[PATCH] arch/alpha: Use ARRAY_SIZE macro
[pohmelfs.git] / net / ipv4 / tcp_output.c
blob5c08ea20a18dfb8fa965fc23070729c176aba4f0
1 /*
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 $
10 * Authors: Ross Biro
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
27 * : AF independence
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
39 #include <net/tcp.h>
41 #include <linux/compiler.h>
42 #include <linux/module.h>
43 #include <linux/smp_lock.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows = 0;
53 /* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
57 int sysctl_tcp_tso_win_divisor = 3;
59 int sysctl_tcp_mtu_probing = 0;
60 int sysctl_tcp_base_mss = 512;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle = 1;
65 static void update_send_head(struct sock *sk, struct tcp_sock *tp,
66 struct sk_buff *skb)
68 sk->sk_send_head = skb->next;
69 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue)
70 sk->sk_send_head = NULL;
71 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
72 tcp_packets_out_inc(sk, tp, skb);
75 /* SND.NXT, if window was not shrunk.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp)
83 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt))
84 return tp->snd_nxt;
85 else
86 return tp->snd_una+tp->snd_wnd;
89 /* Calculate mss to advertise in SYN segment.
90 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 * 1. It is independent of path mtu.
93 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
94 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
95 * attached devices, because some buggy hosts are confused by
96 * large MSS.
97 * 4. We do not make 3, we advertise MSS, calculated from first
98 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
99 * This may be overridden via information stored in routing table.
100 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
101 * probably even Jumbo".
103 static __u16 tcp_advertise_mss(struct sock *sk)
105 struct tcp_sock *tp = tcp_sk(sk);
106 struct dst_entry *dst = __sk_dst_get(sk);
107 int mss = tp->advmss;
109 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
110 mss = dst_metric(dst, RTAX_ADVMSS);
111 tp->advmss = mss;
114 return (__u16)mss;
117 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
118 * This is the first part of cwnd validation mechanism. */
119 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
121 struct tcp_sock *tp = tcp_sk(sk);
122 s32 delta = tcp_time_stamp - tp->lsndtime;
123 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
124 u32 cwnd = tp->snd_cwnd;
126 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
128 tp->snd_ssthresh = tcp_current_ssthresh(sk);
129 restart_cwnd = min(restart_cwnd, cwnd);
131 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
132 cwnd >>= 1;
133 tp->snd_cwnd = max(cwnd, restart_cwnd);
134 tp->snd_cwnd_stamp = tcp_time_stamp;
135 tp->snd_cwnd_used = 0;
138 static void tcp_event_data_sent(struct tcp_sock *tp,
139 struct sk_buff *skb, struct sock *sk)
141 struct inet_connection_sock *icsk = inet_csk(sk);
142 const u32 now = tcp_time_stamp;
144 if (sysctl_tcp_slow_start_after_idle &&
145 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
146 tcp_cwnd_restart(sk, __sk_dst_get(sk));
148 tp->lsndtime = now;
150 /* If it is a reply for ato after last received
151 * packet, enter pingpong mode.
153 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
154 icsk->icsk_ack.pingpong = 1;
157 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
159 tcp_dec_quickack_mode(sk, pkts);
160 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
163 /* Determine a window scaling and initial window to offer.
164 * Based on the assumption that the given amount of space
165 * will be offered. Store the results in the tp structure.
166 * NOTE: for smooth operation initial space offering should
167 * be a multiple of mss if possible. We assume here that mss >= 1.
168 * This MUST be enforced by all callers.
170 void tcp_select_initial_window(int __space, __u32 mss,
171 __u32 *rcv_wnd, __u32 *window_clamp,
172 int wscale_ok, __u8 *rcv_wscale)
174 unsigned int space = (__space < 0 ? 0 : __space);
176 /* If no clamp set the clamp to the max possible scaled window */
177 if (*window_clamp == 0)
178 (*window_clamp) = (65535 << 14);
179 space = min(*window_clamp, space);
181 /* Quantize space offering to a multiple of mss if possible. */
182 if (space > mss)
183 space = (space / mss) * mss;
185 /* NOTE: offering an initial window larger than 32767
186 * will break some buggy TCP stacks. If the admin tells us
187 * it is likely we could be speaking with such a buggy stack
188 * we will truncate our initial window offering to 32K-1
189 * unless the remote has sent us a window scaling option,
190 * which we interpret as a sign the remote TCP is not
191 * misinterpreting the window field as a signed quantity.
193 if (sysctl_tcp_workaround_signed_windows)
194 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
195 else
196 (*rcv_wnd) = space;
198 (*rcv_wscale) = 0;
199 if (wscale_ok) {
200 /* Set window scaling on max possible window
201 * See RFC1323 for an explanation of the limit to 14
203 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
204 while (space > 65535 && (*rcv_wscale) < 14) {
205 space >>= 1;
206 (*rcv_wscale)++;
210 /* Set initial window to value enough for senders,
211 * following RFC2414. Senders, not following this RFC,
212 * will be satisfied with 2.
214 if (mss > (1<<*rcv_wscale)) {
215 int init_cwnd = 4;
216 if (mss > 1460*3)
217 init_cwnd = 2;
218 else if (mss > 1460)
219 init_cwnd = 3;
220 if (*rcv_wnd > init_cwnd*mss)
221 *rcv_wnd = init_cwnd*mss;
224 /* Set the clamp no higher than max representable value */
225 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
228 /* Chose a new window to advertise, update state in tcp_sock for the
229 * socket, and return result with RFC1323 scaling applied. The return
230 * value can be stuffed directly into th->window for an outgoing
231 * frame.
233 static u16 tcp_select_window(struct sock *sk)
235 struct tcp_sock *tp = tcp_sk(sk);
236 u32 cur_win = tcp_receive_window(tp);
237 u32 new_win = __tcp_select_window(sk);
239 /* Never shrink the offered window */
240 if(new_win < cur_win) {
241 /* Danger Will Robinson!
242 * Don't update rcv_wup/rcv_wnd here or else
243 * we will not be able to advertise a zero
244 * window in time. --DaveM
246 * Relax Will Robinson.
248 new_win = cur_win;
250 tp->rcv_wnd = new_win;
251 tp->rcv_wup = tp->rcv_nxt;
253 /* Make sure we do not exceed the maximum possible
254 * scaled window.
256 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
257 new_win = min(new_win, MAX_TCP_WINDOW);
258 else
259 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
261 /* RFC1323 scaling applied */
262 new_win >>= tp->rx_opt.rcv_wscale;
264 /* If we advertise zero window, disable fast path. */
265 if (new_win == 0)
266 tp->pred_flags = 0;
268 return new_win;
271 static void tcp_build_and_update_options(__u32 *ptr, struct tcp_sock *tp,
272 __u32 tstamp)
274 if (tp->rx_opt.tstamp_ok) {
275 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) |
276 (TCPOPT_NOP << 16) |
277 (TCPOPT_TIMESTAMP << 8) |
278 TCPOLEN_TIMESTAMP);
279 *ptr++ = htonl(tstamp);
280 *ptr++ = htonl(tp->rx_opt.ts_recent);
282 if (tp->rx_opt.eff_sacks) {
283 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
284 int this_sack;
286 *ptr++ = htonl((TCPOPT_NOP << 24) |
287 (TCPOPT_NOP << 16) |
288 (TCPOPT_SACK << 8) |
289 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
290 TCPOLEN_SACK_PERBLOCK)));
291 for(this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
292 *ptr++ = htonl(sp[this_sack].start_seq);
293 *ptr++ = htonl(sp[this_sack].end_seq);
295 if (tp->rx_opt.dsack) {
296 tp->rx_opt.dsack = 0;
297 tp->rx_opt.eff_sacks--;
302 /* Construct a tcp options header for a SYN or SYN_ACK packet.
303 * If this is every changed make sure to change the definition of
304 * MAX_SYN_SIZE to match the new maximum number of options that you
305 * can generate.
307 static void tcp_syn_build_options(__u32 *ptr, int mss, int ts, int sack,
308 int offer_wscale, int wscale, __u32 tstamp,
309 __u32 ts_recent)
311 /* We always get an MSS option.
312 * The option bytes which will be seen in normal data
313 * packets should timestamps be used, must be in the MSS
314 * advertised. But we subtract them from tp->mss_cache so
315 * that calculations in tcp_sendmsg are simpler etc.
316 * So account for this fact here if necessary. If we
317 * don't do this correctly, as a receiver we won't
318 * recognize data packets as being full sized when we
319 * should, and thus we won't abide by the delayed ACK
320 * rules correctly.
321 * SACKs don't matter, we never delay an ACK when we
322 * have any of those going out.
324 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
325 if (ts) {
326 if(sack)
327 *ptr++ = __constant_htonl((TCPOPT_SACK_PERM << 24) | (TCPOLEN_SACK_PERM << 16) |
328 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
329 else
330 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
331 (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP);
332 *ptr++ = htonl(tstamp); /* TSVAL */
333 *ptr++ = htonl(ts_recent); /* TSECR */
334 } else if(sack)
335 *ptr++ = __constant_htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
336 (TCPOPT_SACK_PERM << 8) | TCPOLEN_SACK_PERM);
337 if (offer_wscale)
338 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_WINDOW << 16) | (TCPOLEN_WINDOW << 8) | (wscale));
341 /* This routine actually transmits TCP packets queued in by
342 * tcp_do_sendmsg(). This is used by both the initial
343 * transmission and possible later retransmissions.
344 * All SKB's seen here are completely headerless. It is our
345 * job to build the TCP header, and pass the packet down to
346 * IP so it can do the same plus pass the packet off to the
347 * device.
349 * We are working here with either a clone of the original
350 * SKB, or a fresh unique copy made by the retransmit engine.
352 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
354 const struct inet_connection_sock *icsk = inet_csk(sk);
355 struct inet_sock *inet;
356 struct tcp_sock *tp;
357 struct tcp_skb_cb *tcb;
358 int tcp_header_size;
359 struct tcphdr *th;
360 int sysctl_flags;
361 int err;
363 BUG_ON(!skb || !tcp_skb_pcount(skb));
365 /* If congestion control is doing timestamping, we must
366 * take such a timestamp before we potentially clone/copy.
368 if (icsk->icsk_ca_ops->rtt_sample)
369 __net_timestamp(skb);
371 if (likely(clone_it)) {
372 if (unlikely(skb_cloned(skb)))
373 skb = pskb_copy(skb, gfp_mask);
374 else
375 skb = skb_clone(skb, gfp_mask);
376 if (unlikely(!skb))
377 return -ENOBUFS;
380 inet = inet_sk(sk);
381 tp = tcp_sk(sk);
382 tcb = TCP_SKB_CB(skb);
383 tcp_header_size = tp->tcp_header_len;
385 #define SYSCTL_FLAG_TSTAMPS 0x1
386 #define SYSCTL_FLAG_WSCALE 0x2
387 #define SYSCTL_FLAG_SACK 0x4
389 sysctl_flags = 0;
390 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
391 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
392 if(sysctl_tcp_timestamps) {
393 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
394 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
396 if (sysctl_tcp_window_scaling) {
397 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
398 sysctl_flags |= SYSCTL_FLAG_WSCALE;
400 if (sysctl_tcp_sack) {
401 sysctl_flags |= SYSCTL_FLAG_SACK;
402 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
403 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
405 } else if (unlikely(tp->rx_opt.eff_sacks)) {
406 /* A SACK is 2 pad bytes, a 2 byte header, plus
407 * 2 32-bit sequence numbers for each SACK block.
409 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
410 (tp->rx_opt.eff_sacks *
411 TCPOLEN_SACK_PERBLOCK));
414 if (tcp_packets_in_flight(tp) == 0)
415 tcp_ca_event(sk, CA_EVENT_TX_START);
417 th = (struct tcphdr *) skb_push(skb, tcp_header_size);
418 skb->h.th = th;
419 skb_set_owner_w(skb, sk);
421 /* Build TCP header and checksum it. */
422 th->source = inet->sport;
423 th->dest = inet->dport;
424 th->seq = htonl(tcb->seq);
425 th->ack_seq = htonl(tp->rcv_nxt);
426 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
427 tcb->flags);
429 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
430 /* RFC1323: The window in SYN & SYN/ACK segments
431 * is never scaled.
433 th->window = htons(tp->rcv_wnd);
434 } else {
435 th->window = htons(tcp_select_window(sk));
437 th->check = 0;
438 th->urg_ptr = 0;
440 if (unlikely(tp->urg_mode &&
441 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) {
442 th->urg_ptr = htons(tp->snd_up-tcb->seq);
443 th->urg = 1;
446 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
447 tcp_syn_build_options((__u32 *)(th + 1),
448 tcp_advertise_mss(sk),
449 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
450 (sysctl_flags & SYSCTL_FLAG_SACK),
451 (sysctl_flags & SYSCTL_FLAG_WSCALE),
452 tp->rx_opt.rcv_wscale,
453 tcb->when,
454 tp->rx_opt.ts_recent);
455 } else {
456 tcp_build_and_update_options((__u32 *)(th + 1),
457 tp, tcb->when);
458 TCP_ECN_send(sk, tp, skb, tcp_header_size);
461 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
463 if (likely(tcb->flags & TCPCB_FLAG_ACK))
464 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
466 if (skb->len != tcp_header_size)
467 tcp_event_data_sent(tp, skb, sk);
469 TCP_INC_STATS(TCP_MIB_OUTSEGS);
471 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
472 if (likely(err <= 0))
473 return err;
475 tcp_enter_cwr(sk);
477 /* NET_XMIT_CN is special. It does not guarantee,
478 * that this packet is lost. It tells that device
479 * is about to start to drop packets or already
480 * drops some packets of the same priority and
481 * invokes us to send less aggressively.
483 return err == NET_XMIT_CN ? 0 : err;
485 #undef SYSCTL_FLAG_TSTAMPS
486 #undef SYSCTL_FLAG_WSCALE
487 #undef SYSCTL_FLAG_SACK
491 /* This routine just queue's the buffer
493 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
494 * otherwise socket can stall.
496 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
498 struct tcp_sock *tp = tcp_sk(sk);
500 /* Advance write_seq and place onto the write_queue. */
501 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
502 skb_header_release(skb);
503 __skb_queue_tail(&sk->sk_write_queue, skb);
504 sk_charge_skb(sk, skb);
506 /* Queue it, remembering where we must start sending. */
507 if (sk->sk_send_head == NULL)
508 sk->sk_send_head = skb;
511 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
513 if (skb->len <= mss_now || !sk_can_gso(sk)) {
514 /* Avoid the costly divide in the normal
515 * non-TSO case.
517 skb_shinfo(skb)->gso_segs = 1;
518 skb_shinfo(skb)->gso_size = 0;
519 skb_shinfo(skb)->gso_type = 0;
520 } else {
521 unsigned int factor;
523 factor = skb->len + (mss_now - 1);
524 factor /= mss_now;
525 skb_shinfo(skb)->gso_segs = factor;
526 skb_shinfo(skb)->gso_size = mss_now;
527 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
531 /* Function to create two new TCP segments. Shrinks the given segment
532 * to the specified size and appends a new segment with the rest of the
533 * packet to the list. This won't be called frequently, I hope.
534 * Remember, these are still headerless SKBs at this point.
536 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now)
538 struct tcp_sock *tp = tcp_sk(sk);
539 struct sk_buff *buff;
540 int nsize, old_factor;
541 int nlen;
542 u16 flags;
544 BUG_ON(len > skb->len);
546 clear_all_retrans_hints(tp);
547 nsize = skb_headlen(skb) - len;
548 if (nsize < 0)
549 nsize = 0;
551 if (skb_cloned(skb) &&
552 skb_is_nonlinear(skb) &&
553 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
554 return -ENOMEM;
556 /* Get a new skb... force flag on. */
557 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
558 if (buff == NULL)
559 return -ENOMEM; /* We'll just try again later. */
561 sk_charge_skb(sk, buff);
562 nlen = skb->len - len - nsize;
563 buff->truesize += nlen;
564 skb->truesize -= nlen;
566 /* Correct the sequence numbers. */
567 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
568 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
569 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
571 /* PSH and FIN should only be set in the second packet. */
572 flags = TCP_SKB_CB(skb)->flags;
573 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
574 TCP_SKB_CB(buff)->flags = flags;
575 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
576 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
578 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) {
579 /* Copy and checksum data tail into the new buffer. */
580 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize),
581 nsize, 0);
583 skb_trim(skb, len);
585 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
586 } else {
587 skb->ip_summed = CHECKSUM_HW;
588 skb_split(skb, buff, len);
591 buff->ip_summed = skb->ip_summed;
593 /* Looks stupid, but our code really uses when of
594 * skbs, which it never sent before. --ANK
596 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
597 buff->tstamp = skb->tstamp;
599 old_factor = tcp_skb_pcount(skb);
601 /* Fix up tso_factor for both original and new SKB. */
602 tcp_set_skb_tso_segs(sk, skb, mss_now);
603 tcp_set_skb_tso_segs(sk, buff, mss_now);
605 /* If this packet has been sent out already, we must
606 * adjust the various packet counters.
608 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
609 int diff = old_factor - tcp_skb_pcount(skb) -
610 tcp_skb_pcount(buff);
612 tp->packets_out -= diff;
614 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
615 tp->sacked_out -= diff;
616 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
617 tp->retrans_out -= diff;
619 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) {
620 tp->lost_out -= diff;
621 tp->left_out -= diff;
624 if (diff > 0) {
625 /* Adjust Reno SACK estimate. */
626 if (!tp->rx_opt.sack_ok) {
627 tp->sacked_out -= diff;
628 if ((int)tp->sacked_out < 0)
629 tp->sacked_out = 0;
630 tcp_sync_left_out(tp);
633 tp->fackets_out -= diff;
634 if ((int)tp->fackets_out < 0)
635 tp->fackets_out = 0;
639 /* Link BUFF into the send queue. */
640 skb_header_release(buff);
641 __skb_append(skb, buff, &sk->sk_write_queue);
643 return 0;
646 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
647 * eventually). The difference is that pulled data not copied, but
648 * immediately discarded.
650 static void __pskb_trim_head(struct sk_buff *skb, int len)
652 int i, k, eat;
654 eat = len;
655 k = 0;
656 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
657 if (skb_shinfo(skb)->frags[i].size <= eat) {
658 put_page(skb_shinfo(skb)->frags[i].page);
659 eat -= skb_shinfo(skb)->frags[i].size;
660 } else {
661 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
662 if (eat) {
663 skb_shinfo(skb)->frags[k].page_offset += eat;
664 skb_shinfo(skb)->frags[k].size -= eat;
665 eat = 0;
667 k++;
670 skb_shinfo(skb)->nr_frags = k;
672 skb->tail = skb->data;
673 skb->data_len -= len;
674 skb->len = skb->data_len;
677 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
679 if (skb_cloned(skb) &&
680 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
681 return -ENOMEM;
683 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
684 if (unlikely(len < skb_headlen(skb)))
685 __skb_pull(skb, len);
686 else
687 __pskb_trim_head(skb, len - skb_headlen(skb));
689 TCP_SKB_CB(skb)->seq += len;
690 skb->ip_summed = CHECKSUM_HW;
692 skb->truesize -= len;
693 sk->sk_wmem_queued -= len;
694 sk->sk_forward_alloc += len;
695 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
697 /* Any change of skb->len requires recalculation of tso
698 * factor and mss.
700 if (tcp_skb_pcount(skb) > 1)
701 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
703 return 0;
706 /* Not accounting for SACKs here. */
707 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
709 struct tcp_sock *tp = tcp_sk(sk);
710 struct inet_connection_sock *icsk = inet_csk(sk);
711 int mss_now;
713 /* Calculate base mss without TCP options:
714 It is MMS_S - sizeof(tcphdr) of rfc1122
716 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
718 /* Clamp it (mss_clamp does not include tcp options) */
719 if (mss_now > tp->rx_opt.mss_clamp)
720 mss_now = tp->rx_opt.mss_clamp;
722 /* Now subtract optional transport overhead */
723 mss_now -= icsk->icsk_ext_hdr_len;
725 /* Then reserve room for full set of TCP options and 8 bytes of data */
726 if (mss_now < 48)
727 mss_now = 48;
729 /* Now subtract TCP options size, not including SACKs */
730 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
732 return mss_now;
735 /* Inverse of above */
736 int tcp_mss_to_mtu(struct sock *sk, int mss)
738 struct tcp_sock *tp = tcp_sk(sk);
739 struct inet_connection_sock *icsk = inet_csk(sk);
740 int mtu;
742 mtu = mss +
743 tp->tcp_header_len +
744 icsk->icsk_ext_hdr_len +
745 icsk->icsk_af_ops->net_header_len;
747 return mtu;
750 void tcp_mtup_init(struct sock *sk)
752 struct tcp_sock *tp = tcp_sk(sk);
753 struct inet_connection_sock *icsk = inet_csk(sk);
755 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
756 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
757 icsk->icsk_af_ops->net_header_len;
758 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
759 icsk->icsk_mtup.probe_size = 0;
762 /* This function synchronize snd mss to current pmtu/exthdr set.
764 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
765 for TCP options, but includes only bare TCP header.
767 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
768 It is minimum of user_mss and mss received with SYN.
769 It also does not include TCP options.
771 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
773 tp->mss_cache is current effective sending mss, including
774 all tcp options except for SACKs. It is evaluated,
775 taking into account current pmtu, but never exceeds
776 tp->rx_opt.mss_clamp.
778 NOTE1. rfc1122 clearly states that advertised MSS
779 DOES NOT include either tcp or ip options.
781 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
782 are READ ONLY outside this function. --ANK (980731)
785 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
787 struct tcp_sock *tp = tcp_sk(sk);
788 struct inet_connection_sock *icsk = inet_csk(sk);
789 int mss_now;
791 if (icsk->icsk_mtup.search_high > pmtu)
792 icsk->icsk_mtup.search_high = pmtu;
794 mss_now = tcp_mtu_to_mss(sk, pmtu);
796 /* Bound mss with half of window */
797 if (tp->max_window && mss_now > (tp->max_window>>1))
798 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
800 /* And store cached results */
801 icsk->icsk_pmtu_cookie = pmtu;
802 if (icsk->icsk_mtup.enabled)
803 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
804 tp->mss_cache = mss_now;
806 return mss_now;
809 /* Compute the current effective MSS, taking SACKs and IP options,
810 * and even PMTU discovery events into account.
812 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
813 * cannot be large. However, taking into account rare use of URG, this
814 * is not a big flaw.
816 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
818 struct tcp_sock *tp = tcp_sk(sk);
819 struct dst_entry *dst = __sk_dst_get(sk);
820 u32 mss_now;
821 u16 xmit_size_goal;
822 int doing_tso = 0;
824 mss_now = tp->mss_cache;
826 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
827 doing_tso = 1;
829 if (dst) {
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;
841 if (doing_tso) {
842 xmit_size_goal = (65535 -
843 inet_csk(sk)->icsk_af_ops->net_header_len -
844 inet_csk(sk)->icsk_ext_hdr_len -
845 tp->tcp_header_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;
856 return mss_now;
859 /* Congestion window validation. (RFC2861) */
861 static void tcp_cwnd_validate(struct sock *sk, struct tcp_sock *tp)
863 __u32 packets_out = tp->packets_out;
865 if (packets_out >= tp->snd_cwnd) {
866 /* Network is feed fully. */
867 tp->snd_cwnd_used = 0;
868 tp->snd_cwnd_stamp = tcp_time_stamp;
869 } else {
870 /* Network starves. */
871 if (tp->packets_out > tp->snd_cwnd_used)
872 tp->snd_cwnd_used = tp->packets_out;
874 if ((s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
875 tcp_cwnd_application_limited(sk);
879 static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd)
881 u32 window, cwnd_len;
883 window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq);
884 cwnd_len = mss_now * cwnd;
885 return min(window, cwnd_len);
888 /* Can at least one segment of SKB be sent right now, according to the
889 * congestion window rules? If so, return how many segments are allowed.
891 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb)
893 u32 in_flight, cwnd;
895 /* Don't be strict about the congestion window for the final FIN. */
896 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
897 return 1;
899 in_flight = tcp_packets_in_flight(tp);
900 cwnd = tp->snd_cwnd;
901 if (in_flight < cwnd)
902 return (cwnd - in_flight);
904 return 0;
907 /* This must be invoked the first time we consider transmitting
908 * SKB onto the wire.
910 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now)
912 int tso_segs = tcp_skb_pcount(skb);
914 if (!tso_segs ||
915 (tso_segs > 1 &&
916 tcp_skb_mss(skb) != mss_now)) {
917 tcp_set_skb_tso_segs(sk, skb, mss_now);
918 tso_segs = tcp_skb_pcount(skb);
920 return tso_segs;
923 static inline int tcp_minshall_check(const struct tcp_sock *tp)
925 return after(tp->snd_sml,tp->snd_una) &&
926 !after(tp->snd_sml, tp->snd_nxt);
929 /* Return 0, if packet can be sent now without violation Nagle's rules:
930 * 1. It is full sized.
931 * 2. Or it contains FIN. (already checked by caller)
932 * 3. Or TCP_NODELAY was set.
933 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
934 * With Minshall's modification: all sent small packets are ACKed.
937 static inline int tcp_nagle_check(const struct tcp_sock *tp,
938 const struct sk_buff *skb,
939 unsigned mss_now, int nonagle)
941 return (skb->len < mss_now &&
942 ((nonagle&TCP_NAGLE_CORK) ||
943 (!nonagle &&
944 tp->packets_out &&
945 tcp_minshall_check(tp))));
948 /* Return non-zero if the Nagle test allows this packet to be
949 * sent now.
951 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
952 unsigned int cur_mss, int nonagle)
954 /* Nagle rule does not apply to frames, which sit in the middle of the
955 * write_queue (they have no chances to get new data).
957 * This is implemented in the callers, where they modify the 'nonagle'
958 * argument based upon the location of SKB in the send queue.
960 if (nonagle & TCP_NAGLE_PUSH)
961 return 1;
963 /* Don't use the nagle rule for urgent data (or for the final FIN). */
964 if (tp->urg_mode ||
965 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
966 return 1;
968 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
969 return 1;
971 return 0;
974 /* Does at least the first segment of SKB fit into the send window? */
975 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
977 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
979 if (skb->len > cur_mss)
980 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
982 return !after(end_seq, tp->snd_una + tp->snd_wnd);
985 /* This checks if the data bearing packet SKB (usually sk->sk_send_head)
986 * should be put on the wire right now. If so, it returns the number of
987 * packets allowed by the congestion window.
989 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
990 unsigned int cur_mss, int nonagle)
992 struct tcp_sock *tp = tcp_sk(sk);
993 unsigned int cwnd_quota;
995 tcp_init_tso_segs(sk, skb, cur_mss);
997 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
998 return 0;
1000 cwnd_quota = tcp_cwnd_test(tp, skb);
1001 if (cwnd_quota &&
1002 !tcp_snd_wnd_test(tp, skb, cur_mss))
1003 cwnd_quota = 0;
1005 return cwnd_quota;
1008 static inline int tcp_skb_is_last(const struct sock *sk,
1009 const struct sk_buff *skb)
1011 return skb->next == (struct sk_buff *)&sk->sk_write_queue;
1014 int tcp_may_send_now(struct sock *sk, struct tcp_sock *tp)
1016 struct sk_buff *skb = sk->sk_send_head;
1018 return (skb &&
1019 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1020 (tcp_skb_is_last(sk, skb) ?
1021 TCP_NAGLE_PUSH :
1022 tp->nonagle)));
1025 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1026 * which is put after SKB on the list. It is very much like
1027 * tcp_fragment() except that it may make several kinds of assumptions
1028 * in order to speed up the splitting operation. In particular, we
1029 * know that all the data is in scatter-gather pages, and that the
1030 * packet has never been sent out before (and thus is not cloned).
1032 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1034 struct sk_buff *buff;
1035 int nlen = skb->len - len;
1036 u16 flags;
1038 /* All of a TSO frame must be composed of paged data. */
1039 if (skb->len != skb->data_len)
1040 return tcp_fragment(sk, skb, len, mss_now);
1042 buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC);
1043 if (unlikely(buff == NULL))
1044 return -ENOMEM;
1046 sk_charge_skb(sk, buff);
1047 buff->truesize += nlen;
1048 skb->truesize -= nlen;
1050 /* Correct the sequence numbers. */
1051 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1052 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1053 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1055 /* PSH and FIN should only be set in the second packet. */
1056 flags = TCP_SKB_CB(skb)->flags;
1057 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1058 TCP_SKB_CB(buff)->flags = flags;
1060 /* This packet was never sent out yet, so no SACK bits. */
1061 TCP_SKB_CB(buff)->sacked = 0;
1063 buff->ip_summed = skb->ip_summed = CHECKSUM_HW;
1064 skb_split(skb, buff, len);
1066 /* Fix up tso_factor for both original and new SKB. */
1067 tcp_set_skb_tso_segs(sk, skb, mss_now);
1068 tcp_set_skb_tso_segs(sk, buff, mss_now);
1070 /* Link BUFF into the send queue. */
1071 skb_header_release(buff);
1072 __skb_append(skb, buff, &sk->sk_write_queue);
1074 return 0;
1077 /* Try to defer sending, if possible, in order to minimize the amount
1078 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1080 * This algorithm is from John Heffner.
1082 static int tcp_tso_should_defer(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb)
1084 const struct inet_connection_sock *icsk = inet_csk(sk);
1085 u32 send_win, cong_win, limit, in_flight;
1087 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1088 return 0;
1090 if (icsk->icsk_ca_state != TCP_CA_Open)
1091 return 0;
1093 in_flight = tcp_packets_in_flight(tp);
1095 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1096 (tp->snd_cwnd <= in_flight));
1098 send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq;
1100 /* From in_flight test above, we know that cwnd > in_flight. */
1101 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1103 limit = min(send_win, cong_win);
1105 /* If a full-sized TSO skb can be sent, do it. */
1106 if (limit >= 65536)
1107 return 0;
1109 if (sysctl_tcp_tso_win_divisor) {
1110 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1112 /* If at least some fraction of a window is available,
1113 * just use it.
1115 chunk /= sysctl_tcp_tso_win_divisor;
1116 if (limit >= chunk)
1117 return 0;
1118 } else {
1119 /* Different approach, try not to defer past a single
1120 * ACK. Receiver should ACK every other full sized
1121 * frame, so if we have space for more than 3 frames
1122 * then send now.
1124 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1125 return 0;
1128 /* Ok, it looks like it is advisable to defer. */
1129 return 1;
1132 /* Create a new MTU probe if we are ready.
1133 * Returns 0 if we should wait to probe (no cwnd available),
1134 * 1 if a probe was sent,
1135 * -1 otherwise */
1136 static int tcp_mtu_probe(struct sock *sk)
1138 struct tcp_sock *tp = tcp_sk(sk);
1139 struct inet_connection_sock *icsk = inet_csk(sk);
1140 struct sk_buff *skb, *nskb, *next;
1141 int len;
1142 int probe_size;
1143 unsigned int pif;
1144 int copy;
1145 int mss_now;
1147 /* Not currently probing/verifying,
1148 * not in recovery,
1149 * have enough cwnd, and
1150 * not SACKing (the variable headers throw things off) */
1151 if (!icsk->icsk_mtup.enabled ||
1152 icsk->icsk_mtup.probe_size ||
1153 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1154 tp->snd_cwnd < 11 ||
1155 tp->rx_opt.eff_sacks)
1156 return -1;
1158 /* Very simple search strategy: just double the MSS. */
1159 mss_now = tcp_current_mss(sk, 0);
1160 probe_size = 2*tp->mss_cache;
1161 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1162 /* TODO: set timer for probe_converge_event */
1163 return -1;
1166 /* Have enough data in the send queue to probe? */
1167 len = 0;
1168 if ((skb = sk->sk_send_head) == NULL)
1169 return -1;
1170 while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb))
1171 skb = skb->next;
1172 if (len < probe_size)
1173 return -1;
1175 /* Receive window check. */
1176 if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) {
1177 if (tp->snd_wnd < probe_size)
1178 return -1;
1179 else
1180 return 0;
1183 /* Do we need to wait to drain cwnd? */
1184 pif = tcp_packets_in_flight(tp);
1185 if (pif + 2 > tp->snd_cwnd) {
1186 /* With no packets in flight, don't stall. */
1187 if (pif == 0)
1188 return -1;
1189 else
1190 return 0;
1193 /* We're allowed to probe. Build it now. */
1194 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1195 return -1;
1196 sk_charge_skb(sk, nskb);
1198 skb = sk->sk_send_head;
1199 __skb_insert(nskb, skb->prev, skb, &sk->sk_write_queue);
1200 sk->sk_send_head = nskb;
1202 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1203 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1204 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1205 TCP_SKB_CB(nskb)->sacked = 0;
1206 nskb->csum = 0;
1207 if (skb->ip_summed == CHECKSUM_HW)
1208 nskb->ip_summed = CHECKSUM_HW;
1210 len = 0;
1211 while (len < probe_size) {
1212 next = skb->next;
1214 copy = min_t(int, skb->len, probe_size - len);
1215 if (nskb->ip_summed)
1216 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1217 else
1218 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1219 skb_put(nskb, copy), copy, nskb->csum);
1221 if (skb->len <= copy) {
1222 /* We've eaten all the data from this skb.
1223 * Throw it away. */
1224 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1225 __skb_unlink(skb, &sk->sk_write_queue);
1226 sk_stream_free_skb(sk, skb);
1227 } else {
1228 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1229 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1230 if (!skb_shinfo(skb)->nr_frags) {
1231 skb_pull(skb, copy);
1232 if (skb->ip_summed != CHECKSUM_HW)
1233 skb->csum = csum_partial(skb->data, skb->len, 0);
1234 } else {
1235 __pskb_trim_head(skb, copy);
1236 tcp_set_skb_tso_segs(sk, skb, mss_now);
1238 TCP_SKB_CB(skb)->seq += copy;
1241 len += copy;
1242 skb = next;
1244 tcp_init_tso_segs(sk, nskb, nskb->len);
1246 /* We're ready to send. If this fails, the probe will
1247 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1248 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1249 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1250 /* Decrement cwnd here because we are sending
1251 * effectively two packets. */
1252 tp->snd_cwnd--;
1253 update_send_head(sk, tp, nskb);
1255 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1256 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1257 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1259 return 1;
1262 return -1;
1266 /* This routine writes packets to the network. It advances the
1267 * send_head. This happens as incoming acks open up the remote
1268 * window for us.
1270 * Returns 1, if no segments are in flight and we have queued segments, but
1271 * cannot send anything now because of SWS or another problem.
1273 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1275 struct tcp_sock *tp = tcp_sk(sk);
1276 struct sk_buff *skb;
1277 unsigned int tso_segs, sent_pkts;
1278 int cwnd_quota;
1279 int result;
1281 /* If we are closed, the bytes will have to remain here.
1282 * In time closedown will finish, we empty the write queue and all
1283 * will be happy.
1285 if (unlikely(sk->sk_state == TCP_CLOSE))
1286 return 0;
1288 sent_pkts = 0;
1290 /* Do MTU probing. */
1291 if ((result = tcp_mtu_probe(sk)) == 0) {
1292 return 0;
1293 } else if (result > 0) {
1294 sent_pkts = 1;
1297 while ((skb = sk->sk_send_head)) {
1298 unsigned int limit;
1300 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1301 BUG_ON(!tso_segs);
1303 cwnd_quota = tcp_cwnd_test(tp, skb);
1304 if (!cwnd_quota)
1305 break;
1307 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1308 break;
1310 if (tso_segs == 1) {
1311 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1312 (tcp_skb_is_last(sk, skb) ?
1313 nonagle : TCP_NAGLE_PUSH))))
1314 break;
1315 } else {
1316 if (tcp_tso_should_defer(sk, tp, skb))
1317 break;
1320 limit = mss_now;
1321 if (tso_segs > 1) {
1322 limit = tcp_window_allows(tp, skb,
1323 mss_now, cwnd_quota);
1325 if (skb->len < limit) {
1326 unsigned int trim = skb->len % mss_now;
1328 if (trim)
1329 limit = skb->len - trim;
1333 if (skb->len > limit &&
1334 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1335 break;
1337 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1339 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1340 break;
1342 /* Advance the send_head. This one is sent out.
1343 * This call will increment packets_out.
1345 update_send_head(sk, tp, skb);
1347 tcp_minshall_update(tp, mss_now, skb);
1348 sent_pkts++;
1351 if (likely(sent_pkts)) {
1352 tcp_cwnd_validate(sk, tp);
1353 return 0;
1355 return !tp->packets_out && sk->sk_send_head;
1358 /* Push out any pending frames which were held back due to
1359 * TCP_CORK or attempt at coalescing tiny packets.
1360 * The socket must be locked by the caller.
1362 void __tcp_push_pending_frames(struct sock *sk, struct tcp_sock *tp,
1363 unsigned int cur_mss, int nonagle)
1365 struct sk_buff *skb = sk->sk_send_head;
1367 if (skb) {
1368 if (tcp_write_xmit(sk, cur_mss, nonagle))
1369 tcp_check_probe_timer(sk, tp);
1373 /* Send _single_ skb sitting at the send head. This function requires
1374 * true push pending frames to setup probe timer etc.
1376 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1378 struct tcp_sock *tp = tcp_sk(sk);
1379 struct sk_buff *skb = sk->sk_send_head;
1380 unsigned int tso_segs, cwnd_quota;
1382 BUG_ON(!skb || skb->len < mss_now);
1384 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1385 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1387 if (likely(cwnd_quota)) {
1388 unsigned int limit;
1390 BUG_ON(!tso_segs);
1392 limit = mss_now;
1393 if (tso_segs > 1) {
1394 limit = tcp_window_allows(tp, skb,
1395 mss_now, cwnd_quota);
1397 if (skb->len < limit) {
1398 unsigned int trim = skb->len % mss_now;
1400 if (trim)
1401 limit = skb->len - trim;
1405 if (skb->len > limit &&
1406 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1407 return;
1409 /* Send it out now. */
1410 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1412 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1413 update_send_head(sk, tp, skb);
1414 tcp_cwnd_validate(sk, tp);
1415 return;
1420 /* This function returns the amount that we can raise the
1421 * usable window based on the following constraints
1423 * 1. The window can never be shrunk once it is offered (RFC 793)
1424 * 2. We limit memory per socket
1426 * RFC 1122:
1427 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1428 * RECV.NEXT + RCV.WIN fixed until:
1429 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1431 * i.e. don't raise the right edge of the window until you can raise
1432 * it at least MSS bytes.
1434 * Unfortunately, the recommended algorithm breaks header prediction,
1435 * since header prediction assumes th->window stays fixed.
1437 * Strictly speaking, keeping th->window fixed violates the receiver
1438 * side SWS prevention criteria. The problem is that under this rule
1439 * a stream of single byte packets will cause the right side of the
1440 * window to always advance by a single byte.
1442 * Of course, if the sender implements sender side SWS prevention
1443 * then this will not be a problem.
1445 * BSD seems to make the following compromise:
1447 * If the free space is less than the 1/4 of the maximum
1448 * space available and the free space is less than 1/2 mss,
1449 * then set the window to 0.
1450 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1451 * Otherwise, just prevent the window from shrinking
1452 * and from being larger than the largest representable value.
1454 * This prevents incremental opening of the window in the regime
1455 * where TCP is limited by the speed of the reader side taking
1456 * data out of the TCP receive queue. It does nothing about
1457 * those cases where the window is constrained on the sender side
1458 * because the pipeline is full.
1460 * BSD also seems to "accidentally" limit itself to windows that are a
1461 * multiple of MSS, at least until the free space gets quite small.
1462 * This would appear to be a side effect of the mbuf implementation.
1463 * Combining these two algorithms results in the observed behavior
1464 * of having a fixed window size at almost all times.
1466 * Below we obtain similar behavior by forcing the offered window to
1467 * a multiple of the mss when it is feasible to do so.
1469 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1470 * Regular options like TIMESTAMP are taken into account.
1472 u32 __tcp_select_window(struct sock *sk)
1474 struct inet_connection_sock *icsk = inet_csk(sk);
1475 struct tcp_sock *tp = tcp_sk(sk);
1476 /* MSS for the peer's data. Previous versions used mss_clamp
1477 * here. I don't know if the value based on our guesses
1478 * of peer's MSS is better for the performance. It's more correct
1479 * but may be worse for the performance because of rcv_mss
1480 * fluctuations. --SAW 1998/11/1
1482 int mss = icsk->icsk_ack.rcv_mss;
1483 int free_space = tcp_space(sk);
1484 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1485 int window;
1487 if (mss > full_space)
1488 mss = full_space;
1490 if (free_space < full_space/2) {
1491 icsk->icsk_ack.quick = 0;
1493 if (tcp_memory_pressure)
1494 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1496 if (free_space < mss)
1497 return 0;
1500 if (free_space > tp->rcv_ssthresh)
1501 free_space = tp->rcv_ssthresh;
1503 /* Don't do rounding if we are using window scaling, since the
1504 * scaled window will not line up with the MSS boundary anyway.
1506 window = tp->rcv_wnd;
1507 if (tp->rx_opt.rcv_wscale) {
1508 window = free_space;
1510 /* Advertise enough space so that it won't get scaled away.
1511 * Import case: prevent zero window announcement if
1512 * 1<<rcv_wscale > mss.
1514 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1515 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1516 << tp->rx_opt.rcv_wscale);
1517 } else {
1518 /* Get the largest window that is a nice multiple of mss.
1519 * Window clamp already applied above.
1520 * If our current window offering is within 1 mss of the
1521 * free space we just keep it. This prevents the divide
1522 * and multiply from happening most of the time.
1523 * We also don't do any window rounding when the free space
1524 * is too small.
1526 if (window <= free_space - mss || window > free_space)
1527 window = (free_space/mss)*mss;
1530 return window;
1533 /* Attempt to collapse two adjacent SKB's during retransmission. */
1534 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1536 struct tcp_sock *tp = tcp_sk(sk);
1537 struct sk_buff *next_skb = skb->next;
1539 /* The first test we must make is that neither of these two
1540 * SKB's are still referenced by someone else.
1542 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1543 int skb_size = skb->len, next_skb_size = next_skb->len;
1544 u16 flags = TCP_SKB_CB(skb)->flags;
1546 /* Also punt if next skb has been SACK'd. */
1547 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1548 return;
1550 /* Next skb is out of window. */
1551 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd))
1552 return;
1554 /* Punt if not enough space exists in the first SKB for
1555 * the data in the second, or the total combined payload
1556 * would exceed the MSS.
1558 if ((next_skb_size > skb_tailroom(skb)) ||
1559 ((skb_size + next_skb_size) > mss_now))
1560 return;
1562 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1563 tcp_skb_pcount(next_skb) != 1);
1565 /* changing transmit queue under us so clear hints */
1566 clear_all_retrans_hints(tp);
1568 /* Ok. We will be able to collapse the packet. */
1569 __skb_unlink(next_skb, &sk->sk_write_queue);
1571 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size);
1573 if (next_skb->ip_summed == CHECKSUM_HW)
1574 skb->ip_summed = CHECKSUM_HW;
1576 if (skb->ip_summed != CHECKSUM_HW)
1577 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1579 /* Update sequence range on original skb. */
1580 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1582 /* Merge over control information. */
1583 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1584 TCP_SKB_CB(skb)->flags = flags;
1586 /* All done, get rid of second SKB and account for it so
1587 * packet counting does not break.
1589 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1590 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1591 tp->retrans_out -= tcp_skb_pcount(next_skb);
1592 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) {
1593 tp->lost_out -= tcp_skb_pcount(next_skb);
1594 tp->left_out -= tcp_skb_pcount(next_skb);
1596 /* Reno case is special. Sigh... */
1597 if (!tp->rx_opt.sack_ok && tp->sacked_out) {
1598 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1599 tp->left_out -= tcp_skb_pcount(next_skb);
1602 /* Not quite right: it can be > snd.fack, but
1603 * it is better to underestimate fackets.
1605 tcp_dec_pcount_approx(&tp->fackets_out, next_skb);
1606 tcp_packets_out_dec(tp, next_skb);
1607 sk_stream_free_skb(sk, next_skb);
1611 /* Do a simple retransmit without using the backoff mechanisms in
1612 * tcp_timer. This is used for path mtu discovery.
1613 * The socket is already locked here.
1615 void tcp_simple_retransmit(struct sock *sk)
1617 const struct inet_connection_sock *icsk = inet_csk(sk);
1618 struct tcp_sock *tp = tcp_sk(sk);
1619 struct sk_buff *skb;
1620 unsigned int mss = tcp_current_mss(sk, 0);
1621 int lost = 0;
1623 sk_stream_for_retrans_queue(skb, sk) {
1624 if (skb->len > mss &&
1625 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1626 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1627 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1628 tp->retrans_out -= tcp_skb_pcount(skb);
1630 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1631 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1632 tp->lost_out += tcp_skb_pcount(skb);
1633 lost = 1;
1638 clear_all_retrans_hints(tp);
1640 if (!lost)
1641 return;
1643 tcp_sync_left_out(tp);
1645 /* Don't muck with the congestion window here.
1646 * Reason is that we do not increase amount of _data_
1647 * in network, but units changed and effective
1648 * cwnd/ssthresh really reduced now.
1650 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1651 tp->high_seq = tp->snd_nxt;
1652 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1653 tp->prior_ssthresh = 0;
1654 tp->undo_marker = 0;
1655 tcp_set_ca_state(sk, TCP_CA_Loss);
1657 tcp_xmit_retransmit_queue(sk);
1660 /* This retransmits one SKB. Policy decisions and retransmit queue
1661 * state updates are done by the caller. Returns non-zero if an
1662 * error occurred which prevented the send.
1664 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1666 struct tcp_sock *tp = tcp_sk(sk);
1667 struct inet_connection_sock *icsk = inet_csk(sk);
1668 unsigned int cur_mss = tcp_current_mss(sk, 0);
1669 int err;
1671 /* Inconslusive MTU probe */
1672 if (icsk->icsk_mtup.probe_size) {
1673 icsk->icsk_mtup.probe_size = 0;
1676 /* Do not sent more than we queued. 1/4 is reserved for possible
1677 * copying overhead: fragmentation, tunneling, mangling etc.
1679 if (atomic_read(&sk->sk_wmem_alloc) >
1680 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1681 return -EAGAIN;
1683 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1684 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1685 BUG();
1686 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1687 return -ENOMEM;
1690 /* If receiver has shrunk his window, and skb is out of
1691 * new window, do not retransmit it. The exception is the
1692 * case, when window is shrunk to zero. In this case
1693 * our retransmit serves as a zero window probe.
1695 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)
1696 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1697 return -EAGAIN;
1699 if (skb->len > cur_mss) {
1700 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1701 return -ENOMEM; /* We'll try again later. */
1704 /* Collapse two adjacent packets if worthwhile and we can. */
1705 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1706 (skb->len < (cur_mss >> 1)) &&
1707 (skb->next != sk->sk_send_head) &&
1708 (skb->next != (struct sk_buff *)&sk->sk_write_queue) &&
1709 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) &&
1710 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) &&
1711 (sysctl_tcp_retrans_collapse != 0))
1712 tcp_retrans_try_collapse(sk, skb, cur_mss);
1714 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1715 return -EHOSTUNREACH; /* Routing failure or similar. */
1717 /* Some Solaris stacks overoptimize and ignore the FIN on a
1718 * retransmit when old data is attached. So strip it off
1719 * since it is cheap to do so and saves bytes on the network.
1721 if(skb->len > 0 &&
1722 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1723 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1724 if (!pskb_trim(skb, 0)) {
1725 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1726 skb_shinfo(skb)->gso_segs = 1;
1727 skb_shinfo(skb)->gso_size = 0;
1728 skb_shinfo(skb)->gso_type = 0;
1729 skb->ip_summed = CHECKSUM_NONE;
1730 skb->csum = 0;
1734 /* Make a copy, if the first transmission SKB clone we made
1735 * is still in somebody's hands, else make a clone.
1737 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1739 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1741 if (err == 0) {
1742 /* Update global TCP statistics. */
1743 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1745 tp->total_retrans++;
1747 #if FASTRETRANS_DEBUG > 0
1748 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1749 if (net_ratelimit())
1750 printk(KERN_DEBUG "retrans_out leaked.\n");
1752 #endif
1753 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1754 tp->retrans_out += tcp_skb_pcount(skb);
1756 /* Save stamp of the first retransmit. */
1757 if (!tp->retrans_stamp)
1758 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1760 tp->undo_retrans++;
1762 /* snd_nxt is stored to detect loss of retransmitted segment,
1763 * see tcp_input.c tcp_sacktag_write_queue().
1765 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1767 return err;
1770 /* This gets called after a retransmit timeout, and the initially
1771 * retransmitted data is acknowledged. It tries to continue
1772 * resending the rest of the retransmit queue, until either
1773 * we've sent it all or the congestion window limit is reached.
1774 * If doing SACK, the first ACK which comes back for a timeout
1775 * based retransmit packet might feed us FACK information again.
1776 * If so, we use it to avoid unnecessarily retransmissions.
1778 void tcp_xmit_retransmit_queue(struct sock *sk)
1780 const struct inet_connection_sock *icsk = inet_csk(sk);
1781 struct tcp_sock *tp = tcp_sk(sk);
1782 struct sk_buff *skb;
1783 int packet_cnt;
1785 if (tp->retransmit_skb_hint) {
1786 skb = tp->retransmit_skb_hint;
1787 packet_cnt = tp->retransmit_cnt_hint;
1788 }else{
1789 skb = sk->sk_write_queue.next;
1790 packet_cnt = 0;
1793 /* First pass: retransmit lost packets. */
1794 if (tp->lost_out) {
1795 sk_stream_for_retrans_queue_from(skb, sk) {
1796 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1798 /* we could do better than to assign each time */
1799 tp->retransmit_skb_hint = skb;
1800 tp->retransmit_cnt_hint = packet_cnt;
1802 /* Assume this retransmit will generate
1803 * only one packet for congestion window
1804 * calculation purposes. This works because
1805 * tcp_retransmit_skb() will chop up the
1806 * packet to be MSS sized and all the
1807 * packet counting works out.
1809 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1810 return;
1812 if (sacked & TCPCB_LOST) {
1813 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1814 if (tcp_retransmit_skb(sk, skb)) {
1815 tp->retransmit_skb_hint = NULL;
1816 return;
1818 if (icsk->icsk_ca_state != TCP_CA_Loss)
1819 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1820 else
1821 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1823 if (skb ==
1824 skb_peek(&sk->sk_write_queue))
1825 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1826 inet_csk(sk)->icsk_rto,
1827 TCP_RTO_MAX);
1830 packet_cnt += tcp_skb_pcount(skb);
1831 if (packet_cnt >= tp->lost_out)
1832 break;
1837 /* OK, demanded retransmission is finished. */
1839 /* Forward retransmissions are possible only during Recovery. */
1840 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1841 return;
1843 /* No forward retransmissions in Reno are possible. */
1844 if (!tp->rx_opt.sack_ok)
1845 return;
1847 /* Yeah, we have to make difficult choice between forward transmission
1848 * and retransmission... Both ways have their merits...
1850 * For now we do not retransmit anything, while we have some new
1851 * segments to send.
1854 if (tcp_may_send_now(sk, tp))
1855 return;
1857 if (tp->forward_skb_hint) {
1858 skb = tp->forward_skb_hint;
1859 packet_cnt = tp->forward_cnt_hint;
1860 } else{
1861 skb = sk->sk_write_queue.next;
1862 packet_cnt = 0;
1865 sk_stream_for_retrans_queue_from(skb, sk) {
1866 tp->forward_cnt_hint = packet_cnt;
1867 tp->forward_skb_hint = skb;
1869 /* Similar to the retransmit loop above we
1870 * can pretend that the retransmitted SKB
1871 * we send out here will be composed of one
1872 * real MSS sized packet because tcp_retransmit_skb()
1873 * will fragment it if necessary.
1875 if (++packet_cnt > tp->fackets_out)
1876 break;
1878 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1879 break;
1881 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
1882 continue;
1884 /* Ok, retransmit it. */
1885 if (tcp_retransmit_skb(sk, skb)) {
1886 tp->forward_skb_hint = NULL;
1887 break;
1890 if (skb == skb_peek(&sk->sk_write_queue))
1891 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1892 inet_csk(sk)->icsk_rto,
1893 TCP_RTO_MAX);
1895 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
1900 /* Send a fin. The caller locks the socket for us. This cannot be
1901 * allowed to fail queueing a FIN frame under any circumstances.
1903 void tcp_send_fin(struct sock *sk)
1905 struct tcp_sock *tp = tcp_sk(sk);
1906 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue);
1907 int mss_now;
1909 /* Optimization, tack on the FIN if we have a queue of
1910 * unsent frames. But be careful about outgoing SACKS
1911 * and IP options.
1913 mss_now = tcp_current_mss(sk, 1);
1915 if (sk->sk_send_head != NULL) {
1916 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
1917 TCP_SKB_CB(skb)->end_seq++;
1918 tp->write_seq++;
1919 } else {
1920 /* Socket is locked, keep trying until memory is available. */
1921 for (;;) {
1922 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
1923 if (skb)
1924 break;
1925 yield();
1928 /* Reserve space for headers and prepare control bits. */
1929 skb_reserve(skb, MAX_TCP_HEADER);
1930 skb->csum = 0;
1931 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
1932 TCP_SKB_CB(skb)->sacked = 0;
1933 skb_shinfo(skb)->gso_segs = 1;
1934 skb_shinfo(skb)->gso_size = 0;
1935 skb_shinfo(skb)->gso_type = 0;
1937 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
1938 TCP_SKB_CB(skb)->seq = tp->write_seq;
1939 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
1940 tcp_queue_skb(sk, skb);
1942 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF);
1945 /* We get here when a process closes a file descriptor (either due to
1946 * an explicit close() or as a byproduct of exit()'ing) and there
1947 * was unread data in the receive queue. This behavior is recommended
1948 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM
1950 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1952 struct tcp_sock *tp = tcp_sk(sk);
1953 struct sk_buff *skb;
1955 /* NOTE: No TCP options attached and we never retransmit this. */
1956 skb = alloc_skb(MAX_TCP_HEADER, priority);
1957 if (!skb) {
1958 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1959 return;
1962 /* Reserve space for headers and prepare control bits. */
1963 skb_reserve(skb, MAX_TCP_HEADER);
1964 skb->csum = 0;
1965 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
1966 TCP_SKB_CB(skb)->sacked = 0;
1967 skb_shinfo(skb)->gso_segs = 1;
1968 skb_shinfo(skb)->gso_size = 0;
1969 skb_shinfo(skb)->gso_type = 0;
1971 /* Send it off. */
1972 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp);
1973 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
1974 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1975 if (tcp_transmit_skb(sk, skb, 0, priority))
1976 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
1979 /* WARNING: This routine must only be called when we have already sent
1980 * a SYN packet that crossed the incoming SYN that caused this routine
1981 * to get called. If this assumption fails then the initial rcv_wnd
1982 * and rcv_wscale values will not be correct.
1984 int tcp_send_synack(struct sock *sk)
1986 struct sk_buff* skb;
1988 skb = skb_peek(&sk->sk_write_queue);
1989 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
1990 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
1991 return -EFAULT;
1993 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
1994 if (skb_cloned(skb)) {
1995 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
1996 if (nskb == NULL)
1997 return -ENOMEM;
1998 __skb_unlink(skb, &sk->sk_write_queue);
1999 skb_header_release(nskb);
2000 __skb_queue_head(&sk->sk_write_queue, nskb);
2001 sk_stream_free_skb(sk, skb);
2002 sk_charge_skb(sk, nskb);
2003 skb = nskb;
2006 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2007 TCP_ECN_send_synack(tcp_sk(sk), skb);
2009 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2010 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2014 * Prepare a SYN-ACK.
2016 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2017 struct request_sock *req)
2019 struct inet_request_sock *ireq = inet_rsk(req);
2020 struct tcp_sock *tp = tcp_sk(sk);
2021 struct tcphdr *th;
2022 int tcp_header_size;
2023 struct sk_buff *skb;
2025 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2026 if (skb == NULL)
2027 return NULL;
2029 /* Reserve space for headers. */
2030 skb_reserve(skb, MAX_TCP_HEADER);
2032 skb->dst = dst_clone(dst);
2034 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2035 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2036 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2037 /* SACK_PERM is in the place of NOP NOP of TS */
2038 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2039 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size);
2041 memset(th, 0, sizeof(struct tcphdr));
2042 th->syn = 1;
2043 th->ack = 1;
2044 TCP_ECN_make_synack(req, th);
2045 th->source = inet_sk(sk)->sport;
2046 th->dest = ireq->rmt_port;
2047 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2048 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2049 TCP_SKB_CB(skb)->sacked = 0;
2050 skb_shinfo(skb)->gso_segs = 1;
2051 skb_shinfo(skb)->gso_size = 0;
2052 skb_shinfo(skb)->gso_type = 0;
2053 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2054 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2055 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2056 __u8 rcv_wscale;
2057 /* Set this up on the first call only */
2058 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2059 /* tcp_full_space because it is guaranteed to be the first packet */
2060 tcp_select_initial_window(tcp_full_space(sk),
2061 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2062 &req->rcv_wnd,
2063 &req->window_clamp,
2064 ireq->wscale_ok,
2065 &rcv_wscale);
2066 ireq->rcv_wscale = rcv_wscale;
2069 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2070 th->window = htons(req->rcv_wnd);
2072 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2073 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2074 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2075 TCP_SKB_CB(skb)->when,
2076 req->ts_recent);
2078 skb->csum = 0;
2079 th->doff = (tcp_header_size >> 2);
2080 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2081 return skb;
2085 * Do all connect socket setups that can be done AF independent.
2087 static void tcp_connect_init(struct sock *sk)
2089 struct dst_entry *dst = __sk_dst_get(sk);
2090 struct tcp_sock *tp = tcp_sk(sk);
2091 __u8 rcv_wscale;
2093 /* We'll fix this up when we get a response from the other end.
2094 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2096 tp->tcp_header_len = sizeof(struct tcphdr) +
2097 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2099 /* If user gave his TCP_MAXSEG, record it to clamp */
2100 if (tp->rx_opt.user_mss)
2101 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2102 tp->max_window = 0;
2103 tcp_mtup_init(sk);
2104 tcp_sync_mss(sk, dst_mtu(dst));
2106 if (!tp->window_clamp)
2107 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2108 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2109 tcp_initialize_rcv_mss(sk);
2111 tcp_select_initial_window(tcp_full_space(sk),
2112 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2113 &tp->rcv_wnd,
2114 &tp->window_clamp,
2115 sysctl_tcp_window_scaling,
2116 &rcv_wscale);
2118 tp->rx_opt.rcv_wscale = rcv_wscale;
2119 tp->rcv_ssthresh = tp->rcv_wnd;
2121 sk->sk_err = 0;
2122 sock_reset_flag(sk, SOCK_DONE);
2123 tp->snd_wnd = 0;
2124 tcp_init_wl(tp, tp->write_seq, 0);
2125 tp->snd_una = tp->write_seq;
2126 tp->snd_sml = tp->write_seq;
2127 tp->rcv_nxt = 0;
2128 tp->rcv_wup = 0;
2129 tp->copied_seq = 0;
2131 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2132 inet_csk(sk)->icsk_retransmits = 0;
2133 tcp_clear_retrans(tp);
2137 * Build a SYN and send it off.
2139 int tcp_connect(struct sock *sk)
2141 struct tcp_sock *tp = tcp_sk(sk);
2142 struct sk_buff *buff;
2144 tcp_connect_init(sk);
2146 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2147 if (unlikely(buff == NULL))
2148 return -ENOBUFS;
2150 /* Reserve space for headers. */
2151 skb_reserve(buff, MAX_TCP_HEADER);
2153 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2154 TCP_ECN_send_syn(sk, tp, buff);
2155 TCP_SKB_CB(buff)->sacked = 0;
2156 skb_shinfo(buff)->gso_segs = 1;
2157 skb_shinfo(buff)->gso_size = 0;
2158 skb_shinfo(buff)->gso_type = 0;
2159 buff->csum = 0;
2160 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2161 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2162 tp->snd_nxt = tp->write_seq;
2163 tp->pushed_seq = tp->write_seq;
2165 /* Send it off. */
2166 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2167 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2168 skb_header_release(buff);
2169 __skb_queue_tail(&sk->sk_write_queue, buff);
2170 sk_charge_skb(sk, buff);
2171 tp->packets_out += tcp_skb_pcount(buff);
2172 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2173 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2175 /* Timer for repeating the SYN until an answer. */
2176 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2177 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2178 return 0;
2181 /* Send out a delayed ack, the caller does the policy checking
2182 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2183 * for details.
2185 void tcp_send_delayed_ack(struct sock *sk)
2187 struct inet_connection_sock *icsk = inet_csk(sk);
2188 int ato = icsk->icsk_ack.ato;
2189 unsigned long timeout;
2191 if (ato > TCP_DELACK_MIN) {
2192 const struct tcp_sock *tp = tcp_sk(sk);
2193 int max_ato = HZ/2;
2195 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2196 max_ato = TCP_DELACK_MAX;
2198 /* Slow path, intersegment interval is "high". */
2200 /* If some rtt estimate is known, use it to bound delayed ack.
2201 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2202 * directly.
2204 if (tp->srtt) {
2205 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2207 if (rtt < max_ato)
2208 max_ato = rtt;
2211 ato = min(ato, max_ato);
2214 /* Stay within the limit we were given */
2215 timeout = jiffies + ato;
2217 /* Use new timeout only if there wasn't a older one earlier. */
2218 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2219 /* If delack timer was blocked or is about to expire,
2220 * send ACK now.
2222 if (icsk->icsk_ack.blocked ||
2223 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2224 tcp_send_ack(sk);
2225 return;
2228 if (!time_before(timeout, icsk->icsk_ack.timeout))
2229 timeout = icsk->icsk_ack.timeout;
2231 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2232 icsk->icsk_ack.timeout = timeout;
2233 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2236 /* This routine sends an ack and also updates the window. */
2237 void tcp_send_ack(struct sock *sk)
2239 /* If we have been reset, we may not send again. */
2240 if (sk->sk_state != TCP_CLOSE) {
2241 struct tcp_sock *tp = tcp_sk(sk);
2242 struct sk_buff *buff;
2244 /* We are not putting this on the write queue, so
2245 * tcp_transmit_skb() will set the ownership to this
2246 * sock.
2248 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2249 if (buff == NULL) {
2250 inet_csk_schedule_ack(sk);
2251 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2252 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2253 TCP_DELACK_MAX, TCP_RTO_MAX);
2254 return;
2257 /* Reserve space for headers and prepare control bits. */
2258 skb_reserve(buff, MAX_TCP_HEADER);
2259 buff->csum = 0;
2260 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2261 TCP_SKB_CB(buff)->sacked = 0;
2262 skb_shinfo(buff)->gso_segs = 1;
2263 skb_shinfo(buff)->gso_size = 0;
2264 skb_shinfo(buff)->gso_type = 0;
2266 /* Send it off, this clears delayed acks for us. */
2267 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp);
2268 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2269 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2273 /* This routine sends a packet with an out of date sequence
2274 * number. It assumes the other end will try to ack it.
2276 * Question: what should we make while urgent mode?
2277 * 4.4BSD forces sending single byte of data. We cannot send
2278 * out of window data, because we have SND.NXT==SND.MAX...
2280 * Current solution: to send TWO zero-length segments in urgent mode:
2281 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2282 * out-of-date with SND.UNA-1 to probe window.
2284 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2286 struct tcp_sock *tp = tcp_sk(sk);
2287 struct sk_buff *skb;
2289 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2290 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2291 if (skb == NULL)
2292 return -1;
2294 /* Reserve space for headers and set control bits. */
2295 skb_reserve(skb, MAX_TCP_HEADER);
2296 skb->csum = 0;
2297 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2298 TCP_SKB_CB(skb)->sacked = urgent;
2299 skb_shinfo(skb)->gso_segs = 1;
2300 skb_shinfo(skb)->gso_size = 0;
2301 skb_shinfo(skb)->gso_type = 0;
2303 /* Use a previous sequence. This should cause the other
2304 * end to send an ack. Don't queue or clone SKB, just
2305 * send it.
2307 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2308 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2309 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2310 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2313 int tcp_write_wakeup(struct sock *sk)
2315 if (sk->sk_state != TCP_CLOSE) {
2316 struct tcp_sock *tp = tcp_sk(sk);
2317 struct sk_buff *skb;
2319 if ((skb = sk->sk_send_head) != NULL &&
2320 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) {
2321 int err;
2322 unsigned int mss = tcp_current_mss(sk, 0);
2323 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq;
2325 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2326 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2328 /* We are probing the opening of a window
2329 * but the window size is != 0
2330 * must have been a result SWS avoidance ( sender )
2332 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2333 skb->len > mss) {
2334 seg_size = min(seg_size, mss);
2335 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2336 if (tcp_fragment(sk, skb, seg_size, mss))
2337 return -1;
2338 } else if (!tcp_skb_pcount(skb))
2339 tcp_set_skb_tso_segs(sk, skb, mss);
2341 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2342 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2343 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2344 if (!err) {
2345 update_send_head(sk, tp, skb);
2347 return err;
2348 } else {
2349 if (tp->urg_mode &&
2350 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2351 tcp_xmit_probe_skb(sk, TCPCB_URG);
2352 return tcp_xmit_probe_skb(sk, 0);
2355 return -1;
2358 /* A window probe timeout has occurred. If window is not closed send
2359 * a partial packet else a zero probe.
2361 void tcp_send_probe0(struct sock *sk)
2363 struct inet_connection_sock *icsk = inet_csk(sk);
2364 struct tcp_sock *tp = tcp_sk(sk);
2365 int err;
2367 err = tcp_write_wakeup(sk);
2369 if (tp->packets_out || !sk->sk_send_head) {
2370 /* Cancel probe timer, if it is not required. */
2371 icsk->icsk_probes_out = 0;
2372 icsk->icsk_backoff = 0;
2373 return;
2376 if (err <= 0) {
2377 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2378 icsk->icsk_backoff++;
2379 icsk->icsk_probes_out++;
2380 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2381 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2382 TCP_RTO_MAX);
2383 } else {
2384 /* If packet was not sent due to local congestion,
2385 * do not backoff and do not remember icsk_probes_out.
2386 * Let local senders to fight for local resources.
2388 * Use accumulated backoff yet.
2390 if (!icsk->icsk_probes_out)
2391 icsk->icsk_probes_out = 1;
2392 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2393 min(icsk->icsk_rto << icsk->icsk_backoff,
2394 TCP_RESOURCE_PROBE_INTERVAL),
2395 TCP_RTO_MAX);
2399 EXPORT_SYMBOL(tcp_connect);
2400 EXPORT_SYMBOL(tcp_make_synack);
2401 EXPORT_SYMBOL(tcp_simple_retransmit);
2402 EXPORT_SYMBOL(tcp_sync_mss);
2403 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2404 EXPORT_SYMBOL(tcp_mtup_init);