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>
44 /* People can turn this off for buggy TCP's found in printers etc. */
45 int sysctl_tcp_retrans_collapse __read_mostly = 1;
47 /* People can turn this on to work with those rare, broken TCPs that
48 * interpret the window field as a signed quantity.
50 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52 /* This limits the percentage of the congestion window which we
53 * will allow a single TSO frame to consume. Building TSO frames
54 * which are too large can cause TCP streams to be bursty.
56 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
58 int sysctl_tcp_mtu_probing __read_mostly = 0;
59 int sysctl_tcp_base_mss __read_mostly = 512;
61 /* By default, RFC2861 behavior. */
62 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66 struct tcp_sock *tp = tcp_sk(sk);
67 unsigned int prior_packets = tp->packets_out;
69 tcp_advance_send_head(sk, skb);
70 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
72 /* Don't override Nagle indefinately with F-RTO */
73 if (tp->frto_counter == 2)
76 tp->packets_out += tcp_skb_pcount(skb);
78 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
79 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
82 /* SND.NXT, if window was not shrunk.
83 * If window has been shrunk, what should we make? It is not clear at all.
84 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
85 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
86 * invalid. OK, let's make this for now:
88 static inline __u32 tcp_acceptable_seq(struct sock *sk)
90 struct tcp_sock *tp = tcp_sk(sk);
92 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
95 return tcp_wnd_end(tp);
98 /* Calculate mss to advertise in SYN segment.
99 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
101 * 1. It is independent of path mtu.
102 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
103 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
104 * attached devices, because some buggy hosts are confused by
106 * 4. We do not make 3, we advertise MSS, calculated from first
107 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
108 * This may be overridden via information stored in routing table.
109 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
110 * probably even Jumbo".
112 static __u16 tcp_advertise_mss(struct sock *sk)
114 struct tcp_sock *tp = tcp_sk(sk);
115 struct dst_entry *dst = __sk_dst_get(sk);
116 int mss = tp->advmss;
118 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
119 mss = dst_metric(dst, RTAX_ADVMSS);
126 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
127 * This is the first part of cwnd validation mechanism. */
128 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
130 struct tcp_sock *tp = tcp_sk(sk);
131 s32 delta = tcp_time_stamp - tp->lsndtime;
132 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
133 u32 cwnd = tp->snd_cwnd;
135 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
137 tp->snd_ssthresh = tcp_current_ssthresh(sk);
138 restart_cwnd = min(restart_cwnd, cwnd);
140 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
142 tp->snd_cwnd = max(cwnd, restart_cwnd);
143 tp->snd_cwnd_stamp = tcp_time_stamp;
144 tp->snd_cwnd_used = 0;
147 static void tcp_event_data_sent(struct tcp_sock *tp,
148 struct sk_buff *skb, struct sock *sk)
150 struct inet_connection_sock *icsk = inet_csk(sk);
151 const u32 now = tcp_time_stamp;
153 if (sysctl_tcp_slow_start_after_idle &&
154 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
155 tcp_cwnd_restart(sk, __sk_dst_get(sk));
159 /* If it is a reply for ato after last received
160 * packet, enter pingpong mode.
162 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
163 icsk->icsk_ack.pingpong = 1;
166 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
168 tcp_dec_quickack_mode(sk, pkts);
169 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
172 /* Determine a window scaling and initial window to offer.
173 * Based on the assumption that the given amount of space
174 * will be offered. Store the results in the tp structure.
175 * NOTE: for smooth operation initial space offering should
176 * be a multiple of mss if possible. We assume here that mss >= 1.
177 * This MUST be enforced by all callers.
179 void tcp_select_initial_window(int __space, __u32 mss,
180 __u32 *rcv_wnd, __u32 *window_clamp,
181 int wscale_ok, __u8 *rcv_wscale)
183 unsigned int space = (__space < 0 ? 0 : __space);
185 /* If no clamp set the clamp to the max possible scaled window */
186 if (*window_clamp == 0)
187 (*window_clamp) = (65535 << 14);
188 space = min(*window_clamp, space);
190 /* Quantize space offering to a multiple of mss if possible. */
192 space = (space / mss) * mss;
194 /* NOTE: offering an initial window larger than 32767
195 * will break some buggy TCP stacks. If the admin tells us
196 * it is likely we could be speaking with such a buggy stack
197 * we will truncate our initial window offering to 32K-1
198 * unless the remote has sent us a window scaling option,
199 * which we interpret as a sign the remote TCP is not
200 * misinterpreting the window field as a signed quantity.
202 if (sysctl_tcp_workaround_signed_windows)
203 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
209 /* Set window scaling on max possible window
210 * See RFC1323 for an explanation of the limit to 14
212 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
213 space = min_t(u32, space, *window_clamp);
214 while (space > 65535 && (*rcv_wscale) < 14) {
220 /* Set initial window to value enough for senders,
221 * following RFC2414. Senders, not following this RFC,
222 * will be satisfied with 2.
224 if (mss > (1 << *rcv_wscale)) {
230 if (*rcv_wnd > init_cwnd * mss)
231 *rcv_wnd = init_cwnd * mss;
234 /* Set the clamp no higher than max representable value */
235 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
238 /* Chose a new window to advertise, update state in tcp_sock for the
239 * socket, and return result with RFC1323 scaling applied. The return
240 * value can be stuffed directly into th->window for an outgoing
243 static u16 tcp_select_window(struct sock *sk)
245 struct tcp_sock *tp = tcp_sk(sk);
246 u32 cur_win = tcp_receive_window(tp);
247 u32 new_win = __tcp_select_window(sk);
249 /* Never shrink the offered window */
250 if (new_win < cur_win) {
251 /* Danger Will Robinson!
252 * Don't update rcv_wup/rcv_wnd here or else
253 * we will not be able to advertise a zero
254 * window in time. --DaveM
256 * Relax Will Robinson.
260 tp->rcv_wnd = new_win;
261 tp->rcv_wup = tp->rcv_nxt;
263 /* Make sure we do not exceed the maximum possible
266 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
267 new_win = min(new_win, MAX_TCP_WINDOW);
269 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
271 /* RFC1323 scaling applied */
272 new_win >>= tp->rx_opt.rcv_wscale;
274 /* If we advertise zero window, disable fast path. */
281 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
283 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
284 if (!(tp->ecn_flags & TCP_ECN_OK))
285 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
288 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
290 struct tcp_sock *tp = tcp_sk(sk);
293 if (sysctl_tcp_ecn) {
294 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
295 tp->ecn_flags = TCP_ECN_OK;
299 static __inline__ void
300 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
302 if (inet_rsk(req)->ecn_ok)
306 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
309 struct tcp_sock *tp = tcp_sk(sk);
311 if (tp->ecn_flags & TCP_ECN_OK) {
312 /* Not-retransmitted data segment: set ECT and inject CWR. */
313 if (skb->len != tcp_header_len &&
314 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
316 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
317 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
318 tcp_hdr(skb)->cwr = 1;
319 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
322 /* ACK or retransmitted segment: clear ECT|CE */
323 INET_ECN_dontxmit(sk);
325 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
326 tcp_hdr(skb)->ece = 1;
330 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
331 __u32 tstamp, __u8 **md5_hash)
333 if (tp->rx_opt.tstamp_ok) {
334 *ptr++ = htonl((TCPOPT_NOP << 24) |
336 (TCPOPT_TIMESTAMP << 8) |
338 *ptr++ = htonl(tstamp);
339 *ptr++ = htonl(tp->rx_opt.ts_recent);
341 if (tp->rx_opt.eff_sacks) {
342 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
345 *ptr++ = htonl((TCPOPT_NOP << 24) |
348 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
349 TCPOLEN_SACK_PERBLOCK)));
351 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
352 *ptr++ = htonl(sp[this_sack].start_seq);
353 *ptr++ = htonl(sp[this_sack].end_seq);
356 if (tp->rx_opt.dsack) {
357 tp->rx_opt.dsack = 0;
358 tp->rx_opt.eff_sacks--;
361 #ifdef CONFIG_TCP_MD5SIG
363 *ptr++ = htonl((TCPOPT_NOP << 24) |
365 (TCPOPT_MD5SIG << 8) |
367 *md5_hash = (__u8 *)ptr;
372 /* Construct a tcp options header for a SYN or SYN_ACK packet.
373 * If this is every changed make sure to change the definition of
374 * MAX_SYN_SIZE to match the new maximum number of options that you
377 * Note - that with the RFC2385 TCP option, we make room for the
378 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
379 * location to be filled is passed back up.
381 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
382 int offer_wscale, int wscale, __u32 tstamp,
383 __u32 ts_recent, __u8 **md5_hash)
385 /* We always get an MSS option.
386 * The option bytes which will be seen in normal data
387 * packets should timestamps be used, must be in the MSS
388 * advertised. But we subtract them from tp->mss_cache so
389 * that calculations in tcp_sendmsg are simpler etc.
390 * So account for this fact here if necessary. If we
391 * don't do this correctly, as a receiver we won't
392 * recognize data packets as being full sized when we
393 * should, and thus we won't abide by the delayed ACK
395 * SACKs don't matter, we never delay an ACK when we
396 * have any of those going out.
398 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
401 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
402 (TCPOLEN_SACK_PERM << 16) |
403 (TCPOPT_TIMESTAMP << 8) |
406 *ptr++ = htonl((TCPOPT_NOP << 24) |
408 (TCPOPT_TIMESTAMP << 8) |
410 *ptr++ = htonl(tstamp); /* TSVAL */
411 *ptr++ = htonl(ts_recent); /* TSECR */
413 *ptr++ = htonl((TCPOPT_NOP << 24) |
415 (TCPOPT_SACK_PERM << 8) |
418 *ptr++ = htonl((TCPOPT_NOP << 24) |
419 (TCPOPT_WINDOW << 16) |
420 (TCPOLEN_WINDOW << 8) |
422 #ifdef CONFIG_TCP_MD5SIG
424 * If MD5 is enabled, then we set the option, and include the size
425 * (always 18). The actual MD5 hash is added just before the
429 *ptr++ = htonl((TCPOPT_NOP << 24) |
431 (TCPOPT_MD5SIG << 8) |
433 *md5_hash = (__u8 *)ptr;
438 /* This routine actually transmits TCP packets queued in by
439 * tcp_do_sendmsg(). This is used by both the initial
440 * transmission and possible later retransmissions.
441 * All SKB's seen here are completely headerless. It is our
442 * job to build the TCP header, and pass the packet down to
443 * IP so it can do the same plus pass the packet off to the
446 * We are working here with either a clone of the original
447 * SKB, or a fresh unique copy made by the retransmit engine.
449 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
452 const struct inet_connection_sock *icsk = inet_csk(sk);
453 struct inet_sock *inet;
455 struct tcp_skb_cb *tcb;
457 #ifdef CONFIG_TCP_MD5SIG
458 struct tcp_md5sig_key *md5;
459 __u8 *md5_hash_location;
465 BUG_ON(!skb || !tcp_skb_pcount(skb));
467 /* If congestion control is doing timestamping, we must
468 * take such a timestamp before we potentially clone/copy.
470 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
471 __net_timestamp(skb);
473 if (likely(clone_it)) {
474 if (unlikely(skb_cloned(skb)))
475 skb = pskb_copy(skb, gfp_mask);
477 skb = skb_clone(skb, gfp_mask);
484 tcb = TCP_SKB_CB(skb);
485 tcp_header_size = tp->tcp_header_len;
487 #define SYSCTL_FLAG_TSTAMPS 0x1
488 #define SYSCTL_FLAG_WSCALE 0x2
489 #define SYSCTL_FLAG_SACK 0x4
492 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
493 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
494 if (sysctl_tcp_timestamps) {
495 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
496 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
498 if (sysctl_tcp_window_scaling) {
499 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
500 sysctl_flags |= SYSCTL_FLAG_WSCALE;
502 if (sysctl_tcp_sack) {
503 sysctl_flags |= SYSCTL_FLAG_SACK;
504 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
505 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
507 } else if (unlikely(tp->rx_opt.eff_sacks)) {
508 /* A SACK is 2 pad bytes, a 2 byte header, plus
509 * 2 32-bit sequence numbers for each SACK block.
511 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
512 (tp->rx_opt.eff_sacks *
513 TCPOLEN_SACK_PERBLOCK));
516 if (tcp_packets_in_flight(tp) == 0)
517 tcp_ca_event(sk, CA_EVENT_TX_START);
519 #ifdef CONFIG_TCP_MD5SIG
521 * Are we doing MD5 on this segment? If so - make
524 md5 = tp->af_specific->md5_lookup(sk, sk);
526 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
529 skb_push(skb, tcp_header_size);
530 skb_reset_transport_header(skb);
531 skb_set_owner_w(skb, sk);
533 /* Build TCP header and checksum it. */
535 th->source = inet->sport;
536 th->dest = inet->dport;
537 th->seq = htonl(tcb->seq);
538 th->ack_seq = htonl(tp->rcv_nxt);
539 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
542 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
543 /* RFC1323: The window in SYN & SYN/ACK segments
546 th->window = htons(min(tp->rcv_wnd, 65535U));
548 th->window = htons(tcp_select_window(sk));
553 if (unlikely(tp->urg_mode &&
554 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
555 th->urg_ptr = htons(tp->snd_up - tcb->seq);
559 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
560 tcp_syn_build_options((__be32 *)(th + 1),
561 tcp_advertise_mss(sk),
562 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
563 (sysctl_flags & SYSCTL_FLAG_SACK),
564 (sysctl_flags & SYSCTL_FLAG_WSCALE),
565 tp->rx_opt.rcv_wscale,
567 tp->rx_opt.ts_recent,
569 #ifdef CONFIG_TCP_MD5SIG
570 md5 ? &md5_hash_location :
574 tcp_build_and_update_options((__be32 *)(th + 1),
576 #ifdef CONFIG_TCP_MD5SIG
577 md5 ? &md5_hash_location :
580 TCP_ECN_send(sk, skb, tcp_header_size);
583 #ifdef CONFIG_TCP_MD5SIG
584 /* Calculate the MD5 hash, as we have all we need now */
586 tp->af_specific->calc_md5_hash(md5_hash_location,
595 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
597 if (likely(tcb->flags & TCPCB_FLAG_ACK))
598 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
600 if (skb->len != tcp_header_size)
601 tcp_event_data_sent(tp, skb, sk);
603 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
604 TCP_INC_STATS(TCP_MIB_OUTSEGS);
606 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
607 if (likely(err <= 0))
610 tcp_enter_cwr(sk, 1);
612 return net_xmit_eval(err);
614 #undef SYSCTL_FLAG_TSTAMPS
615 #undef SYSCTL_FLAG_WSCALE
616 #undef SYSCTL_FLAG_SACK
619 /* This routine just queue's the buffer
621 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
622 * otherwise socket can stall.
624 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
626 struct tcp_sock *tp = tcp_sk(sk);
628 /* Advance write_seq and place onto the write_queue. */
629 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
630 skb_header_release(skb);
631 tcp_add_write_queue_tail(sk, skb);
632 sk->sk_wmem_queued += skb->truesize;
633 sk_mem_charge(sk, skb->truesize);
636 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
637 unsigned int mss_now)
639 if (skb->len <= mss_now || !sk_can_gso(sk)) {
640 /* Avoid the costly divide in the normal
643 skb_shinfo(skb)->gso_segs = 1;
644 skb_shinfo(skb)->gso_size = 0;
645 skb_shinfo(skb)->gso_type = 0;
647 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
648 skb_shinfo(skb)->gso_size = mss_now;
649 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
653 /* When a modification to fackets out becomes necessary, we need to check
654 * skb is counted to fackets_out or not.
656 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
659 struct tcp_sock *tp = tcp_sk(sk);
661 if (!tp->sacked_out || tcp_is_reno(tp))
664 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
665 tp->fackets_out -= decr;
668 /* Function to create two new TCP segments. Shrinks the given segment
669 * to the specified size and appends a new segment with the rest of the
670 * packet to the list. This won't be called frequently, I hope.
671 * Remember, these are still headerless SKBs at this point.
673 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
674 unsigned int mss_now)
676 struct tcp_sock *tp = tcp_sk(sk);
677 struct sk_buff *buff;
678 int nsize, old_factor;
682 BUG_ON(len > skb->len);
684 tcp_clear_retrans_hints_partial(tp);
685 nsize = skb_headlen(skb) - len;
689 if (skb_cloned(skb) &&
690 skb_is_nonlinear(skb) &&
691 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
694 /* Get a new skb... force flag on. */
695 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
697 return -ENOMEM; /* We'll just try again later. */
699 sk->sk_wmem_queued += buff->truesize;
700 sk_mem_charge(sk, buff->truesize);
701 nlen = skb->len - len - nsize;
702 buff->truesize += nlen;
703 skb->truesize -= nlen;
705 /* Correct the sequence numbers. */
706 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
707 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
708 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
710 /* PSH and FIN should only be set in the second packet. */
711 flags = TCP_SKB_CB(skb)->flags;
712 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
713 TCP_SKB_CB(buff)->flags = flags;
714 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
716 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
717 /* Copy and checksum data tail into the new buffer. */
718 buff->csum = csum_partial_copy_nocheck(skb->data + len,
719 skb_put(buff, nsize),
724 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
726 skb->ip_summed = CHECKSUM_PARTIAL;
727 skb_split(skb, buff, len);
730 buff->ip_summed = skb->ip_summed;
732 /* Looks stupid, but our code really uses when of
733 * skbs, which it never sent before. --ANK
735 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
736 buff->tstamp = skb->tstamp;
738 old_factor = tcp_skb_pcount(skb);
740 /* Fix up tso_factor for both original and new SKB. */
741 tcp_set_skb_tso_segs(sk, skb, mss_now);
742 tcp_set_skb_tso_segs(sk, buff, mss_now);
744 /* If this packet has been sent out already, we must
745 * adjust the various packet counters.
747 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
748 int diff = old_factor - tcp_skb_pcount(skb) -
749 tcp_skb_pcount(buff);
751 tp->packets_out -= diff;
753 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
754 tp->sacked_out -= diff;
755 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
756 tp->retrans_out -= diff;
758 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
759 tp->lost_out -= diff;
761 /* Adjust Reno SACK estimate. */
762 if (tcp_is_reno(tp) && diff > 0) {
763 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
764 tcp_verify_left_out(tp);
766 tcp_adjust_fackets_out(sk, skb, diff);
769 /* Link BUFF into the send queue. */
770 skb_header_release(buff);
771 tcp_insert_write_queue_after(skb, buff, sk);
776 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
777 * eventually). The difference is that pulled data not copied, but
778 * immediately discarded.
780 static void __pskb_trim_head(struct sk_buff *skb, int len)
786 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
787 if (skb_shinfo(skb)->frags[i].size <= eat) {
788 put_page(skb_shinfo(skb)->frags[i].page);
789 eat -= skb_shinfo(skb)->frags[i].size;
791 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
793 skb_shinfo(skb)->frags[k].page_offset += eat;
794 skb_shinfo(skb)->frags[k].size -= eat;
800 skb_shinfo(skb)->nr_frags = k;
802 skb_reset_tail_pointer(skb);
803 skb->data_len -= len;
804 skb->len = skb->data_len;
807 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
809 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
812 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
813 if (unlikely(len < skb_headlen(skb)))
814 __skb_pull(skb, len);
816 __pskb_trim_head(skb, len - skb_headlen(skb));
818 TCP_SKB_CB(skb)->seq += len;
819 skb->ip_summed = CHECKSUM_PARTIAL;
821 skb->truesize -= len;
822 sk->sk_wmem_queued -= len;
823 sk_mem_uncharge(sk, len);
824 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
826 /* Any change of skb->len requires recalculation of tso
829 if (tcp_skb_pcount(skb) > 1)
830 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
835 /* Not accounting for SACKs here. */
836 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
838 struct tcp_sock *tp = tcp_sk(sk);
839 struct inet_connection_sock *icsk = inet_csk(sk);
842 /* Calculate base mss without TCP options:
843 It is MMS_S - sizeof(tcphdr) of rfc1122
845 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
847 /* Clamp it (mss_clamp does not include tcp options) */
848 if (mss_now > tp->rx_opt.mss_clamp)
849 mss_now = tp->rx_opt.mss_clamp;
851 /* Now subtract optional transport overhead */
852 mss_now -= icsk->icsk_ext_hdr_len;
854 /* Then reserve room for full set of TCP options and 8 bytes of data */
858 /* Now subtract TCP options size, not including SACKs */
859 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
864 /* Inverse of above */
865 int tcp_mss_to_mtu(struct sock *sk, int mss)
867 struct tcp_sock *tp = tcp_sk(sk);
868 struct inet_connection_sock *icsk = inet_csk(sk);
873 icsk->icsk_ext_hdr_len +
874 icsk->icsk_af_ops->net_header_len;
879 void tcp_mtup_init(struct sock *sk)
881 struct tcp_sock *tp = tcp_sk(sk);
882 struct inet_connection_sock *icsk = inet_csk(sk);
884 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
885 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
886 icsk->icsk_af_ops->net_header_len;
887 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
888 icsk->icsk_mtup.probe_size = 0;
891 /* Bound MSS / TSO packet size with the half of the window */
892 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
894 if (tp->max_window && pktsize > (tp->max_window >> 1))
895 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
900 /* This function synchronize snd mss to current pmtu/exthdr set.
902 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
903 for TCP options, but includes only bare TCP header.
905 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
906 It is minimum of user_mss and mss received with SYN.
907 It also does not include TCP options.
909 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
911 tp->mss_cache is current effective sending mss, including
912 all tcp options except for SACKs. It is evaluated,
913 taking into account current pmtu, but never exceeds
914 tp->rx_opt.mss_clamp.
916 NOTE1. rfc1122 clearly states that advertised MSS
917 DOES NOT include either tcp or ip options.
919 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
920 are READ ONLY outside this function. --ANK (980731)
922 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
924 struct tcp_sock *tp = tcp_sk(sk);
925 struct inet_connection_sock *icsk = inet_csk(sk);
928 if (icsk->icsk_mtup.search_high > pmtu)
929 icsk->icsk_mtup.search_high = pmtu;
931 mss_now = tcp_mtu_to_mss(sk, pmtu);
932 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
934 /* And store cached results */
935 icsk->icsk_pmtu_cookie = pmtu;
936 if (icsk->icsk_mtup.enabled)
937 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
938 tp->mss_cache = mss_now;
943 /* Compute the current effective MSS, taking SACKs and IP options,
944 * and even PMTU discovery events into account.
946 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
947 * cannot be large. However, taking into account rare use of URG, this
950 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
952 struct tcp_sock *tp = tcp_sk(sk);
953 struct dst_entry *dst = __sk_dst_get(sk);
958 mss_now = tp->mss_cache;
960 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
964 u32 mtu = dst_mtu(dst);
965 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
966 mss_now = tcp_sync_mss(sk, mtu);
969 if (tp->rx_opt.eff_sacks)
970 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
971 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
973 #ifdef CONFIG_TCP_MD5SIG
974 if (tp->af_specific->md5_lookup(sk, sk))
975 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
978 xmit_size_goal = mss_now;
981 xmit_size_goal = (65535 -
982 inet_csk(sk)->icsk_af_ops->net_header_len -
983 inet_csk(sk)->icsk_ext_hdr_len -
986 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
987 xmit_size_goal -= (xmit_size_goal % mss_now);
989 tp->xmit_size_goal = xmit_size_goal;
994 /* Congestion window validation. (RFC2861) */
995 static void tcp_cwnd_validate(struct sock *sk)
997 struct tcp_sock *tp = tcp_sk(sk);
998 __u32 packets_out = tp->packets_out;
1000 if (packets_out >= tp->snd_cwnd) {
1001 /* Network is feed fully. */
1002 tp->snd_cwnd_used = 0;
1003 tp->snd_cwnd_stamp = tcp_time_stamp;
1005 /* Network starves. */
1006 if (tp->packets_out > tp->snd_cwnd_used)
1007 tp->snd_cwnd_used = tp->packets_out;
1009 if (sysctl_tcp_slow_start_after_idle &&
1010 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1011 tcp_cwnd_application_limited(sk);
1015 /* Returns the portion of skb which can be sent right away without
1016 * introducing MSS oddities to segment boundaries. In rare cases where
1017 * mss_now != mss_cache, we will request caller to create a small skb
1018 * per input skb which could be mostly avoided here (if desired).
1020 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1021 unsigned int mss_now, unsigned int cwnd)
1023 struct tcp_sock *tp = tcp_sk(sk);
1024 u32 needed, window, cwnd_len;
1026 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1027 cwnd_len = mss_now * cwnd;
1029 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1032 if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len)
1035 needed = min(skb->len, window);
1036 return needed - needed % mss_now;
1039 /* Can at least one segment of SKB be sent right now, according to the
1040 * congestion window rules? If so, return how many segments are allowed.
1042 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1043 struct sk_buff *skb)
1045 u32 in_flight, cwnd;
1047 /* Don't be strict about the congestion window for the final FIN. */
1048 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1049 tcp_skb_pcount(skb) == 1)
1052 in_flight = tcp_packets_in_flight(tp);
1053 cwnd = tp->snd_cwnd;
1054 if (in_flight < cwnd)
1055 return (cwnd - in_flight);
1060 /* This must be invoked the first time we consider transmitting
1061 * SKB onto the wire.
1063 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1064 unsigned int mss_now)
1066 int tso_segs = tcp_skb_pcount(skb);
1068 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1069 tcp_set_skb_tso_segs(sk, skb, mss_now);
1070 tso_segs = tcp_skb_pcount(skb);
1075 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1077 return after(tp->snd_sml,tp->snd_una) &&
1078 !after(tp->snd_sml, tp->snd_nxt);
1081 /* Return 0, if packet can be sent now without violation Nagle's rules:
1082 * 1. It is full sized.
1083 * 2. Or it contains FIN. (already checked by caller)
1084 * 3. Or TCP_NODELAY was set.
1085 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1086 * With Minshall's modification: all sent small packets are ACKed.
1088 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1089 const struct sk_buff *skb,
1090 unsigned mss_now, int nonagle)
1092 return (skb->len < mss_now &&
1093 ((nonagle & TCP_NAGLE_CORK) ||
1094 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1097 /* Return non-zero if the Nagle test allows this packet to be
1100 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1101 unsigned int cur_mss, int nonagle)
1103 /* Nagle rule does not apply to frames, which sit in the middle of the
1104 * write_queue (they have no chances to get new data).
1106 * This is implemented in the callers, where they modify the 'nonagle'
1107 * argument based upon the location of SKB in the send queue.
1109 if (nonagle & TCP_NAGLE_PUSH)
1112 /* Don't use the nagle rule for urgent data (or for the final FIN).
1113 * Nagle can be ignored during F-RTO too (see RFC4138).
1115 if (tp->urg_mode || (tp->frto_counter == 2) ||
1116 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1119 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1125 /* Does at least the first segment of SKB fit into the send window? */
1126 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1127 unsigned int cur_mss)
1129 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1131 if (skb->len > cur_mss)
1132 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1134 return !after(end_seq, tcp_wnd_end(tp));
1137 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1138 * should be put on the wire right now. If so, it returns the number of
1139 * packets allowed by the congestion window.
1141 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1142 unsigned int cur_mss, int nonagle)
1144 struct tcp_sock *tp = tcp_sk(sk);
1145 unsigned int cwnd_quota;
1147 tcp_init_tso_segs(sk, skb, cur_mss);
1149 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1152 cwnd_quota = tcp_cwnd_test(tp, skb);
1153 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1159 int tcp_may_send_now(struct sock *sk)
1161 struct tcp_sock *tp = tcp_sk(sk);
1162 struct sk_buff *skb = tcp_send_head(sk);
1165 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1166 (tcp_skb_is_last(sk, skb) ?
1167 tp->nonagle : TCP_NAGLE_PUSH)));
1170 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1171 * which is put after SKB on the list. It is very much like
1172 * tcp_fragment() except that it may make several kinds of assumptions
1173 * in order to speed up the splitting operation. In particular, we
1174 * know that all the data is in scatter-gather pages, and that the
1175 * packet has never been sent out before (and thus is not cloned).
1177 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1178 unsigned int mss_now)
1180 struct sk_buff *buff;
1181 int nlen = skb->len - len;
1184 /* All of a TSO frame must be composed of paged data. */
1185 if (skb->len != skb->data_len)
1186 return tcp_fragment(sk, skb, len, mss_now);
1188 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1189 if (unlikely(buff == NULL))
1192 sk->sk_wmem_queued += buff->truesize;
1193 sk_mem_charge(sk, buff->truesize);
1194 buff->truesize += nlen;
1195 skb->truesize -= nlen;
1197 /* Correct the sequence numbers. */
1198 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1199 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1200 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1202 /* PSH and FIN should only be set in the second packet. */
1203 flags = TCP_SKB_CB(skb)->flags;
1204 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1205 TCP_SKB_CB(buff)->flags = flags;
1207 /* This packet was never sent out yet, so no SACK bits. */
1208 TCP_SKB_CB(buff)->sacked = 0;
1210 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1211 skb_split(skb, buff, len);
1213 /* Fix up tso_factor for both original and new SKB. */
1214 tcp_set_skb_tso_segs(sk, skb, mss_now);
1215 tcp_set_skb_tso_segs(sk, buff, mss_now);
1217 /* Link BUFF into the send queue. */
1218 skb_header_release(buff);
1219 tcp_insert_write_queue_after(skb, buff, sk);
1224 /* Try to defer sending, if possible, in order to minimize the amount
1225 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1227 * This algorithm is from John Heffner.
1229 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1231 struct tcp_sock *tp = tcp_sk(sk);
1232 const struct inet_connection_sock *icsk = inet_csk(sk);
1233 u32 send_win, cong_win, limit, in_flight;
1235 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1238 if (icsk->icsk_ca_state != TCP_CA_Open)
1241 /* Defer for less than two clock ticks. */
1242 if (tp->tso_deferred &&
1243 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1246 in_flight = tcp_packets_in_flight(tp);
1248 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1250 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1252 /* From in_flight test above, we know that cwnd > in_flight. */
1253 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1255 limit = min(send_win, cong_win);
1257 /* If a full-sized TSO skb can be sent, do it. */
1261 if (sysctl_tcp_tso_win_divisor) {
1262 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1264 /* If at least some fraction of a window is available,
1267 chunk /= sysctl_tcp_tso_win_divisor;
1271 /* Different approach, try not to defer past a single
1272 * ACK. Receiver should ACK every other full sized
1273 * frame, so if we have space for more than 3 frames
1276 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1280 /* Ok, it looks like it is advisable to defer. */
1281 tp->tso_deferred = 1 | (jiffies << 1);
1286 tp->tso_deferred = 0;
1290 /* Create a new MTU probe if we are ready.
1291 * Returns 0 if we should wait to probe (no cwnd available),
1292 * 1 if a probe was sent,
1295 static int tcp_mtu_probe(struct sock *sk)
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct inet_connection_sock *icsk = inet_csk(sk);
1299 struct sk_buff *skb, *nskb, *next;
1306 /* Not currently probing/verifying,
1308 * have enough cwnd, and
1309 * not SACKing (the variable headers throw things off) */
1310 if (!icsk->icsk_mtup.enabled ||
1311 icsk->icsk_mtup.probe_size ||
1312 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1313 tp->snd_cwnd < 11 ||
1314 tp->rx_opt.eff_sacks)
1317 /* Very simple search strategy: just double the MSS. */
1318 mss_now = tcp_current_mss(sk, 0);
1319 probe_size = 2 * tp->mss_cache;
1320 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1321 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1322 /* TODO: set timer for probe_converge_event */
1326 /* Have enough data in the send queue to probe? */
1327 if (tp->write_seq - tp->snd_nxt < size_needed)
1330 if (tp->snd_wnd < size_needed)
1332 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1335 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1336 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1337 if (!tcp_packets_in_flight(tp))
1343 /* We're allowed to probe. Build it now. */
1344 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1346 sk->sk_wmem_queued += nskb->truesize;
1347 sk_mem_charge(sk, nskb->truesize);
1349 skb = tcp_send_head(sk);
1351 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1352 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1353 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1354 TCP_SKB_CB(nskb)->sacked = 0;
1356 nskb->ip_summed = skb->ip_summed;
1358 tcp_insert_write_queue_before(nskb, skb, sk);
1361 tcp_for_write_queue_from_safe(skb, next, sk) {
1362 copy = min_t(int, skb->len, probe_size - len);
1363 if (nskb->ip_summed)
1364 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1366 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1367 skb_put(nskb, copy),
1370 if (skb->len <= copy) {
1371 /* We've eaten all the data from this skb.
1373 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1374 tcp_unlink_write_queue(skb, sk);
1375 sk_wmem_free_skb(sk, skb);
1377 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1378 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1379 if (!skb_shinfo(skb)->nr_frags) {
1380 skb_pull(skb, copy);
1381 if (skb->ip_summed != CHECKSUM_PARTIAL)
1382 skb->csum = csum_partial(skb->data,
1385 __pskb_trim_head(skb, copy);
1386 tcp_set_skb_tso_segs(sk, skb, mss_now);
1388 TCP_SKB_CB(skb)->seq += copy;
1393 if (len >= probe_size)
1396 tcp_init_tso_segs(sk, nskb, nskb->len);
1398 /* We're ready to send. If this fails, the probe will
1399 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1400 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1401 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1402 /* Decrement cwnd here because we are sending
1403 * effectively two packets. */
1405 tcp_event_new_data_sent(sk, nskb);
1407 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1408 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1409 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1417 /* This routine writes packets to the network. It advances the
1418 * send_head. This happens as incoming acks open up the remote
1421 * Returns 1, if no segments are in flight and we have queued segments, but
1422 * cannot send anything now because of SWS or another problem.
1424 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1426 struct tcp_sock *tp = tcp_sk(sk);
1427 struct sk_buff *skb;
1428 unsigned int tso_segs, sent_pkts;
1432 /* If we are closed, the bytes will have to remain here.
1433 * In time closedown will finish, we empty the write queue and all
1436 if (unlikely(sk->sk_state == TCP_CLOSE))
1441 /* Do MTU probing. */
1442 if ((result = tcp_mtu_probe(sk)) == 0) {
1444 } else if (result > 0) {
1448 while ((skb = tcp_send_head(sk))) {
1451 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1454 cwnd_quota = tcp_cwnd_test(tp, skb);
1458 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1461 if (tso_segs == 1) {
1462 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1463 (tcp_skb_is_last(sk, skb) ?
1464 nonagle : TCP_NAGLE_PUSH))))
1467 if (tcp_tso_should_defer(sk, skb))
1473 limit = tcp_mss_split_point(sk, skb, mss_now,
1476 if (skb->len > limit &&
1477 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1480 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1482 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1485 /* Advance the send_head. This one is sent out.
1486 * This call will increment packets_out.
1488 tcp_event_new_data_sent(sk, skb);
1490 tcp_minshall_update(tp, mss_now, skb);
1494 if (likely(sent_pkts)) {
1495 tcp_cwnd_validate(sk);
1498 return !tp->packets_out && tcp_send_head(sk);
1501 /* Push out any pending frames which were held back due to
1502 * TCP_CORK or attempt at coalescing tiny packets.
1503 * The socket must be locked by the caller.
1505 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1508 struct sk_buff *skb = tcp_send_head(sk);
1511 if (tcp_write_xmit(sk, cur_mss, nonagle))
1512 tcp_check_probe_timer(sk);
1516 /* Send _single_ skb sitting at the send head. This function requires
1517 * true push pending frames to setup probe timer etc.
1519 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1521 struct sk_buff *skb = tcp_send_head(sk);
1522 unsigned int tso_segs, cwnd_quota;
1524 BUG_ON(!skb || skb->len < mss_now);
1526 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1527 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1529 if (likely(cwnd_quota)) {
1536 limit = tcp_mss_split_point(sk, skb, mss_now,
1539 if (skb->len > limit &&
1540 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1543 /* Send it out now. */
1544 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1546 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1547 tcp_event_new_data_sent(sk, skb);
1548 tcp_cwnd_validate(sk);
1554 /* This function returns the amount that we can raise the
1555 * usable window based on the following constraints
1557 * 1. The window can never be shrunk once it is offered (RFC 793)
1558 * 2. We limit memory per socket
1561 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1562 * RECV.NEXT + RCV.WIN fixed until:
1563 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1565 * i.e. don't raise the right edge of the window until you can raise
1566 * it at least MSS bytes.
1568 * Unfortunately, the recommended algorithm breaks header prediction,
1569 * since header prediction assumes th->window stays fixed.
1571 * Strictly speaking, keeping th->window fixed violates the receiver
1572 * side SWS prevention criteria. The problem is that under this rule
1573 * a stream of single byte packets will cause the right side of the
1574 * window to always advance by a single byte.
1576 * Of course, if the sender implements sender side SWS prevention
1577 * then this will not be a problem.
1579 * BSD seems to make the following compromise:
1581 * If the free space is less than the 1/4 of the maximum
1582 * space available and the free space is less than 1/2 mss,
1583 * then set the window to 0.
1584 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1585 * Otherwise, just prevent the window from shrinking
1586 * and from being larger than the largest representable value.
1588 * This prevents incremental opening of the window in the regime
1589 * where TCP is limited by the speed of the reader side taking
1590 * data out of the TCP receive queue. It does nothing about
1591 * those cases where the window is constrained on the sender side
1592 * because the pipeline is full.
1594 * BSD also seems to "accidentally" limit itself to windows that are a
1595 * multiple of MSS, at least until the free space gets quite small.
1596 * This would appear to be a side effect of the mbuf implementation.
1597 * Combining these two algorithms results in the observed behavior
1598 * of having a fixed window size at almost all times.
1600 * Below we obtain similar behavior by forcing the offered window to
1601 * a multiple of the mss when it is feasible to do so.
1603 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1604 * Regular options like TIMESTAMP are taken into account.
1606 u32 __tcp_select_window(struct sock *sk)
1608 struct inet_connection_sock *icsk = inet_csk(sk);
1609 struct tcp_sock *tp = tcp_sk(sk);
1610 /* MSS for the peer's data. Previous versions used mss_clamp
1611 * here. I don't know if the value based on our guesses
1612 * of peer's MSS is better for the performance. It's more correct
1613 * but may be worse for the performance because of rcv_mss
1614 * fluctuations. --SAW 1998/11/1
1616 int mss = icsk->icsk_ack.rcv_mss;
1617 int free_space = tcp_space(sk);
1618 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1621 if (mss > full_space)
1624 if (free_space < (full_space >> 1)) {
1625 icsk->icsk_ack.quick = 0;
1627 if (tcp_memory_pressure)
1628 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1631 if (free_space < mss)
1635 if (free_space > tp->rcv_ssthresh)
1636 free_space = tp->rcv_ssthresh;
1638 /* Don't do rounding if we are using window scaling, since the
1639 * scaled window will not line up with the MSS boundary anyway.
1641 window = tp->rcv_wnd;
1642 if (tp->rx_opt.rcv_wscale) {
1643 window = free_space;
1645 /* Advertise enough space so that it won't get scaled away.
1646 * Import case: prevent zero window announcement if
1647 * 1<<rcv_wscale > mss.
1649 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1650 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1651 << tp->rx_opt.rcv_wscale);
1653 /* Get the largest window that is a nice multiple of mss.
1654 * Window clamp already applied above.
1655 * If our current window offering is within 1 mss of the
1656 * free space we just keep it. This prevents the divide
1657 * and multiply from happening most of the time.
1658 * We also don't do any window rounding when the free space
1661 if (window <= free_space - mss || window > free_space)
1662 window = (free_space / mss) * mss;
1663 else if (mss == full_space &&
1664 free_space > window + (full_space >> 1))
1665 window = free_space;
1671 /* Attempt to collapse two adjacent SKB's during retransmission. */
1672 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1675 struct tcp_sock *tp = tcp_sk(sk);
1676 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1677 int skb_size, next_skb_size;
1680 /* The first test we must make is that neither of these two
1681 * SKB's are still referenced by someone else.
1683 if (skb_cloned(skb) || skb_cloned(next_skb))
1686 skb_size = skb->len;
1687 next_skb_size = next_skb->len;
1688 flags = TCP_SKB_CB(skb)->flags;
1690 /* Also punt if next skb has been SACK'd. */
1691 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1694 /* Next skb is out of window. */
1695 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1698 /* Punt if not enough space exists in the first SKB for
1699 * the data in the second, or the total combined payload
1700 * would exceed the MSS.
1702 if ((next_skb_size > skb_tailroom(skb)) ||
1703 ((skb_size + next_skb_size) > mss_now))
1706 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1708 tcp_highest_sack_combine(sk, next_skb, skb);
1710 /* Ok. We will be able to collapse the packet. */
1711 tcp_unlink_write_queue(next_skb, sk);
1713 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1716 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1717 skb->ip_summed = CHECKSUM_PARTIAL;
1719 if (skb->ip_summed != CHECKSUM_PARTIAL)
1720 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1722 /* Update sequence range on original skb. */
1723 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1725 /* Merge over control information. */
1726 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1727 TCP_SKB_CB(skb)->flags = flags;
1729 /* All done, get rid of second SKB and account for it so
1730 * packet counting does not break.
1732 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1733 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1734 tp->retrans_out -= tcp_skb_pcount(next_skb);
1735 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1736 tp->lost_out -= tcp_skb_pcount(next_skb);
1737 /* Reno case is special. Sigh... */
1738 if (tcp_is_reno(tp) && tp->sacked_out)
1739 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1741 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1742 tp->packets_out -= tcp_skb_pcount(next_skb);
1744 /* changed transmit queue under us so clear hints */
1745 tcp_clear_retrans_hints_partial(tp);
1747 sk_wmem_free_skb(sk, next_skb);
1750 /* Do a simple retransmit without using the backoff mechanisms in
1751 * tcp_timer. This is used for path mtu discovery.
1752 * The socket is already locked here.
1754 void tcp_simple_retransmit(struct sock *sk)
1756 const struct inet_connection_sock *icsk = inet_csk(sk);
1757 struct tcp_sock *tp = tcp_sk(sk);
1758 struct sk_buff *skb;
1759 unsigned int mss = tcp_current_mss(sk, 0);
1762 tcp_for_write_queue(skb, sk) {
1763 if (skb == tcp_send_head(sk))
1765 if (skb->len > mss &&
1766 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1767 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1768 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1769 tp->retrans_out -= tcp_skb_pcount(skb);
1771 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
1772 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1773 tp->lost_out += tcp_skb_pcount(skb);
1779 tcp_clear_all_retrans_hints(tp);
1784 tcp_verify_left_out(tp);
1786 /* Don't muck with the congestion window here.
1787 * Reason is that we do not increase amount of _data_
1788 * in network, but units changed and effective
1789 * cwnd/ssthresh really reduced now.
1791 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1792 tp->high_seq = tp->snd_nxt;
1793 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1794 tp->prior_ssthresh = 0;
1795 tp->undo_marker = 0;
1796 tcp_set_ca_state(sk, TCP_CA_Loss);
1798 tcp_xmit_retransmit_queue(sk);
1801 /* This retransmits one SKB. Policy decisions and retransmit queue
1802 * state updates are done by the caller. Returns non-zero if an
1803 * error occurred which prevented the send.
1805 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1807 struct tcp_sock *tp = tcp_sk(sk);
1808 struct inet_connection_sock *icsk = inet_csk(sk);
1809 unsigned int cur_mss = tcp_current_mss(sk, 0);
1812 /* Inconslusive MTU probe */
1813 if (icsk->icsk_mtup.probe_size) {
1814 icsk->icsk_mtup.probe_size = 0;
1817 /* Do not sent more than we queued. 1/4 is reserved for possible
1818 * copying overhead: fragmentation, tunneling, mangling etc.
1820 if (atomic_read(&sk->sk_wmem_alloc) >
1821 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1824 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1825 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1827 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1831 /* If receiver has shrunk his window, and skb is out of
1832 * new window, do not retransmit it. The exception is the
1833 * case, when window is shrunk to zero. In this case
1834 * our retransmit serves as a zero window probe.
1836 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1837 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1840 if (skb->len > cur_mss) {
1841 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1842 return -ENOMEM; /* We'll try again later. */
1845 /* Collapse two adjacent packets if worthwhile and we can. */
1846 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1847 (skb->len < (cur_mss >> 1)) &&
1848 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1849 (!tcp_skb_is_last(sk, skb)) &&
1850 (skb_shinfo(skb)->nr_frags == 0 &&
1851 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1852 (tcp_skb_pcount(skb) == 1 &&
1853 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1854 (sysctl_tcp_retrans_collapse != 0))
1855 tcp_retrans_try_collapse(sk, skb, cur_mss);
1857 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1858 return -EHOSTUNREACH; /* Routing failure or similar. */
1860 /* Some Solaris stacks overoptimize and ignore the FIN on a
1861 * retransmit when old data is attached. So strip it off
1862 * since it is cheap to do so and saves bytes on the network.
1865 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1866 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1867 if (!pskb_trim(skb, 0)) {
1868 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1869 skb_shinfo(skb)->gso_segs = 1;
1870 skb_shinfo(skb)->gso_size = 0;
1871 skb_shinfo(skb)->gso_type = 0;
1872 skb->ip_summed = CHECKSUM_NONE;
1877 /* Make a copy, if the first transmission SKB clone we made
1878 * is still in somebody's hands, else make a clone.
1880 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1882 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1885 /* Update global TCP statistics. */
1886 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1888 tp->total_retrans++;
1890 #if FASTRETRANS_DEBUG > 0
1891 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1892 if (net_ratelimit())
1893 printk(KERN_DEBUG "retrans_out leaked.\n");
1896 if (!tp->retrans_out)
1897 tp->lost_retrans_low = tp->snd_nxt;
1898 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1899 tp->retrans_out += tcp_skb_pcount(skb);
1901 /* Save stamp of the first retransmit. */
1902 if (!tp->retrans_stamp)
1903 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1907 /* snd_nxt is stored to detect loss of retransmitted segment,
1908 * see tcp_input.c tcp_sacktag_write_queue().
1910 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1915 /* This gets called after a retransmit timeout, and the initially
1916 * retransmitted data is acknowledged. It tries to continue
1917 * resending the rest of the retransmit queue, until either
1918 * we've sent it all or the congestion window limit is reached.
1919 * If doing SACK, the first ACK which comes back for a timeout
1920 * based retransmit packet might feed us FACK information again.
1921 * If so, we use it to avoid unnecessarily retransmissions.
1923 void tcp_xmit_retransmit_queue(struct sock *sk)
1925 const struct inet_connection_sock *icsk = inet_csk(sk);
1926 struct tcp_sock *tp = tcp_sk(sk);
1927 struct sk_buff *skb;
1930 if (tp->retransmit_skb_hint) {
1931 skb = tp->retransmit_skb_hint;
1932 packet_cnt = tp->retransmit_cnt_hint;
1934 skb = tcp_write_queue_head(sk);
1938 /* First pass: retransmit lost packets. */
1940 tcp_for_write_queue_from(skb, sk) {
1941 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1943 if (skb == tcp_send_head(sk))
1945 /* we could do better than to assign each time */
1946 tp->retransmit_skb_hint = skb;
1947 tp->retransmit_cnt_hint = packet_cnt;
1949 /* Assume this retransmit will generate
1950 * only one packet for congestion window
1951 * calculation purposes. This works because
1952 * tcp_retransmit_skb() will chop up the
1953 * packet to be MSS sized and all the
1954 * packet counting works out.
1956 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1959 if (sacked & TCPCB_LOST) {
1960 if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1961 if (tcp_retransmit_skb(sk, skb)) {
1962 tp->retransmit_skb_hint = NULL;
1965 if (icsk->icsk_ca_state != TCP_CA_Loss)
1966 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1968 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1970 if (skb == tcp_write_queue_head(sk))
1971 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1972 inet_csk(sk)->icsk_rto,
1976 packet_cnt += tcp_skb_pcount(skb);
1977 if (packet_cnt >= tp->lost_out)
1983 /* OK, demanded retransmission is finished. */
1985 /* Forward retransmissions are possible only during Recovery. */
1986 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1989 /* No forward retransmissions in Reno are possible. */
1990 if (tcp_is_reno(tp))
1993 /* Yeah, we have to make difficult choice between forward transmission
1994 * and retransmission... Both ways have their merits...
1996 * For now we do not retransmit anything, while we have some new
1997 * segments to send. In the other cases, follow rule 3 for
1998 * NextSeg() specified in RFC3517.
2001 if (tcp_may_send_now(sk))
2004 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2005 if (!tp->sacked_out)
2008 if (tp->forward_skb_hint)
2009 skb = tp->forward_skb_hint;
2011 skb = tcp_write_queue_head(sk);
2013 tcp_for_write_queue_from(skb, sk) {
2014 if (skb == tcp_send_head(sk))
2016 tp->forward_skb_hint = skb;
2018 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2021 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2024 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2027 /* Ok, retransmit it. */
2028 if (tcp_retransmit_skb(sk, skb)) {
2029 tp->forward_skb_hint = NULL;
2033 if (skb == tcp_write_queue_head(sk))
2034 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2035 inet_csk(sk)->icsk_rto,
2038 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2042 /* Send a fin. The caller locks the socket for us. This cannot be
2043 * allowed to fail queueing a FIN frame under any circumstances.
2045 void tcp_send_fin(struct sock *sk)
2047 struct tcp_sock *tp = tcp_sk(sk);
2048 struct sk_buff *skb = tcp_write_queue_tail(sk);
2051 /* Optimization, tack on the FIN if we have a queue of
2052 * unsent frames. But be careful about outgoing SACKS
2055 mss_now = tcp_current_mss(sk, 1);
2057 if (tcp_send_head(sk) != NULL) {
2058 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2059 TCP_SKB_CB(skb)->end_seq++;
2062 /* Socket is locked, keep trying until memory is available. */
2064 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2070 /* Reserve space for headers and prepare control bits. */
2071 skb_reserve(skb, MAX_TCP_HEADER);
2073 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2074 TCP_SKB_CB(skb)->sacked = 0;
2075 skb_shinfo(skb)->gso_segs = 1;
2076 skb_shinfo(skb)->gso_size = 0;
2077 skb_shinfo(skb)->gso_type = 0;
2079 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2080 TCP_SKB_CB(skb)->seq = tp->write_seq;
2081 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2082 tcp_queue_skb(sk, skb);
2084 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2087 /* We get here when a process closes a file descriptor (either due to
2088 * an explicit close() or as a byproduct of exit()'ing) and there
2089 * was unread data in the receive queue. This behavior is recommended
2090 * by RFC 2525, section 2.17. -DaveM
2092 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2094 struct sk_buff *skb;
2096 /* NOTE: No TCP options attached and we never retransmit this. */
2097 skb = alloc_skb(MAX_TCP_HEADER, priority);
2099 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2103 /* Reserve space for headers and prepare control bits. */
2104 skb_reserve(skb, MAX_TCP_HEADER);
2106 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2107 TCP_SKB_CB(skb)->sacked = 0;
2108 skb_shinfo(skb)->gso_segs = 1;
2109 skb_shinfo(skb)->gso_size = 0;
2110 skb_shinfo(skb)->gso_type = 0;
2113 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
2114 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2115 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2116 if (tcp_transmit_skb(sk, skb, 0, priority))
2117 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2120 /* WARNING: This routine must only be called when we have already sent
2121 * a SYN packet that crossed the incoming SYN that caused this routine
2122 * to get called. If this assumption fails then the initial rcv_wnd
2123 * and rcv_wscale values will not be correct.
2125 int tcp_send_synack(struct sock *sk)
2127 struct sk_buff *skb;
2129 skb = tcp_write_queue_head(sk);
2130 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2131 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2134 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2135 if (skb_cloned(skb)) {
2136 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2139 tcp_unlink_write_queue(skb, sk);
2140 skb_header_release(nskb);
2141 __tcp_add_write_queue_head(sk, nskb);
2142 sk_wmem_free_skb(sk, skb);
2143 sk->sk_wmem_queued += nskb->truesize;
2144 sk_mem_charge(sk, nskb->truesize);
2148 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2149 TCP_ECN_send_synack(tcp_sk(sk), skb);
2151 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2152 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2156 * Prepare a SYN-ACK.
2158 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2159 struct request_sock *req)
2161 struct inet_request_sock *ireq = inet_rsk(req);
2162 struct tcp_sock *tp = tcp_sk(sk);
2164 int tcp_header_size;
2165 struct sk_buff *skb;
2166 #ifdef CONFIG_TCP_MD5SIG
2167 struct tcp_md5sig_key *md5;
2168 __u8 *md5_hash_location;
2171 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2175 /* Reserve space for headers. */
2176 skb_reserve(skb, MAX_TCP_HEADER);
2178 skb->dst = dst_clone(dst);
2180 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2181 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2182 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2183 /* SACK_PERM is in the place of NOP NOP of TS */
2184 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2186 #ifdef CONFIG_TCP_MD5SIG
2187 /* Are we doing MD5 on this segment? If so - make room for it */
2188 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2190 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2192 skb_push(skb, tcp_header_size);
2193 skb_reset_transport_header(skb);
2196 memset(th, 0, sizeof(struct tcphdr));
2199 TCP_ECN_make_synack(req, th);
2200 th->source = inet_sk(sk)->sport;
2201 th->dest = ireq->rmt_port;
2202 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2203 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2204 TCP_SKB_CB(skb)->sacked = 0;
2205 skb_shinfo(skb)->gso_segs = 1;
2206 skb_shinfo(skb)->gso_size = 0;
2207 skb_shinfo(skb)->gso_type = 0;
2208 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2209 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2210 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2212 /* Set this up on the first call only */
2213 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2214 /* tcp_full_space because it is guaranteed to be the first packet */
2215 tcp_select_initial_window(tcp_full_space(sk),
2216 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2221 ireq->rcv_wscale = rcv_wscale;
2224 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2225 th->window = htons(min(req->rcv_wnd, 65535U));
2227 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2228 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2229 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2230 TCP_SKB_CB(skb)->when,
2233 #ifdef CONFIG_TCP_MD5SIG
2234 md5 ? &md5_hash_location :
2240 th->doff = (tcp_header_size >> 2);
2241 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2243 #ifdef CONFIG_TCP_MD5SIG
2244 /* Okay, we have all we need - do the md5 hash if needed */
2246 tp->af_specific->calc_md5_hash(md5_hash_location,
2249 tcp_hdr(skb), sk->sk_protocol,
2258 * Do all connect socket setups that can be done AF independent.
2260 static void tcp_connect_init(struct sock *sk)
2262 struct dst_entry *dst = __sk_dst_get(sk);
2263 struct tcp_sock *tp = tcp_sk(sk);
2266 /* We'll fix this up when we get a response from the other end.
2267 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2269 tp->tcp_header_len = sizeof(struct tcphdr) +
2270 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2272 #ifdef CONFIG_TCP_MD5SIG
2273 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2274 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2277 /* If user gave his TCP_MAXSEG, record it to clamp */
2278 if (tp->rx_opt.user_mss)
2279 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2282 tcp_sync_mss(sk, dst_mtu(dst));
2284 if (!tp->window_clamp)
2285 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2286 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2287 tcp_initialize_rcv_mss(sk);
2289 tcp_select_initial_window(tcp_full_space(sk),
2290 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2293 sysctl_tcp_window_scaling,
2296 tp->rx_opt.rcv_wscale = rcv_wscale;
2297 tp->rcv_ssthresh = tp->rcv_wnd;
2300 sock_reset_flag(sk, SOCK_DONE);
2302 tcp_init_wl(tp, tp->write_seq, 0);
2303 tp->snd_una = tp->write_seq;
2304 tp->snd_sml = tp->write_seq;
2309 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2310 inet_csk(sk)->icsk_retransmits = 0;
2311 tcp_clear_retrans(tp);
2315 * Build a SYN and send it off.
2317 int tcp_connect(struct sock *sk)
2319 struct tcp_sock *tp = tcp_sk(sk);
2320 struct sk_buff *buff;
2322 tcp_connect_init(sk);
2324 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2325 if (unlikely(buff == NULL))
2328 /* Reserve space for headers. */
2329 skb_reserve(buff, MAX_TCP_HEADER);
2331 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2332 TCP_ECN_send_syn(sk, buff);
2333 TCP_SKB_CB(buff)->sacked = 0;
2334 skb_shinfo(buff)->gso_segs = 1;
2335 skb_shinfo(buff)->gso_size = 0;
2336 skb_shinfo(buff)->gso_type = 0;
2338 tp->snd_nxt = tp->write_seq;
2339 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2340 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2343 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2344 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2345 skb_header_release(buff);
2346 __tcp_add_write_queue_tail(sk, buff);
2347 sk->sk_wmem_queued += buff->truesize;
2348 sk_mem_charge(sk, buff->truesize);
2349 tp->packets_out += tcp_skb_pcount(buff);
2350 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2352 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2353 * in order to make this packet get counted in tcpOutSegs.
2355 tp->snd_nxt = tp->write_seq;
2356 tp->pushed_seq = tp->write_seq;
2357 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2359 /* Timer for repeating the SYN until an answer. */
2360 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2361 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2365 /* Send out a delayed ack, the caller does the policy checking
2366 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2369 void tcp_send_delayed_ack(struct sock *sk)
2371 struct inet_connection_sock *icsk = inet_csk(sk);
2372 int ato = icsk->icsk_ack.ato;
2373 unsigned long timeout;
2375 if (ato > TCP_DELACK_MIN) {
2376 const struct tcp_sock *tp = tcp_sk(sk);
2377 int max_ato = HZ / 2;
2379 if (icsk->icsk_ack.pingpong ||
2380 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2381 max_ato = TCP_DELACK_MAX;
2383 /* Slow path, intersegment interval is "high". */
2385 /* If some rtt estimate is known, use it to bound delayed ack.
2386 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2390 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2396 ato = min(ato, max_ato);
2399 /* Stay within the limit we were given */
2400 timeout = jiffies + ato;
2402 /* Use new timeout only if there wasn't a older one earlier. */
2403 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2404 /* If delack timer was blocked or is about to expire,
2407 if (icsk->icsk_ack.blocked ||
2408 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2413 if (!time_before(timeout, icsk->icsk_ack.timeout))
2414 timeout = icsk->icsk_ack.timeout;
2416 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2417 icsk->icsk_ack.timeout = timeout;
2418 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2421 /* This routine sends an ack and also updates the window. */
2422 void tcp_send_ack(struct sock *sk)
2424 struct sk_buff *buff;
2426 /* If we have been reset, we may not send again. */
2427 if (sk->sk_state == TCP_CLOSE)
2430 /* We are not putting this on the write queue, so
2431 * tcp_transmit_skb() will set the ownership to this
2434 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2436 inet_csk_schedule_ack(sk);
2437 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2438 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2439 TCP_DELACK_MAX, TCP_RTO_MAX);
2443 /* Reserve space for headers and prepare control bits. */
2444 skb_reserve(buff, MAX_TCP_HEADER);
2446 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2447 TCP_SKB_CB(buff)->sacked = 0;
2448 skb_shinfo(buff)->gso_segs = 1;
2449 skb_shinfo(buff)->gso_size = 0;
2450 skb_shinfo(buff)->gso_type = 0;
2452 /* Send it off, this clears delayed acks for us. */
2453 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
2454 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2455 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2458 /* This routine sends a packet with an out of date sequence
2459 * number. It assumes the other end will try to ack it.
2461 * Question: what should we make while urgent mode?
2462 * 4.4BSD forces sending single byte of data. We cannot send
2463 * out of window data, because we have SND.NXT==SND.MAX...
2465 * Current solution: to send TWO zero-length segments in urgent mode:
2466 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2467 * out-of-date with SND.UNA-1 to probe window.
2469 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2471 struct tcp_sock *tp = tcp_sk(sk);
2472 struct sk_buff *skb;
2474 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2475 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2479 /* Reserve space for headers and set control bits. */
2480 skb_reserve(skb, MAX_TCP_HEADER);
2482 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2483 TCP_SKB_CB(skb)->sacked = 0;
2484 skb_shinfo(skb)->gso_segs = 1;
2485 skb_shinfo(skb)->gso_size = 0;
2486 skb_shinfo(skb)->gso_type = 0;
2488 /* Use a previous sequence. This should cause the other
2489 * end to send an ack. Don't queue or clone SKB, just
2492 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2493 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2494 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2495 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2498 int tcp_write_wakeup(struct sock *sk)
2500 struct tcp_sock *tp = tcp_sk(sk);
2501 struct sk_buff *skb;
2503 if (sk->sk_state == TCP_CLOSE)
2506 if ((skb = tcp_send_head(sk)) != NULL &&
2507 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2509 unsigned int mss = tcp_current_mss(sk, 0);
2510 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2512 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2513 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2515 /* We are probing the opening of a window
2516 * but the window size is != 0
2517 * must have been a result SWS avoidance ( sender )
2519 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2521 seg_size = min(seg_size, mss);
2522 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2523 if (tcp_fragment(sk, skb, seg_size, mss))
2525 } else if (!tcp_skb_pcount(skb))
2526 tcp_set_skb_tso_segs(sk, skb, mss);
2528 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2529 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2530 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2532 tcp_event_new_data_sent(sk, skb);
2536 between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2537 tcp_xmit_probe_skb(sk, 1);
2538 return tcp_xmit_probe_skb(sk, 0);
2542 /* A window probe timeout has occurred. If window is not closed send
2543 * a partial packet else a zero probe.
2545 void tcp_send_probe0(struct sock *sk)
2547 struct inet_connection_sock *icsk = inet_csk(sk);
2548 struct tcp_sock *tp = tcp_sk(sk);
2551 err = tcp_write_wakeup(sk);
2553 if (tp->packets_out || !tcp_send_head(sk)) {
2554 /* Cancel probe timer, if it is not required. */
2555 icsk->icsk_probes_out = 0;
2556 icsk->icsk_backoff = 0;
2561 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2562 icsk->icsk_backoff++;
2563 icsk->icsk_probes_out++;
2564 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2565 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2568 /* If packet was not sent due to local congestion,
2569 * do not backoff and do not remember icsk_probes_out.
2570 * Let local senders to fight for local resources.
2572 * Use accumulated backoff yet.
2574 if (!icsk->icsk_probes_out)
2575 icsk->icsk_probes_out = 1;
2576 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2577 min(icsk->icsk_rto << icsk->icsk_backoff,
2578 TCP_RESOURCE_PROBE_INTERVAL),
2583 EXPORT_SYMBOL(tcp_connect);
2584 EXPORT_SYMBOL(tcp_make_synack);
2585 EXPORT_SYMBOL(tcp_simple_retransmit);
2586 EXPORT_SYMBOL(tcp_sync_mss);
2587 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2588 EXPORT_SYMBOL(tcp_mtup_init);