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 tcp_sock *tp, struct dst_entry *dst)
110 s32 delta = tcp_time_stamp - tp->lsndtime;
111 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
112 u32 cwnd = tp->snd_cwnd;
114 tcp_ca_event(tp, CA_EVENT_CWND_RESTART);
116 tp->snd_ssthresh = tcp_current_ssthresh(tp);
117 restart_cwnd = min(restart_cwnd, cwnd);
119 while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd)
121 tp->snd_cwnd = max(cwnd, restart_cwnd);
122 tp->snd_cwnd_stamp = tcp_time_stamp;
123 tp->snd_cwnd_used = 0;
126 static inline void tcp_event_data_sent(struct tcp_sock *tp,
127 struct sk_buff *skb, struct sock *sk)
129 u32 now = tcp_time_stamp;
131 if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto)
132 tcp_cwnd_restart(tp, __sk_dst_get(sk));
136 /* If it is a reply for ato after last received
137 * packet, enter pingpong mode.
139 if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato)
140 tp->ack.pingpong = 1;
143 static __inline__ void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
145 struct tcp_sock *tp = tcp_sk(sk);
147 tcp_dec_quickack_mode(tp, pkts);
148 tcp_clear_xmit_timer(sk, TCP_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 struct inet_sock *inet = inet_sk(sk);
269 struct tcp_sock *tp = tcp_sk(sk);
270 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
271 int tcp_header_size = tp->tcp_header_len;
276 BUG_ON(!tcp_skb_pcount(skb));
278 #define SYSCTL_FLAG_TSTAMPS 0x1
279 #define SYSCTL_FLAG_WSCALE 0x2
280 #define SYSCTL_FLAG_SACK 0x4
282 /* If congestion control is doing timestamping */
283 if (tp->ca_ops->rtt_sample)
284 do_gettimeofday(&skb->stamp);
287 if (tcb->flags & TCPCB_FLAG_SYN) {
288 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
289 if(sysctl_tcp_timestamps) {
290 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
291 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
293 if(sysctl_tcp_window_scaling) {
294 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
295 sysctl_flags |= SYSCTL_FLAG_WSCALE;
297 if(sysctl_tcp_sack) {
298 sysctl_flags |= SYSCTL_FLAG_SACK;
299 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
300 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
302 } else if (tp->rx_opt.eff_sacks) {
303 /* A SACK is 2 pad bytes, a 2 byte header, plus
304 * 2 32-bit sequence numbers for each SACK block.
306 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
307 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
310 if (tcp_packets_in_flight(tp) == 0)
311 tcp_ca_event(tp, CA_EVENT_TX_START);
313 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
315 skb_set_owner_w(skb, sk);
317 /* Build TCP header and checksum it. */
318 th->source = inet->sport;
319 th->dest = inet->dport;
320 th->seq = htonl(tcb->seq);
321 th->ack_seq = htonl(tp->rcv_nxt);
322 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
323 if (tcb->flags & TCPCB_FLAG_SYN) {
324 /* RFC1323: The window in SYN & SYN/ACK segments
327 th->window = htons(tp->rcv_wnd);
329 th->window = htons(tcp_select_window(sk));
335 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
336 th->urg_ptr = htons(tp->snd_up-tcb->seq);
340 if (tcb->flags & TCPCB_FLAG_SYN) {
341 tcp_syn_build_options((__u32 *)(th + 1),
342 tcp_advertise_mss(sk),
343 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
344 (sysctl_flags & SYSCTL_FLAG_SACK),
345 (sysctl_flags & SYSCTL_FLAG_WSCALE),
346 tp->rx_opt.rcv_wscale,
348 tp->rx_opt.ts_recent);
350 tcp_build_and_update_options((__u32 *)(th + 1),
353 TCP_ECN_send(sk, tp, skb, tcp_header_size);
355 tp->af_specific->send_check(sk, th, skb->len, skb);
357 if (tcb->flags & TCPCB_FLAG_ACK)
358 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
360 if (skb->len != tcp_header_size)
361 tcp_event_data_sent(tp, skb, sk);
363 TCP_INC_STATS(TCP_MIB_OUTSEGS);
365 err = tp->af_specific->queue_xmit(skb, 0);
371 /* NET_XMIT_CN is special. It does not guarantee,
372 * that this packet is lost. It tells that device
373 * is about to start to drop packets or already
374 * drops some packets of the same priority and
375 * invokes us to send less aggressively.
377 return err == NET_XMIT_CN ? 0 : err;
380 #undef SYSCTL_FLAG_TSTAMPS
381 #undef SYSCTL_FLAG_WSCALE
382 #undef SYSCTL_FLAG_SACK
386 /* This routine just queue's the buffer
388 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
389 * otherwise socket can stall.
391 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
393 struct tcp_sock *tp = tcp_sk(sk);
395 /* Advance write_seq and place onto the write_queue. */
396 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
397 skb_header_release(skb);
398 __skb_queue_tail(&sk->sk_write_queue, skb);
399 sk_charge_skb(sk, skb);
401 /* Queue it, remembering where we must start sending. */
402 if (sk->sk_send_head == NULL)
403 sk->sk_send_head = skb;
406 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb)
408 struct tcp_sock *tp = tcp_sk(sk);
410 if (skb->len <= tp->mss_cache ||
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 + (tp->mss_cache - 1);
421 factor /= tp->mss_cache;
422 skb_shinfo(skb)->tso_segs = factor;
423 skb_shinfo(skb)->tso_size = tp->mss_cache;
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 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len)
434 struct tcp_sock *tp = tcp_sk(sk);
435 struct sk_buff *buff;
439 nsize = skb_headlen(skb) - len;
443 if (skb_cloned(skb) &&
444 skb_is_nonlinear(skb) &&
445 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
448 /* Get a new skb... force flag on. */
449 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
451 return -ENOMEM; /* We'll just try again later. */
452 sk_charge_skb(sk, buff);
454 /* Correct the sequence numbers. */
455 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
456 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
457 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
459 /* PSH and FIN should only be set in the second packet. */
460 flags = TCP_SKB_CB(skb)->flags;
461 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
462 TCP_SKB_CB(buff)->flags = flags;
463 TCP_SKB_CB(buff)->sacked =
464 (TCP_SKB_CB(skb)->sacked &
465 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
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->stamp = skb->stamp;
489 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
490 tp->lost_out -= tcp_skb_pcount(skb);
491 tp->left_out -= tcp_skb_pcount(skb);
494 /* Fix up tso_factor for both original and new SKB. */
495 tcp_set_skb_tso_segs(sk, skb);
496 tcp_set_skb_tso_segs(sk, buff);
498 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
499 tp->lost_out += tcp_skb_pcount(skb);
500 tp->left_out += tcp_skb_pcount(skb);
503 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) {
504 tp->lost_out += tcp_skb_pcount(buff);
505 tp->left_out += tcp_skb_pcount(buff);
508 /* Link BUFF into the send queue. */
509 skb_header_release(buff);
510 __skb_append(skb, buff);
515 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
516 * eventually). The difference is that pulled data not copied, but
517 * immediately discarded.
519 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
525 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
526 if (skb_shinfo(skb)->frags[i].size <= eat) {
527 put_page(skb_shinfo(skb)->frags[i].page);
528 eat -= skb_shinfo(skb)->frags[i].size;
530 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
532 skb_shinfo(skb)->frags[k].page_offset += eat;
533 skb_shinfo(skb)->frags[k].size -= eat;
539 skb_shinfo(skb)->nr_frags = k;
541 skb->tail = skb->data;
542 skb->data_len -= len;
543 skb->len = skb->data_len;
547 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
549 if (skb_cloned(skb) &&
550 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
553 if (len <= skb_headlen(skb)) {
554 __skb_pull(skb, len);
556 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
560 TCP_SKB_CB(skb)->seq += len;
561 skb->ip_summed = CHECKSUM_HW;
563 skb->truesize -= len;
564 sk->sk_wmem_queued -= len;
565 sk->sk_forward_alloc += len;
566 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
568 /* Any change of skb->len requires recalculation of tso
571 if (tcp_skb_pcount(skb) > 1)
572 tcp_set_skb_tso_segs(sk, skb);
577 /* This function synchronize snd mss to current pmtu/exthdr set.
579 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
580 for TCP options, but includes only bare TCP header.
582 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
583 It is minumum of user_mss and mss received with SYN.
584 It also does not include TCP options.
586 tp->pmtu_cookie is last pmtu, seen by this function.
588 tp->mss_cache is current effective sending mss, including
589 all tcp options except for SACKs. It is evaluated,
590 taking into account current pmtu, but never exceeds
591 tp->rx_opt.mss_clamp.
593 NOTE1. rfc1122 clearly states that advertised MSS
594 DOES NOT include either tcp or ip options.
596 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
597 this function. --ANK (980731)
600 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
602 struct tcp_sock *tp = tcp_sk(sk);
605 /* Calculate base mss without TCP options:
606 It is MMS_S - sizeof(tcphdr) of rfc1122
608 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
610 /* Clamp it (mss_clamp does not include tcp options) */
611 if (mss_now > tp->rx_opt.mss_clamp)
612 mss_now = tp->rx_opt.mss_clamp;
614 /* Now subtract optional transport overhead */
615 mss_now -= tp->ext_header_len;
617 /* Then reserve room for full set of TCP options and 8 bytes of data */
621 /* Now subtract TCP options size, not including SACKs */
622 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
624 /* Bound mss with half of window */
625 if (tp->max_window && mss_now > (tp->max_window>>1))
626 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
628 /* And store cached results */
629 tp->pmtu_cookie = pmtu;
630 tp->mss_cache = mss_now;
635 /* Compute the current effective MSS, taking SACKs and IP options,
636 * and even PMTU discovery events into account.
638 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
639 * cannot be large. However, taking into account rare use of URG, this
642 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
644 struct tcp_sock *tp = tcp_sk(sk);
645 struct dst_entry *dst = __sk_dst_get(sk);
650 mss_now = tp->mss_cache;
653 (sk->sk_route_caps & NETIF_F_TSO) &&
658 u32 mtu = dst_mtu(dst);
659 if (mtu != tp->pmtu_cookie)
660 mss_now = tcp_sync_mss(sk, mtu);
663 if (tp->rx_opt.eff_sacks)
664 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
665 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
667 xmit_size_goal = mss_now;
670 xmit_size_goal = 65535 -
671 tp->af_specific->net_header_len -
672 tp->ext_header_len - tp->tcp_header_len;
674 if (tp->max_window &&
675 (xmit_size_goal > (tp->max_window >> 1)))
676 xmit_size_goal = max((tp->max_window >> 1),
677 68U - tp->tcp_header_len);
679 xmit_size_goal -= (xmit_size_goal % mss_now);
681 tp->xmit_size_goal = xmit_size_goal;
686 /* Congestion window validation. (RFC2861) */
688 static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
690 __u32 packets_out = tp->packets_out;
692 if (packets_out >= tp->snd_cwnd) {
693 /* Network is feed fully. */
694 tp->snd_cwnd_used = 0;
695 tp->snd_cwnd_stamp = tcp_time_stamp;
697 /* Network starves. */
698 if (tp->packets_out > tp->snd_cwnd_used)
699 tp->snd_cwnd_used = tp->packets_out;
701 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= tp->rto)
702 tcp_cwnd_application_limited(sk);
706 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
708 u32 window, cwnd_len;
710 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
711 cwnd_len = mss_now * cwnd;
712 return min(window, cwnd_len);
715 /* Can at least one segment of SKB be sent right now, according to the
716 * congestion window rules? If so, return how many segments are allowed.
718 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
722 /* Don't be strict about the congestion window for the final FIN. */
723 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
726 in_flight = tcp_packets_in_flight(tp);
728 if (in_flight < cwnd)
729 return (cwnd - in_flight);
734 /* This must be invoked the first time we consider transmitting
737 static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb)
739 int tso_segs = tcp_skb_pcount(skb);
742 tcp_set_skb_tso_segs(sk, skb);
743 tso_segs = tcp_skb_pcount(skb);
748 static inline int tcp_minshall_check(const struct tcp_sock *tp)
750 return after(tp->snd_sml,tp->snd_una) &&
751 !after(tp->snd_sml, tp->snd_nxt);
754 /* Return 0, if packet can be sent now without violation Nagle's rules:
755 * 1. It is full sized.
756 * 2. Or it contains FIN. (already checked by caller)
757 * 3. Or TCP_NODELAY was set.
758 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
759 * With Minshall's modification: all sent small packets are ACKed.
762 static inline int tcp_nagle_check(const struct tcp_sock *tp,
763 const struct sk_buff *skb,
764 unsigned mss_now, int nonagle)
766 return (skb->len < mss_now &&
767 ((nonagle&TCP_NAGLE_CORK) ||
770 tcp_minshall_check(tp))));
773 /* Return non-zero if the Nagle test allows this packet to be
776 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
777 unsigned int cur_mss, int nonagle)
779 /* Nagle rule does not apply to frames, which sit in the middle of the
780 * write_queue (they have no chances to get new data).
782 * This is implemented in the callers, where they modify the 'nonagle'
783 * argument based upon the location of SKB in the send queue.
785 if (nonagle & TCP_NAGLE_PUSH)
788 /* Don't use the nagle rule for urgent data (or for the final FIN). */
790 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
793 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
799 /* Does at least the first segment of SKB fit into the send window? */
800 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
802 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
804 if (skb->len > cur_mss)
805 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
807 return !after(end_seq, tp->snd_una + tp->snd_wnd);
810 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
811 * should be put on the wire right now. If so, it returns the number of
812 * packets allowed by the congestion window.
814 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
815 unsigned int cur_mss, int nonagle)
817 struct tcp_sock *tp = tcp_sk(sk);
818 unsigned int cwnd_quota;
820 tcp_init_tso_segs(sk, skb);
822 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
825 cwnd_quota = tcp_cwnd_test(tp, skb);
827 !tcp_snd_wnd_test(tp, skb, cur_mss))
833 static inline int tcp_skb_is_last(const struct sock *sk,
834 const struct sk_buff *skb)
836 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
839 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
841 struct sk_buff *skb = sk->sk_send_head;
844 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
845 (tcp_skb_is_last(sk, skb) ?
850 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
851 * which is put after SKB on the list. It is very much like
852 * tcp_fragment() except that it may make several kinds of assumptions
853 * in order to speed up the splitting operation. In particular, we
854 * know that all the data is in scatter-gather pages, and that the
855 * packet has never been sent out before (and thus is not cloned).
857 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len)
859 struct sk_buff *buff;
860 int nlen = skb->len - len;
863 /* All of a TSO frame must be composed of paged data. */
864 BUG_ON(skb->len != skb->data_len);
866 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
867 if (unlikely(buff == NULL))
870 buff->truesize = nlen;
871 skb->truesize -= nlen;
873 /* Correct the sequence numbers. */
874 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
875 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
876 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
878 /* PSH and FIN should only be set in the second packet. */
879 flags = TCP_SKB_CB(skb)->flags;
880 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
881 TCP_SKB_CB(buff)->flags = flags;
883 /* This packet was never sent out yet, so no SACK bits. */
884 TCP_SKB_CB(buff)->sacked = 0;
886 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
887 skb_split(skb, buff, len);
889 /* Fix up tso_factor for both original and new SKB. */
890 tcp_set_skb_tso_segs(sk, skb);
891 tcp_set_skb_tso_segs(sk, buff);
893 /* Link BUFF into the send queue. */
894 skb_header_release(buff);
895 __skb_append(skb, buff);
900 /* Try to defer sending, if possible, in order to minimize the amount
901 * of TSO splitting we do. View it as a kind of TSO Nagle test.
903 * This algorithm is from John Heffner.
905 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
907 u32 send_win, cong_win, limit, in_flight;
909 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
912 in_flight = tcp_packets_in_flight(tp);
914 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
915 (tp->snd_cwnd <= in_flight));
917 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
919 /* From in_flight test above, we know that cwnd > in_flight. */
920 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
922 limit = min(send_win, cong_win);
924 /* If sk_send_head can be sent fully now, just do it. */
925 if (skb->len <= limit)
928 if (sysctl_tcp_tso_win_divisor) {
929 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
931 /* If at least some fraction of a window is available,
934 chunk /= sysctl_tcp_tso_win_divisor;
938 /* Different approach, try not to defer past a single
939 * ACK. Receiver should ACK every other full sized
940 * frame, so if we have space for more than 3 frames
943 if (limit > tcp_max_burst(tp) * tp->mss_cache)
947 /* Ok, it looks like it is advisable to defer. */
951 /* This routine writes packets to the network. It advances the
952 * send_head. This happens as incoming acks open up the remote
955 * Returns 1, if no segments are in flight and we have queued segments, but
956 * cannot send anything now because of SWS or another problem.
958 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
960 struct tcp_sock *tp = tcp_sk(sk);
962 unsigned int tso_segs, sent_pkts;
965 /* If we are closed, the bytes will have to remain here.
966 * In time closedown will finish, we empty the write queue and all
969 if (unlikely(sk->sk_state == TCP_CLOSE))
972 skb = sk->sk_send_head;
976 tso_segs = tcp_init_tso_segs(sk, skb);
977 cwnd_quota = tcp_cwnd_test(tp, skb);
978 if (unlikely(!cwnd_quota))
982 while (likely(tcp_snd_wnd_test(tp, skb, mss_now))) {
986 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
987 (tcp_skb_is_last(sk, skb) ?
988 nonagle : TCP_NAGLE_PUSH))))
991 if (tcp_tso_should_defer(sk, tp, skb))
996 u32 limit = tcp_window_allows(tp, skb,
997 mss_now, cwnd_quota);
999 if (skb->len < limit) {
1000 unsigned int trim = skb->len % mss_now;
1003 limit = skb->len - trim;
1005 if (skb->len > limit) {
1006 if (tso_fragment(sk, skb, limit))
1009 } else if (unlikely(skb->len > mss_now)) {
1010 if (unlikely(tcp_fragment(sk, skb, mss_now)))
1014 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1016 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
1019 /* Advance the send_head. This one is sent out.
1020 * This call will increment packets_out.
1022 update_send_head(sk, tp, skb);
1024 tcp_minshall_update(tp, mss_now, skb);
1027 /* Do not optimize this to use tso_segs. If we chopped up
1028 * the packet above, tso_segs will no longer be valid.
1030 cwnd_quota -= tcp_skb_pcount(skb);
1032 BUG_ON(cwnd_quota < 0);
1036 skb = sk->sk_send_head;
1039 tso_segs = tcp_init_tso_segs(sk, skb);
1042 if (likely(sent_pkts)) {
1043 tcp_cwnd_validate(sk, tp);
1047 return !tp->packets_out && sk->sk_send_head;
1050 /* Push out any pending frames which were held back due to
1051 * TCP_CORK or attempt at coalescing tiny packets.
1052 * The socket must be locked by the caller.
1054 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1055 unsigned int cur_mss, int nonagle)
1057 struct sk_buff *skb = sk->sk_send_head;
1060 if (tcp_write_xmit(sk, cur_mss, nonagle))
1061 tcp_check_probe_timer(sk, tp);
1065 /* Send _single_ skb sitting at the send head. This function requires
1066 * true push pending frames to setup probe timer etc.
1068 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1070 struct tcp_sock *tp = tcp_sk(sk);
1071 struct sk_buff *skb = sk->sk_send_head;
1072 unsigned int tso_segs, cwnd_quota;
1074 BUG_ON(!skb || skb->len < mss_now);
1076 tso_segs = tcp_init_tso_segs(sk, skb);
1077 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1079 if (likely(cwnd_quota)) {
1083 u32 limit = tcp_window_allows(tp, skb,
1084 mss_now, cwnd_quota);
1086 if (skb->len < limit) {
1087 unsigned int trim = skb->len % mss_now;
1090 limit = skb->len - trim;
1092 if (skb->len > limit) {
1093 if (unlikely(tso_fragment(sk, skb, limit)))
1096 } else if (unlikely(skb->len > mss_now)) {
1097 if (unlikely(tcp_fragment(sk, skb, mss_now)))
1101 /* Send it out now. */
1102 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1104 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1105 update_send_head(sk, tp, skb);
1106 tcp_cwnd_validate(sk, tp);
1112 /* This function returns the amount that we can raise the
1113 * usable window based on the following constraints
1115 * 1. The window can never be shrunk once it is offered (RFC 793)
1116 * 2. We limit memory per socket
1119 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1120 * RECV.NEXT + RCV.WIN fixed until:
1121 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1123 * i.e. don't raise the right edge of the window until you can raise
1124 * it at least MSS bytes.
1126 * Unfortunately, the recommended algorithm breaks header prediction,
1127 * since header prediction assumes th->window stays fixed.
1129 * Strictly speaking, keeping th->window fixed violates the receiver
1130 * side SWS prevention criteria. The problem is that under this rule
1131 * a stream of single byte packets will cause the right side of the
1132 * window to always advance by a single byte.
1134 * Of course, if the sender implements sender side SWS prevention
1135 * then this will not be a problem.
1137 * BSD seems to make the following compromise:
1139 * If the free space is less than the 1/4 of the maximum
1140 * space available and the free space is less than 1/2 mss,
1141 * then set the window to 0.
1142 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1143 * Otherwise, just prevent the window from shrinking
1144 * and from being larger than the largest representable value.
1146 * This prevents incremental opening of the window in the regime
1147 * where TCP is limited by the speed of the reader side taking
1148 * data out of the TCP receive queue. It does nothing about
1149 * those cases where the window is constrained on the sender side
1150 * because the pipeline is full.
1152 * BSD also seems to "accidentally" limit itself to windows that are a
1153 * multiple of MSS, at least until the free space gets quite small.
1154 * This would appear to be a side effect of the mbuf implementation.
1155 * Combining these two algorithms results in the observed behavior
1156 * of having a fixed window size at almost all times.
1158 * Below we obtain similar behavior by forcing the offered window to
1159 * a multiple of the mss when it is feasible to do so.
1161 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1162 * Regular options like TIMESTAMP are taken into account.
1164 u32 __tcp_select_window(struct sock *sk)
1166 struct tcp_sock *tp = tcp_sk(sk);
1167 /* MSS for the peer's data. Previous verions used mss_clamp
1168 * here. I don't know if the value based on our guesses
1169 * of peer's MSS is better for the performance. It's more correct
1170 * but may be worse for the performance because of rcv_mss
1171 * fluctuations. --SAW 1998/11/1
1173 int mss = tp->ack.rcv_mss;
1174 int free_space = tcp_space(sk);
1175 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1178 if (mss > full_space)
1181 if (free_space < full_space/2) {
1184 if (tcp_memory_pressure)
1185 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1187 if (free_space < mss)
1191 if (free_space > tp->rcv_ssthresh)
1192 free_space = tp->rcv_ssthresh;
1194 /* Don't do rounding if we are using window scaling, since the
1195 * scaled window will not line up with the MSS boundary anyway.
1197 window = tp->rcv_wnd;
1198 if (tp->rx_opt.rcv_wscale) {
1199 window = free_space;
1201 /* Advertise enough space so that it won't get scaled away.
1202 * Import case: prevent zero window announcement if
1203 * 1<<rcv_wscale > mss.
1205 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1206 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1207 << tp->rx_opt.rcv_wscale);
1209 /* Get the largest window that is a nice multiple of mss.
1210 * Window clamp already applied above.
1211 * If our current window offering is within 1 mss of the
1212 * free space we just keep it. This prevents the divide
1213 * and multiply from happening most of the time.
1214 * We also don't do any window rounding when the free space
1217 if (window <= free_space - mss || window > free_space)
1218 window = (free_space/mss)*mss;
1224 /* Attempt to collapse two adjacent SKB's during retransmission. */
1225 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1227 struct tcp_sock *tp = tcp_sk(sk);
1228 struct sk_buff *next_skb = skb->next;
1230 /* The first test we must make is that neither of these two
1231 * SKB's are still referenced by someone else.
1233 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1234 int skb_size = skb->len, next_skb_size = next_skb->len;
1235 u16 flags = TCP_SKB_CB(skb)->flags;
1237 /* Also punt if next skb has been SACK'd. */
1238 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1241 /* Next skb is out of window. */
1242 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1245 /* Punt if not enough space exists in the first SKB for
1246 * the data in the second, or the total combined payload
1247 * would exceed the MSS.
1249 if ((next_skb_size > skb_tailroom(skb)) ||
1250 ((skb_size + next_skb_size) > mss_now))
1253 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1254 tcp_skb_pcount(next_skb) != 1);
1256 /* Ok. We will be able to collapse the packet. */
1257 __skb_unlink(next_skb, next_skb->list);
1259 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1261 if (next_skb->ip_summed == CHECKSUM_HW)
1262 skb->ip_summed = CHECKSUM_HW;
1264 if (skb->ip_summed != CHECKSUM_HW)
1265 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1267 /* Update sequence range on original skb. */
1268 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1270 /* Merge over control information. */
1271 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1272 TCP_SKB_CB(skb)->flags = flags;
1274 /* All done, get rid of second SKB and account for it so
1275 * packet counting does not break.
1277 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1278 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1279 tp->retrans_out -= tcp_skb_pcount(next_skb);
1280 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1281 tp->lost_out -= tcp_skb_pcount(next_skb);
1282 tp->left_out -= tcp_skb_pcount(next_skb);
1284 /* Reno case is special. Sigh... */
1285 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1286 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1287 tp->left_out -= tcp_skb_pcount(next_skb);
1290 /* Not quite right: it can be > snd.fack, but
1291 * it is better to underestimate fackets.
1293 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1294 tcp_packets_out_dec(tp, next_skb);
1295 sk_stream_free_skb(sk, next_skb);
1299 /* Do a simple retransmit without using the backoff mechanisms in
1300 * tcp_timer. This is used for path mtu discovery.
1301 * The socket is already locked here.
1303 void tcp_simple_retransmit(struct sock *sk)
1305 struct tcp_sock *tp = tcp_sk(sk);
1306 struct sk_buff *skb;
1307 unsigned int mss = tcp_current_mss(sk, 0);
1310 sk_stream_for_retrans_queue(skb, sk) {
1311 if (skb->len > mss &&
1312 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1313 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1314 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1315 tp->retrans_out -= tcp_skb_pcount(skb);
1317 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1318 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1319 tp->lost_out += tcp_skb_pcount(skb);
1328 tcp_sync_left_out(tp);
1330 /* Don't muck with the congestion window here.
1331 * Reason is that we do not increase amount of _data_
1332 * in network, but units changed and effective
1333 * cwnd/ssthresh really reduced now.
1335 if (tp->ca_state != TCP_CA_Loss) {
1336 tp->high_seq = tp->snd_nxt;
1337 tp->snd_ssthresh = tcp_current_ssthresh(tp);
1338 tp->prior_ssthresh = 0;
1339 tp->undo_marker = 0;
1340 tcp_set_ca_state(tp, TCP_CA_Loss);
1342 tcp_xmit_retransmit_queue(sk);
1345 /* This retransmits one SKB. Policy decisions and retransmit queue
1346 * state updates are done by the caller. Returns non-zero if an
1347 * error occurred which prevented the send.
1349 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1351 struct tcp_sock *tp = tcp_sk(sk);
1352 unsigned int cur_mss = tcp_current_mss(sk, 0);
1355 /* Do not sent more than we queued. 1/4 is reserved for possible
1356 * copying overhead: frgagmentation, tunneling, mangling etc.
1358 if (atomic_read(&sk->sk_wmem_alloc) >
1359 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1362 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1363 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1366 if (sk->sk_route_caps & NETIF_F_TSO) {
1367 sk->sk_route_caps &= ~NETIF_F_TSO;
1368 sock_set_flag(sk, SOCK_NO_LARGESEND);
1371 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1375 /* If receiver has shrunk his window, and skb is out of
1376 * new window, do not retransmit it. The exception is the
1377 * case, when window is shrunk to zero. In this case
1378 * our retransmit serves as a zero window probe.
1380 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1381 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1384 if (skb->len > cur_mss) {
1385 int old_factor = tcp_skb_pcount(skb);
1388 if (tcp_fragment(sk, skb, cur_mss))
1389 return -ENOMEM; /* We'll try again later. */
1391 /* New SKB created, account for it. */
1392 new_factor = tcp_skb_pcount(skb);
1393 tp->packets_out -= old_factor - new_factor;
1394 tp->packets_out += tcp_skb_pcount(skb->next);
1397 /* Collapse two adjacent packets if worthwhile and we can. */
1398 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1399 (skb->len < (cur_mss >> 1)) &&
1400 (skb->next != sk->sk_send_head) &&
1401 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1402 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1403 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1404 (sysctl_tcp_retrans_collapse != 0))
1405 tcp_retrans_try_collapse(sk, skb, cur_mss);
1407 if(tp->af_specific->rebuild_header(sk))
1408 return -EHOSTUNREACH; /* Routing failure or similar. */
1410 /* Some Solaris stacks overoptimize and ignore the FIN on a
1411 * retransmit when old data is attached. So strip it off
1412 * since it is cheap to do so and saves bytes on the network.
1415 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1416 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1417 if (!pskb_trim(skb, 0)) {
1418 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1419 skb_shinfo(skb)->tso_segs = 1;
1420 skb_shinfo(skb)->tso_size = 0;
1421 skb->ip_summed = CHECKSUM_NONE;
1426 /* Make a copy, if the first transmission SKB clone we made
1427 * is still in somebody's hands, else make a clone.
1429 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1431 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1432 pskb_copy(skb, GFP_ATOMIC):
1433 skb_clone(skb, GFP_ATOMIC)));
1436 /* Update global TCP statistics. */
1437 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1439 tp->total_retrans++;
1441 #if FASTRETRANS_DEBUG > 0
1442 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1443 if (net_ratelimit())
1444 printk(KERN_DEBUG "retrans_out leaked.\n");
1447 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1448 tp->retrans_out += tcp_skb_pcount(skb);
1450 /* Save stamp of the first retransmit. */
1451 if (!tp->retrans_stamp)
1452 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1456 /* snd_nxt is stored to detect loss of retransmitted segment,
1457 * see tcp_input.c tcp_sacktag_write_queue().
1459 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1464 /* This gets called after a retransmit timeout, and the initially
1465 * retransmitted data is acknowledged. It tries to continue
1466 * resending the rest of the retransmit queue, until either
1467 * we've sent it all or the congestion window limit is reached.
1468 * If doing SACK, the first ACK which comes back for a timeout
1469 * based retransmit packet might feed us FACK information again.
1470 * If so, we use it to avoid unnecessarily retransmissions.
1472 void tcp_xmit_retransmit_queue(struct sock *sk)
1474 struct tcp_sock *tp = tcp_sk(sk);
1475 struct sk_buff *skb;
1476 int packet_cnt = tp->lost_out;
1478 /* First pass: retransmit lost packets. */
1480 sk_stream_for_retrans_queue(skb, sk) {
1481 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1483 /* Assume this retransmit will generate
1484 * only one packet for congestion window
1485 * calculation purposes. This works because
1486 * tcp_retransmit_skb() will chop up the
1487 * packet to be MSS sized and all the
1488 * packet counting works out.
1490 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1493 if (sacked&TCPCB_LOST) {
1494 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1495 if (tcp_retransmit_skb(sk, skb))
1497 if (tp->ca_state != TCP_CA_Loss)
1498 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1500 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1503 skb_peek(&sk->sk_write_queue))
1504 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1507 packet_cnt -= tcp_skb_pcount(skb);
1508 if (packet_cnt <= 0)
1514 /* OK, demanded retransmission is finished. */
1516 /* Forward retransmissions are possible only during Recovery. */
1517 if (tp->ca_state != TCP_CA_Recovery)
1520 /* No forward retransmissions in Reno are possible. */
1521 if (!tp->rx_opt.sack_ok)
1524 /* Yeah, we have to make difficult choice between forward transmission
1525 * and retransmission... Both ways have their merits...
1527 * For now we do not retransmit anything, while we have some new
1531 if (tcp_may_send_now(sk, tp))
1536 sk_stream_for_retrans_queue(skb, sk) {
1537 /* Similar to the retransmit loop above we
1538 * can pretend that the retransmitted SKB
1539 * we send out here will be composed of one
1540 * real MSS sized packet because tcp_retransmit_skb()
1541 * will fragment it if necessary.
1543 if (++packet_cnt > tp->fackets_out)
1546 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1549 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1552 /* Ok, retransmit it. */
1553 if (tcp_retransmit_skb(sk, skb))
1556 if (skb == skb_peek(&sk->sk_write_queue))
1557 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1559 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1564 /* Send a fin. The caller locks the socket for us. This cannot be
1565 * allowed to fail queueing a FIN frame under any circumstances.
1567 void tcp_send_fin(struct sock *sk)
1569 struct tcp_sock *tp = tcp_sk(sk);
1570 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1573 /* Optimization, tack on the FIN if we have a queue of
1574 * unsent frames. But be careful about outgoing SACKS
1577 mss_now = tcp_current_mss(sk, 1);
1579 if (sk->sk_send_head != NULL) {
1580 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1581 TCP_SKB_CB(skb)->end_seq++;
1584 /* Socket is locked, keep trying until memory is available. */
1586 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL);
1592 /* Reserve space for headers and prepare control bits. */
1593 skb_reserve(skb, MAX_TCP_HEADER);
1595 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1596 TCP_SKB_CB(skb)->sacked = 0;
1597 skb_shinfo(skb)->tso_segs = 1;
1598 skb_shinfo(skb)->tso_size = 0;
1600 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1601 TCP_SKB_CB(skb)->seq = tp->write_seq;
1602 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1603 tcp_queue_skb(sk, skb);
1605 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1608 /* We get here when a process closes a file descriptor (either due to
1609 * an explicit close() or as a byproduct of exit()'ing) and there
1610 * was unread data in the receive queue. This behavior is recommended
1611 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1613 void tcp_send_active_reset(struct sock *sk, int priority)
1615 struct tcp_sock *tp = tcp_sk(sk);
1616 struct sk_buff *skb;
1618 /* NOTE: No TCP options attached and we never retransmit this. */
1619 skb = alloc_skb(MAX_TCP_HEADER, priority);
1621 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1625 /* Reserve space for headers and prepare control bits. */
1626 skb_reserve(skb, MAX_TCP_HEADER);
1628 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1629 TCP_SKB_CB(skb)->sacked = 0;
1630 skb_shinfo(skb)->tso_segs = 1;
1631 skb_shinfo(skb)->tso_size = 0;
1634 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1635 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1636 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1637 if (tcp_transmit_skb(sk, skb))
1638 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1641 /* WARNING: This routine must only be called when we have already sent
1642 * a SYN packet that crossed the incoming SYN that caused this routine
1643 * to get called. If this assumption fails then the initial rcv_wnd
1644 * and rcv_wscale values will not be correct.
1646 int tcp_send_synack(struct sock *sk)
1648 struct sk_buff* skb;
1650 skb = skb_peek(&sk->sk_write_queue);
1651 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1652 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1655 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1656 if (skb_cloned(skb)) {
1657 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1660 __skb_unlink(skb, &sk->sk_write_queue);
1661 skb_header_release(nskb);
1662 __skb_queue_head(&sk->sk_write_queue, nskb);
1663 sk_stream_free_skb(sk, skb);
1664 sk_charge_skb(sk, nskb);
1668 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1669 TCP_ECN_send_synack(tcp_sk(sk), skb);
1671 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1672 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1676 * Prepare a SYN-ACK.
1678 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1679 struct request_sock *req)
1681 struct inet_request_sock *ireq = inet_rsk(req);
1682 struct tcp_sock *tp = tcp_sk(sk);
1684 int tcp_header_size;
1685 struct sk_buff *skb;
1687 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1691 /* Reserve space for headers. */
1692 skb_reserve(skb, MAX_TCP_HEADER);
1694 skb->dst = dst_clone(dst);
1696 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1697 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1698 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1699 /* SACK_PERM is in the place of NOP NOP of TS */
1700 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1701 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1703 memset(th, 0, sizeof(struct tcphdr));
1706 if (dst->dev->features&NETIF_F_TSO)
1708 TCP_ECN_make_synack(req, th);
1709 th->source = inet_sk(sk)->sport;
1710 th->dest = ireq->rmt_port;
1711 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
1712 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1713 TCP_SKB_CB(skb)->sacked = 0;
1714 skb_shinfo(skb)->tso_segs = 1;
1715 skb_shinfo(skb)->tso_size = 0;
1716 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1717 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
1718 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1720 /* Set this up on the first call only */
1721 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1722 /* tcp_full_space because it is guaranteed to be the first packet */
1723 tcp_select_initial_window(tcp_full_space(sk),
1724 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1729 ireq->rcv_wscale = rcv_wscale;
1732 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1733 th->window = htons(req->rcv_wnd);
1735 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1736 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1737 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
1738 TCP_SKB_CB(skb)->when,
1742 th->doff = (tcp_header_size >> 2);
1743 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1748 * Do all connect socket setups that can be done AF independent.
1750 static inline void tcp_connect_init(struct sock *sk)
1752 struct dst_entry *dst = __sk_dst_get(sk);
1753 struct tcp_sock *tp = tcp_sk(sk);
1756 /* We'll fix this up when we get a response from the other end.
1757 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1759 tp->tcp_header_len = sizeof(struct tcphdr) +
1760 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1762 /* If user gave his TCP_MAXSEG, record it to clamp */
1763 if (tp->rx_opt.user_mss)
1764 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1766 tcp_sync_mss(sk, dst_mtu(dst));
1768 if (!tp->window_clamp)
1769 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1770 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1771 tcp_initialize_rcv_mss(sk);
1773 tcp_select_initial_window(tcp_full_space(sk),
1774 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1777 sysctl_tcp_window_scaling,
1780 tp->rx_opt.rcv_wscale = rcv_wscale;
1781 tp->rcv_ssthresh = tp->rcv_wnd;
1784 sock_reset_flag(sk, SOCK_DONE);
1786 tcp_init_wl(tp, tp->write_seq, 0);
1787 tp->snd_una = tp->write_seq;
1788 tp->snd_sml = tp->write_seq;
1793 tp->rto = TCP_TIMEOUT_INIT;
1794 tp->retransmits = 0;
1795 tcp_clear_retrans(tp);
1799 * Build a SYN and send it off.
1801 int tcp_connect(struct sock *sk)
1803 struct tcp_sock *tp = tcp_sk(sk);
1804 struct sk_buff *buff;
1806 tcp_connect_init(sk);
1808 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation);
1809 if (unlikely(buff == NULL))
1812 /* Reserve space for headers. */
1813 skb_reserve(buff, MAX_TCP_HEADER);
1815 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1816 TCP_ECN_send_syn(sk, tp, buff);
1817 TCP_SKB_CB(buff)->sacked = 0;
1818 skb_shinfo(buff)->tso_segs = 1;
1819 skb_shinfo(buff)->tso_size = 0;
1821 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1822 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1823 tp->snd_nxt = tp->write_seq;
1824 tp->pushed_seq = tp->write_seq;
1827 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1828 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1829 skb_header_release(buff);
1830 __skb_queue_tail(&sk->sk_write_queue, buff);
1831 sk_charge_skb(sk, buff);
1832 tp->packets_out += tcp_skb_pcount(buff);
1833 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1834 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1836 /* Timer for repeating the SYN until an answer. */
1837 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1841 /* Send out a delayed ack, the caller does the policy checking
1842 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1845 void tcp_send_delayed_ack(struct sock *sk)
1847 struct tcp_sock *tp = tcp_sk(sk);
1848 int ato = tp->ack.ato;
1849 unsigned long timeout;
1851 if (ato > TCP_DELACK_MIN) {
1854 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED))
1855 max_ato = TCP_DELACK_MAX;
1857 /* Slow path, intersegment interval is "high". */
1859 /* If some rtt estimate is known, use it to bound delayed ack.
1860 * Do not use tp->rto here, use results of rtt measurements
1864 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1870 ato = min(ato, max_ato);
1873 /* Stay within the limit we were given */
1874 timeout = jiffies + ato;
1876 /* Use new timeout only if there wasn't a older one earlier. */
1877 if (tp->ack.pending&TCP_ACK_TIMER) {
1878 /* If delack timer was blocked or is about to expire,
1881 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) {
1886 if (!time_before(timeout, tp->ack.timeout))
1887 timeout = tp->ack.timeout;
1889 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER;
1890 tp->ack.timeout = timeout;
1891 sk_reset_timer(sk, &tp->delack_timer, timeout);
1894 /* This routine sends an ack and also updates the window. */
1895 void tcp_send_ack(struct sock *sk)
1897 /* If we have been reset, we may not send again. */
1898 if (sk->sk_state != TCP_CLOSE) {
1899 struct tcp_sock *tp = tcp_sk(sk);
1900 struct sk_buff *buff;
1902 /* We are not putting this on the write queue, so
1903 * tcp_transmit_skb() will set the ownership to this
1906 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1908 tcp_schedule_ack(tp);
1909 tp->ack.ato = TCP_ATO_MIN;
1910 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
1914 /* Reserve space for headers and prepare control bits. */
1915 skb_reserve(buff, MAX_TCP_HEADER);
1917 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1918 TCP_SKB_CB(buff)->sacked = 0;
1919 skb_shinfo(buff)->tso_segs = 1;
1920 skb_shinfo(buff)->tso_size = 0;
1922 /* Send it off, this clears delayed acks for us. */
1923 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1924 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1925 tcp_transmit_skb(sk, buff);
1929 /* This routine sends a packet with an out of date sequence
1930 * number. It assumes the other end will try to ack it.
1932 * Question: what should we make while urgent mode?
1933 * 4.4BSD forces sending single byte of data. We cannot send
1934 * out of window data, because we have SND.NXT==SND.MAX...
1936 * Current solution: to send TWO zero-length segments in urgent mode:
1937 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1938 * out-of-date with SND.UNA-1 to probe window.
1940 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1942 struct tcp_sock *tp = tcp_sk(sk);
1943 struct sk_buff *skb;
1945 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1946 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1950 /* Reserve space for headers and set control bits. */
1951 skb_reserve(skb, MAX_TCP_HEADER);
1953 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1954 TCP_SKB_CB(skb)->sacked = urgent;
1955 skb_shinfo(skb)->tso_segs = 1;
1956 skb_shinfo(skb)->tso_size = 0;
1958 /* Use a previous sequence. This should cause the other
1959 * end to send an ack. Don't queue or clone SKB, just
1962 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1963 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1964 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1965 return tcp_transmit_skb(sk, skb);
1968 int tcp_write_wakeup(struct sock *sk)
1970 if (sk->sk_state != TCP_CLOSE) {
1971 struct tcp_sock *tp = tcp_sk(sk);
1972 struct sk_buff *skb;
1974 if ((skb = sk->sk_send_head) != NULL &&
1975 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1977 unsigned int mss = tcp_current_mss(sk, 0);
1978 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1980 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1981 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1983 /* We are probing the opening of a window
1984 * but the window size is != 0
1985 * must have been a result SWS avoidance ( sender )
1987 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1989 seg_size = min(seg_size, mss);
1990 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1991 if (tcp_fragment(sk, skb, seg_size))
1993 /* SWS override triggered forced fragmentation.
1994 * Disable TSO, the connection is too sick. */
1995 if (sk->sk_route_caps & NETIF_F_TSO) {
1996 sock_set_flag(sk, SOCK_NO_LARGESEND);
1997 sk->sk_route_caps &= ~NETIF_F_TSO;
1999 } else if (!tcp_skb_pcount(skb))
2000 tcp_set_skb_tso_segs(sk, skb);
2002 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2003 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2004 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
2006 update_send_head(sk, tp, skb);
2011 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2012 tcp_xmit_probe_skb(sk, TCPCB_URG);
2013 return tcp_xmit_probe_skb(sk, 0);
2019 /* A window probe timeout has occurred. If window is not closed send
2020 * a partial packet else a zero probe.
2022 void tcp_send_probe0(struct sock *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. */
2037 if (tp->backoff < sysctl_tcp_retries2)
2040 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
2041 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2043 /* If packet was not sent due to local congestion,
2044 * do not backoff and do not remember probes_out.
2045 * Let local senders to fight for local resources.
2047 * Use accumulated backoff yet.
2049 if (!tp->probes_out)
2051 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
2052 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL));
2056 EXPORT_SYMBOL(tcp_connect);
2057 EXPORT_SYMBOL(tcp_make_synack);
2058 EXPORT_SYMBOL(tcp_simple_retransmit);
2059 EXPORT_SYMBOL(tcp_sync_mss);