2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
8 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
13 * Corey Minyard <wf-rch!minyard@relay.EU.net>
14 * Florian La Roche, <flla@stud.uni-sb.de>
15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
16 * Linus Torvalds, <torvalds@cs.helsinki.fi>
17 * Alan Cox, <gw4pts@gw4pts.ampr.org>
18 * Matthew Dillon, <dillon@apollo.west.oic.com>
19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
20 * Jorge Cwik, <jorge@laser.satlink.net>
24 * Changes: Pedro Roque : Retransmit queue handled by TCP.
25 * : Fragmentation on mtu decrease
26 * : Segment collapse on retransmit
29 * Linus Torvalds : send_delayed_ack
30 * David S. Miller : Charge memory using the right skb
31 * during syn/ack processing.
32 * David S. Miller : Output engine completely rewritten.
33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
34 * Cacophonix Gaul : draft-minshall-nagle-01
35 * J Hadi Salim : ECN support
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse = 1;
48 /* This limits the percentage of the congestion window which we
49 * will allow a single TSO frame to consume. Building TSO frames
50 * which are too large can cause TCP streams to be bursty.
52 int sysctl_tcp_tso_win_divisor = 3;
54 static inline void update_send_head(struct sock *sk, struct tcp_sock *tp,
57 sk->sk_send_head = skb->next;
58 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
59 sk->sk_send_head = NULL;
60 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
61 tcp_packets_out_inc(sk, tp, skb);
64 /* SND.NXT, if window was not shrunk.
65 * If window has been shrunk, what should we make? It is not clear at all.
66 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
67 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
68 * invalid. OK, let's make this for now:
70 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
72 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
75 return tp->snd_una+tp->snd_wnd;
78 /* Calculate mss to advertise in SYN segment.
79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
81 * 1. It is independent of path mtu.
82 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
83 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
84 * attached devices, because some buggy hosts are confused by
86 * 4. We do not make 3, we advertise MSS, calculated from first
87 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
88 * This may be overridden via information stored in routing table.
89 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
90 * probably even Jumbo".
92 static __u16 tcp_advertise_mss(struct sock *sk)
94 struct tcp_sock *tp = tcp_sk(sk);
95 struct dst_entry *dst = __sk_dst_get(sk);
98 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
99 mss = dst_metric(dst, RTAX_ADVMSS);
106 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
107 * This is the first part of cwnd validation mechanism. */
108 static void tcp_cwnd_restart(struct tcp_sock *tp, struct dst_entry *dst)
110 s32 delta = tcp_time_stamp - tp->lsndtime;
111 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
112 u32 cwnd = tp->snd_cwnd;
114 tcp_ca_event(tp, CA_EVENT_CWND_RESTART);
116 tp->snd_ssthresh = tcp_current_ssthresh(tp);
117 restart_cwnd = min(restart_cwnd, cwnd);
119 while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd)
121 tp->snd_cwnd = max(cwnd, restart_cwnd);
122 tp->snd_cwnd_stamp = tcp_time_stamp;
123 tp->snd_cwnd_used = 0;
126 static inline void tcp_event_data_sent(struct tcp_sock *tp,
127 struct sk_buff *skb, struct sock *sk)
129 u32 now = tcp_time_stamp;
131 if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto)
132 tcp_cwnd_restart(tp, __sk_dst_get(sk));
136 /* If it is a reply for ato after last received
137 * packet, enter pingpong mode.
139 if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato)
140 tp->ack.pingpong = 1;
143 static __inline__ void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
145 struct tcp_sock *tp = tcp_sk(sk);
147 tcp_dec_quickack_mode(tp, pkts);
148 tcp_clear_xmit_timer(sk, TCP_TIME_DACK);
151 /* Determine a window scaling and initial window to offer.
152 * Based on the assumption that the given amount of space
153 * will be offered. Store the results in the tp structure.
154 * NOTE: for smooth operation initial space offering should
155 * be a multiple of mss if possible. We assume here that mss >= 1.
156 * This MUST be enforced by all callers.
158 void tcp_select_initial_window(int __space, __u32 mss,
159 __u32 *rcv_wnd, __u32 *window_clamp,
160 int wscale_ok, __u8 *rcv_wscale)
162 unsigned int space = (__space < 0 ? 0 : __space);
164 /* If no clamp set the clamp to the max possible scaled window */
165 if (*window_clamp == 0)
166 (*window_clamp) = (65535 << 14);
167 space = min(*window_clamp, space);
169 /* Quantize space offering to a multiple of mss if possible. */
171 space = (space / mss) * mss;
173 /* NOTE: offering an initial window larger than 32767
174 * will break some buggy TCP stacks. We try to be nice.
175 * If we are not window scaling, then this truncates
176 * our initial window offering to 32k. There should also
177 * be a sysctl option to stop being nice.
179 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
182 /* Set window scaling on max possible window
183 * See RFC1323 for an explanation of the limit to 14
185 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
186 while (space > 65535 && (*rcv_wscale) < 14) {
192 /* Set initial window to value enough for senders,
193 * following RFC1414. Senders, not following this RFC,
194 * will be satisfied with 2.
196 if (mss > (1<<*rcv_wscale)) {
202 if (*rcv_wnd > init_cwnd*mss)
203 *rcv_wnd = init_cwnd*mss;
206 /* Set the clamp no higher than max representable value */
207 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
210 /* Chose a new window to advertise, update state in tcp_sock for the
211 * socket, and return result with RFC1323 scaling applied. The return
212 * value can be stuffed directly into th->window for an outgoing
215 static __inline__ u16 tcp_select_window(struct sock *sk)
217 struct tcp_sock *tp = tcp_sk(sk);
218 u32 cur_win = tcp_receive_window(tp);
219 u32 new_win = __tcp_select_window(sk);
221 /* Never shrink the offered window */
222 if(new_win < cur_win) {
223 /* Danger Will Robinson!
224 * Don't update rcv_wup/rcv_wnd here or else
225 * we will not be able to advertise a zero
226 * window in time. --DaveM
228 * Relax Will Robinson.
232 tp->rcv_wnd = new_win;
233 tp->rcv_wup = tp->rcv_nxt;
235 /* Make sure we do not exceed the maximum possible
238 if (!tp->rx_opt.rcv_wscale)
239 new_win = min(new_win, MAX_TCP_WINDOW);
241 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
243 /* RFC1323 scaling applied */
244 new_win >>= tp->rx_opt.rcv_wscale;
246 /* If we advertise zero window, disable fast path. */
254 /* This routine actually transmits TCP packets queued in by
255 * tcp_do_sendmsg(). This is used by both the initial
256 * transmission and possible later retransmissions.
257 * All SKB's seen here are completely headerless. It is our
258 * job to build the TCP header, and pass the packet down to
259 * IP so it can do the same plus pass the packet off to the
262 * We are working here with either a clone of the original
263 * SKB, or a fresh unique copy made by the retransmit engine.
265 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb)
268 struct inet_sock *inet = inet_sk(sk);
269 struct tcp_sock *tp = tcp_sk(sk);
270 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
271 int tcp_header_size = tp->tcp_header_len;
276 BUG_ON(!tcp_skb_pcount(skb));
278 #define SYSCTL_FLAG_TSTAMPS 0x1
279 #define SYSCTL_FLAG_WSCALE 0x2
280 #define SYSCTL_FLAG_SACK 0x4
282 /* If congestion control is doing timestamping */
283 if (tp->ca_ops->rtt_sample)
284 do_gettimeofday(&skb->stamp);
287 if (tcb->flags & TCPCB_FLAG_SYN) {
288 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
289 if(sysctl_tcp_timestamps) {
290 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
291 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
293 if(sysctl_tcp_window_scaling) {
294 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
295 sysctl_flags |= SYSCTL_FLAG_WSCALE;
297 if(sysctl_tcp_sack) {
298 sysctl_flags |= SYSCTL_FLAG_SACK;
299 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
300 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
302 } else if (tp->rx_opt.eff_sacks) {
303 /* A SACK is 2 pad bytes, a 2 byte header, plus
304 * 2 32-bit sequence numbers for each SACK block.
306 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
307 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
310 if (tcp_packets_in_flight(tp) == 0)
311 tcp_ca_event(tp, CA_EVENT_TX_START);
313 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
315 skb_set_owner_w(skb, sk);
317 /* Build TCP header and checksum it. */
318 th->source = inet->sport;
319 th->dest = inet->dport;
320 th->seq = htonl(tcb->seq);
321 th->ack_seq = htonl(tp->rcv_nxt);
322 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags);
323 if (tcb->flags & TCPCB_FLAG_SYN) {
324 /* RFC1323: The window in SYN & SYN/ACK segments
327 th->window = htons(tp->rcv_wnd);
329 th->window = htons(tcp_select_window(sk));
335 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) {
336 th->urg_ptr = htons(tp->snd_up-tcb->seq);
340 if (tcb->flags & TCPCB_FLAG_SYN) {
341 tcp_syn_build_options((__u32 *)(th + 1),
342 tcp_advertise_mss(sk),
343 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
344 (sysctl_flags & SYSCTL_FLAG_SACK),
345 (sysctl_flags & SYSCTL_FLAG_WSCALE),
346 tp->rx_opt.rcv_wscale,
348 tp->rx_opt.ts_recent);
350 tcp_build_and_update_options((__u32 *)(th + 1),
353 TCP_ECN_send(sk, tp, skb, tcp_header_size);
355 tp->af_specific->send_check(sk, th, skb->len, skb);
357 if (tcb->flags & TCPCB_FLAG_ACK)
358 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
360 if (skb->len != tcp_header_size)
361 tcp_event_data_sent(tp, skb, sk);
363 TCP_INC_STATS(TCP_MIB_OUTSEGS);
365 err = tp->af_specific->queue_xmit(skb, 0);
371 /* NET_XMIT_CN is special. It does not guarantee,
372 * that this packet is lost. It tells that device
373 * is about to start to drop packets or already
374 * drops some packets of the same priority and
375 * invokes us to send less aggressively.
377 return err == NET_XMIT_CN ? 0 : err;
380 #undef SYSCTL_FLAG_TSTAMPS
381 #undef SYSCTL_FLAG_WSCALE
382 #undef SYSCTL_FLAG_SACK
386 /* This routine just queue's the buffer
388 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
389 * otherwise socket can stall.
391 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
393 struct tcp_sock *tp = tcp_sk(sk);
395 /* Advance write_seq and place onto the write_queue. */
396 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
397 skb_header_release(skb);
398 __skb_queue_tail(&sk->sk_write_queue, skb);
399 sk_charge_skb(sk, skb);
401 /* Queue it, remembering where we must start sending. */
402 if (sk->sk_send_head == NULL)
403 sk->sk_send_head = skb;
406 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
408 if (skb->len <= mss_now ||
409 !(sk->sk_route_caps & NETIF_F_TSO)) {
410 /* Avoid the costly divide in the normal
413 skb_shinfo(skb)->tso_segs = 1;
414 skb_shinfo(skb)->tso_size = 0;
418 factor = skb->len + (mss_now - 1);
420 skb_shinfo(skb)->tso_segs = factor;
421 skb_shinfo(skb)->tso_size = mss_now;
425 /* Function to create two new TCP segments. Shrinks the given segment
426 * to the specified size and appends a new segment with the rest of the
427 * packet to the list. This won't be called frequently, I hope.
428 * Remember, these are still headerless SKBs at this point.
430 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
432 struct tcp_sock *tp = tcp_sk(sk);
433 struct sk_buff *buff;
437 nsize = skb_headlen(skb) - len;
441 if (skb_cloned(skb) &&
442 skb_is_nonlinear(skb) &&
443 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
446 /* Get a new skb... force flag on. */
447 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
449 return -ENOMEM; /* We'll just try again later. */
450 sk_charge_skb(sk, buff);
452 /* Correct the sequence numbers. */
453 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
454 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
455 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
457 /* PSH and FIN should only be set in the second packet. */
458 flags = TCP_SKB_CB(skb)->flags;
459 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
460 TCP_SKB_CB(buff)->flags = flags;
461 TCP_SKB_CB(buff)->sacked =
462 (TCP_SKB_CB(skb)->sacked &
463 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL));
464 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
466 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
467 /* Copy and checksum data tail into the new buffer. */
468 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
473 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
475 skb->ip_summed = CHECKSUM_HW;
476 skb_split(skb, buff, len);
479 buff->ip_summed = skb->ip_summed;
481 /* Looks stupid, but our code really uses when of
482 * skbs, which it never sent before. --ANK
484 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
485 buff->stamp = skb->stamp;
487 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
488 tp->lost_out -= tcp_skb_pcount(skb);
489 tp->left_out -= tcp_skb_pcount(skb);
492 /* Fix up tso_factor for both original and new SKB. */
493 tcp_set_skb_tso_segs(sk, skb, mss_now);
494 tcp_set_skb_tso_segs(sk, buff, mss_now);
496 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
497 tp->lost_out += tcp_skb_pcount(skb);
498 tp->left_out += tcp_skb_pcount(skb);
501 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) {
502 tp->lost_out += tcp_skb_pcount(buff);
503 tp->left_out += tcp_skb_pcount(buff);
506 /* Link BUFF into the send queue. */
507 skb_header_release(buff);
508 __skb_append(skb, buff);
513 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
514 * eventually). The difference is that pulled data not copied, but
515 * immediately discarded.
517 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len)
523 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
524 if (skb_shinfo(skb)->frags[i].size <= eat) {
525 put_page(skb_shinfo(skb)->frags[i].page);
526 eat -= skb_shinfo(skb)->frags[i].size;
528 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
530 skb_shinfo(skb)->frags[k].page_offset += eat;
531 skb_shinfo(skb)->frags[k].size -= eat;
537 skb_shinfo(skb)->nr_frags = k;
539 skb->tail = skb->data;
540 skb->data_len -= len;
541 skb->len = skb->data_len;
545 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
547 if (skb_cloned(skb) &&
548 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
551 if (len <= skb_headlen(skb)) {
552 __skb_pull(skb, len);
554 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL)
558 TCP_SKB_CB(skb)->seq += len;
559 skb->ip_summed = CHECKSUM_HW;
561 skb->truesize -= len;
562 sk->sk_wmem_queued -= len;
563 sk->sk_forward_alloc += len;
564 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
566 /* Any change of skb->len requires recalculation of tso
569 if (tcp_skb_pcount(skb) > 1)
570 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
575 /* This function synchronize snd mss to current pmtu/exthdr set.
577 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
578 for TCP options, but includes only bare TCP header.
580 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
581 It is minumum of user_mss and mss received with SYN.
582 It also does not include TCP options.
584 tp->pmtu_cookie is last pmtu, seen by this function.
586 tp->mss_cache is current effective sending mss, including
587 all tcp options except for SACKs. It is evaluated,
588 taking into account current pmtu, but never exceeds
589 tp->rx_opt.mss_clamp.
591 NOTE1. rfc1122 clearly states that advertised MSS
592 DOES NOT include either tcp or ip options.
594 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside
595 this function. --ANK (980731)
598 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
600 struct tcp_sock *tp = tcp_sk(sk);
603 /* Calculate base mss without TCP options:
604 It is MMS_S - sizeof(tcphdr) of rfc1122
606 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr);
608 /* Clamp it (mss_clamp does not include tcp options) */
609 if (mss_now > tp->rx_opt.mss_clamp)
610 mss_now = tp->rx_opt.mss_clamp;
612 /* Now subtract optional transport overhead */
613 mss_now -= tp->ext_header_len;
615 /* Then reserve room for full set of TCP options and 8 bytes of data */
619 /* Now subtract TCP options size, not including SACKs */
620 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
622 /* Bound mss with half of window */
623 if (tp->max_window && mss_now > (tp->max_window>>1))
624 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
626 /* And store cached results */
627 tp->pmtu_cookie = pmtu;
628 tp->mss_cache = mss_now;
633 /* Compute the current effective MSS, taking SACKs and IP options,
634 * and even PMTU discovery events into account.
636 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
637 * cannot be large. However, taking into account rare use of URG, this
640 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
642 struct tcp_sock *tp = tcp_sk(sk);
643 struct dst_entry *dst = __sk_dst_get(sk);
648 mss_now = tp->mss_cache;
651 (sk->sk_route_caps & NETIF_F_TSO) &&
656 u32 mtu = dst_mtu(dst);
657 if (mtu != tp->pmtu_cookie)
658 mss_now = tcp_sync_mss(sk, mtu);
661 if (tp->rx_opt.eff_sacks)
662 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
663 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
665 xmit_size_goal = mss_now;
668 xmit_size_goal = 65535 -
669 tp->af_specific->net_header_len -
670 tp->ext_header_len - tp->tcp_header_len;
672 if (tp->max_window &&
673 (xmit_size_goal > (tp->max_window >> 1)))
674 xmit_size_goal = max((tp->max_window >> 1),
675 68U - tp->tcp_header_len);
677 xmit_size_goal -= (xmit_size_goal % mss_now);
679 tp->xmit_size_goal = xmit_size_goal;
684 /* Congestion window validation. (RFC2861) */
686 static inline void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
688 __u32 packets_out = tp->packets_out;
690 if (packets_out >= tp->snd_cwnd) {
691 /* Network is feed fully. */
692 tp->snd_cwnd_used = 0;
693 tp->snd_cwnd_stamp = tcp_time_stamp;
695 /* Network starves. */
696 if (tp->packets_out > tp->snd_cwnd_used)
697 tp->snd_cwnd_used = tp->packets_out;
699 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= tp->rto)
700 tcp_cwnd_application_limited(sk);
704 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
706 u32 window, cwnd_len;
708 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
709 cwnd_len = mss_now * cwnd;
710 return min(window, cwnd_len);
713 /* Can at least one segment of SKB be sent right now, according to the
714 * congestion window rules? If so, return how many segments are allowed.
716 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
720 /* Don't be strict about the congestion window for the final FIN. */
721 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
724 in_flight = tcp_packets_in_flight(tp);
726 if (in_flight < cwnd)
727 return (cwnd - in_flight);
732 /* This must be invoked the first time we consider transmitting
735 static inline int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
737 int tso_segs = tcp_skb_pcount(skb);
741 skb_shinfo(skb)->tso_size != mss_now)) {
742 tcp_set_skb_tso_segs(sk, skb, mss_now);
743 tso_segs = tcp_skb_pcount(skb);
748 static inline int tcp_minshall_check(const struct tcp_sock *tp)
750 return after(tp->snd_sml,tp->snd_una) &&
751 !after(tp->snd_sml, tp->snd_nxt);
754 /* Return 0, if packet can be sent now without violation Nagle's rules:
755 * 1. It is full sized.
756 * 2. Or it contains FIN. (already checked by caller)
757 * 3. Or TCP_NODELAY was set.
758 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
759 * With Minshall's modification: all sent small packets are ACKed.
762 static inline int tcp_nagle_check(const struct tcp_sock *tp,
763 const struct sk_buff *skb,
764 unsigned mss_now, int nonagle)
766 return (skb->len < mss_now &&
767 ((nonagle&TCP_NAGLE_CORK) ||
770 tcp_minshall_check(tp))));
773 /* Return non-zero if the Nagle test allows this packet to be
776 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
777 unsigned int cur_mss, int nonagle)
779 /* Nagle rule does not apply to frames, which sit in the middle of the
780 * write_queue (they have no chances to get new data).
782 * This is implemented in the callers, where they modify the 'nonagle'
783 * argument based upon the location of SKB in the send queue.
785 if (nonagle & TCP_NAGLE_PUSH)
788 /* Don't use the nagle rule for urgent data (or for the final FIN). */
790 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
793 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
799 /* Does at least the first segment of SKB fit into the send window? */
800 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
802 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
804 if (skb->len > cur_mss)
805 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
807 return !after(end_seq, tp->snd_una + tp->snd_wnd);
810 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
811 * should be put on the wire right now. If so, it returns the number of
812 * packets allowed by the congestion window.
814 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
815 unsigned int cur_mss, int nonagle)
817 struct tcp_sock *tp = tcp_sk(sk);
818 unsigned int cwnd_quota;
820 tcp_init_tso_segs(sk, skb, cur_mss);
822 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
825 cwnd_quota = tcp_cwnd_test(tp, skb);
827 !tcp_snd_wnd_test(tp, skb, cur_mss))
833 static inline int tcp_skb_is_last(const struct sock *sk,
834 const struct sk_buff *skb)
836 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
839 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
841 struct sk_buff *skb = sk->sk_send_head;
844 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
845 (tcp_skb_is_last(sk, skb) ?
850 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
851 * which is put after SKB on the list. It is very much like
852 * tcp_fragment() except that it may make several kinds of assumptions
853 * in order to speed up the splitting operation. In particular, we
854 * know that all the data is in scatter-gather pages, and that the
855 * packet has never been sent out before (and thus is not cloned).
857 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
859 struct sk_buff *buff;
860 int nlen = skb->len - len;
863 /* All of a TSO frame must be composed of paged data. */
864 BUG_ON(skb->len != skb->data_len);
866 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
867 if (unlikely(buff == NULL))
870 buff->truesize = nlen;
871 skb->truesize -= nlen;
873 /* Correct the sequence numbers. */
874 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
875 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
876 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
878 /* PSH and FIN should only be set in the second packet. */
879 flags = TCP_SKB_CB(skb)->flags;
880 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
881 TCP_SKB_CB(buff)->flags = flags;
883 /* This packet was never sent out yet, so no SACK bits. */
884 TCP_SKB_CB(buff)->sacked = 0;
886 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
887 skb_split(skb, buff, len);
889 /* Fix up tso_factor for both original and new SKB. */
890 tcp_set_skb_tso_segs(sk, skb, mss_now);
891 tcp_set_skb_tso_segs(sk, buff, mss_now);
893 /* Link BUFF into the send queue. */
894 skb_header_release(buff);
895 __skb_append(skb, buff);
900 /* Try to defer sending, if possible, in order to minimize the amount
901 * of TSO splitting we do. View it as a kind of TSO Nagle test.
903 * This algorithm is from John Heffner.
905 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
907 u32 send_win, cong_win, limit, in_flight;
909 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
912 if (tp->ca_state != TCP_CA_Open)
915 in_flight = tcp_packets_in_flight(tp);
917 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
918 (tp->snd_cwnd <= in_flight));
920 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
922 /* From in_flight test above, we know that cwnd > in_flight. */
923 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
925 limit = min(send_win, cong_win);
927 /* If sk_send_head can be sent fully now, just do it. */
928 if (skb->len <= limit)
931 if (sysctl_tcp_tso_win_divisor) {
932 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
934 /* If at least some fraction of a window is available,
937 chunk /= sysctl_tcp_tso_win_divisor;
941 /* Different approach, try not to defer past a single
942 * ACK. Receiver should ACK every other full sized
943 * frame, so if we have space for more than 3 frames
946 if (limit > tcp_max_burst(tp) * tp->mss_cache)
950 /* Ok, it looks like it is advisable to defer. */
954 /* This routine writes packets to the network. It advances the
955 * send_head. This happens as incoming acks open up the remote
958 * Returns 1, if no segments are in flight and we have queued segments, but
959 * cannot send anything now because of SWS or another problem.
961 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
963 struct tcp_sock *tp = tcp_sk(sk);
965 unsigned int tso_segs, sent_pkts;
968 /* If we are closed, the bytes will have to remain here.
969 * In time closedown will finish, we empty the write queue and all
972 if (unlikely(sk->sk_state == TCP_CLOSE))
976 while ((skb = sk->sk_send_head)) {
977 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
980 cwnd_quota = tcp_cwnd_test(tp, skb);
984 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
988 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
989 (tcp_skb_is_last(sk, skb) ?
990 nonagle : TCP_NAGLE_PUSH))))
993 if (tcp_tso_should_defer(sk, tp, skb))
998 u32 limit = tcp_window_allows(tp, skb,
999 mss_now, cwnd_quota);
1001 if (skb->len < limit) {
1002 unsigned int trim = skb->len % mss_now;
1005 limit = skb->len - trim;
1007 if (skb->len > limit) {
1008 if (tso_fragment(sk, skb, limit, mss_now))
1011 } else if (unlikely(skb->len > mss_now)) {
1012 if (unlikely(tcp_fragment(sk, skb, mss_now, mss_now)))
1016 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1018 if (unlikely(tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))))
1021 /* Advance the send_head. This one is sent out.
1022 * This call will increment packets_out.
1024 update_send_head(sk, tp, skb);
1026 tcp_minshall_update(tp, mss_now, skb);
1030 if (likely(sent_pkts)) {
1031 tcp_cwnd_validate(sk, tp);
1034 return !tp->packets_out && sk->sk_send_head;
1037 /* Push out any pending frames which were held back due to
1038 * TCP_CORK or attempt at coalescing tiny packets.
1039 * The socket must be locked by the caller.
1041 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1042 unsigned int cur_mss, int nonagle)
1044 struct sk_buff *skb = sk->sk_send_head;
1047 if (tcp_write_xmit(sk, cur_mss, nonagle))
1048 tcp_check_probe_timer(sk, tp);
1052 /* Send _single_ skb sitting at the send head. This function requires
1053 * true push pending frames to setup probe timer etc.
1055 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1057 struct tcp_sock *tp = tcp_sk(sk);
1058 struct sk_buff *skb = sk->sk_send_head;
1059 unsigned int tso_segs, cwnd_quota;
1061 BUG_ON(!skb || skb->len < mss_now);
1063 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1064 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1066 if (likely(cwnd_quota)) {
1070 u32 limit = tcp_window_allows(tp, skb,
1071 mss_now, cwnd_quota);
1073 if (skb->len < limit) {
1074 unsigned int trim = skb->len % mss_now;
1077 limit = skb->len - trim;
1079 if (skb->len > limit) {
1080 if (unlikely(tso_fragment(sk, skb, limit, mss_now)))
1083 } else if (unlikely(skb->len > mss_now)) {
1084 if (unlikely(tcp_fragment(sk, skb, mss_now, mss_now)))
1088 /* Send it out now. */
1089 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1091 if (likely(!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation)))) {
1092 update_send_head(sk, tp, skb);
1093 tcp_cwnd_validate(sk, tp);
1099 /* This function returns the amount that we can raise the
1100 * usable window based on the following constraints
1102 * 1. The window can never be shrunk once it is offered (RFC 793)
1103 * 2. We limit memory per socket
1106 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1107 * RECV.NEXT + RCV.WIN fixed until:
1108 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1110 * i.e. don't raise the right edge of the window until you can raise
1111 * it at least MSS bytes.
1113 * Unfortunately, the recommended algorithm breaks header prediction,
1114 * since header prediction assumes th->window stays fixed.
1116 * Strictly speaking, keeping th->window fixed violates the receiver
1117 * side SWS prevention criteria. The problem is that under this rule
1118 * a stream of single byte packets will cause the right side of the
1119 * window to always advance by a single byte.
1121 * Of course, if the sender implements sender side SWS prevention
1122 * then this will not be a problem.
1124 * BSD seems to make the following compromise:
1126 * If the free space is less than the 1/4 of the maximum
1127 * space available and the free space is less than 1/2 mss,
1128 * then set the window to 0.
1129 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1130 * Otherwise, just prevent the window from shrinking
1131 * and from being larger than the largest representable value.
1133 * This prevents incremental opening of the window in the regime
1134 * where TCP is limited by the speed of the reader side taking
1135 * data out of the TCP receive queue. It does nothing about
1136 * those cases where the window is constrained on the sender side
1137 * because the pipeline is full.
1139 * BSD also seems to "accidentally" limit itself to windows that are a
1140 * multiple of MSS, at least until the free space gets quite small.
1141 * This would appear to be a side effect of the mbuf implementation.
1142 * Combining these two algorithms results in the observed behavior
1143 * of having a fixed window size at almost all times.
1145 * Below we obtain similar behavior by forcing the offered window to
1146 * a multiple of the mss when it is feasible to do so.
1148 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1149 * Regular options like TIMESTAMP are taken into account.
1151 u32 __tcp_select_window(struct sock *sk)
1153 struct tcp_sock *tp = tcp_sk(sk);
1154 /* MSS for the peer's data. Previous verions used mss_clamp
1155 * here. I don't know if the value based on our guesses
1156 * of peer's MSS is better for the performance. It's more correct
1157 * but may be worse for the performance because of rcv_mss
1158 * fluctuations. --SAW 1998/11/1
1160 int mss = tp->ack.rcv_mss;
1161 int free_space = tcp_space(sk);
1162 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1165 if (mss > full_space)
1168 if (free_space < full_space/2) {
1171 if (tcp_memory_pressure)
1172 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1174 if (free_space < mss)
1178 if (free_space > tp->rcv_ssthresh)
1179 free_space = tp->rcv_ssthresh;
1181 /* Don't do rounding if we are using window scaling, since the
1182 * scaled window will not line up with the MSS boundary anyway.
1184 window = tp->rcv_wnd;
1185 if (tp->rx_opt.rcv_wscale) {
1186 window = free_space;
1188 /* Advertise enough space so that it won't get scaled away.
1189 * Import case: prevent zero window announcement if
1190 * 1<<rcv_wscale > mss.
1192 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1193 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1194 << tp->rx_opt.rcv_wscale);
1196 /* Get the largest window that is a nice multiple of mss.
1197 * Window clamp already applied above.
1198 * If our current window offering is within 1 mss of the
1199 * free space we just keep it. This prevents the divide
1200 * and multiply from happening most of the time.
1201 * We also don't do any window rounding when the free space
1204 if (window <= free_space - mss || window > free_space)
1205 window = (free_space/mss)*mss;
1211 /* Attempt to collapse two adjacent SKB's during retransmission. */
1212 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1214 struct tcp_sock *tp = tcp_sk(sk);
1215 struct sk_buff *next_skb = skb->next;
1217 /* The first test we must make is that neither of these two
1218 * SKB's are still referenced by someone else.
1220 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1221 int skb_size = skb->len, next_skb_size = next_skb->len;
1222 u16 flags = TCP_SKB_CB(skb)->flags;
1224 /* Also punt if next skb has been SACK'd. */
1225 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1228 /* Next skb is out of window. */
1229 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1232 /* Punt if not enough space exists in the first SKB for
1233 * the data in the second, or the total combined payload
1234 * would exceed the MSS.
1236 if ((next_skb_size > skb_tailroom(skb)) ||
1237 ((skb_size + next_skb_size) > mss_now))
1240 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1241 tcp_skb_pcount(next_skb) != 1);
1243 /* Ok. We will be able to collapse the packet. */
1244 __skb_unlink(next_skb, next_skb->list);
1246 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1248 if (next_skb->ip_summed == CHECKSUM_HW)
1249 skb->ip_summed = CHECKSUM_HW;
1251 if (skb->ip_summed != CHECKSUM_HW)
1252 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1254 /* Update sequence range on original skb. */
1255 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1257 /* Merge over control information. */
1258 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1259 TCP_SKB_CB(skb)->flags = flags;
1261 /* All done, get rid of second SKB and account for it so
1262 * packet counting does not break.
1264 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1265 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1266 tp->retrans_out -= tcp_skb_pcount(next_skb);
1267 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1268 tp->lost_out -= tcp_skb_pcount(next_skb);
1269 tp->left_out -= tcp_skb_pcount(next_skb);
1271 /* Reno case is special. Sigh... */
1272 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1273 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1274 tp->left_out -= tcp_skb_pcount(next_skb);
1277 /* Not quite right: it can be > snd.fack, but
1278 * it is better to underestimate fackets.
1280 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1281 tcp_packets_out_dec(tp, next_skb);
1282 sk_stream_free_skb(sk, next_skb);
1286 /* Do a simple retransmit without using the backoff mechanisms in
1287 * tcp_timer. This is used for path mtu discovery.
1288 * The socket is already locked here.
1290 void tcp_simple_retransmit(struct sock *sk)
1292 struct tcp_sock *tp = tcp_sk(sk);
1293 struct sk_buff *skb;
1294 unsigned int mss = tcp_current_mss(sk, 0);
1297 sk_stream_for_retrans_queue(skb, sk) {
1298 if (skb->len > mss &&
1299 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1300 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1301 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1302 tp->retrans_out -= tcp_skb_pcount(skb);
1304 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1305 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1306 tp->lost_out += tcp_skb_pcount(skb);
1315 tcp_sync_left_out(tp);
1317 /* Don't muck with the congestion window here.
1318 * Reason is that we do not increase amount of _data_
1319 * in network, but units changed and effective
1320 * cwnd/ssthresh really reduced now.
1322 if (tp->ca_state != TCP_CA_Loss) {
1323 tp->high_seq = tp->snd_nxt;
1324 tp->snd_ssthresh = tcp_current_ssthresh(tp);
1325 tp->prior_ssthresh = 0;
1326 tp->undo_marker = 0;
1327 tcp_set_ca_state(tp, TCP_CA_Loss);
1329 tcp_xmit_retransmit_queue(sk);
1332 /* This retransmits one SKB. Policy decisions and retransmit queue
1333 * state updates are done by the caller. Returns non-zero if an
1334 * error occurred which prevented the send.
1336 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1338 struct tcp_sock *tp = tcp_sk(sk);
1339 unsigned int cur_mss = tcp_current_mss(sk, 0);
1342 /* Do not sent more than we queued. 1/4 is reserved for possible
1343 * copying overhead: frgagmentation, tunneling, mangling etc.
1345 if (atomic_read(&sk->sk_wmem_alloc) >
1346 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1349 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1350 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1353 if (sk->sk_route_caps & NETIF_F_TSO) {
1354 sk->sk_route_caps &= ~NETIF_F_TSO;
1355 sock_set_flag(sk, SOCK_NO_LARGESEND);
1358 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1362 /* If receiver has shrunk his window, and skb is out of
1363 * new window, do not retransmit it. The exception is the
1364 * case, when window is shrunk to zero. In this case
1365 * our retransmit serves as a zero window probe.
1367 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1368 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1371 if (skb->len > cur_mss) {
1372 int old_factor = tcp_skb_pcount(skb);
1375 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1376 return -ENOMEM; /* We'll try again later. */
1378 /* New SKB created, account for it. */
1379 diff = old_factor - tcp_skb_pcount(skb) -
1380 tcp_skb_pcount(skb->next);
1381 tp->packets_out -= diff;
1384 tp->fackets_out -= diff;
1385 if ((int)tp->fackets_out < 0)
1386 tp->fackets_out = 0;
1390 /* Collapse two adjacent packets if worthwhile and we can. */
1391 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1392 (skb->len < (cur_mss >> 1)) &&
1393 (skb->next != sk->sk_send_head) &&
1394 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1395 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1396 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1397 (sysctl_tcp_retrans_collapse != 0))
1398 tcp_retrans_try_collapse(sk, skb, cur_mss);
1400 if(tp->af_specific->rebuild_header(sk))
1401 return -EHOSTUNREACH; /* Routing failure or similar. */
1403 /* Some Solaris stacks overoptimize and ignore the FIN on a
1404 * retransmit when old data is attached. So strip it off
1405 * since it is cheap to do so and saves bytes on the network.
1408 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1409 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1410 if (!pskb_trim(skb, 0)) {
1411 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1412 skb_shinfo(skb)->tso_segs = 1;
1413 skb_shinfo(skb)->tso_size = 0;
1414 skb->ip_summed = CHECKSUM_NONE;
1419 /* Make a copy, if the first transmission SKB clone we made
1420 * is still in somebody's hands, else make a clone.
1422 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1424 err = tcp_transmit_skb(sk, (skb_cloned(skb) ?
1425 pskb_copy(skb, GFP_ATOMIC):
1426 skb_clone(skb, GFP_ATOMIC)));
1429 /* Update global TCP statistics. */
1430 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1432 tp->total_retrans++;
1434 #if FASTRETRANS_DEBUG > 0
1435 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1436 if (net_ratelimit())
1437 printk(KERN_DEBUG "retrans_out leaked.\n");
1440 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1441 tp->retrans_out += tcp_skb_pcount(skb);
1443 /* Save stamp of the first retransmit. */
1444 if (!tp->retrans_stamp)
1445 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1449 /* snd_nxt is stored to detect loss of retransmitted segment,
1450 * see tcp_input.c tcp_sacktag_write_queue().
1452 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1457 /* This gets called after a retransmit timeout, and the initially
1458 * retransmitted data is acknowledged. It tries to continue
1459 * resending the rest of the retransmit queue, until either
1460 * we've sent it all or the congestion window limit is reached.
1461 * If doing SACK, the first ACK which comes back for a timeout
1462 * based retransmit packet might feed us FACK information again.
1463 * If so, we use it to avoid unnecessarily retransmissions.
1465 void tcp_xmit_retransmit_queue(struct sock *sk)
1467 struct tcp_sock *tp = tcp_sk(sk);
1468 struct sk_buff *skb;
1469 int packet_cnt = tp->lost_out;
1471 /* First pass: retransmit lost packets. */
1473 sk_stream_for_retrans_queue(skb, sk) {
1474 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1476 /* Assume this retransmit will generate
1477 * only one packet for congestion window
1478 * calculation purposes. This works because
1479 * tcp_retransmit_skb() will chop up the
1480 * packet to be MSS sized and all the
1481 * packet counting works out.
1483 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1486 if (sacked&TCPCB_LOST) {
1487 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1488 if (tcp_retransmit_skb(sk, skb))
1490 if (tp->ca_state != TCP_CA_Loss)
1491 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1493 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1496 skb_peek(&sk->sk_write_queue))
1497 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1500 packet_cnt -= tcp_skb_pcount(skb);
1501 if (packet_cnt <= 0)
1507 /* OK, demanded retransmission is finished. */
1509 /* Forward retransmissions are possible only during Recovery. */
1510 if (tp->ca_state != TCP_CA_Recovery)
1513 /* No forward retransmissions in Reno are possible. */
1514 if (!tp->rx_opt.sack_ok)
1517 /* Yeah, we have to make difficult choice between forward transmission
1518 * and retransmission... Both ways have their merits...
1520 * For now we do not retransmit anything, while we have some new
1524 if (tcp_may_send_now(sk, tp))
1529 sk_stream_for_retrans_queue(skb, sk) {
1530 /* Similar to the retransmit loop above we
1531 * can pretend that the retransmitted SKB
1532 * we send out here will be composed of one
1533 * real MSS sized packet because tcp_retransmit_skb()
1534 * will fragment it if necessary.
1536 if (++packet_cnt > tp->fackets_out)
1539 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1542 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1545 /* Ok, retransmit it. */
1546 if (tcp_retransmit_skb(sk, skb))
1549 if (skb == skb_peek(&sk->sk_write_queue))
1550 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1552 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1557 /* Send a fin. The caller locks the socket for us. This cannot be
1558 * allowed to fail queueing a FIN frame under any circumstances.
1560 void tcp_send_fin(struct sock *sk)
1562 struct tcp_sock *tp = tcp_sk(sk);
1563 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1566 /* Optimization, tack on the FIN if we have a queue of
1567 * unsent frames. But be careful about outgoing SACKS
1570 mss_now = tcp_current_mss(sk, 1);
1572 if (sk->sk_send_head != NULL) {
1573 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1574 TCP_SKB_CB(skb)->end_seq++;
1577 /* Socket is locked, keep trying until memory is available. */
1579 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL);
1585 /* Reserve space for headers and prepare control bits. */
1586 skb_reserve(skb, MAX_TCP_HEADER);
1588 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1589 TCP_SKB_CB(skb)->sacked = 0;
1590 skb_shinfo(skb)->tso_segs = 1;
1591 skb_shinfo(skb)->tso_size = 0;
1593 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1594 TCP_SKB_CB(skb)->seq = tp->write_seq;
1595 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1596 tcp_queue_skb(sk, skb);
1598 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1601 /* We get here when a process closes a file descriptor (either due to
1602 * an explicit close() or as a byproduct of exit()'ing) and there
1603 * was unread data in the receive queue. This behavior is recommended
1604 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1606 void tcp_send_active_reset(struct sock *sk, unsigned int __nocast priority)
1608 struct tcp_sock *tp = tcp_sk(sk);
1609 struct sk_buff *skb;
1611 /* NOTE: No TCP options attached and we never retransmit this. */
1612 skb = alloc_skb(MAX_TCP_HEADER, priority);
1614 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1618 /* Reserve space for headers and prepare control bits. */
1619 skb_reserve(skb, MAX_TCP_HEADER);
1621 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1622 TCP_SKB_CB(skb)->sacked = 0;
1623 skb_shinfo(skb)->tso_segs = 1;
1624 skb_shinfo(skb)->tso_size = 0;
1627 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1628 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1629 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1630 if (tcp_transmit_skb(sk, skb))
1631 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1634 /* WARNING: This routine must only be called when we have already sent
1635 * a SYN packet that crossed the incoming SYN that caused this routine
1636 * to get called. If this assumption fails then the initial rcv_wnd
1637 * and rcv_wscale values will not be correct.
1639 int tcp_send_synack(struct sock *sk)
1641 struct sk_buff* skb;
1643 skb = skb_peek(&sk->sk_write_queue);
1644 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1645 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1648 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1649 if (skb_cloned(skb)) {
1650 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1653 __skb_unlink(skb, &sk->sk_write_queue);
1654 skb_header_release(nskb);
1655 __skb_queue_head(&sk->sk_write_queue, nskb);
1656 sk_stream_free_skb(sk, skb);
1657 sk_charge_skb(sk, nskb);
1661 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
1662 TCP_ECN_send_synack(tcp_sk(sk), skb);
1664 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1665 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1669 * Prepare a SYN-ACK.
1671 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
1672 struct request_sock *req)
1674 struct inet_request_sock *ireq = inet_rsk(req);
1675 struct tcp_sock *tp = tcp_sk(sk);
1677 int tcp_header_size;
1678 struct sk_buff *skb;
1680 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
1684 /* Reserve space for headers. */
1685 skb_reserve(skb, MAX_TCP_HEADER);
1687 skb->dst = dst_clone(dst);
1689 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
1690 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
1691 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
1692 /* SACK_PERM is in the place of NOP NOP of TS */
1693 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
1694 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
1696 memset(th, 0, sizeof(struct tcphdr));
1699 if (dst->dev->features&NETIF_F_TSO)
1701 TCP_ECN_make_synack(req, th);
1702 th->source = inet_sk(sk)->sport;
1703 th->dest = ireq->rmt_port;
1704 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
1705 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1706 TCP_SKB_CB(skb)->sacked = 0;
1707 skb_shinfo(skb)->tso_segs = 1;
1708 skb_shinfo(skb)->tso_size = 0;
1709 th->seq = htonl(TCP_SKB_CB(skb)->seq);
1710 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
1711 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
1713 /* Set this up on the first call only */
1714 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
1715 /* tcp_full_space because it is guaranteed to be the first packet */
1716 tcp_select_initial_window(tcp_full_space(sk),
1717 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
1722 ireq->rcv_wscale = rcv_wscale;
1725 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
1726 th->window = htons(req->rcv_wnd);
1728 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1729 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
1730 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
1731 TCP_SKB_CB(skb)->when,
1735 th->doff = (tcp_header_size >> 2);
1736 TCP_INC_STATS(TCP_MIB_OUTSEGS);
1741 * Do all connect socket setups that can be done AF independent.
1743 static inline void tcp_connect_init(struct sock *sk)
1745 struct dst_entry *dst = __sk_dst_get(sk);
1746 struct tcp_sock *tp = tcp_sk(sk);
1749 /* We'll fix this up when we get a response from the other end.
1750 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
1752 tp->tcp_header_len = sizeof(struct tcphdr) +
1753 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
1755 /* If user gave his TCP_MAXSEG, record it to clamp */
1756 if (tp->rx_opt.user_mss)
1757 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
1759 tcp_sync_mss(sk, dst_mtu(dst));
1761 if (!tp->window_clamp)
1762 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
1763 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
1764 tcp_initialize_rcv_mss(sk);
1766 tcp_select_initial_window(tcp_full_space(sk),
1767 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
1770 sysctl_tcp_window_scaling,
1773 tp->rx_opt.rcv_wscale = rcv_wscale;
1774 tp->rcv_ssthresh = tp->rcv_wnd;
1777 sock_reset_flag(sk, SOCK_DONE);
1779 tcp_init_wl(tp, tp->write_seq, 0);
1780 tp->snd_una = tp->write_seq;
1781 tp->snd_sml = tp->write_seq;
1786 tp->rto = TCP_TIMEOUT_INIT;
1787 tp->retransmits = 0;
1788 tcp_clear_retrans(tp);
1792 * Build a SYN and send it off.
1794 int tcp_connect(struct sock *sk)
1796 struct tcp_sock *tp = tcp_sk(sk);
1797 struct sk_buff *buff;
1799 tcp_connect_init(sk);
1801 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation);
1802 if (unlikely(buff == NULL))
1805 /* Reserve space for headers. */
1806 skb_reserve(buff, MAX_TCP_HEADER);
1808 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
1809 TCP_ECN_send_syn(sk, tp, buff);
1810 TCP_SKB_CB(buff)->sacked = 0;
1811 skb_shinfo(buff)->tso_segs = 1;
1812 skb_shinfo(buff)->tso_size = 0;
1814 TCP_SKB_CB(buff)->seq = tp->write_seq++;
1815 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
1816 tp->snd_nxt = tp->write_seq;
1817 tp->pushed_seq = tp->write_seq;
1820 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1821 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
1822 skb_header_release(buff);
1823 __skb_queue_tail(&sk->sk_write_queue, buff);
1824 sk_charge_skb(sk, buff);
1825 tp->packets_out += tcp_skb_pcount(buff);
1826 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL));
1827 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
1829 /* Timer for repeating the SYN until an answer. */
1830 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto);
1834 /* Send out a delayed ack, the caller does the policy checking
1835 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
1838 void tcp_send_delayed_ack(struct sock *sk)
1840 struct tcp_sock *tp = tcp_sk(sk);
1841 int ato = tp->ack.ato;
1842 unsigned long timeout;
1844 if (ato > TCP_DELACK_MIN) {
1847 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED))
1848 max_ato = TCP_DELACK_MAX;
1850 /* Slow path, intersegment interval is "high". */
1852 /* If some rtt estimate is known, use it to bound delayed ack.
1853 * Do not use tp->rto here, use results of rtt measurements
1857 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
1863 ato = min(ato, max_ato);
1866 /* Stay within the limit we were given */
1867 timeout = jiffies + ato;
1869 /* Use new timeout only if there wasn't a older one earlier. */
1870 if (tp->ack.pending&TCP_ACK_TIMER) {
1871 /* If delack timer was blocked or is about to expire,
1874 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) {
1879 if (!time_before(timeout, tp->ack.timeout))
1880 timeout = tp->ack.timeout;
1882 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER;
1883 tp->ack.timeout = timeout;
1884 sk_reset_timer(sk, &tp->delack_timer, timeout);
1887 /* This routine sends an ack and also updates the window. */
1888 void tcp_send_ack(struct sock *sk)
1890 /* If we have been reset, we may not send again. */
1891 if (sk->sk_state != TCP_CLOSE) {
1892 struct tcp_sock *tp = tcp_sk(sk);
1893 struct sk_buff *buff;
1895 /* We are not putting this on the write queue, so
1896 * tcp_transmit_skb() will set the ownership to this
1899 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1901 tcp_schedule_ack(tp);
1902 tp->ack.ato = TCP_ATO_MIN;
1903 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX);
1907 /* Reserve space for headers and prepare control bits. */
1908 skb_reserve(buff, MAX_TCP_HEADER);
1910 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
1911 TCP_SKB_CB(buff)->sacked = 0;
1912 skb_shinfo(buff)->tso_segs = 1;
1913 skb_shinfo(buff)->tso_size = 0;
1915 /* Send it off, this clears delayed acks for us. */
1916 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
1917 TCP_SKB_CB(buff)->when = tcp_time_stamp;
1918 tcp_transmit_skb(sk, buff);
1922 /* This routine sends a packet with an out of date sequence
1923 * number. It assumes the other end will try to ack it.
1925 * Question: what should we make while urgent mode?
1926 * 4.4BSD forces sending single byte of data. We cannot send
1927 * out of window data, because we have SND.NXT==SND.MAX...
1929 * Current solution: to send TWO zero-length segments in urgent mode:
1930 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
1931 * out-of-date with SND.UNA-1 to probe window.
1933 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
1935 struct tcp_sock *tp = tcp_sk(sk);
1936 struct sk_buff *skb;
1938 /* We don't queue it, tcp_transmit_skb() sets ownership. */
1939 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
1943 /* Reserve space for headers and set control bits. */
1944 skb_reserve(skb, MAX_TCP_HEADER);
1946 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
1947 TCP_SKB_CB(skb)->sacked = urgent;
1948 skb_shinfo(skb)->tso_segs = 1;
1949 skb_shinfo(skb)->tso_size = 0;
1951 /* Use a previous sequence. This should cause the other
1952 * end to send an ack. Don't queue or clone SKB, just
1955 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
1956 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1957 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1958 return tcp_transmit_skb(sk, skb);
1961 int tcp_write_wakeup(struct sock *sk)
1963 if (sk->sk_state != TCP_CLOSE) {
1964 struct tcp_sock *tp = tcp_sk(sk);
1965 struct sk_buff *skb;
1967 if ((skb = sk->sk_send_head) != NULL &&
1968 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
1970 unsigned int mss = tcp_current_mss(sk, 0);
1971 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
1973 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
1974 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
1976 /* We are probing the opening of a window
1977 * but the window size is != 0
1978 * must have been a result SWS avoidance ( sender )
1980 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
1982 seg_size = min(seg_size, mss);
1983 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1984 if (tcp_fragment(sk, skb, seg_size, mss))
1986 /* SWS override triggered forced fragmentation.
1987 * Disable TSO, the connection is too sick. */
1988 if (sk->sk_route_caps & NETIF_F_TSO) {
1989 sock_set_flag(sk, SOCK_NO_LARGESEND);
1990 sk->sk_route_caps &= ~NETIF_F_TSO;
1992 } else if (!tcp_skb_pcount(skb))
1993 tcp_set_skb_tso_segs(sk, skb, mss);
1995 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
1996 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1997 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC));
1999 update_send_head(sk, tp, skb);
2004 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2005 tcp_xmit_probe_skb(sk, TCPCB_URG);
2006 return tcp_xmit_probe_skb(sk, 0);
2012 /* A window probe timeout has occurred. If window is not closed send
2013 * a partial packet else a zero probe.
2015 void tcp_send_probe0(struct sock *sk)
2017 struct tcp_sock *tp = tcp_sk(sk);
2020 err = tcp_write_wakeup(sk);
2022 if (tp->packets_out || !sk->sk_send_head) {
2023 /* Cancel probe timer, if it is not required. */
2030 if (tp->backoff < sysctl_tcp_retries2)
2033 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
2034 min(tp->rto << tp->backoff, TCP_RTO_MAX));
2036 /* If packet was not sent due to local congestion,
2037 * do not backoff and do not remember probes_out.
2038 * Let local senders to fight for local resources.
2040 * Use accumulated backoff yet.
2042 if (!tp->probes_out)
2044 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0,
2045 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL));
2049 EXPORT_SYMBOL(tcp_connect);
2050 EXPORT_SYMBOL(tcp_make_synack);
2051 EXPORT_SYMBOL(tcp_simple_retransmit);
2052 EXPORT_SYMBOL(tcp_sync_mss);