2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows = 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor = 3;
59 int sysctl_tcp_mtu_probing = 0;
60 int sysctl_tcp_base_mss = 512;
62 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
65 sk->sk_send_head = skb->next;
66 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
67 sk->sk_send_head = NULL;
68 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
69 tcp_packets_out_inc(sk, tp, skb);
72 /* SND.NXT, if window was not shrunk.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
78 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
80 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
83 return tp->snd_una+tp->snd_wnd;
86 /* Calculate mss to advertise in SYN segment.
87 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
89 * 1. It is independent of path mtu.
90 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
91 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
92 * attached devices, because some buggy hosts are confused by
94 * 4. We do not make 3, we advertise MSS, calculated from first
95 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
96 * This may be overridden via information stored in routing table.
97 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
98 * probably even Jumbo".
100 static __u16 tcp_advertise_mss(struct sock *sk)
102 struct tcp_sock *tp = tcp_sk(sk);
103 struct dst_entry *dst = __sk_dst_get(sk);
104 int mss = tp->advmss;
106 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
107 mss = dst_metric(dst, RTAX_ADVMSS);
114 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
115 * This is the first part of cwnd validation mechanism. */
116 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
118 struct tcp_sock *tp = tcp_sk(sk);
119 s32 delta = tcp_time_stamp - tp->lsndtime;
120 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
121 u32 cwnd = tp->snd_cwnd;
123 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
125 tp->snd_ssthresh = tcp_current_ssthresh(sk);
126 restart_cwnd = min(restart_cwnd, cwnd);
128 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
130 tp->snd_cwnd = max(cwnd, restart_cwnd);
131 tp->snd_cwnd_stamp = tcp_time_stamp;
132 tp->snd_cwnd_used = 0;
135 static void tcp_event_data_sent(struct tcp_sock *tp,
136 struct sk_buff *skb, struct sock *sk)
138 struct inet_connection_sock *icsk = inet_csk(sk);
139 const u32 now = tcp_time_stamp;
141 if (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)
142 tcp_cwnd_restart(sk, __sk_dst_get(sk));
146 /* If it is a reply for ato after last received
147 * packet, enter pingpong mode.
149 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
150 icsk->icsk_ack.pingpong = 1;
153 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
155 tcp_dec_quickack_mode(sk, pkts);
156 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
159 /* Determine a window scaling and initial window to offer.
160 * Based on the assumption that the given amount of space
161 * will be offered. Store the results in the tp structure.
162 * NOTE: for smooth operation initial space offering should
163 * be a multiple of mss if possible. We assume here that mss >= 1.
164 * This MUST be enforced by all callers.
166 void tcp_select_initial_window(int __space, __u32 mss,
167 __u32 *rcv_wnd, __u32 *window_clamp,
168 int wscale_ok, __u8 *rcv_wscale)
170 unsigned int space = (__space < 0 ? 0 : __space);
172 /* If no clamp set the clamp to the max possible scaled window */
173 if (*window_clamp == 0)
174 (*window_clamp) = (65535 << 14);
175 space = min(*window_clamp, space);
177 /* Quantize space offering to a multiple of mss if possible. */
179 space = (space / mss) * mss;
181 /* NOTE: offering an initial window larger than 32767
182 * will break some buggy TCP stacks. If the admin tells us
183 * it is likely we could be speaking with such a buggy stack
184 * we will truncate our initial window offering to 32K-1
185 * unless the remote has sent us a window scaling option,
186 * which we interpret as a sign the remote TCP is not
187 * misinterpreting the window field as a signed quantity.
189 if (sysctl_tcp_workaround_signed_windows)
190 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
196 /* Set window scaling on max possible window
197 * See RFC1323 for an explanation of the limit to 14
199 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
200 space = min_t(u32, space, *window_clamp);
201 while (space > 65535 && (*rcv_wscale) < 14) {
207 /* Set initial window to value enough for senders,
208 * following RFC2414. Senders, not following this RFC,
209 * will be satisfied with 2.
211 if (mss > (1<<*rcv_wscale)) {
217 if (*rcv_wnd > init_cwnd*mss)
218 *rcv_wnd = init_cwnd*mss;
221 /* Set the clamp no higher than max representable value */
222 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
225 /* Chose a new window to advertise, update state in tcp_sock for the
226 * socket, and return result with RFC1323 scaling applied. The return
227 * value can be stuffed directly into th->window for an outgoing
230 static u16 tcp_select_window(struct sock *sk)
232 struct tcp_sock *tp = tcp_sk(sk);
233 u32 cur_win = tcp_receive_window(tp);
234 u32 new_win = __tcp_select_window(sk);
236 /* Never shrink the offered window */
237 if(new_win < cur_win) {
238 /* Danger Will Robinson!
239 * Don't update rcv_wup/rcv_wnd here or else
240 * we will not be able to advertise a zero
241 * window in time. --DaveM
243 * Relax Will Robinson.
247 tp->rcv_wnd = new_win;
248 tp->rcv_wup = tp->rcv_nxt;
250 /* Make sure we do not exceed the maximum possible
253 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
254 new_win = min(new_win, MAX_TCP_WINDOW);
256 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
258 /* RFC1323 scaling applied */
259 new_win >>= tp->rx_opt.rcv_wscale;
261 /* If we advertise zero window, disable fast path. */
268 static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp,
271 if (tp->rx_opt.tstamp_ok) {
272 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
274 (TCPOPT_TIMESTAMP << 8) |
276 *ptr++ = htonl(tstamp);
277 *ptr++ = htonl(tp->rx_opt.ts_recent);
279 if (tp->rx_opt.eff_sacks) {
280 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
283 *ptr++ = htonl((TCPOPT_NOP << 24) |
286 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
287 TCPOLEN_SACK_PERBLOCK)));
288 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
289 *ptr++ = htonl(sp[this_sack].start_seq);
290 *ptr++ = htonl(sp[this_sack].end_seq);
292 if (tp->rx_opt.dsack) {
293 tp->rx_opt.dsack = 0;
294 tp->rx_opt.eff_sacks--;
299 /* Construct a tcp options header for a SYN or SYN_ACK packet.
300 * If this is every changed make sure to change the definition of
301 * MAX_SYN_SIZE to match the new maximum number of options that you
304 static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
305 int offer_wscale, int wscale, __u32 tstamp,
308 /* We always get an MSS option.
309 * The option bytes which will be seen in normal data
310 * packets should timestamps be used, must be in the MSS
311 * advertised. But we subtract them from tp->mss_cache so
312 * that calculations in tcp_sendmsg are simpler etc.
313 * So account for this fact here if necessary. If we
314 * don't do this correctly, as a receiver we won't
315 * recognize data packets as being full sized when we
316 * should, and thus we won't abide by the delayed ACK
318 * SACKs don't matter, we never delay an ACK when we
319 * have any of those going out.
321 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
324 *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
325 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
327 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
328 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
329 *ptr++ = htonl(tstamp); /* TSVAL */
330 *ptr++ = htonl(ts_recent); /* TSECR */
332 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
333 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
335 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
338 /* This routine actually transmits TCP packets queued in by
339 * tcp_do_sendmsg(). This is used by both the initial
340 * transmission and possible later retransmissions.
341 * All SKB's seen here are completely headerless. It is our
342 * job to build the TCP header, and pass the packet down to
343 * IP so it can do the same plus pass the packet off to the
346 * We are working here with either a clone of the original
347 * SKB, or a fresh unique copy made by the retransmit engine.
349 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
351 const struct inet_connection_sock *icsk = inet_csk(sk);
352 struct inet_sock *inet;
354 struct tcp_skb_cb *tcb;
360 BUG_ON(!skb || !tcp_skb_pcount(skb));
362 /* If congestion control is doing timestamping, we must
363 * take such a timestamp before we potentially clone/copy.
365 if (icsk->icsk_ca_ops->rtt_sample)
366 __net_timestamp(skb);
368 if (likely(clone_it)) {
369 if (unlikely(skb_cloned(skb)))
370 skb = pskb_copy(skb, gfp_mask);
372 skb = skb_clone(skb, gfp_mask);
379 tcb = TCP_SKB_CB(skb);
380 tcp_header_size = tp->tcp_header_len;
382 #define SYSCTL_FLAG_TSTAMPS 0x1
383 #define SYSCTL_FLAG_WSCALE 0x2
384 #define SYSCTL_FLAG_SACK 0x4
387 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
388 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
389 if(sysctl_tcp_timestamps) {
390 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
391 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
393 if (sysctl_tcp_window_scaling) {
394 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
395 sysctl_flags |= SYSCTL_FLAG_WSCALE;
397 if (sysctl_tcp_sack) {
398 sysctl_flags |= SYSCTL_FLAG_SACK;
399 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
400 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
402 } else if (unlikely(tp->rx_opt.eff_sacks)) {
403 /* A SACK is 2 pad bytes, a 2 byte header, plus
404 * 2 32-bit sequence numbers for each SACK block.
406 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
407 (tp->rx_opt.eff_sacks *
408 TCPOLEN_SACK_PERBLOCK));
411 if (tcp_packets_in_flight(tp) == 0)
412 tcp_ca_event(sk, CA_EVENT_TX_START);
414 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
416 skb_set_owner_w(skb, sk);
418 /* Build TCP header and checksum it. */
419 th->source = inet->sport;
420 th->dest = inet->dport;
421 th->seq = htonl(tcb->seq);
422 th->ack_seq = htonl(tp->rcv_nxt);
423 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
426 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
427 /* RFC1323: The window in SYN & SYN/ACK segments
430 th->window = htons(tp->rcv_wnd);
432 th->window = htons(tcp_select_window(sk));
437 if (unlikely(tp->urg_mode &&
438 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
439 th->urg_ptr = htons(tp->snd_up-tcb->seq);
443 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
444 tcp_syn_build_options((__u32 *)(th + 1),
445 tcp_advertise_mss(sk),
446 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
447 (sysctl_flags & SYSCTL_FLAG_SACK),
448 (sysctl_flags & SYSCTL_FLAG_WSCALE),
449 tp->rx_opt.rcv_wscale,
451 tp->rx_opt.ts_recent);
453 tcp_build_and_update_options((__u32 *)(th + 1),
455 TCP_ECN_send(sk, tp, skb, tcp_header_size);
458 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
460 if (likely(tcb->flags & TCPCB_FLAG_ACK))
461 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
463 if (skb->len != tcp_header_size)
464 tcp_event_data_sent(tp, skb, sk);
466 TCP_INC_STATS(TCP_MIB_OUTSEGS);
468 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
469 if (likely(err <= 0))
474 /* NET_XMIT_CN is special. It does not guarantee,
475 * that this packet is lost. It tells that device
476 * is about to start to drop packets or already
477 * drops some packets of the same priority and
478 * invokes us to send less aggressively.
480 return err == NET_XMIT_CN ? 0 : err;
482 #undef SYSCTL_FLAG_TSTAMPS
483 #undef SYSCTL_FLAG_WSCALE
484 #undef SYSCTL_FLAG_SACK
488 /* This routine just queue's the buffer
490 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
491 * otherwise socket can stall.
493 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
495 struct tcp_sock *tp = tcp_sk(sk);
497 /* Advance write_seq and place onto the write_queue. */
498 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
499 skb_header_release(skb);
500 __skb_queue_tail(&sk->sk_write_queue, skb);
501 sk_charge_skb(sk, skb);
503 /* Queue it, remembering where we must start sending. */
504 if (sk->sk_send_head == NULL)
505 sk->sk_send_head = skb;
508 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
510 if (skb->len <= mss_now ||
511 !(sk->sk_route_caps & NETIF_F_TSO)) {
512 /* Avoid the costly divide in the normal
515 skb_shinfo(skb)->gso_segs = 1;
516 skb_shinfo(skb)->gso_size = 0;
517 skb_shinfo(skb)->gso_type = 0;
521 factor = skb->len + (mss_now - 1);
523 skb_shinfo(skb)->gso_segs = factor;
524 skb_shinfo(skb)->gso_size = mss_now;
525 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
529 /* Function to create two new TCP segments. Shrinks the given segment
530 * to the specified size and appends a new segment with the rest of the
531 * packet to the list. This won't be called frequently, I hope.
532 * Remember, these are still headerless SKBs at this point.
534 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
536 struct tcp_sock *tp = tcp_sk(sk);
537 struct sk_buff *buff;
538 int nsize, old_factor;
542 BUG_ON(len > skb->len);
544 clear_all_retrans_hints(tp);
545 nsize = skb_headlen(skb) - len;
549 if (skb_cloned(skb) &&
550 skb_is_nonlinear(skb) &&
551 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
554 /* Get a new skb... force flag on. */
555 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
557 return -ENOMEM; /* We'll just try again later. */
559 sk_charge_skb(sk, buff);
560 nlen = skb->len - len - nsize;
561 buff->truesize += nlen;
562 skb->truesize -= nlen;
564 /* Correct the sequence numbers. */
565 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
566 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
567 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
569 /* PSH and FIN should only be set in the second packet. */
570 flags = TCP_SKB_CB(skb)->flags;
571 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
572 TCP_SKB_CB(buff)->flags = flags;
573 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
574 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
576 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
577 /* Copy and checksum data tail into the new buffer. */
578 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
583 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
585 skb->ip_summed = CHECKSUM_HW;
586 skb_split(skb, buff, len);
589 buff->ip_summed = skb->ip_summed;
591 /* Looks stupid, but our code really uses when of
592 * skbs, which it never sent before. --ANK
594 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
595 buff->tstamp = skb->tstamp;
597 old_factor = tcp_skb_pcount(skb);
599 /* Fix up tso_factor for both original and new SKB. */
600 tcp_set_skb_tso_segs(sk, skb, mss_now);
601 tcp_set_skb_tso_segs(sk, buff, mss_now);
603 /* If this packet has been sent out already, we must
604 * adjust the various packet counters.
606 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
607 int diff = old_factor - tcp_skb_pcount(skb) -
608 tcp_skb_pcount(buff);
610 tp->packets_out -= diff;
612 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
613 tp->sacked_out -= diff;
614 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
615 tp->retrans_out -= diff;
617 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
618 tp->lost_out -= diff;
619 tp->left_out -= diff;
623 /* Adjust Reno SACK estimate. */
624 if (!tp->rx_opt.sack_ok) {
625 tp->sacked_out -= diff;
626 if ((int)tp->sacked_out < 0)
628 tcp_sync_left_out(tp);
631 tp->fackets_out -= diff;
632 if ((int)tp->fackets_out < 0)
637 /* Link BUFF into the send queue. */
638 skb_header_release(buff);
639 __skb_append(skb, buff, &sk->sk_write_queue);
644 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
645 * eventually). The difference is that pulled data not copied, but
646 * immediately discarded.
648 static void __pskb_trim_head(struct sk_buff *skb, int len)
654 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
655 if (skb_shinfo(skb)->frags[i].size <= eat) {
656 put_page(skb_shinfo(skb)->frags[i].page);
657 eat -= skb_shinfo(skb)->frags[i].size;
659 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
661 skb_shinfo(skb)->frags[k].page_offset += eat;
662 skb_shinfo(skb)->frags[k].size -= eat;
668 skb_shinfo(skb)->nr_frags = k;
670 skb->tail = skb->data;
671 skb->data_len -= len;
672 skb->len = skb->data_len;
675 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
677 if (skb_cloned(skb) &&
678 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
681 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
682 if (unlikely(len < skb_headlen(skb)))
683 __skb_pull(skb, len);
685 __pskb_trim_head(skb, len - skb_headlen(skb));
687 TCP_SKB_CB(skb)->seq += len;
688 skb->ip_summed = CHECKSUM_HW;
690 skb->truesize -= len;
691 sk->sk_wmem_queued -= len;
692 sk->sk_forward_alloc += len;
693 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
695 /* Any change of skb->len requires recalculation of tso
698 if (tcp_skb_pcount(skb) > 1)
699 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
704 /* Not accounting for SACKs here. */
705 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
707 struct tcp_sock *tp = tcp_sk(sk);
708 struct inet_connection_sock *icsk = inet_csk(sk);
711 /* Calculate base mss without TCP options:
712 It is MMS_S - sizeof(tcphdr) of rfc1122
714 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
716 /* Clamp it (mss_clamp does not include tcp options) */
717 if (mss_now > tp->rx_opt.mss_clamp)
718 mss_now = tp->rx_opt.mss_clamp;
720 /* Now subtract optional transport overhead */
721 mss_now -= icsk->icsk_ext_hdr_len;
723 /* Then reserve room for full set of TCP options and 8 bytes of data */
727 /* Now subtract TCP options size, not including SACKs */
728 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
733 /* Inverse of above */
734 int tcp_mss_to_mtu(struct sock *sk, int mss)
736 struct tcp_sock *tp = tcp_sk(sk);
737 struct inet_connection_sock *icsk = inet_csk(sk);
742 icsk->icsk_ext_hdr_len +
743 icsk->icsk_af_ops->net_header_len;
748 void tcp_mtup_init(struct sock *sk)
750 struct tcp_sock *tp = tcp_sk(sk);
751 struct inet_connection_sock *icsk = inet_csk(sk);
753 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
754 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
755 icsk->icsk_af_ops->net_header_len;
756 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
757 icsk->icsk_mtup.probe_size = 0;
760 /* This function synchronize snd mss to current pmtu/exthdr set.
762 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
763 for TCP options, but includes only bare TCP header.
765 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
766 It is minimum of user_mss and mss received with SYN.
767 It also does not include TCP options.
769 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
771 tp->mss_cache is current effective sending mss, including
772 all tcp options except for SACKs. It is evaluated,
773 taking into account current pmtu, but never exceeds
774 tp->rx_opt.mss_clamp.
776 NOTE1. rfc1122 clearly states that advertised MSS
777 DOES NOT include either tcp or ip options.
779 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
780 are READ ONLY outside this function. --ANK (980731)
783 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
785 struct tcp_sock *tp = tcp_sk(sk);
786 struct inet_connection_sock *icsk = inet_csk(sk);
789 if (icsk->icsk_mtup.search_high > pmtu)
790 icsk->icsk_mtup.search_high = pmtu;
792 mss_now = tcp_mtu_to_mss(sk, pmtu);
794 /* Bound mss with half of window */
795 if (tp->max_window && mss_now > (tp->max_window>>1))
796 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
798 /* And store cached results */
799 icsk->icsk_pmtu_cookie = pmtu;
800 if (icsk->icsk_mtup.enabled)
801 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
802 tp->mss_cache = mss_now;
807 /* Compute the current effective MSS, taking SACKs and IP options,
808 * and even PMTU discovery events into account.
810 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
811 * cannot be large. However, taking into account rare use of URG, this
814 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
816 struct tcp_sock *tp = tcp_sk(sk);
817 struct dst_entry *dst = __sk_dst_get(sk);
822 mss_now = tp->mss_cache;
825 (sk->sk_route_caps & NETIF_F_TSO) &&
830 u32 mtu = dst_mtu(dst);
831 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
832 mss_now = tcp_sync_mss(sk, mtu);
835 if (tp->rx_opt.eff_sacks)
836 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
837 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
839 xmit_size_goal = mss_now;
842 xmit_size_goal = (65535 -
843 inet_csk(sk)->icsk_af_ops->net_header_len -
844 inet_csk(sk)->icsk_ext_hdr_len -
847 if (tp->max_window &&
848 (xmit_size_goal > (tp->max_window >> 1)))
849 xmit_size_goal = max((tp->max_window >> 1),
850 68U - tp->tcp_header_len);
852 xmit_size_goal -= (xmit_size_goal % mss_now);
854 tp->xmit_size_goal = xmit_size_goal;
859 EXPORT_SYMBOL_GPL(tcp_current_mss);
861 /* Congestion window validation. (RFC2861) */
863 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
865 __u32 packets_out = tp->packets_out;
867 if (packets_out >= tp->snd_cwnd) {
868 /* Network is feed fully. */
869 tp->snd_cwnd_used = 0;
870 tp->snd_cwnd_stamp = tcp_time_stamp;
872 /* Network starves. */
873 if (tp->packets_out > tp->snd_cwnd_used)
874 tp->snd_cwnd_used = tp->packets_out;
876 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
877 tcp_cwnd_application_limited(sk);
881 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
883 u32 window, cwnd_len;
885 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
886 cwnd_len = mss_now * cwnd;
887 return min(window, cwnd_len);
890 /* Can at least one segment of SKB be sent right now, according to the
891 * congestion window rules? If so, return how many segments are allowed.
893 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
897 /* Don't be strict about the congestion window for the final FIN. */
898 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
901 in_flight = tcp_packets_in_flight(tp);
903 if (in_flight < cwnd)
904 return (cwnd - in_flight);
909 /* This must be invoked the first time we consider transmitting
912 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
914 int tso_segs = tcp_skb_pcount(skb);
918 tcp_skb_mss(skb) != mss_now)) {
919 tcp_set_skb_tso_segs(sk, skb, mss_now);
920 tso_segs = tcp_skb_pcount(skb);
925 static inline int tcp_minshall_check(const struct tcp_sock *tp)
927 return after(tp->snd_sml,tp->snd_una) &&
928 !after(tp->snd_sml, tp->snd_nxt);
931 /* Return 0, if packet can be sent now without violation Nagle's rules:
932 * 1. It is full sized.
933 * 2. Or it contains FIN. (already checked by caller)
934 * 3. Or TCP_NODELAY was set.
935 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
936 * With Minshall's modification: all sent small packets are ACKed.
939 static inline int tcp_nagle_check(const struct tcp_sock *tp,
940 const struct sk_buff *skb,
941 unsigned mss_now, int nonagle)
943 return (skb->len < mss_now &&
944 ((nonagle&TCP_NAGLE_CORK) ||
947 tcp_minshall_check(tp))));
950 /* Return non-zero if the Nagle test allows this packet to be
953 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
954 unsigned int cur_mss, int nonagle)
956 /* Nagle rule does not apply to frames, which sit in the middle of the
957 * write_queue (they have no chances to get new data).
959 * This is implemented in the callers, where they modify the 'nonagle'
960 * argument based upon the location of SKB in the send queue.
962 if (nonagle & TCP_NAGLE_PUSH)
965 /* Don't use the nagle rule for urgent data (or for the final FIN). */
967 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
970 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
976 /* Does at least the first segment of SKB fit into the send window? */
977 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
979 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
981 if (skb->len > cur_mss)
982 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
984 return !after(end_seq, tp->snd_una + tp->snd_wnd);
987 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
988 * should be put on the wire right now. If so, it returns the number of
989 * packets allowed by the congestion window.
991 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
992 unsigned int cur_mss, int nonagle)
994 struct tcp_sock *tp = tcp_sk(sk);
995 unsigned int cwnd_quota;
997 tcp_init_tso_segs(sk, skb, cur_mss);
999 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1002 cwnd_quota = tcp_cwnd_test(tp, skb);
1004 !tcp_snd_wnd_test(tp, skb, cur_mss))
1010 static inline int tcp_skb_is_last(const struct sock *sk,
1011 const struct sk_buff *skb)
1013 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1016 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1018 struct sk_buff *skb = sk->sk_send_head;
1021 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1022 (tcp_skb_is_last(sk, skb) ?
1027 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1028 * which is put after SKB on the list. It is very much like
1029 * tcp_fragment() except that it may make several kinds of assumptions
1030 * in order to speed up the splitting operation. In particular, we
1031 * know that all the data is in scatter-gather pages, and that the
1032 * packet has never been sent out before (and thus is not cloned).
1034 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1036 struct sk_buff *buff;
1037 int nlen = skb->len - len;
1040 /* All of a TSO frame must be composed of paged data. */
1041 if (skb->len != skb->data_len)
1042 return tcp_fragment(sk, skb, len, mss_now);
1044 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1045 if (unlikely(buff == NULL))
1048 sk_charge_skb(sk, buff);
1049 buff->truesize += nlen;
1050 skb->truesize -= nlen;
1052 /* Correct the sequence numbers. */
1053 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1054 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1055 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1057 /* PSH and FIN should only be set in the second packet. */
1058 flags = TCP_SKB_CB(skb)->flags;
1059 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1060 TCP_SKB_CB(buff)->flags = flags;
1062 /* This packet was never sent out yet, so no SACK bits. */
1063 TCP_SKB_CB(buff)->sacked = 0;
1065 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
1066 skb_split(skb, buff, len);
1068 /* Fix up tso_factor for both original and new SKB. */
1069 tcp_set_skb_tso_segs(sk, skb, mss_now);
1070 tcp_set_skb_tso_segs(sk, buff, mss_now);
1072 /* Link BUFF into the send queue. */
1073 skb_header_release(buff);
1074 __skb_append(skb, buff, &sk->sk_write_queue);
1079 /* Try to defer sending, if possible, in order to minimize the amount
1080 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1082 * This algorithm is from John Heffner.
1084 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1086 const struct inet_connection_sock *icsk = inet_csk(sk);
1087 u32 send_win, cong_win, limit, in_flight;
1089 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1092 if (icsk->icsk_ca_state != TCP_CA_Open)
1095 in_flight = tcp_packets_in_flight(tp);
1097 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1098 (tp->snd_cwnd <= in_flight));
1100 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1102 /* From in_flight test above, we know that cwnd > in_flight. */
1103 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1105 limit = min(send_win, cong_win);
1107 /* If a full-sized TSO skb can be sent, do it. */
1111 if (sysctl_tcp_tso_win_divisor) {
1112 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1114 /* If at least some fraction of a window is available,
1117 chunk /= sysctl_tcp_tso_win_divisor;
1121 /* Different approach, try not to defer past a single
1122 * ACK. Receiver should ACK every other full sized
1123 * frame, so if we have space for more than 3 frames
1126 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1130 /* Ok, it looks like it is advisable to defer. */
1134 /* Create a new MTU probe if we are ready.
1135 * Returns 0 if we should wait to probe (no cwnd available),
1136 * 1 if a probe was sent,
1138 static int tcp_mtu_probe(struct sock *sk)
1140 struct tcp_sock *tp = tcp_sk(sk);
1141 struct inet_connection_sock *icsk = inet_csk(sk);
1142 struct sk_buff *skb, *nskb, *next;
1149 /* Not currently probing/verifying,
1151 * have enough cwnd, and
1152 * not SACKing (the variable headers throw things off) */
1153 if (!icsk->icsk_mtup.enabled ||
1154 icsk->icsk_mtup.probe_size ||
1155 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1156 tp->snd_cwnd < 11 ||
1157 tp->rx_opt.eff_sacks)
1160 /* Very simple search strategy: just double the MSS. */
1161 mss_now = tcp_current_mss(sk, 0);
1162 probe_size = 2*tp->mss_cache;
1163 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1164 /* TODO: set timer for probe_converge_event */
1168 /* Have enough data in the send queue to probe? */
1170 if ((skb = sk->sk_send_head) == NULL)
1172 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1174 if (len < probe_size)
1177 /* Receive window check. */
1178 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1179 if (tp->snd_wnd < probe_size)
1185 /* Do we need to wait to drain cwnd? */
1186 pif = tcp_packets_in_flight(tp);
1187 if (pif + 2 > tp->snd_cwnd) {
1188 /* With no packets in flight, don't stall. */
1195 /* We're allowed to probe. Build it now. */
1196 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1198 sk_charge_skb(sk, nskb);
1200 skb = sk->sk_send_head;
1201 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1202 sk->sk_send_head = nskb;
1204 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1205 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1206 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1207 TCP_SKB_CB(nskb)->sacked = 0;
1209 if (skb->ip_summed == CHECKSUM_HW)
1210 nskb->ip_summed = CHECKSUM_HW;
1213 while (len < probe_size) {
1216 copy = min_t(int, skb->len, probe_size - len);
1217 if (nskb->ip_summed)
1218 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1220 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1221 skb_put(nskb, copy), copy, nskb->csum);
1223 if (skb->len <= copy) {
1224 /* We've eaten all the data from this skb.
1226 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1227 __skb_unlink(skb, &sk->sk_write_queue);
1228 sk_stream_free_skb(sk, skb);
1230 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1231 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1232 if (!skb_shinfo(skb)->nr_frags) {
1233 skb_pull(skb, copy);
1234 if (skb->ip_summed != CHECKSUM_HW)
1235 skb->csum = csum_partial(skb->data, skb->len, 0);
1237 __pskb_trim_head(skb, copy);
1238 tcp_set_skb_tso_segs(sk, skb, mss_now);
1240 TCP_SKB_CB(skb)->seq += copy;
1246 tcp_init_tso_segs(sk, nskb, nskb->len);
1248 /* We're ready to send. If this fails, the probe will
1249 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1250 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1251 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1252 /* Decrement cwnd here because we are sending
1253 * effectively two packets. */
1255 update_send_head(sk, tp, nskb);
1257 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1258 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1259 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1268 /* This routine writes packets to the network. It advances the
1269 * send_head. This happens as incoming acks open up the remote
1272 * Returns 1, if no segments are in flight and we have queued segments, but
1273 * cannot send anything now because of SWS or another problem.
1275 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1277 struct tcp_sock *tp = tcp_sk(sk);
1278 struct sk_buff *skb;
1279 unsigned int tso_segs, sent_pkts;
1283 /* If we are closed, the bytes will have to remain here.
1284 * In time closedown will finish, we empty the write queue and all
1287 if (unlikely(sk->sk_state == TCP_CLOSE))
1292 /* Do MTU probing. */
1293 if ((result = tcp_mtu_probe(sk)) == 0) {
1295 } else if (result > 0) {
1299 while ((skb = sk->sk_send_head)) {
1302 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1305 cwnd_quota = tcp_cwnd_test(tp, skb);
1309 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1312 if (tso_segs == 1) {
1313 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1314 (tcp_skb_is_last(sk, skb) ?
1315 nonagle : TCP_NAGLE_PUSH))))
1318 if (tcp_tso_should_defer(sk, tp, skb))
1324 limit = tcp_window_allows(tp, skb,
1325 mss_now, cwnd_quota);
1327 if (skb->len < limit) {
1328 unsigned int trim = skb->len % mss_now;
1331 limit = skb->len - trim;
1335 if (skb->len > limit &&
1336 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1339 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1341 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1344 /* Advance the send_head. This one is sent out.
1345 * This call will increment packets_out.
1347 update_send_head(sk, tp, skb);
1349 tcp_minshall_update(tp, mss_now, skb);
1353 if (likely(sent_pkts)) {
1354 tcp_cwnd_validate(sk, tp);
1357 return !tp->packets_out && sk->sk_send_head;
1360 /* Push out any pending frames which were held back due to
1361 * TCP_CORK or attempt at coalescing tiny packets.
1362 * The socket must be locked by the caller.
1364 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1365 unsigned int cur_mss, int nonagle)
1367 struct sk_buff *skb = sk->sk_send_head;
1370 if (tcp_write_xmit(sk, cur_mss, nonagle))
1371 tcp_check_probe_timer(sk, tp);
1374 EXPORT_SYMBOL_GPL(__tcp_push_pending_frames);
1376 /* Send _single_ skb sitting at the send head. This function requires
1377 * true push pending frames to setup probe timer etc.
1379 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1381 struct tcp_sock *tp = tcp_sk(sk);
1382 struct sk_buff *skb = sk->sk_send_head;
1383 unsigned int tso_segs, cwnd_quota;
1385 BUG_ON(!skb || skb->len < mss_now);
1387 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1388 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1390 if (likely(cwnd_quota)) {
1397 limit = tcp_window_allows(tp, skb,
1398 mss_now, cwnd_quota);
1400 if (skb->len < limit) {
1401 unsigned int trim = skb->len % mss_now;
1404 limit = skb->len - trim;
1408 if (skb->len > limit &&
1409 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1412 /* Send it out now. */
1413 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1415 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1416 update_send_head(sk, tp, skb);
1417 tcp_cwnd_validate(sk, tp);
1423 /* This function returns the amount that we can raise the
1424 * usable window based on the following constraints
1426 * 1. The window can never be shrunk once it is offered (RFC 793)
1427 * 2. We limit memory per socket
1430 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1431 * RECV.NEXT + RCV.WIN fixed until:
1432 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1434 * i.e. don't raise the right edge of the window until you can raise
1435 * it at least MSS bytes.
1437 * Unfortunately, the recommended algorithm breaks header prediction,
1438 * since header prediction assumes th->window stays fixed.
1440 * Strictly speaking, keeping th->window fixed violates the receiver
1441 * side SWS prevention criteria. The problem is that under this rule
1442 * a stream of single byte packets will cause the right side of the
1443 * window to always advance by a single byte.
1445 * Of course, if the sender implements sender side SWS prevention
1446 * then this will not be a problem.
1448 * BSD seems to make the following compromise:
1450 * If the free space is less than the 1/4 of the maximum
1451 * space available and the free space is less than 1/2 mss,
1452 * then set the window to 0.
1453 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1454 * Otherwise, just prevent the window from shrinking
1455 * and from being larger than the largest representable value.
1457 * This prevents incremental opening of the window in the regime
1458 * where TCP is limited by the speed of the reader side taking
1459 * data out of the TCP receive queue. It does nothing about
1460 * those cases where the window is constrained on the sender side
1461 * because the pipeline is full.
1463 * BSD also seems to "accidentally" limit itself to windows that are a
1464 * multiple of MSS, at least until the free space gets quite small.
1465 * This would appear to be a side effect of the mbuf implementation.
1466 * Combining these two algorithms results in the observed behavior
1467 * of having a fixed window size at almost all times.
1469 * Below we obtain similar behavior by forcing the offered window to
1470 * a multiple of the mss when it is feasible to do so.
1472 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1473 * Regular options like TIMESTAMP are taken into account.
1475 u32 __tcp_select_window(struct sock *sk)
1477 struct inet_connection_sock *icsk = inet_csk(sk);
1478 struct tcp_sock *tp = tcp_sk(sk);
1479 /* MSS for the peer's data. Previous versions used mss_clamp
1480 * here. I don't know if the value based on our guesses
1481 * of peer's MSS is better for the performance. It's more correct
1482 * but may be worse for the performance because of rcv_mss
1483 * fluctuations. --SAW 1998/11/1
1485 int mss = icsk->icsk_ack.rcv_mss;
1486 int free_space = tcp_space(sk);
1487 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1490 if (mss > full_space)
1493 if (free_space < full_space/2) {
1494 icsk->icsk_ack.quick = 0;
1496 if (tcp_memory_pressure)
1497 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1499 if (free_space < mss)
1503 if (free_space > tp->rcv_ssthresh)
1504 free_space = tp->rcv_ssthresh;
1506 /* Don't do rounding if we are using window scaling, since the
1507 * scaled window will not line up with the MSS boundary anyway.
1509 window = tp->rcv_wnd;
1510 if (tp->rx_opt.rcv_wscale) {
1511 window = free_space;
1513 /* Advertise enough space so that it won't get scaled away.
1514 * Import case: prevent zero window announcement if
1515 * 1<<rcv_wscale > mss.
1517 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1518 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1519 << tp->rx_opt.rcv_wscale);
1521 /* Get the largest window that is a nice multiple of mss.
1522 * Window clamp already applied above.
1523 * If our current window offering is within 1 mss of the
1524 * free space we just keep it. This prevents the divide
1525 * and multiply from happening most of the time.
1526 * We also don't do any window rounding when the free space
1529 if (window <= free_space - mss || window > free_space)
1530 window = (free_space/mss)*mss;
1536 /* Attempt to collapse two adjacent SKB's during retransmission. */
1537 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *next_skb = skb->next;
1542 /* The first test we must make is that neither of these two
1543 * SKB's are still referenced by someone else.
1545 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1546 int skb_size = skb->len, next_skb_size = next_skb->len;
1547 u16 flags = TCP_SKB_CB(skb)->flags;
1549 /* Also punt if next skb has been SACK'd. */
1550 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1553 /* Next skb is out of window. */
1554 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1557 /* Punt if not enough space exists in the first SKB for
1558 * the data in the second, or the total combined payload
1559 * would exceed the MSS.
1561 if ((next_skb_size > skb_tailroom(skb)) ||
1562 ((skb_size + next_skb_size) > mss_now))
1565 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1566 tcp_skb_pcount(next_skb) != 1);
1568 /* changing transmit queue under us so clear hints */
1569 clear_all_retrans_hints(tp);
1571 /* Ok. We will be able to collapse the packet. */
1572 __skb_unlink(next_skb, &sk->sk_write_queue);
1574 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1576 if (next_skb->ip_summed == CHECKSUM_HW)
1577 skb->ip_summed = CHECKSUM_HW;
1579 if (skb->ip_summed != CHECKSUM_HW)
1580 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1582 /* Update sequence range on original skb. */
1583 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1585 /* Merge over control information. */
1586 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1587 TCP_SKB_CB(skb)->flags = flags;
1589 /* All done, get rid of second SKB and account for it so
1590 * packet counting does not break.
1592 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1593 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1594 tp->retrans_out -= tcp_skb_pcount(next_skb);
1595 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1596 tp->lost_out -= tcp_skb_pcount(next_skb);
1597 tp->left_out -= tcp_skb_pcount(next_skb);
1599 /* Reno case is special. Sigh... */
1600 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1601 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1602 tp->left_out -= tcp_skb_pcount(next_skb);
1605 /* Not quite right: it can be > snd.fack, but
1606 * it is better to underestimate fackets.
1608 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1609 tcp_packets_out_dec(tp, next_skb);
1610 sk_stream_free_skb(sk, next_skb);
1614 /* Do a simple retransmit without using the backoff mechanisms in
1615 * tcp_timer. This is used for path mtu discovery.
1616 * The socket is already locked here.
1618 void tcp_simple_retransmit(struct sock *sk)
1620 const struct inet_connection_sock *icsk = inet_csk(sk);
1621 struct tcp_sock *tp = tcp_sk(sk);
1622 struct sk_buff *skb;
1623 unsigned int mss = tcp_current_mss(sk, 0);
1626 sk_stream_for_retrans_queue(skb, sk) {
1627 if (skb->len > mss &&
1628 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1629 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1630 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1631 tp->retrans_out -= tcp_skb_pcount(skb);
1633 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1634 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1635 tp->lost_out += tcp_skb_pcount(skb);
1641 clear_all_retrans_hints(tp);
1646 tcp_sync_left_out(tp);
1648 /* Don't muck with the congestion window here.
1649 * Reason is that we do not increase amount of _data_
1650 * in network, but units changed and effective
1651 * cwnd/ssthresh really reduced now.
1653 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1654 tp->high_seq = tp->snd_nxt;
1655 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1656 tp->prior_ssthresh = 0;
1657 tp->undo_marker = 0;
1658 tcp_set_ca_state(sk, TCP_CA_Loss);
1660 tcp_xmit_retransmit_queue(sk);
1663 /* This retransmits one SKB. Policy decisions and retransmit queue
1664 * state updates are done by the caller. Returns non-zero if an
1665 * error occurred which prevented the send.
1667 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1669 struct tcp_sock *tp = tcp_sk(sk);
1670 struct inet_connection_sock *icsk = inet_csk(sk);
1671 unsigned int cur_mss = tcp_current_mss(sk, 0);
1674 /* Inconslusive MTU probe */
1675 if (icsk->icsk_mtup.probe_size) {
1676 icsk->icsk_mtup.probe_size = 0;
1679 /* Do not sent more than we queued. 1/4 is reserved for possible
1680 * copying overhead: fragmentation, tunneling, mangling etc.
1682 if (atomic_read(&sk->sk_wmem_alloc) >
1683 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1686 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1687 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1689 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1693 /* If receiver has shrunk his window, and skb is out of
1694 * new window, do not retransmit it. The exception is the
1695 * case, when window is shrunk to zero. In this case
1696 * our retransmit serves as a zero window probe.
1698 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1699 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1702 if (skb->len > cur_mss) {
1703 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1704 return -ENOMEM; /* We'll try again later. */
1707 /* Collapse two adjacent packets if worthwhile and we can. */
1708 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1709 (skb->len < (cur_mss >> 1)) &&
1710 (skb->next != sk->sk_send_head) &&
1711 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1712 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1713 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1714 (sysctl_tcp_retrans_collapse != 0))
1715 tcp_retrans_try_collapse(sk, skb, cur_mss);
1717 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1718 return -EHOSTUNREACH; /* Routing failure or similar. */
1720 /* Some Solaris stacks overoptimize and ignore the FIN on a
1721 * retransmit when old data is attached. So strip it off
1722 * since it is cheap to do so and saves bytes on the network.
1725 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1726 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1727 if (!pskb_trim(skb, 0)) {
1728 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1729 skb_shinfo(skb)->gso_segs = 1;
1730 skb_shinfo(skb)->gso_size = 0;
1731 skb_shinfo(skb)->gso_type = 0;
1732 skb->ip_summed = CHECKSUM_NONE;
1737 /* Make a copy, if the first transmission SKB clone we made
1738 * is still in somebody's hands, else make a clone.
1740 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1742 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1745 /* Update global TCP statistics. */
1746 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1748 tp->total_retrans++;
1750 #if FASTRETRANS_DEBUG > 0
1751 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1752 if (net_ratelimit())
1753 printk(KERN_DEBUG "retrans_out leaked.\n");
1756 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1757 tp->retrans_out += tcp_skb_pcount(skb);
1759 /* Save stamp of the first retransmit. */
1760 if (!tp->retrans_stamp)
1761 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1765 /* snd_nxt is stored to detect loss of retransmitted segment,
1766 * see tcp_input.c tcp_sacktag_write_queue().
1768 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1773 /* This gets called after a retransmit timeout, and the initially
1774 * retransmitted data is acknowledged. It tries to continue
1775 * resending the rest of the retransmit queue, until either
1776 * we've sent it all or the congestion window limit is reached.
1777 * If doing SACK, the first ACK which comes back for a timeout
1778 * based retransmit packet might feed us FACK information again.
1779 * If so, we use it to avoid unnecessarily retransmissions.
1781 void tcp_xmit_retransmit_queue(struct sock *sk)
1783 const struct inet_connection_sock *icsk = inet_csk(sk);
1784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct sk_buff *skb;
1788 if (tp->retransmit_skb_hint) {
1789 skb = tp->retransmit_skb_hint;
1790 packet_cnt = tp->retransmit_cnt_hint;
1792 skb = sk->sk_write_queue.next;
1796 /* First pass: retransmit lost packets. */
1798 sk_stream_for_retrans_queue_from(skb, sk) {
1799 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1801 /* we could do better than to assign each time */
1802 tp->retransmit_skb_hint = skb;
1803 tp->retransmit_cnt_hint = packet_cnt;
1805 /* Assume this retransmit will generate
1806 * only one packet for congestion window
1807 * calculation purposes. This works because
1808 * tcp_retransmit_skb() will chop up the
1809 * packet to be MSS sized and all the
1810 * packet counting works out.
1812 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1815 if (sacked & TCPCB_LOST) {
1816 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1817 if (tcp_retransmit_skb(sk, skb)) {
1818 tp->retransmit_skb_hint = NULL;
1821 if (icsk->icsk_ca_state != TCP_CA_Loss)
1822 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1824 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1827 skb_peek(&sk->sk_write_queue))
1828 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1829 inet_csk(sk)->icsk_rto,
1833 packet_cnt += tcp_skb_pcount(skb);
1834 if (packet_cnt >= tp->lost_out)
1840 /* OK, demanded retransmission is finished. */
1842 /* Forward retransmissions are possible only during Recovery. */
1843 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1846 /* No forward retransmissions in Reno are possible. */
1847 if (!tp->rx_opt.sack_ok)
1850 /* Yeah, we have to make difficult choice between forward transmission
1851 * and retransmission... Both ways have their merits...
1853 * For now we do not retransmit anything, while we have some new
1857 if (tcp_may_send_now(sk, tp))
1860 if (tp->forward_skb_hint) {
1861 skb = tp->forward_skb_hint;
1862 packet_cnt = tp->forward_cnt_hint;
1864 skb = sk->sk_write_queue.next;
1868 sk_stream_for_retrans_queue_from(skb, sk) {
1869 tp->forward_cnt_hint = packet_cnt;
1870 tp->forward_skb_hint = skb;
1872 /* Similar to the retransmit loop above we
1873 * can pretend that the retransmitted SKB
1874 * we send out here will be composed of one
1875 * real MSS sized packet because tcp_retransmit_skb()
1876 * will fragment it if necessary.
1878 if (++packet_cnt > tp->fackets_out)
1881 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1884 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1887 /* Ok, retransmit it. */
1888 if (tcp_retransmit_skb(sk, skb)) {
1889 tp->forward_skb_hint = NULL;
1893 if (skb == skb_peek(&sk->sk_write_queue))
1894 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1895 inet_csk(sk)->icsk_rto,
1898 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1903 /* Send a fin. The caller locks the socket for us. This cannot be
1904 * allowed to fail queueing a FIN frame under any circumstances.
1906 void tcp_send_fin(struct sock *sk)
1908 struct tcp_sock *tp = tcp_sk(sk);
1909 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1912 /* Optimization, tack on the FIN if we have a queue of
1913 * unsent frames. But be careful about outgoing SACKS
1916 mss_now = tcp_current_mss(sk, 1);
1918 if (sk->sk_send_head != NULL) {
1919 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1920 TCP_SKB_CB(skb)->end_seq++;
1923 /* Socket is locked, keep trying until memory is available. */
1925 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1931 /* Reserve space for headers and prepare control bits. */
1932 skb_reserve(skb, MAX_TCP_HEADER);
1934 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1935 TCP_SKB_CB(skb)->sacked = 0;
1936 skb_shinfo(skb)->gso_segs = 1;
1937 skb_shinfo(skb)->gso_size = 0;
1938 skb_shinfo(skb)->gso_type = 0;
1940 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1941 TCP_SKB_CB(skb)->seq = tp->write_seq;
1942 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1943 tcp_queue_skb(sk, skb);
1945 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1948 /* We get here when a process closes a file descriptor (either due to
1949 * an explicit close() or as a byproduct of exit()'ing) and there
1950 * was unread data in the receive queue. This behavior is recommended
1951 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1953 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1955 struct tcp_sock *tp = tcp_sk(sk);
1956 struct sk_buff *skb;
1958 /* NOTE: No TCP options attached and we never retransmit this. */
1959 skb = alloc_skb(MAX_TCP_HEADER, priority);
1961 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1965 /* Reserve space for headers and prepare control bits. */
1966 skb_reserve(skb, MAX_TCP_HEADER);
1968 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1969 TCP_SKB_CB(skb)->sacked = 0;
1970 skb_shinfo(skb)->gso_segs = 1;
1971 skb_shinfo(skb)->gso_size = 0;
1972 skb_shinfo(skb)->gso_type = 0;
1975 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1976 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1977 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1978 if (tcp_transmit_skb(sk, skb, 0, priority))
1979 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1982 /* WARNING: This routine must only be called when we have already sent
1983 * a SYN packet that crossed the incoming SYN that caused this routine
1984 * to get called. If this assumption fails then the initial rcv_wnd
1985 * and rcv_wscale values will not be correct.
1987 int tcp_send_synack(struct sock *sk)
1989 struct sk_buff* skb;
1991 skb = skb_peek(&sk->sk_write_queue);
1992 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1993 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1996 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1997 if (skb_cloned(skb)) {
1998 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2001 __skb_unlink(skb, &sk->sk_write_queue);
2002 skb_header_release(nskb);
2003 __skb_queue_head(&sk->sk_write_queue, nskb);
2004 sk_stream_free_skb(sk, skb);
2005 sk_charge_skb(sk, nskb);
2009 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2010 TCP_ECN_send_synack(tcp_sk(sk), skb);
2012 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2013 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2017 * Prepare a SYN-ACK.
2019 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2020 struct request_sock *req)
2022 struct inet_request_sock *ireq = inet_rsk(req);
2023 struct tcp_sock *tp = tcp_sk(sk);
2025 int tcp_header_size;
2026 struct sk_buff *skb;
2028 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2032 /* Reserve space for headers. */
2033 skb_reserve(skb, MAX_TCP_HEADER);
2035 skb->dst = dst_clone(dst);
2037 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2038 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2039 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2040 /* SACK_PERM is in the place of NOP NOP of TS */
2041 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2042 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2044 memset(th, 0, sizeof(struct tcphdr));
2047 if (dst->dev->features&NETIF_F_TSO)
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 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2166 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2167 tp->snd_nxt = tp->write_seq;
2168 tp->pushed_seq = tp->write_seq;
2171 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2172 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2173 skb_header_release(buff);
2174 __skb_queue_tail(&sk->sk_write_queue, buff);
2175 sk_charge_skb(sk, buff);
2176 tp->packets_out += tcp_skb_pcount(buff);
2177 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2178 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2180 /* Timer for repeating the SYN until an answer. */
2181 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2182 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2186 /* Send out a delayed ack, the caller does the policy checking
2187 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2190 void tcp_send_delayed_ack(struct sock *sk)
2192 struct inet_connection_sock *icsk = inet_csk(sk);
2193 int ato = icsk->icsk_ack.ato;
2194 unsigned long timeout;
2196 if (ato > TCP_DELACK_MIN) {
2197 const struct tcp_sock *tp = tcp_sk(sk);
2200 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2201 max_ato = TCP_DELACK_MAX;
2203 /* Slow path, intersegment interval is "high". */
2205 /* If some rtt estimate is known, use it to bound delayed ack.
2206 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2210 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2216 ato = min(ato, max_ato);
2219 /* Stay within the limit we were given */
2220 timeout = jiffies + ato;
2222 /* Use new timeout only if there wasn't a older one earlier. */
2223 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2224 /* If delack timer was blocked or is about to expire,
2227 if (icsk->icsk_ack.blocked ||
2228 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2233 if (!time_before(timeout, icsk->icsk_ack.timeout))
2234 timeout = icsk->icsk_ack.timeout;
2236 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2237 icsk->icsk_ack.timeout = timeout;
2238 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2241 /* This routine sends an ack and also updates the window. */
2242 void tcp_send_ack(struct sock *sk)
2244 /* If we have been reset, we may not send again. */
2245 if (sk->sk_state != TCP_CLOSE) {
2246 struct tcp_sock *tp = tcp_sk(sk);
2247 struct sk_buff *buff;
2249 /* We are not putting this on the write queue, so
2250 * tcp_transmit_skb() will set the ownership to this
2253 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2255 inet_csk_schedule_ack(sk);
2256 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2257 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2258 TCP_DELACK_MAX, TCP_RTO_MAX);
2262 /* Reserve space for headers and prepare control bits. */
2263 skb_reserve(buff, MAX_TCP_HEADER);
2265 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2266 TCP_SKB_CB(buff)->sacked = 0;
2267 skb_shinfo(buff)->gso_segs = 1;
2268 skb_shinfo(buff)->gso_size = 0;
2269 skb_shinfo(buff)->gso_type = 0;
2271 /* Send it off, this clears delayed acks for us. */
2272 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2273 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2274 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2278 /* This routine sends a packet with an out of date sequence
2279 * number. It assumes the other end will try to ack it.
2281 * Question: what should we make while urgent mode?
2282 * 4.4BSD forces sending single byte of data. We cannot send
2283 * out of window data, because we have SND.NXT==SND.MAX...
2285 * Current solution: to send TWO zero-length segments in urgent mode:
2286 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2287 * out-of-date with SND.UNA-1 to probe window.
2289 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2291 struct tcp_sock *tp = tcp_sk(sk);
2292 struct sk_buff *skb;
2294 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2295 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2299 /* Reserve space for headers and set control bits. */
2300 skb_reserve(skb, MAX_TCP_HEADER);
2302 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2303 TCP_SKB_CB(skb)->sacked = urgent;
2304 skb_shinfo(skb)->gso_segs = 1;
2305 skb_shinfo(skb)->gso_size = 0;
2306 skb_shinfo(skb)->gso_type = 0;
2308 /* Use a previous sequence. This should cause the other
2309 * end to send an ack. Don't queue or clone SKB, just
2312 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2313 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2314 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2315 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2318 int tcp_write_wakeup(struct sock *sk)
2320 if (sk->sk_state != TCP_CLOSE) {
2321 struct tcp_sock *tp = tcp_sk(sk);
2322 struct sk_buff *skb;
2324 if ((skb = sk->sk_send_head) != NULL &&
2325 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2327 unsigned int mss = tcp_current_mss(sk, 0);
2328 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2330 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2331 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2333 /* We are probing the opening of a window
2334 * but the window size is != 0
2335 * must have been a result SWS avoidance ( sender )
2337 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2339 seg_size = min(seg_size, mss);
2340 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2341 if (tcp_fragment(sk, skb, seg_size, mss))
2343 } else if (!tcp_skb_pcount(skb))
2344 tcp_set_skb_tso_segs(sk, skb, mss);
2346 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2347 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2348 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2350 update_send_head(sk, tp, skb);
2355 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2356 tcp_xmit_probe_skb(sk, TCPCB_URG);
2357 return tcp_xmit_probe_skb(sk, 0);
2363 /* A window probe timeout has occurred. If window is not closed send
2364 * a partial packet else a zero probe.
2366 void tcp_send_probe0(struct sock *sk)
2368 struct inet_connection_sock *icsk = inet_csk(sk);
2369 struct tcp_sock *tp = tcp_sk(sk);
2372 err = tcp_write_wakeup(sk);
2374 if (tp->packets_out || !sk->sk_send_head) {
2375 /* Cancel probe timer, if it is not required. */
2376 icsk->icsk_probes_out = 0;
2377 icsk->icsk_backoff = 0;
2382 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2383 icsk->icsk_backoff++;
2384 icsk->icsk_probes_out++;
2385 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2386 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2389 /* If packet was not sent due to local congestion,
2390 * do not backoff and do not remember icsk_probes_out.
2391 * Let local senders to fight for local resources.
2393 * Use accumulated backoff yet.
2395 if (!icsk->icsk_probes_out)
2396 icsk->icsk_probes_out = 1;
2397 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2398 min(icsk->icsk_rto << icsk->icsk_backoff,
2399 TCP_RESOURCE_PROBE_INTERVAL),
2404 EXPORT_SYMBOL(tcp_connect);
2405 EXPORT_SYMBOL(tcp_make_synack);
2406 EXPORT_SYMBOL(tcp_simple_retransmit);
2407 EXPORT_SYMBOL(tcp_sync_mss);
2408 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2409 EXPORT_SYMBOL(tcp_mtup_init);