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 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows = 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor = 3;
59 int sysctl_tcp_mtu_probing = 0;
60 int sysctl_tcp_base_mss = 512;
62 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
65 sk->sk_send_head = skb->next;
66 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
67 sk->sk_send_head = NULL;
68 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
69 tcp_packets_out_inc(sk, tp, skb);
72 /* SND.NXT, if window was not shrunk.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
80 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
83 return tp->snd_una+tp->snd_wnd;
86 /* Calculate mss to advertise in SYN segment.
87 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
89 * 1. It is independent of path mtu.
90 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
91 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
92 * attached devices, because some buggy hosts are confused by
94 * 4. We do not make 3, we advertise MSS, calculated from first
95 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
96 * This may be overridden via information stored in routing table.
97 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
98 * probably even Jumbo".
100 static __u16 tcp_advertise_mss(struct sock *sk)
102 struct tcp_sock *tp = tcp_sk(sk);
103 struct dst_entry *dst = __sk_dst_get(sk);
104 int mss = tp->advmss;
106 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
107 mss = dst_metric(dst, RTAX_ADVMSS);
114 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
115 * This is the first part of cwnd validation mechanism. */
116 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
118 struct tcp_sock *tp = tcp_sk(sk);
119 s32 delta = tcp_time_stamp - tp->lsndtime;
120 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
121 u32 cwnd = tp->snd_cwnd;
123 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
125 tp->snd_ssthresh = tcp_current_ssthresh(sk);
126 restart_cwnd = min(restart_cwnd, cwnd);
128 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
130 tp->snd_cwnd = max(cwnd, restart_cwnd);
131 tp->snd_cwnd_stamp = tcp_time_stamp;
132 tp->snd_cwnd_used = 0;
135 static void tcp_event_data_sent(struct tcp_sock *tp,
136 struct sk_buff *skb, struct sock *sk)
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const u32 now = tcp_time_stamp;
141 if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)
142 tcp_cwnd_restart(sk, __sk_dst_get(sk));
146 /* If it is a reply for ato after last received
147 * packet, enter pingpong mode.
149 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
150 icsk->icsk_ack.pingpong = 1;
153 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
155 tcp_dec_quickack_mode(sk, pkts);
156 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
159 /* Determine a window scaling and initial window to offer.
160 * Based on the assumption that the given amount of space
161 * will be offered. Store the results in the tp structure.
162 * NOTE: for smooth operation initial space offering should
163 * be a multiple of mss if possible. We assume here that mss >= 1.
164 * This MUST be enforced by all callers.
166 void tcp_select_initial_window(int __space, __u32 mss,
167 __u32 *rcv_wnd, __u32 *window_clamp,
168 int wscale_ok, __u8 *rcv_wscale)
170 unsigned int space = (__space < 0 ? 0 : __space);
172 /* If no clamp set the clamp to the max possible scaled window */
173 if (*window_clamp == 0)
174 (*window_clamp) = (65535 << 14);
175 space = min(*window_clamp, space);
177 /* Quantize space offering to a multiple of mss if possible. */
179 space = (space / mss) * mss;
181 /* NOTE: offering an initial window larger than 32767
182 * will break some buggy TCP stacks. If the admin tells us
183 * it is likely we could be speaking with such a buggy stack
184 * we will truncate our initial window offering to 32K-1
185 * unless the remote has sent us a window scaling option,
186 * which we interpret as a sign the remote TCP is not
187 * misinterpreting the window field as a signed quantity.
189 if (sysctl_tcp_workaround_signed_windows)
190 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
196 /* Set window scaling on max possible window
197 * See RFC1323 for an explanation of the limit to 14
199 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
200 while (space > 65535 && (*rcv_wscale) < 14) {
206 /* Set initial window to value enough for senders,
207 * following RFC2414. Senders, not following this RFC,
208 * will be satisfied with 2.
210 if (mss > (1<<*rcv_wscale)) {
216 if (*rcv_wnd > init_cwnd*mss)
217 *rcv_wnd = init_cwnd*mss;
220 /* Set the clamp no higher than max representable value */
221 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
224 /* Chose a new window to advertise, update state in tcp_sock for the
225 * socket, and return result with RFC1323 scaling applied. The return
226 * value can be stuffed directly into th->window for an outgoing
229 static u16 tcp_select_window(struct sock *sk)
231 struct tcp_sock *tp = tcp_sk(sk);
232 u32 cur_win = tcp_receive_window(tp);
233 u32 new_win = __tcp_select_window(sk);
235 /* Never shrink the offered window */
236 if(new_win < cur_win) {
237 /* Danger Will Robinson!
238 * Don't update rcv_wup/rcv_wnd here or else
239 * we will not be able to advertise a zero
240 * window in time. --DaveM
242 * Relax Will Robinson.
246 tp->rcv_wnd = new_win;
247 tp->rcv_wup = tp->rcv_nxt;
249 /* Make sure we do not exceed the maximum possible
252 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
253 new_win = min(new_win, MAX_TCP_WINDOW);
255 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
257 /* RFC1323 scaling applied */
258 new_win >>= tp->rx_opt.rcv_wscale;
260 /* If we advertise zero window, disable fast path. */
267 static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp,
270 if (tp->rx_opt.tstamp_ok) {
271 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
273 (TCPOPT_TIMESTAMP << 8) |
275 *ptr++ = htonl(tstamp);
276 *ptr++ = htonl(tp->rx_opt.ts_recent);
278 if (tp->rx_opt.eff_sacks) {
279 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
282 *ptr++ = htonl((TCPOPT_NOP << 24) |
285 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
286 TCPOLEN_SACK_PERBLOCK)));
287 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
288 *ptr++ = htonl(sp[this_sack].start_seq);
289 *ptr++ = htonl(sp[this_sack].end_seq);
291 if (tp->rx_opt.dsack) {
292 tp->rx_opt.dsack = 0;
293 tp->rx_opt.eff_sacks--;
298 /* Construct a tcp options header for a SYN or SYN_ACK packet.
299 * If this is every changed make sure to change the definition of
300 * MAX_SYN_SIZE to match the new maximum number of options that you
303 static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
304 int offer_wscale, int wscale, __u32 tstamp,
307 /* We always get an MSS option.
308 * The option bytes which will be seen in normal data
309 * packets should timestamps be used, must be in the MSS
310 * advertised. But we subtract them from tp->mss_cache so
311 * that calculations in tcp_sendmsg are simpler etc.
312 * So account for this fact here if necessary. If we
313 * don't do this correctly, as a receiver we won't
314 * recognize data packets as being full sized when we
315 * should, and thus we won't abide by the delayed ACK
317 * SACKs don't matter, we never delay an ACK when we
318 * have any of those going out.
320 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
323 *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
324 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
326 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
327 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
328 *ptr++ = htonl(tstamp); /* TSVAL */
329 *ptr++ = htonl(ts_recent); /* TSECR */
331 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
332 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
334 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
337 /* This routine actually transmits TCP packets queued in by
338 * tcp_do_sendmsg(). This is used by both the initial
339 * transmission and possible later retransmissions.
340 * All SKB's seen here are completely headerless. It is our
341 * job to build the TCP header, and pass the packet down to
342 * IP so it can do the same plus pass the packet off to the
345 * We are working here with either a clone of the original
346 * SKB, or a fresh unique copy made by the retransmit engine.
348 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
350 const struct inet_connection_sock *icsk = inet_csk(sk);
351 struct inet_sock *inet;
353 struct tcp_skb_cb *tcb;
359 BUG_ON(!skb || !tcp_skb_pcount(skb));
361 /* If congestion control is doing timestamping, we must
362 * take such a timestamp before we potentially clone/copy.
364 if (icsk->icsk_ca_ops->rtt_sample)
365 __net_timestamp(skb);
367 if (likely(clone_it)) {
368 if (unlikely(skb_cloned(skb)))
369 skb = pskb_copy(skb, gfp_mask);
371 skb = skb_clone(skb, gfp_mask);
378 tcb = TCP_SKB_CB(skb);
379 tcp_header_size = tp->tcp_header_len;
381 #define SYSCTL_FLAG_TSTAMPS 0x1
382 #define SYSCTL_FLAG_WSCALE 0x2
383 #define SYSCTL_FLAG_SACK 0x4
386 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
387 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
388 if(sysctl_tcp_timestamps) {
389 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
390 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
392 if (sysctl_tcp_window_scaling) {
393 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
394 sysctl_flags |= SYSCTL_FLAG_WSCALE;
396 if (sysctl_tcp_sack) {
397 sysctl_flags |= SYSCTL_FLAG_SACK;
398 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
399 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
401 } else if (unlikely(tp->rx_opt.eff_sacks)) {
402 /* A SACK is 2 pad bytes, a 2 byte header, plus
403 * 2 32-bit sequence numbers for each SACK block.
405 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
406 (tp->rx_opt.eff_sacks *
407 TCPOLEN_SACK_PERBLOCK));
410 if (tcp_packets_in_flight(tp) == 0)
411 tcp_ca_event(sk, CA_EVENT_TX_START);
413 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
415 skb_set_owner_w(skb, sk);
417 /* Build TCP header and checksum it. */
418 th->source = inet->sport;
419 th->dest = inet->dport;
420 th->seq = htonl(tcb->seq);
421 th->ack_seq = htonl(tp->rcv_nxt);
422 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
425 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
426 /* RFC1323: The window in SYN & SYN/ACK segments
429 th->window = htons(tp->rcv_wnd);
431 th->window = htons(tcp_select_window(sk));
436 if (unlikely(tp->urg_mode &&
437 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
438 th->urg_ptr = htons(tp->snd_up-tcb->seq);
442 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
443 tcp_syn_build_options((__u32 *)(th + 1),
444 tcp_advertise_mss(sk),
445 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
446 (sysctl_flags & SYSCTL_FLAG_SACK),
447 (sysctl_flags & SYSCTL_FLAG_WSCALE),
448 tp->rx_opt.rcv_wscale,
450 tp->rx_opt.ts_recent);
452 tcp_build_and_update_options((__u32 *)(th + 1),
454 TCP_ECN_send(sk, tp, skb, tcp_header_size);
457 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
459 if (likely(tcb->flags & TCPCB_FLAG_ACK))
460 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
462 if (skb->len != tcp_header_size)
463 tcp_event_data_sent(tp, skb, sk);
465 TCP_INC_STATS(TCP_MIB_OUTSEGS);
467 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
468 if (unlikely(err <= 0))
473 /* NET_XMIT_CN is special. It does not guarantee,
474 * that this packet is lost. It tells that device
475 * is about to start to drop packets or already
476 * drops some packets of the same priority and
477 * invokes us to send less aggressively.
479 return err == NET_XMIT_CN ? 0 : err;
481 #undef SYSCTL_FLAG_TSTAMPS
482 #undef SYSCTL_FLAG_WSCALE
483 #undef SYSCTL_FLAG_SACK
487 /* This routine just queue's the buffer
489 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
490 * otherwise socket can stall.
492 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
494 struct tcp_sock *tp = tcp_sk(sk);
496 /* Advance write_seq and place onto the write_queue. */
497 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
498 skb_header_release(skb);
499 __skb_queue_tail(&sk->sk_write_queue, skb);
500 sk_charge_skb(sk, skb);
502 /* Queue it, remembering where we must start sending. */
503 if (sk->sk_send_head == NULL)
504 sk->sk_send_head = skb;
507 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
509 if (skb->len <= mss_now ||
510 !(sk->sk_route_caps & NETIF_F_TSO)) {
511 /* Avoid the costly divide in the normal
514 skb_shinfo(skb)->tso_segs = 1;
515 skb_shinfo(skb)->tso_size = 0;
519 factor = skb->len + (mss_now - 1);
521 skb_shinfo(skb)->tso_segs = factor;
522 skb_shinfo(skb)->tso_size = mss_now;
526 /* Function to create two new TCP segments. Shrinks the given segment
527 * to the specified size and appends a new segment with the rest of the
528 * packet to the list. This won't be called frequently, I hope.
529 * Remember, these are still headerless SKBs at this point.
531 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
533 struct tcp_sock *tp = tcp_sk(sk);
534 struct sk_buff *buff;
535 int nsize, old_factor;
538 BUG_ON(len > skb->len);
540 clear_all_retrans_hints(tp);
541 nsize = skb_headlen(skb) - len;
545 if (skb_cloned(skb) &&
546 skb_is_nonlinear(skb) &&
547 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
550 /* Get a new skb... force flag on. */
551 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
553 return -ENOMEM; /* We'll just try again later. */
554 sk_charge_skb(sk, buff);
556 /* Correct the sequence numbers. */
557 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
558 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
559 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
561 /* PSH and FIN should only be set in the second packet. */
562 flags = TCP_SKB_CB(skb)->flags;
563 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
564 TCP_SKB_CB(buff)->flags = flags;
565 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
566 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
568 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
569 /* Copy and checksum data tail into the new buffer. */
570 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
575 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
577 skb->ip_summed = CHECKSUM_HW;
578 skb_split(skb, buff, len);
581 buff->ip_summed = skb->ip_summed;
583 /* Looks stupid, but our code really uses when of
584 * skbs, which it never sent before. --ANK
586 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
587 buff->tstamp = skb->tstamp;
589 old_factor = tcp_skb_pcount(skb);
591 /* Fix up tso_factor for both original and new SKB. */
592 tcp_set_skb_tso_segs(sk, skb, mss_now);
593 tcp_set_skb_tso_segs(sk, buff, mss_now);
595 /* If this packet has been sent out already, we must
596 * adjust the various packet counters.
598 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
599 int diff = old_factor - tcp_skb_pcount(skb) -
600 tcp_skb_pcount(buff);
602 tp->packets_out -= diff;
604 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
605 tp->sacked_out -= diff;
606 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
607 tp->retrans_out -= diff;
609 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
610 tp->lost_out -= diff;
611 tp->left_out -= diff;
615 /* Adjust Reno SACK estimate. */
616 if (!tp->rx_opt.sack_ok) {
617 tp->sacked_out -= diff;
618 if ((int)tp->sacked_out < 0)
620 tcp_sync_left_out(tp);
623 tp->fackets_out -= diff;
624 if ((int)tp->fackets_out < 0)
629 /* Link BUFF into the send queue. */
630 skb_header_release(buff);
631 __skb_append(skb, buff, &sk->sk_write_queue);
636 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
637 * eventually). The difference is that pulled data not copied, but
638 * immediately discarded.
640 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
646 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
647 if (skb_shinfo(skb)->frags[i].size <= eat) {
648 put_page(skb_shinfo(skb)->frags[i].page);
649 eat -= skb_shinfo(skb)->frags[i].size;
651 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
653 skb_shinfo(skb)->frags[k].page_offset += eat;
654 skb_shinfo(skb)->frags[k].size -= eat;
660 skb_shinfo(skb)->nr_frags = k;
662 skb->tail = skb->data;
663 skb->data_len -= len;
664 skb->len = skb->data_len;
668 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
670 if (skb_cloned(skb) &&
671 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
674 if (len <= skb_headlen(skb)) {
675 __skb_pull(skb, len);
677 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
681 TCP_SKB_CB(skb)->seq += len;
682 skb->ip_summed = CHECKSUM_HW;
684 skb->truesize -= len;
685 sk->sk_wmem_queued -= len;
686 sk->sk_forward_alloc += len;
687 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
689 /* Any change of skb->len requires recalculation of tso
692 if (tcp_skb_pcount(skb) > 1)
693 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
698 /* Not accounting for SACKs here. */
699 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
701 struct tcp_sock *tp = tcp_sk(sk);
702 struct inet_connection_sock *icsk = inet_csk(sk);
705 /* Calculate base mss without TCP options:
706 It is MMS_S - sizeof(tcphdr) of rfc1122
708 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
710 /* Clamp it (mss_clamp does not include tcp options) */
711 if (mss_now > tp->rx_opt.mss_clamp)
712 mss_now = tp->rx_opt.mss_clamp;
714 /* Now subtract optional transport overhead */
715 mss_now -= icsk->icsk_ext_hdr_len;
717 /* Then reserve room for full set of TCP options and 8 bytes of data */
721 /* Now subtract TCP options size, not including SACKs */
722 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
727 /* Inverse of above */
728 int tcp_mss_to_mtu(struct sock *sk, int mss)
730 struct tcp_sock *tp = tcp_sk(sk);
731 struct inet_connection_sock *icsk = inet_csk(sk);
736 icsk->icsk_ext_hdr_len +
737 icsk->icsk_af_ops->net_header_len;
742 void tcp_mtup_init(struct sock *sk)
744 struct tcp_sock *tp = tcp_sk(sk);
745 struct inet_connection_sock *icsk = inet_csk(sk);
747 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
748 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
749 icsk->icsk_af_ops->net_header_len;
750 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
751 icsk->icsk_mtup.probe_size = 0;
754 /* This function synchronize snd mss to current pmtu/exthdr set.
756 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
757 for TCP options, but includes only bare TCP header.
759 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
760 It is minimum of user_mss and mss received with SYN.
761 It also does not include TCP options.
763 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
765 tp->mss_cache is current effective sending mss, including
766 all tcp options except for SACKs. It is evaluated,
767 taking into account current pmtu, but never exceeds
768 tp->rx_opt.mss_clamp.
770 NOTE1. rfc1122 clearly states that advertised MSS
771 DOES NOT include either tcp or ip options.
773 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
774 are READ ONLY outside this function. --ANK (980731)
777 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
779 struct tcp_sock *tp = tcp_sk(sk);
780 struct inet_connection_sock *icsk = inet_csk(sk);
783 if (icsk->icsk_mtup.search_high > pmtu)
784 icsk->icsk_mtup.search_high = pmtu;
786 mss_now = tcp_mtu_to_mss(sk, pmtu);
788 /* Bound mss with half of window */
789 if (tp->max_window && mss_now > (tp->max_window>>1))
790 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
792 /* And store cached results */
793 icsk->icsk_pmtu_cookie = pmtu;
794 if (icsk->icsk_mtup.enabled)
795 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
796 tp->mss_cache = mss_now;
801 /* Compute the current effective MSS, taking SACKs and IP options,
802 * and even PMTU discovery events into account.
804 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
805 * cannot be large. However, taking into account rare use of URG, this
808 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
810 struct tcp_sock *tp = tcp_sk(sk);
811 struct dst_entry *dst = __sk_dst_get(sk);
816 mss_now = tp->mss_cache;
819 (sk->sk_route_caps & NETIF_F_TSO) &&
824 u32 mtu = dst_mtu(dst);
825 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
826 mss_now = tcp_sync_mss(sk, mtu);
829 if (tp->rx_opt.eff_sacks)
830 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
831 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
833 xmit_size_goal = mss_now;
836 xmit_size_goal = (65535 -
837 inet_csk(sk)->icsk_af_ops->net_header_len -
838 inet_csk(sk)->icsk_ext_hdr_len -
841 if (tp->max_window &&
842 (xmit_size_goal > (tp->max_window >> 1)))
843 xmit_size_goal = max((tp->max_window >> 1),
844 68U - tp->tcp_header_len);
846 xmit_size_goal -= (xmit_size_goal % mss_now);
848 tp->xmit_size_goal = xmit_size_goal;
853 /* Congestion window validation. (RFC2861) */
855 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
857 __u32 packets_out = tp->packets_out;
859 if (packets_out >= tp->snd_cwnd) {
860 /* Network is feed fully. */
861 tp->snd_cwnd_used = 0;
862 tp->snd_cwnd_stamp = tcp_time_stamp;
864 /* Network starves. */
865 if (tp->packets_out > tp->snd_cwnd_used)
866 tp->snd_cwnd_used = tp->packets_out;
868 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
869 tcp_cwnd_application_limited(sk);
873 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
875 u32 window, cwnd_len;
877 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
878 cwnd_len = mss_now * cwnd;
879 return min(window, cwnd_len);
882 /* Can at least one segment of SKB be sent right now, according to the
883 * congestion window rules? If so, return how many segments are allowed.
885 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
889 /* Don't be strict about the congestion window for the final FIN. */
890 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
893 in_flight = tcp_packets_in_flight(tp);
895 if (in_flight < cwnd)
896 return (cwnd - in_flight);
901 /* This must be invoked the first time we consider transmitting
904 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
906 int tso_segs = tcp_skb_pcount(skb);
910 skb_shinfo(skb)->tso_size != mss_now)) {
911 tcp_set_skb_tso_segs(sk, skb, mss_now);
912 tso_segs = tcp_skb_pcount(skb);
917 static inline int tcp_minshall_check(const struct tcp_sock *tp)
919 return after(tp->snd_sml,tp->snd_una) &&
920 !after(tp->snd_sml, tp->snd_nxt);
923 /* Return 0, if packet can be sent now without violation Nagle's rules:
924 * 1. It is full sized.
925 * 2. Or it contains FIN. (already checked by caller)
926 * 3. Or TCP_NODELAY was set.
927 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
928 * With Minshall's modification: all sent small packets are ACKed.
931 static inline int tcp_nagle_check(const struct tcp_sock *tp,
932 const struct sk_buff *skb,
933 unsigned mss_now, int nonagle)
935 return (skb->len < mss_now &&
936 ((nonagle&TCP_NAGLE_CORK) ||
939 tcp_minshall_check(tp))));
942 /* Return non-zero if the Nagle test allows this packet to be
945 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
946 unsigned int cur_mss, int nonagle)
948 /* Nagle rule does not apply to frames, which sit in the middle of the
949 * write_queue (they have no chances to get new data).
951 * This is implemented in the callers, where they modify the 'nonagle'
952 * argument based upon the location of SKB in the send queue.
954 if (nonagle & TCP_NAGLE_PUSH)
957 /* Don't use the nagle rule for urgent data (or for the final FIN). */
959 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
962 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
968 /* Does at least the first segment of SKB fit into the send window? */
969 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
971 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
973 if (skb->len > cur_mss)
974 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
976 return !after(end_seq, tp->snd_una + tp->snd_wnd);
979 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
980 * should be put on the wire right now. If so, it returns the number of
981 * packets allowed by the congestion window.
983 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
984 unsigned int cur_mss, int nonagle)
986 struct tcp_sock *tp = tcp_sk(sk);
987 unsigned int cwnd_quota;
989 tcp_init_tso_segs(sk, skb, cur_mss);
991 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
994 cwnd_quota = tcp_cwnd_test(tp, skb);
996 !tcp_snd_wnd_test(tp, skb, cur_mss))
1002 static inline int tcp_skb_is_last(const struct sock *sk,
1003 const struct sk_buff *skb)
1005 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1008 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1010 struct sk_buff *skb = sk->sk_send_head;
1013 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1014 (tcp_skb_is_last(sk, skb) ?
1019 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1020 * which is put after SKB on the list. It is very much like
1021 * tcp_fragment() except that it may make several kinds of assumptions
1022 * in order to speed up the splitting operation. In particular, we
1023 * know that all the data is in scatter-gather pages, and that the
1024 * packet has never been sent out before (and thus is not cloned).
1026 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1028 struct sk_buff *buff;
1029 int nlen = skb->len - len;
1032 /* All of a TSO frame must be composed of paged data. */
1033 if (skb->len != skb->data_len)
1034 return tcp_fragment(sk, skb, len, mss_now);
1036 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1037 if (unlikely(buff == NULL))
1040 buff->truesize = nlen;
1041 skb->truesize -= nlen;
1043 /* Correct the sequence numbers. */
1044 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1045 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1046 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1048 /* PSH and FIN should only be set in the second packet. */
1049 flags = TCP_SKB_CB(skb)->flags;
1050 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1051 TCP_SKB_CB(buff)->flags = flags;
1053 /* This packet was never sent out yet, so no SACK bits. */
1054 TCP_SKB_CB(buff)->sacked = 0;
1056 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
1057 skb_split(skb, buff, len);
1059 /* Fix up tso_factor for both original and new SKB. */
1060 tcp_set_skb_tso_segs(sk, skb, mss_now);
1061 tcp_set_skb_tso_segs(sk, buff, mss_now);
1063 /* Link BUFF into the send queue. */
1064 skb_header_release(buff);
1065 __skb_append(skb, buff, &sk->sk_write_queue);
1070 /* Try to defer sending, if possible, in order to minimize the amount
1071 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1073 * This algorithm is from John Heffner.
1075 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1077 const struct inet_connection_sock *icsk = inet_csk(sk);
1078 u32 send_win, cong_win, limit, in_flight;
1080 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1083 if (icsk->icsk_ca_state != TCP_CA_Open)
1086 in_flight = tcp_packets_in_flight(tp);
1088 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1089 (tp->snd_cwnd <= in_flight));
1091 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1093 /* From in_flight test above, we know that cwnd > in_flight. */
1094 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1096 limit = min(send_win, cong_win);
1098 /* If a full-sized TSO skb can be sent, do it. */
1102 if (sysctl_tcp_tso_win_divisor) {
1103 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1105 /* If at least some fraction of a window is available,
1108 chunk /= sysctl_tcp_tso_win_divisor;
1112 /* Different approach, try not to defer past a single
1113 * ACK. Receiver should ACK every other full sized
1114 * frame, so if we have space for more than 3 frames
1117 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1121 /* Ok, it looks like it is advisable to defer. */
1125 /* Create a new MTU probe if we are ready.
1126 * Returns 0 if we should wait to probe (no cwnd available),
1127 * 1 if a probe was sent,
1129 static int tcp_mtu_probe(struct sock *sk)
1131 struct tcp_sock *tp = tcp_sk(sk);
1132 struct inet_connection_sock *icsk = inet_csk(sk);
1133 struct sk_buff *skb, *nskb, *next;
1140 /* Not currently probing/verifying,
1142 * have enough cwnd, and
1143 * not SACKing (the variable headers throw things off) */
1144 if (!icsk->icsk_mtup.enabled ||
1145 icsk->icsk_mtup.probe_size ||
1146 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1147 tp->snd_cwnd < 11 ||
1148 tp->rx_opt.eff_sacks)
1151 /* Very simple search strategy: just double the MSS. */
1152 mss_now = tcp_current_mss(sk, 0);
1153 probe_size = 2*tp->mss_cache;
1154 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1155 /* TODO: set timer for probe_converge_event */
1159 /* Have enough data in the send queue to probe? */
1161 if ((skb = sk->sk_send_head) == NULL)
1163 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1165 if (len < probe_size)
1168 /* Receive window check. */
1169 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1170 if (tp->snd_wnd < probe_size)
1176 /* Do we need to wait to drain cwnd? */
1177 pif = tcp_packets_in_flight(tp);
1178 if (pif + 2 > tp->snd_cwnd) {
1179 /* With no packets in flight, don't stall. */
1186 /* We're allowed to probe. Build it now. */
1187 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1189 sk_charge_skb(sk, nskb);
1191 skb = sk->sk_send_head;
1192 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1193 sk->sk_send_head = nskb;
1195 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1196 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1197 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1198 TCP_SKB_CB(nskb)->sacked = 0;
1200 if (skb->ip_summed == CHECKSUM_HW)
1201 nskb->ip_summed = CHECKSUM_HW;
1204 while (len < probe_size) {
1207 copy = min_t(int, skb->len, probe_size - len);
1208 if (nskb->ip_summed)
1209 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1211 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1212 skb_put(nskb, copy), copy, nskb->csum);
1214 if (skb->len <= copy) {
1215 /* We've eaten all the data from this skb.
1217 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1218 __skb_unlink(skb, &sk->sk_write_queue);
1219 sk_stream_free_skb(sk, skb);
1221 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1222 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1223 if (!skb_shinfo(skb)->nr_frags) {
1224 skb_pull(skb, copy);
1225 if (skb->ip_summed != CHECKSUM_HW)
1226 skb->csum = csum_partial(skb->data, skb->len, 0);
1228 __pskb_trim_head(skb, copy);
1229 tcp_set_skb_tso_segs(sk, skb, mss_now);
1231 TCP_SKB_CB(skb)->seq += copy;
1237 tcp_init_tso_segs(sk, nskb, nskb->len);
1239 /* We're ready to send. If this fails, the probe will
1240 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1241 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1242 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1243 /* Decrement cwnd here because we are sending
1244 * effectively two packets. */
1246 update_send_head(sk, tp, nskb);
1248 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1249 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1250 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1259 /* This routine writes packets to the network. It advances the
1260 * send_head. This happens as incoming acks open up the remote
1263 * Returns 1, if no segments are in flight and we have queued segments, but
1264 * cannot send anything now because of SWS or another problem.
1266 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1268 struct tcp_sock *tp = tcp_sk(sk);
1269 struct sk_buff *skb;
1270 unsigned int tso_segs, sent_pkts;
1274 /* If we are closed, the bytes will have to remain here.
1275 * In time closedown will finish, we empty the write queue and all
1278 if (unlikely(sk->sk_state == TCP_CLOSE))
1283 /* Do MTU probing. */
1284 if ((result = tcp_mtu_probe(sk)) == 0) {
1286 } else if (result > 0) {
1290 while ((skb = sk->sk_send_head)) {
1293 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1296 cwnd_quota = tcp_cwnd_test(tp, skb);
1300 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1303 if (tso_segs == 1) {
1304 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1305 (tcp_skb_is_last(sk, skb) ?
1306 nonagle : TCP_NAGLE_PUSH))))
1309 if (tcp_tso_should_defer(sk, tp, skb))
1315 limit = tcp_window_allows(tp, skb,
1316 mss_now, cwnd_quota);
1318 if (skb->len < limit) {
1319 unsigned int trim = skb->len % mss_now;
1322 limit = skb->len - trim;
1326 if (skb->len > limit &&
1327 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1330 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1332 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1335 /* Advance the send_head. This one is sent out.
1336 * This call will increment packets_out.
1338 update_send_head(sk, tp, skb);
1340 tcp_minshall_update(tp, mss_now, skb);
1344 if (likely(sent_pkts)) {
1345 tcp_cwnd_validate(sk, tp);
1348 return !tp->packets_out && sk->sk_send_head;
1351 /* Push out any pending frames which were held back due to
1352 * TCP_CORK or attempt at coalescing tiny packets.
1353 * The socket must be locked by the caller.
1355 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1356 unsigned int cur_mss, int nonagle)
1358 struct sk_buff *skb = sk->sk_send_head;
1361 if (tcp_write_xmit(sk, cur_mss, nonagle))
1362 tcp_check_probe_timer(sk, tp);
1366 /* Send _single_ skb sitting at the send head. This function requires
1367 * true push pending frames to setup probe timer etc.
1369 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1371 struct tcp_sock *tp = tcp_sk(sk);
1372 struct sk_buff *skb = sk->sk_send_head;
1373 unsigned int tso_segs, cwnd_quota;
1375 BUG_ON(!skb || skb->len < mss_now);
1377 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1378 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1380 if (likely(cwnd_quota)) {
1387 limit = tcp_window_allows(tp, skb,
1388 mss_now, cwnd_quota);
1390 if (skb->len < limit) {
1391 unsigned int trim = skb->len % mss_now;
1394 limit = skb->len - trim;
1398 if (skb->len > limit &&
1399 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1402 /* Send it out now. */
1403 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1405 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1406 update_send_head(sk, tp, skb);
1407 tcp_cwnd_validate(sk, tp);
1413 /* This function returns the amount that we can raise the
1414 * usable window based on the following constraints
1416 * 1. The window can never be shrunk once it is offered (RFC 793)
1417 * 2. We limit memory per socket
1420 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1421 * RECV.NEXT + RCV.WIN fixed until:
1422 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1424 * i.e. don't raise the right edge of the window until you can raise
1425 * it at least MSS bytes.
1427 * Unfortunately, the recommended algorithm breaks header prediction,
1428 * since header prediction assumes th->window stays fixed.
1430 * Strictly speaking, keeping th->window fixed violates the receiver
1431 * side SWS prevention criteria. The problem is that under this rule
1432 * a stream of single byte packets will cause the right side of the
1433 * window to always advance by a single byte.
1435 * Of course, if the sender implements sender side SWS prevention
1436 * then this will not be a problem.
1438 * BSD seems to make the following compromise:
1440 * If the free space is less than the 1/4 of the maximum
1441 * space available and the free space is less than 1/2 mss,
1442 * then set the window to 0.
1443 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1444 * Otherwise, just prevent the window from shrinking
1445 * and from being larger than the largest representable value.
1447 * This prevents incremental opening of the window in the regime
1448 * where TCP is limited by the speed of the reader side taking
1449 * data out of the TCP receive queue. It does nothing about
1450 * those cases where the window is constrained on the sender side
1451 * because the pipeline is full.
1453 * BSD also seems to "accidentally" limit itself to windows that are a
1454 * multiple of MSS, at least until the free space gets quite small.
1455 * This would appear to be a side effect of the mbuf implementation.
1456 * Combining these two algorithms results in the observed behavior
1457 * of having a fixed window size at almost all times.
1459 * Below we obtain similar behavior by forcing the offered window to
1460 * a multiple of the mss when it is feasible to do so.
1462 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1463 * Regular options like TIMESTAMP are taken into account.
1465 u32 __tcp_select_window(struct sock *sk)
1467 struct inet_connection_sock *icsk = inet_csk(sk);
1468 struct tcp_sock *tp = tcp_sk(sk);
1469 /* MSS for the peer's data. Previous versions used mss_clamp
1470 * here. I don't know if the value based on our guesses
1471 * of peer's MSS is better for the performance. It's more correct
1472 * but may be worse for the performance because of rcv_mss
1473 * fluctuations. --SAW 1998/11/1
1475 int mss = icsk->icsk_ack.rcv_mss;
1476 int free_space = tcp_space(sk);
1477 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1480 if (mss > full_space)
1483 if (free_space < full_space/2) {
1484 icsk->icsk_ack.quick = 0;
1486 if (tcp_memory_pressure)
1487 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1489 if (free_space < mss)
1493 if (free_space > tp->rcv_ssthresh)
1494 free_space = tp->rcv_ssthresh;
1496 /* Don't do rounding if we are using window scaling, since the
1497 * scaled window will not line up with the MSS boundary anyway.
1499 window = tp->rcv_wnd;
1500 if (tp->rx_opt.rcv_wscale) {
1501 window = free_space;
1503 /* Advertise enough space so that it won't get scaled away.
1504 * Import case: prevent zero window announcement if
1505 * 1<<rcv_wscale > mss.
1507 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1508 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1509 << tp->rx_opt.rcv_wscale);
1511 /* Get the largest window that is a nice multiple of mss.
1512 * Window clamp already applied above.
1513 * If our current window offering is within 1 mss of the
1514 * free space we just keep it. This prevents the divide
1515 * and multiply from happening most of the time.
1516 * We also don't do any window rounding when the free space
1519 if (window <= free_space - mss || window > free_space)
1520 window = (free_space/mss)*mss;
1526 /* Attempt to collapse two adjacent SKB's during retransmission. */
1527 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1529 struct tcp_sock *tp = tcp_sk(sk);
1530 struct sk_buff *next_skb = skb->next;
1532 /* The first test we must make is that neither of these two
1533 * SKB's are still referenced by someone else.
1535 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1536 int skb_size = skb->len, next_skb_size = next_skb->len;
1537 u16 flags = TCP_SKB_CB(skb)->flags;
1539 /* Also punt if next skb has been SACK'd. */
1540 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1543 /* Next skb is out of window. */
1544 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1547 /* Punt if not enough space exists in the first SKB for
1548 * the data in the second, or the total combined payload
1549 * would exceed the MSS.
1551 if ((next_skb_size > skb_tailroom(skb)) ||
1552 ((skb_size + next_skb_size) > mss_now))
1555 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1556 tcp_skb_pcount(next_skb) != 1);
1558 /* changing transmit queue under us so clear hints */
1559 clear_all_retrans_hints(tp);
1561 /* Ok. We will be able to collapse the packet. */
1562 __skb_unlink(next_skb, &sk->sk_write_queue);
1564 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1566 if (next_skb->ip_summed == CHECKSUM_HW)
1567 skb->ip_summed = CHECKSUM_HW;
1569 if (skb->ip_summed != CHECKSUM_HW)
1570 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1572 /* Update sequence range on original skb. */
1573 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1575 /* Merge over control information. */
1576 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1577 TCP_SKB_CB(skb)->flags = flags;
1579 /* All done, get rid of second SKB and account for it so
1580 * packet counting does not break.
1582 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1583 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1584 tp->retrans_out -= tcp_skb_pcount(next_skb);
1585 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1586 tp->lost_out -= tcp_skb_pcount(next_skb);
1587 tp->left_out -= tcp_skb_pcount(next_skb);
1589 /* Reno case is special. Sigh... */
1590 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1591 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1592 tp->left_out -= tcp_skb_pcount(next_skb);
1595 /* Not quite right: it can be > snd.fack, but
1596 * it is better to underestimate fackets.
1598 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1599 tcp_packets_out_dec(tp, next_skb);
1600 sk_stream_free_skb(sk, next_skb);
1604 /* Do a simple retransmit without using the backoff mechanisms in
1605 * tcp_timer. This is used for path mtu discovery.
1606 * The socket is already locked here.
1608 void tcp_simple_retransmit(struct sock *sk)
1610 const struct inet_connection_sock *icsk = inet_csk(sk);
1611 struct tcp_sock *tp = tcp_sk(sk);
1612 struct sk_buff *skb;
1613 unsigned int mss = tcp_current_mss(sk, 0);
1616 sk_stream_for_retrans_queue(skb, sk) {
1617 if (skb->len > mss &&
1618 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1619 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1620 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1621 tp->retrans_out -= tcp_skb_pcount(skb);
1623 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1624 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1625 tp->lost_out += tcp_skb_pcount(skb);
1631 clear_all_retrans_hints(tp);
1636 tcp_sync_left_out(tp);
1638 /* Don't muck with the congestion window here.
1639 * Reason is that we do not increase amount of _data_
1640 * in network, but units changed and effective
1641 * cwnd/ssthresh really reduced now.
1643 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1644 tp->high_seq = tp->snd_nxt;
1645 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1646 tp->prior_ssthresh = 0;
1647 tp->undo_marker = 0;
1648 tcp_set_ca_state(sk, TCP_CA_Loss);
1650 tcp_xmit_retransmit_queue(sk);
1653 /* This retransmits one SKB. Policy decisions and retransmit queue
1654 * state updates are done by the caller. Returns non-zero if an
1655 * error occurred which prevented the send.
1657 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1659 struct tcp_sock *tp = tcp_sk(sk);
1660 struct inet_connection_sock *icsk = inet_csk(sk);
1661 unsigned int cur_mss = tcp_current_mss(sk, 0);
1664 /* Inconslusive MTU probe */
1665 if (icsk->icsk_mtup.probe_size) {
1666 icsk->icsk_mtup.probe_size = 0;
1669 /* Do not sent more than we queued. 1/4 is reserved for possible
1670 * copying overhead: fragmentation, tunneling, mangling etc.
1672 if (atomic_read(&sk->sk_wmem_alloc) >
1673 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1676 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1677 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1679 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1683 /* If receiver has shrunk his window, and skb is out of
1684 * new window, do not retransmit it. The exception is the
1685 * case, when window is shrunk to zero. In this case
1686 * our retransmit serves as a zero window probe.
1688 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1689 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1692 if (skb->len > cur_mss) {
1693 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1694 return -ENOMEM; /* We'll try again later. */
1697 /* Collapse two adjacent packets if worthwhile and we can. */
1698 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1699 (skb->len < (cur_mss >> 1)) &&
1700 (skb->next != sk->sk_send_head) &&
1701 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1702 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1703 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1704 (sysctl_tcp_retrans_collapse != 0))
1705 tcp_retrans_try_collapse(sk, skb, cur_mss);
1707 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1708 return -EHOSTUNREACH; /* Routing failure or similar. */
1710 /* Some Solaris stacks overoptimize and ignore the FIN on a
1711 * retransmit when old data is attached. So strip it off
1712 * since it is cheap to do so and saves bytes on the network.
1715 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1716 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1717 if (!pskb_trim(skb, 0)) {
1718 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1719 skb_shinfo(skb)->tso_segs = 1;
1720 skb_shinfo(skb)->tso_size = 0;
1721 skb->ip_summed = CHECKSUM_NONE;
1726 /* Make a copy, if the first transmission SKB clone we made
1727 * is still in somebody's hands, else make a clone.
1729 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1731 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1734 /* Update global TCP statistics. */
1735 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1737 tp->total_retrans++;
1739 #if FASTRETRANS_DEBUG > 0
1740 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1741 if (net_ratelimit())
1742 printk(KERN_DEBUG "retrans_out leaked.\n");
1745 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1746 tp->retrans_out += tcp_skb_pcount(skb);
1748 /* Save stamp of the first retransmit. */
1749 if (!tp->retrans_stamp)
1750 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1754 /* snd_nxt is stored to detect loss of retransmitted segment,
1755 * see tcp_input.c tcp_sacktag_write_queue().
1757 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1762 /* This gets called after a retransmit timeout, and the initially
1763 * retransmitted data is acknowledged. It tries to continue
1764 * resending the rest of the retransmit queue, until either
1765 * we've sent it all or the congestion window limit is reached.
1766 * If doing SACK, the first ACK which comes back for a timeout
1767 * based retransmit packet might feed us FACK information again.
1768 * If so, we use it to avoid unnecessarily retransmissions.
1770 void tcp_xmit_retransmit_queue(struct sock *sk)
1772 const struct inet_connection_sock *icsk = inet_csk(sk);
1773 struct tcp_sock *tp = tcp_sk(sk);
1774 struct sk_buff *skb;
1777 if (tp->retransmit_skb_hint) {
1778 skb = tp->retransmit_skb_hint;
1779 packet_cnt = tp->retransmit_cnt_hint;
1781 skb = sk->sk_write_queue.next;
1785 /* First pass: retransmit lost packets. */
1787 sk_stream_for_retrans_queue_from(skb, sk) {
1788 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1790 /* we could do better than to assign each time */
1791 tp->retransmit_skb_hint = skb;
1792 tp->retransmit_cnt_hint = packet_cnt;
1794 /* Assume this retransmit will generate
1795 * only one packet for congestion window
1796 * calculation purposes. This works because
1797 * tcp_retransmit_skb() will chop up the
1798 * packet to be MSS sized and all the
1799 * packet counting works out.
1801 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1804 if (sacked & TCPCB_LOST) {
1805 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1806 if (tcp_retransmit_skb(sk, skb)) {
1807 tp->retransmit_skb_hint = NULL;
1810 if (icsk->icsk_ca_state != TCP_CA_Loss)
1811 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1813 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1816 skb_peek(&sk->sk_write_queue))
1817 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1818 inet_csk(sk)->icsk_rto,
1822 packet_cnt += tcp_skb_pcount(skb);
1823 if (packet_cnt >= tp->lost_out)
1829 /* OK, demanded retransmission is finished. */
1831 /* Forward retransmissions are possible only during Recovery. */
1832 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1835 /* No forward retransmissions in Reno are possible. */
1836 if (!tp->rx_opt.sack_ok)
1839 /* Yeah, we have to make difficult choice between forward transmission
1840 * and retransmission... Both ways have their merits...
1842 * For now we do not retransmit anything, while we have some new
1846 if (tcp_may_send_now(sk, tp))
1849 if (tp->forward_skb_hint) {
1850 skb = tp->forward_skb_hint;
1851 packet_cnt = tp->forward_cnt_hint;
1853 skb = sk->sk_write_queue.next;
1857 sk_stream_for_retrans_queue_from(skb, sk) {
1858 tp->forward_cnt_hint = packet_cnt;
1859 tp->forward_skb_hint = skb;
1861 /* Similar to the retransmit loop above we
1862 * can pretend that the retransmitted SKB
1863 * we send out here will be composed of one
1864 * real MSS sized packet because tcp_retransmit_skb()
1865 * will fragment it if necessary.
1867 if (++packet_cnt > tp->fackets_out)
1870 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1873 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1876 /* Ok, retransmit it. */
1877 if (tcp_retransmit_skb(sk, skb)) {
1878 tp->forward_skb_hint = NULL;
1882 if (skb == skb_peek(&sk->sk_write_queue))
1883 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1884 inet_csk(sk)->icsk_rto,
1887 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1892 /* Send a fin. The caller locks the socket for us. This cannot be
1893 * allowed to fail queueing a FIN frame under any circumstances.
1895 void tcp_send_fin(struct sock *sk)
1897 struct tcp_sock *tp = tcp_sk(sk);
1898 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1901 /* Optimization, tack on the FIN if we have a queue of
1902 * unsent frames. But be careful about outgoing SACKS
1905 mss_now = tcp_current_mss(sk, 1);
1907 if (sk->sk_send_head != NULL) {
1908 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1909 TCP_SKB_CB(skb)->end_seq++;
1912 /* Socket is locked, keep trying until memory is available. */
1914 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1920 /* Reserve space for headers and prepare control bits. */
1921 skb_reserve(skb, MAX_TCP_HEADER);
1923 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1924 TCP_SKB_CB(skb)->sacked = 0;
1925 skb_shinfo(skb)->tso_segs = 1;
1926 skb_shinfo(skb)->tso_size = 0;
1928 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1929 TCP_SKB_CB(skb)->seq = tp->write_seq;
1930 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1931 tcp_queue_skb(sk, skb);
1933 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1936 /* We get here when a process closes a file descriptor (either due to
1937 * an explicit close() or as a byproduct of exit()'ing) and there
1938 * was unread data in the receive queue. This behavior is recommended
1939 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1941 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1943 struct tcp_sock *tp = tcp_sk(sk);
1944 struct sk_buff *skb;
1946 /* NOTE: No TCP options attached and we never retransmit this. */
1947 skb = alloc_skb(MAX_TCP_HEADER, priority);
1949 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1953 /* Reserve space for headers and prepare control bits. */
1954 skb_reserve(skb, MAX_TCP_HEADER);
1956 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1957 TCP_SKB_CB(skb)->sacked = 0;
1958 skb_shinfo(skb)->tso_segs = 1;
1959 skb_shinfo(skb)->tso_size = 0;
1962 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1963 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1964 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1965 if (tcp_transmit_skb(sk, skb, 0, priority))
1966 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1969 /* WARNING: This routine must only be called when we have already sent
1970 * a SYN packet that crossed the incoming SYN that caused this routine
1971 * to get called. If this assumption fails then the initial rcv_wnd
1972 * and rcv_wscale values will not be correct.
1974 int tcp_send_synack(struct sock *sk)
1976 struct sk_buff* skb;
1978 skb = skb_peek(&sk->sk_write_queue);
1979 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1980 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1983 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1984 if (skb_cloned(skb)) {
1985 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1988 __skb_unlink(skb, &sk->sk_write_queue);
1989 skb_header_release(nskb);
1990 __skb_queue_head(&sk->sk_write_queue, nskb);
1991 sk_stream_free_skb(sk, skb);
1992 sk_charge_skb(sk, nskb);
1996 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1997 TCP_ECN_send_synack(tcp_sk(sk), skb);
1999 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2000 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2004 * Prepare a SYN-ACK.
2006 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2007 struct request_sock *req)
2009 struct inet_request_sock *ireq = inet_rsk(req);
2010 struct tcp_sock *tp = tcp_sk(sk);
2012 int tcp_header_size;
2013 struct sk_buff *skb;
2015 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2019 /* Reserve space for headers. */
2020 skb_reserve(skb, MAX_TCP_HEADER);
2022 skb->dst = dst_clone(dst);
2024 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2025 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2026 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2027 /* SACK_PERM is in the place of NOP NOP of TS */
2028 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2029 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2031 memset(th, 0, sizeof(struct tcphdr));
2034 if (dst->dev->features&NETIF_F_TSO)
2036 TCP_ECN_make_synack(req, th);
2037 th->source = inet_sk(sk)->sport;
2038 th->dest = ireq->rmt_port;
2039 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2040 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2041 TCP_SKB_CB(skb)->sacked = 0;
2042 skb_shinfo(skb)->tso_segs = 1;
2043 skb_shinfo(skb)->tso_size = 0;
2044 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2045 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2046 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2048 /* Set this up on the first call only */
2049 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2050 /* tcp_full_space because it is guaranteed to be the first packet */
2051 tcp_select_initial_window(tcp_full_space(sk),
2052 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2057 ireq->rcv_wscale = rcv_wscale;
2060 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2061 th->window = htons(req->rcv_wnd);
2063 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2064 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2065 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2066 TCP_SKB_CB(skb)->when,
2070 th->doff = (tcp_header_size >> 2);
2071 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2076 * Do all connect socket setups that can be done AF independent.
2078 static void tcp_connect_init(struct sock *sk)
2080 struct dst_entry *dst = __sk_dst_get(sk);
2081 struct tcp_sock *tp = tcp_sk(sk);
2084 /* We'll fix this up when we get a response from the other end.
2085 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2087 tp->tcp_header_len = sizeof(struct tcphdr) +
2088 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2090 /* If user gave his TCP_MAXSEG, record it to clamp */
2091 if (tp->rx_opt.user_mss)
2092 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2095 tcp_sync_mss(sk, dst_mtu(dst));
2097 if (!tp->window_clamp)
2098 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2099 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2100 tcp_initialize_rcv_mss(sk);
2102 tcp_select_initial_window(tcp_full_space(sk),
2103 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2106 sysctl_tcp_window_scaling,
2109 tp->rx_opt.rcv_wscale = rcv_wscale;
2110 tp->rcv_ssthresh = tp->rcv_wnd;
2113 sock_reset_flag(sk, SOCK_DONE);
2115 tcp_init_wl(tp, tp->write_seq, 0);
2116 tp->snd_una = tp->write_seq;
2117 tp->snd_sml = tp->write_seq;
2122 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2123 inet_csk(sk)->icsk_retransmits = 0;
2124 tcp_clear_retrans(tp);
2128 * Build a SYN and send it off.
2130 int tcp_connect(struct sock *sk)
2132 struct tcp_sock *tp = tcp_sk(sk);
2133 struct sk_buff *buff;
2135 tcp_connect_init(sk);
2137 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2138 if (unlikely(buff == NULL))
2141 /* Reserve space for headers. */
2142 skb_reserve(buff, MAX_TCP_HEADER);
2144 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2145 TCP_ECN_send_syn(sk, tp, buff);
2146 TCP_SKB_CB(buff)->sacked = 0;
2147 skb_shinfo(buff)->tso_segs = 1;
2148 skb_shinfo(buff)->tso_size = 0;
2150 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2151 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2152 tp->snd_nxt = tp->write_seq;
2153 tp->pushed_seq = tp->write_seq;
2156 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2157 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2158 skb_header_release(buff);
2159 __skb_queue_tail(&sk->sk_write_queue, buff);
2160 sk_charge_skb(sk, buff);
2161 tp->packets_out += tcp_skb_pcount(buff);
2162 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2163 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2165 /* Timer for repeating the SYN until an answer. */
2166 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2167 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2171 /* Send out a delayed ack, the caller does the policy checking
2172 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2175 void tcp_send_delayed_ack(struct sock *sk)
2177 struct inet_connection_sock *icsk = inet_csk(sk);
2178 int ato = icsk->icsk_ack.ato;
2179 unsigned long timeout;
2181 if (ato > TCP_DELACK_MIN) {
2182 const struct tcp_sock *tp = tcp_sk(sk);
2185 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2186 max_ato = TCP_DELACK_MAX;
2188 /* Slow path, intersegment interval is "high". */
2190 /* If some rtt estimate is known, use it to bound delayed ack.
2191 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2195 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2201 ato = min(ato, max_ato);
2204 /* Stay within the limit we were given */
2205 timeout = jiffies + ato;
2207 /* Use new timeout only if there wasn't a older one earlier. */
2208 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2209 /* If delack timer was blocked or is about to expire,
2212 if (icsk->icsk_ack.blocked ||
2213 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2218 if (!time_before(timeout, icsk->icsk_ack.timeout))
2219 timeout = icsk->icsk_ack.timeout;
2221 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2222 icsk->icsk_ack.timeout = timeout;
2223 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2226 /* This routine sends an ack and also updates the window. */
2227 void tcp_send_ack(struct sock *sk)
2229 /* If we have been reset, we may not send again. */
2230 if (sk->sk_state != TCP_CLOSE) {
2231 struct tcp_sock *tp = tcp_sk(sk);
2232 struct sk_buff *buff;
2234 /* We are not putting this on the write queue, so
2235 * tcp_transmit_skb() will set the ownership to this
2238 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2240 inet_csk_schedule_ack(sk);
2241 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2242 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2243 TCP_DELACK_MAX, TCP_RTO_MAX);
2247 /* Reserve space for headers and prepare control bits. */
2248 skb_reserve(buff, MAX_TCP_HEADER);
2250 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2251 TCP_SKB_CB(buff)->sacked = 0;
2252 skb_shinfo(buff)->tso_segs = 1;
2253 skb_shinfo(buff)->tso_size = 0;
2255 /* Send it off, this clears delayed acks for us. */
2256 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2257 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2258 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2262 /* This routine sends a packet with an out of date sequence
2263 * number. It assumes the other end will try to ack it.
2265 * Question: what should we make while urgent mode?
2266 * 4.4BSD forces sending single byte of data. We cannot send
2267 * out of window data, because we have SND.NXT==SND.MAX...
2269 * Current solution: to send TWO zero-length segments in urgent mode:
2270 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2271 * out-of-date with SND.UNA-1 to probe window.
2273 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2275 struct tcp_sock *tp = tcp_sk(sk);
2276 struct sk_buff *skb;
2278 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2279 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2283 /* Reserve space for headers and set control bits. */
2284 skb_reserve(skb, MAX_TCP_HEADER);
2286 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2287 TCP_SKB_CB(skb)->sacked = urgent;
2288 skb_shinfo(skb)->tso_segs = 1;
2289 skb_shinfo(skb)->tso_size = 0;
2291 /* Use a previous sequence. This should cause the other
2292 * end to send an ack. Don't queue or clone SKB, just
2295 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2296 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2297 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2298 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2301 int tcp_write_wakeup(struct sock *sk)
2303 if (sk->sk_state != TCP_CLOSE) {
2304 struct tcp_sock *tp = tcp_sk(sk);
2305 struct sk_buff *skb;
2307 if ((skb = sk->sk_send_head) != NULL &&
2308 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2310 unsigned int mss = tcp_current_mss(sk, 0);
2311 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2313 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2314 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2316 /* We are probing the opening of a window
2317 * but the window size is != 0
2318 * must have been a result SWS avoidance ( sender )
2320 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2322 seg_size = min(seg_size, mss);
2323 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2324 if (tcp_fragment(sk, skb, seg_size, mss))
2326 } else if (!tcp_skb_pcount(skb))
2327 tcp_set_skb_tso_segs(sk, skb, mss);
2329 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2330 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2331 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2333 update_send_head(sk, tp, skb);
2338 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2339 tcp_xmit_probe_skb(sk, TCPCB_URG);
2340 return tcp_xmit_probe_skb(sk, 0);
2346 /* A window probe timeout has occurred. If window is not closed send
2347 * a partial packet else a zero probe.
2349 void tcp_send_probe0(struct sock *sk)
2351 struct inet_connection_sock *icsk = inet_csk(sk);
2352 struct tcp_sock *tp = tcp_sk(sk);
2355 err = tcp_write_wakeup(sk);
2357 if (tp->packets_out || !sk->sk_send_head) {
2358 /* Cancel probe timer, if it is not required. */
2359 icsk->icsk_probes_out = 0;
2360 icsk->icsk_backoff = 0;
2365 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2366 icsk->icsk_backoff++;
2367 icsk->icsk_probes_out++;
2368 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2369 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2372 /* If packet was not sent due to local congestion,
2373 * do not backoff and do not remember icsk_probes_out.
2374 * Let local senders to fight for local resources.
2376 * Use accumulated backoff yet.
2378 if (!icsk->icsk_probes_out)
2379 icsk->icsk_probes_out = 1;
2380 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2381 min(icsk->icsk_rto << icsk->icsk_backoff,
2382 TCP_RESOURCE_PROBE_INTERVAL),
2387 EXPORT_SYMBOL(tcp_connect);
2388 EXPORT_SYMBOL(tcp_make_synack);
2389 EXPORT_SYMBOL(tcp_simple_retransmit);
2390 EXPORT_SYMBOL(tcp_sync_mss);
2391 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2392 EXPORT_SYMBOL(tcp_mtup_init);