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 __read_mostly = 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 __read_mostly = 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 __read_mostly = 3;
59 int sysctl_tcp_mtu_probing __read_mostly = 0;
60 int sysctl_tcp_base_mss __read_mostly = 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
68 sk->sk_send_head = skb->next;
69 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
70 sk->sk_send_head = NULL;
71 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
72 tcp_packets_out_inc(sk, tp, skb);
75 /* SND.NXT, if window was not shrunk.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
83 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
86 return tp->snd_una+tp->snd_wnd;
89 /* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
103 static __u16 tcp_advertise_mss(struct sock *sk)
105 struct tcp_sock *tp = tcp_sk(sk);
106 struct dst_entry *dst = __sk_dst_get(sk);
107 int mss = tp->advmss;
109 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
110 mss = dst_metric(dst, RTAX_ADVMSS);
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
121 struct tcp_sock *tp = tcp_sk(sk);
122 s32 delta = tcp_time_stamp - tp->lsndtime;
123 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
124 u32 cwnd = tp->snd_cwnd;
126 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
128 tp->snd_ssthresh = tcp_current_ssthresh(sk);
129 restart_cwnd = min(restart_cwnd, cwnd);
131 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
133 tp->snd_cwnd = max(cwnd, restart_cwnd);
134 tp->snd_cwnd_stamp = tcp_time_stamp;
135 tp->snd_cwnd_used = 0;
138 static void tcp_event_data_sent(struct tcp_sock *tp,
139 struct sk_buff *skb, struct sock *sk)
141 struct inet_connection_sock *icsk = inet_csk(sk);
142 const u32 now = tcp_time_stamp;
144 if (sysctl_tcp_slow_start_after_idle &&
145 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
146 tcp_cwnd_restart(sk, __sk_dst_get(sk));
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
153 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
154 icsk->icsk_ack.pingpong = 1;
157 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
159 tcp_dec_quickack_mode(sk, pkts);
160 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
163 /* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
170 void tcp_select_initial_window(int __space, __u32 mss,
171 __u32 *rcv_wnd, __u32 *window_clamp,
172 int wscale_ok, __u8 *rcv_wscale)
174 unsigned int space = (__space < 0 ? 0 : __space);
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp == 0)
178 (*window_clamp) = (65535 << 14);
179 space = min(*window_clamp, space);
181 /* Quantize space offering to a multiple of mss if possible. */
183 space = (space / mss) * mss;
185 /* NOTE: offering an initial window larger than 32767
186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
193 if (sysctl_tcp_workaround_signed_windows)
194 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
203 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
204 space = min_t(u32, space, *window_clamp);
205 while (space > 65535 && (*rcv_wscale) < 14) {
211 /* Set initial window to value enough for senders,
212 * following RFC2414. Senders, not following this RFC,
213 * will be satisfied with 2.
215 if (mss > (1<<*rcv_wscale)) {
221 if (*rcv_wnd > init_cwnd*mss)
222 *rcv_wnd = init_cwnd*mss;
225 /* Set the clamp no higher than max representable value */
226 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
229 /* Chose a new window to advertise, update state in tcp_sock for the
230 * socket, and return result with RFC1323 scaling applied. The return
231 * value can be stuffed directly into th->window for an outgoing
234 static u16 tcp_select_window(struct sock *sk)
236 struct tcp_sock *tp = tcp_sk(sk);
237 u32 cur_win = tcp_receive_window(tp);
238 u32 new_win = __tcp_select_window(sk);
240 /* Never shrink the offered window */
241 if(new_win < cur_win) {
242 /* Danger Will Robinson!
243 * Don't update rcv_wup/rcv_wnd here or else
244 * we will not be able to advertise a zero
245 * window in time. --DaveM
247 * Relax Will Robinson.
251 tp->rcv_wnd = new_win;
252 tp->rcv_wup = tp->rcv_nxt;
254 /* Make sure we do not exceed the maximum possible
257 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
258 new_win = min(new_win, MAX_TCP_WINDOW);
260 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
262 /* RFC1323 scaling applied */
263 new_win >>= tp->rx_opt.rcv_wscale;
265 /* If we advertise zero window, disable fast path. */
272 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
275 if (tp->rx_opt.tstamp_ok) {
276 *ptr++ = htonl((TCPOPT_NOP << 24) |
278 (TCPOPT_TIMESTAMP << 8) |
280 *ptr++ = htonl(tstamp);
281 *ptr++ = htonl(tp->rx_opt.ts_recent);
283 if (tp->rx_opt.eff_sacks) {
284 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
287 *ptr++ = htonl((TCPOPT_NOP << 24) |
290 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
291 TCPOLEN_SACK_PERBLOCK)));
292 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
293 *ptr++ = htonl(sp[this_sack].start_seq);
294 *ptr++ = htonl(sp[this_sack].end_seq);
296 if (tp->rx_opt.dsack) {
297 tp->rx_opt.dsack = 0;
298 tp->rx_opt.eff_sacks--;
303 /* Construct a tcp options header for a SYN or SYN_ACK packet.
304 * If this is every changed make sure to change the definition of
305 * MAX_SYN_SIZE to match the new maximum number of options that you
308 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
309 int offer_wscale, int wscale, __u32 tstamp,
312 /* We always get an MSS option.
313 * The option bytes which will be seen in normal data
314 * packets should timestamps be used, must be in the MSS
315 * advertised. But we subtract them from tp->mss_cache so
316 * that calculations in tcp_sendmsg are simpler etc.
317 * So account for this fact here if necessary. If we
318 * don't do this correctly, as a receiver we won't
319 * recognize data packets as being full sized when we
320 * should, and thus we won't abide by the delayed ACK
322 * SACKs don't matter, we never delay an ACK when we
323 * have any of those going out.
325 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
328 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
329 (TCPOLEN_SACK_PERM << 16) |
330 (TCPOPT_TIMESTAMP << 8) |
333 *ptr++ = htonl((TCPOPT_NOP << 24) |
335 (TCPOPT_TIMESTAMP << 8) |
337 *ptr++ = htonl(tstamp); /* TSVAL */
338 *ptr++ = htonl(ts_recent); /* TSECR */
340 *ptr++ = htonl((TCPOPT_NOP << 24) |
342 (TCPOPT_SACK_PERM << 8) |
345 *ptr++ = htonl((TCPOPT_NOP << 24) |
346 (TCPOPT_WINDOW << 16) |
347 (TCPOLEN_WINDOW << 8) |
351 /* This routine actually transmits TCP packets queued in by
352 * tcp_do_sendmsg(). This is used by both the initial
353 * transmission and possible later retransmissions.
354 * All SKB's seen here are completely headerless. It is our
355 * job to build the TCP header, and pass the packet down to
356 * IP so it can do the same plus pass the packet off to the
359 * We are working here with either a clone of the original
360 * SKB, or a fresh unique copy made by the retransmit engine.
362 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
364 const struct inet_connection_sock *icsk = inet_csk(sk);
365 struct inet_sock *inet;
367 struct tcp_skb_cb *tcb;
373 BUG_ON(!skb || !tcp_skb_pcount(skb));
375 /* If congestion control is doing timestamping, we must
376 * take such a timestamp before we potentially clone/copy.
378 if (icsk->icsk_ca_ops->rtt_sample)
379 __net_timestamp(skb);
381 if (likely(clone_it)) {
382 if (unlikely(skb_cloned(skb)))
383 skb = pskb_copy(skb, gfp_mask);
385 skb = skb_clone(skb, gfp_mask);
392 tcb = TCP_SKB_CB(skb);
393 tcp_header_size = tp->tcp_header_len;
395 #define SYSCTL_FLAG_TSTAMPS 0x1
396 #define SYSCTL_FLAG_WSCALE 0x2
397 #define SYSCTL_FLAG_SACK 0x4
400 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
401 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
402 if(sysctl_tcp_timestamps) {
403 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
404 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
406 if (sysctl_tcp_window_scaling) {
407 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
408 sysctl_flags |= SYSCTL_FLAG_WSCALE;
410 if (sysctl_tcp_sack) {
411 sysctl_flags |= SYSCTL_FLAG_SACK;
412 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
413 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
415 } else if (unlikely(tp->rx_opt.eff_sacks)) {
416 /* A SACK is 2 pad bytes, a 2 byte header, plus
417 * 2 32-bit sequence numbers for each SACK block.
419 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
420 (tp->rx_opt.eff_sacks *
421 TCPOLEN_SACK_PERBLOCK));
424 if (tcp_packets_in_flight(tp) == 0)
425 tcp_ca_event(sk, CA_EVENT_TX_START);
427 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
429 skb_set_owner_w(skb, sk);
431 /* Build TCP header and checksum it. */
432 th->source = inet->sport;
433 th->dest = inet->dport;
434 th->seq = htonl(tcb->seq);
435 th->ack_seq = htonl(tp->rcv_nxt);
436 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
439 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
440 /* RFC1323: The window in SYN & SYN/ACK segments
443 th->window = htons(tp->rcv_wnd);
445 th->window = htons(tcp_select_window(sk));
450 if (unlikely(tp->urg_mode &&
451 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
452 th->urg_ptr = htons(tp->snd_up-tcb->seq);
456 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
457 tcp_syn_build_options((__be32 *)(th + 1),
458 tcp_advertise_mss(sk),
459 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
460 (sysctl_flags & SYSCTL_FLAG_SACK),
461 (sysctl_flags & SYSCTL_FLAG_WSCALE),
462 tp->rx_opt.rcv_wscale,
464 tp->rx_opt.ts_recent);
466 tcp_build_and_update_options((__be32 *)(th + 1),
468 TCP_ECN_send(sk, tp, skb, tcp_header_size);
471 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
473 if (likely(tcb->flags & TCPCB_FLAG_ACK))
474 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
476 if (skb->len != tcp_header_size)
477 tcp_event_data_sent(tp, skb, sk);
479 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
480 TCP_INC_STATS(TCP_MIB_OUTSEGS);
482 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
483 if (likely(err <= 0))
488 /* NET_XMIT_CN is special. It does not guarantee,
489 * that this packet is lost. It tells that device
490 * is about to start to drop packets or already
491 * drops some packets of the same priority and
492 * invokes us to send less aggressively.
494 return err == NET_XMIT_CN ? 0 : err;
496 #undef SYSCTL_FLAG_TSTAMPS
497 #undef SYSCTL_FLAG_WSCALE
498 #undef SYSCTL_FLAG_SACK
502 /* This routine just queue's the buffer
504 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
505 * otherwise socket can stall.
507 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
509 struct tcp_sock *tp = tcp_sk(sk);
511 /* Advance write_seq and place onto the write_queue. */
512 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
513 skb_header_release(skb);
514 __skb_queue_tail(&sk->sk_write_queue, skb);
515 sk_charge_skb(sk, skb);
517 /* Queue it, remembering where we must start sending. */
518 if (sk->sk_send_head == NULL)
519 sk->sk_send_head = skb;
522 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
524 if (skb->len <= mss_now || !sk_can_gso(sk)) {
525 /* Avoid the costly divide in the normal
528 skb_shinfo(skb)->gso_segs = 1;
529 skb_shinfo(skb)->gso_size = 0;
530 skb_shinfo(skb)->gso_type = 0;
534 factor = skb->len + (mss_now - 1);
536 skb_shinfo(skb)->gso_segs = factor;
537 skb_shinfo(skb)->gso_size = mss_now;
538 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
542 /* Function to create two new TCP segments. Shrinks the given segment
543 * to the specified size and appends a new segment with the rest of the
544 * packet to the list. This won't be called frequently, I hope.
545 * Remember, these are still headerless SKBs at this point.
547 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
549 struct tcp_sock *tp = tcp_sk(sk);
550 struct sk_buff *buff;
551 int nsize, old_factor;
555 BUG_ON(len > skb->len);
557 clear_all_retrans_hints(tp);
558 nsize = skb_headlen(skb) - len;
562 if (skb_cloned(skb) &&
563 skb_is_nonlinear(skb) &&
564 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
567 /* Get a new skb... force flag on. */
568 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
570 return -ENOMEM; /* We'll just try again later. */
572 sk_charge_skb(sk, buff);
573 nlen = skb->len - len - nsize;
574 buff->truesize += nlen;
575 skb->truesize -= nlen;
577 /* Correct the sequence numbers. */
578 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
579 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
580 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
582 /* PSH and FIN should only be set in the second packet. */
583 flags = TCP_SKB_CB(skb)->flags;
584 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
585 TCP_SKB_CB(buff)->flags = flags;
586 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
587 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
589 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
590 /* Copy and checksum data tail into the new buffer. */
591 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
596 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
598 skb->ip_summed = CHECKSUM_PARTIAL;
599 skb_split(skb, buff, len);
602 buff->ip_summed = skb->ip_summed;
604 /* Looks stupid, but our code really uses when of
605 * skbs, which it never sent before. --ANK
607 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
608 buff->tstamp = skb->tstamp;
610 old_factor = tcp_skb_pcount(skb);
612 /* Fix up tso_factor for both original and new SKB. */
613 tcp_set_skb_tso_segs(sk, skb, mss_now);
614 tcp_set_skb_tso_segs(sk, buff, mss_now);
616 /* If this packet has been sent out already, we must
617 * adjust the various packet counters.
619 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
620 int diff = old_factor - tcp_skb_pcount(skb) -
621 tcp_skb_pcount(buff);
623 tp->packets_out -= diff;
625 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
626 tp->sacked_out -= diff;
627 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
628 tp->retrans_out -= diff;
630 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
631 tp->lost_out -= diff;
632 tp->left_out -= diff;
636 /* Adjust Reno SACK estimate. */
637 if (!tp->rx_opt.sack_ok) {
638 tp->sacked_out -= diff;
639 if ((int)tp->sacked_out < 0)
641 tcp_sync_left_out(tp);
644 tp->fackets_out -= diff;
645 if ((int)tp->fackets_out < 0)
650 /* Link BUFF into the send queue. */
651 skb_header_release(buff);
652 __skb_append(skb, buff, &sk->sk_write_queue);
657 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
658 * eventually). The difference is that pulled data not copied, but
659 * immediately discarded.
661 static void __pskb_trim_head(struct sk_buff *skb, int len)
667 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
668 if (skb_shinfo(skb)->frags[i].size <= eat) {
669 put_page(skb_shinfo(skb)->frags[i].page);
670 eat -= skb_shinfo(skb)->frags[i].size;
672 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
674 skb_shinfo(skb)->frags[k].page_offset += eat;
675 skb_shinfo(skb)->frags[k].size -= eat;
681 skb_shinfo(skb)->nr_frags = k;
683 skb->tail = skb->data;
684 skb->data_len -= len;
685 skb->len = skb->data_len;
688 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
690 if (skb_cloned(skb) &&
691 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
694 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
695 if (unlikely(len < skb_headlen(skb)))
696 __skb_pull(skb, len);
698 __pskb_trim_head(skb, len - skb_headlen(skb));
700 TCP_SKB_CB(skb)->seq += len;
701 skb->ip_summed = CHECKSUM_PARTIAL;
703 skb->truesize -= len;
704 sk->sk_wmem_queued -= len;
705 sk->sk_forward_alloc += len;
706 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
708 /* Any change of skb->len requires recalculation of tso
711 if (tcp_skb_pcount(skb) > 1)
712 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
717 /* Not accounting for SACKs here. */
718 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
720 struct tcp_sock *tp = tcp_sk(sk);
721 struct inet_connection_sock *icsk = inet_csk(sk);
724 /* Calculate base mss without TCP options:
725 It is MMS_S - sizeof(tcphdr) of rfc1122
727 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
729 /* Clamp it (mss_clamp does not include tcp options) */
730 if (mss_now > tp->rx_opt.mss_clamp)
731 mss_now = tp->rx_opt.mss_clamp;
733 /* Now subtract optional transport overhead */
734 mss_now -= icsk->icsk_ext_hdr_len;
736 /* Then reserve room for full set of TCP options and 8 bytes of data */
740 /* Now subtract TCP options size, not including SACKs */
741 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
746 /* Inverse of above */
747 int tcp_mss_to_mtu(struct sock *sk, int mss)
749 struct tcp_sock *tp = tcp_sk(sk);
750 struct inet_connection_sock *icsk = inet_csk(sk);
755 icsk->icsk_ext_hdr_len +
756 icsk->icsk_af_ops->net_header_len;
761 void tcp_mtup_init(struct sock *sk)
763 struct tcp_sock *tp = tcp_sk(sk);
764 struct inet_connection_sock *icsk = inet_csk(sk);
766 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
767 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
768 icsk->icsk_af_ops->net_header_len;
769 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
770 icsk->icsk_mtup.probe_size = 0;
773 /* This function synchronize snd mss to current pmtu/exthdr set.
775 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
776 for TCP options, but includes only bare TCP header.
778 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
779 It is minimum of user_mss and mss received with SYN.
780 It also does not include TCP options.
782 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
784 tp->mss_cache is current effective sending mss, including
785 all tcp options except for SACKs. It is evaluated,
786 taking into account current pmtu, but never exceeds
787 tp->rx_opt.mss_clamp.
789 NOTE1. rfc1122 clearly states that advertised MSS
790 DOES NOT include either tcp or ip options.
792 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
793 are READ ONLY outside this function. --ANK (980731)
796 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
798 struct tcp_sock *tp = tcp_sk(sk);
799 struct inet_connection_sock *icsk = inet_csk(sk);
802 if (icsk->icsk_mtup.search_high > pmtu)
803 icsk->icsk_mtup.search_high = pmtu;
805 mss_now = tcp_mtu_to_mss(sk, pmtu);
807 /* Bound mss with half of window */
808 if (tp->max_window && mss_now > (tp->max_window>>1))
809 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
811 /* And store cached results */
812 icsk->icsk_pmtu_cookie = pmtu;
813 if (icsk->icsk_mtup.enabled)
814 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
815 tp->mss_cache = mss_now;
820 /* Compute the current effective MSS, taking SACKs and IP options,
821 * and even PMTU discovery events into account.
823 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
824 * cannot be large. However, taking into account rare use of URG, this
827 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
829 struct tcp_sock *tp = tcp_sk(sk);
830 struct dst_entry *dst = __sk_dst_get(sk);
835 mss_now = tp->mss_cache;
837 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
841 u32 mtu = dst_mtu(dst);
842 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
843 mss_now = tcp_sync_mss(sk, mtu);
846 if (tp->rx_opt.eff_sacks)
847 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
848 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
850 xmit_size_goal = mss_now;
853 xmit_size_goal = (65535 -
854 inet_csk(sk)->icsk_af_ops->net_header_len -
855 inet_csk(sk)->icsk_ext_hdr_len -
858 if (tp->max_window &&
859 (xmit_size_goal > (tp->max_window >> 1)))
860 xmit_size_goal = max((tp->max_window >> 1),
861 68U - tp->tcp_header_len);
863 xmit_size_goal -= (xmit_size_goal % mss_now);
865 tp->xmit_size_goal = xmit_size_goal;
870 /* Congestion window validation. (RFC2861) */
872 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
874 __u32 packets_out = tp->packets_out;
876 if (packets_out >= tp->snd_cwnd) {
877 /* Network is feed fully. */
878 tp->snd_cwnd_used = 0;
879 tp->snd_cwnd_stamp = tcp_time_stamp;
881 /* Network starves. */
882 if (tp->packets_out > tp->snd_cwnd_used)
883 tp->snd_cwnd_used = tp->packets_out;
885 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
886 tcp_cwnd_application_limited(sk);
890 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
892 u32 window, cwnd_len;
894 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
895 cwnd_len = mss_now * cwnd;
896 return min(window, cwnd_len);
899 /* Can at least one segment of SKB be sent right now, according to the
900 * congestion window rules? If so, return how many segments are allowed.
902 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
906 /* Don't be strict about the congestion window for the final FIN. */
907 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
910 in_flight = tcp_packets_in_flight(tp);
912 if (in_flight < cwnd)
913 return (cwnd - in_flight);
918 /* This must be invoked the first time we consider transmitting
921 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
923 int tso_segs = tcp_skb_pcount(skb);
927 tcp_skb_mss(skb) != mss_now)) {
928 tcp_set_skb_tso_segs(sk, skb, mss_now);
929 tso_segs = tcp_skb_pcount(skb);
934 static inline int tcp_minshall_check(const struct tcp_sock *tp)
936 return after(tp->snd_sml,tp->snd_una) &&
937 !after(tp->snd_sml, tp->snd_nxt);
940 /* Return 0, if packet can be sent now without violation Nagle's rules:
941 * 1. It is full sized.
942 * 2. Or it contains FIN. (already checked by caller)
943 * 3. Or TCP_NODELAY was set.
944 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
945 * With Minshall's modification: all sent small packets are ACKed.
948 static inline int tcp_nagle_check(const struct tcp_sock *tp,
949 const struct sk_buff *skb,
950 unsigned mss_now, int nonagle)
952 return (skb->len < mss_now &&
953 ((nonagle&TCP_NAGLE_CORK) ||
956 tcp_minshall_check(tp))));
959 /* Return non-zero if the Nagle test allows this packet to be
962 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
963 unsigned int cur_mss, int nonagle)
965 /* Nagle rule does not apply to frames, which sit in the middle of the
966 * write_queue (they have no chances to get new data).
968 * This is implemented in the callers, where they modify the 'nonagle'
969 * argument based upon the location of SKB in the send queue.
971 if (nonagle & TCP_NAGLE_PUSH)
974 /* Don't use the nagle rule for urgent data (or for the final FIN). */
976 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
979 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
985 /* Does at least the first segment of SKB fit into the send window? */
986 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
988 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
990 if (skb->len > cur_mss)
991 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
993 return !after(end_seq, tp->snd_una + tp->snd_wnd);
996 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
997 * should be put on the wire right now. If so, it returns the number of
998 * packets allowed by the congestion window.
1000 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1001 unsigned int cur_mss, int nonagle)
1003 struct tcp_sock *tp = tcp_sk(sk);
1004 unsigned int cwnd_quota;
1006 tcp_init_tso_segs(sk, skb, cur_mss);
1008 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1011 cwnd_quota = tcp_cwnd_test(tp, skb);
1013 !tcp_snd_wnd_test(tp, skb, cur_mss))
1019 static inline int tcp_skb_is_last(const struct sock *sk,
1020 const struct sk_buff *skb)
1022 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1025 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1027 struct sk_buff *skb = sk->sk_send_head;
1030 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1031 (tcp_skb_is_last(sk, skb) ?
1036 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1037 * which is put after SKB on the list. It is very much like
1038 * tcp_fragment() except that it may make several kinds of assumptions
1039 * in order to speed up the splitting operation. In particular, we
1040 * know that all the data is in scatter-gather pages, and that the
1041 * packet has never been sent out before (and thus is not cloned).
1043 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1045 struct sk_buff *buff;
1046 int nlen = skb->len - len;
1049 /* All of a TSO frame must be composed of paged data. */
1050 if (skb->len != skb->data_len)
1051 return tcp_fragment(sk, skb, len, mss_now);
1053 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1054 if (unlikely(buff == NULL))
1057 sk_charge_skb(sk, buff);
1058 buff->truesize += nlen;
1059 skb->truesize -= nlen;
1061 /* Correct the sequence numbers. */
1062 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1063 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1064 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1066 /* PSH and FIN should only be set in the second packet. */
1067 flags = TCP_SKB_CB(skb)->flags;
1068 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1069 TCP_SKB_CB(buff)->flags = flags;
1071 /* This packet was never sent out yet, so no SACK bits. */
1072 TCP_SKB_CB(buff)->sacked = 0;
1074 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1075 skb_split(skb, buff, len);
1077 /* Fix up tso_factor for both original and new SKB. */
1078 tcp_set_skb_tso_segs(sk, skb, mss_now);
1079 tcp_set_skb_tso_segs(sk, buff, mss_now);
1081 /* Link BUFF into the send queue. */
1082 skb_header_release(buff);
1083 __skb_append(skb, buff, &sk->sk_write_queue);
1088 /* Try to defer sending, if possible, in order to minimize the amount
1089 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1091 * This algorithm is from John Heffner.
1093 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1095 const struct inet_connection_sock *icsk = inet_csk(sk);
1096 u32 send_win, cong_win, limit, in_flight;
1098 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1101 if (icsk->icsk_ca_state != TCP_CA_Open)
1104 in_flight = tcp_packets_in_flight(tp);
1106 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1107 (tp->snd_cwnd <= in_flight));
1109 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1111 /* From in_flight test above, we know that cwnd > in_flight. */
1112 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1114 limit = min(send_win, cong_win);
1116 /* If a full-sized TSO skb can be sent, do it. */
1120 if (sysctl_tcp_tso_win_divisor) {
1121 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1123 /* If at least some fraction of a window is available,
1126 chunk /= sysctl_tcp_tso_win_divisor;
1130 /* Different approach, try not to defer past a single
1131 * ACK. Receiver should ACK every other full sized
1132 * frame, so if we have space for more than 3 frames
1135 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1139 /* Ok, it looks like it is advisable to defer. */
1143 /* Create a new MTU probe if we are ready.
1144 * Returns 0 if we should wait to probe (no cwnd available),
1145 * 1 if a probe was sent,
1147 static int tcp_mtu_probe(struct sock *sk)
1149 struct tcp_sock *tp = tcp_sk(sk);
1150 struct inet_connection_sock *icsk = inet_csk(sk);
1151 struct sk_buff *skb, *nskb, *next;
1158 /* Not currently probing/verifying,
1160 * have enough cwnd, and
1161 * not SACKing (the variable headers throw things off) */
1162 if (!icsk->icsk_mtup.enabled ||
1163 icsk->icsk_mtup.probe_size ||
1164 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1165 tp->snd_cwnd < 11 ||
1166 tp->rx_opt.eff_sacks)
1169 /* Very simple search strategy: just double the MSS. */
1170 mss_now = tcp_current_mss(sk, 0);
1171 probe_size = 2*tp->mss_cache;
1172 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1173 /* TODO: set timer for probe_converge_event */
1177 /* Have enough data in the send queue to probe? */
1179 if ((skb = sk->sk_send_head) == NULL)
1181 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1183 if (len < probe_size)
1186 /* Receive window check. */
1187 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1188 if (tp->snd_wnd < probe_size)
1194 /* Do we need to wait to drain cwnd? */
1195 pif = tcp_packets_in_flight(tp);
1196 if (pif + 2 > tp->snd_cwnd) {
1197 /* With no packets in flight, don't stall. */
1204 /* We're allowed to probe. Build it now. */
1205 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1207 sk_charge_skb(sk, nskb);
1209 skb = sk->sk_send_head;
1210 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1211 sk->sk_send_head = nskb;
1213 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1214 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1215 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1216 TCP_SKB_CB(nskb)->sacked = 0;
1218 nskb->ip_summed = skb->ip_summed;
1221 while (len < probe_size) {
1224 copy = min_t(int, skb->len, probe_size - len);
1225 if (nskb->ip_summed)
1226 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1228 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1229 skb_put(nskb, copy), copy, nskb->csum);
1231 if (skb->len <= copy) {
1232 /* We've eaten all the data from this skb.
1234 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1235 __skb_unlink(skb, &sk->sk_write_queue);
1236 sk_stream_free_skb(sk, skb);
1238 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1239 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1240 if (!skb_shinfo(skb)->nr_frags) {
1241 skb_pull(skb, copy);
1242 if (skb->ip_summed != CHECKSUM_PARTIAL)
1243 skb->csum = csum_partial(skb->data, skb->len, 0);
1245 __pskb_trim_head(skb, copy);
1246 tcp_set_skb_tso_segs(sk, skb, mss_now);
1248 TCP_SKB_CB(skb)->seq += copy;
1254 tcp_init_tso_segs(sk, nskb, nskb->len);
1256 /* We're ready to send. If this fails, the probe will
1257 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1258 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1259 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1260 /* Decrement cwnd here because we are sending
1261 * effectively two packets. */
1263 update_send_head(sk, tp, nskb);
1265 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1266 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1267 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1276 /* This routine writes packets to the network. It advances the
1277 * send_head. This happens as incoming acks open up the remote
1280 * Returns 1, if no segments are in flight and we have queued segments, but
1281 * cannot send anything now because of SWS or another problem.
1283 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1285 struct tcp_sock *tp = tcp_sk(sk);
1286 struct sk_buff *skb;
1287 unsigned int tso_segs, sent_pkts;
1291 /* If we are closed, the bytes will have to remain here.
1292 * In time closedown will finish, we empty the write queue and all
1295 if (unlikely(sk->sk_state == TCP_CLOSE))
1300 /* Do MTU probing. */
1301 if ((result = tcp_mtu_probe(sk)) == 0) {
1303 } else if (result > 0) {
1307 while ((skb = sk->sk_send_head)) {
1310 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1313 cwnd_quota = tcp_cwnd_test(tp, skb);
1317 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1320 if (tso_segs == 1) {
1321 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1322 (tcp_skb_is_last(sk, skb) ?
1323 nonagle : TCP_NAGLE_PUSH))))
1326 if (tcp_tso_should_defer(sk, tp, skb))
1332 limit = tcp_window_allows(tp, skb,
1333 mss_now, cwnd_quota);
1335 if (skb->len < limit) {
1336 unsigned int trim = skb->len % mss_now;
1339 limit = skb->len - trim;
1343 if (skb->len > limit &&
1344 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1347 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1349 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1352 /* Advance the send_head. This one is sent out.
1353 * This call will increment packets_out.
1355 update_send_head(sk, tp, skb);
1357 tcp_minshall_update(tp, mss_now, skb);
1361 if (likely(sent_pkts)) {
1362 tcp_cwnd_validate(sk, tp);
1365 return !tp->packets_out && sk->sk_send_head;
1368 /* Push out any pending frames which were held back due to
1369 * TCP_CORK or attempt at coalescing tiny packets.
1370 * The socket must be locked by the caller.
1372 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1373 unsigned int cur_mss, int nonagle)
1375 struct sk_buff *skb = sk->sk_send_head;
1378 if (tcp_write_xmit(sk, cur_mss, nonagle))
1379 tcp_check_probe_timer(sk, tp);
1383 /* Send _single_ skb sitting at the send head. This function requires
1384 * true push pending frames to setup probe timer etc.
1386 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1388 struct tcp_sock *tp = tcp_sk(sk);
1389 struct sk_buff *skb = sk->sk_send_head;
1390 unsigned int tso_segs, cwnd_quota;
1392 BUG_ON(!skb || skb->len < mss_now);
1394 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1395 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1397 if (likely(cwnd_quota)) {
1404 limit = tcp_window_allows(tp, skb,
1405 mss_now, cwnd_quota);
1407 if (skb->len < limit) {
1408 unsigned int trim = skb->len % mss_now;
1411 limit = skb->len - trim;
1415 if (skb->len > limit &&
1416 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1419 /* Send it out now. */
1420 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1422 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1423 update_send_head(sk, tp, skb);
1424 tcp_cwnd_validate(sk, tp);
1430 /* This function returns the amount that we can raise the
1431 * usable window based on the following constraints
1433 * 1. The window can never be shrunk once it is offered (RFC 793)
1434 * 2. We limit memory per socket
1437 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1438 * RECV.NEXT + RCV.WIN fixed until:
1439 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1441 * i.e. don't raise the right edge of the window until you can raise
1442 * it at least MSS bytes.
1444 * Unfortunately, the recommended algorithm breaks header prediction,
1445 * since header prediction assumes th->window stays fixed.
1447 * Strictly speaking, keeping th->window fixed violates the receiver
1448 * side SWS prevention criteria. The problem is that under this rule
1449 * a stream of single byte packets will cause the right side of the
1450 * window to always advance by a single byte.
1452 * Of course, if the sender implements sender side SWS prevention
1453 * then this will not be a problem.
1455 * BSD seems to make the following compromise:
1457 * If the free space is less than the 1/4 of the maximum
1458 * space available and the free space is less than 1/2 mss,
1459 * then set the window to 0.
1460 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1461 * Otherwise, just prevent the window from shrinking
1462 * and from being larger than the largest representable value.
1464 * This prevents incremental opening of the window in the regime
1465 * where TCP is limited by the speed of the reader side taking
1466 * data out of the TCP receive queue. It does nothing about
1467 * those cases where the window is constrained on the sender side
1468 * because the pipeline is full.
1470 * BSD also seems to "accidentally" limit itself to windows that are a
1471 * multiple of MSS, at least until the free space gets quite small.
1472 * This would appear to be a side effect of the mbuf implementation.
1473 * Combining these two algorithms results in the observed behavior
1474 * of having a fixed window size at almost all times.
1476 * Below we obtain similar behavior by forcing the offered window to
1477 * a multiple of the mss when it is feasible to do so.
1479 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1480 * Regular options like TIMESTAMP are taken into account.
1482 u32 __tcp_select_window(struct sock *sk)
1484 struct inet_connection_sock *icsk = inet_csk(sk);
1485 struct tcp_sock *tp = tcp_sk(sk);
1486 /* MSS for the peer's data. Previous versions used mss_clamp
1487 * here. I don't know if the value based on our guesses
1488 * of peer's MSS is better for the performance. It's more correct
1489 * but may be worse for the performance because of rcv_mss
1490 * fluctuations. --SAW 1998/11/1
1492 int mss = icsk->icsk_ack.rcv_mss;
1493 int free_space = tcp_space(sk);
1494 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1497 if (mss > full_space)
1500 if (free_space < full_space/2) {
1501 icsk->icsk_ack.quick = 0;
1503 if (tcp_memory_pressure)
1504 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1506 if (free_space < mss)
1510 if (free_space > tp->rcv_ssthresh)
1511 free_space = tp->rcv_ssthresh;
1513 /* Don't do rounding if we are using window scaling, since the
1514 * scaled window will not line up with the MSS boundary anyway.
1516 window = tp->rcv_wnd;
1517 if (tp->rx_opt.rcv_wscale) {
1518 window = free_space;
1520 /* Advertise enough space so that it won't get scaled away.
1521 * Import case: prevent zero window announcement if
1522 * 1<<rcv_wscale > mss.
1524 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1525 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1526 << tp->rx_opt.rcv_wscale);
1528 /* Get the largest window that is a nice multiple of mss.
1529 * Window clamp already applied above.
1530 * If our current window offering is within 1 mss of the
1531 * free space we just keep it. This prevents the divide
1532 * and multiply from happening most of the time.
1533 * We also don't do any window rounding when the free space
1536 if (window <= free_space - mss || window > free_space)
1537 window = (free_space/mss)*mss;
1543 /* Attempt to collapse two adjacent SKB's during retransmission. */
1544 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1546 struct tcp_sock *tp = tcp_sk(sk);
1547 struct sk_buff *next_skb = skb->next;
1549 /* The first test we must make is that neither of these two
1550 * SKB's are still referenced by someone else.
1552 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1553 int skb_size = skb->len, next_skb_size = next_skb->len;
1554 u16 flags = TCP_SKB_CB(skb)->flags;
1556 /* Also punt if next skb has been SACK'd. */
1557 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1560 /* Next skb is out of window. */
1561 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1564 /* Punt if not enough space exists in the first SKB for
1565 * the data in the second, or the total combined payload
1566 * would exceed the MSS.
1568 if ((next_skb_size > skb_tailroom(skb)) ||
1569 ((skb_size + next_skb_size) > mss_now))
1572 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1573 tcp_skb_pcount(next_skb) != 1);
1575 /* changing transmit queue under us so clear hints */
1576 clear_all_retrans_hints(tp);
1578 /* Ok. We will be able to collapse the packet. */
1579 __skb_unlink(next_skb, &sk->sk_write_queue);
1581 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1583 skb->ip_summed = next_skb->ip_summed;
1585 if (skb->ip_summed != CHECKSUM_PARTIAL)
1586 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1588 /* Update sequence range on original skb. */
1589 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1591 /* Merge over control information. */
1592 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1593 TCP_SKB_CB(skb)->flags = flags;
1595 /* All done, get rid of second SKB and account for it so
1596 * packet counting does not break.
1598 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1599 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1600 tp->retrans_out -= tcp_skb_pcount(next_skb);
1601 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1602 tp->lost_out -= tcp_skb_pcount(next_skb);
1603 tp->left_out -= tcp_skb_pcount(next_skb);
1605 /* Reno case is special. Sigh... */
1606 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1607 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1608 tp->left_out -= tcp_skb_pcount(next_skb);
1611 /* Not quite right: it can be > snd.fack, but
1612 * it is better to underestimate fackets.
1614 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1615 tcp_packets_out_dec(tp, next_skb);
1616 sk_stream_free_skb(sk, next_skb);
1620 /* Do a simple retransmit without using the backoff mechanisms in
1621 * tcp_timer. This is used for path mtu discovery.
1622 * The socket is already locked here.
1624 void tcp_simple_retransmit(struct sock *sk)
1626 const struct inet_connection_sock *icsk = inet_csk(sk);
1627 struct tcp_sock *tp = tcp_sk(sk);
1628 struct sk_buff *skb;
1629 unsigned int mss = tcp_current_mss(sk, 0);
1632 sk_stream_for_retrans_queue(skb, sk) {
1633 if (skb->len > mss &&
1634 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1635 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1636 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1637 tp->retrans_out -= tcp_skb_pcount(skb);
1639 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1640 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1641 tp->lost_out += tcp_skb_pcount(skb);
1647 clear_all_retrans_hints(tp);
1652 tcp_sync_left_out(tp);
1654 /* Don't muck with the congestion window here.
1655 * Reason is that we do not increase amount of _data_
1656 * in network, but units changed and effective
1657 * cwnd/ssthresh really reduced now.
1659 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1660 tp->high_seq = tp->snd_nxt;
1661 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1662 tp->prior_ssthresh = 0;
1663 tp->undo_marker = 0;
1664 tcp_set_ca_state(sk, TCP_CA_Loss);
1666 tcp_xmit_retransmit_queue(sk);
1669 /* This retransmits one SKB. Policy decisions and retransmit queue
1670 * state updates are done by the caller. Returns non-zero if an
1671 * error occurred which prevented the send.
1673 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1675 struct tcp_sock *tp = tcp_sk(sk);
1676 struct inet_connection_sock *icsk = inet_csk(sk);
1677 unsigned int cur_mss = tcp_current_mss(sk, 0);
1680 /* Inconslusive MTU probe */
1681 if (icsk->icsk_mtup.probe_size) {
1682 icsk->icsk_mtup.probe_size = 0;
1685 /* Do not sent more than we queued. 1/4 is reserved for possible
1686 * copying overhead: fragmentation, tunneling, mangling etc.
1688 if (atomic_read(&sk->sk_wmem_alloc) >
1689 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1692 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1693 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1695 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1699 /* If receiver has shrunk his window, and skb is out of
1700 * new window, do not retransmit it. The exception is the
1701 * case, when window is shrunk to zero. In this case
1702 * our retransmit serves as a zero window probe.
1704 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1705 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1708 if (skb->len > cur_mss) {
1709 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1710 return -ENOMEM; /* We'll try again later. */
1713 /* Collapse two adjacent packets if worthwhile and we can. */
1714 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1715 (skb->len < (cur_mss >> 1)) &&
1716 (skb->next != sk->sk_send_head) &&
1717 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1718 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1719 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1720 (sysctl_tcp_retrans_collapse != 0))
1721 tcp_retrans_try_collapse(sk, skb, cur_mss);
1723 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1724 return -EHOSTUNREACH; /* Routing failure or similar. */
1726 /* Some Solaris stacks overoptimize and ignore the FIN on a
1727 * retransmit when old data is attached. So strip it off
1728 * since it is cheap to do so and saves bytes on the network.
1731 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1732 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1733 if (!pskb_trim(skb, 0)) {
1734 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1735 skb_shinfo(skb)->gso_segs = 1;
1736 skb_shinfo(skb)->gso_size = 0;
1737 skb_shinfo(skb)->gso_type = 0;
1738 skb->ip_summed = CHECKSUM_NONE;
1743 /* Make a copy, if the first transmission SKB clone we made
1744 * is still in somebody's hands, else make a clone.
1746 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1748 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1751 /* Update global TCP statistics. */
1752 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1754 tp->total_retrans++;
1756 #if FASTRETRANS_DEBUG > 0
1757 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1758 if (net_ratelimit())
1759 printk(KERN_DEBUG "retrans_out leaked.\n");
1762 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1763 tp->retrans_out += tcp_skb_pcount(skb);
1765 /* Save stamp of the first retransmit. */
1766 if (!tp->retrans_stamp)
1767 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1771 /* snd_nxt is stored to detect loss of retransmitted segment,
1772 * see tcp_input.c tcp_sacktag_write_queue().
1774 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1779 /* This gets called after a retransmit timeout, and the initially
1780 * retransmitted data is acknowledged. It tries to continue
1781 * resending the rest of the retransmit queue, until either
1782 * we've sent it all or the congestion window limit is reached.
1783 * If doing SACK, the first ACK which comes back for a timeout
1784 * based retransmit packet might feed us FACK information again.
1785 * If so, we use it to avoid unnecessarily retransmissions.
1787 void tcp_xmit_retransmit_queue(struct sock *sk)
1789 const struct inet_connection_sock *icsk = inet_csk(sk);
1790 struct tcp_sock *tp = tcp_sk(sk);
1791 struct sk_buff *skb;
1794 if (tp->retransmit_skb_hint) {
1795 skb = tp->retransmit_skb_hint;
1796 packet_cnt = tp->retransmit_cnt_hint;
1798 skb = sk->sk_write_queue.next;
1802 /* First pass: retransmit lost packets. */
1804 sk_stream_for_retrans_queue_from(skb, sk) {
1805 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1807 /* we could do better than to assign each time */
1808 tp->retransmit_skb_hint = skb;
1809 tp->retransmit_cnt_hint = packet_cnt;
1811 /* Assume this retransmit will generate
1812 * only one packet for congestion window
1813 * calculation purposes. This works because
1814 * tcp_retransmit_skb() will chop up the
1815 * packet to be MSS sized and all the
1816 * packet counting works out.
1818 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1821 if (sacked & TCPCB_LOST) {
1822 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1823 if (tcp_retransmit_skb(sk, skb)) {
1824 tp->retransmit_skb_hint = NULL;
1827 if (icsk->icsk_ca_state != TCP_CA_Loss)
1828 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1830 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1833 skb_peek(&sk->sk_write_queue))
1834 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1835 inet_csk(sk)->icsk_rto,
1839 packet_cnt += tcp_skb_pcount(skb);
1840 if (packet_cnt >= tp->lost_out)
1846 /* OK, demanded retransmission is finished. */
1848 /* Forward retransmissions are possible only during Recovery. */
1849 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1852 /* No forward retransmissions in Reno are possible. */
1853 if (!tp->rx_opt.sack_ok)
1856 /* Yeah, we have to make difficult choice between forward transmission
1857 * and retransmission... Both ways have their merits...
1859 * For now we do not retransmit anything, while we have some new
1863 if (tcp_may_send_now(sk, tp))
1866 if (tp->forward_skb_hint) {
1867 skb = tp->forward_skb_hint;
1868 packet_cnt = tp->forward_cnt_hint;
1870 skb = sk->sk_write_queue.next;
1874 sk_stream_for_retrans_queue_from(skb, sk) {
1875 tp->forward_cnt_hint = packet_cnt;
1876 tp->forward_skb_hint = skb;
1878 /* Similar to the retransmit loop above we
1879 * can pretend that the retransmitted SKB
1880 * we send out here will be composed of one
1881 * real MSS sized packet because tcp_retransmit_skb()
1882 * will fragment it if necessary.
1884 if (++packet_cnt > tp->fackets_out)
1887 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1890 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1893 /* Ok, retransmit it. */
1894 if (tcp_retransmit_skb(sk, skb)) {
1895 tp->forward_skb_hint = NULL;
1899 if (skb == skb_peek(&sk->sk_write_queue))
1900 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1901 inet_csk(sk)->icsk_rto,
1904 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1909 /* Send a fin. The caller locks the socket for us. This cannot be
1910 * allowed to fail queueing a FIN frame under any circumstances.
1912 void tcp_send_fin(struct sock *sk)
1914 struct tcp_sock *tp = tcp_sk(sk);
1915 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1918 /* Optimization, tack on the FIN if we have a queue of
1919 * unsent frames. But be careful about outgoing SACKS
1922 mss_now = tcp_current_mss(sk, 1);
1924 if (sk->sk_send_head != NULL) {
1925 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1926 TCP_SKB_CB(skb)->end_seq++;
1929 /* Socket is locked, keep trying until memory is available. */
1931 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1937 /* Reserve space for headers and prepare control bits. */
1938 skb_reserve(skb, MAX_TCP_HEADER);
1940 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1941 TCP_SKB_CB(skb)->sacked = 0;
1942 skb_shinfo(skb)->gso_segs = 1;
1943 skb_shinfo(skb)->gso_size = 0;
1944 skb_shinfo(skb)->gso_type = 0;
1946 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1947 TCP_SKB_CB(skb)->seq = tp->write_seq;
1948 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1949 tcp_queue_skb(sk, skb);
1951 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1954 /* We get here when a process closes a file descriptor (either due to
1955 * an explicit close() or as a byproduct of exit()'ing) and there
1956 * was unread data in the receive queue. This behavior is recommended
1957 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1959 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1961 struct tcp_sock *tp = tcp_sk(sk);
1962 struct sk_buff *skb;
1964 /* NOTE: No TCP options attached and we never retransmit this. */
1965 skb = alloc_skb(MAX_TCP_HEADER, priority);
1967 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1971 /* Reserve space for headers and prepare control bits. */
1972 skb_reserve(skb, MAX_TCP_HEADER);
1974 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1975 TCP_SKB_CB(skb)->sacked = 0;
1976 skb_shinfo(skb)->gso_segs = 1;
1977 skb_shinfo(skb)->gso_size = 0;
1978 skb_shinfo(skb)->gso_type = 0;
1981 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1982 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1983 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1984 if (tcp_transmit_skb(sk, skb, 0, priority))
1985 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1988 /* WARNING: This routine must only be called when we have already sent
1989 * a SYN packet that crossed the incoming SYN that caused this routine
1990 * to get called. If this assumption fails then the initial rcv_wnd
1991 * and rcv_wscale values will not be correct.
1993 int tcp_send_synack(struct sock *sk)
1995 struct sk_buff* skb;
1997 skb = skb_peek(&sk->sk_write_queue);
1998 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1999 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2002 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2003 if (skb_cloned(skb)) {
2004 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2007 __skb_unlink(skb, &sk->sk_write_queue);
2008 skb_header_release(nskb);
2009 __skb_queue_head(&sk->sk_write_queue, nskb);
2010 sk_stream_free_skb(sk, skb);
2011 sk_charge_skb(sk, nskb);
2015 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2016 TCP_ECN_send_synack(tcp_sk(sk), skb);
2018 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2019 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2023 * Prepare a SYN-ACK.
2025 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2026 struct request_sock *req)
2028 struct inet_request_sock *ireq = inet_rsk(req);
2029 struct tcp_sock *tp = tcp_sk(sk);
2031 int tcp_header_size;
2032 struct sk_buff *skb;
2034 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2038 /* Reserve space for headers. */
2039 skb_reserve(skb, MAX_TCP_HEADER);
2041 skb->dst = dst_clone(dst);
2043 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2044 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2045 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2046 /* SACK_PERM is in the place of NOP NOP of TS */
2047 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2048 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2050 memset(th, 0, sizeof(struct tcphdr));
2053 TCP_ECN_make_synack(req, th);
2054 th->source = inet_sk(sk)->sport;
2055 th->dest = ireq->rmt_port;
2056 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2057 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2058 TCP_SKB_CB(skb)->sacked = 0;
2059 skb_shinfo(skb)->gso_segs = 1;
2060 skb_shinfo(skb)->gso_size = 0;
2061 skb_shinfo(skb)->gso_type = 0;
2062 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2063 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2064 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2066 /* Set this up on the first call only */
2067 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2068 /* tcp_full_space because it is guaranteed to be the first packet */
2069 tcp_select_initial_window(tcp_full_space(sk),
2070 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2075 ireq->rcv_wscale = rcv_wscale;
2078 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2079 th->window = htons(req->rcv_wnd);
2081 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2082 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2083 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2084 TCP_SKB_CB(skb)->when,
2088 th->doff = (tcp_header_size >> 2);
2089 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2094 * Do all connect socket setups that can be done AF independent.
2096 static void tcp_connect_init(struct sock *sk)
2098 struct dst_entry *dst = __sk_dst_get(sk);
2099 struct tcp_sock *tp = tcp_sk(sk);
2102 /* We'll fix this up when we get a response from the other end.
2103 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2105 tp->tcp_header_len = sizeof(struct tcphdr) +
2106 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2108 /* If user gave his TCP_MAXSEG, record it to clamp */
2109 if (tp->rx_opt.user_mss)
2110 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2113 tcp_sync_mss(sk, dst_mtu(dst));
2115 if (!tp->window_clamp)
2116 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2117 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2118 tcp_initialize_rcv_mss(sk);
2120 tcp_select_initial_window(tcp_full_space(sk),
2121 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2124 sysctl_tcp_window_scaling,
2127 tp->rx_opt.rcv_wscale = rcv_wscale;
2128 tp->rcv_ssthresh = tp->rcv_wnd;
2131 sock_reset_flag(sk, SOCK_DONE);
2133 tcp_init_wl(tp, tp->write_seq, 0);
2134 tp->snd_una = tp->write_seq;
2135 tp->snd_sml = tp->write_seq;
2140 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2141 inet_csk(sk)->icsk_retransmits = 0;
2142 tcp_clear_retrans(tp);
2146 * Build a SYN and send it off.
2148 int tcp_connect(struct sock *sk)
2150 struct tcp_sock *tp = tcp_sk(sk);
2151 struct sk_buff *buff;
2153 tcp_connect_init(sk);
2155 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2156 if (unlikely(buff == NULL))
2159 /* Reserve space for headers. */
2160 skb_reserve(buff, MAX_TCP_HEADER);
2162 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2163 TCP_ECN_send_syn(sk, tp, buff);
2164 TCP_SKB_CB(buff)->sacked = 0;
2165 skb_shinfo(buff)->gso_segs = 1;
2166 skb_shinfo(buff)->gso_size = 0;
2167 skb_shinfo(buff)->gso_type = 0;
2169 tp->snd_nxt = tp->write_seq;
2170 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2171 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2174 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2175 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2176 skb_header_release(buff);
2177 __skb_queue_tail(&sk->sk_write_queue, buff);
2178 sk_charge_skb(sk, buff);
2179 tp->packets_out += tcp_skb_pcount(buff);
2180 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2182 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2183 * in order to make this packet get counted in tcpOutSegs.
2185 tp->snd_nxt = tp->write_seq;
2186 tp->pushed_seq = tp->write_seq;
2187 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2189 /* Timer for repeating the SYN until an answer. */
2190 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2191 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2195 /* Send out a delayed ack, the caller does the policy checking
2196 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2199 void tcp_send_delayed_ack(struct sock *sk)
2201 struct inet_connection_sock *icsk = inet_csk(sk);
2202 int ato = icsk->icsk_ack.ato;
2203 unsigned long timeout;
2205 if (ato > TCP_DELACK_MIN) {
2206 const struct tcp_sock *tp = tcp_sk(sk);
2209 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2210 max_ato = TCP_DELACK_MAX;
2212 /* Slow path, intersegment interval is "high". */
2214 /* If some rtt estimate is known, use it to bound delayed ack.
2215 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2219 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2225 ato = min(ato, max_ato);
2228 /* Stay within the limit we were given */
2229 timeout = jiffies + ato;
2231 /* Use new timeout only if there wasn't a older one earlier. */
2232 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2233 /* If delack timer was blocked or is about to expire,
2236 if (icsk->icsk_ack.blocked ||
2237 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2242 if (!time_before(timeout, icsk->icsk_ack.timeout))
2243 timeout = icsk->icsk_ack.timeout;
2245 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2246 icsk->icsk_ack.timeout = timeout;
2247 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2250 /* This routine sends an ack and also updates the window. */
2251 void tcp_send_ack(struct sock *sk)
2253 /* If we have been reset, we may not send again. */
2254 if (sk->sk_state != TCP_CLOSE) {
2255 struct tcp_sock *tp = tcp_sk(sk);
2256 struct sk_buff *buff;
2258 /* We are not putting this on the write queue, so
2259 * tcp_transmit_skb() will set the ownership to this
2262 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2264 inet_csk_schedule_ack(sk);
2265 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2266 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2267 TCP_DELACK_MAX, TCP_RTO_MAX);
2271 /* Reserve space for headers and prepare control bits. */
2272 skb_reserve(buff, MAX_TCP_HEADER);
2274 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2275 TCP_SKB_CB(buff)->sacked = 0;
2276 skb_shinfo(buff)->gso_segs = 1;
2277 skb_shinfo(buff)->gso_size = 0;
2278 skb_shinfo(buff)->gso_type = 0;
2280 /* Send it off, this clears delayed acks for us. */
2281 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2282 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2283 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2287 /* This routine sends a packet with an out of date sequence
2288 * number. It assumes the other end will try to ack it.
2290 * Question: what should we make while urgent mode?
2291 * 4.4BSD forces sending single byte of data. We cannot send
2292 * out of window data, because we have SND.NXT==SND.MAX...
2294 * Current solution: to send TWO zero-length segments in urgent mode:
2295 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2296 * out-of-date with SND.UNA-1 to probe window.
2298 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2300 struct tcp_sock *tp = tcp_sk(sk);
2301 struct sk_buff *skb;
2303 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2304 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2308 /* Reserve space for headers and set control bits. */
2309 skb_reserve(skb, MAX_TCP_HEADER);
2311 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2312 TCP_SKB_CB(skb)->sacked = urgent;
2313 skb_shinfo(skb)->gso_segs = 1;
2314 skb_shinfo(skb)->gso_size = 0;
2315 skb_shinfo(skb)->gso_type = 0;
2317 /* Use a previous sequence. This should cause the other
2318 * end to send an ack. Don't queue or clone SKB, just
2321 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2322 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2323 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2324 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2327 int tcp_write_wakeup(struct sock *sk)
2329 if (sk->sk_state != TCP_CLOSE) {
2330 struct tcp_sock *tp = tcp_sk(sk);
2331 struct sk_buff *skb;
2333 if ((skb = sk->sk_send_head) != NULL &&
2334 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2336 unsigned int mss = tcp_current_mss(sk, 0);
2337 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2339 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2340 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2342 /* We are probing the opening of a window
2343 * but the window size is != 0
2344 * must have been a result SWS avoidance ( sender )
2346 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2348 seg_size = min(seg_size, mss);
2349 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2350 if (tcp_fragment(sk, skb, seg_size, mss))
2352 } else if (!tcp_skb_pcount(skb))
2353 tcp_set_skb_tso_segs(sk, skb, mss);
2355 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2356 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2357 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2359 update_send_head(sk, tp, skb);
2364 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2365 tcp_xmit_probe_skb(sk, TCPCB_URG);
2366 return tcp_xmit_probe_skb(sk, 0);
2372 /* A window probe timeout has occurred. If window is not closed send
2373 * a partial packet else a zero probe.
2375 void tcp_send_probe0(struct sock *sk)
2377 struct inet_connection_sock *icsk = inet_csk(sk);
2378 struct tcp_sock *tp = tcp_sk(sk);
2381 err = tcp_write_wakeup(sk);
2383 if (tp->packets_out || !sk->sk_send_head) {
2384 /* Cancel probe timer, if it is not required. */
2385 icsk->icsk_probes_out = 0;
2386 icsk->icsk_backoff = 0;
2391 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2392 icsk->icsk_backoff++;
2393 icsk->icsk_probes_out++;
2394 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2395 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2398 /* If packet was not sent due to local congestion,
2399 * do not backoff and do not remember icsk_probes_out.
2400 * Let local senders to fight for local resources.
2402 * Use accumulated backoff yet.
2404 if (!icsk->icsk_probes_out)
2405 icsk->icsk_probes_out = 1;
2406 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2407 min(icsk->icsk_rto << icsk->icsk_backoff,
2408 TCP_RESOURCE_PROBE_INTERVAL),
2413 EXPORT_SYMBOL(tcp_connect);
2414 EXPORT_SYMBOL(tcp_make_synack);
2415 EXPORT_SYMBOL(tcp_simple_retransmit);
2416 EXPORT_SYMBOL(tcp_sync_mss);
2417 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2418 EXPORT_SYMBOL(tcp_mtup_init);