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,
284 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
285 if (!(tp->ecn_flags&TCP_ECN_OK))
286 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
289 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
291 struct tcp_sock *tp = tcp_sk(sk);
294 if (sysctl_tcp_ecn) {
295 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR;
296 tp->ecn_flags = TCP_ECN_OK;
300 static __inline__ void
301 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
303 if (inet_rsk(req)->ecn_ok)
307 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
310 struct tcp_sock *tp = tcp_sk(sk);
312 if (tp->ecn_flags & TCP_ECN_OK) {
313 /* Not-retransmitted data segment: set ECT and inject CWR. */
314 if (skb->len != tcp_header_len &&
315 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
317 if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) {
318 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
319 tcp_hdr(skb)->cwr = 1;
320 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
323 /* ACK or retransmitted segment: clear ECT|CE */
324 INET_ECN_dontxmit(sk);
326 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
327 tcp_hdr(skb)->ece = 1;
331 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
332 __u32 tstamp, __u8 **md5_hash)
334 if (tp->rx_opt.tstamp_ok) {
335 *ptr++ = htonl((TCPOPT_NOP << 24) |
337 (TCPOPT_TIMESTAMP << 8) |
339 *ptr++ = htonl(tstamp);
340 *ptr++ = htonl(tp->rx_opt.ts_recent);
342 if (tp->rx_opt.eff_sacks) {
343 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
346 *ptr++ = htonl((TCPOPT_NOP << 24) |
349 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
350 TCPOLEN_SACK_PERBLOCK)));
352 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
353 *ptr++ = htonl(sp[this_sack].start_seq);
354 *ptr++ = htonl(sp[this_sack].end_seq);
357 if (tp->rx_opt.dsack) {
358 tp->rx_opt.dsack = 0;
359 tp->rx_opt.eff_sacks--;
362 #ifdef CONFIG_TCP_MD5SIG
364 *ptr++ = htonl((TCPOPT_NOP << 24) |
366 (TCPOPT_MD5SIG << 8) |
368 *md5_hash = (__u8 *)ptr;
373 /* Construct a tcp options header for a SYN or SYN_ACK packet.
374 * If this is every changed make sure to change the definition of
375 * MAX_SYN_SIZE to match the new maximum number of options that you
378 * Note - that with the RFC2385 TCP option, we make room for the
379 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
380 * location to be filled is passed back up.
382 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
383 int offer_wscale, int wscale, __u32 tstamp,
384 __u32 ts_recent, __u8 **md5_hash)
386 /* We always get an MSS option.
387 * The option bytes which will be seen in normal data
388 * packets should timestamps be used, must be in the MSS
389 * advertised. But we subtract them from tp->mss_cache so
390 * that calculations in tcp_sendmsg are simpler etc.
391 * So account for this fact here if necessary. If we
392 * don't do this correctly, as a receiver we won't
393 * recognize data packets as being full sized when we
394 * should, and thus we won't abide by the delayed ACK
396 * SACKs don't matter, we never delay an ACK when we
397 * have any of those going out.
399 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
402 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
403 (TCPOLEN_SACK_PERM << 16) |
404 (TCPOPT_TIMESTAMP << 8) |
407 *ptr++ = htonl((TCPOPT_NOP << 24) |
409 (TCPOPT_TIMESTAMP << 8) |
411 *ptr++ = htonl(tstamp); /* TSVAL */
412 *ptr++ = htonl(ts_recent); /* TSECR */
414 *ptr++ = htonl((TCPOPT_NOP << 24) |
416 (TCPOPT_SACK_PERM << 8) |
419 *ptr++ = htonl((TCPOPT_NOP << 24) |
420 (TCPOPT_WINDOW << 16) |
421 (TCPOLEN_WINDOW << 8) |
423 #ifdef CONFIG_TCP_MD5SIG
425 * If MD5 is enabled, then we set the option, and include the size
426 * (always 18). The actual MD5 hash is added just before the
430 *ptr++ = htonl((TCPOPT_NOP << 24) |
432 (TCPOPT_MD5SIG << 8) |
434 *md5_hash = (__u8 *) ptr;
439 /* This routine actually transmits TCP packets queued in by
440 * tcp_do_sendmsg(). This is used by both the initial
441 * transmission and possible later retransmissions.
442 * All SKB's seen here are completely headerless. It is our
443 * job to build the TCP header, and pass the packet down to
444 * IP so it can do the same plus pass the packet off to the
447 * We are working here with either a clone of the original
448 * SKB, or a fresh unique copy made by the retransmit engine.
450 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask)
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
620 /* This routine just queue's the buffer
622 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
623 * otherwise socket can stall.
625 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
627 struct tcp_sock *tp = tcp_sk(sk);
629 /* Advance write_seq and place onto the write_queue. */
630 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
631 skb_header_release(skb);
632 tcp_add_write_queue_tail(sk, skb);
633 sk->sk_wmem_queued += skb->truesize;
634 sk_mem_charge(sk, skb->truesize);
637 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 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, unsigned int mss_now)
675 struct tcp_sock *tp = tcp_sk(sk);
676 struct sk_buff *buff;
677 int nsize, old_factor;
681 BUG_ON(len > skb->len);
683 tcp_clear_retrans_hints_partial(tp);
684 nsize = skb_headlen(skb) - len;
688 if (skb_cloned(skb) &&
689 skb_is_nonlinear(skb) &&
690 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
693 /* Get a new skb... force flag on. */
694 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
696 return -ENOMEM; /* We'll just try again later. */
698 sk->sk_wmem_queued += buff->truesize;
699 sk_mem_charge(sk, buff->truesize);
700 nlen = skb->len - len - nsize;
701 buff->truesize += nlen;
702 skb->truesize -= nlen;
704 /* Correct the sequence numbers. */
705 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
706 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
707 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
709 /* PSH and FIN should only be set in the second packet. */
710 flags = TCP_SKB_CB(skb)->flags;
711 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
712 TCP_SKB_CB(buff)->flags = flags;
713 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
714 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL;
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, skb_put(buff, nsize),
723 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
725 skb->ip_summed = CHECKSUM_PARTIAL;
726 skb_split(skb, buff, len);
729 buff->ip_summed = skb->ip_summed;
731 /* Looks stupid, but our code really uses when of
732 * skbs, which it never sent before. --ANK
734 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
735 buff->tstamp = skb->tstamp;
737 old_factor = tcp_skb_pcount(skb);
739 /* Fix up tso_factor for both original and new SKB. */
740 tcp_set_skb_tso_segs(sk, skb, mss_now);
741 tcp_set_skb_tso_segs(sk, buff, mss_now);
743 /* If this packet has been sent out already, we must
744 * adjust the various packet counters.
746 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
747 int diff = old_factor - tcp_skb_pcount(skb) -
748 tcp_skb_pcount(buff);
750 tp->packets_out -= diff;
752 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
753 tp->sacked_out -= diff;
754 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
755 tp->retrans_out -= diff;
757 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
758 tp->lost_out -= diff;
760 /* Adjust Reno SACK estimate. */
761 if (tcp_is_reno(tp) && diff > 0) {
762 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
763 tcp_verify_left_out(tp);
765 tcp_adjust_fackets_out(sk, skb, diff);
768 /* Link BUFF into the send queue. */
769 skb_header_release(buff);
770 tcp_insert_write_queue_after(skb, buff, sk);
775 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
776 * eventually). The difference is that pulled data not copied, but
777 * immediately discarded.
779 static void __pskb_trim_head(struct sk_buff *skb, int len)
785 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) {
786 if (skb_shinfo(skb)->frags[i].size <= eat) {
787 put_page(skb_shinfo(skb)->frags[i].page);
788 eat -= skb_shinfo(skb)->frags[i].size;
790 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
792 skb_shinfo(skb)->frags[k].page_offset += eat;
793 skb_shinfo(skb)->frags[k].size -= eat;
799 skb_shinfo(skb)->nr_frags = k;
801 skb_reset_tail_pointer(skb);
802 skb->data_len -= len;
803 skb->len = skb->data_len;
806 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
808 if (skb_cloned(skb) &&
809 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 /* This function synchronize snd mss to current pmtu/exthdr set.
893 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
894 for TCP options, but includes only bare TCP header.
896 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
897 It is minimum of user_mss and mss received with SYN.
898 It also does not include TCP options.
900 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
902 tp->mss_cache is current effective sending mss, including
903 all tcp options except for SACKs. It is evaluated,
904 taking into account current pmtu, but never exceeds
905 tp->rx_opt.mss_clamp.
907 NOTE1. rfc1122 clearly states that advertised MSS
908 DOES NOT include either tcp or ip options.
910 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
911 are READ ONLY outside this function. --ANK (980731)
914 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
916 struct tcp_sock *tp = tcp_sk(sk);
917 struct inet_connection_sock *icsk = inet_csk(sk);
920 if (icsk->icsk_mtup.search_high > pmtu)
921 icsk->icsk_mtup.search_high = pmtu;
923 mss_now = tcp_mtu_to_mss(sk, pmtu);
925 /* Bound mss with half of window */
926 if (tp->max_window && mss_now > (tp->max_window>>1))
927 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len);
929 /* And store cached results */
930 icsk->icsk_pmtu_cookie = pmtu;
931 if (icsk->icsk_mtup.enabled)
932 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
933 tp->mss_cache = mss_now;
938 /* Compute the current effective MSS, taking SACKs and IP options,
939 * and even PMTU discovery events into account.
941 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
942 * cannot be large. However, taking into account rare use of URG, this
945 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
947 struct tcp_sock *tp = tcp_sk(sk);
948 struct dst_entry *dst = __sk_dst_get(sk);
953 mss_now = tp->mss_cache;
955 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
959 u32 mtu = dst_mtu(dst);
960 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
961 mss_now = tcp_sync_mss(sk, mtu);
964 if (tp->rx_opt.eff_sacks)
965 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
966 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
968 #ifdef CONFIG_TCP_MD5SIG
969 if (tp->af_specific->md5_lookup(sk, sk))
970 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
973 xmit_size_goal = mss_now;
976 xmit_size_goal = (65535 -
977 inet_csk(sk)->icsk_af_ops->net_header_len -
978 inet_csk(sk)->icsk_ext_hdr_len -
981 if (tp->max_window &&
982 (xmit_size_goal > (tp->max_window >> 1)))
983 xmit_size_goal = max((tp->max_window >> 1),
984 68U - tp->tcp_header_len);
986 xmit_size_goal -= (xmit_size_goal % mss_now);
988 tp->xmit_size_goal = xmit_size_goal;
993 /* 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,
1024 struct tcp_sock *tp = tcp_sk(sk);
1025 u32 needed, window, cwnd_len;
1027 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1028 cwnd_len = mss_now * cwnd;
1030 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1033 if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len)
1036 needed = min(skb->len, window);
1037 return needed - needed % mss_now;
1040 /* Can at least one segment of SKB be sent right now, according to the
1041 * congestion window rules? If so, return how many segments are allowed.
1043 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 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, unsigned int mss_now)
1065 int tso_segs = tcp_skb_pcount(skb);
1069 tcp_skb_mss(skb) != mss_now)) {
1070 tcp_set_skb_tso_segs(sk, skb, mss_now);
1071 tso_segs = tcp_skb_pcount(skb);
1076 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1078 return after(tp->snd_sml,tp->snd_una) &&
1079 !after(tp->snd_sml, tp->snd_nxt);
1082 /* Return 0, if packet can be sent now without violation Nagle's rules:
1083 * 1. It is full sized.
1084 * 2. Or it contains FIN. (already checked by caller)
1085 * 3. Or TCP_NODELAY was set.
1086 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1087 * With Minshall's modification: all sent small packets are ACKed.
1090 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1091 const struct sk_buff *skb,
1092 unsigned mss_now, int nonagle)
1094 return (skb->len < mss_now &&
1095 ((nonagle&TCP_NAGLE_CORK) ||
1098 tcp_minshall_check(tp))));
1101 /* Return non-zero if the Nagle test allows this packet to be
1104 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1105 unsigned int cur_mss, int nonagle)
1107 /* Nagle rule does not apply to frames, which sit in the middle of the
1108 * write_queue (they have no chances to get new data).
1110 * This is implemented in the callers, where they modify the 'nonagle'
1111 * argument based upon the location of SKB in the send queue.
1113 if (nonagle & TCP_NAGLE_PUSH)
1116 /* Don't use the nagle rule for urgent data (or for the final FIN).
1117 * Nagle can be ignored during F-RTO too (see RFC4138).
1119 if (tp->urg_mode || (tp->frto_counter == 2) ||
1120 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1123 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1129 /* Does at least the first segment of SKB fit into the send window? */
1130 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss)
1132 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1134 if (skb->len > cur_mss)
1135 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1137 return !after(end_seq, tcp_wnd_end(tp));
1140 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1141 * should be put on the wire right now. If so, it returns the number of
1142 * packets allowed by the congestion window.
1144 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1145 unsigned int cur_mss, int nonagle)
1147 struct tcp_sock *tp = tcp_sk(sk);
1148 unsigned int cwnd_quota;
1150 tcp_init_tso_segs(sk, skb, cur_mss);
1152 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1155 cwnd_quota = tcp_cwnd_test(tp, skb);
1157 !tcp_snd_wnd_test(tp, skb, cur_mss))
1163 int tcp_may_send_now(struct sock *sk)
1165 struct tcp_sock *tp = tcp_sk(sk);
1166 struct sk_buff *skb = tcp_send_head(sk);
1169 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1170 (tcp_skb_is_last(sk, skb) ?
1171 tp->nonagle : TCP_NAGLE_PUSH)));
1174 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1175 * which is put after SKB on the list. It is very much like
1176 * tcp_fragment() except that it may make several kinds of assumptions
1177 * in order to speed up the splitting operation. In particular, we
1178 * know that all the data is in scatter-gather pages, and that the
1179 * packet has never been sent out before (and thus is not cloned).
1181 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now)
1183 struct sk_buff *buff;
1184 int nlen = skb->len - len;
1187 /* All of a TSO frame must be composed of paged data. */
1188 if (skb->len != skb->data_len)
1189 return tcp_fragment(sk, skb, len, mss_now);
1191 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1192 if (unlikely(buff == NULL))
1195 sk->sk_wmem_queued += buff->truesize;
1196 sk_mem_charge(sk, buff->truesize);
1197 buff->truesize += nlen;
1198 skb->truesize -= nlen;
1200 /* Correct the sequence numbers. */
1201 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1202 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1203 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1205 /* PSH and FIN should only be set in the second packet. */
1206 flags = TCP_SKB_CB(skb)->flags;
1207 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1208 TCP_SKB_CB(buff)->flags = flags;
1210 /* This packet was never sent out yet, so no SACK bits. */
1211 TCP_SKB_CB(buff)->sacked = 0;
1213 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1214 skb_split(skb, buff, len);
1216 /* Fix up tso_factor for both original and new SKB. */
1217 tcp_set_skb_tso_segs(sk, skb, mss_now);
1218 tcp_set_skb_tso_segs(sk, buff, mss_now);
1220 /* Link BUFF into the send queue. */
1221 skb_header_release(buff);
1222 tcp_insert_write_queue_after(skb, buff, sk);
1227 /* Try to defer sending, if possible, in order to minimize the amount
1228 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1230 * This algorithm is from John Heffner.
1232 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1234 struct tcp_sock *tp = tcp_sk(sk);
1235 const struct inet_connection_sock *icsk = inet_csk(sk);
1236 u32 send_win, cong_win, limit, in_flight;
1238 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1241 if (icsk->icsk_ca_state != TCP_CA_Open)
1244 /* Defer for less than two clock ticks. */
1245 if (tp->tso_deferred &&
1246 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1249 in_flight = tcp_packets_in_flight(tp);
1251 BUG_ON(tcp_skb_pcount(skb) <= 1 ||
1252 (tp->snd_cwnd <= in_flight));
1254 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1256 /* From in_flight test above, we know that cwnd > in_flight. */
1257 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1259 limit = min(send_win, cong_win);
1261 /* If a full-sized TSO skb can be sent, do it. */
1265 if (sysctl_tcp_tso_win_divisor) {
1266 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1268 /* If at least some fraction of a window is available,
1271 chunk /= sysctl_tcp_tso_win_divisor;
1275 /* Different approach, try not to defer past a single
1276 * ACK. Receiver should ACK every other full sized
1277 * frame, so if we have space for more than 3 frames
1280 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1284 /* Ok, it looks like it is advisable to defer. */
1285 tp->tso_deferred = 1 | (jiffies<<1);
1290 tp->tso_deferred = 0;
1294 /* Create a new MTU probe if we are ready.
1295 * Returns 0 if we should wait to probe (no cwnd available),
1296 * 1 if a probe was sent,
1298 static int tcp_mtu_probe(struct sock *sk)
1300 struct tcp_sock *tp = tcp_sk(sk);
1301 struct inet_connection_sock *icsk = inet_csk(sk);
1302 struct sk_buff *skb, *nskb, *next;
1309 /* Not currently probing/verifying,
1311 * have enough cwnd, and
1312 * not SACKing (the variable headers throw things off) */
1313 if (!icsk->icsk_mtup.enabled ||
1314 icsk->icsk_mtup.probe_size ||
1315 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1316 tp->snd_cwnd < 11 ||
1317 tp->rx_opt.eff_sacks)
1320 /* Very simple search strategy: just double the MSS. */
1321 mss_now = tcp_current_mss(sk, 0);
1322 probe_size = 2*tp->mss_cache;
1323 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1324 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1325 /* TODO: set timer for probe_converge_event */
1329 /* Have enough data in the send queue to probe? */
1330 if (tp->write_seq - tp->snd_nxt < size_needed)
1333 if (tp->snd_wnd < size_needed)
1335 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1338 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1339 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1340 if (!tcp_packets_in_flight(tp))
1346 /* We're allowed to probe. Build it now. */
1347 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1349 sk->sk_wmem_queued += nskb->truesize;
1350 sk_mem_charge(sk, nskb->truesize);
1352 skb = tcp_send_head(sk);
1354 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1355 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1356 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1357 TCP_SKB_CB(nskb)->sacked = 0;
1359 nskb->ip_summed = skb->ip_summed;
1361 tcp_insert_write_queue_before(nskb, skb, sk);
1364 tcp_for_write_queue_from_safe(skb, next, sk) {
1365 copy = min_t(int, skb->len, probe_size - len);
1366 if (nskb->ip_summed)
1367 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1369 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1370 skb_put(nskb, copy), copy, nskb->csum);
1372 if (skb->len <= copy) {
1373 /* We've eaten all the data from this skb.
1375 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1376 tcp_unlink_write_queue(skb, sk);
1377 sk_wmem_free_skb(sk, skb);
1379 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1380 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1381 if (!skb_shinfo(skb)->nr_frags) {
1382 skb_pull(skb, copy);
1383 if (skb->ip_summed != CHECKSUM_PARTIAL)
1384 skb->csum = csum_partial(skb->data, skb->len, 0);
1386 __pskb_trim_head(skb, copy);
1387 tcp_set_skb_tso_segs(sk, skb, mss_now);
1389 TCP_SKB_CB(skb)->seq += copy;
1394 if (len >= probe_size)
1397 tcp_init_tso_segs(sk, nskb, nskb->len);
1399 /* We're ready to send. If this fails, the probe will
1400 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1401 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1402 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1403 /* Decrement cwnd here because we are sending
1404 * effectively two packets. */
1406 tcp_event_new_data_sent(sk, nskb);
1408 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1409 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1410 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1419 /* This routine writes packets to the network. It advances the
1420 * send_head. This happens as incoming acks open up the remote
1423 * Returns 1, if no segments are in flight and we have queued segments, but
1424 * cannot send anything now because of SWS or another problem.
1426 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1428 struct tcp_sock *tp = tcp_sk(sk);
1429 struct sk_buff *skb;
1430 unsigned int tso_segs, sent_pkts;
1434 /* If we are closed, the bytes will have to remain here.
1435 * In time closedown will finish, we empty the write queue and all
1438 if (unlikely(sk->sk_state == TCP_CLOSE))
1443 /* Do MTU probing. */
1444 if ((result = tcp_mtu_probe(sk)) == 0) {
1446 } else if (result > 0) {
1450 while ((skb = tcp_send_head(sk))) {
1453 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1456 cwnd_quota = tcp_cwnd_test(tp, skb);
1460 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1463 if (tso_segs == 1) {
1464 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1465 (tcp_skb_is_last(sk, skb) ?
1466 nonagle : TCP_NAGLE_PUSH))))
1469 if (tcp_tso_should_defer(sk, skb))
1475 limit = tcp_mss_split_point(sk, skb, mss_now,
1478 if (skb->len > limit &&
1479 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1482 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1484 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1487 /* Advance the send_head. This one is sent out.
1488 * This call will increment packets_out.
1490 tcp_event_new_data_sent(sk, skb);
1492 tcp_minshall_update(tp, mss_now, skb);
1496 if (likely(sent_pkts)) {
1497 tcp_cwnd_validate(sk);
1500 return !tp->packets_out && tcp_send_head(sk);
1503 /* Push out any pending frames which were held back due to
1504 * TCP_CORK or attempt at coalescing tiny packets.
1505 * The socket must be locked by the caller.
1507 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1510 struct sk_buff *skb = tcp_send_head(sk);
1513 if (tcp_write_xmit(sk, cur_mss, nonagle))
1514 tcp_check_probe_timer(sk);
1518 /* Send _single_ skb sitting at the send head. This function requires
1519 * true push pending frames to setup probe timer etc.
1521 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1523 struct sk_buff *skb = tcp_send_head(sk);
1524 unsigned int tso_segs, cwnd_quota;
1526 BUG_ON(!skb || skb->len < mss_now);
1528 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1529 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1531 if (likely(cwnd_quota)) {
1538 limit = tcp_mss_split_point(sk, skb, mss_now,
1541 if (skb->len > limit &&
1542 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1545 /* Send it out now. */
1546 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1548 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1549 tcp_event_new_data_sent(sk, skb);
1550 tcp_cwnd_validate(sk);
1556 /* This function returns the amount that we can raise the
1557 * usable window based on the following constraints
1559 * 1. The window can never be shrunk once it is offered (RFC 793)
1560 * 2. We limit memory per socket
1563 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1564 * RECV.NEXT + RCV.WIN fixed until:
1565 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1567 * i.e. don't raise the right edge of the window until you can raise
1568 * it at least MSS bytes.
1570 * Unfortunately, the recommended algorithm breaks header prediction,
1571 * since header prediction assumes th->window stays fixed.
1573 * Strictly speaking, keeping th->window fixed violates the receiver
1574 * side SWS prevention criteria. The problem is that under this rule
1575 * a stream of single byte packets will cause the right side of the
1576 * window to always advance by a single byte.
1578 * Of course, if the sender implements sender side SWS prevention
1579 * then this will not be a problem.
1581 * BSD seems to make the following compromise:
1583 * If the free space is less than the 1/4 of the maximum
1584 * space available and the free space is less than 1/2 mss,
1585 * then set the window to 0.
1586 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1587 * Otherwise, just prevent the window from shrinking
1588 * and from being larger than the largest representable value.
1590 * This prevents incremental opening of the window in the regime
1591 * where TCP is limited by the speed of the reader side taking
1592 * data out of the TCP receive queue. It does nothing about
1593 * those cases where the window is constrained on the sender side
1594 * because the pipeline is full.
1596 * BSD also seems to "accidentally" limit itself to windows that are a
1597 * multiple of MSS, at least until the free space gets quite small.
1598 * This would appear to be a side effect of the mbuf implementation.
1599 * Combining these two algorithms results in the observed behavior
1600 * of having a fixed window size at almost all times.
1602 * Below we obtain similar behavior by forcing the offered window to
1603 * a multiple of the mss when it is feasible to do so.
1605 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1606 * Regular options like TIMESTAMP are taken into account.
1608 u32 __tcp_select_window(struct sock *sk)
1610 struct inet_connection_sock *icsk = inet_csk(sk);
1611 struct tcp_sock *tp = tcp_sk(sk);
1612 /* MSS for the peer's data. Previous versions used mss_clamp
1613 * here. I don't know if the value based on our guesses
1614 * of peer's MSS is better for the performance. It's more correct
1615 * but may be worse for the performance because of rcv_mss
1616 * fluctuations. --SAW 1998/11/1
1618 int mss = icsk->icsk_ack.rcv_mss;
1619 int free_space = tcp_space(sk);
1620 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1623 if (mss > full_space)
1626 if (free_space < (full_space >> 1)) {
1627 icsk->icsk_ack.quick = 0;
1629 if (tcp_memory_pressure)
1630 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss);
1632 if (free_space < mss)
1636 if (free_space > tp->rcv_ssthresh)
1637 free_space = tp->rcv_ssthresh;
1639 /* Don't do rounding if we are using window scaling, since the
1640 * scaled window will not line up with the MSS boundary anyway.
1642 window = tp->rcv_wnd;
1643 if (tp->rx_opt.rcv_wscale) {
1644 window = free_space;
1646 /* Advertise enough space so that it won't get scaled away.
1647 * Import case: prevent zero window announcement if
1648 * 1<<rcv_wscale > mss.
1650 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1651 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1652 << tp->rx_opt.rcv_wscale);
1654 /* Get the largest window that is a nice multiple of mss.
1655 * Window clamp already applied above.
1656 * If our current window offering is within 1 mss of the
1657 * free space we just keep it. This prevents the divide
1658 * and multiply from happening most of the time.
1659 * We also don't do any window rounding when the free space
1662 if (window <= free_space - mss || window > free_space)
1663 window = (free_space/mss)*mss;
1664 else if (mss == full_space &&
1665 free_space > window + (full_space >> 1))
1666 window = free_space;
1672 /* Attempt to collapse two adjacent SKB's during retransmission. */
1673 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now)
1675 struct tcp_sock *tp = tcp_sk(sk);
1676 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1678 /* The first test we must make is that neither of these two
1679 * SKB's are still referenced by someone else.
1681 if (!skb_cloned(skb) && !skb_cloned(next_skb)) {
1682 int skb_size = skb->len, next_skb_size = next_skb->len;
1683 u16 flags = TCP_SKB_CB(skb)->flags;
1685 /* Also punt if next skb has been SACK'd. */
1686 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1689 /* Next skb is out of window. */
1690 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1693 /* Punt if not enough space exists in the first SKB for
1694 * the data in the second, or the total combined payload
1695 * would exceed the MSS.
1697 if ((next_skb_size > skb_tailroom(skb)) ||
1698 ((skb_size + next_skb_size) > mss_now))
1701 BUG_ON(tcp_skb_pcount(skb) != 1 ||
1702 tcp_skb_pcount(next_skb) != 1);
1704 tcp_highest_sack_combine(sk, next_skb, skb);
1706 /* Ok. We will be able to collapse the packet. */
1707 tcp_unlink_write_queue(next_skb, sk);
1709 skb_copy_from_linear_data(next_skb,
1710 skb_put(skb, next_skb_size),
1713 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1714 skb->ip_summed = CHECKSUM_PARTIAL;
1716 if (skb->ip_summed != CHECKSUM_PARTIAL)
1717 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1719 /* Update sequence range on original skb. */
1720 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1722 /* Merge over control information. */
1723 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1724 TCP_SKB_CB(skb)->flags = flags;
1726 /* All done, get rid of second SKB and account for it so
1727 * packet counting does not break.
1729 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL);
1730 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS)
1731 tp->retrans_out -= tcp_skb_pcount(next_skb);
1732 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST)
1733 tp->lost_out -= tcp_skb_pcount(next_skb);
1734 /* Reno case is special. Sigh... */
1735 if (tcp_is_reno(tp) && tp->sacked_out)
1736 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1738 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1739 tp->packets_out -= tcp_skb_pcount(next_skb);
1741 /* changed transmit queue under us so clear hints */
1742 tcp_clear_retrans_hints_partial(tp);
1744 sk_wmem_free_skb(sk, next_skb);
1748 /* Do a simple retransmit without using the backoff mechanisms in
1749 * tcp_timer. This is used for path mtu discovery.
1750 * The socket is already locked here.
1752 void tcp_simple_retransmit(struct sock *sk)
1754 const struct inet_connection_sock *icsk = inet_csk(sk);
1755 struct tcp_sock *tp = tcp_sk(sk);
1756 struct sk_buff *skb;
1757 unsigned int mss = tcp_current_mss(sk, 0);
1760 tcp_for_write_queue(skb, sk) {
1761 if (skb == tcp_send_head(sk))
1763 if (skb->len > mss &&
1764 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) {
1765 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1766 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1767 tp->retrans_out -= tcp_skb_pcount(skb);
1769 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) {
1770 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1771 tp->lost_out += tcp_skb_pcount(skb);
1777 tcp_clear_all_retrans_hints(tp);
1782 tcp_verify_left_out(tp);
1784 /* Don't muck with the congestion window here.
1785 * Reason is that we do not increase amount of _data_
1786 * in network, but units changed and effective
1787 * cwnd/ssthresh really reduced now.
1789 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1790 tp->high_seq = tp->snd_nxt;
1791 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1792 tp->prior_ssthresh = 0;
1793 tp->undo_marker = 0;
1794 tcp_set_ca_state(sk, TCP_CA_Loss);
1796 tcp_xmit_retransmit_queue(sk);
1799 /* This retransmits one SKB. Policy decisions and retransmit queue
1800 * state updates are done by the caller. Returns non-zero if an
1801 * error occurred which prevented the send.
1803 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1805 struct tcp_sock *tp = tcp_sk(sk);
1806 struct inet_connection_sock *icsk = inet_csk(sk);
1807 unsigned int cur_mss = tcp_current_mss(sk, 0);
1810 /* Inconslusive MTU probe */
1811 if (icsk->icsk_mtup.probe_size) {
1812 icsk->icsk_mtup.probe_size = 0;
1815 /* Do not sent more than we queued. 1/4 is reserved for possible
1816 * copying overhead: fragmentation, tunneling, mangling etc.
1818 if (atomic_read(&sk->sk_wmem_alloc) >
1819 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1822 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1823 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1825 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1829 /* If receiver has shrunk his window, and skb is out of
1830 * new window, do not retransmit it. The exception is the
1831 * case, when window is shrunk to zero. In this case
1832 * our retransmit serves as a zero window probe.
1834 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1835 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1838 if (skb->len > cur_mss) {
1839 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1840 return -ENOMEM; /* We'll try again later. */
1843 /* Collapse two adjacent packets if worthwhile and we can. */
1844 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1845 (skb->len < (cur_mss >> 1)) &&
1846 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1847 (!tcp_skb_is_last(sk, skb)) &&
1848 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1849 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1850 (sysctl_tcp_retrans_collapse != 0))
1851 tcp_retrans_try_collapse(sk, skb, cur_mss);
1853 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1854 return -EHOSTUNREACH; /* Routing failure or similar. */
1856 /* Some Solaris stacks overoptimize and ignore the FIN on a
1857 * retransmit when old data is attached. So strip it off
1858 * since it is cheap to do so and saves bytes on the network.
1861 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1862 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1863 if (!pskb_trim(skb, 0)) {
1864 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1;
1865 skb_shinfo(skb)->gso_segs = 1;
1866 skb_shinfo(skb)->gso_size = 0;
1867 skb_shinfo(skb)->gso_type = 0;
1868 skb->ip_summed = CHECKSUM_NONE;
1873 /* Make a copy, if the first transmission SKB clone we made
1874 * is still in somebody's hands, else make a clone.
1876 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1878 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1881 /* Update global TCP statistics. */
1882 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1884 tp->total_retrans++;
1886 #if FASTRETRANS_DEBUG > 0
1887 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) {
1888 if (net_ratelimit())
1889 printk(KERN_DEBUG "retrans_out leaked.\n");
1892 if (!tp->retrans_out)
1893 tp->lost_retrans_low = tp->snd_nxt;
1894 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1895 tp->retrans_out += tcp_skb_pcount(skb);
1897 /* Save stamp of the first retransmit. */
1898 if (!tp->retrans_stamp)
1899 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1903 /* snd_nxt is stored to detect loss of retransmitted segment,
1904 * see tcp_input.c tcp_sacktag_write_queue().
1906 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1911 /* This gets called after a retransmit timeout, and the initially
1912 * retransmitted data is acknowledged. It tries to continue
1913 * resending the rest of the retransmit queue, until either
1914 * we've sent it all or the congestion window limit is reached.
1915 * If doing SACK, the first ACK which comes back for a timeout
1916 * based retransmit packet might feed us FACK information again.
1917 * If so, we use it to avoid unnecessarily retransmissions.
1919 void tcp_xmit_retransmit_queue(struct sock *sk)
1921 const struct inet_connection_sock *icsk = inet_csk(sk);
1922 struct tcp_sock *tp = tcp_sk(sk);
1923 struct sk_buff *skb;
1926 if (tp->retransmit_skb_hint) {
1927 skb = tp->retransmit_skb_hint;
1928 packet_cnt = tp->retransmit_cnt_hint;
1930 skb = tcp_write_queue_head(sk);
1934 /* First pass: retransmit lost packets. */
1936 tcp_for_write_queue_from(skb, sk) {
1937 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1939 if (skb == tcp_send_head(sk))
1941 /* we could do better than to assign each time */
1942 tp->retransmit_skb_hint = skb;
1943 tp->retransmit_cnt_hint = packet_cnt;
1945 /* Assume this retransmit will generate
1946 * only one packet for congestion window
1947 * calculation purposes. This works because
1948 * tcp_retransmit_skb() will chop up the
1949 * packet to be MSS sized and all the
1950 * packet counting works out.
1952 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1955 if (sacked & TCPCB_LOST) {
1956 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1957 if (tcp_retransmit_skb(sk, skb)) {
1958 tp->retransmit_skb_hint = NULL;
1961 if (icsk->icsk_ca_state != TCP_CA_Loss)
1962 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1964 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1966 if (skb == tcp_write_queue_head(sk))
1967 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1968 inet_csk(sk)->icsk_rto,
1972 packet_cnt += tcp_skb_pcount(skb);
1973 if (packet_cnt >= tp->lost_out)
1979 /* OK, demanded retransmission is finished. */
1981 /* Forward retransmissions are possible only during Recovery. */
1982 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1985 /* No forward retransmissions in Reno are possible. */
1986 if (tcp_is_reno(tp))
1989 /* Yeah, we have to make difficult choice between forward transmission
1990 * and retransmission... Both ways have their merits...
1992 * For now we do not retransmit anything, while we have some new
1993 * segments to send. In the other cases, follow rule 3 for
1994 * NextSeg() specified in RFC3517.
1997 if (tcp_may_send_now(sk))
2000 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2001 if (!tp->sacked_out)
2004 if (tp->forward_skb_hint)
2005 skb = tp->forward_skb_hint;
2007 skb = tcp_write_queue_head(sk);
2009 tcp_for_write_queue_from(skb, sk) {
2010 if (skb == tcp_send_head(sk))
2012 tp->forward_skb_hint = skb;
2014 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2017 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2020 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2023 /* Ok, retransmit it. */
2024 if (tcp_retransmit_skb(sk, skb)) {
2025 tp->forward_skb_hint = NULL;
2029 if (skb == tcp_write_queue_head(sk))
2030 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2031 inet_csk(sk)->icsk_rto,
2034 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2039 /* Send a fin. The caller locks the socket for us. This cannot be
2040 * allowed to fail queueing a FIN frame under any circumstances.
2042 void tcp_send_fin(struct sock *sk)
2044 struct tcp_sock *tp = tcp_sk(sk);
2045 struct sk_buff *skb = tcp_write_queue_tail(sk);
2048 /* Optimization, tack on the FIN if we have a queue of
2049 * unsent frames. But be careful about outgoing SACKS
2052 mss_now = tcp_current_mss(sk, 1);
2054 if (tcp_send_head(sk) != NULL) {
2055 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2056 TCP_SKB_CB(skb)->end_seq++;
2059 /* Socket is locked, keep trying until memory is available. */
2061 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2067 /* Reserve space for headers and prepare control bits. */
2068 skb_reserve(skb, MAX_TCP_HEADER);
2070 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2071 TCP_SKB_CB(skb)->sacked = 0;
2072 skb_shinfo(skb)->gso_segs = 1;
2073 skb_shinfo(skb)->gso_size = 0;
2074 skb_shinfo(skb)->gso_type = 0;
2076 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2077 TCP_SKB_CB(skb)->seq = tp->write_seq;
2078 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2079 tcp_queue_skb(sk, skb);
2081 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2084 /* We get here when a process closes a file descriptor (either due to
2085 * an explicit close() or as a byproduct of exit()'ing) and there
2086 * was unread data in the receive queue. This behavior is recommended
2087 * by RFC 2525, section 2.17. -DaveM
2089 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2091 struct sk_buff *skb;
2093 /* NOTE: No TCP options attached and we never retransmit this. */
2094 skb = alloc_skb(MAX_TCP_HEADER, priority);
2096 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2100 /* Reserve space for headers and prepare control bits. */
2101 skb_reserve(skb, MAX_TCP_HEADER);
2103 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2104 TCP_SKB_CB(skb)->sacked = 0;
2105 skb_shinfo(skb)->gso_segs = 1;
2106 skb_shinfo(skb)->gso_size = 0;
2107 skb_shinfo(skb)->gso_type = 0;
2110 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk);
2111 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2112 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2113 if (tcp_transmit_skb(sk, skb, 0, priority))
2114 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2117 /* WARNING: This routine must only be called when we have already sent
2118 * a SYN packet that crossed the incoming SYN that caused this routine
2119 * to get called. If this assumption fails then the initial rcv_wnd
2120 * and rcv_wscale values will not be correct.
2122 int tcp_send_synack(struct sock *sk)
2124 struct sk_buff* skb;
2126 skb = tcp_write_queue_head(sk);
2127 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) {
2128 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2131 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) {
2132 if (skb_cloned(skb)) {
2133 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2136 tcp_unlink_write_queue(skb, sk);
2137 skb_header_release(nskb);
2138 __tcp_add_write_queue_head(sk, nskb);
2139 sk_wmem_free_skb(sk, skb);
2140 sk->sk_wmem_queued += nskb->truesize;
2141 sk_mem_charge(sk, nskb->truesize);
2145 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2146 TCP_ECN_send_synack(tcp_sk(sk), skb);
2148 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2149 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2153 * Prepare a SYN-ACK.
2155 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2156 struct request_sock *req)
2158 struct inet_request_sock *ireq = inet_rsk(req);
2159 struct tcp_sock *tp = tcp_sk(sk);
2161 int tcp_header_size;
2162 struct sk_buff *skb;
2163 #ifdef CONFIG_TCP_MD5SIG
2164 struct tcp_md5sig_key *md5;
2165 __u8 *md5_hash_location;
2168 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2172 /* Reserve space for headers. */
2173 skb_reserve(skb, MAX_TCP_HEADER);
2175 skb->dst = dst_clone(dst);
2177 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2178 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2179 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2180 /* SACK_PERM is in the place of NOP NOP of TS */
2181 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2183 #ifdef CONFIG_TCP_MD5SIG
2184 /* Are we doing MD5 on this segment? If so - make room for it */
2185 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2187 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2189 skb_push(skb, tcp_header_size);
2190 skb_reset_transport_header(skb);
2193 memset(th, 0, sizeof(struct tcphdr));
2196 TCP_ECN_make_synack(req, th);
2197 th->source = inet_sk(sk)->sport;
2198 th->dest = ireq->rmt_port;
2199 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn;
2200 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1;
2201 TCP_SKB_CB(skb)->sacked = 0;
2202 skb_shinfo(skb)->gso_segs = 1;
2203 skb_shinfo(skb)->gso_size = 0;
2204 skb_shinfo(skb)->gso_type = 0;
2205 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2206 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2207 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2209 /* Set this up on the first call only */
2210 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2211 /* tcp_full_space because it is guaranteed to be the first packet */
2212 tcp_select_initial_window(tcp_full_space(sk),
2213 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2218 ireq->rcv_wscale = rcv_wscale;
2221 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2222 th->window = htons(min(req->rcv_wnd, 65535U));
2224 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2225 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2226 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2227 TCP_SKB_CB(skb)->when,
2230 #ifdef CONFIG_TCP_MD5SIG
2231 md5 ? &md5_hash_location :
2237 th->doff = (tcp_header_size >> 2);
2238 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2240 #ifdef CONFIG_TCP_MD5SIG
2241 /* Okay, we have all we need - do the md5 hash if needed */
2243 tp->af_specific->calc_md5_hash(md5_hash_location,
2246 tcp_hdr(skb), sk->sk_protocol,
2255 * Do all connect socket setups that can be done AF independent.
2257 static void tcp_connect_init(struct sock *sk)
2259 struct dst_entry *dst = __sk_dst_get(sk);
2260 struct tcp_sock *tp = tcp_sk(sk);
2263 /* We'll fix this up when we get a response from the other end.
2264 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2266 tp->tcp_header_len = sizeof(struct tcphdr) +
2267 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2269 #ifdef CONFIG_TCP_MD5SIG
2270 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2271 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2274 /* If user gave his TCP_MAXSEG, record it to clamp */
2275 if (tp->rx_opt.user_mss)
2276 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2279 tcp_sync_mss(sk, dst_mtu(dst));
2281 if (!tp->window_clamp)
2282 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2283 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2284 tcp_initialize_rcv_mss(sk);
2286 tcp_select_initial_window(tcp_full_space(sk),
2287 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2290 sysctl_tcp_window_scaling,
2293 tp->rx_opt.rcv_wscale = rcv_wscale;
2294 tp->rcv_ssthresh = tp->rcv_wnd;
2297 sock_reset_flag(sk, SOCK_DONE);
2299 tcp_init_wl(tp, tp->write_seq, 0);
2300 tp->snd_una = tp->write_seq;
2301 tp->snd_sml = tp->write_seq;
2306 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2307 inet_csk(sk)->icsk_retransmits = 0;
2308 tcp_clear_retrans(tp);
2312 * Build a SYN and send it off.
2314 int tcp_connect(struct sock *sk)
2316 struct tcp_sock *tp = tcp_sk(sk);
2317 struct sk_buff *buff;
2319 tcp_connect_init(sk);
2321 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2322 if (unlikely(buff == NULL))
2325 /* Reserve space for headers. */
2326 skb_reserve(buff, MAX_TCP_HEADER);
2328 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN;
2329 TCP_ECN_send_syn(sk, buff);
2330 TCP_SKB_CB(buff)->sacked = 0;
2331 skb_shinfo(buff)->gso_segs = 1;
2332 skb_shinfo(buff)->gso_size = 0;
2333 skb_shinfo(buff)->gso_type = 0;
2335 tp->snd_nxt = tp->write_seq;
2336 TCP_SKB_CB(buff)->seq = tp->write_seq++;
2337 TCP_SKB_CB(buff)->end_seq = tp->write_seq;
2340 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2341 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2342 skb_header_release(buff);
2343 __tcp_add_write_queue_tail(sk, buff);
2344 sk->sk_wmem_queued += buff->truesize;
2345 sk_mem_charge(sk, buff->truesize);
2346 tp->packets_out += tcp_skb_pcount(buff);
2347 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2349 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2350 * in order to make this packet get counted in tcpOutSegs.
2352 tp->snd_nxt = tp->write_seq;
2353 tp->pushed_seq = tp->write_seq;
2354 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2356 /* Timer for repeating the SYN until an answer. */
2357 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2358 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2362 /* Send out a delayed ack, the caller does the policy checking
2363 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2366 void tcp_send_delayed_ack(struct sock *sk)
2368 struct inet_connection_sock *icsk = inet_csk(sk);
2369 int ato = icsk->icsk_ack.ato;
2370 unsigned long timeout;
2372 if (ato > TCP_DELACK_MIN) {
2373 const struct tcp_sock *tp = tcp_sk(sk);
2376 if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2377 max_ato = TCP_DELACK_MAX;
2379 /* Slow path, intersegment interval is "high". */
2381 /* If some rtt estimate is known, use it to bound delayed ack.
2382 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2386 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN);
2392 ato = min(ato, max_ato);
2395 /* Stay within the limit we were given */
2396 timeout = jiffies + ato;
2398 /* Use new timeout only if there wasn't a older one earlier. */
2399 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2400 /* If delack timer was blocked or is about to expire,
2403 if (icsk->icsk_ack.blocked ||
2404 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2409 if (!time_before(timeout, icsk->icsk_ack.timeout))
2410 timeout = icsk->icsk_ack.timeout;
2412 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2413 icsk->icsk_ack.timeout = timeout;
2414 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2417 /* This routine sends an ack and also updates the window. */
2418 void tcp_send_ack(struct sock *sk)
2420 /* If we have been reset, we may not send again. */
2421 if (sk->sk_state != TCP_CLOSE) {
2422 struct sk_buff *buff;
2424 /* We are not putting this on the write queue, so
2425 * tcp_transmit_skb() will set the ownership to this
2428 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2430 inet_csk_schedule_ack(sk);
2431 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2432 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2433 TCP_DELACK_MAX, TCP_RTO_MAX);
2437 /* Reserve space for headers and prepare control bits. */
2438 skb_reserve(buff, MAX_TCP_HEADER);
2440 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK;
2441 TCP_SKB_CB(buff)->sacked = 0;
2442 skb_shinfo(buff)->gso_segs = 1;
2443 skb_shinfo(buff)->gso_size = 0;
2444 skb_shinfo(buff)->gso_type = 0;
2446 /* Send it off, this clears delayed acks for us. */
2447 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk);
2448 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2449 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2453 /* This routine sends a packet with an out of date sequence
2454 * number. It assumes the other end will try to ack it.
2456 * Question: what should we make while urgent mode?
2457 * 4.4BSD forces sending single byte of data. We cannot send
2458 * out of window data, because we have SND.NXT==SND.MAX...
2460 * Current solution: to send TWO zero-length segments in urgent mode:
2461 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2462 * out-of-date with SND.UNA-1 to probe window.
2464 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2466 struct tcp_sock *tp = tcp_sk(sk);
2467 struct sk_buff *skb;
2469 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2470 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2474 /* Reserve space for headers and set control bits. */
2475 skb_reserve(skb, MAX_TCP_HEADER);
2477 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK;
2478 TCP_SKB_CB(skb)->sacked = urgent;
2479 skb_shinfo(skb)->gso_segs = 1;
2480 skb_shinfo(skb)->gso_size = 0;
2481 skb_shinfo(skb)->gso_type = 0;
2483 /* Use a previous sequence. This should cause the other
2484 * end to send an ack. Don't queue or clone SKB, just
2487 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1;
2488 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq;
2489 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2490 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2493 int tcp_write_wakeup(struct sock *sk)
2495 if (sk->sk_state != TCP_CLOSE) {
2496 struct tcp_sock *tp = tcp_sk(sk);
2497 struct sk_buff *skb;
2499 if ((skb = tcp_send_head(sk)) != NULL &&
2500 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2502 unsigned int mss = tcp_current_mss(sk, 0);
2503 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2505 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2506 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2508 /* We are probing the opening of a window
2509 * but the window size is != 0
2510 * must have been a result SWS avoidance ( sender )
2512 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2514 seg_size = min(seg_size, mss);
2515 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2516 if (tcp_fragment(sk, skb, seg_size, mss))
2518 } else if (!tcp_skb_pcount(skb))
2519 tcp_set_skb_tso_segs(sk, skb, mss);
2521 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2522 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2523 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2525 tcp_event_new_data_sent(sk, skb);
2529 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF))
2530 tcp_xmit_probe_skb(sk, TCPCB_URG);
2531 return tcp_xmit_probe_skb(sk, 0);
2537 /* A window probe timeout has occurred. If window is not closed send
2538 * a partial packet else a zero probe.
2540 void tcp_send_probe0(struct sock *sk)
2542 struct inet_connection_sock *icsk = inet_csk(sk);
2543 struct tcp_sock *tp = tcp_sk(sk);
2546 err = tcp_write_wakeup(sk);
2548 if (tp->packets_out || !tcp_send_head(sk)) {
2549 /* Cancel probe timer, if it is not required. */
2550 icsk->icsk_probes_out = 0;
2551 icsk->icsk_backoff = 0;
2556 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2557 icsk->icsk_backoff++;
2558 icsk->icsk_probes_out++;
2559 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2560 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2563 /* If packet was not sent due to local congestion,
2564 * do not backoff and do not remember icsk_probes_out.
2565 * Let local senders to fight for local resources.
2567 * Use accumulated backoff yet.
2569 if (!icsk->icsk_probes_out)
2570 icsk->icsk_probes_out = 1;
2571 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2572 min(icsk->icsk_rto << icsk->icsk_backoff,
2573 TCP_RESOURCE_PROBE_INTERVAL),
2578 EXPORT_SYMBOL(tcp_connect);
2579 EXPORT_SYMBOL(tcp_make_synack);
2580 EXPORT_SYMBOL(tcp_simple_retransmit);
2581 EXPORT_SYMBOL(tcp_sync_mss);
2582 EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor);
2583 EXPORT_SYMBOL(tcp_mtup_init);