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.
258 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
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 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
331 * auto increment end seqno.
333 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
337 TCP_SKB_CB(skb)->flags = flags;
338 TCP_SKB_CB(skb)->sacked = 0;
340 skb_shinfo(skb)->gso_segs = 1;
341 skb_shinfo(skb)->gso_size = 0;
342 skb_shinfo(skb)->gso_type = 0;
344 TCP_SKB_CB(skb)->seq = seq;
345 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
347 TCP_SKB_CB(skb)->end_seq = seq;
350 static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp,
351 __u32 tstamp, __u8 **md5_hash)
353 if (tp->rx_opt.tstamp_ok) {
354 *ptr++ = htonl((TCPOPT_NOP << 24) |
356 (TCPOPT_TIMESTAMP << 8) |
358 *ptr++ = htonl(tstamp);
359 *ptr++ = htonl(tp->rx_opt.ts_recent);
361 if (tp->rx_opt.eff_sacks) {
362 struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks;
365 *ptr++ = htonl((TCPOPT_NOP << 24) |
368 (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks *
369 TCPOLEN_SACK_PERBLOCK)));
371 for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) {
372 *ptr++ = htonl(sp[this_sack].start_seq);
373 *ptr++ = htonl(sp[this_sack].end_seq);
376 if (tp->rx_opt.dsack) {
377 tp->rx_opt.dsack = 0;
378 tp->rx_opt.eff_sacks--;
381 #ifdef CONFIG_TCP_MD5SIG
383 *ptr++ = htonl((TCPOPT_NOP << 24) |
385 (TCPOPT_MD5SIG << 8) |
387 *md5_hash = (__u8 *)ptr;
392 /* Construct a tcp options header for a SYN or SYN_ACK packet.
393 * If this is every changed make sure to change the definition of
394 * MAX_SYN_SIZE to match the new maximum number of options that you
397 * Note - that with the RFC2385 TCP option, we make room for the
398 * 16 byte MD5 hash. This will be filled in later, so the pointer for the
399 * location to be filled is passed back up.
401 static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack,
402 int offer_wscale, int wscale, __u32 tstamp,
403 __u32 ts_recent, __u8 **md5_hash)
405 /* We always get an MSS option.
406 * The option bytes which will be seen in normal data
407 * packets should timestamps be used, must be in the MSS
408 * advertised. But we subtract them from tp->mss_cache so
409 * that calculations in tcp_sendmsg are simpler etc.
410 * So account for this fact here if necessary. If we
411 * don't do this correctly, as a receiver we won't
412 * recognize data packets as being full sized when we
413 * should, and thus we won't abide by the delayed ACK
415 * SACKs don't matter, we never delay an ACK when we
416 * have any of those going out.
418 *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss);
421 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
422 (TCPOLEN_SACK_PERM << 16) |
423 (TCPOPT_TIMESTAMP << 8) |
426 *ptr++ = htonl((TCPOPT_NOP << 24) |
428 (TCPOPT_TIMESTAMP << 8) |
430 *ptr++ = htonl(tstamp); /* TSVAL */
431 *ptr++ = htonl(ts_recent); /* TSECR */
433 *ptr++ = htonl((TCPOPT_NOP << 24) |
435 (TCPOPT_SACK_PERM << 8) |
438 *ptr++ = htonl((TCPOPT_NOP << 24) |
439 (TCPOPT_WINDOW << 16) |
440 (TCPOLEN_WINDOW << 8) |
442 #ifdef CONFIG_TCP_MD5SIG
444 * If MD5 is enabled, then we set the option, and include the size
445 * (always 18). The actual MD5 hash is added just before the
449 *ptr++ = htonl((TCPOPT_NOP << 24) |
451 (TCPOPT_MD5SIG << 8) |
453 *md5_hash = (__u8 *)ptr;
458 /* This routine actually transmits TCP packets queued in by
459 * tcp_do_sendmsg(). This is used by both the initial
460 * transmission and possible later retransmissions.
461 * All SKB's seen here are completely headerless. It is our
462 * job to build the TCP header, and pass the packet down to
463 * IP so it can do the same plus pass the packet off to the
466 * We are working here with either a clone of the original
467 * SKB, or a fresh unique copy made by the retransmit engine.
469 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
472 const struct inet_connection_sock *icsk = inet_csk(sk);
473 struct inet_sock *inet;
475 struct tcp_skb_cb *tcb;
477 #ifdef CONFIG_TCP_MD5SIG
478 struct tcp_md5sig_key *md5;
479 __u8 *md5_hash_location;
485 BUG_ON(!skb || !tcp_skb_pcount(skb));
487 /* If congestion control is doing timestamping, we must
488 * take such a timestamp before we potentially clone/copy.
490 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
491 __net_timestamp(skb);
493 if (likely(clone_it)) {
494 if (unlikely(skb_cloned(skb)))
495 skb = pskb_copy(skb, gfp_mask);
497 skb = skb_clone(skb, gfp_mask);
504 tcb = TCP_SKB_CB(skb);
505 tcp_header_size = tp->tcp_header_len;
507 #define SYSCTL_FLAG_TSTAMPS 0x1
508 #define SYSCTL_FLAG_WSCALE 0x2
509 #define SYSCTL_FLAG_SACK 0x4
512 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
513 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS;
514 if (sysctl_tcp_timestamps) {
515 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED;
516 sysctl_flags |= SYSCTL_FLAG_TSTAMPS;
518 if (sysctl_tcp_window_scaling) {
519 tcp_header_size += TCPOLEN_WSCALE_ALIGNED;
520 sysctl_flags |= SYSCTL_FLAG_WSCALE;
522 if (sysctl_tcp_sack) {
523 sysctl_flags |= SYSCTL_FLAG_SACK;
524 if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS))
525 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED;
527 } else if (unlikely(tp->rx_opt.eff_sacks)) {
528 /* A SACK is 2 pad bytes, a 2 byte header, plus
529 * 2 32-bit sequence numbers for each SACK block.
531 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED +
532 (tp->rx_opt.eff_sacks *
533 TCPOLEN_SACK_PERBLOCK));
536 if (tcp_packets_in_flight(tp) == 0)
537 tcp_ca_event(sk, CA_EVENT_TX_START);
539 #ifdef CONFIG_TCP_MD5SIG
541 * Are we doing MD5 on this segment? If so - make
544 md5 = tp->af_specific->md5_lookup(sk, sk);
546 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
549 skb_push(skb, tcp_header_size);
550 skb_reset_transport_header(skb);
551 skb_set_owner_w(skb, sk);
553 /* Build TCP header and checksum it. */
555 th->source = inet->sport;
556 th->dest = inet->dport;
557 th->seq = htonl(tcb->seq);
558 th->ack_seq = htonl(tp->rcv_nxt);
559 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
562 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
563 /* RFC1323: The window in SYN & SYN/ACK segments
566 th->window = htons(min(tp->rcv_wnd, 65535U));
568 th->window = htons(tcp_select_window(sk));
573 if (unlikely(tp->urg_mode &&
574 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
575 th->urg_ptr = htons(tp->snd_up - tcb->seq);
579 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
580 tcp_syn_build_options((__be32 *)(th + 1),
581 tcp_advertise_mss(sk),
582 (sysctl_flags & SYSCTL_FLAG_TSTAMPS),
583 (sysctl_flags & SYSCTL_FLAG_SACK),
584 (sysctl_flags & SYSCTL_FLAG_WSCALE),
585 tp->rx_opt.rcv_wscale,
587 tp->rx_opt.ts_recent,
589 #ifdef CONFIG_TCP_MD5SIG
590 md5 ? &md5_hash_location :
594 tcp_build_and_update_options((__be32 *)(th + 1),
596 #ifdef CONFIG_TCP_MD5SIG
597 md5 ? &md5_hash_location :
600 TCP_ECN_send(sk, skb, tcp_header_size);
603 #ifdef CONFIG_TCP_MD5SIG
604 /* Calculate the MD5 hash, as we have all we need now */
606 tp->af_specific->calc_md5_hash(md5_hash_location,
614 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
616 if (likely(tcb->flags & TCPCB_FLAG_ACK))
617 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
619 if (skb->len != tcp_header_size)
620 tcp_event_data_sent(tp, skb, sk);
622 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
623 TCP_INC_STATS(TCP_MIB_OUTSEGS);
625 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
626 if (likely(err <= 0))
629 tcp_enter_cwr(sk, 1);
631 return net_xmit_eval(err);
633 #undef SYSCTL_FLAG_TSTAMPS
634 #undef SYSCTL_FLAG_WSCALE
635 #undef SYSCTL_FLAG_SACK
638 /* This routine just queue's the buffer
640 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
641 * otherwise socket can stall.
643 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
645 struct tcp_sock *tp = tcp_sk(sk);
647 /* Advance write_seq and place onto the write_queue. */
648 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
649 skb_header_release(skb);
650 tcp_add_write_queue_tail(sk, skb);
651 sk->sk_wmem_queued += skb->truesize;
652 sk_mem_charge(sk, skb->truesize);
655 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
656 unsigned int mss_now)
658 if (skb->len <= mss_now || !sk_can_gso(sk)) {
659 /* Avoid the costly divide in the normal
662 skb_shinfo(skb)->gso_segs = 1;
663 skb_shinfo(skb)->gso_size = 0;
664 skb_shinfo(skb)->gso_type = 0;
666 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
667 skb_shinfo(skb)->gso_size = mss_now;
668 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
672 /* When a modification to fackets out becomes necessary, we need to check
673 * skb is counted to fackets_out or not.
675 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
678 struct tcp_sock *tp = tcp_sk(sk);
680 if (!tp->sacked_out || tcp_is_reno(tp))
683 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
684 tp->fackets_out -= decr;
687 /* Function to create two new TCP segments. Shrinks the given segment
688 * to the specified size and appends a new segment with the rest of the
689 * packet to the list. This won't be called frequently, I hope.
690 * Remember, these are still headerless SKBs at this point.
692 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
693 unsigned int mss_now)
695 struct tcp_sock *tp = tcp_sk(sk);
696 struct sk_buff *buff;
697 int nsize, old_factor;
701 BUG_ON(len > skb->len);
703 tcp_clear_retrans_hints_partial(tp);
704 nsize = skb_headlen(skb) - len;
708 if (skb_cloned(skb) &&
709 skb_is_nonlinear(skb) &&
710 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
713 /* Get a new skb... force flag on. */
714 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
716 return -ENOMEM; /* We'll just try again later. */
718 sk->sk_wmem_queued += buff->truesize;
719 sk_mem_charge(sk, buff->truesize);
720 nlen = skb->len - len - nsize;
721 buff->truesize += nlen;
722 skb->truesize -= nlen;
724 /* Correct the sequence numbers. */
725 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
726 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
727 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
729 /* PSH and FIN should only be set in the second packet. */
730 flags = TCP_SKB_CB(skb)->flags;
731 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
732 TCP_SKB_CB(buff)->flags = flags;
733 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
735 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
736 /* Copy and checksum data tail into the new buffer. */
737 buff->csum = csum_partial_copy_nocheck(skb->data + len,
738 skb_put(buff, nsize),
743 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
745 skb->ip_summed = CHECKSUM_PARTIAL;
746 skb_split(skb, buff, len);
749 buff->ip_summed = skb->ip_summed;
751 /* Looks stupid, but our code really uses when of
752 * skbs, which it never sent before. --ANK
754 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
755 buff->tstamp = skb->tstamp;
757 old_factor = tcp_skb_pcount(skb);
759 /* Fix up tso_factor for both original and new SKB. */
760 tcp_set_skb_tso_segs(sk, skb, mss_now);
761 tcp_set_skb_tso_segs(sk, buff, mss_now);
763 /* If this packet has been sent out already, we must
764 * adjust the various packet counters.
766 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
767 int diff = old_factor - tcp_skb_pcount(skb) -
768 tcp_skb_pcount(buff);
770 tp->packets_out -= diff;
772 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
773 tp->sacked_out -= diff;
774 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
775 tp->retrans_out -= diff;
777 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
778 tp->lost_out -= diff;
780 /* Adjust Reno SACK estimate. */
781 if (tcp_is_reno(tp) && diff > 0) {
782 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
783 tcp_verify_left_out(tp);
785 tcp_adjust_fackets_out(sk, skb, diff);
788 /* Link BUFF into the send queue. */
789 skb_header_release(buff);
790 tcp_insert_write_queue_after(skb, buff, sk);
795 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
796 * eventually). The difference is that pulled data not copied, but
797 * immediately discarded.
799 static void __pskb_trim_head(struct sk_buff *skb, int len)
805 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
806 if (skb_shinfo(skb)->frags[i].size <= eat) {
807 put_page(skb_shinfo(skb)->frags[i].page);
808 eat -= skb_shinfo(skb)->frags[i].size;
810 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
812 skb_shinfo(skb)->frags[k].page_offset += eat;
813 skb_shinfo(skb)->frags[k].size -= eat;
819 skb_shinfo(skb)->nr_frags = k;
821 skb_reset_tail_pointer(skb);
822 skb->data_len -= len;
823 skb->len = skb->data_len;
826 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
828 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
831 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
832 if (unlikely(len < skb_headlen(skb)))
833 __skb_pull(skb, len);
835 __pskb_trim_head(skb, len - skb_headlen(skb));
837 TCP_SKB_CB(skb)->seq += len;
838 skb->ip_summed = CHECKSUM_PARTIAL;
840 skb->truesize -= len;
841 sk->sk_wmem_queued -= len;
842 sk_mem_uncharge(sk, len);
843 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
845 /* Any change of skb->len requires recalculation of tso
848 if (tcp_skb_pcount(skb) > 1)
849 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
854 /* Not accounting for SACKs here. */
855 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
857 struct tcp_sock *tp = tcp_sk(sk);
858 struct inet_connection_sock *icsk = inet_csk(sk);
861 /* Calculate base mss without TCP options:
862 It is MMS_S - sizeof(tcphdr) of rfc1122
864 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
866 /* Clamp it (mss_clamp does not include tcp options) */
867 if (mss_now > tp->rx_opt.mss_clamp)
868 mss_now = tp->rx_opt.mss_clamp;
870 /* Now subtract optional transport overhead */
871 mss_now -= icsk->icsk_ext_hdr_len;
873 /* Then reserve room for full set of TCP options and 8 bytes of data */
877 /* Now subtract TCP options size, not including SACKs */
878 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
883 /* Inverse of above */
884 int tcp_mss_to_mtu(struct sock *sk, int mss)
886 struct tcp_sock *tp = tcp_sk(sk);
887 struct inet_connection_sock *icsk = inet_csk(sk);
892 icsk->icsk_ext_hdr_len +
893 icsk->icsk_af_ops->net_header_len;
898 void tcp_mtup_init(struct sock *sk)
900 struct tcp_sock *tp = tcp_sk(sk);
901 struct inet_connection_sock *icsk = inet_csk(sk);
903 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
904 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
905 icsk->icsk_af_ops->net_header_len;
906 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
907 icsk->icsk_mtup.probe_size = 0;
910 /* Bound MSS / TSO packet size with the half of the window */
911 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
913 if (tp->max_window && pktsize > (tp->max_window >> 1))
914 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
919 /* This function synchronize snd mss to current pmtu/exthdr set.
921 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
922 for TCP options, but includes only bare TCP header.
924 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
925 It is minimum of user_mss and mss received with SYN.
926 It also does not include TCP options.
928 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
930 tp->mss_cache is current effective sending mss, including
931 all tcp options except for SACKs. It is evaluated,
932 taking into account current pmtu, but never exceeds
933 tp->rx_opt.mss_clamp.
935 NOTE1. rfc1122 clearly states that advertised MSS
936 DOES NOT include either tcp or ip options.
938 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
939 are READ ONLY outside this function. --ANK (980731)
941 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
943 struct tcp_sock *tp = tcp_sk(sk);
944 struct inet_connection_sock *icsk = inet_csk(sk);
947 if (icsk->icsk_mtup.search_high > pmtu)
948 icsk->icsk_mtup.search_high = pmtu;
950 mss_now = tcp_mtu_to_mss(sk, pmtu);
951 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
953 /* And store cached results */
954 icsk->icsk_pmtu_cookie = pmtu;
955 if (icsk->icsk_mtup.enabled)
956 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
957 tp->mss_cache = mss_now;
962 /* Compute the current effective MSS, taking SACKs and IP options,
963 * and even PMTU discovery events into account.
965 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
966 * cannot be large. However, taking into account rare use of URG, this
969 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
971 struct tcp_sock *tp = tcp_sk(sk);
972 struct dst_entry *dst = __sk_dst_get(sk);
977 mss_now = tp->mss_cache;
979 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
983 u32 mtu = dst_mtu(dst);
984 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
985 mss_now = tcp_sync_mss(sk, mtu);
988 if (tp->rx_opt.eff_sacks)
989 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
990 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
992 #ifdef CONFIG_TCP_MD5SIG
993 if (tp->af_specific->md5_lookup(sk, sk))
994 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
997 xmit_size_goal = mss_now;
1000 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1001 inet_csk(sk)->icsk_af_ops->net_header_len -
1002 inet_csk(sk)->icsk_ext_hdr_len -
1003 tp->tcp_header_len);
1005 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1006 xmit_size_goal -= (xmit_size_goal % mss_now);
1008 tp->xmit_size_goal = xmit_size_goal;
1013 /* Congestion window validation. (RFC2861) */
1014 static void tcp_cwnd_validate(struct sock *sk)
1016 struct tcp_sock *tp = tcp_sk(sk);
1018 if (tp->packets_out >= tp->snd_cwnd) {
1019 /* Network is feed fully. */
1020 tp->snd_cwnd_used = 0;
1021 tp->snd_cwnd_stamp = tcp_time_stamp;
1023 /* Network starves. */
1024 if (tp->packets_out > tp->snd_cwnd_used)
1025 tp->snd_cwnd_used = tp->packets_out;
1027 if (sysctl_tcp_slow_start_after_idle &&
1028 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1029 tcp_cwnd_application_limited(sk);
1033 /* Returns the portion of skb which can be sent right away without
1034 * introducing MSS oddities to segment boundaries. In rare cases where
1035 * mss_now != mss_cache, we will request caller to create a small skb
1036 * per input skb which could be mostly avoided here (if desired).
1038 * We explicitly want to create a request for splitting write queue tail
1039 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1040 * thus all the complexity (cwnd_len is always MSS multiple which we
1041 * return whenever allowed by the other factors). Basically we need the
1042 * modulo only when the receiver window alone is the limiting factor or
1043 * when we would be allowed to send the split-due-to-Nagle skb fully.
1045 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1046 unsigned int mss_now, unsigned int cwnd)
1048 struct tcp_sock *tp = tcp_sk(sk);
1049 u32 needed, window, cwnd_len;
1051 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1052 cwnd_len = mss_now * cwnd;
1054 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1057 needed = min(skb->len, window);
1059 if (cwnd_len <= needed)
1062 return needed - needed % mss_now;
1065 /* Can at least one segment of SKB be sent right now, according to the
1066 * congestion window rules? If so, return how many segments are allowed.
1068 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1069 struct sk_buff *skb)
1071 u32 in_flight, cwnd;
1073 /* Don't be strict about the congestion window for the final FIN. */
1074 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1075 tcp_skb_pcount(skb) == 1)
1078 in_flight = tcp_packets_in_flight(tp);
1079 cwnd = tp->snd_cwnd;
1080 if (in_flight < cwnd)
1081 return (cwnd - in_flight);
1086 /* This must be invoked the first time we consider transmitting
1087 * SKB onto the wire.
1089 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1090 unsigned int mss_now)
1092 int tso_segs = tcp_skb_pcount(skb);
1094 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1095 tcp_set_skb_tso_segs(sk, skb, mss_now);
1096 tso_segs = tcp_skb_pcount(skb);
1101 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1103 return after(tp->snd_sml,tp->snd_una) &&
1104 !after(tp->snd_sml, tp->snd_nxt);
1107 /* Return 0, if packet can be sent now without violation Nagle's rules:
1108 * 1. It is full sized.
1109 * 2. Or it contains FIN. (already checked by caller)
1110 * 3. Or TCP_NODELAY was set.
1111 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1112 * With Minshall's modification: all sent small packets are ACKed.
1114 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1115 const struct sk_buff *skb,
1116 unsigned mss_now, int nonagle)
1118 return (skb->len < mss_now &&
1119 ((nonagle & TCP_NAGLE_CORK) ||
1120 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1123 /* Return non-zero if the Nagle test allows this packet to be
1126 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1127 unsigned int cur_mss, int nonagle)
1129 /* Nagle rule does not apply to frames, which sit in the middle of the
1130 * write_queue (they have no chances to get new data).
1132 * This is implemented in the callers, where they modify the 'nonagle'
1133 * argument based upon the location of SKB in the send queue.
1135 if (nonagle & TCP_NAGLE_PUSH)
1138 /* Don't use the nagle rule for urgent data (or for the final FIN).
1139 * Nagle can be ignored during F-RTO too (see RFC4138).
1141 if (tp->urg_mode || (tp->frto_counter == 2) ||
1142 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1145 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1151 /* Does at least the first segment of SKB fit into the send window? */
1152 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1153 unsigned int cur_mss)
1155 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1157 if (skb->len > cur_mss)
1158 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1160 return !after(end_seq, tcp_wnd_end(tp));
1163 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1164 * should be put on the wire right now. If so, it returns the number of
1165 * packets allowed by the congestion window.
1167 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1168 unsigned int cur_mss, int nonagle)
1170 struct tcp_sock *tp = tcp_sk(sk);
1171 unsigned int cwnd_quota;
1173 tcp_init_tso_segs(sk, skb, cur_mss);
1175 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1178 cwnd_quota = tcp_cwnd_test(tp, skb);
1179 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1185 int tcp_may_send_now(struct sock *sk)
1187 struct tcp_sock *tp = tcp_sk(sk);
1188 struct sk_buff *skb = tcp_send_head(sk);
1191 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1192 (tcp_skb_is_last(sk, skb) ?
1193 tp->nonagle : TCP_NAGLE_PUSH)));
1196 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1197 * which is put after SKB on the list. It is very much like
1198 * tcp_fragment() except that it may make several kinds of assumptions
1199 * in order to speed up the splitting operation. In particular, we
1200 * know that all the data is in scatter-gather pages, and that the
1201 * packet has never been sent out before (and thus is not cloned).
1203 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1204 unsigned int mss_now)
1206 struct sk_buff *buff;
1207 int nlen = skb->len - len;
1210 /* All of a TSO frame must be composed of paged data. */
1211 if (skb->len != skb->data_len)
1212 return tcp_fragment(sk, skb, len, mss_now);
1214 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1215 if (unlikely(buff == NULL))
1218 sk->sk_wmem_queued += buff->truesize;
1219 sk_mem_charge(sk, buff->truesize);
1220 buff->truesize += nlen;
1221 skb->truesize -= nlen;
1223 /* Correct the sequence numbers. */
1224 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1225 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1226 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1228 /* PSH and FIN should only be set in the second packet. */
1229 flags = TCP_SKB_CB(skb)->flags;
1230 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1231 TCP_SKB_CB(buff)->flags = flags;
1233 /* This packet was never sent out yet, so no SACK bits. */
1234 TCP_SKB_CB(buff)->sacked = 0;
1236 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1237 skb_split(skb, buff, len);
1239 /* Fix up tso_factor for both original and new SKB. */
1240 tcp_set_skb_tso_segs(sk, skb, mss_now);
1241 tcp_set_skb_tso_segs(sk, buff, mss_now);
1243 /* Link BUFF into the send queue. */
1244 skb_header_release(buff);
1245 tcp_insert_write_queue_after(skb, buff, sk);
1250 /* Try to defer sending, if possible, in order to minimize the amount
1251 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1253 * This algorithm is from John Heffner.
1255 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1257 struct tcp_sock *tp = tcp_sk(sk);
1258 const struct inet_connection_sock *icsk = inet_csk(sk);
1259 u32 send_win, cong_win, limit, in_flight;
1261 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1264 if (icsk->icsk_ca_state != TCP_CA_Open)
1267 /* Defer for less than two clock ticks. */
1268 if (tp->tso_deferred &&
1269 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1272 in_flight = tcp_packets_in_flight(tp);
1274 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1276 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1278 /* From in_flight test above, we know that cwnd > in_flight. */
1279 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1281 limit = min(send_win, cong_win);
1283 /* If a full-sized TSO skb can be sent, do it. */
1284 if (limit >= sk->sk_gso_max_size)
1287 if (sysctl_tcp_tso_win_divisor) {
1288 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1290 /* If at least some fraction of a window is available,
1293 chunk /= sysctl_tcp_tso_win_divisor;
1297 /* Different approach, try not to defer past a single
1298 * ACK. Receiver should ACK every other full sized
1299 * frame, so if we have space for more than 3 frames
1302 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1306 /* Ok, it looks like it is advisable to defer. */
1307 tp->tso_deferred = 1 | (jiffies << 1);
1312 tp->tso_deferred = 0;
1316 /* Create a new MTU probe if we are ready.
1317 * Returns 0 if we should wait to probe (no cwnd available),
1318 * 1 if a probe was sent,
1321 static int tcp_mtu_probe(struct sock *sk)
1323 struct tcp_sock *tp = tcp_sk(sk);
1324 struct inet_connection_sock *icsk = inet_csk(sk);
1325 struct sk_buff *skb, *nskb, *next;
1332 /* Not currently probing/verifying,
1334 * have enough cwnd, and
1335 * not SACKing (the variable headers throw things off) */
1336 if (!icsk->icsk_mtup.enabled ||
1337 icsk->icsk_mtup.probe_size ||
1338 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1339 tp->snd_cwnd < 11 ||
1340 tp->rx_opt.eff_sacks)
1343 /* Very simple search strategy: just double the MSS. */
1344 mss_now = tcp_current_mss(sk, 0);
1345 probe_size = 2 * tp->mss_cache;
1346 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1347 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1348 /* TODO: set timer for probe_converge_event */
1352 /* Have enough data in the send queue to probe? */
1353 if (tp->write_seq - tp->snd_nxt < size_needed)
1356 if (tp->snd_wnd < size_needed)
1358 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1361 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1362 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1363 if (!tcp_packets_in_flight(tp))
1369 /* We're allowed to probe. Build it now. */
1370 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1372 sk->sk_wmem_queued += nskb->truesize;
1373 sk_mem_charge(sk, nskb->truesize);
1375 skb = tcp_send_head(sk);
1377 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1378 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1379 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1380 TCP_SKB_CB(nskb)->sacked = 0;
1382 nskb->ip_summed = skb->ip_summed;
1384 tcp_insert_write_queue_before(nskb, skb, sk);
1387 tcp_for_write_queue_from_safe(skb, next, sk) {
1388 copy = min_t(int, skb->len, probe_size - len);
1389 if (nskb->ip_summed)
1390 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1392 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1393 skb_put(nskb, copy),
1396 if (skb->len <= copy) {
1397 /* We've eaten all the data from this skb.
1399 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1400 tcp_unlink_write_queue(skb, sk);
1401 sk_wmem_free_skb(sk, skb);
1403 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1404 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1405 if (!skb_shinfo(skb)->nr_frags) {
1406 skb_pull(skb, copy);
1407 if (skb->ip_summed != CHECKSUM_PARTIAL)
1408 skb->csum = csum_partial(skb->data,
1411 __pskb_trim_head(skb, copy);
1412 tcp_set_skb_tso_segs(sk, skb, mss_now);
1414 TCP_SKB_CB(skb)->seq += copy;
1419 if (len >= probe_size)
1422 tcp_init_tso_segs(sk, nskb, nskb->len);
1424 /* We're ready to send. If this fails, the probe will
1425 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1426 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1427 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1428 /* Decrement cwnd here because we are sending
1429 * effectively two packets. */
1431 tcp_event_new_data_sent(sk, nskb);
1433 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1434 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1435 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1443 /* This routine writes packets to the network. It advances the
1444 * send_head. This happens as incoming acks open up the remote
1447 * Returns 1, if no segments are in flight and we have queued segments, but
1448 * cannot send anything now because of SWS or another problem.
1450 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1452 struct tcp_sock *tp = tcp_sk(sk);
1453 struct sk_buff *skb;
1454 unsigned int tso_segs, sent_pkts;
1458 /* If we are closed, the bytes will have to remain here.
1459 * In time closedown will finish, we empty the write queue and all
1462 if (unlikely(sk->sk_state == TCP_CLOSE))
1467 /* Do MTU probing. */
1468 if ((result = tcp_mtu_probe(sk)) == 0) {
1470 } else if (result > 0) {
1474 while ((skb = tcp_send_head(sk))) {
1477 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1480 cwnd_quota = tcp_cwnd_test(tp, skb);
1484 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1487 if (tso_segs == 1) {
1488 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1489 (tcp_skb_is_last(sk, skb) ?
1490 nonagle : TCP_NAGLE_PUSH))))
1493 if (tcp_tso_should_defer(sk, skb))
1499 limit = tcp_mss_split_point(sk, skb, mss_now,
1502 if (skb->len > limit &&
1503 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1506 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1508 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1511 /* Advance the send_head. This one is sent out.
1512 * This call will increment packets_out.
1514 tcp_event_new_data_sent(sk, skb);
1516 tcp_minshall_update(tp, mss_now, skb);
1520 if (likely(sent_pkts)) {
1521 tcp_cwnd_validate(sk);
1524 return !tp->packets_out && tcp_send_head(sk);
1527 /* Push out any pending frames which were held back due to
1528 * TCP_CORK or attempt at coalescing tiny packets.
1529 * The socket must be locked by the caller.
1531 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1534 struct sk_buff *skb = tcp_send_head(sk);
1537 if (tcp_write_xmit(sk, cur_mss, nonagle))
1538 tcp_check_probe_timer(sk);
1542 /* Send _single_ skb sitting at the send head. This function requires
1543 * true push pending frames to setup probe timer etc.
1545 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1547 struct sk_buff *skb = tcp_send_head(sk);
1548 unsigned int tso_segs, cwnd_quota;
1550 BUG_ON(!skb || skb->len < mss_now);
1552 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1553 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1555 if (likely(cwnd_quota)) {
1562 limit = tcp_mss_split_point(sk, skb, mss_now,
1565 if (skb->len > limit &&
1566 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1569 /* Send it out now. */
1570 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1572 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1573 tcp_event_new_data_sent(sk, skb);
1574 tcp_cwnd_validate(sk);
1580 /* This function returns the amount that we can raise the
1581 * usable window based on the following constraints
1583 * 1. The window can never be shrunk once it is offered (RFC 793)
1584 * 2. We limit memory per socket
1587 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1588 * RECV.NEXT + RCV.WIN fixed until:
1589 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1591 * i.e. don't raise the right edge of the window until you can raise
1592 * it at least MSS bytes.
1594 * Unfortunately, the recommended algorithm breaks header prediction,
1595 * since header prediction assumes th->window stays fixed.
1597 * Strictly speaking, keeping th->window fixed violates the receiver
1598 * side SWS prevention criteria. The problem is that under this rule
1599 * a stream of single byte packets will cause the right side of the
1600 * window to always advance by a single byte.
1602 * Of course, if the sender implements sender side SWS prevention
1603 * then this will not be a problem.
1605 * BSD seems to make the following compromise:
1607 * If the free space is less than the 1/4 of the maximum
1608 * space available and the free space is less than 1/2 mss,
1609 * then set the window to 0.
1610 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1611 * Otherwise, just prevent the window from shrinking
1612 * and from being larger than the largest representable value.
1614 * This prevents incremental opening of the window in the regime
1615 * where TCP is limited by the speed of the reader side taking
1616 * data out of the TCP receive queue. It does nothing about
1617 * those cases where the window is constrained on the sender side
1618 * because the pipeline is full.
1620 * BSD also seems to "accidentally" limit itself to windows that are a
1621 * multiple of MSS, at least until the free space gets quite small.
1622 * This would appear to be a side effect of the mbuf implementation.
1623 * Combining these two algorithms results in the observed behavior
1624 * of having a fixed window size at almost all times.
1626 * Below we obtain similar behavior by forcing the offered window to
1627 * a multiple of the mss when it is feasible to do so.
1629 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1630 * Regular options like TIMESTAMP are taken into account.
1632 u32 __tcp_select_window(struct sock *sk)
1634 struct inet_connection_sock *icsk = inet_csk(sk);
1635 struct tcp_sock *tp = tcp_sk(sk);
1636 /* MSS for the peer's data. Previous versions used mss_clamp
1637 * here. I don't know if the value based on our guesses
1638 * of peer's MSS is better for the performance. It's more correct
1639 * but may be worse for the performance because of rcv_mss
1640 * fluctuations. --SAW 1998/11/1
1642 int mss = icsk->icsk_ack.rcv_mss;
1643 int free_space = tcp_space(sk);
1644 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1647 if (mss > full_space)
1650 if (free_space < (full_space >> 1)) {
1651 icsk->icsk_ack.quick = 0;
1653 if (tcp_memory_pressure)
1654 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1657 if (free_space < mss)
1661 if (free_space > tp->rcv_ssthresh)
1662 free_space = tp->rcv_ssthresh;
1664 /* Don't do rounding if we are using window scaling, since the
1665 * scaled window will not line up with the MSS boundary anyway.
1667 window = tp->rcv_wnd;
1668 if (tp->rx_opt.rcv_wscale) {
1669 window = free_space;
1671 /* Advertise enough space so that it won't get scaled away.
1672 * Import case: prevent zero window announcement if
1673 * 1<<rcv_wscale > mss.
1675 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1676 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1677 << tp->rx_opt.rcv_wscale);
1679 /* Get the largest window that is a nice multiple of mss.
1680 * Window clamp already applied above.
1681 * If our current window offering is within 1 mss of the
1682 * free space we just keep it. This prevents the divide
1683 * and multiply from happening most of the time.
1684 * We also don't do any window rounding when the free space
1687 if (window <= free_space - mss || window > free_space)
1688 window = (free_space / mss) * mss;
1689 else if (mss == full_space &&
1690 free_space > window + (full_space >> 1))
1691 window = free_space;
1697 /* Attempt to collapse two adjacent SKB's during retransmission. */
1698 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1701 struct tcp_sock *tp = tcp_sk(sk);
1702 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1703 int skb_size, next_skb_size;
1706 /* The first test we must make is that neither of these two
1707 * SKB's are still referenced by someone else.
1709 if (skb_cloned(skb) || skb_cloned(next_skb))
1712 skb_size = skb->len;
1713 next_skb_size = next_skb->len;
1714 flags = TCP_SKB_CB(skb)->flags;
1716 /* Also punt if next skb has been SACK'd. */
1717 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1720 /* Next skb is out of window. */
1721 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1724 /* Punt if not enough space exists in the first SKB for
1725 * the data in the second, or the total combined payload
1726 * would exceed the MSS.
1728 if ((next_skb_size > skb_tailroom(skb)) ||
1729 ((skb_size + next_skb_size) > mss_now))
1732 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1734 tcp_highest_sack_combine(sk, next_skb, skb);
1736 /* Ok. We will be able to collapse the packet. */
1737 tcp_unlink_write_queue(next_skb, sk);
1739 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1742 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1743 skb->ip_summed = CHECKSUM_PARTIAL;
1745 if (skb->ip_summed != CHECKSUM_PARTIAL)
1746 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1748 /* Update sequence range on original skb. */
1749 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1751 /* Merge over control information. */
1752 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1753 TCP_SKB_CB(skb)->flags = flags;
1755 /* All done, get rid of second SKB and account for it so
1756 * packet counting does not break.
1758 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1759 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1760 tp->retrans_out -= tcp_skb_pcount(next_skb);
1761 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1762 tp->lost_out -= tcp_skb_pcount(next_skb);
1763 /* Reno case is special. Sigh... */
1764 if (tcp_is_reno(tp) && tp->sacked_out)
1765 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1767 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1768 tp->packets_out -= tcp_skb_pcount(next_skb);
1770 /* changed transmit queue under us so clear hints */
1771 tcp_clear_retrans_hints_partial(tp);
1773 sk_wmem_free_skb(sk, next_skb);
1776 /* Do a simple retransmit without using the backoff mechanisms in
1777 * tcp_timer. This is used for path mtu discovery.
1778 * The socket is already locked here.
1780 void tcp_simple_retransmit(struct sock *sk)
1782 const struct inet_connection_sock *icsk = inet_csk(sk);
1783 struct tcp_sock *tp = tcp_sk(sk);
1784 struct sk_buff *skb;
1785 unsigned int mss = tcp_current_mss(sk, 0);
1788 tcp_for_write_queue(skb, sk) {
1789 if (skb == tcp_send_head(sk))
1791 if (skb->len > mss &&
1792 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1793 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1794 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1795 tp->retrans_out -= tcp_skb_pcount(skb);
1797 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
1798 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1799 tp->lost_out += tcp_skb_pcount(skb);
1805 tcp_clear_all_retrans_hints(tp);
1810 if (tcp_is_reno(tp))
1811 tcp_limit_reno_sacked(tp);
1813 tcp_verify_left_out(tp);
1815 /* Don't muck with the congestion window here.
1816 * Reason is that we do not increase amount of _data_
1817 * in network, but units changed and effective
1818 * cwnd/ssthresh really reduced now.
1820 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1821 tp->high_seq = tp->snd_nxt;
1822 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1823 tp->prior_ssthresh = 0;
1824 tp->undo_marker = 0;
1825 tcp_set_ca_state(sk, TCP_CA_Loss);
1827 tcp_xmit_retransmit_queue(sk);
1830 /* This retransmits one SKB. Policy decisions and retransmit queue
1831 * state updates are done by the caller. Returns non-zero if an
1832 * error occurred which prevented the send.
1834 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct inet_connection_sock *icsk = inet_csk(sk);
1838 unsigned int cur_mss;
1841 /* Inconslusive MTU probe */
1842 if (icsk->icsk_mtup.probe_size) {
1843 icsk->icsk_mtup.probe_size = 0;
1846 /* Do not sent more than we queued. 1/4 is reserved for possible
1847 * copying overhead: fragmentation, tunneling, mangling etc.
1849 if (atomic_read(&sk->sk_wmem_alloc) >
1850 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1853 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1854 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1856 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1860 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1861 return -EHOSTUNREACH; /* Routing failure or similar. */
1863 cur_mss = tcp_current_mss(sk, 0);
1865 /* If receiver has shrunk his window, and skb is out of
1866 * new window, do not retransmit it. The exception is the
1867 * case, when window is shrunk to zero. In this case
1868 * our retransmit serves as a zero window probe.
1870 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1871 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1874 if (skb->len > cur_mss) {
1875 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1876 return -ENOMEM; /* We'll try again later. */
1879 /* Collapse two adjacent packets if worthwhile and we can. */
1880 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1881 (skb->len < (cur_mss >> 1)) &&
1882 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1883 (!tcp_skb_is_last(sk, skb)) &&
1884 (skb_shinfo(skb)->nr_frags == 0 &&
1885 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1886 (tcp_skb_pcount(skb) == 1 &&
1887 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1888 (sysctl_tcp_retrans_collapse != 0))
1889 tcp_retrans_try_collapse(sk, skb, cur_mss);
1891 /* Some Solaris stacks overoptimize and ignore the FIN on a
1892 * retransmit when old data is attached. So strip it off
1893 * since it is cheap to do so and saves bytes on the network.
1896 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1897 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1898 if (!pskb_trim(skb, 0)) {
1899 /* Reuse, even though it does some unnecessary work */
1900 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1901 TCP_SKB_CB(skb)->flags);
1902 skb->ip_summed = CHECKSUM_NONE;
1906 /* Make a copy, if the first transmission SKB clone we made
1907 * is still in somebody's hands, else make a clone.
1909 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1911 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1914 /* Update global TCP statistics. */
1915 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1917 tp->total_retrans++;
1919 #if FASTRETRANS_DEBUG > 0
1920 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1921 if (net_ratelimit())
1922 printk(KERN_DEBUG "retrans_out leaked.\n");
1925 if (!tp->retrans_out)
1926 tp->lost_retrans_low = tp->snd_nxt;
1927 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1928 tp->retrans_out += tcp_skb_pcount(skb);
1930 /* Save stamp of the first retransmit. */
1931 if (!tp->retrans_stamp)
1932 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1936 /* snd_nxt is stored to detect loss of retransmitted segment,
1937 * see tcp_input.c tcp_sacktag_write_queue().
1939 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1944 /* This gets called after a retransmit timeout, and the initially
1945 * retransmitted data is acknowledged. It tries to continue
1946 * resending the rest of the retransmit queue, until either
1947 * we've sent it all or the congestion window limit is reached.
1948 * If doing SACK, the first ACK which comes back for a timeout
1949 * based retransmit packet might feed us FACK information again.
1950 * If so, we use it to avoid unnecessarily retransmissions.
1952 void tcp_xmit_retransmit_queue(struct sock *sk)
1954 const struct inet_connection_sock *icsk = inet_csk(sk);
1955 struct tcp_sock *tp = tcp_sk(sk);
1956 struct sk_buff *skb;
1959 if (tp->retransmit_skb_hint) {
1960 skb = tp->retransmit_skb_hint;
1961 packet_cnt = tp->retransmit_cnt_hint;
1963 skb = tcp_write_queue_head(sk);
1967 /* First pass: retransmit lost packets. */
1969 tcp_for_write_queue_from(skb, sk) {
1970 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1972 if (skb == tcp_send_head(sk))
1974 /* we could do better than to assign each time */
1975 tp->retransmit_skb_hint = skb;
1976 tp->retransmit_cnt_hint = packet_cnt;
1978 /* Assume this retransmit will generate
1979 * only one packet for congestion window
1980 * calculation purposes. This works because
1981 * tcp_retransmit_skb() will chop up the
1982 * packet to be MSS sized and all the
1983 * packet counting works out.
1985 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1988 if (sacked & TCPCB_LOST) {
1989 if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1990 if (tcp_retransmit_skb(sk, skb)) {
1991 tp->retransmit_skb_hint = NULL;
1994 if (icsk->icsk_ca_state != TCP_CA_Loss)
1995 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1997 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1999 if (skb == tcp_write_queue_head(sk))
2000 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2001 inet_csk(sk)->icsk_rto,
2005 packet_cnt += tcp_skb_pcount(skb);
2006 if (packet_cnt >= tp->lost_out)
2012 /* OK, demanded retransmission is finished. */
2014 /* Forward retransmissions are possible only during Recovery. */
2015 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2018 /* No forward retransmissions in Reno are possible. */
2019 if (tcp_is_reno(tp))
2022 /* Yeah, we have to make difficult choice between forward transmission
2023 * and retransmission... Both ways have their merits...
2025 * For now we do not retransmit anything, while we have some new
2026 * segments to send. In the other cases, follow rule 3 for
2027 * NextSeg() specified in RFC3517.
2030 if (tcp_may_send_now(sk))
2033 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2034 if (!tp->sacked_out)
2037 if (tp->forward_skb_hint)
2038 skb = tp->forward_skb_hint;
2040 skb = tcp_write_queue_head(sk);
2042 tcp_for_write_queue_from(skb, sk) {
2043 if (skb == tcp_send_head(sk))
2045 tp->forward_skb_hint = skb;
2047 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2050 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2053 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2056 /* Ok, retransmit it. */
2057 if (tcp_retransmit_skb(sk, skb)) {
2058 tp->forward_skb_hint = NULL;
2062 if (skb == tcp_write_queue_head(sk))
2063 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2064 inet_csk(sk)->icsk_rto,
2067 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2071 /* Send a fin. The caller locks the socket for us. This cannot be
2072 * allowed to fail queueing a FIN frame under any circumstances.
2074 void tcp_send_fin(struct sock *sk)
2076 struct tcp_sock *tp = tcp_sk(sk);
2077 struct sk_buff *skb = tcp_write_queue_tail(sk);
2080 /* Optimization, tack on the FIN if we have a queue of
2081 * unsent frames. But be careful about outgoing SACKS
2084 mss_now = tcp_current_mss(sk, 1);
2086 if (tcp_send_head(sk) != NULL) {
2087 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2088 TCP_SKB_CB(skb)->end_seq++;
2091 /* Socket is locked, keep trying until memory is available. */
2093 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2099 /* Reserve space for headers and prepare control bits. */
2100 skb_reserve(skb, MAX_TCP_HEADER);
2101 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2102 tcp_init_nondata_skb(skb, tp->write_seq,
2103 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2104 tcp_queue_skb(sk, skb);
2106 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2109 /* We get here when a process closes a file descriptor (either due to
2110 * an explicit close() or as a byproduct of exit()'ing) and there
2111 * was unread data in the receive queue. This behavior is recommended
2112 * by RFC 2525, section 2.17. -DaveM
2114 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2116 struct sk_buff *skb;
2118 /* NOTE: No TCP options attached and we never retransmit this. */
2119 skb = alloc_skb(MAX_TCP_HEADER, priority);
2121 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2125 /* Reserve space for headers and prepare control bits. */
2126 skb_reserve(skb, MAX_TCP_HEADER);
2127 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2128 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2130 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2131 if (tcp_transmit_skb(sk, skb, 0, priority))
2132 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2134 TCP_INC_STATS(TCP_MIB_OUTRSTS);
2137 /* WARNING: This routine must only be called when we have already sent
2138 * a SYN packet that crossed the incoming SYN that caused this routine
2139 * to get called. If this assumption fails then the initial rcv_wnd
2140 * and rcv_wscale values will not be correct.
2142 int tcp_send_synack(struct sock *sk)
2144 struct sk_buff *skb;
2146 skb = tcp_write_queue_head(sk);
2147 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2148 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2151 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2152 if (skb_cloned(skb)) {
2153 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2156 tcp_unlink_write_queue(skb, sk);
2157 skb_header_release(nskb);
2158 __tcp_add_write_queue_head(sk, nskb);
2159 sk_wmem_free_skb(sk, skb);
2160 sk->sk_wmem_queued += nskb->truesize;
2161 sk_mem_charge(sk, nskb->truesize);
2165 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2166 TCP_ECN_send_synack(tcp_sk(sk), skb);
2168 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2169 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2173 * Prepare a SYN-ACK.
2175 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2176 struct request_sock *req)
2178 struct inet_request_sock *ireq = inet_rsk(req);
2179 struct tcp_sock *tp = tcp_sk(sk);
2181 int tcp_header_size;
2182 struct sk_buff *skb;
2183 #ifdef CONFIG_TCP_MD5SIG
2184 struct tcp_md5sig_key *md5;
2185 __u8 *md5_hash_location;
2188 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2192 /* Reserve space for headers. */
2193 skb_reserve(skb, MAX_TCP_HEADER);
2195 skb->dst = dst_clone(dst);
2197 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2198 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2199 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2200 /* SACK_PERM is in the place of NOP NOP of TS */
2201 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2203 #ifdef CONFIG_TCP_MD5SIG
2204 /* Are we doing MD5 on this segment? If so - make room for it */
2205 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2207 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2209 skb_push(skb, tcp_header_size);
2210 skb_reset_transport_header(skb);
2213 memset(th, 0, sizeof(struct tcphdr));
2216 TCP_ECN_make_synack(req, th);
2217 th->source = inet_sk(sk)->sport;
2218 th->dest = ireq->rmt_port;
2219 /* Setting of flags are superfluous here for callers (and ECE is
2220 * not even correctly set)
2222 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2223 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2224 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2225 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2226 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2228 /* Set this up on the first call only */
2229 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2230 /* tcp_full_space because it is guaranteed to be the first packet */
2231 tcp_select_initial_window(tcp_full_space(sk),
2232 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2237 ireq->rcv_wscale = rcv_wscale;
2240 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2241 th->window = htons(min(req->rcv_wnd, 65535U));
2242 #ifdef CONFIG_SYN_COOKIES
2243 if (unlikely(req->cookie_ts))
2244 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2247 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2248 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2249 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2250 TCP_SKB_CB(skb)->when,
2253 #ifdef CONFIG_TCP_MD5SIG
2254 md5 ? &md5_hash_location :
2259 th->doff = (tcp_header_size >> 2);
2260 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2262 #ifdef CONFIG_TCP_MD5SIG
2263 /* Okay, we have all we need - do the md5 hash if needed */
2265 tp->af_specific->calc_md5_hash(md5_hash_location,
2277 * Do all connect socket setups that can be done AF independent.
2279 static void tcp_connect_init(struct sock *sk)
2281 struct dst_entry *dst = __sk_dst_get(sk);
2282 struct tcp_sock *tp = tcp_sk(sk);
2285 /* We'll fix this up when we get a response from the other end.
2286 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2288 tp->tcp_header_len = sizeof(struct tcphdr) +
2289 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2291 #ifdef CONFIG_TCP_MD5SIG
2292 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2293 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2296 /* If user gave his TCP_MAXSEG, record it to clamp */
2297 if (tp->rx_opt.user_mss)
2298 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2301 tcp_sync_mss(sk, dst_mtu(dst));
2303 if (!tp->window_clamp)
2304 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2305 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2306 tcp_initialize_rcv_mss(sk);
2308 tcp_select_initial_window(tcp_full_space(sk),
2309 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2312 sysctl_tcp_window_scaling,
2315 tp->rx_opt.rcv_wscale = rcv_wscale;
2316 tp->rcv_ssthresh = tp->rcv_wnd;
2319 sock_reset_flag(sk, SOCK_DONE);
2321 tcp_init_wl(tp, tp->write_seq, 0);
2322 tp->snd_una = tp->write_seq;
2323 tp->snd_sml = tp->write_seq;
2328 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2329 inet_csk(sk)->icsk_retransmits = 0;
2330 tcp_clear_retrans(tp);
2334 * Build a SYN and send it off.
2336 int tcp_connect(struct sock *sk)
2338 struct tcp_sock *tp = tcp_sk(sk);
2339 struct sk_buff *buff;
2341 tcp_connect_init(sk);
2343 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2344 if (unlikely(buff == NULL))
2347 /* Reserve space for headers. */
2348 skb_reserve(buff, MAX_TCP_HEADER);
2350 tp->snd_nxt = tp->write_seq;
2351 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2352 TCP_ECN_send_syn(sk, buff);
2355 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2356 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2357 skb_header_release(buff);
2358 __tcp_add_write_queue_tail(sk, buff);
2359 sk->sk_wmem_queued += buff->truesize;
2360 sk_mem_charge(sk, buff->truesize);
2361 tp->packets_out += tcp_skb_pcount(buff);
2362 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2364 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2365 * in order to make this packet get counted in tcpOutSegs.
2367 tp->snd_nxt = tp->write_seq;
2368 tp->pushed_seq = tp->write_seq;
2369 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2371 /* Timer for repeating the SYN until an answer. */
2372 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2373 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2377 /* Send out a delayed ack, the caller does the policy checking
2378 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2381 void tcp_send_delayed_ack(struct sock *sk)
2383 struct inet_connection_sock *icsk = inet_csk(sk);
2384 int ato = icsk->icsk_ack.ato;
2385 unsigned long timeout;
2387 if (ato > TCP_DELACK_MIN) {
2388 const struct tcp_sock *tp = tcp_sk(sk);
2389 int max_ato = HZ / 2;
2391 if (icsk->icsk_ack.pingpong ||
2392 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2393 max_ato = TCP_DELACK_MAX;
2395 /* Slow path, intersegment interval is "high". */
2397 /* If some rtt estimate is known, use it to bound delayed ack.
2398 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2402 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2408 ato = min(ato, max_ato);
2411 /* Stay within the limit we were given */
2412 timeout = jiffies + ato;
2414 /* Use new timeout only if there wasn't a older one earlier. */
2415 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2416 /* If delack timer was blocked or is about to expire,
2419 if (icsk->icsk_ack.blocked ||
2420 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2425 if (!time_before(timeout, icsk->icsk_ack.timeout))
2426 timeout = icsk->icsk_ack.timeout;
2428 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2429 icsk->icsk_ack.timeout = timeout;
2430 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2433 /* This routine sends an ack and also updates the window. */
2434 void tcp_send_ack(struct sock *sk)
2436 struct sk_buff *buff;
2438 /* If we have been reset, we may not send again. */
2439 if (sk->sk_state == TCP_CLOSE)
2442 /* We are not putting this on the write queue, so
2443 * tcp_transmit_skb() will set the ownership to this
2446 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2448 inet_csk_schedule_ack(sk);
2449 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2450 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2451 TCP_DELACK_MAX, TCP_RTO_MAX);
2455 /* Reserve space for headers and prepare control bits. */
2456 skb_reserve(buff, MAX_TCP_HEADER);
2457 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2459 /* Send it off, this clears delayed acks for us. */
2460 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2461 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2464 /* This routine sends a packet with an out of date sequence
2465 * number. It assumes the other end will try to ack it.
2467 * Question: what should we make while urgent mode?
2468 * 4.4BSD forces sending single byte of data. We cannot send
2469 * out of window data, because we have SND.NXT==SND.MAX...
2471 * Current solution: to send TWO zero-length segments in urgent mode:
2472 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2473 * out-of-date with SND.UNA-1 to probe window.
2475 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2477 struct tcp_sock *tp = tcp_sk(sk);
2478 struct sk_buff *skb;
2480 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2481 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2485 /* Reserve space for headers and set control bits. */
2486 skb_reserve(skb, MAX_TCP_HEADER);
2487 /* Use a previous sequence. This should cause the other
2488 * end to send an ack. Don't queue or clone SKB, just
2491 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2492 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2493 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2496 int tcp_write_wakeup(struct sock *sk)
2498 struct tcp_sock *tp = tcp_sk(sk);
2499 struct sk_buff *skb;
2501 if (sk->sk_state == TCP_CLOSE)
2504 if ((skb = tcp_send_head(sk)) != NULL &&
2505 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2507 unsigned int mss = tcp_current_mss(sk, 0);
2508 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2510 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2511 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2513 /* We are probing the opening of a window
2514 * but the window size is != 0
2515 * must have been a result SWS avoidance ( sender )
2517 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2519 seg_size = min(seg_size, mss);
2520 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2521 if (tcp_fragment(sk, skb, seg_size, mss))
2523 } else if (!tcp_skb_pcount(skb))
2524 tcp_set_skb_tso_segs(sk, skb, mss);
2526 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2527 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2528 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2530 tcp_event_new_data_sent(sk, skb);
2534 between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2535 tcp_xmit_probe_skb(sk, 1);
2536 return tcp_xmit_probe_skb(sk, 0);
2540 /* A window probe timeout has occurred. If window is not closed send
2541 * a partial packet else a zero probe.
2543 void tcp_send_probe0(struct sock *sk)
2545 struct inet_connection_sock *icsk = inet_csk(sk);
2546 struct tcp_sock *tp = tcp_sk(sk);
2549 err = tcp_write_wakeup(sk);
2551 if (tp->packets_out || !tcp_send_head(sk)) {
2552 /* Cancel probe timer, if it is not required. */
2553 icsk->icsk_probes_out = 0;
2554 icsk->icsk_backoff = 0;
2559 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2560 icsk->icsk_backoff++;
2561 icsk->icsk_probes_out++;
2562 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2563 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2566 /* If packet was not sent due to local congestion,
2567 * do not backoff and do not remember icsk_probes_out.
2568 * Let local senders to fight for local resources.
2570 * Use accumulated backoff yet.
2572 if (!icsk->icsk_probes_out)
2573 icsk->icsk_probes_out = 1;
2574 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2575 min(icsk->icsk_rto << icsk->icsk_backoff,
2576 TCP_RESOURCE_PROBE_INTERVAL),
2581 EXPORT_SYMBOL(tcp_select_initial_window);
2582 EXPORT_SYMBOL(tcp_connect);
2583 EXPORT_SYMBOL(tcp_make_synack);
2584 EXPORT_SYMBOL(tcp_simple_retransmit);
2585 EXPORT_SYMBOL(tcp_sync_mss);
2586 EXPORT_SYMBOL(tcp_mtup_init);