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 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle = 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(__u32 *ptr, struct tcp_sock *tp,
275 if (tp->rx_opt.tstamp_ok) {
276 *ptr++ = __constant_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(__u32 *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++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
329 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
331 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
332 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
333 *ptr++ = htonl(tstamp); /* TSVAL */
334 *ptr++ = htonl(ts_recent); /* TSECR */
336 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
337 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
339 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
342 /* This routine actually transmits TCP packets queued in by
343 * tcp_do_sendmsg(). This is used by both the initial
344 * transmission and possible later retransmissions.
345 * All SKB's seen here are completely headerless. It is our
346 * job to build the TCP header, and pass the packet down to
347 * IP so it can do the same plus pass the packet off to the
350 * We are working here with either a clone of the original
351 * SKB, or a fresh unique copy made by the retransmit engine.
353 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
355 const struct inet_connection_sock *icsk = inet_csk(sk);
356 struct inet_sock *inet;
358 struct tcp_skb_cb *tcb;
364 BUG_ON(!skb || !tcp_skb_pcount(skb));
366 /* If congestion control is doing timestamping, we must
367 * take such a timestamp before we potentially clone/copy.
369 if (icsk->icsk_ca_ops->rtt_sample)
370 __net_timestamp(skb);
372 if (likely(clone_it)) {
373 if (unlikely(skb_cloned(skb)))
374 skb = pskb_copy(skb, gfp_mask);
376 skb = skb_clone(skb, gfp_mask);
383 tcb = TCP_SKB_CB(skb);
384 tcp_header_size = tp->tcp_header_len;
386 #define SYSCTL_FLAG_TSTAMPS 0x1
387 #define SYSCTL_FLAG_WSCALE 0x2
388 #define SYSCTL_FLAG_SACK 0x4
391 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
392 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
393 if(sysctl_tcp_timestamps) {
394 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
395 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
397 if (sysctl_tcp_window_scaling) {
398 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
399 sysctl_flags |= SYSCTL_FLAG_WSCALE;
401 if (sysctl_tcp_sack) {
402 sysctl_flags |= SYSCTL_FLAG_SACK;
403 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
404 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
406 } else if (unlikely(tp->rx_opt.eff_sacks)) {
407 /* A SACK is 2 pad bytes, a 2 byte header, plus
408 * 2 32-bit sequence numbers for each SACK block.
410 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
411 (tp->rx_opt.eff_sacks *
412 TCPOLEN_SACK_PERBLOCK));
415 if (tcp_packets_in_flight(tp) == 0)
416 tcp_ca_event(sk, CA_EVENT_TX_START);
418 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
420 skb_set_owner_w(skb, sk);
422 /* Build TCP header and checksum it. */
423 th->source = inet->sport;
424 th->dest = inet->dport;
425 th->seq = htonl(tcb->seq);
426 th->ack_seq = htonl(tp->rcv_nxt);
427 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
430 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
431 /* RFC1323: The window in SYN & SYN/ACK segments
434 th->window = htons(tp->rcv_wnd);
436 th->window = htons(tcp_select_window(sk));
441 if (unlikely(tp->urg_mode &&
442 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
443 th->urg_ptr = htons(tp->snd_up-tcb->seq);
447 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
448 tcp_syn_build_options((__u32 *)(th + 1),
449 tcp_advertise_mss(sk),
450 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
451 (sysctl_flags & SYSCTL_FLAG_SACK),
452 (sysctl_flags & SYSCTL_FLAG_WSCALE),
453 tp->rx_opt.rcv_wscale,
455 tp->rx_opt.ts_recent);
457 tcp_build_and_update_options((__u32 *)(th + 1),
459 TCP_ECN_send(sk, tp, skb, tcp_header_size);
462 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
464 if (likely(tcb->flags & TCPCB_FLAG_ACK))
465 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
467 if (skb->len != tcp_header_size)
468 tcp_event_data_sent(tp, skb, sk);
470 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
471 TCP_INC_STATS(TCP_MIB_OUTSEGS);
473 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
474 if (likely(err <= 0))
479 /* NET_XMIT_CN is special. It does not guarantee,
480 * that this packet is lost. It tells that device
481 * is about to start to drop packets or already
482 * drops some packets of the same priority and
483 * invokes us to send less aggressively.
485 return err == NET_XMIT_CN ? 0 : err;
487 #undef SYSCTL_FLAG_TSTAMPS
488 #undef SYSCTL_FLAG_WSCALE
489 #undef SYSCTL_FLAG_SACK
493 /* This routine just queue's the buffer
495 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
496 * otherwise socket can stall.
498 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
500 struct tcp_sock *tp = tcp_sk(sk);
502 /* Advance write_seq and place onto the write_queue. */
503 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
504 skb_header_release(skb);
505 __skb_queue_tail(&sk->sk_write_queue, skb);
506 sk_charge_skb(sk, skb);
508 /* Queue it, remembering where we must start sending. */
509 if (sk->sk_send_head == NULL)
510 sk->sk_send_head = skb;
513 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
515 if (skb->len <= mss_now || !sk_can_gso(sk)) {
516 /* Avoid the costly divide in the normal
519 skb_shinfo(skb)->gso_segs = 1;
520 skb_shinfo(skb)->gso_size = 0;
521 skb_shinfo(skb)->gso_type = 0;
525 factor = skb->len + (mss_now - 1);
527 skb_shinfo(skb)->gso_segs = factor;
528 skb_shinfo(skb)->gso_size = mss_now;
529 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
533 /* Function to create two new TCP segments. Shrinks the given segment
534 * to the specified size and appends a new segment with the rest of the
535 * packet to the list. This won't be called frequently, I hope.
536 * Remember, these are still headerless SKBs at this point.
538 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
540 struct tcp_sock *tp = tcp_sk(sk);
541 struct sk_buff *buff;
542 int nsize, old_factor;
546 BUG_ON(len > skb->len);
548 clear_all_retrans_hints(tp);
549 nsize = skb_headlen(skb) - len;
553 if (skb_cloned(skb) &&
554 skb_is_nonlinear(skb) &&
555 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
558 /* Get a new skb... force flag on. */
559 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
561 return -ENOMEM; /* We'll just try again later. */
563 sk_charge_skb(sk, buff);
564 nlen = skb->len - len - nsize;
565 buff->truesize += nlen;
566 skb->truesize -= nlen;
568 /* Correct the sequence numbers. */
569 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
570 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
571 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
573 /* PSH and FIN should only be set in the second packet. */
574 flags = TCP_SKB_CB(skb)->flags;
575 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
576 TCP_SKB_CB(buff)->flags = flags;
577 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
578 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
580 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
581 /* Copy and checksum data tail into the new buffer. */
582 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
587 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
589 skb->ip_summed = CHECKSUM_HW;
590 skb_split(skb, buff, len);
593 buff->ip_summed = skb->ip_summed;
595 /* Looks stupid, but our code really uses when of
596 * skbs, which it never sent before. --ANK
598 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
599 buff->tstamp = skb->tstamp;
601 old_factor = tcp_skb_pcount(skb);
603 /* Fix up tso_factor for both original and new SKB. */
604 tcp_set_skb_tso_segs(sk, skb, mss_now);
605 tcp_set_skb_tso_segs(sk, buff, mss_now);
607 /* If this packet has been sent out already, we must
608 * adjust the various packet counters.
610 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
611 int diff = old_factor - tcp_skb_pcount(skb) -
612 tcp_skb_pcount(buff);
614 tp->packets_out -= diff;
616 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
617 tp->sacked_out -= diff;
618 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
619 tp->retrans_out -= diff;
621 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
622 tp->lost_out -= diff;
623 tp->left_out -= diff;
627 /* Adjust Reno SACK estimate. */
628 if (!tp->rx_opt.sack_ok) {
629 tp->sacked_out -= diff;
630 if ((int)tp->sacked_out < 0)
632 tcp_sync_left_out(tp);
635 tp->fackets_out -= diff;
636 if ((int)tp->fackets_out < 0)
641 /* Link BUFF into the send queue. */
642 skb_header_release(buff);
643 __skb_append(skb, buff, &sk->sk_write_queue);
648 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
649 * eventually). The difference is that pulled data not copied, but
650 * immediately discarded.
652 static void __pskb_trim_head(struct sk_buff *skb, int len)
658 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
659 if (skb_shinfo(skb)->frags[i].size <= eat) {
660 put_page(skb_shinfo(skb)->frags[i].page);
661 eat -= skb_shinfo(skb)->frags[i].size;
663 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
665 skb_shinfo(skb)->frags[k].page_offset += eat;
666 skb_shinfo(skb)->frags[k].size -= eat;
672 skb_shinfo(skb)->nr_frags = k;
674 skb->tail = skb->data;
675 skb->data_len -= len;
676 skb->len = skb->data_len;
679 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
681 if (skb_cloned(skb) &&
682 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
685 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
686 if (unlikely(len < skb_headlen(skb)))
687 __skb_pull(skb, len);
689 __pskb_trim_head(skb, len - skb_headlen(skb));
691 TCP_SKB_CB(skb)->seq += len;
692 skb->ip_summed = CHECKSUM_HW;
694 skb->truesize -= len;
695 sk->sk_wmem_queued -= len;
696 sk->sk_forward_alloc += len;
697 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
699 /* Any change of skb->len requires recalculation of tso
702 if (tcp_skb_pcount(skb) > 1)
703 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
708 /* Not accounting for SACKs here. */
709 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
711 struct tcp_sock *tp = tcp_sk(sk);
712 struct inet_connection_sock *icsk = inet_csk(sk);
715 /* Calculate base mss without TCP options:
716 It is MMS_S - sizeof(tcphdr) of rfc1122
718 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
720 /* Clamp it (mss_clamp does not include tcp options) */
721 if (mss_now > tp->rx_opt.mss_clamp)
722 mss_now = tp->rx_opt.mss_clamp;
724 /* Now subtract optional transport overhead */
725 mss_now -= icsk->icsk_ext_hdr_len;
727 /* Then reserve room for full set of TCP options and 8 bytes of data */
731 /* Now subtract TCP options size, not including SACKs */
732 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
737 /* Inverse of above */
738 int tcp_mss_to_mtu(struct sock *sk, int mss)
740 struct tcp_sock *tp = tcp_sk(sk);
741 struct inet_connection_sock *icsk = inet_csk(sk);
746 icsk->icsk_ext_hdr_len +
747 icsk->icsk_af_ops->net_header_len;
752 void tcp_mtup_init(struct sock *sk)
754 struct tcp_sock *tp = tcp_sk(sk);
755 struct inet_connection_sock *icsk = inet_csk(sk);
757 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
758 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
759 icsk->icsk_af_ops->net_header_len;
760 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
761 icsk->icsk_mtup.probe_size = 0;
764 /* This function synchronize snd mss to current pmtu/exthdr set.
766 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
767 for TCP options, but includes only bare TCP header.
769 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
770 It is minimum of user_mss and mss received with SYN.
771 It also does not include TCP options.
773 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
775 tp->mss_cache is current effective sending mss, including
776 all tcp options except for SACKs. It is evaluated,
777 taking into account current pmtu, but never exceeds
778 tp->rx_opt.mss_clamp.
780 NOTE1. rfc1122 clearly states that advertised MSS
781 DOES NOT include either tcp or ip options.
783 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
784 are READ ONLY outside this function. --ANK (980731)
787 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
789 struct tcp_sock *tp = tcp_sk(sk);
790 struct inet_connection_sock *icsk = inet_csk(sk);
793 if (icsk->icsk_mtup.search_high > pmtu)
794 icsk->icsk_mtup.search_high = pmtu;
796 mss_now = tcp_mtu_to_mss(sk, pmtu);
798 /* Bound mss with half of window */
799 if (tp->max_window && mss_now > (tp->max_window>>1))
800 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
802 /* And store cached results */
803 icsk->icsk_pmtu_cookie = pmtu;
804 if (icsk->icsk_mtup.enabled)
805 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
806 tp->mss_cache = mss_now;
811 /* Compute the current effective MSS, taking SACKs and IP options,
812 * and even PMTU discovery events into account.
814 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
815 * cannot be large. However, taking into account rare use of URG, this
818 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
820 struct tcp_sock *tp = tcp_sk(sk);
821 struct dst_entry *dst = __sk_dst_get(sk);
826 mss_now = tp->mss_cache;
828 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
832 u32 mtu = dst_mtu(dst);
833 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
834 mss_now = tcp_sync_mss(sk, mtu);
837 if (tp->rx_opt.eff_sacks)
838 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
839 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
841 xmit_size_goal = mss_now;
844 xmit_size_goal = (65535 -
845 inet_csk(sk)->icsk_af_ops->net_header_len -
846 inet_csk(sk)->icsk_ext_hdr_len -
849 if (tp->max_window &&
850 (xmit_size_goal > (tp->max_window >> 1)))
851 xmit_size_goal = max((tp->max_window >> 1),
852 68U - tp->tcp_header_len);
854 xmit_size_goal -= (xmit_size_goal % mss_now);
856 tp->xmit_size_goal = xmit_size_goal;
861 EXPORT_SYMBOL_GPL(tcp_current_mss);
863 /* Congestion window validation. (RFC2861) */
865 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
867 __u32 packets_out = tp->packets_out;
869 if (packets_out >= tp->snd_cwnd) {
870 /* Network is feed fully. */
871 tp->snd_cwnd_used = 0;
872 tp->snd_cwnd_stamp = tcp_time_stamp;
874 /* Network starves. */
875 if (tp->packets_out > tp->snd_cwnd_used)
876 tp->snd_cwnd_used = tp->packets_out;
878 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
879 tcp_cwnd_application_limited(sk);
883 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
885 u32 window, cwnd_len;
887 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
888 cwnd_len = mss_now * cwnd;
889 return min(window, cwnd_len);
892 /* Can at least one segment of SKB be sent right now, according to the
893 * congestion window rules? If so, return how many segments are allowed.
895 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
899 /* Don't be strict about the congestion window for the final FIN. */
900 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
903 in_flight = tcp_packets_in_flight(tp);
905 if (in_flight < cwnd)
906 return (cwnd - in_flight);
911 /* This must be invoked the first time we consider transmitting
914 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
916 int tso_segs = tcp_skb_pcount(skb);
920 tcp_skb_mss(skb) != mss_now)) {
921 tcp_set_skb_tso_segs(sk, skb, mss_now);
922 tso_segs = tcp_skb_pcount(skb);
927 static inline int tcp_minshall_check(const struct tcp_sock *tp)
929 return after(tp->snd_sml,tp->snd_una) &&
930 !after(tp->snd_sml, tp->snd_nxt);
933 /* Return 0, if packet can be sent now without violation Nagle's rules:
934 * 1. It is full sized.
935 * 2. Or it contains FIN. (already checked by caller)
936 * 3. Or TCP_NODELAY was set.
937 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
938 * With Minshall's modification: all sent small packets are ACKed.
941 static inline int tcp_nagle_check(const struct tcp_sock *tp,
942 const struct sk_buff *skb,
943 unsigned mss_now, int nonagle)
945 return (skb->len < mss_now &&
946 ((nonagle&TCP_NAGLE_CORK) ||
949 tcp_minshall_check(tp))));
952 /* Return non-zero if the Nagle test allows this packet to be
955 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
956 unsigned int cur_mss, int nonagle)
958 /* Nagle rule does not apply to frames, which sit in the middle of the
959 * write_queue (they have no chances to get new data).
961 * This is implemented in the callers, where they modify the 'nonagle'
962 * argument based upon the location of SKB in the send queue.
964 if (nonagle & TCP_NAGLE_PUSH)
967 /* Don't use the nagle rule for urgent data (or for the final FIN). */
969 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
972 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
978 /* Does at least the first segment of SKB fit into the send window? */
979 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
981 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
983 if (skb->len > cur_mss)
984 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
986 return !after(end_seq, tp->snd_una + tp->snd_wnd);
989 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
990 * should be put on the wire right now. If so, it returns the number of
991 * packets allowed by the congestion window.
993 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
994 unsigned int cur_mss, int nonagle)
996 struct tcp_sock *tp = tcp_sk(sk);
997 unsigned int cwnd_quota;
999 tcp_init_tso_segs(sk, skb, cur_mss);
1001 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1004 cwnd_quota = tcp_cwnd_test(tp, skb);
1006 !tcp_snd_wnd_test(tp, skb, cur_mss))
1012 static inline int tcp_skb_is_last(const struct sock *sk,
1013 const struct sk_buff *skb)
1015 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1018 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1020 struct sk_buff *skb = sk->sk_send_head;
1023 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1024 (tcp_skb_is_last(sk, skb) ?
1029 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1030 * which is put after SKB on the list. It is very much like
1031 * tcp_fragment() except that it may make several kinds of assumptions
1032 * in order to speed up the splitting operation. In particular, we
1033 * know that all the data is in scatter-gather pages, and that the
1034 * packet has never been sent out before (and thus is not cloned).
1036 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1038 struct sk_buff *buff;
1039 int nlen = skb->len - len;
1042 /* All of a TSO frame must be composed of paged data. */
1043 if (skb->len != skb->data_len)
1044 return tcp_fragment(sk, skb, len, mss_now);
1046 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1047 if (unlikely(buff == NULL))
1050 sk_charge_skb(sk, buff);
1051 buff->truesize += nlen;
1052 skb->truesize -= nlen;
1054 /* Correct the sequence numbers. */
1055 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1056 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1057 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1059 /* PSH and FIN should only be set in the second packet. */
1060 flags = TCP_SKB_CB(skb)->flags;
1061 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1062 TCP_SKB_CB(buff)->flags = flags;
1064 /* This packet was never sent out yet, so no SACK bits. */
1065 TCP_SKB_CB(buff)->sacked = 0;
1067 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
1068 skb_split(skb, buff, len);
1070 /* Fix up tso_factor for both original and new SKB. */
1071 tcp_set_skb_tso_segs(sk, skb, mss_now);
1072 tcp_set_skb_tso_segs(sk, buff, mss_now);
1074 /* Link BUFF into the send queue. */
1075 skb_header_release(buff);
1076 __skb_append(skb, buff, &sk->sk_write_queue);
1081 /* Try to defer sending, if possible, in order to minimize the amount
1082 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1084 * This algorithm is from John Heffner.
1086 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1088 const struct inet_connection_sock *icsk = inet_csk(sk);
1089 u32 send_win, cong_win, limit, in_flight;
1091 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1094 if (icsk->icsk_ca_state != TCP_CA_Open)
1097 in_flight = tcp_packets_in_flight(tp);
1099 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1100 (tp->snd_cwnd <= in_flight));
1102 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1104 /* From in_flight test above, we know that cwnd > in_flight. */
1105 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1107 limit = min(send_win, cong_win);
1109 /* If a full-sized TSO skb can be sent, do it. */
1113 if (sysctl_tcp_tso_win_divisor) {
1114 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1116 /* If at least some fraction of a window is available,
1119 chunk /= sysctl_tcp_tso_win_divisor;
1123 /* Different approach, try not to defer past a single
1124 * ACK. Receiver should ACK every other full sized
1125 * frame, so if we have space for more than 3 frames
1128 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1132 /* Ok, it looks like it is advisable to defer. */
1136 /* Create a new MTU probe if we are ready.
1137 * Returns 0 if we should wait to probe (no cwnd available),
1138 * 1 if a probe was sent,
1140 static int tcp_mtu_probe(struct sock *sk)
1142 struct tcp_sock *tp = tcp_sk(sk);
1143 struct inet_connection_sock *icsk = inet_csk(sk);
1144 struct sk_buff *skb, *nskb, *next;
1151 /* Not currently probing/verifying,
1153 * have enough cwnd, and
1154 * not SACKing (the variable headers throw things off) */
1155 if (!icsk->icsk_mtup.enabled ||
1156 icsk->icsk_mtup.probe_size ||
1157 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1158 tp->snd_cwnd < 11 ||
1159 tp->rx_opt.eff_sacks)
1162 /* Very simple search strategy: just double the MSS. */
1163 mss_now = tcp_current_mss(sk, 0);
1164 probe_size = 2*tp->mss_cache;
1165 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1166 /* TODO: set timer for probe_converge_event */
1170 /* Have enough data in the send queue to probe? */
1172 if ((skb = sk->sk_send_head) == NULL)
1174 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1176 if (len < probe_size)
1179 /* Receive window check. */
1180 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1181 if (tp->snd_wnd < probe_size)
1187 /* Do we need to wait to drain cwnd? */
1188 pif = tcp_packets_in_flight(tp);
1189 if (pif + 2 > tp->snd_cwnd) {
1190 /* With no packets in flight, don't stall. */
1197 /* We're allowed to probe. Build it now. */
1198 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1200 sk_charge_skb(sk, nskb);
1202 skb = sk->sk_send_head;
1203 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1204 sk->sk_send_head = nskb;
1206 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1207 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1208 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1209 TCP_SKB_CB(nskb)->sacked = 0;
1211 if (skb->ip_summed == CHECKSUM_HW)
1212 nskb->ip_summed = CHECKSUM_HW;
1215 while (len < probe_size) {
1218 copy = min_t(int, skb->len, probe_size - len);
1219 if (nskb->ip_summed)
1220 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1222 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1223 skb_put(nskb, copy), copy, nskb->csum);
1225 if (skb->len <= copy) {
1226 /* We've eaten all the data from this skb.
1228 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1229 __skb_unlink(skb, &sk->sk_write_queue);
1230 sk_stream_free_skb(sk, skb);
1232 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1233 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1234 if (!skb_shinfo(skb)->nr_frags) {
1235 skb_pull(skb, copy);
1236 if (skb->ip_summed != CHECKSUM_HW)
1237 skb->csum = csum_partial(skb->data, skb->len, 0);
1239 __pskb_trim_head(skb, copy);
1240 tcp_set_skb_tso_segs(sk, skb, mss_now);
1242 TCP_SKB_CB(skb)->seq += copy;
1248 tcp_init_tso_segs(sk, nskb, nskb->len);
1250 /* We're ready to send. If this fails, the probe will
1251 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1252 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1253 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1254 /* Decrement cwnd here because we are sending
1255 * effectively two packets. */
1257 update_send_head(sk, tp, nskb);
1259 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1260 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1261 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1270 /* This routine writes packets to the network. It advances the
1271 * send_head. This happens as incoming acks open up the remote
1274 * Returns 1, if no segments are in flight and we have queued segments, but
1275 * cannot send anything now because of SWS or another problem.
1277 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1279 struct tcp_sock *tp = tcp_sk(sk);
1280 struct sk_buff *skb;
1281 unsigned int tso_segs, sent_pkts;
1285 /* If we are closed, the bytes will have to remain here.
1286 * In time closedown will finish, we empty the write queue and all
1289 if (unlikely(sk->sk_state == TCP_CLOSE))
1294 /* Do MTU probing. */
1295 if ((result = tcp_mtu_probe(sk)) == 0) {
1297 } else if (result > 0) {
1301 while ((skb = sk->sk_send_head)) {
1304 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1307 cwnd_quota = tcp_cwnd_test(tp, skb);
1311 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1314 if (tso_segs == 1) {
1315 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1316 (tcp_skb_is_last(sk, skb) ?
1317 nonagle : TCP_NAGLE_PUSH))))
1320 if (tcp_tso_should_defer(sk, tp, skb))
1326 limit = tcp_window_allows(tp, skb,
1327 mss_now, cwnd_quota);
1329 if (skb->len < limit) {
1330 unsigned int trim = skb->len % mss_now;
1333 limit = skb->len - trim;
1337 if (skb->len > limit &&
1338 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1341 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1343 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1346 /* Advance the send_head. This one is sent out.
1347 * This call will increment packets_out.
1349 update_send_head(sk, tp, skb);
1351 tcp_minshall_update(tp, mss_now, skb);
1355 if (likely(sent_pkts)) {
1356 tcp_cwnd_validate(sk, tp);
1359 return !tp->packets_out && sk->sk_send_head;
1362 /* Push out any pending frames which were held back due to
1363 * TCP_CORK or attempt at coalescing tiny packets.
1364 * The socket must be locked by the caller.
1366 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1367 unsigned int cur_mss, int nonagle)
1369 struct sk_buff *skb = sk->sk_send_head;
1372 if (tcp_write_xmit(sk, cur_mss, nonagle))
1373 tcp_check_probe_timer(sk, tp);
1376 EXPORT_SYMBOL_GPL(__tcp_push_pending_frames);
1378 /* Send _single_ skb sitting at the send head. This function requires
1379 * true push pending frames to setup probe timer etc.
1381 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1383 struct tcp_sock *tp = tcp_sk(sk);
1384 struct sk_buff *skb = sk->sk_send_head;
1385 unsigned int tso_segs, cwnd_quota;
1387 BUG_ON(!skb || skb->len < mss_now);
1389 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1390 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1392 if (likely(cwnd_quota)) {
1399 limit = tcp_window_allows(tp, skb,
1400 mss_now, cwnd_quota);
1402 if (skb->len < limit) {
1403 unsigned int trim = skb->len % mss_now;
1406 limit = skb->len - trim;
1410 if (skb->len > limit &&
1411 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1414 /* Send it out now. */
1415 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1417 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1418 update_send_head(sk, tp, skb);
1419 tcp_cwnd_validate(sk, tp);
1425 /* This function returns the amount that we can raise the
1426 * usable window based on the following constraints
1428 * 1. The window can never be shrunk once it is offered (RFC 793)
1429 * 2. We limit memory per socket
1432 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1433 * RECV.NEXT + RCV.WIN fixed until:
1434 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1436 * i.e. don't raise the right edge of the window until you can raise
1437 * it at least MSS bytes.
1439 * Unfortunately, the recommended algorithm breaks header prediction,
1440 * since header prediction assumes th->window stays fixed.
1442 * Strictly speaking, keeping th->window fixed violates the receiver
1443 * side SWS prevention criteria. The problem is that under this rule
1444 * a stream of single byte packets will cause the right side of the
1445 * window to always advance by a single byte.
1447 * Of course, if the sender implements sender side SWS prevention
1448 * then this will not be a problem.
1450 * BSD seems to make the following compromise:
1452 * If the free space is less than the 1/4 of the maximum
1453 * space available and the free space is less than 1/2 mss,
1454 * then set the window to 0.
1455 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1456 * Otherwise, just prevent the window from shrinking
1457 * and from being larger than the largest representable value.
1459 * This prevents incremental opening of the window in the regime
1460 * where TCP is limited by the speed of the reader side taking
1461 * data out of the TCP receive queue. It does nothing about
1462 * those cases where the window is constrained on the sender side
1463 * because the pipeline is full.
1465 * BSD also seems to "accidentally" limit itself to windows that are a
1466 * multiple of MSS, at least until the free space gets quite small.
1467 * This would appear to be a side effect of the mbuf implementation.
1468 * Combining these two algorithms results in the observed behavior
1469 * of having a fixed window size at almost all times.
1471 * Below we obtain similar behavior by forcing the offered window to
1472 * a multiple of the mss when it is feasible to do so.
1474 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1475 * Regular options like TIMESTAMP are taken into account.
1477 u32 __tcp_select_window(struct sock *sk)
1479 struct inet_connection_sock *icsk = inet_csk(sk);
1480 struct tcp_sock *tp = tcp_sk(sk);
1481 /* MSS for the peer's data. Previous versions used mss_clamp
1482 * here. I don't know if the value based on our guesses
1483 * of peer's MSS is better for the performance. It's more correct
1484 * but may be worse for the performance because of rcv_mss
1485 * fluctuations. --SAW 1998/11/1
1487 int mss = icsk->icsk_ack.rcv_mss;
1488 int free_space = tcp_space(sk);
1489 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1492 if (mss > full_space)
1495 if (free_space < full_space/2) {
1496 icsk->icsk_ack.quick = 0;
1498 if (tcp_memory_pressure)
1499 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1501 if (free_space < mss)
1505 if (free_space > tp->rcv_ssthresh)
1506 free_space = tp->rcv_ssthresh;
1508 /* Don't do rounding if we are using window scaling, since the
1509 * scaled window will not line up with the MSS boundary anyway.
1511 window = tp->rcv_wnd;
1512 if (tp->rx_opt.rcv_wscale) {
1513 window = free_space;
1515 /* Advertise enough space so that it won't get scaled away.
1516 * Import case: prevent zero window announcement if
1517 * 1<<rcv_wscale > mss.
1519 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1520 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1521 << tp->rx_opt.rcv_wscale);
1523 /* Get the largest window that is a nice multiple of mss.
1524 * Window clamp already applied above.
1525 * If our current window offering is within 1 mss of the
1526 * free space we just keep it. This prevents the divide
1527 * and multiply from happening most of the time.
1528 * We also don't do any window rounding when the free space
1531 if (window <= free_space - mss || window > free_space)
1532 window = (free_space/mss)*mss;
1538 /* Attempt to collapse two adjacent SKB's during retransmission. */
1539 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 struct sk_buff *next_skb = skb->next;
1544 /* The first test we must make is that neither of these two
1545 * SKB's are still referenced by someone else.
1547 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1548 int skb_size = skb->len, next_skb_size = next_skb->len;
1549 u16 flags = TCP_SKB_CB(skb)->flags;
1551 /* Also punt if next skb has been SACK'd. */
1552 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1555 /* Next skb is out of window. */
1556 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1559 /* Punt if not enough space exists in the first SKB for
1560 * the data in the second, or the total combined payload
1561 * would exceed the MSS.
1563 if ((next_skb_size > skb_tailroom(skb)) ||
1564 ((skb_size + next_skb_size) > mss_now))
1567 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1568 tcp_skb_pcount(next_skb) != 1);
1570 /* changing transmit queue under us so clear hints */
1571 clear_all_retrans_hints(tp);
1573 /* Ok. We will be able to collapse the packet. */
1574 __skb_unlink(next_skb, &sk->sk_write_queue);
1576 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1578 if (next_skb->ip_summed == CHECKSUM_HW)
1579 skb->ip_summed = CHECKSUM_HW;
1581 if (skb->ip_summed != CHECKSUM_HW)
1582 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1584 /* Update sequence range on original skb. */
1585 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1587 /* Merge over control information. */
1588 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1589 TCP_SKB_CB(skb)->flags = flags;
1591 /* All done, get rid of second SKB and account for it so
1592 * packet counting does not break.
1594 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1595 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1596 tp->retrans_out -= tcp_skb_pcount(next_skb);
1597 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1598 tp->lost_out -= tcp_skb_pcount(next_skb);
1599 tp->left_out -= tcp_skb_pcount(next_skb);
1601 /* Reno case is special. Sigh... */
1602 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1603 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1604 tp->left_out -= tcp_skb_pcount(next_skb);
1607 /* Not quite right: it can be > snd.fack, but
1608 * it is better to underestimate fackets.
1610 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1611 tcp_packets_out_dec(tp, next_skb);
1612 sk_stream_free_skb(sk, next_skb);
1616 /* Do a simple retransmit without using the backoff mechanisms in
1617 * tcp_timer. This is used for path mtu discovery.
1618 * The socket is already locked here.
1620 void tcp_simple_retransmit(struct sock *sk)
1622 const struct inet_connection_sock *icsk = inet_csk(sk);
1623 struct tcp_sock *tp = tcp_sk(sk);
1624 struct sk_buff *skb;
1625 unsigned int mss = tcp_current_mss(sk, 0);
1628 sk_stream_for_retrans_queue(skb, sk) {
1629 if (skb->len > mss &&
1630 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1631 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1632 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1633 tp->retrans_out -= tcp_skb_pcount(skb);
1635 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1636 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1637 tp->lost_out += tcp_skb_pcount(skb);
1643 clear_all_retrans_hints(tp);
1648 tcp_sync_left_out(tp);
1650 /* Don't muck with the congestion window here.
1651 * Reason is that we do not increase amount of _data_
1652 * in network, but units changed and effective
1653 * cwnd/ssthresh really reduced now.
1655 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1656 tp->high_seq = tp->snd_nxt;
1657 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1658 tp->prior_ssthresh = 0;
1659 tp->undo_marker = 0;
1660 tcp_set_ca_state(sk, TCP_CA_Loss);
1662 tcp_xmit_retransmit_queue(sk);
1665 /* This retransmits one SKB. Policy decisions and retransmit queue
1666 * state updates are done by the caller. Returns non-zero if an
1667 * error occurred which prevented the send.
1669 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1671 struct tcp_sock *tp = tcp_sk(sk);
1672 struct inet_connection_sock *icsk = inet_csk(sk);
1673 unsigned int cur_mss = tcp_current_mss(sk, 0);
1676 /* Inconslusive MTU probe */
1677 if (icsk->icsk_mtup.probe_size) {
1678 icsk->icsk_mtup.probe_size = 0;
1681 /* Do not sent more than we queued. 1/4 is reserved for possible
1682 * copying overhead: fragmentation, tunneling, mangling etc.
1684 if (atomic_read(&sk->sk_wmem_alloc) >
1685 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1688 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1689 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1691 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1695 /* If receiver has shrunk his window, and skb is out of
1696 * new window, do not retransmit it. The exception is the
1697 * case, when window is shrunk to zero. In this case
1698 * our retransmit serves as a zero window probe.
1700 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1701 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1704 if (skb->len > cur_mss) {
1705 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1706 return -ENOMEM; /* We'll try again later. */
1709 /* Collapse two adjacent packets if worthwhile and we can. */
1710 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1711 (skb->len < (cur_mss >> 1)) &&
1712 (skb->next != sk->sk_send_head) &&
1713 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1714 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1715 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1716 (sysctl_tcp_retrans_collapse != 0))
1717 tcp_retrans_try_collapse(sk, skb, cur_mss);
1719 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1720 return -EHOSTUNREACH; /* Routing failure or similar. */
1722 /* Some Solaris stacks overoptimize and ignore the FIN on a
1723 * retransmit when old data is attached. So strip it off
1724 * since it is cheap to do so and saves bytes on the network.
1727 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1728 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1729 if (!pskb_trim(skb, 0)) {
1730 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1731 skb_shinfo(skb)->gso_segs = 1;
1732 skb_shinfo(skb)->gso_size = 0;
1733 skb_shinfo(skb)->gso_type = 0;
1734 skb->ip_summed = CHECKSUM_NONE;
1739 /* Make a copy, if the first transmission SKB clone we made
1740 * is still in somebody's hands, else make a clone.
1742 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1744 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1747 /* Update global TCP statistics. */
1748 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1750 tp->total_retrans++;
1752 #if FASTRETRANS_DEBUG > 0
1753 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1754 if (net_ratelimit())
1755 printk(KERN_DEBUG "retrans_out leaked.\n");
1758 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1759 tp->retrans_out += tcp_skb_pcount(skb);
1761 /* Save stamp of the first retransmit. */
1762 if (!tp->retrans_stamp)
1763 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1767 /* snd_nxt is stored to detect loss of retransmitted segment,
1768 * see tcp_input.c tcp_sacktag_write_queue().
1770 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1775 /* This gets called after a retransmit timeout, and the initially
1776 * retransmitted data is acknowledged. It tries to continue
1777 * resending the rest of the retransmit queue, until either
1778 * we've sent it all or the congestion window limit is reached.
1779 * If doing SACK, the first ACK which comes back for a timeout
1780 * based retransmit packet might feed us FACK information again.
1781 * If so, we use it to avoid unnecessarily retransmissions.
1783 void tcp_xmit_retransmit_queue(struct sock *sk)
1785 const struct inet_connection_sock *icsk = inet_csk(sk);
1786 struct tcp_sock *tp = tcp_sk(sk);
1787 struct sk_buff *skb;
1790 if (tp->retransmit_skb_hint) {
1791 skb = tp->retransmit_skb_hint;
1792 packet_cnt = tp->retransmit_cnt_hint;
1794 skb = sk->sk_write_queue.next;
1798 /* First pass: retransmit lost packets. */
1800 sk_stream_for_retrans_queue_from(skb, sk) {
1801 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1803 /* we could do better than to assign each time */
1804 tp->retransmit_skb_hint = skb;
1805 tp->retransmit_cnt_hint = packet_cnt;
1807 /* Assume this retransmit will generate
1808 * only one packet for congestion window
1809 * calculation purposes. This works because
1810 * tcp_retransmit_skb() will chop up the
1811 * packet to be MSS sized and all the
1812 * packet counting works out.
1814 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1817 if (sacked & TCPCB_LOST) {
1818 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1819 if (tcp_retransmit_skb(sk, skb)) {
1820 tp->retransmit_skb_hint = NULL;
1823 if (icsk->icsk_ca_state != TCP_CA_Loss)
1824 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1826 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1829 skb_peek(&sk->sk_write_queue))
1830 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1831 inet_csk(sk)->icsk_rto,
1835 packet_cnt += tcp_skb_pcount(skb);
1836 if (packet_cnt >= tp->lost_out)
1842 /* OK, demanded retransmission is finished. */
1844 /* Forward retransmissions are possible only during Recovery. */
1845 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1848 /* No forward retransmissions in Reno are possible. */
1849 if (!tp->rx_opt.sack_ok)
1852 /* Yeah, we have to make difficult choice between forward transmission
1853 * and retransmission... Both ways have their merits...
1855 * For now we do not retransmit anything, while we have some new
1859 if (tcp_may_send_now(sk, tp))
1862 if (tp->forward_skb_hint) {
1863 skb = tp->forward_skb_hint;
1864 packet_cnt = tp->forward_cnt_hint;
1866 skb = sk->sk_write_queue.next;
1870 sk_stream_for_retrans_queue_from(skb, sk) {
1871 tp->forward_cnt_hint = packet_cnt;
1872 tp->forward_skb_hint = skb;
1874 /* Similar to the retransmit loop above we
1875 * can pretend that the retransmitted SKB
1876 * we send out here will be composed of one
1877 * real MSS sized packet because tcp_retransmit_skb()
1878 * will fragment it if necessary.
1880 if (++packet_cnt > tp->fackets_out)
1883 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1886 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1889 /* Ok, retransmit it. */
1890 if (tcp_retransmit_skb(sk, skb)) {
1891 tp->forward_skb_hint = NULL;
1895 if (skb == skb_peek(&sk->sk_write_queue))
1896 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1897 inet_csk(sk)->icsk_rto,
1900 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1905 /* Send a fin. The caller locks the socket for us. This cannot be
1906 * allowed to fail queueing a FIN frame under any circumstances.
1908 void tcp_send_fin(struct sock *sk)
1910 struct tcp_sock *tp = tcp_sk(sk);
1911 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1914 /* Optimization, tack on the FIN if we have a queue of
1915 * unsent frames. But be careful about outgoing SACKS
1918 mss_now = tcp_current_mss(sk, 1);
1920 if (sk->sk_send_head != NULL) {
1921 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1922 TCP_SKB_CB(skb)->end_seq++;
1925 /* Socket is locked, keep trying until memory is available. */
1927 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1933 /* Reserve space for headers and prepare control bits. */
1934 skb_reserve(skb, MAX_TCP_HEADER);
1936 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1937 TCP_SKB_CB(skb)->sacked = 0;
1938 skb_shinfo(skb)->gso_segs = 1;
1939 skb_shinfo(skb)->gso_size = 0;
1940 skb_shinfo(skb)->gso_type = 0;
1942 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1943 TCP_SKB_CB(skb)->seq = tp->write_seq;
1944 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1945 tcp_queue_skb(sk, skb);
1947 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1950 /* We get here when a process closes a file descriptor (either due to
1951 * an explicit close() or as a byproduct of exit()'ing) and there
1952 * was unread data in the receive queue. This behavior is recommended
1953 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1955 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 struct sk_buff *skb;
1960 /* NOTE: No TCP options attached and we never retransmit this. */
1961 skb = alloc_skb(MAX_TCP_HEADER, priority);
1963 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1967 /* Reserve space for headers and prepare control bits. */
1968 skb_reserve(skb, MAX_TCP_HEADER);
1970 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1971 TCP_SKB_CB(skb)->sacked = 0;
1972 skb_shinfo(skb)->gso_segs = 1;
1973 skb_shinfo(skb)->gso_size = 0;
1974 skb_shinfo(skb)->gso_type = 0;
1977 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1978 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1979 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1980 if (tcp_transmit_skb(sk, skb, 0, priority))
1981 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1984 /* WARNING: This routine must only be called when we have already sent
1985 * a SYN packet that crossed the incoming SYN that caused this routine
1986 * to get called. If this assumption fails then the initial rcv_wnd
1987 * and rcv_wscale values will not be correct.
1989 int tcp_send_synack(struct sock *sk)
1991 struct sk_buff* skb;
1993 skb = skb_peek(&sk->sk_write_queue);
1994 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1995 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1998 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1999 if (skb_cloned(skb)) {
2000 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2003 __skb_unlink(skb, &sk->sk_write_queue);
2004 skb_header_release(nskb);
2005 __skb_queue_head(&sk->sk_write_queue, nskb);
2006 sk_stream_free_skb(sk, skb);
2007 sk_charge_skb(sk, nskb);
2011 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2012 TCP_ECN_send_synack(tcp_sk(sk), skb);
2014 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2015 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2019 * Prepare a SYN-ACK.
2021 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2022 struct request_sock *req)
2024 struct inet_request_sock *ireq = inet_rsk(req);
2025 struct tcp_sock *tp = tcp_sk(sk);
2027 int tcp_header_size;
2028 struct sk_buff *skb;
2030 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2034 /* Reserve space for headers. */
2035 skb_reserve(skb, MAX_TCP_HEADER);
2037 skb->dst = dst_clone(dst);
2039 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2040 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2041 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2042 /* SACK_PERM is in the place of NOP NOP of TS */
2043 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2044 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2046 memset(th, 0, sizeof(struct tcphdr));
2049 TCP_ECN_make_synack(req, th);
2050 th->source = inet_sk(sk)->sport;
2051 th->dest = ireq->rmt_port;
2052 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2053 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2054 TCP_SKB_CB(skb)->sacked = 0;
2055 skb_shinfo(skb)->gso_segs = 1;
2056 skb_shinfo(skb)->gso_size = 0;
2057 skb_shinfo(skb)->gso_type = 0;
2058 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2059 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2060 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2062 /* Set this up on the first call only */
2063 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2064 /* tcp_full_space because it is guaranteed to be the first packet */
2065 tcp_select_initial_window(tcp_full_space(sk),
2066 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2071 ireq->rcv_wscale = rcv_wscale;
2074 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2075 th->window = htons(req->rcv_wnd);
2077 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2078 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2079 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2080 TCP_SKB_CB(skb)->when,
2084 th->doff = (tcp_header_size >> 2);
2085 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2090 * Do all connect socket setups that can be done AF independent.
2092 static void tcp_connect_init(struct sock *sk)
2094 struct dst_entry *dst = __sk_dst_get(sk);
2095 struct tcp_sock *tp = tcp_sk(sk);
2098 /* We'll fix this up when we get a response from the other end.
2099 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2101 tp->tcp_header_len = sizeof(struct tcphdr) +
2102 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2104 /* If user gave his TCP_MAXSEG, record it to clamp */
2105 if (tp->rx_opt.user_mss)
2106 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2109 tcp_sync_mss(sk, dst_mtu(dst));
2111 if (!tp->window_clamp)
2112 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2113 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2114 tcp_initialize_rcv_mss(sk);
2116 tcp_select_initial_window(tcp_full_space(sk),
2117 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2120 sysctl_tcp_window_scaling,
2123 tp->rx_opt.rcv_wscale = rcv_wscale;
2124 tp->rcv_ssthresh = tp->rcv_wnd;
2127 sock_reset_flag(sk, SOCK_DONE);
2129 tcp_init_wl(tp, tp->write_seq, 0);
2130 tp->snd_una = tp->write_seq;
2131 tp->snd_sml = tp->write_seq;
2136 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2137 inet_csk(sk)->icsk_retransmits = 0;
2138 tcp_clear_retrans(tp);
2142 * Build a SYN and send it off.
2144 int tcp_connect(struct sock *sk)
2146 struct tcp_sock *tp = tcp_sk(sk);
2147 struct sk_buff *buff;
2149 tcp_connect_init(sk);
2151 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2152 if (unlikely(buff == NULL))
2155 /* Reserve space for headers. */
2156 skb_reserve(buff, MAX_TCP_HEADER);
2158 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2159 TCP_ECN_send_syn(sk, tp, buff);
2160 TCP_SKB_CB(buff)->sacked = 0;
2161 skb_shinfo(buff)->gso_segs = 1;
2162 skb_shinfo(buff)->gso_size = 0;
2163 skb_shinfo(buff)->gso_type = 0;
2165 tp->snd_nxt = tp->write_seq;
2166 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2167 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2170 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2171 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2172 skb_header_release(buff);
2173 __skb_queue_tail(&sk->sk_write_queue, buff);
2174 sk_charge_skb(sk, buff);
2175 tp->packets_out += tcp_skb_pcount(buff);
2176 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2178 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2179 * in order to make this packet get counted in tcpOutSegs.
2181 tp->snd_nxt = tp->write_seq;
2182 tp->pushed_seq = tp->write_seq;
2183 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2185 /* Timer for repeating the SYN until an answer. */
2186 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2187 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2191 /* Send out a delayed ack, the caller does the policy checking
2192 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2195 void tcp_send_delayed_ack(struct sock *sk)
2197 struct inet_connection_sock *icsk = inet_csk(sk);
2198 int ato = icsk->icsk_ack.ato;
2199 unsigned long timeout;
2201 if (ato > TCP_DELACK_MIN) {
2202 const struct tcp_sock *tp = tcp_sk(sk);
2205 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2206 max_ato = TCP_DELACK_MAX;
2208 /* Slow path, intersegment interval is "high". */
2210 /* If some rtt estimate is known, use it to bound delayed ack.
2211 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2215 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2221 ato = min(ato, max_ato);
2224 /* Stay within the limit we were given */
2225 timeout = jiffies + ato;
2227 /* Use new timeout only if there wasn't a older one earlier. */
2228 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2229 /* If delack timer was blocked or is about to expire,
2232 if (icsk->icsk_ack.blocked ||
2233 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2238 if (!time_before(timeout, icsk->icsk_ack.timeout))
2239 timeout = icsk->icsk_ack.timeout;
2241 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2242 icsk->icsk_ack.timeout = timeout;
2243 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2246 /* This routine sends an ack and also updates the window. */
2247 void tcp_send_ack(struct sock *sk)
2249 /* If we have been reset, we may not send again. */
2250 if (sk->sk_state != TCP_CLOSE) {
2251 struct tcp_sock *tp = tcp_sk(sk);
2252 struct sk_buff *buff;
2254 /* We are not putting this on the write queue, so
2255 * tcp_transmit_skb() will set the ownership to this
2258 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2260 inet_csk_schedule_ack(sk);
2261 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2262 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2263 TCP_DELACK_MAX, TCP_RTO_MAX);
2267 /* Reserve space for headers and prepare control bits. */
2268 skb_reserve(buff, MAX_TCP_HEADER);
2270 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2271 TCP_SKB_CB(buff)->sacked = 0;
2272 skb_shinfo(buff)->gso_segs = 1;
2273 skb_shinfo(buff)->gso_size = 0;
2274 skb_shinfo(buff)->gso_type = 0;
2276 /* Send it off, this clears delayed acks for us. */
2277 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2278 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2279 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2283 /* This routine sends a packet with an out of date sequence
2284 * number. It assumes the other end will try to ack it.
2286 * Question: what should we make while urgent mode?
2287 * 4.4BSD forces sending single byte of data. We cannot send
2288 * out of window data, because we have SND.NXT==SND.MAX...
2290 * Current solution: to send TWO zero-length segments in urgent mode:
2291 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2292 * out-of-date with SND.UNA-1 to probe window.
2294 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb;
2299 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2300 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2304 /* Reserve space for headers and set control bits. */
2305 skb_reserve(skb, MAX_TCP_HEADER);
2307 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2308 TCP_SKB_CB(skb)->sacked = urgent;
2309 skb_shinfo(skb)->gso_segs = 1;
2310 skb_shinfo(skb)->gso_size = 0;
2311 skb_shinfo(skb)->gso_type = 0;
2313 /* Use a previous sequence. This should cause the other
2314 * end to send an ack. Don't queue or clone SKB, just
2317 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2318 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2319 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2320 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2323 int tcp_write_wakeup(struct sock *sk)
2325 if (sk->sk_state != TCP_CLOSE) {
2326 struct tcp_sock *tp = tcp_sk(sk);
2327 struct sk_buff *skb;
2329 if ((skb = sk->sk_send_head) != NULL &&
2330 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2332 unsigned int mss = tcp_current_mss(sk, 0);
2333 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2335 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2336 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2338 /* We are probing the opening of a window
2339 * but the window size is != 0
2340 * must have been a result SWS avoidance ( sender )
2342 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2344 seg_size = min(seg_size, mss);
2345 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2346 if (tcp_fragment(sk, skb, seg_size, mss))
2348 } else if (!tcp_skb_pcount(skb))
2349 tcp_set_skb_tso_segs(sk, skb, mss);
2351 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2352 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2353 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2355 update_send_head(sk, tp, skb);
2360 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2361 tcp_xmit_probe_skb(sk, TCPCB_URG);
2362 return tcp_xmit_probe_skb(sk, 0);
2368 /* A window probe timeout has occurred. If window is not closed send
2369 * a partial packet else a zero probe.
2371 void tcp_send_probe0(struct sock *sk)
2373 struct inet_connection_sock *icsk = inet_csk(sk);
2374 struct tcp_sock *tp = tcp_sk(sk);
2377 err = tcp_write_wakeup(sk);
2379 if (tp->packets_out || !sk->sk_send_head) {
2380 /* Cancel probe timer, if it is not required. */
2381 icsk->icsk_probes_out = 0;
2382 icsk->icsk_backoff = 0;
2387 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2388 icsk->icsk_backoff++;
2389 icsk->icsk_probes_out++;
2390 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2391 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2394 /* If packet was not sent due to local congestion,
2395 * do not backoff and do not remember icsk_probes_out.
2396 * Let local senders to fight for local resources.
2398 * Use accumulated backoff yet.
2400 if (!icsk->icsk_probes_out)
2401 icsk->icsk_probes_out = 1;
2402 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2403 min(icsk->icsk_rto << icsk->icsk_backoff,
2404 TCP_RESOURCE_PROBE_INTERVAL),
2409 EXPORT_SYMBOL(tcp_connect);
2410 EXPORT_SYMBOL(tcp_make_synack);
2411 EXPORT_SYMBOL(tcp_simple_retransmit);
2412 EXPORT_SYMBOL(tcp_sync_mss);
2413 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2414 EXPORT_SYMBOL(tcp_mtup_init);