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,
615 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
617 if (likely(tcb->flags & TCPCB_FLAG_ACK))
618 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
620 if (skb->len != tcp_header_size)
621 tcp_event_data_sent(tp, skb, sk);
623 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
624 TCP_INC_STATS(TCP_MIB_OUTSEGS);
626 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
627 if (likely(err <= 0))
630 tcp_enter_cwr(sk, 1);
632 return net_xmit_eval(err);
634 #undef SYSCTL_FLAG_TSTAMPS
635 #undef SYSCTL_FLAG_WSCALE
636 #undef SYSCTL_FLAG_SACK
639 /* This routine just queue's the buffer
641 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
642 * otherwise socket can stall.
644 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
646 struct tcp_sock *tp = tcp_sk(sk);
648 /* Advance write_seq and place onto the write_queue. */
649 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
650 skb_header_release(skb);
651 tcp_add_write_queue_tail(sk, skb);
652 sk->sk_wmem_queued += skb->truesize;
653 sk_mem_charge(sk, skb->truesize);
656 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
657 unsigned int mss_now)
659 if (skb->len <= mss_now || !sk_can_gso(sk)) {
660 /* Avoid the costly divide in the normal
663 skb_shinfo(skb)->gso_segs = 1;
664 skb_shinfo(skb)->gso_size = 0;
665 skb_shinfo(skb)->gso_type = 0;
667 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
668 skb_shinfo(skb)->gso_size = mss_now;
669 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
673 /* When a modification to fackets out becomes necessary, we need to check
674 * skb is counted to fackets_out or not.
676 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
679 struct tcp_sock *tp = tcp_sk(sk);
681 if (!tp->sacked_out || tcp_is_reno(tp))
684 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
685 tp->fackets_out -= decr;
688 /* Function to create two new TCP segments. Shrinks the given segment
689 * to the specified size and appends a new segment with the rest of the
690 * packet to the list. This won't be called frequently, I hope.
691 * Remember, these are still headerless SKBs at this point.
693 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
694 unsigned int mss_now)
696 struct tcp_sock *tp = tcp_sk(sk);
697 struct sk_buff *buff;
698 int nsize, old_factor;
702 BUG_ON(len > skb->len);
704 tcp_clear_retrans_hints_partial(tp);
705 nsize = skb_headlen(skb) - len;
709 if (skb_cloned(skb) &&
710 skb_is_nonlinear(skb) &&
711 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
714 /* Get a new skb... force flag on. */
715 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
717 return -ENOMEM; /* We'll just try again later. */
719 sk->sk_wmem_queued += buff->truesize;
720 sk_mem_charge(sk, buff->truesize);
721 nlen = skb->len - len - nsize;
722 buff->truesize += nlen;
723 skb->truesize -= nlen;
725 /* Correct the sequence numbers. */
726 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
727 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
728 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
730 /* PSH and FIN should only be set in the second packet. */
731 flags = TCP_SKB_CB(skb)->flags;
732 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
733 TCP_SKB_CB(buff)->flags = flags;
734 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
736 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
737 /* Copy and checksum data tail into the new buffer. */
738 buff->csum = csum_partial_copy_nocheck(skb->data + len,
739 skb_put(buff, nsize),
744 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
746 skb->ip_summed = CHECKSUM_PARTIAL;
747 skb_split(skb, buff, len);
750 buff->ip_summed = skb->ip_summed;
752 /* Looks stupid, but our code really uses when of
753 * skbs, which it never sent before. --ANK
755 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
756 buff->tstamp = skb->tstamp;
758 old_factor = tcp_skb_pcount(skb);
760 /* Fix up tso_factor for both original and new SKB. */
761 tcp_set_skb_tso_segs(sk, skb, mss_now);
762 tcp_set_skb_tso_segs(sk, buff, mss_now);
764 /* If this packet has been sent out already, we must
765 * adjust the various packet counters.
767 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
768 int diff = old_factor - tcp_skb_pcount(skb) -
769 tcp_skb_pcount(buff);
771 tp->packets_out -= diff;
773 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
774 tp->sacked_out -= diff;
775 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
776 tp->retrans_out -= diff;
778 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
779 tp->lost_out -= diff;
781 /* Adjust Reno SACK estimate. */
782 if (tcp_is_reno(tp) && diff > 0) {
783 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
784 tcp_verify_left_out(tp);
786 tcp_adjust_fackets_out(sk, skb, diff);
789 /* Link BUFF into the send queue. */
790 skb_header_release(buff);
791 tcp_insert_write_queue_after(skb, buff, sk);
796 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
797 * eventually). The difference is that pulled data not copied, but
798 * immediately discarded.
800 static void __pskb_trim_head(struct sk_buff *skb, int len)
806 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
807 if (skb_shinfo(skb)->frags[i].size <= eat) {
808 put_page(skb_shinfo(skb)->frags[i].page);
809 eat -= skb_shinfo(skb)->frags[i].size;
811 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
813 skb_shinfo(skb)->frags[k].page_offset += eat;
814 skb_shinfo(skb)->frags[k].size -= eat;
820 skb_shinfo(skb)->nr_frags = k;
822 skb_reset_tail_pointer(skb);
823 skb->data_len -= len;
824 skb->len = skb->data_len;
827 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
829 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
832 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
833 if (unlikely(len < skb_headlen(skb)))
834 __skb_pull(skb, len);
836 __pskb_trim_head(skb, len - skb_headlen(skb));
838 TCP_SKB_CB(skb)->seq += len;
839 skb->ip_summed = CHECKSUM_PARTIAL;
841 skb->truesize -= len;
842 sk->sk_wmem_queued -= len;
843 sk_mem_uncharge(sk, len);
844 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
846 /* Any change of skb->len requires recalculation of tso
849 if (tcp_skb_pcount(skb) > 1)
850 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
855 /* Not accounting for SACKs here. */
856 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
858 struct tcp_sock *tp = tcp_sk(sk);
859 struct inet_connection_sock *icsk = inet_csk(sk);
862 /* Calculate base mss without TCP options:
863 It is MMS_S - sizeof(tcphdr) of rfc1122
865 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
867 /* Clamp it (mss_clamp does not include tcp options) */
868 if (mss_now > tp->rx_opt.mss_clamp)
869 mss_now = tp->rx_opt.mss_clamp;
871 /* Now subtract optional transport overhead */
872 mss_now -= icsk->icsk_ext_hdr_len;
874 /* Then reserve room for full set of TCP options and 8 bytes of data */
878 /* Now subtract TCP options size, not including SACKs */
879 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
884 /* Inverse of above */
885 int tcp_mss_to_mtu(struct sock *sk, int mss)
887 struct tcp_sock *tp = tcp_sk(sk);
888 struct inet_connection_sock *icsk = inet_csk(sk);
893 icsk->icsk_ext_hdr_len +
894 icsk->icsk_af_ops->net_header_len;
899 void tcp_mtup_init(struct sock *sk)
901 struct tcp_sock *tp = tcp_sk(sk);
902 struct inet_connection_sock *icsk = inet_csk(sk);
904 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
905 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
906 icsk->icsk_af_ops->net_header_len;
907 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
908 icsk->icsk_mtup.probe_size = 0;
911 /* Bound MSS / TSO packet size with the half of the window */
912 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
914 if (tp->max_window && pktsize > (tp->max_window >> 1))
915 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
920 /* This function synchronize snd mss to current pmtu/exthdr set.
922 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
923 for TCP options, but includes only bare TCP header.
925 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
926 It is minimum of user_mss and mss received with SYN.
927 It also does not include TCP options.
929 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
931 tp->mss_cache is current effective sending mss, including
932 all tcp options except for SACKs. It is evaluated,
933 taking into account current pmtu, but never exceeds
934 tp->rx_opt.mss_clamp.
936 NOTE1. rfc1122 clearly states that advertised MSS
937 DOES NOT include either tcp or ip options.
939 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
940 are READ ONLY outside this function. --ANK (980731)
942 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
944 struct tcp_sock *tp = tcp_sk(sk);
945 struct inet_connection_sock *icsk = inet_csk(sk);
948 if (icsk->icsk_mtup.search_high > pmtu)
949 icsk->icsk_mtup.search_high = pmtu;
951 mss_now = tcp_mtu_to_mss(sk, pmtu);
952 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
954 /* And store cached results */
955 icsk->icsk_pmtu_cookie = pmtu;
956 if (icsk->icsk_mtup.enabled)
957 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
958 tp->mss_cache = mss_now;
963 /* Compute the current effective MSS, taking SACKs and IP options,
964 * and even PMTU discovery events into account.
966 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up
967 * cannot be large. However, taking into account rare use of URG, this
970 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
972 struct tcp_sock *tp = tcp_sk(sk);
973 struct dst_entry *dst = __sk_dst_get(sk);
978 mss_now = tp->mss_cache;
980 if (large_allowed && sk_can_gso(sk) && !tp->urg_mode)
984 u32 mtu = dst_mtu(dst);
985 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
986 mss_now = tcp_sync_mss(sk, mtu);
989 if (tp->rx_opt.eff_sacks)
990 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED +
991 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK));
993 #ifdef CONFIG_TCP_MD5SIG
994 if (tp->af_specific->md5_lookup(sk, sk))
995 mss_now -= TCPOLEN_MD5SIG_ALIGNED;
998 xmit_size_goal = mss_now;
1001 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1002 inet_csk(sk)->icsk_af_ops->net_header_len -
1003 inet_csk(sk)->icsk_ext_hdr_len -
1004 tp->tcp_header_len);
1006 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1007 xmit_size_goal -= (xmit_size_goal % mss_now);
1009 tp->xmit_size_goal = xmit_size_goal;
1014 /* Congestion window validation. (RFC2861) */
1015 static void tcp_cwnd_validate(struct sock *sk)
1017 struct tcp_sock *tp = tcp_sk(sk);
1019 if (tp->packets_out >= tp->snd_cwnd) {
1020 /* Network is feed fully. */
1021 tp->snd_cwnd_used = 0;
1022 tp->snd_cwnd_stamp = tcp_time_stamp;
1024 /* Network starves. */
1025 if (tp->packets_out > tp->snd_cwnd_used)
1026 tp->snd_cwnd_used = tp->packets_out;
1028 if (sysctl_tcp_slow_start_after_idle &&
1029 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1030 tcp_cwnd_application_limited(sk);
1034 /* Returns the portion of skb which can be sent right away without
1035 * introducing MSS oddities to segment boundaries. In rare cases where
1036 * mss_now != mss_cache, we will request caller to create a small skb
1037 * per input skb which could be mostly avoided here (if desired).
1039 * We explicitly want to create a request for splitting write queue tail
1040 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1041 * thus all the complexity (cwnd_len is always MSS multiple which we
1042 * return whenever allowed by the other factors). Basically we need the
1043 * modulo only when the receiver window alone is the limiting factor or
1044 * when we would be allowed to send the split-due-to-Nagle skb fully.
1046 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1047 unsigned int mss_now, unsigned int cwnd)
1049 struct tcp_sock *tp = tcp_sk(sk);
1050 u32 needed, window, cwnd_len;
1052 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1053 cwnd_len = mss_now * cwnd;
1055 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1058 needed = min(skb->len, window);
1060 if (cwnd_len <= needed)
1063 return needed - needed % mss_now;
1066 /* Can at least one segment of SKB be sent right now, according to the
1067 * congestion window rules? If so, return how many segments are allowed.
1069 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1070 struct sk_buff *skb)
1072 u32 in_flight, cwnd;
1074 /* Don't be strict about the congestion window for the final FIN. */
1075 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1076 tcp_skb_pcount(skb) == 1)
1079 in_flight = tcp_packets_in_flight(tp);
1080 cwnd = tp->snd_cwnd;
1081 if (in_flight < cwnd)
1082 return (cwnd - in_flight);
1087 /* This must be invoked the first time we consider transmitting
1088 * SKB onto the wire.
1090 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1091 unsigned int mss_now)
1093 int tso_segs = tcp_skb_pcount(skb);
1095 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1096 tcp_set_skb_tso_segs(sk, skb, mss_now);
1097 tso_segs = tcp_skb_pcount(skb);
1102 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1104 return after(tp->snd_sml,tp->snd_una) &&
1105 !after(tp->snd_sml, tp->snd_nxt);
1108 /* Return 0, if packet can be sent now without violation Nagle's rules:
1109 * 1. It is full sized.
1110 * 2. Or it contains FIN. (already checked by caller)
1111 * 3. Or TCP_NODELAY was set.
1112 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1113 * With Minshall's modification: all sent small packets are ACKed.
1115 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1116 const struct sk_buff *skb,
1117 unsigned mss_now, int nonagle)
1119 return (skb->len < mss_now &&
1120 ((nonagle & TCP_NAGLE_CORK) ||
1121 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1124 /* Return non-zero if the Nagle test allows this packet to be
1127 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1128 unsigned int cur_mss, int nonagle)
1130 /* Nagle rule does not apply to frames, which sit in the middle of the
1131 * write_queue (they have no chances to get new data).
1133 * This is implemented in the callers, where they modify the 'nonagle'
1134 * argument based upon the location of SKB in the send queue.
1136 if (nonagle & TCP_NAGLE_PUSH)
1139 /* Don't use the nagle rule for urgent data (or for the final FIN).
1140 * Nagle can be ignored during F-RTO too (see RFC4138).
1142 if (tp->urg_mode || (tp->frto_counter == 2) ||
1143 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1146 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1152 /* Does at least the first segment of SKB fit into the send window? */
1153 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1154 unsigned int cur_mss)
1156 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1158 if (skb->len > cur_mss)
1159 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1161 return !after(end_seq, tcp_wnd_end(tp));
1164 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1165 * should be put on the wire right now. If so, it returns the number of
1166 * packets allowed by the congestion window.
1168 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1169 unsigned int cur_mss, int nonagle)
1171 struct tcp_sock *tp = tcp_sk(sk);
1172 unsigned int cwnd_quota;
1174 tcp_init_tso_segs(sk, skb, cur_mss);
1176 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1179 cwnd_quota = tcp_cwnd_test(tp, skb);
1180 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1186 int tcp_may_send_now(struct sock *sk)
1188 struct tcp_sock *tp = tcp_sk(sk);
1189 struct sk_buff *skb = tcp_send_head(sk);
1192 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1193 (tcp_skb_is_last(sk, skb) ?
1194 tp->nonagle : TCP_NAGLE_PUSH)));
1197 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1198 * which is put after SKB on the list. It is very much like
1199 * tcp_fragment() except that it may make several kinds of assumptions
1200 * in order to speed up the splitting operation. In particular, we
1201 * know that all the data is in scatter-gather pages, and that the
1202 * packet has never been sent out before (and thus is not cloned).
1204 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1205 unsigned int mss_now)
1207 struct sk_buff *buff;
1208 int nlen = skb->len - len;
1211 /* All of a TSO frame must be composed of paged data. */
1212 if (skb->len != skb->data_len)
1213 return tcp_fragment(sk, skb, len, mss_now);
1215 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1216 if (unlikely(buff == NULL))
1219 sk->sk_wmem_queued += buff->truesize;
1220 sk_mem_charge(sk, buff->truesize);
1221 buff->truesize += nlen;
1222 skb->truesize -= nlen;
1224 /* Correct the sequence numbers. */
1225 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1226 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1227 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1229 /* PSH and FIN should only be set in the second packet. */
1230 flags = TCP_SKB_CB(skb)->flags;
1231 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1232 TCP_SKB_CB(buff)->flags = flags;
1234 /* This packet was never sent out yet, so no SACK bits. */
1235 TCP_SKB_CB(buff)->sacked = 0;
1237 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1238 skb_split(skb, buff, len);
1240 /* Fix up tso_factor for both original and new SKB. */
1241 tcp_set_skb_tso_segs(sk, skb, mss_now);
1242 tcp_set_skb_tso_segs(sk, buff, mss_now);
1244 /* Link BUFF into the send queue. */
1245 skb_header_release(buff);
1246 tcp_insert_write_queue_after(skb, buff, sk);
1251 /* Try to defer sending, if possible, in order to minimize the amount
1252 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1254 * This algorithm is from John Heffner.
1256 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1258 struct tcp_sock *tp = tcp_sk(sk);
1259 const struct inet_connection_sock *icsk = inet_csk(sk);
1260 u32 send_win, cong_win, limit, in_flight;
1262 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1265 if (icsk->icsk_ca_state != TCP_CA_Open)
1268 /* Defer for less than two clock ticks. */
1269 if (tp->tso_deferred &&
1270 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1273 in_flight = tcp_packets_in_flight(tp);
1275 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1277 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1279 /* From in_flight test above, we know that cwnd > in_flight. */
1280 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1282 limit = min(send_win, cong_win);
1284 /* If a full-sized TSO skb can be sent, do it. */
1285 if (limit >= sk->sk_gso_max_size)
1288 if (sysctl_tcp_tso_win_divisor) {
1289 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1291 /* If at least some fraction of a window is available,
1294 chunk /= sysctl_tcp_tso_win_divisor;
1298 /* Different approach, try not to defer past a single
1299 * ACK. Receiver should ACK every other full sized
1300 * frame, so if we have space for more than 3 frames
1303 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1307 /* Ok, it looks like it is advisable to defer. */
1308 tp->tso_deferred = 1 | (jiffies << 1);
1313 tp->tso_deferred = 0;
1317 /* Create a new MTU probe if we are ready.
1318 * Returns 0 if we should wait to probe (no cwnd available),
1319 * 1 if a probe was sent,
1322 static int tcp_mtu_probe(struct sock *sk)
1324 struct tcp_sock *tp = tcp_sk(sk);
1325 struct inet_connection_sock *icsk = inet_csk(sk);
1326 struct sk_buff *skb, *nskb, *next;
1333 /* Not currently probing/verifying,
1335 * have enough cwnd, and
1336 * not SACKing (the variable headers throw things off) */
1337 if (!icsk->icsk_mtup.enabled ||
1338 icsk->icsk_mtup.probe_size ||
1339 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1340 tp->snd_cwnd < 11 ||
1341 tp->rx_opt.eff_sacks)
1344 /* Very simple search strategy: just double the MSS. */
1345 mss_now = tcp_current_mss(sk, 0);
1346 probe_size = 2 * tp->mss_cache;
1347 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1348 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1349 /* TODO: set timer for probe_converge_event */
1353 /* Have enough data in the send queue to probe? */
1354 if (tp->write_seq - tp->snd_nxt < size_needed)
1357 if (tp->snd_wnd < size_needed)
1359 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1362 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1363 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1364 if (!tcp_packets_in_flight(tp))
1370 /* We're allowed to probe. Build it now. */
1371 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1373 sk->sk_wmem_queued += nskb->truesize;
1374 sk_mem_charge(sk, nskb->truesize);
1376 skb = tcp_send_head(sk);
1378 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1379 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1380 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1381 TCP_SKB_CB(nskb)->sacked = 0;
1383 nskb->ip_summed = skb->ip_summed;
1385 tcp_insert_write_queue_before(nskb, skb, sk);
1388 tcp_for_write_queue_from_safe(skb, next, sk) {
1389 copy = min_t(int, skb->len, probe_size - len);
1390 if (nskb->ip_summed)
1391 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1393 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1394 skb_put(nskb, copy),
1397 if (skb->len <= copy) {
1398 /* We've eaten all the data from this skb.
1400 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1401 tcp_unlink_write_queue(skb, sk);
1402 sk_wmem_free_skb(sk, skb);
1404 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1405 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1406 if (!skb_shinfo(skb)->nr_frags) {
1407 skb_pull(skb, copy);
1408 if (skb->ip_summed != CHECKSUM_PARTIAL)
1409 skb->csum = csum_partial(skb->data,
1412 __pskb_trim_head(skb, copy);
1413 tcp_set_skb_tso_segs(sk, skb, mss_now);
1415 TCP_SKB_CB(skb)->seq += copy;
1420 if (len >= probe_size)
1423 tcp_init_tso_segs(sk, nskb, nskb->len);
1425 /* We're ready to send. If this fails, the probe will
1426 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1427 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1428 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1429 /* Decrement cwnd here because we are sending
1430 * effectively two packets. */
1432 tcp_event_new_data_sent(sk, nskb);
1434 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1435 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1436 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1444 /* This routine writes packets to the network. It advances the
1445 * send_head. This happens as incoming acks open up the remote
1448 * Returns 1, if no segments are in flight and we have queued segments, but
1449 * cannot send anything now because of SWS or another problem.
1451 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1453 struct tcp_sock *tp = tcp_sk(sk);
1454 struct sk_buff *skb;
1455 unsigned int tso_segs, sent_pkts;
1459 /* If we are closed, the bytes will have to remain here.
1460 * In time closedown will finish, we empty the write queue and all
1463 if (unlikely(sk->sk_state == TCP_CLOSE))
1468 /* Do MTU probing. */
1469 if ((result = tcp_mtu_probe(sk)) == 0) {
1471 } else if (result > 0) {
1475 while ((skb = tcp_send_head(sk))) {
1478 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1481 cwnd_quota = tcp_cwnd_test(tp, skb);
1485 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1488 if (tso_segs == 1) {
1489 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1490 (tcp_skb_is_last(sk, skb) ?
1491 nonagle : TCP_NAGLE_PUSH))))
1494 if (tcp_tso_should_defer(sk, skb))
1500 limit = tcp_mss_split_point(sk, skb, mss_now,
1503 if (skb->len > limit &&
1504 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1507 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1509 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1512 /* Advance the send_head. This one is sent out.
1513 * This call will increment packets_out.
1515 tcp_event_new_data_sent(sk, skb);
1517 tcp_minshall_update(tp, mss_now, skb);
1521 if (likely(sent_pkts)) {
1522 tcp_cwnd_validate(sk);
1525 return !tp->packets_out && tcp_send_head(sk);
1528 /* Push out any pending frames which were held back due to
1529 * TCP_CORK or attempt at coalescing tiny packets.
1530 * The socket must be locked by the caller.
1532 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1535 struct sk_buff *skb = tcp_send_head(sk);
1538 if (tcp_write_xmit(sk, cur_mss, nonagle))
1539 tcp_check_probe_timer(sk);
1543 /* Send _single_ skb sitting at the send head. This function requires
1544 * true push pending frames to setup probe timer etc.
1546 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1548 struct sk_buff *skb = tcp_send_head(sk);
1549 unsigned int tso_segs, cwnd_quota;
1551 BUG_ON(!skb || skb->len < mss_now);
1553 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1554 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1556 if (likely(cwnd_quota)) {
1563 limit = tcp_mss_split_point(sk, skb, mss_now,
1566 if (skb->len > limit &&
1567 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1570 /* Send it out now. */
1571 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1573 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1574 tcp_event_new_data_sent(sk, skb);
1575 tcp_cwnd_validate(sk);
1581 /* This function returns the amount that we can raise the
1582 * usable window based on the following constraints
1584 * 1. The window can never be shrunk once it is offered (RFC 793)
1585 * 2. We limit memory per socket
1588 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1589 * RECV.NEXT + RCV.WIN fixed until:
1590 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1592 * i.e. don't raise the right edge of the window until you can raise
1593 * it at least MSS bytes.
1595 * Unfortunately, the recommended algorithm breaks header prediction,
1596 * since header prediction assumes th->window stays fixed.
1598 * Strictly speaking, keeping th->window fixed violates the receiver
1599 * side SWS prevention criteria. The problem is that under this rule
1600 * a stream of single byte packets will cause the right side of the
1601 * window to always advance by a single byte.
1603 * Of course, if the sender implements sender side SWS prevention
1604 * then this will not be a problem.
1606 * BSD seems to make the following compromise:
1608 * If the free space is less than the 1/4 of the maximum
1609 * space available and the free space is less than 1/2 mss,
1610 * then set the window to 0.
1611 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1612 * Otherwise, just prevent the window from shrinking
1613 * and from being larger than the largest representable value.
1615 * This prevents incremental opening of the window in the regime
1616 * where TCP is limited by the speed of the reader side taking
1617 * data out of the TCP receive queue. It does nothing about
1618 * those cases where the window is constrained on the sender side
1619 * because the pipeline is full.
1621 * BSD also seems to "accidentally" limit itself to windows that are a
1622 * multiple of MSS, at least until the free space gets quite small.
1623 * This would appear to be a side effect of the mbuf implementation.
1624 * Combining these two algorithms results in the observed behavior
1625 * of having a fixed window size at almost all times.
1627 * Below we obtain similar behavior by forcing the offered window to
1628 * a multiple of the mss when it is feasible to do so.
1630 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1631 * Regular options like TIMESTAMP are taken into account.
1633 u32 __tcp_select_window(struct sock *sk)
1635 struct inet_connection_sock *icsk = inet_csk(sk);
1636 struct tcp_sock *tp = tcp_sk(sk);
1637 /* MSS for the peer's data. Previous versions used mss_clamp
1638 * here. I don't know if the value based on our guesses
1639 * of peer's MSS is better for the performance. It's more correct
1640 * but may be worse for the performance because of rcv_mss
1641 * fluctuations. --SAW 1998/11/1
1643 int mss = icsk->icsk_ack.rcv_mss;
1644 int free_space = tcp_space(sk);
1645 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1648 if (mss > full_space)
1651 if (free_space < (full_space >> 1)) {
1652 icsk->icsk_ack.quick = 0;
1654 if (tcp_memory_pressure)
1655 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1658 if (free_space < mss)
1662 if (free_space > tp->rcv_ssthresh)
1663 free_space = tp->rcv_ssthresh;
1665 /* Don't do rounding if we are using window scaling, since the
1666 * scaled window will not line up with the MSS boundary anyway.
1668 window = tp->rcv_wnd;
1669 if (tp->rx_opt.rcv_wscale) {
1670 window = free_space;
1672 /* Advertise enough space so that it won't get scaled away.
1673 * Import case: prevent zero window announcement if
1674 * 1<<rcv_wscale > mss.
1676 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1677 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1678 << tp->rx_opt.rcv_wscale);
1680 /* Get the largest window that is a nice multiple of mss.
1681 * Window clamp already applied above.
1682 * If our current window offering is within 1 mss of the
1683 * free space we just keep it. This prevents the divide
1684 * and multiply from happening most of the time.
1685 * We also don't do any window rounding when the free space
1688 if (window <= free_space - mss || window > free_space)
1689 window = (free_space / mss) * mss;
1690 else if (mss == full_space &&
1691 free_space > window + (full_space >> 1))
1692 window = free_space;
1698 /* Attempt to collapse two adjacent SKB's during retransmission. */
1699 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1702 struct tcp_sock *tp = tcp_sk(sk);
1703 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1704 int skb_size, next_skb_size;
1707 /* The first test we must make is that neither of these two
1708 * SKB's are still referenced by someone else.
1710 if (skb_cloned(skb) || skb_cloned(next_skb))
1713 skb_size = skb->len;
1714 next_skb_size = next_skb->len;
1715 flags = TCP_SKB_CB(skb)->flags;
1717 /* Also punt if next skb has been SACK'd. */
1718 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1721 /* Next skb is out of window. */
1722 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1725 /* Punt if not enough space exists in the first SKB for
1726 * the data in the second, or the total combined payload
1727 * would exceed the MSS.
1729 if ((next_skb_size > skb_tailroom(skb)) ||
1730 ((skb_size + next_skb_size) > mss_now))
1733 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1735 tcp_highest_sack_combine(sk, next_skb, skb);
1737 /* Ok. We will be able to collapse the packet. */
1738 tcp_unlink_write_queue(next_skb, sk);
1740 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1743 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1744 skb->ip_summed = CHECKSUM_PARTIAL;
1746 if (skb->ip_summed != CHECKSUM_PARTIAL)
1747 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1749 /* Update sequence range on original skb. */
1750 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1752 /* Merge over control information. */
1753 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1754 TCP_SKB_CB(skb)->flags = flags;
1756 /* All done, get rid of second SKB and account for it so
1757 * packet counting does not break.
1759 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1760 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1761 tp->retrans_out -= tcp_skb_pcount(next_skb);
1762 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1763 tp->lost_out -= tcp_skb_pcount(next_skb);
1764 /* Reno case is special. Sigh... */
1765 if (tcp_is_reno(tp) && tp->sacked_out)
1766 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1768 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1769 tp->packets_out -= tcp_skb_pcount(next_skb);
1771 /* changed transmit queue under us so clear hints */
1772 tcp_clear_retrans_hints_partial(tp);
1774 sk_wmem_free_skb(sk, next_skb);
1777 /* Do a simple retransmit without using the backoff mechanisms in
1778 * tcp_timer. This is used for path mtu discovery.
1779 * The socket is already locked here.
1781 void tcp_simple_retransmit(struct sock *sk)
1783 const struct inet_connection_sock *icsk = inet_csk(sk);
1784 struct tcp_sock *tp = tcp_sk(sk);
1785 struct sk_buff *skb;
1786 unsigned int mss = tcp_current_mss(sk, 0);
1789 tcp_for_write_queue(skb, sk) {
1790 if (skb == tcp_send_head(sk))
1792 if (skb->len > mss &&
1793 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1794 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1795 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1796 tp->retrans_out -= tcp_skb_pcount(skb);
1798 if (!(TCP_SKB_CB(skb)->sacked & TCPCB_LOST)) {
1799 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1800 tp->lost_out += tcp_skb_pcount(skb);
1806 tcp_clear_all_retrans_hints(tp);
1811 if (tcp_is_reno(tp))
1812 tcp_limit_reno_sacked(tp);
1814 tcp_verify_left_out(tp);
1816 /* Don't muck with the congestion window here.
1817 * Reason is that we do not increase amount of _data_
1818 * in network, but units changed and effective
1819 * cwnd/ssthresh really reduced now.
1821 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1822 tp->high_seq = tp->snd_nxt;
1823 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1824 tp->prior_ssthresh = 0;
1825 tp->undo_marker = 0;
1826 tcp_set_ca_state(sk, TCP_CA_Loss);
1828 tcp_xmit_retransmit_queue(sk);
1831 /* This retransmits one SKB. Policy decisions and retransmit queue
1832 * state updates are done by the caller. Returns non-zero if an
1833 * error occurred which prevented the send.
1835 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1837 struct tcp_sock *tp = tcp_sk(sk);
1838 struct inet_connection_sock *icsk = inet_csk(sk);
1839 unsigned int cur_mss = tcp_current_mss(sk, 0);
1842 /* Inconslusive MTU probe */
1843 if (icsk->icsk_mtup.probe_size) {
1844 icsk->icsk_mtup.probe_size = 0;
1847 /* Do not sent more than we queued. 1/4 is reserved for possible
1848 * copying overhead: fragmentation, tunneling, mangling etc.
1850 if (atomic_read(&sk->sk_wmem_alloc) >
1851 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1854 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1855 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1857 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1861 /* If receiver has shrunk his window, and skb is out of
1862 * new window, do not retransmit it. The exception is the
1863 * case, when window is shrunk to zero. In this case
1864 * our retransmit serves as a zero window probe.
1866 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1867 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1870 if (skb->len > cur_mss) {
1871 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1872 return -ENOMEM; /* We'll try again later. */
1875 /* Collapse two adjacent packets if worthwhile and we can. */
1876 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1877 (skb->len < (cur_mss >> 1)) &&
1878 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1879 (!tcp_skb_is_last(sk, skb)) &&
1880 (skb_shinfo(skb)->nr_frags == 0 &&
1881 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1882 (tcp_skb_pcount(skb) == 1 &&
1883 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1884 (sysctl_tcp_retrans_collapse != 0))
1885 tcp_retrans_try_collapse(sk, skb, cur_mss);
1887 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1888 return -EHOSTUNREACH; /* Routing failure or similar. */
1890 /* Some Solaris stacks overoptimize and ignore the FIN on a
1891 * retransmit when old data is attached. So strip it off
1892 * since it is cheap to do so and saves bytes on the network.
1895 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1896 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1897 if (!pskb_trim(skb, 0)) {
1898 /* Reuse, even though it does some unnecessary work */
1899 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1900 TCP_SKB_CB(skb)->flags);
1901 skb->ip_summed = CHECKSUM_NONE;
1905 /* Make a copy, if the first transmission SKB clone we made
1906 * is still in somebody's hands, else make a clone.
1908 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1910 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1913 /* Update global TCP statistics. */
1914 TCP_INC_STATS(TCP_MIB_RETRANSSEGS);
1916 tp->total_retrans++;
1918 #if FASTRETRANS_DEBUG > 0
1919 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1920 if (net_ratelimit())
1921 printk(KERN_DEBUG "retrans_out leaked.\n");
1924 if (!tp->retrans_out)
1925 tp->lost_retrans_low = tp->snd_nxt;
1926 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1927 tp->retrans_out += tcp_skb_pcount(skb);
1929 /* Save stamp of the first retransmit. */
1930 if (!tp->retrans_stamp)
1931 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1935 /* snd_nxt is stored to detect loss of retransmitted segment,
1936 * see tcp_input.c tcp_sacktag_write_queue().
1938 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1943 /* This gets called after a retransmit timeout, and the initially
1944 * retransmitted data is acknowledged. It tries to continue
1945 * resending the rest of the retransmit queue, until either
1946 * we've sent it all or the congestion window limit is reached.
1947 * If doing SACK, the first ACK which comes back for a timeout
1948 * based retransmit packet might feed us FACK information again.
1949 * If so, we use it to avoid unnecessarily retransmissions.
1951 void tcp_xmit_retransmit_queue(struct sock *sk)
1953 const struct inet_connection_sock *icsk = inet_csk(sk);
1954 struct tcp_sock *tp = tcp_sk(sk);
1955 struct sk_buff *skb;
1958 if (tp->retransmit_skb_hint) {
1959 skb = tp->retransmit_skb_hint;
1960 packet_cnt = tp->retransmit_cnt_hint;
1962 skb = tcp_write_queue_head(sk);
1966 /* First pass: retransmit lost packets. */
1968 tcp_for_write_queue_from(skb, sk) {
1969 __u8 sacked = TCP_SKB_CB(skb)->sacked;
1971 if (skb == tcp_send_head(sk))
1973 /* we could do better than to assign each time */
1974 tp->retransmit_skb_hint = skb;
1975 tp->retransmit_cnt_hint = packet_cnt;
1977 /* Assume this retransmit will generate
1978 * only one packet for congestion window
1979 * calculation purposes. This works because
1980 * tcp_retransmit_skb() will chop up the
1981 * packet to be MSS sized and all the
1982 * packet counting works out.
1984 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
1987 if (sacked & TCPCB_LOST) {
1988 if (!(sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) {
1989 if (tcp_retransmit_skb(sk, skb)) {
1990 tp->retransmit_skb_hint = NULL;
1993 if (icsk->icsk_ca_state != TCP_CA_Loss)
1994 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS);
1996 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS);
1998 if (skb == tcp_write_queue_head(sk))
1999 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2000 inet_csk(sk)->icsk_rto,
2004 packet_cnt += tcp_skb_pcount(skb);
2005 if (packet_cnt >= tp->lost_out)
2011 /* OK, demanded retransmission is finished. */
2013 /* Forward retransmissions are possible only during Recovery. */
2014 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2017 /* No forward retransmissions in Reno are possible. */
2018 if (tcp_is_reno(tp))
2021 /* Yeah, we have to make difficult choice between forward transmission
2022 * and retransmission... Both ways have their merits...
2024 * For now we do not retransmit anything, while we have some new
2025 * segments to send. In the other cases, follow rule 3 for
2026 * NextSeg() specified in RFC3517.
2029 if (tcp_may_send_now(sk))
2032 /* If nothing is SACKed, highest_sack in the loop won't be valid */
2033 if (!tp->sacked_out)
2036 if (tp->forward_skb_hint)
2037 skb = tp->forward_skb_hint;
2039 skb = tcp_write_queue_head(sk);
2041 tcp_for_write_queue_from(skb, sk) {
2042 if (skb == tcp_send_head(sk))
2044 tp->forward_skb_hint = skb;
2046 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2049 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2052 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS)
2055 /* Ok, retransmit it. */
2056 if (tcp_retransmit_skb(sk, skb)) {
2057 tp->forward_skb_hint = NULL;
2061 if (skb == tcp_write_queue_head(sk))
2062 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2063 inet_csk(sk)->icsk_rto,
2066 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS);
2070 /* Send a fin. The caller locks the socket for us. This cannot be
2071 * allowed to fail queueing a FIN frame under any circumstances.
2073 void tcp_send_fin(struct sock *sk)
2075 struct tcp_sock *tp = tcp_sk(sk);
2076 struct sk_buff *skb = tcp_write_queue_tail(sk);
2079 /* Optimization, tack on the FIN if we have a queue of
2080 * unsent frames. But be careful about outgoing SACKS
2083 mss_now = tcp_current_mss(sk, 1);
2085 if (tcp_send_head(sk) != NULL) {
2086 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2087 TCP_SKB_CB(skb)->end_seq++;
2090 /* Socket is locked, keep trying until memory is available. */
2092 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2098 /* Reserve space for headers and prepare control bits. */
2099 skb_reserve(skb, MAX_TCP_HEADER);
2100 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2101 tcp_init_nondata_skb(skb, tp->write_seq,
2102 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2103 tcp_queue_skb(sk, skb);
2105 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2108 /* We get here when a process closes a file descriptor (either due to
2109 * an explicit close() or as a byproduct of exit()'ing) and there
2110 * was unread data in the receive queue. This behavior is recommended
2111 * by RFC 2525, section 2.17. -DaveM
2113 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2115 struct sk_buff *skb;
2117 /* NOTE: No TCP options attached and we never retransmit this. */
2118 skb = alloc_skb(MAX_TCP_HEADER, priority);
2120 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2124 /* Reserve space for headers and prepare control bits. */
2125 skb_reserve(skb, MAX_TCP_HEADER);
2126 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2127 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2129 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2130 if (tcp_transmit_skb(sk, skb, 0, priority))
2131 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED);
2134 /* WARNING: This routine must only be called when we have already sent
2135 * a SYN packet that crossed the incoming SYN that caused this routine
2136 * to get called. If this assumption fails then the initial rcv_wnd
2137 * and rcv_wscale values will not be correct.
2139 int tcp_send_synack(struct sock *sk)
2141 struct sk_buff *skb;
2143 skb = tcp_write_queue_head(sk);
2144 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2145 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2148 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2149 if (skb_cloned(skb)) {
2150 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2153 tcp_unlink_write_queue(skb, sk);
2154 skb_header_release(nskb);
2155 __tcp_add_write_queue_head(sk, nskb);
2156 sk_wmem_free_skb(sk, skb);
2157 sk->sk_wmem_queued += nskb->truesize;
2158 sk_mem_charge(sk, nskb->truesize);
2162 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2163 TCP_ECN_send_synack(tcp_sk(sk), skb);
2165 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2166 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2170 * Prepare a SYN-ACK.
2172 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2173 struct request_sock *req)
2175 struct inet_request_sock *ireq = inet_rsk(req);
2176 struct tcp_sock *tp = tcp_sk(sk);
2178 int tcp_header_size;
2179 struct sk_buff *skb;
2180 #ifdef CONFIG_TCP_MD5SIG
2181 struct tcp_md5sig_key *md5;
2182 __u8 *md5_hash_location;
2185 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2189 /* Reserve space for headers. */
2190 skb_reserve(skb, MAX_TCP_HEADER);
2192 skb->dst = dst_clone(dst);
2194 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS +
2195 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) +
2196 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) +
2197 /* SACK_PERM is in the place of NOP NOP of TS */
2198 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0));
2200 #ifdef CONFIG_TCP_MD5SIG
2201 /* Are we doing MD5 on this segment? If so - make room for it */
2202 md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
2204 tcp_header_size += TCPOLEN_MD5SIG_ALIGNED;
2206 skb_push(skb, tcp_header_size);
2207 skb_reset_transport_header(skb);
2210 memset(th, 0, sizeof(struct tcphdr));
2213 TCP_ECN_make_synack(req, th);
2214 th->source = inet_sk(sk)->sport;
2215 th->dest = ireq->rmt_port;
2216 /* Setting of flags are superfluous here for callers (and ECE is
2217 * not even correctly set)
2219 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2220 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2221 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2222 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2223 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2225 /* Set this up on the first call only */
2226 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2227 /* tcp_full_space because it is guaranteed to be the first packet */
2228 tcp_select_initial_window(tcp_full_space(sk),
2229 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2234 ireq->rcv_wscale = rcv_wscale;
2237 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2238 th->window = htons(min(req->rcv_wnd, 65535U));
2239 #ifdef CONFIG_SYN_COOKIES
2240 if (unlikely(req->cookie_ts))
2241 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2244 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2245 tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok,
2246 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale,
2247 TCP_SKB_CB(skb)->when,
2250 #ifdef CONFIG_TCP_MD5SIG
2251 md5 ? &md5_hash_location :
2256 th->doff = (tcp_header_size >> 2);
2257 TCP_INC_STATS(TCP_MIB_OUTSEGS);
2259 #ifdef CONFIG_TCP_MD5SIG
2260 /* Okay, we have all we need - do the md5 hash if needed */
2262 tp->af_specific->calc_md5_hash(md5_hash_location,
2265 tcp_hdr(skb), sk->sk_protocol,
2274 * Do all connect socket setups that can be done AF independent.
2276 static void tcp_connect_init(struct sock *sk)
2278 struct dst_entry *dst = __sk_dst_get(sk);
2279 struct tcp_sock *tp = tcp_sk(sk);
2282 /* We'll fix this up when we get a response from the other end.
2283 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2285 tp->tcp_header_len = sizeof(struct tcphdr) +
2286 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2288 #ifdef CONFIG_TCP_MD5SIG
2289 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2290 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2293 /* If user gave his TCP_MAXSEG, record it to clamp */
2294 if (tp->rx_opt.user_mss)
2295 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2298 tcp_sync_mss(sk, dst_mtu(dst));
2300 if (!tp->window_clamp)
2301 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2302 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2303 tcp_initialize_rcv_mss(sk);
2305 tcp_select_initial_window(tcp_full_space(sk),
2306 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2309 sysctl_tcp_window_scaling,
2312 tp->rx_opt.rcv_wscale = rcv_wscale;
2313 tp->rcv_ssthresh = tp->rcv_wnd;
2316 sock_reset_flag(sk, SOCK_DONE);
2318 tcp_init_wl(tp, tp->write_seq, 0);
2319 tp->snd_una = tp->write_seq;
2320 tp->snd_sml = tp->write_seq;
2325 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2326 inet_csk(sk)->icsk_retransmits = 0;
2327 tcp_clear_retrans(tp);
2331 * Build a SYN and send it off.
2333 int tcp_connect(struct sock *sk)
2335 struct tcp_sock *tp = tcp_sk(sk);
2336 struct sk_buff *buff;
2338 tcp_connect_init(sk);
2340 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2341 if (unlikely(buff == NULL))
2344 /* Reserve space for headers. */
2345 skb_reserve(buff, MAX_TCP_HEADER);
2347 tp->snd_nxt = tp->write_seq;
2348 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2349 TCP_ECN_send_syn(sk, buff);
2352 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2353 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2354 skb_header_release(buff);
2355 __tcp_add_write_queue_tail(sk, buff);
2356 sk->sk_wmem_queued += buff->truesize;
2357 sk_mem_charge(sk, buff->truesize);
2358 tp->packets_out += tcp_skb_pcount(buff);
2359 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2361 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2362 * in order to make this packet get counted in tcpOutSegs.
2364 tp->snd_nxt = tp->write_seq;
2365 tp->pushed_seq = tp->write_seq;
2366 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS);
2368 /* Timer for repeating the SYN until an answer. */
2369 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2370 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2374 /* Send out a delayed ack, the caller does the policy checking
2375 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2378 void tcp_send_delayed_ack(struct sock *sk)
2380 struct inet_connection_sock *icsk = inet_csk(sk);
2381 int ato = icsk->icsk_ack.ato;
2382 unsigned long timeout;
2384 if (ato > TCP_DELACK_MIN) {
2385 const struct tcp_sock *tp = tcp_sk(sk);
2386 int max_ato = HZ / 2;
2388 if (icsk->icsk_ack.pingpong ||
2389 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2390 max_ato = TCP_DELACK_MAX;
2392 /* Slow path, intersegment interval is "high". */
2394 /* If some rtt estimate is known, use it to bound delayed ack.
2395 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2399 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2405 ato = min(ato, max_ato);
2408 /* Stay within the limit we were given */
2409 timeout = jiffies + ato;
2411 /* Use new timeout only if there wasn't a older one earlier. */
2412 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2413 /* If delack timer was blocked or is about to expire,
2416 if (icsk->icsk_ack.blocked ||
2417 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2422 if (!time_before(timeout, icsk->icsk_ack.timeout))
2423 timeout = icsk->icsk_ack.timeout;
2425 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2426 icsk->icsk_ack.timeout = timeout;
2427 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2430 /* This routine sends an ack and also updates the window. */
2431 void tcp_send_ack(struct sock *sk)
2433 struct sk_buff *buff;
2435 /* If we have been reset, we may not send again. */
2436 if (sk->sk_state == TCP_CLOSE)
2439 /* We are not putting this on the write queue, so
2440 * tcp_transmit_skb() will set the ownership to this
2443 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2445 inet_csk_schedule_ack(sk);
2446 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2447 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2448 TCP_DELACK_MAX, TCP_RTO_MAX);
2452 /* Reserve space for headers and prepare control bits. */
2453 skb_reserve(buff, MAX_TCP_HEADER);
2454 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2456 /* Send it off, this clears delayed acks for us. */
2457 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2458 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2461 /* This routine sends a packet with an out of date sequence
2462 * number. It assumes the other end will try to ack it.
2464 * Question: what should we make while urgent mode?
2465 * 4.4BSD forces sending single byte of data. We cannot send
2466 * out of window data, because we have SND.NXT==SND.MAX...
2468 * Current solution: to send TWO zero-length segments in urgent mode:
2469 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2470 * out-of-date with SND.UNA-1 to probe window.
2472 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2474 struct tcp_sock *tp = tcp_sk(sk);
2475 struct sk_buff *skb;
2477 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2478 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2482 /* Reserve space for headers and set control bits. */
2483 skb_reserve(skb, MAX_TCP_HEADER);
2484 /* Use a previous sequence. This should cause the other
2485 * end to send an ack. Don't queue or clone SKB, just
2488 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
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 struct tcp_sock *tp = tcp_sk(sk);
2496 struct sk_buff *skb;
2498 if (sk->sk_state == TCP_CLOSE)
2501 if ((skb = tcp_send_head(sk)) != NULL &&
2502 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2504 unsigned int mss = tcp_current_mss(sk, 0);
2505 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2507 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2508 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2510 /* We are probing the opening of a window
2511 * but the window size is != 0
2512 * must have been a result SWS avoidance ( sender )
2514 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2516 seg_size = min(seg_size, mss);
2517 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2518 if (tcp_fragment(sk, skb, seg_size, mss))
2520 } else if (!tcp_skb_pcount(skb))
2521 tcp_set_skb_tso_segs(sk, skb, mss);
2523 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2524 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2525 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2527 tcp_event_new_data_sent(sk, skb);
2531 between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2532 tcp_xmit_probe_skb(sk, 1);
2533 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_select_initial_window);
2579 EXPORT_SYMBOL(tcp_connect);
2580 EXPORT_SYMBOL(tcp_make_synack);
2581 EXPORT_SYMBOL(tcp_simple_retransmit);
2582 EXPORT_SYMBOL(tcp_sync_mss);
2583 EXPORT_SYMBOL(tcp_mtup_init);