonagle&TCP_NAGLE_CORK) ||
777 tcp_minshall_check(tp))));
780 /* Return non-zero if the Nagle test allows this packet to be
783 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
784 unsigned int cur_mss, int nonagle)
786 /* Nagle rule does not apply to frames, which sit in the middle of the
787 * write_queue (they have no chances to get new data).
789 * This is implemented in the callers, where they modify the 'nonagle'
790 * argument based upon the location of SKB in the send queue.
792 if (nonagle & TCP_NAGLE_PUSH)
795 /* Don't use the nagle rule for urgent data (or for the final FIN). */
797 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
800 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
806 /* Does at least the first segment of SKB fit into the send window? */
807 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
809 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
811 if (skb->len > cur_mss)
812 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
814 return !after(end_seq, tp->snd_una + tp->snd_wnd);
817 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
818 * should be put on the wire right now. If so, it returns the number of
819 * packets allowed by the congestion window.
821 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
822 unsigned int cur_mss, int nonagle)
824 struct tcp_sock *tp = tcp_sk(sk);
825 unsigned int cwnd_quota;
827 tcp_init_tso_segs(sk, skb, cur_mss);
829 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
832 cwnd_quota = tcp_cwnd_test(tp, skb);
834 !tcp_snd_wnd_test(tp, skb, cur_mss))
840 static inline int tcp_skb_is_last(const struct sock *sk,
841 const struct sk_buff *skb)
843 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
846 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
848 struct sk_buff *skb = sk->sk_send_head;
851 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
852 (tcp_skb_is_last(sk, skb) ?
857 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
858 * which is put after SKB on the list. It is very much like
859 * tcp_fragment() except that it may make several kinds of assumptions
860 * in order to speed up the splitting operation. In particular, we
861 * know that all the data is in scatter-gather pages, and that the
862 * packet has never been sent out before (and thus is not cloned).
864 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
866 struct sk_buff *buff;
867 int nlen = skb->len - len;
870 /* All of a TSO frame must be composed of paged data. */
871 if (skb->len != skb->data_len)
872 return tcp_fragment(sk, skb, len, mss_now);
874 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
875 if (unlikely(buff == NULL))
878 buff->truesize = nlen;
879 skb->truesize -= nlen;
881 /* Correct the sequence numbers. */
882 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
883 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
884 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
886 /* PSH and FIN should only be set in the second packet. */
887 flags = TCP_SKB_CB(skb)->flags;
888 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
889 TCP_SKB_CB(buff)->flags = flags;
891 /* This packet was never sent out yet, so no SACK bits. */
892 TCP_SKB_CB(buff)->sacked = 0;
894 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
895 skb_split(skb, buff, len);
897 /* Fix up tso_factor for both original and new SKB. */
898 tcp_set_skb_tso_segs(sk, skb, mss_now);
899 tcp_set_skb_tso_segs(sk, buff, mss_now);
901 /* Link BUFF into the send queue. */
902 skb_header_release(buff);
903 __skb_append(skb, buff, &sk->sk_write_queue);
908 /* Try to defer sending, if possible, in order to minimize the amount
909 * of TSO splitting we do. View it as a kind of TSO Nagle test.
911 * This algorithm is from John Heffner.
913 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
915 const struct inet_connection_sock *icsk = inet_csk(sk);
916 u32 send_win, cong_win, limit, in_flight;
918 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
921 if (icsk->icsk_ca_state != TCP_CA_Open)
924 in_flight = tcp_packets_in_flight(tp);
926 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
927 (tp->snd_cwnd <= in_flight));
929 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
931 /* From in_flight test above, we know that cwnd > in_flight. */
932 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
934 limit = min(send_win, cong_win);
936 if (sysctl_tcp_tso_win_divisor) {
937 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
939 /* If at least some fraction of a window is available,
942 chunk /= sysctl_tcp_tso_win_divisor;
946 /* Different approach, try not to defer past a single
947 * ACK. Receiver should ACK every other full sized
948 * frame, so if we have space for more than 3 frames
951 if (limit > tcp_max_burst(tp) * tp->mss_cache)
955 /* Ok, it looks like it is advisable to defer. */
959 /* This routine writes packets to the network. It advances the
960 * send_head. This happens as incoming acks open up the remote
963 * Returns 1, if no segments are in flight and we have queued segments, but
964 * cannot send anything now because of SWS or another problem.
966 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
968 struct tcp_sock *tp = tcp_sk(sk);
970 unsigned int tso_segs, sent_pkts;
973 /* If we are closed, the bytes will have to remain here.
974 * In time closedown will finish, we empty the write queue and all
977 if (unlikely(sk->sk_state == TCP_CLOSE))
981 while ((skb = sk->sk_send_head)) {
984 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
987 cwnd_quota = tcp_cwnd_test(tp, skb);
991 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
995 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
996 (tcp_skb_is_last(sk, skb) ?
997 nonagle : TCP_NAGLE_PUSH))))
1000 if (tcp_tso_should_defer(sk, tp, skb))
1006 limit = tcp_window_allows(tp, skb,
1007 mss_now, cwnd_quota);
1009 if (skb->len < limit) {
1010 unsigned int trim = skb->len % mss_now;
1013 limit = skb->len - trim;
1017 if (skb->len > limit &&
1018 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1021 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1023 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
1026 /* Advance the send_head. This one is sent out.
1027 * This call will increment packets_out.
1029 update_send_head(sk, tp, skb);
1031 tcp_minshall_update(tp, mss_now, skb);
1035 if (likely(sent_pkts)) {
1036 tcp_cwnd_validate(sk, tp);
1039 return !tp->packets_out && sk->sk_send_head;
1042 /* Push out any pending frames which were held back due to
1043 * TCP_CORK or attempt at coalescing tiny packets.
1044 * The socket must be locked by the caller.
1046 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1047 unsigned int cur_mss, int nonagle)
1049 struct sk_buff *skb = sk->sk_send_head;
1052 if (tcp_write_xmit(sk, cur_mss, nonagle))
1053 tcp_check_probe_timer(sk, tp);
1057 /* Send _single_ skb sitting at the send head. This function requires
1058 * true push pending frames to setup probe timer etc.
1060 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1062 struct tcp_sock *tp = tcp_sk(sk);
1063 struct sk_buff *skb = sk->sk_send_head;
1064 unsigned int tso_segs, cwnd_quota;
1066 BUG_ON(!skb || skb->len < mss_now);
1068 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1069 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1071 if (likely(cwnd_quota)) {
1078 limit = tcp_window_allows(tp, skb,
1079 mss_now, cwnd_quota);
1081 if (skb->len < limit) {
1082 unsigned int trim = skb->len % mss_now;
1085 limit = skb->len - trim;
1089 if (skb->len > limit &&
1090 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1093 /* Send it out now. */
1094 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1096 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1097 update_send_head(sk, tp, skb);
1098 tcp_cwnd_validate(sk, tp);
1104 /* This function returns the amount that we can raise the
1105 * usable window based on the following constraints
1107 * 1. The window can never be shrunk once it is offered (RFC 793)
1108 * 2. We limit memory per socket
1111 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1112 * RECV.NEXT + RCV.WIN fixed until:
1113 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1115 * i.e. don't raise the right edge of the window until you can raise
1116 * it at least MSS bytes.
1118 * Unfortunately, the recommended algorithm breaks header prediction,
1119 * since header prediction assumes th->window stays fixed.
1121 * Strictly speaking, keeping th->window fixed violates the receiver
1122 * side SWS prevention criteria. The problem is that under this rule
1123 * a stream of single byte packets will cause the right side of the
1124 * window to always advance by a single byte.
1126 * Of course, if the sender implements sender side SWS prevention
1127 * then this will not be a problem.
1129 * BSD seems to make the following compromise:
1131 * If the free space is less than the 1/4 of the maximum
1132 * space available and the free space is less than 1/2 mss,
1133 * then set the window to 0.
1134 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1135 * Otherwise, just prevent the window from shrinking
1136 * and from being larger than the largest representable value.
1138 * This prevents incremental opening of the window in the regime
1139 * where TCP is limited by the speed of the reader side taking
1140 * data out of the TCP receive queue. It does nothing about
1141 * those cases where the window is constrained on the sender side
1142 * because the pipeline is full.
1144 * BSD also seems to "accidentally" limit itself to windows that are a
1145 * multiple of MSS, at least until the free space gets quite small.
1146 * This would appear to be a side effect of the mbuf implementation.
1147 * Combining these two algorithms results in the observed behavior
1148 * of having a fixed window size at almost all times.
1150 * Below we obtain similar behavior by forcing the offered window to
1151 * a multiple of the mss when it is feasible to do so.
1153 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1154 * Regular options like TIMESTAMP are taken into account.
1156 u32 __tcp_select_window(struct sock *sk)
1158 struct inet_connection_sock *icsk = inet_csk(sk);
1159 struct tcp_sock *tp = tcp_sk(sk);
1160 /* MSS for the peer's data. Previous verions used mss_clamp
1161 * here. I don't know if the value based on our guesses
1162 * of peer's MSS is better for the performance. It's more correct
1163 * but may be worse for the performance because of rcv_mss
1164 * fluctuations. --SAW 1998/11/1
1166 int mss = icsk->icsk_ack.rcv_mss;
1167 int free_space = tcp_space(sk);
1168 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1171 if (mss > full_space)
1174 if (free_space < full_space/2) {
1175 icsk->icsk_ack.quick = 0;
1177 if (tcp_memory_pressure)
1178 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1180 if (free_space < mss)
1184 if (free_space > tp->rcv_ssthresh)
1185 free_space = tp->rcv_ssthresh;
1187 /* Don't do rounding if we are using window scaling, since the
1188 * scaled window will not line up with the MSS boundary anyway.
1190 window = tp->rcv_wnd;
1191 if (tp->rx_opt.rcv_wscale) {
1192 window = free_space;
1194 /* Advertise enough space so that it won't get scaled away.
1195 * Import case: prevent zero window announcement if
1196 * 1<<rcv_wscale > mss.
1198 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1199 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1200 << tp->rx_opt.rcv_wscale);
1202 /* Get the largest window that is a nice multiple of mss.
1203 * Window clamp already applied above.
1204 * If our current window offering is within 1 mss of the
1205 * free space we just keep it. This prevents the divide
1206 * and multiply from happening most of the time.
1207 * We also don't do any window rounding when the free space
1210 if (window <= free_space - mss || window > free_space)
1211 window = (free_space/mss)*mss;
1217 /* Attempt to collapse two adjacent SKB's during retransmission. */
1218 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1220 struct tcp_sock *tp = tcp_sk(sk);
1221 struct sk_buff *next_skb = skb->next;
1223 /* The first test we must make is that neither of these two
1224 * SKB's are still referenced by someone else.
1226 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1227 int skb_size = skb->len, next_skb_size = next_skb->len;
1228 u16 flags = TCP_SKB_CB(skb)->flags;
1230 /* Also punt if next skb has been SACK'd. */
1231 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1234 /* Next skb is out of window. */
1235 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1238 /* Punt if not enough space exists in the first SKB for
1239 * the data in the second, or the total combined payload
1240 * would exceed the MSS.
1242 if ((next_skb_size > skb_tailroom(skb)) ||
1243 ((skb_size + next_skb_size) > mss_now))
1246 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1247 tcp_skb_pcount(next_skb) != 1);
1249 /* Ok. We will be able to collapse the packet. */
1250 __skb_unlink(next_skb, &sk->sk_write_queue);
1252 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1254 if (next_skb->ip_summed == CHECKSUM_HW)
1255 skb->ip_summed = CHECKSUM_HW;
1257 if (skb->ip_summed != CHECKSUM_HW)
1258 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1260 /* Update sequence range on original skb. */
1261 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1263 /* Merge over control information. */
1264 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1265 TCP_SKB_CB(skb)->flags = flags;
1267 /* All done, get rid of second SKB and account for it so
1268 * packet counting does not break.
1270 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1271 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1272 tp->retrans_out -= tcp_skb_pcount(next_skb);
1273 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1274 tp->lost_out -= tcp_skb_pcount(next_skb);
1275 tp->left_out -= tcp_skb_pcount(next_skb);
1277 /* Reno case is special. Sigh... */
1278 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1279 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1280 tp->left_out -= tcp_skb_pcount(next_skb);
1283 /* Not quite right: it can be > snd.fack, but
1284 * it is better to underestimate fackets.
1286 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1287 tcp_packets_out_dec(tp, next_skb);
1288 sk_stream_free_skb(sk, next_skb);
1292 /* Do a simple retransmit without using the backoff mechanisms in
1293 * tcp_timer. This is used for path mtu discovery.
1294 * The socket is already locked here.
1296 void tcp_simple_retransmit(struct sock *sk)
1298 const struct inet_connection_sock *icsk = inet_csk(sk);
1299 struct tcp_sock *tp = tcp_sk(sk);
1300 struct sk_buff *skb;
1301 unsigned int mss = tcp_current_mss(sk, 0);
1304 sk_stream_for_retrans_queue(skb, sk) {
1305 if (skb->len > mss &&
1306 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1307 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1308 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1309 tp->retrans_out -= tcp_skb_pcount(skb);
1311 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1312 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1313 tp->lost_out += tcp_skb_pcount(skb);
1322 tcp_sync_left_out(tp);
1324 /* Don't muck with the congestion window here.
1325 * Reason is that we do not increase amount of _data_
1326 * in network, but units changed and effective
1327 * cwnd/ssthresh really reduced now.
1329 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1330 tp->high_seq = tp->snd_nxt;
1331 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1332 tp->prior_ssthresh = 0;
1333 tp->undo_marker = 0;
1334 tcp_set_ca_state(sk, TCP_CA_Loss);
1336 tcp_xmit_retransmit_queue(sk);
1339 /* This retransmits one SKB. Policy decisions and retransmit queue
1340 * state updates are done by the caller. Returns non-zero if an
1341 * error occurred which prevented the send.
1343 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1345 struct tcp_sock *tp = tcp_sk(sk);
1346 unsigned int cur_mss = tcp_current_mss(sk, 0);
1349 /* Do not sent more than we queued. 1/4 is reserved for possible
1350 * copying overhead: frgagmentation, tunneling, mangling etc.
1352 if (atomic_read(&sk->sk_wmem_alloc) >
1353 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1356 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1357 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1359 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1363 /* If receiver has shrunk his window, and skb is out of
1364 * new window, do not retransmit it. The exception is the
1365 * case, when window is shrunk to zero. In this case
1366 * our retransmit serves as a zero window probe.
1368 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1369 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1372 if (skb->len > cur_mss) {
1373 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1374 return -ENOMEM; /* We'll try again later. */
1377 /* Collapse two adjacent packets if worthwhile and we can. */
1378 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1379 (skb->len < (cur_mss >> 1)) &&
1380 (skb->next != sk->sk_send_head) &&
1381 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1382 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1383 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1384 (sysctl_tcp_retrans_collapse != 0))
1385 tcp_retrans_try_collapse(sk, skb, cur_mss);
1387 if(tp->af_specific->rebuild_header(sk))
1388 return -EHOSTUNREACH; /* Routing failure or similar. */
1390 /* Some Solaris stacks overoptimize and ignore the FIN on a
1391 * retransmit when old data is attached. So strip it off
1392 * since it is cheap to do so and saves bytes on the network.
1395 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1396 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1397 if (!pskb_trim(skb, 0)) {
1398 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1399 skb_shinfo(skb)->tso_segs = 1;
1400 skb_shinfo(skb)->tso_size = 0;
1401 skb->ip_summed = CHECKSUM_NONE;
1406 /* Make a copy, if the first transmission SKB clone we made
1407 * is still in somebody's hands, else make a clone.
1409 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1411 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1412 pskb_copy(skb, GFP_ATOMIC):
1413 skb_clone(skb, GFP_ATOMIC)));
1416 /* Update global TCP statistics. */
1417 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1419 tp->total_retrans++;
1421 #if FASTRETRANS_DEBUG > 0
1422 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1423 if (net_ratelimit())
1424 printk(KERN_DEBUG "retrans_out leaked.\n");
1427 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1428 tp->retrans_out += tcp_skb_pcount(skb);
1430 /* Save stamp of the first retransmit. */
1431 if (!tp->retrans_stamp)
1432 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1436 /* snd_nxt is stored to detect loss of retransmitted segment,
1437 * see tcp_input.c tcp_sacktag_write_queue().
1439 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1444 /* This gets called after a retransmit timeout, and the initially
1445 * retransmitted data is acknowledged. It tries to continue
1446 * resending the rest of the retransmit queue, until either
1447 * we've sent it all or the congestion window limit is reached.
1448 * If doing SACK, the first ACK which comes back for a timeout
1449 * based retransmit packet might feed us FACK information again.
1450 * If so, we use it to avoid unnecessarily retransmissions.
1452 void tcp_xmit_retransmit_queue(struct sock *sk)
1454 const struct inet_connection_sock *icsk = inet_csk(sk);
1455 struct tcp_sock *tp = tcp_sk(sk);
1456 struct sk_buff *skb;
1457 int packet_cnt = tp->lost_out;
1459 /* First pass: retransmit lost packets. */
1461 sk_stream_for_retrans_queue(skb, sk) {
1462 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1464 /* Assume this retransmit will generate
1465 * only one packet for congestion window
1466 * calculation purposes. This works because
1467 * tcp_retransmit_skb() will chop up the
1468 * packet to be MSS sized and all the
1469 * packet counting works out.
1471 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1474 if (sacked&TCPCB_LOST) {
1475 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1476 if (tcp_retransmit_skb(sk, skb))
1478 if (icsk->icsk_ca_state != TCP_CA_Loss)
1479 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1481 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1484 skb_peek(&sk->sk_write_queue))
1485 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1486 inet_csk(sk)->icsk_rto,
1490 packet_cnt -= tcp_skb_pcount(skb);
1491 if (packet_cnt <= 0)
1497 /* OK, demanded retransmission is finished. */
1499 /* Forward retransmissions are possible only during Recovery. */
1500 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1503 /* No forward retransmissions in Reno are possible. */
1504 if (!tp->rx_opt.sack_ok)
1507 /* Yeah, we have to make difficult choice between forward transmission
1508 * and retransmission... Both ways have their merits...
1510 * For now we do not retransmit anything, while we have some new
1514 if (tcp_may_send_now(sk, tp))
1519 sk_stream_for_retrans_queue(skb, sk) {
1520 /* Similar to the retransmit loop above we
1521 * can pretend that the retransmitted SKB
1522 * we send out here will be composed of one
1523 * real MSS sized packet because tcp_retransmit_skb()
1524 * will fragment it if necessary.
1526 if (++packet_cnt > tp->fackets_out)
1529 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1532 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1535 /* Ok, retransmit it. */
1536 if (tcp_retransmit_skb(sk, skb))
1539 if (skb == skb_peek(&sk->sk_write_queue))
1540 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1541 inet_csk(sk)->icsk_rto,
1544 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1549 /* Send a fin. The caller locks the socket for us. This cannot be
1550 * allowed to fail queueing a FIN frame under any circumstances.
1552 void tcp_send_fin(struct sock *sk)
1554 struct tcp_sock *tp = tcp_sk(sk);
1555 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1558 /* Optimization, tack on the FIN if we have a queue of
1559 * unsent frames. But be careful about outgoing SACKS
1562 mss_now = tcp_current_mss(sk, 1);
1564 if (sk->sk_send_head != NULL) {
1565 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1566 TCP_SKB_CB(skb)->end_seq++;
1569 /* Socket is locked, keep trying until memory is available. */
1571 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1577 /* Reserve space for headers and prepare control bits. */
1578 skb_reserve(skb, MAX_TCP_HEADER);
1580 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1581 TCP_SKB_CB(skb)->sacked = 0;
1582 skb_shinfo(skb)->tso_segs = 1;
1583 skb_shinfo(skb)->tso_size = 0;
1585 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1586 TCP_SKB_CB(skb)->seq = tp->write_seq;
1587 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1588 tcp_queue_skb(sk, skb);
1590 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1593 /* We get here when a process closes a file descriptor (either due to
1594 * an explicit close() or as a byproduct of exit()'ing) and there
1595 * was unread data in the receive queue. This behavior is recommended
1596 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1598 void tcp_send_active_reset(struct sock *sk, unsigned int __nocast priority)
1600 struct tcp_sock *tp = tcp_sk(sk);
1601 struct sk_buff *skb;
1603 /* NOTE: No TCP options attached and we never retransmit this. */
1604 skb = alloc_skb(MAX_TCP_HEADER, priority);
1606 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1610 /* Reserve space for headers and prepare control bits. */
1611 skb_reserve(skb, MAX_TCP_HEADER);
1613 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1614 TCP_SKB_CB(skb)->sacked = 0;
1615 skb_shinfo(skb)->tso_segs = 1;
1616 skb_shinfo(skb)->tso_size = 0;
1619 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1620 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1621 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1622 if (tcp_transmit_skb(sk, skb))
1623 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1626 /* WARNING: This routine must only be called when we have already sent
1627 * a SYN packet that crossed the incoming SYN that caused this routine
1628 * to get called. If this assumption fails then the initial rcv_wnd
1629 * and rcv_wscale values will not be correct.
1631 int tcp_send_synack(struct sock *sk)
1633 struct sk_buff* skb;
1635 skb = skb_peek(&sk->sk_write_queue);
1636 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1637 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1640 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1641 if (skb_cloned(skb)) {
1642 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1645 __skb_unlink(skb, &sk->sk_write_queue);
1646 skb_header_release(nskb);
1647 __skb_queue_head(&sk->sk_write_queue, nskb);
1648 sk_stream_free_skb(sk, skb);
1649 sk_charge_skb(sk, nskb);
1653 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1654 TCP_ECN_send_synack(tcp_sk(sk), skb);
1656 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1657 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1661 * Prepare a SYN-ACK.
1663 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1664 struct request_sock *req)
1666 struct inet_request_sock *ireq = inet_rsk(req);
1667 struct tcp_sock *tp = tcp_sk(sk);
1669 int tcp_header_size;
1670 struct sk_buff *skb;
1672 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1676 /* Reserve space for headers. */
1677 skb_reserve(skb, MAX_TCP_HEADER);
1679 skb->dst = dst_clone(dst);
1681 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1682 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1683 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1684 /* SACK_PERM is in the place of NOP NOP of TS */
1685 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1686 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1688 memset(th, 0, sizeof(struct tcphdr));
1691 if (dst->dev->features&NETIF_F_TSO)
1693 TCP_ECN_make_synack(req, th);
1694 th->source = inet_sk(sk)->sport;
1695 th->dest = ireq->rmt_port;
1696 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
1697 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1698 TCP_SKB_CB(skb)->sacked = 0;
1699 skb_shinfo(skb)->tso_segs = 1;
1700 skb_shinfo(skb)->tso_size = 0;
1701 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1702 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
1703 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1705 /* Set this up on the first call only */
1706 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1707 /* tcp_full_space because it is guaranteed to be the first packet */
1708 tcp_select_initial_window(tcp_full_space(sk),
1709 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1714 ireq->rcv_wscale = rcv_wscale;
1717 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1718 th->window = htons(req->rcv_wnd);
1720 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1721 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1722 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
1723 TCP_SKB_CB(skb)->when,
1727 th->doff = (tcp_header_size >> 2);
1728 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1733 * Do all connect socket setups that can be done AF independent.
1735 static inline void tcp_connect_init(struct sock *sk)
1737 struct dst_entry *dst = __sk_dst_get(sk);
1738 struct tcp_sock *tp = tcp_sk(sk);
1741 /* We'll fix this up when we get a response from the other end.
1742 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1744 tp->tcp_header_len = sizeof(struct tcphdr) +
1745 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1747 /* If user gave his TCP_MAXSEG, record it to clamp */
1748 if (tp->rx_opt.user_mss)
1749 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1751 tcp_sync_mss(sk, dst_mtu(dst));
1753 if (!tp->window_clamp)
1754 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1755 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1756 tcp_initialize_rcv_mss(sk);
1758 tcp_select_initial_window(tcp_full_space(sk),
1759 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1762 sysctl_tcp_window_scaling,
1765 tp->rx_opt.rcv_wscale = rcv_wscale;
1766 tp->rcv_ssthresh = tp->rcv_wnd;
1769 sock_reset_flag(sk, SOCK_DONE);
1771 tcp_init_wl(tp, tp->write_seq, 0);
1772 tp->snd_una = tp->write_seq;
1773 tp->snd_sml = tp->write_seq;
1778 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
1779 inet_csk(sk)->icsk_retransmits = 0;
1780 tcp_clear_retrans(tp);
1784 * Build a SYN and send it off.
1786 int tcp_connect(struct sock *sk)
1788 struct tcp_sock *tp = tcp_sk(sk);
1789 struct sk_buff *buff;
1791 tcp_connect_init(sk);
1793 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
1794 if (unlikely(buff == NULL))
1797 /* Reserve space for headers. */
1798 skb_reserve(buff, MAX_TCP_HEADER);
1800 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1801 TCP_ECN_send_syn(sk, tp, buff);
1802 TCP_SKB_CB(buff)->sacked = 0;
1803 skb_shinfo(buff)->tso_segs = 1;
1804 skb_shinfo(buff)->tso_size = 0;
1806 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1807 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1808 tp->snd_nxt = tp->write_seq;
1809 tp->pushed_seq = tp->write_seq;
1812 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1813 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1814 skb_header_release(buff);
1815 __skb_queue_tail(&sk->sk_write_queue, buff);
1816 sk_charge_skb(sk, buff);
1817 tp->packets_out += tcp_skb_pcount(buff);
1818 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1819 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1821 /* Timer for repeating the SYN until an answer. */
1822 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1823 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1827 /* Send out a delayed ack, the caller does the policy checking
1828 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1831 void tcp_send_delayed_ack(struct sock *sk)
1833 struct inet_connection_sock *icsk = inet_csk(sk);
1834 int ato = icsk->icsk_ack.ato;
1835 unsigned long timeout;
1837 if (ato > TCP_DELACK_MIN) {
1838 const struct tcp_sock *tp = tcp_sk(sk);
1841 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
1842 max_ato = TCP_DELACK_MAX;
1844 /* Slow path, intersegment interval is "high". */
1846 /* If some rtt estimate is known, use it to bound delayed ack.
1847 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1851 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1857 ato = min(ato, max_ato);
1860 /* Stay within the limit we were given */
1861 timeout = jiffies + ato;
1863 /* Use new timeout only if there wasn't a older one earlier. */
1864 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
1865 /* If delack timer was blocked or is about to expire,
1868 if (icsk->icsk_ack.blocked ||
1869 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
1874 if (!time_before(timeout, icsk->icsk_ack.timeout))
1875 timeout = icsk->icsk_ack.timeout;
1877 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
1878 icsk->icsk_ack.timeout = timeout;
1879 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
1882 /* This routine sends an ack and also updates the window. */
1883 void tcp_send_ack(struct sock *sk)
1885 /* If we have been reset, we may not send again. */
1886 if (sk->sk_state != TCP_CLOSE) {
1887 struct tcp_sock *tp = tcp_sk(sk);
1888 struct sk_buff *buff;
1890 /* We are not putting this on the write queue, so
1891 * tcp_transmit_skb() will set the ownership to this
1894 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1896 inet_csk_schedule_ack(sk);
1897 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
1898 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
1899 TCP_DELACK_MAX, TCP_RTO_MAX);
1903 /* Reserve space for headers and prepare control bits. */
1904 skb_reserve(buff, MAX_TCP_HEADER);
1906 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1907 TCP_SKB_CB(buff)->sacked = 0;
1908 skb_shinfo(buff)->tso_segs = 1;
1909 skb_shinfo(buff)->tso_size = 0;
1911 /* Send it off, this clears delayed acks for us. */
1912 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1913 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1914 tcp_transmit_skb(sk, buff);
1918 /* This routine sends a packet with an out of date sequence
1919 * number. It assumes the other end will try to ack it.
1921 * Question: what should we make while urgent mode?
1922 * 4.4BSD forces sending single byte of data. We cannot send
1923 * out of window data, because we have SND.NXT==SND.MAX...
1925 * Current solution: to send TWO zero-length segments in urgent mode:
1926 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1927 * out-of-date with SND.UNA-1 to probe window.
1929 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1931 struct tcp_sock *tp = tcp_sk(sk);
1932 struct sk_buff *skb;
1934 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1935 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1939 /* Reserve space for headers and set control bits. */
1940 skb_reserve(skb, MAX_TCP_HEADER);
1942 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1943 TCP_SKB_CB(skb)->sacked = urgent;
1944 skb_shinfo(skb)->tso_segs = 1;
1945 skb_shinfo(skb)->tso_size = 0;
1947 /* Use a previous sequence. This should cause the other
1948 * end to send an ack. Don't queue or clone SKB, just
1951 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1952 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1953 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1954 return tcp_transmit_skb(sk, skb);
1957 int tcp_write_wakeup(struct sock *sk)
1959 if (sk->sk_state != TCP_CLOSE) {
1960 struct tcp_sock *tp = tcp_sk(sk);
1961 struct sk_buff *skb;
1963 if ((skb = sk->sk_send_head) != NULL &&
1964 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1966 unsigned int mss = tcp_current_mss(sk, 0);
1967 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1969 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1970 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1972 /* We are probing the opening of a window
1973 * but the window size is != 0
1974 * must have been a result SWS avoidance ( sender )
1976 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1978 seg_size = min(seg_size, mss);
1979 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1980 if (tcp_fragment(sk, skb, seg_size, mss))
1982 } else if (!tcp_skb_pcount(skb))
1983 tcp_set_skb_tso_segs(sk, skb, mss);
1985 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1986 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1987 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1989 update_send_head(sk, tp, skb);
1994 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
1995 tcp_xmit_probe_skb(sk, TCPCB_URG);
1996 return tcp_xmit_probe_skb(sk, 0);
2002 /* A window probe timeout has occurred. If window is not closed send
2003 * a partial packet else a zero probe.
2005 void tcp_send_probe0(struct sock *sk)
2007 struct inet_connection_sock *icsk = inet_csk(sk);
2008 struct tcp_sock *tp = tcp_sk(sk);
2011 err = tcp_write_wakeup(sk);
2013 if (tp->packets_out || !sk->sk_send_head) {
2014 /* Cancel probe timer, if it is not required. */
2015 icsk->icsk_probes_out = 0;
2016 icsk->icsk_backoff = 0;
2021 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2022 icsk->icsk_backoff++;
2023 icsk->icsk_probes_out++;
2024 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2025 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2028 /* If packet was not sent due to local congestion,
2029 * do not backoff and do not remember icsk_probes_out.
2030 * Let local senders to fight for local resources.
2032 * Use accumulated backoff yet.
2034 if (!icsk->icsk_probes_out)
2035 icsk->icsk_probes_out = 1;
2036 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2037 min(icsk->icsk_rto << icsk->icsk_backoff,
2038 TCP_RESOURCE_PROBE_INTERVAL),
2043 EXPORT_SYMBOL(tcp_connect);
2044 EXPORT_SYMBOL(tcp_make_synack);
2045 EXPORT_SYMBOL(tcp_simple_retransmit);
2046 EXPORT_SYMBOL(tcp_sync_mss);