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).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
39 #include <linux/compiler.h>
40 #include <linux/module.h>
42 /* People can turn this off for buggy TCP's found in printers etc. */
43 int sysctl_tcp_retrans_collapse __read_mostly = 1;
45 /* People can turn this on to work with those rare, broken TCPs that
46 * interpret the window field as a signed quantity.
48 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
50 /* This limits the percentage of the congestion window which we
51 * will allow a single TSO frame to consume. Building TSO frames
52 * which are too large can cause TCP streams to be bursty.
54 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
56 int sysctl_tcp_mtu_probing __read_mostly = 0;
57 int sysctl_tcp_base_mss __read_mostly = 512;
59 /* By default, RFC2861 behavior. */
60 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
62 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
64 struct tcp_sock *tp = tcp_sk(sk);
65 unsigned int prior_packets = tp->packets_out;
67 tcp_advance_send_head(sk, skb);
68 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
70 /* Don't override Nagle indefinately with F-RTO */
71 if (tp->frto_counter == 2)
74 tp->packets_out += tcp_skb_pcount(skb);
76 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
77 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
80 /* SND.NXT, if window was not shrunk.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
86 static inline __u32 tcp_acceptable_seq(struct sock *sk)
88 struct tcp_sock *tp = tcp_sk(sk);
90 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
93 return tcp_wnd_end(tp);
96 /* Calculate mss to advertise in SYN segment.
97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
99 * 1. It is independent of path mtu.
100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
102 * attached devices, because some buggy hosts are confused by
104 * 4. We do not make 3, we advertise MSS, calculated from first
105 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
106 * This may be overridden via information stored in routing table.
107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
108 * probably even Jumbo".
110 static __u16 tcp_advertise_mss(struct sock *sk)
112 struct tcp_sock *tp = tcp_sk(sk);
113 struct dst_entry *dst = __sk_dst_get(sk);
114 int mss = tp->advmss;
116 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
117 mss = dst_metric(dst, RTAX_ADVMSS);
124 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
125 * This is the first part of cwnd validation mechanism. */
126 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
128 struct tcp_sock *tp = tcp_sk(sk);
129 s32 delta = tcp_time_stamp - tp->lsndtime;
130 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
131 u32 cwnd = tp->snd_cwnd;
133 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
135 tp->snd_ssthresh = tcp_current_ssthresh(sk);
136 restart_cwnd = min(restart_cwnd, cwnd);
138 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
140 tp->snd_cwnd = max(cwnd, restart_cwnd);
141 tp->snd_cwnd_stamp = tcp_time_stamp;
142 tp->snd_cwnd_used = 0;
145 static void tcp_event_data_sent(struct tcp_sock *tp,
146 struct sk_buff *skb, struct sock *sk)
148 struct inet_connection_sock *icsk = inet_csk(sk);
149 const u32 now = tcp_time_stamp;
151 if (sysctl_tcp_slow_start_after_idle &&
152 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
153 tcp_cwnd_restart(sk, __sk_dst_get(sk));
157 /* If it is a reply for ato after last received
158 * packet, enter pingpong mode.
160 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
161 icsk->icsk_ack.pingpong = 1;
164 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
166 tcp_dec_quickack_mode(sk, pkts);
167 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
170 /* Determine a window scaling and initial window to offer.
171 * Based on the assumption that the given amount of space
172 * will be offered. Store the results in the tp structure.
173 * NOTE: for smooth operation initial space offering should
174 * be a multiple of mss if possible. We assume here that mss >= 1.
175 * This MUST be enforced by all callers.
177 void tcp_select_initial_window(int __space, __u32 mss,
178 __u32 *rcv_wnd, __u32 *window_clamp,
179 int wscale_ok, __u8 *rcv_wscale)
181 unsigned int space = (__space < 0 ? 0 : __space);
183 /* If no clamp set the clamp to the max possible scaled window */
184 if (*window_clamp == 0)
185 (*window_clamp) = (65535 << 14);
186 space = min(*window_clamp, space);
188 /* Quantize space offering to a multiple of mss if possible. */
190 space = (space / mss) * mss;
192 /* NOTE: offering an initial window larger than 32767
193 * will break some buggy TCP stacks. If the admin tells us
194 * it is likely we could be speaking with such a buggy stack
195 * we will truncate our initial window offering to 32K-1
196 * unless the remote has sent us a window scaling option,
197 * which we interpret as a sign the remote TCP is not
198 * misinterpreting the window field as a signed quantity.
200 if (sysctl_tcp_workaround_signed_windows)
201 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
207 /* Set window scaling on max possible window
208 * See RFC1323 for an explanation of the limit to 14
210 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
211 space = min_t(u32, space, *window_clamp);
212 while (space > 65535 && (*rcv_wscale) < 14) {
218 /* Set initial window to value enough for senders,
219 * following RFC2414. Senders, not following this RFC,
220 * will be satisfied with 2.
222 if (mss > (1 << *rcv_wscale)) {
228 if (*rcv_wnd > init_cwnd * mss)
229 *rcv_wnd = init_cwnd * mss;
232 /* Set the clamp no higher than max representable value */
233 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
236 /* Chose a new window to advertise, update state in tcp_sock for the
237 * socket, and return result with RFC1323 scaling applied. The return
238 * value can be stuffed directly into th->window for an outgoing
241 static u16 tcp_select_window(struct sock *sk)
243 struct tcp_sock *tp = tcp_sk(sk);
244 u32 cur_win = tcp_receive_window(tp);
245 u32 new_win = __tcp_select_window(sk);
247 /* Never shrink the offered window */
248 if (new_win < cur_win) {
249 /* Danger Will Robinson!
250 * Don't update rcv_wup/rcv_wnd here or else
251 * we will not be able to advertise a zero
252 * window in time. --DaveM
254 * Relax Will Robinson.
256 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
258 tp->rcv_wnd = new_win;
259 tp->rcv_wup = tp->rcv_nxt;
261 /* Make sure we do not exceed the maximum possible
264 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
265 new_win = min(new_win, MAX_TCP_WINDOW);
267 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
269 /* RFC1323 scaling applied */
270 new_win >>= tp->rx_opt.rcv_wscale;
272 /* If we advertise zero window, disable fast path. */
279 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
281 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR;
282 if (!(tp->ecn_flags & TCP_ECN_OK))
283 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE;
286 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
288 struct tcp_sock *tp = tcp_sk(sk);
291 if (sysctl_tcp_ecn) {
292 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR;
293 tp->ecn_flags = TCP_ECN_OK;
297 static __inline__ void
298 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
300 if (inet_rsk(req)->ecn_ok)
304 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
307 struct tcp_sock *tp = tcp_sk(sk);
309 if (tp->ecn_flags & TCP_ECN_OK) {
310 /* Not-retransmitted data segment: set ECT and inject CWR. */
311 if (skb->len != tcp_header_len &&
312 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
314 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
315 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
316 tcp_hdr(skb)->cwr = 1;
317 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
320 /* ACK or retransmitted segment: clear ECT|CE */
321 INET_ECN_dontxmit(sk);
323 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
324 tcp_hdr(skb)->ece = 1;
328 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
329 * auto increment end seqno.
331 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
335 TCP_SKB_CB(skb)->flags = flags;
336 TCP_SKB_CB(skb)->sacked = 0;
338 skb_shinfo(skb)->gso_segs = 1;
339 skb_shinfo(skb)->gso_size = 0;
340 skb_shinfo(skb)->gso_type = 0;
342 TCP_SKB_CB(skb)->seq = seq;
343 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN))
345 TCP_SKB_CB(skb)->end_seq = seq;
348 static inline int tcp_urg_mode(const struct tcp_sock *tp)
350 return tp->snd_una != tp->snd_up;
353 #define OPTION_SACK_ADVERTISE (1 << 0)
354 #define OPTION_TS (1 << 1)
355 #define OPTION_MD5 (1 << 2)
357 struct tcp_out_options {
358 u8 options; /* bit field of OPTION_* */
359 u8 ws; /* window scale, 0 to disable */
360 u8 num_sack_blocks; /* number of SACK blocks to include */
361 u16 mss; /* 0 to disable */
362 __u32 tsval, tsecr; /* need to include OPTION_TS */
365 /* Beware: Something in the Internet is very sensitive to the ordering of
366 * TCP options, we learned this through the hard way, so be careful here.
367 * Luckily we can at least blame others for their non-compliance but from
368 * inter-operatibility perspective it seems that we're somewhat stuck with
369 * the ordering which we have been using if we want to keep working with
370 * those broken things (not that it currently hurts anybody as there isn't
371 * particular reason why the ordering would need to be changed).
373 * At least SACK_PERM as the first option is known to lead to a disaster
374 * (but it may well be that other scenarios fail similarly).
376 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
377 const struct tcp_out_options *opts,
379 if (unlikely(OPTION_MD5 & opts->options)) {
380 *ptr++ = htonl((TCPOPT_NOP << 24) |
382 (TCPOPT_MD5SIG << 8) |
384 *md5_hash = (__u8 *)ptr;
390 if (unlikely(opts->mss)) {
391 *ptr++ = htonl((TCPOPT_MSS << 24) |
392 (TCPOLEN_MSS << 16) |
396 if (likely(OPTION_TS & opts->options)) {
397 if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) {
398 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
399 (TCPOLEN_SACK_PERM << 16) |
400 (TCPOPT_TIMESTAMP << 8) |
403 *ptr++ = htonl((TCPOPT_NOP << 24) |
405 (TCPOPT_TIMESTAMP << 8) |
408 *ptr++ = htonl(opts->tsval);
409 *ptr++ = htonl(opts->tsecr);
412 if (unlikely(OPTION_SACK_ADVERTISE & opts->options &&
413 !(OPTION_TS & opts->options))) {
414 *ptr++ = htonl((TCPOPT_NOP << 24) |
416 (TCPOPT_SACK_PERM << 8) |
420 if (unlikely(opts->ws)) {
421 *ptr++ = htonl((TCPOPT_NOP << 24) |
422 (TCPOPT_WINDOW << 16) |
423 (TCPOLEN_WINDOW << 8) |
427 if (unlikely(opts->num_sack_blocks)) {
428 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
429 tp->duplicate_sack : tp->selective_acks;
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
435 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
436 TCPOLEN_SACK_PERBLOCK)));
438 for (this_sack = 0; this_sack < opts->num_sack_blocks;
440 *ptr++ = htonl(sp[this_sack].start_seq);
441 *ptr++ = htonl(sp[this_sack].end_seq);
444 tp->rx_opt.dsack = 0;
448 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
449 struct tcp_out_options *opts,
450 struct tcp_md5sig_key **md5) {
451 struct tcp_sock *tp = tcp_sk(sk);
454 #ifdef CONFIG_TCP_MD5SIG
455 *md5 = tp->af_specific->md5_lookup(sk, sk);
457 opts->options |= OPTION_MD5;
458 size += TCPOLEN_MD5SIG_ALIGNED;
464 /* We always get an MSS option. The option bytes which will be seen in
465 * normal data packets should timestamps be used, must be in the MSS
466 * advertised. But we subtract them from tp->mss_cache so that
467 * calculations in tcp_sendmsg are simpler etc. So account for this
468 * fact here if necessary. If we don't do this correctly, as a
469 * receiver we won't recognize data packets as being full sized when we
470 * should, and thus we won't abide by the delayed ACK rules correctly.
471 * SACKs don't matter, we never delay an ACK when we have any of those
473 opts->mss = tcp_advertise_mss(sk);
474 size += TCPOLEN_MSS_ALIGNED;
476 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
477 opts->options |= OPTION_TS;
478 opts->tsval = TCP_SKB_CB(skb)->when;
479 opts->tsecr = tp->rx_opt.ts_recent;
480 size += TCPOLEN_TSTAMP_ALIGNED;
482 if (likely(sysctl_tcp_window_scaling)) {
483 opts->ws = tp->rx_opt.rcv_wscale;
484 if (likely(opts->ws))
485 size += TCPOLEN_WSCALE_ALIGNED;
487 if (likely(sysctl_tcp_sack)) {
488 opts->options |= OPTION_SACK_ADVERTISE;
489 if (unlikely(!(OPTION_TS & opts->options)))
490 size += TCPOLEN_SACKPERM_ALIGNED;
496 static unsigned tcp_synack_options(struct sock *sk,
497 struct request_sock *req,
498 unsigned mss, struct sk_buff *skb,
499 struct tcp_out_options *opts,
500 struct tcp_md5sig_key **md5) {
502 struct inet_request_sock *ireq = inet_rsk(req);
505 #ifdef CONFIG_TCP_MD5SIG
506 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
508 opts->options |= OPTION_MD5;
509 size += TCPOLEN_MD5SIG_ALIGNED;
515 /* we can't fit any SACK blocks in a packet with MD5 + TS
516 options. There was discussion about disabling SACK rather than TS in
517 order to fit in better with old, buggy kernels, but that was deemed
518 to be unnecessary. */
519 doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
522 size += TCPOLEN_MSS_ALIGNED;
524 if (likely(ireq->wscale_ok)) {
525 opts->ws = ireq->rcv_wscale;
526 if (likely(opts->ws))
527 size += TCPOLEN_WSCALE_ALIGNED;
529 if (likely(doing_ts)) {
530 opts->options |= OPTION_TS;
531 opts->tsval = TCP_SKB_CB(skb)->when;
532 opts->tsecr = req->ts_recent;
533 size += TCPOLEN_TSTAMP_ALIGNED;
535 if (likely(ireq->sack_ok)) {
536 opts->options |= OPTION_SACK_ADVERTISE;
537 if (unlikely(!doing_ts))
538 size += TCPOLEN_SACKPERM_ALIGNED;
544 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
545 struct tcp_out_options *opts,
546 struct tcp_md5sig_key **md5) {
547 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
548 struct tcp_sock *tp = tcp_sk(sk);
550 unsigned int eff_sacks;
552 #ifdef CONFIG_TCP_MD5SIG
553 *md5 = tp->af_specific->md5_lookup(sk, sk);
554 if (unlikely(*md5)) {
555 opts->options |= OPTION_MD5;
556 size += TCPOLEN_MD5SIG_ALIGNED;
562 if (likely(tp->rx_opt.tstamp_ok)) {
563 opts->options |= OPTION_TS;
564 opts->tsval = tcb ? tcb->when : 0;
565 opts->tsecr = tp->rx_opt.ts_recent;
566 size += TCPOLEN_TSTAMP_ALIGNED;
569 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
570 if (unlikely(eff_sacks)) {
571 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
572 opts->num_sack_blocks =
573 min_t(unsigned, eff_sacks,
574 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
575 TCPOLEN_SACK_PERBLOCK);
576 size += TCPOLEN_SACK_BASE_ALIGNED +
577 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
583 /* This routine actually transmits TCP packets queued in by
584 * tcp_do_sendmsg(). This is used by both the initial
585 * transmission and possible later retransmissions.
586 * All SKB's seen here are completely headerless. It is our
587 * job to build the TCP header, and pass the packet down to
588 * IP so it can do the same plus pass the packet off to the
591 * We are working here with either a clone of the original
592 * SKB, or a fresh unique copy made by the retransmit engine.
594 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
597 const struct inet_connection_sock *icsk = inet_csk(sk);
598 struct inet_sock *inet;
600 struct tcp_skb_cb *tcb;
601 struct tcp_out_options opts;
602 unsigned tcp_options_size, tcp_header_size;
603 struct tcp_md5sig_key *md5;
604 __u8 *md5_hash_location;
608 BUG_ON(!skb || !tcp_skb_pcount(skb));
610 /* If congestion control is doing timestamping, we must
611 * take such a timestamp before we potentially clone/copy.
613 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
614 __net_timestamp(skb);
616 if (likely(clone_it)) {
617 if (unlikely(skb_cloned(skb)))
618 skb = pskb_copy(skb, gfp_mask);
620 skb = skb_clone(skb, gfp_mask);
627 tcb = TCP_SKB_CB(skb);
628 memset(&opts, 0, sizeof(opts));
630 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
631 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
633 tcp_options_size = tcp_established_options(sk, skb, &opts,
635 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
637 if (tcp_packets_in_flight(tp) == 0)
638 tcp_ca_event(sk, CA_EVENT_TX_START);
640 skb_push(skb, tcp_header_size);
641 skb_reset_transport_header(skb);
642 skb_set_owner_w(skb, sk);
644 /* Build TCP header and checksum it. */
646 th->source = inet->sport;
647 th->dest = inet->dport;
648 th->seq = htonl(tcb->seq);
649 th->ack_seq = htonl(tp->rcv_nxt);
650 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
653 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
654 /* RFC1323: The window in SYN & SYN/ACK segments
657 th->window = htons(min(tp->rcv_wnd, 65535U));
659 th->window = htons(tcp_select_window(sk));
664 /* The urg_mode check is necessary during a below snd_una win probe */
665 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
666 if (before(tp->snd_up, tcb->seq + 0x10000)) {
667 th->urg_ptr = htons(tp->snd_up - tcb->seq);
669 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
670 th->urg_ptr = 0xFFFF;
675 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
676 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
677 TCP_ECN_send(sk, skb, tcp_header_size);
679 #ifdef CONFIG_TCP_MD5SIG
680 /* Calculate the MD5 hash, as we have all we need now */
682 sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
683 tp->af_specific->calc_md5_hash(md5_hash_location,
688 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
690 if (likely(tcb->flags & TCPCB_FLAG_ACK))
691 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
693 if (skb->len != tcp_header_size)
694 tcp_event_data_sent(tp, skb, sk);
696 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
697 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
699 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
700 if (likely(err <= 0))
703 tcp_enter_cwr(sk, 1);
705 return net_xmit_eval(err);
708 /* This routine just queue's the buffer
710 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
711 * otherwise socket can stall.
713 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
715 struct tcp_sock *tp = tcp_sk(sk);
717 /* Advance write_seq and place onto the write_queue. */
718 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
719 skb_header_release(skb);
720 tcp_add_write_queue_tail(sk, skb);
721 sk->sk_wmem_queued += skb->truesize;
722 sk_mem_charge(sk, skb->truesize);
725 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
726 unsigned int mss_now)
728 if (skb->len <= mss_now || !sk_can_gso(sk)) {
729 /* Avoid the costly divide in the normal
732 skb_shinfo(skb)->gso_segs = 1;
733 skb_shinfo(skb)->gso_size = 0;
734 skb_shinfo(skb)->gso_type = 0;
736 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
737 skb_shinfo(skb)->gso_size = mss_now;
738 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
742 /* When a modification to fackets out becomes necessary, we need to check
743 * skb is counted to fackets_out or not.
745 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
748 struct tcp_sock *tp = tcp_sk(sk);
750 if (!tp->sacked_out || tcp_is_reno(tp))
753 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
754 tp->fackets_out -= decr;
757 /* Function to create two new TCP segments. Shrinks the given segment
758 * to the specified size and appends a new segment with the rest of the
759 * packet to the list. This won't be called frequently, I hope.
760 * Remember, these are still headerless SKBs at this point.
762 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
763 unsigned int mss_now)
765 struct tcp_sock *tp = tcp_sk(sk);
766 struct sk_buff *buff;
767 int nsize, old_factor;
771 BUG_ON(len > skb->len);
773 nsize = skb_headlen(skb) - len;
777 if (skb_cloned(skb) &&
778 skb_is_nonlinear(skb) &&
779 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
782 /* Get a new skb... force flag on. */
783 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
785 return -ENOMEM; /* We'll just try again later. */
787 sk->sk_wmem_queued += buff->truesize;
788 sk_mem_charge(sk, buff->truesize);
789 nlen = skb->len - len - nsize;
790 buff->truesize += nlen;
791 skb->truesize -= nlen;
793 /* Correct the sequence numbers. */
794 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
795 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
796 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
798 /* PSH and FIN should only be set in the second packet. */
799 flags = TCP_SKB_CB(skb)->flags;
800 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
801 TCP_SKB_CB(buff)->flags = flags;
802 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
804 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
805 /* Copy and checksum data tail into the new buffer. */
806 buff->csum = csum_partial_copy_nocheck(skb->data + len,
807 skb_put(buff, nsize),
812 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
814 skb->ip_summed = CHECKSUM_PARTIAL;
815 skb_split(skb, buff, len);
818 buff->ip_summed = skb->ip_summed;
820 /* Looks stupid, but our code really uses when of
821 * skbs, which it never sent before. --ANK
823 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
824 buff->tstamp = skb->tstamp;
826 old_factor = tcp_skb_pcount(skb);
828 /* Fix up tso_factor for both original and new SKB. */
829 tcp_set_skb_tso_segs(sk, skb, mss_now);
830 tcp_set_skb_tso_segs(sk, buff, mss_now);
832 /* If this packet has been sent out already, we must
833 * adjust the various packet counters.
835 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
836 int diff = old_factor - tcp_skb_pcount(skb) -
837 tcp_skb_pcount(buff);
839 tp->packets_out -= diff;
841 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
842 tp->sacked_out -= diff;
843 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
844 tp->retrans_out -= diff;
846 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
847 tp->lost_out -= diff;
849 /* Adjust Reno SACK estimate. */
850 if (tcp_is_reno(tp) && diff > 0) {
851 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
852 tcp_verify_left_out(tp);
854 tcp_adjust_fackets_out(sk, skb, diff);
856 if (tp->lost_skb_hint &&
857 before(TCP_SKB_CB(skb)->seq,
858 TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
859 (tcp_is_fack(tp) || TCP_SKB_CB(skb)->sacked))
860 tp->lost_cnt_hint -= diff;
863 /* Link BUFF into the send queue. */
864 skb_header_release(buff);
865 tcp_insert_write_queue_after(skb, buff, sk);
870 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
871 * eventually). The difference is that pulled data not copied, but
872 * immediately discarded.
874 static void __pskb_trim_head(struct sk_buff *skb, int len)
880 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
881 if (skb_shinfo(skb)->frags[i].size <= eat) {
882 put_page(skb_shinfo(skb)->frags[i].page);
883 eat -= skb_shinfo(skb)->frags[i].size;
885 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
887 skb_shinfo(skb)->frags[k].page_offset += eat;
888 skb_shinfo(skb)->frags[k].size -= eat;
894 skb_shinfo(skb)->nr_frags = k;
896 skb_reset_tail_pointer(skb);
897 skb->data_len -= len;
898 skb->len = skb->data_len;
901 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
903 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
906 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
907 if (unlikely(len < skb_headlen(skb)))
908 __skb_pull(skb, len);
910 __pskb_trim_head(skb, len - skb_headlen(skb));
912 TCP_SKB_CB(skb)->seq += len;
913 skb->ip_summed = CHECKSUM_PARTIAL;
915 skb->truesize -= len;
916 sk->sk_wmem_queued -= len;
917 sk_mem_uncharge(sk, len);
918 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
920 /* Any change of skb->len requires recalculation of tso
923 if (tcp_skb_pcount(skb) > 1)
924 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
929 /* Not accounting for SACKs here. */
930 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
932 struct tcp_sock *tp = tcp_sk(sk);
933 struct inet_connection_sock *icsk = inet_csk(sk);
936 /* Calculate base mss without TCP options:
937 It is MMS_S - sizeof(tcphdr) of rfc1122
939 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
941 /* Clamp it (mss_clamp does not include tcp options) */
942 if (mss_now > tp->rx_opt.mss_clamp)
943 mss_now = tp->rx_opt.mss_clamp;
945 /* Now subtract optional transport overhead */
946 mss_now -= icsk->icsk_ext_hdr_len;
948 /* Then reserve room for full set of TCP options and 8 bytes of data */
952 /* Now subtract TCP options size, not including SACKs */
953 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
958 /* Inverse of above */
959 int tcp_mss_to_mtu(struct sock *sk, int mss)
961 struct tcp_sock *tp = tcp_sk(sk);
962 struct inet_connection_sock *icsk = inet_csk(sk);
967 icsk->icsk_ext_hdr_len +
968 icsk->icsk_af_ops->net_header_len;
973 void tcp_mtup_init(struct sock *sk)
975 struct tcp_sock *tp = tcp_sk(sk);
976 struct inet_connection_sock *icsk = inet_csk(sk);
978 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
979 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
980 icsk->icsk_af_ops->net_header_len;
981 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
982 icsk->icsk_mtup.probe_size = 0;
985 /* Bound MSS / TSO packet size with the half of the window */
986 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
988 if (tp->max_window && pktsize > (tp->max_window >> 1))
989 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
994 /* This function synchronize snd mss to current pmtu/exthdr set.
996 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
997 for TCP options, but includes only bare TCP header.
999 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1000 It is minimum of user_mss and mss received with SYN.
1001 It also does not include TCP options.
1003 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1005 tp->mss_cache is current effective sending mss, including
1006 all tcp options except for SACKs. It is evaluated,
1007 taking into account current pmtu, but never exceeds
1008 tp->rx_opt.mss_clamp.
1010 NOTE1. rfc1122 clearly states that advertised MSS
1011 DOES NOT include either tcp or ip options.
1013 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1014 are READ ONLY outside this function. --ANK (980731)
1016 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1018 struct tcp_sock *tp = tcp_sk(sk);
1019 struct inet_connection_sock *icsk = inet_csk(sk);
1022 if (icsk->icsk_mtup.search_high > pmtu)
1023 icsk->icsk_mtup.search_high = pmtu;
1025 mss_now = tcp_mtu_to_mss(sk, pmtu);
1026 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1028 /* And store cached results */
1029 icsk->icsk_pmtu_cookie = pmtu;
1030 if (icsk->icsk_mtup.enabled)
1031 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1032 tp->mss_cache = mss_now;
1037 /* Compute the current effective MSS, taking SACKs and IP options,
1038 * and even PMTU discovery events into account.
1040 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
1042 struct tcp_sock *tp = tcp_sk(sk);
1043 struct dst_entry *dst = __sk_dst_get(sk);
1047 unsigned header_len;
1048 struct tcp_out_options opts;
1049 struct tcp_md5sig_key *md5;
1051 mss_now = tp->mss_cache;
1053 if (large_allowed && sk_can_gso(sk))
1057 u32 mtu = dst_mtu(dst);
1058 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1059 mss_now = tcp_sync_mss(sk, mtu);
1062 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1063 sizeof(struct tcphdr);
1064 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1065 * some common options. If this is an odd packet (because we have SACK
1066 * blocks etc) then our calculated header_len will be different, and
1067 * we have to adjust mss_now correspondingly */
1068 if (header_len != tp->tcp_header_len) {
1069 int delta = (int) header_len - tp->tcp_header_len;
1073 xmit_size_goal = mss_now;
1076 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1077 inet_csk(sk)->icsk_af_ops->net_header_len -
1078 inet_csk(sk)->icsk_ext_hdr_len -
1079 tp->tcp_header_len);
1081 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1082 xmit_size_goal -= (xmit_size_goal % mss_now);
1084 tp->xmit_size_goal = xmit_size_goal;
1089 /* Congestion window validation. (RFC2861) */
1090 static void tcp_cwnd_validate(struct sock *sk)
1092 struct tcp_sock *tp = tcp_sk(sk);
1094 if (tp->packets_out >= tp->snd_cwnd) {
1095 /* Network is feed fully. */
1096 tp->snd_cwnd_used = 0;
1097 tp->snd_cwnd_stamp = tcp_time_stamp;
1099 /* Network starves. */
1100 if (tp->packets_out > tp->snd_cwnd_used)
1101 tp->snd_cwnd_used = tp->packets_out;
1103 if (sysctl_tcp_slow_start_after_idle &&
1104 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1105 tcp_cwnd_application_limited(sk);
1109 /* Returns the portion of skb which can be sent right away without
1110 * introducing MSS oddities to segment boundaries. In rare cases where
1111 * mss_now != mss_cache, we will request caller to create a small skb
1112 * per input skb which could be mostly avoided here (if desired).
1114 * We explicitly want to create a request for splitting write queue tail
1115 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1116 * thus all the complexity (cwnd_len is always MSS multiple which we
1117 * return whenever allowed by the other factors). Basically we need the
1118 * modulo only when the receiver window alone is the limiting factor or
1119 * when we would be allowed to send the split-due-to-Nagle skb fully.
1121 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1122 unsigned int mss_now, unsigned int cwnd)
1124 struct tcp_sock *tp = tcp_sk(sk);
1125 u32 needed, window, cwnd_len;
1127 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1128 cwnd_len = mss_now * cwnd;
1130 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1133 needed = min(skb->len, window);
1135 if (cwnd_len <= needed)
1138 return needed - needed % mss_now;
1141 /* Can at least one segment of SKB be sent right now, according to the
1142 * congestion window rules? If so, return how many segments are allowed.
1144 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1145 struct sk_buff *skb)
1147 u32 in_flight, cwnd;
1149 /* Don't be strict about the congestion window for the final FIN. */
1150 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1151 tcp_skb_pcount(skb) == 1)
1154 in_flight = tcp_packets_in_flight(tp);
1155 cwnd = tp->snd_cwnd;
1156 if (in_flight < cwnd)
1157 return (cwnd - in_flight);
1162 /* This must be invoked the first time we consider transmitting
1163 * SKB onto the wire.
1165 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1166 unsigned int mss_now)
1168 int tso_segs = tcp_skb_pcount(skb);
1170 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1171 tcp_set_skb_tso_segs(sk, skb, mss_now);
1172 tso_segs = tcp_skb_pcount(skb);
1177 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1179 return after(tp->snd_sml, tp->snd_una) &&
1180 !after(tp->snd_sml, tp->snd_nxt);
1183 /* Return 0, if packet can be sent now without violation Nagle's rules:
1184 * 1. It is full sized.
1185 * 2. Or it contains FIN. (already checked by caller)
1186 * 3. Or TCP_NODELAY was set.
1187 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1188 * With Minshall's modification: all sent small packets are ACKed.
1190 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1191 const struct sk_buff *skb,
1192 unsigned mss_now, int nonagle)
1194 return (skb->len < mss_now &&
1195 ((nonagle & TCP_NAGLE_CORK) ||
1196 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1199 /* Return non-zero if the Nagle test allows this packet to be
1202 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1203 unsigned int cur_mss, int nonagle)
1205 /* Nagle rule does not apply to frames, which sit in the middle of the
1206 * write_queue (they have no chances to get new data).
1208 * This is implemented in the callers, where they modify the 'nonagle'
1209 * argument based upon the location of SKB in the send queue.
1211 if (nonagle & TCP_NAGLE_PUSH)
1214 /* Don't use the nagle rule for urgent data (or for the final FIN).
1215 * Nagle can be ignored during F-RTO too (see RFC4138).
1217 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1218 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1221 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1227 /* Does at least the first segment of SKB fit into the send window? */
1228 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1229 unsigned int cur_mss)
1231 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1233 if (skb->len > cur_mss)
1234 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1236 return !after(end_seq, tcp_wnd_end(tp));
1239 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1240 * should be put on the wire right now. If so, it returns the number of
1241 * packets allowed by the congestion window.
1243 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1244 unsigned int cur_mss, int nonagle)
1246 struct tcp_sock *tp = tcp_sk(sk);
1247 unsigned int cwnd_quota;
1249 tcp_init_tso_segs(sk, skb, cur_mss);
1251 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1254 cwnd_quota = tcp_cwnd_test(tp, skb);
1255 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1261 int tcp_may_send_now(struct sock *sk)
1263 struct tcp_sock *tp = tcp_sk(sk);
1264 struct sk_buff *skb = tcp_send_head(sk);
1267 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1268 (tcp_skb_is_last(sk, skb) ?
1269 tp->nonagle : TCP_NAGLE_PUSH)));
1272 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1273 * which is put after SKB on the list. It is very much like
1274 * tcp_fragment() except that it may make several kinds of assumptions
1275 * in order to speed up the splitting operation. In particular, we
1276 * know that all the data is in scatter-gather pages, and that the
1277 * packet has never been sent out before (and thus is not cloned).
1279 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1280 unsigned int mss_now)
1282 struct sk_buff *buff;
1283 int nlen = skb->len - len;
1286 /* All of a TSO frame must be composed of paged data. */
1287 if (skb->len != skb->data_len)
1288 return tcp_fragment(sk, skb, len, mss_now);
1290 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1291 if (unlikely(buff == NULL))
1294 sk->sk_wmem_queued += buff->truesize;
1295 sk_mem_charge(sk, buff->truesize);
1296 buff->truesize += nlen;
1297 skb->truesize -= nlen;
1299 /* Correct the sequence numbers. */
1300 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1301 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1302 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1304 /* PSH and FIN should only be set in the second packet. */
1305 flags = TCP_SKB_CB(skb)->flags;
1306 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1307 TCP_SKB_CB(buff)->flags = flags;
1309 /* This packet was never sent out yet, so no SACK bits. */
1310 TCP_SKB_CB(buff)->sacked = 0;
1312 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1313 skb_split(skb, buff, len);
1315 /* Fix up tso_factor for both original and new SKB. */
1316 tcp_set_skb_tso_segs(sk, skb, mss_now);
1317 tcp_set_skb_tso_segs(sk, buff, mss_now);
1319 /* Link BUFF into the send queue. */
1320 skb_header_release(buff);
1321 tcp_insert_write_queue_after(skb, buff, sk);
1326 /* Try to defer sending, if possible, in order to minimize the amount
1327 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1329 * This algorithm is from John Heffner.
1331 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1333 struct tcp_sock *tp = tcp_sk(sk);
1334 const struct inet_connection_sock *icsk = inet_csk(sk);
1335 u32 send_win, cong_win, limit, in_flight;
1337 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1340 if (icsk->icsk_ca_state != TCP_CA_Open)
1343 /* Defer for less than two clock ticks. */
1344 if (tp->tso_deferred &&
1345 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1348 in_flight = tcp_packets_in_flight(tp);
1350 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1352 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1354 /* From in_flight test above, we know that cwnd > in_flight. */
1355 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1357 limit = min(send_win, cong_win);
1359 /* If a full-sized TSO skb can be sent, do it. */
1360 if (limit >= sk->sk_gso_max_size)
1363 /* Middle in queue won't get any more data, full sendable already? */
1364 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1367 if (sysctl_tcp_tso_win_divisor) {
1368 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1370 /* If at least some fraction of a window is available,
1373 chunk /= sysctl_tcp_tso_win_divisor;
1377 /* Different approach, try not to defer past a single
1378 * ACK. Receiver should ACK every other full sized
1379 * frame, so if we have space for more than 3 frames
1382 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1386 /* Ok, it looks like it is advisable to defer. */
1387 tp->tso_deferred = 1 | (jiffies << 1);
1392 tp->tso_deferred = 0;
1396 /* Create a new MTU probe if we are ready.
1397 * Returns 0 if we should wait to probe (no cwnd available),
1398 * 1 if a probe was sent,
1401 static int tcp_mtu_probe(struct sock *sk)
1403 struct tcp_sock *tp = tcp_sk(sk);
1404 struct inet_connection_sock *icsk = inet_csk(sk);
1405 struct sk_buff *skb, *nskb, *next;
1412 /* Not currently probing/verifying,
1414 * have enough cwnd, and
1415 * not SACKing (the variable headers throw things off) */
1416 if (!icsk->icsk_mtup.enabled ||
1417 icsk->icsk_mtup.probe_size ||
1418 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1419 tp->snd_cwnd < 11 ||
1420 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1423 /* Very simple search strategy: just double the MSS. */
1424 mss_now = tcp_current_mss(sk, 0);
1425 probe_size = 2 * tp->mss_cache;
1426 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1427 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1428 /* TODO: set timer for probe_converge_event */
1432 /* Have enough data in the send queue to probe? */
1433 if (tp->write_seq - tp->snd_nxt < size_needed)
1436 if (tp->snd_wnd < size_needed)
1438 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1441 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1442 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1443 if (!tcp_packets_in_flight(tp))
1449 /* We're allowed to probe. Build it now. */
1450 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1452 sk->sk_wmem_queued += nskb->truesize;
1453 sk_mem_charge(sk, nskb->truesize);
1455 skb = tcp_send_head(sk);
1457 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1458 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1459 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1460 TCP_SKB_CB(nskb)->sacked = 0;
1462 nskb->ip_summed = skb->ip_summed;
1464 tcp_insert_write_queue_before(nskb, skb, sk);
1467 tcp_for_write_queue_from_safe(skb, next, sk) {
1468 copy = min_t(int, skb->len, probe_size - len);
1469 if (nskb->ip_summed)
1470 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1472 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1473 skb_put(nskb, copy),
1476 if (skb->len <= copy) {
1477 /* We've eaten all the data from this skb.
1479 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1480 tcp_unlink_write_queue(skb, sk);
1481 sk_wmem_free_skb(sk, skb);
1483 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1484 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1485 if (!skb_shinfo(skb)->nr_frags) {
1486 skb_pull(skb, copy);
1487 if (skb->ip_summed != CHECKSUM_PARTIAL)
1488 skb->csum = csum_partial(skb->data,
1491 __pskb_trim_head(skb, copy);
1492 tcp_set_skb_tso_segs(sk, skb, mss_now);
1494 TCP_SKB_CB(skb)->seq += copy;
1499 if (len >= probe_size)
1502 tcp_init_tso_segs(sk, nskb, nskb->len);
1504 /* We're ready to send. If this fails, the probe will
1505 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1506 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1507 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1508 /* Decrement cwnd here because we are sending
1509 * effectively two packets. */
1511 tcp_event_new_data_sent(sk, nskb);
1513 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1514 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1515 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1523 /* This routine writes packets to the network. It advances the
1524 * send_head. This happens as incoming acks open up the remote
1527 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1528 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1529 * account rare use of URG, this is not a big flaw.
1531 * Returns 1, if no segments are in flight and we have queued segments, but
1532 * cannot send anything now because of SWS or another problem.
1534 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1535 int push_one, gfp_t gfp)
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct sk_buff *skb;
1539 unsigned int tso_segs, sent_pkts;
1546 /* Do MTU probing. */
1547 result = tcp_mtu_probe(sk);
1550 } else if (result > 0) {
1555 while ((skb = tcp_send_head(sk))) {
1558 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1561 cwnd_quota = tcp_cwnd_test(tp, skb);
1565 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1568 if (tso_segs == 1) {
1569 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1570 (tcp_skb_is_last(sk, skb) ?
1571 nonagle : TCP_NAGLE_PUSH))))
1574 if (!push_one && tcp_tso_should_defer(sk, skb))
1579 if (tso_segs > 1 && !tcp_urg_mode(tp))
1580 limit = tcp_mss_split_point(sk, skb, mss_now,
1583 if (skb->len > limit &&
1584 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1587 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1589 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1592 /* Advance the send_head. This one is sent out.
1593 * This call will increment packets_out.
1595 tcp_event_new_data_sent(sk, skb);
1597 tcp_minshall_update(tp, mss_now, skb);
1604 if (likely(sent_pkts)) {
1605 tcp_cwnd_validate(sk);
1608 return !tp->packets_out && tcp_send_head(sk);
1611 /* Push out any pending frames which were held back due to
1612 * TCP_CORK or attempt at coalescing tiny packets.
1613 * The socket must be locked by the caller.
1615 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1618 struct sk_buff *skb = tcp_send_head(sk);
1623 /* If we are closed, the bytes will have to remain here.
1624 * In time closedown will finish, we empty the write queue and
1625 * all will be happy.
1627 if (unlikely(sk->sk_state == TCP_CLOSE))
1630 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1631 tcp_check_probe_timer(sk);
1634 /* Send _single_ skb sitting at the send head. This function requires
1635 * true push pending frames to setup probe timer etc.
1637 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1639 struct sk_buff *skb = tcp_send_head(sk);
1641 BUG_ON(!skb || skb->len < mss_now);
1643 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1646 /* This function returns the amount that we can raise the
1647 * usable window based on the following constraints
1649 * 1. The window can never be shrunk once it is offered (RFC 793)
1650 * 2. We limit memory per socket
1653 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1654 * RECV.NEXT + RCV.WIN fixed until:
1655 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1657 * i.e. don't raise the right edge of the window until you can raise
1658 * it at least MSS bytes.
1660 * Unfortunately, the recommended algorithm breaks header prediction,
1661 * since header prediction assumes th->window stays fixed.
1663 * Strictly speaking, keeping th->window fixed violates the receiver
1664 * side SWS prevention criteria. The problem is that under this rule
1665 * a stream of single byte packets will cause the right side of the
1666 * window to always advance by a single byte.
1668 * Of course, if the sender implements sender side SWS prevention
1669 * then this will not be a problem.
1671 * BSD seems to make the following compromise:
1673 * If the free space is less than the 1/4 of the maximum
1674 * space available and the free space is less than 1/2 mss,
1675 * then set the window to 0.
1676 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1677 * Otherwise, just prevent the window from shrinking
1678 * and from being larger than the largest representable value.
1680 * This prevents incremental opening of the window in the regime
1681 * where TCP is limited by the speed of the reader side taking
1682 * data out of the TCP receive queue. It does nothing about
1683 * those cases where the window is constrained on the sender side
1684 * because the pipeline is full.
1686 * BSD also seems to "accidentally" limit itself to windows that are a
1687 * multiple of MSS, at least until the free space gets quite small.
1688 * This would appear to be a side effect of the mbuf implementation.
1689 * Combining these two algorithms results in the observed behavior
1690 * of having a fixed window size at almost all times.
1692 * Below we obtain similar behavior by forcing the offered window to
1693 * a multiple of the mss when it is feasible to do so.
1695 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1696 * Regular options like TIMESTAMP are taken into account.
1698 u32 __tcp_select_window(struct sock *sk)
1700 struct inet_connection_sock *icsk = inet_csk(sk);
1701 struct tcp_sock *tp = tcp_sk(sk);
1702 /* MSS for the peer's data. Previous versions used mss_clamp
1703 * here. I don't know if the value based on our guesses
1704 * of peer's MSS is better for the performance. It's more correct
1705 * but may be worse for the performance because of rcv_mss
1706 * fluctuations. --SAW 1998/11/1
1708 int mss = icsk->icsk_ack.rcv_mss;
1709 int free_space = tcp_space(sk);
1710 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1713 if (mss > full_space)
1716 if (free_space < (full_space >> 1)) {
1717 icsk->icsk_ack.quick = 0;
1719 if (tcp_memory_pressure)
1720 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1723 if (free_space < mss)
1727 if (free_space > tp->rcv_ssthresh)
1728 free_space = tp->rcv_ssthresh;
1730 /* Don't do rounding if we are using window scaling, since the
1731 * scaled window will not line up with the MSS boundary anyway.
1733 window = tp->rcv_wnd;
1734 if (tp->rx_opt.rcv_wscale) {
1735 window = free_space;
1737 /* Advertise enough space so that it won't get scaled away.
1738 * Import case: prevent zero window announcement if
1739 * 1<<rcv_wscale > mss.
1741 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1742 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1743 << tp->rx_opt.rcv_wscale);
1745 /* Get the largest window that is a nice multiple of mss.
1746 * Window clamp already applied above.
1747 * If our current window offering is within 1 mss of the
1748 * free space we just keep it. This prevents the divide
1749 * and multiply from happening most of the time.
1750 * We also don't do any window rounding when the free space
1753 if (window <= free_space - mss || window > free_space)
1754 window = (free_space / mss) * mss;
1755 else if (mss == full_space &&
1756 free_space > window + (full_space >> 1))
1757 window = free_space;
1763 /* Collapses two adjacent SKB's during retransmission. */
1764 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1766 struct tcp_sock *tp = tcp_sk(sk);
1767 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1768 int skb_size, next_skb_size;
1770 skb_size = skb->len;
1771 next_skb_size = next_skb->len;
1773 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1775 tcp_highest_sack_combine(sk, next_skb, skb);
1777 tcp_unlink_write_queue(next_skb, sk);
1779 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1782 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1783 skb->ip_summed = CHECKSUM_PARTIAL;
1785 if (skb->ip_summed != CHECKSUM_PARTIAL)
1786 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1788 /* Update sequence range on original skb. */
1789 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1791 /* Merge over control information. This moves PSH/FIN etc. over */
1792 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1794 /* All done, get rid of second SKB and account for it so
1795 * packet counting does not break.
1797 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1798 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1799 tp->retrans_out -= tcp_skb_pcount(next_skb);
1800 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1801 tp->lost_out -= tcp_skb_pcount(next_skb);
1802 /* Reno case is special. Sigh... */
1803 if (tcp_is_reno(tp) && tp->sacked_out)
1804 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1806 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1807 tp->packets_out -= tcp_skb_pcount(next_skb);
1809 /* changed transmit queue under us so clear hints */
1810 tcp_clear_retrans_hints_partial(tp);
1811 if (next_skb == tp->retransmit_skb_hint)
1812 tp->retransmit_skb_hint = skb;
1814 sk_wmem_free_skb(sk, next_skb);
1817 static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1819 if (tcp_skb_pcount(skb) > 1)
1821 /* TODO: SACK collapsing could be used to remove this condition */
1822 if (skb_shinfo(skb)->nr_frags != 0)
1824 if (skb_cloned(skb))
1826 if (skb == tcp_send_head(sk))
1828 /* Some heurestics for collapsing over SACK'd could be invented */
1829 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1835 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
1838 struct tcp_sock *tp = tcp_sk(sk);
1839 struct sk_buff *skb = to, *tmp;
1842 if (!sysctl_tcp_retrans_collapse)
1844 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)
1847 tcp_for_write_queue_from_safe(skb, tmp, sk) {
1848 if (!tcp_can_collapse(sk, skb))
1860 /* Punt if not enough space exists in the first SKB for
1861 * the data in the second
1863 if (skb->len > skb_tailroom(to))
1866 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
1869 tcp_collapse_retrans(sk, to);
1873 /* This retransmits one SKB. Policy decisions and retransmit queue
1874 * state updates are done by the caller. Returns non-zero if an
1875 * error occurred which prevented the send.
1877 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1879 struct tcp_sock *tp = tcp_sk(sk);
1880 struct inet_connection_sock *icsk = inet_csk(sk);
1881 unsigned int cur_mss;
1884 /* Inconslusive MTU probe */
1885 if (icsk->icsk_mtup.probe_size) {
1886 icsk->icsk_mtup.probe_size = 0;
1889 /* Do not sent more than we queued. 1/4 is reserved for possible
1890 * copying overhead: fragmentation, tunneling, mangling etc.
1892 if (atomic_read(&sk->sk_wmem_alloc) >
1893 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1896 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1897 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1899 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1903 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1904 return -EHOSTUNREACH; /* Routing failure or similar. */
1906 cur_mss = tcp_current_mss(sk, 0);
1908 /* If receiver has shrunk his window, and skb is out of
1909 * new window, do not retransmit it. The exception is the
1910 * case, when window is shrunk to zero. In this case
1911 * our retransmit serves as a zero window probe.
1913 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1914 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1917 if (skb->len > cur_mss) {
1918 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1919 return -ENOMEM; /* We'll try again later. */
1921 tcp_init_tso_segs(sk, skb, cur_mss);
1924 tcp_retrans_try_collapse(sk, skb, cur_mss);
1926 /* Some Solaris stacks overoptimize and ignore the FIN on a
1927 * retransmit when old data is attached. So strip it off
1928 * since it is cheap to do so and saves bytes on the network.
1931 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1932 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1933 if (!pskb_trim(skb, 0)) {
1934 /* Reuse, even though it does some unnecessary work */
1935 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1936 TCP_SKB_CB(skb)->flags);
1937 skb->ip_summed = CHECKSUM_NONE;
1941 /* Make a copy, if the first transmission SKB clone we made
1942 * is still in somebody's hands, else make a clone.
1944 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1946 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1949 /* Update global TCP statistics. */
1950 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
1952 tp->total_retrans++;
1954 #if FASTRETRANS_DEBUG > 0
1955 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1956 if (net_ratelimit())
1957 printk(KERN_DEBUG "retrans_out leaked.\n");
1960 if (!tp->retrans_out)
1961 tp->lost_retrans_low = tp->snd_nxt;
1962 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
1963 tp->retrans_out += tcp_skb_pcount(skb);
1965 /* Save stamp of the first retransmit. */
1966 if (!tp->retrans_stamp)
1967 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
1971 /* snd_nxt is stored to detect loss of retransmitted segment,
1972 * see tcp_input.c tcp_sacktag_write_queue().
1974 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
1979 static int tcp_can_forward_retransmit(struct sock *sk)
1981 const struct inet_connection_sock *icsk = inet_csk(sk);
1982 struct tcp_sock *tp = tcp_sk(sk);
1984 /* Forward retransmissions are possible only during Recovery. */
1985 if (icsk->icsk_ca_state != TCP_CA_Recovery)
1988 /* No forward retransmissions in Reno are possible. */
1989 if (tcp_is_reno(tp))
1992 /* Yeah, we have to make difficult choice between forward transmission
1993 * and retransmission... Both ways have their merits...
1995 * For now we do not retransmit anything, while we have some new
1996 * segments to send. In the other cases, follow rule 3 for
1997 * NextSeg() specified in RFC3517.
2000 if (tcp_may_send_now(sk))
2006 /* This gets called after a retransmit timeout, and the initially
2007 * retransmitted data is acknowledged. It tries to continue
2008 * resending the rest of the retransmit queue, until either
2009 * we've sent it all or the congestion window limit is reached.
2010 * If doing SACK, the first ACK which comes back for a timeout
2011 * based retransmit packet might feed us FACK information again.
2012 * If so, we use it to avoid unnecessarily retransmissions.
2014 void tcp_xmit_retransmit_queue(struct sock *sk)
2016 const struct inet_connection_sock *icsk = inet_csk(sk);
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 struct sk_buff *skb;
2019 struct sk_buff *hole = NULL;
2022 int fwd_rexmitting = 0;
2025 tp->retransmit_high = tp->snd_una;
2027 if (tp->retransmit_skb_hint) {
2028 skb = tp->retransmit_skb_hint;
2029 last_lost = TCP_SKB_CB(skb)->end_seq;
2030 if (after(last_lost, tp->retransmit_high))
2031 last_lost = tp->retransmit_high;
2033 skb = tcp_write_queue_head(sk);
2034 last_lost = tp->snd_una;
2037 tcp_for_write_queue_from(skb, sk) {
2038 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2040 if (skb == tcp_send_head(sk))
2042 /* we could do better than to assign each time */
2044 tp->retransmit_skb_hint = skb;
2046 /* Assume this retransmit will generate
2047 * only one packet for congestion window
2048 * calculation purposes. This works because
2049 * tcp_retransmit_skb() will chop up the
2050 * packet to be MSS sized and all the
2051 * packet counting works out.
2053 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2056 if (fwd_rexmitting) {
2058 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2060 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2062 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2063 tp->retransmit_high = last_lost;
2064 if (!tcp_can_forward_retransmit(sk))
2066 /* Backtrack if necessary to non-L'ed skb */
2074 } else if (!(sacked & TCPCB_LOST)) {
2075 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2080 last_lost = TCP_SKB_CB(skb)->end_seq;
2081 if (icsk->icsk_ca_state != TCP_CA_Loss)
2082 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2084 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2087 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2090 if (tcp_retransmit_skb(sk, skb))
2092 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2094 if (skb == tcp_write_queue_head(sk))
2095 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2096 inet_csk(sk)->icsk_rto,
2101 /* Send a fin. The caller locks the socket for us. This cannot be
2102 * allowed to fail queueing a FIN frame under any circumstances.
2104 void tcp_send_fin(struct sock *sk)
2106 struct tcp_sock *tp = tcp_sk(sk);
2107 struct sk_buff *skb = tcp_write_queue_tail(sk);
2110 /* Optimization, tack on the FIN if we have a queue of
2111 * unsent frames. But be careful about outgoing SACKS
2114 mss_now = tcp_current_mss(sk, 1);
2116 if (tcp_send_head(sk) != NULL) {
2117 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2118 TCP_SKB_CB(skb)->end_seq++;
2121 /* Socket is locked, keep trying until memory is available. */
2123 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2129 /* Reserve space for headers and prepare control bits. */
2130 skb_reserve(skb, MAX_TCP_HEADER);
2131 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2132 tcp_init_nondata_skb(skb, tp->write_seq,
2133 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2134 tcp_queue_skb(sk, skb);
2136 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2139 /* We get here when a process closes a file descriptor (either due to
2140 * an explicit close() or as a byproduct of exit()'ing) and there
2141 * was unread data in the receive queue. This behavior is recommended
2142 * by RFC 2525, section 2.17. -DaveM
2144 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2146 struct sk_buff *skb;
2148 /* NOTE: No TCP options attached and we never retransmit this. */
2149 skb = alloc_skb(MAX_TCP_HEADER, priority);
2151 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2155 /* Reserve space for headers and prepare control bits. */
2156 skb_reserve(skb, MAX_TCP_HEADER);
2157 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2158 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2160 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2161 if (tcp_transmit_skb(sk, skb, 0, priority))
2162 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2164 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2167 /* WARNING: This routine must only be called when we have already sent
2168 * a SYN packet that crossed the incoming SYN that caused this routine
2169 * to get called. If this assumption fails then the initial rcv_wnd
2170 * and rcv_wscale values will not be correct.
2172 int tcp_send_synack(struct sock *sk)
2174 struct sk_buff *skb;
2176 skb = tcp_write_queue_head(sk);
2177 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2178 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2181 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2182 if (skb_cloned(skb)) {
2183 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2186 tcp_unlink_write_queue(skb, sk);
2187 skb_header_release(nskb);
2188 __tcp_add_write_queue_head(sk, nskb);
2189 sk_wmem_free_skb(sk, skb);
2190 sk->sk_wmem_queued += nskb->truesize;
2191 sk_mem_charge(sk, nskb->truesize);
2195 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2196 TCP_ECN_send_synack(tcp_sk(sk), skb);
2198 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2199 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2203 * Prepare a SYN-ACK.
2205 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2206 struct request_sock *req)
2208 struct inet_request_sock *ireq = inet_rsk(req);
2209 struct tcp_sock *tp = tcp_sk(sk);
2211 int tcp_header_size;
2212 struct tcp_out_options opts;
2213 struct sk_buff *skb;
2214 struct tcp_md5sig_key *md5;
2215 __u8 *md5_hash_location;
2218 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2222 /* Reserve space for headers. */
2223 skb_reserve(skb, MAX_TCP_HEADER);
2225 skb->dst = dst_clone(dst);
2227 mss = dst_metric(dst, RTAX_ADVMSS);
2228 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2229 mss = tp->rx_opt.user_mss;
2231 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2233 /* Set this up on the first call only */
2234 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2235 /* tcp_full_space because it is guaranteed to be the first packet */
2236 tcp_select_initial_window(tcp_full_space(sk),
2237 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2242 ireq->rcv_wscale = rcv_wscale;
2245 memset(&opts, 0, sizeof(opts));
2246 #ifdef CONFIG_SYN_COOKIES
2247 if (unlikely(req->cookie_ts))
2248 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2251 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2252 tcp_header_size = tcp_synack_options(sk, req, mss,
2254 sizeof(struct tcphdr);
2256 skb_push(skb, tcp_header_size);
2257 skb_reset_transport_header(skb);
2260 memset(th, 0, sizeof(struct tcphdr));
2263 TCP_ECN_make_synack(req, th);
2264 th->source = ireq->loc_port;
2265 th->dest = ireq->rmt_port;
2266 /* Setting of flags are superfluous here for callers (and ECE is
2267 * not even correctly set)
2269 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2270 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2271 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2272 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2274 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2275 th->window = htons(min(req->rcv_wnd, 65535U));
2276 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
2277 th->doff = (tcp_header_size >> 2);
2278 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
2280 #ifdef CONFIG_TCP_MD5SIG
2281 /* Okay, we have all we need - do the md5 hash if needed */
2283 tp->af_specific->calc_md5_hash(md5_hash_location,
2284 md5, NULL, req, skb);
2292 * Do all connect socket setups that can be done AF independent.
2294 static void tcp_connect_init(struct sock *sk)
2296 struct dst_entry *dst = __sk_dst_get(sk);
2297 struct tcp_sock *tp = tcp_sk(sk);
2300 /* We'll fix this up when we get a response from the other end.
2301 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2303 tp->tcp_header_len = sizeof(struct tcphdr) +
2304 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2306 #ifdef CONFIG_TCP_MD5SIG
2307 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2308 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2311 /* If user gave his TCP_MAXSEG, record it to clamp */
2312 if (tp->rx_opt.user_mss)
2313 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2316 tcp_sync_mss(sk, dst_mtu(dst));
2318 if (!tp->window_clamp)
2319 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2320 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2321 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2322 tp->advmss = tp->rx_opt.user_mss;
2324 tcp_initialize_rcv_mss(sk);
2326 tcp_select_initial_window(tcp_full_space(sk),
2327 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2330 sysctl_tcp_window_scaling,
2333 tp->rx_opt.rcv_wscale = rcv_wscale;
2334 tp->rcv_ssthresh = tp->rcv_wnd;
2337 sock_reset_flag(sk, SOCK_DONE);
2340 tp->snd_una = tp->write_seq;
2341 tp->snd_sml = tp->write_seq;
2342 tp->snd_up = tp->write_seq;
2347 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2348 inet_csk(sk)->icsk_retransmits = 0;
2349 tcp_clear_retrans(tp);
2353 * Build a SYN and send it off.
2355 int tcp_connect(struct sock *sk)
2357 struct tcp_sock *tp = tcp_sk(sk);
2358 struct sk_buff *buff;
2360 tcp_connect_init(sk);
2362 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2363 if (unlikely(buff == NULL))
2366 /* Reserve space for headers. */
2367 skb_reserve(buff, MAX_TCP_HEADER);
2369 tp->snd_nxt = tp->write_seq;
2370 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2371 TCP_ECN_send_syn(sk, buff);
2374 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2375 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2376 skb_header_release(buff);
2377 __tcp_add_write_queue_tail(sk, buff);
2378 sk->sk_wmem_queued += buff->truesize;
2379 sk_mem_charge(sk, buff->truesize);
2380 tp->packets_out += tcp_skb_pcount(buff);
2381 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2383 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2384 * in order to make this packet get counted in tcpOutSegs.
2386 tp->snd_nxt = tp->write_seq;
2387 tp->pushed_seq = tp->write_seq;
2388 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2390 /* Timer for repeating the SYN until an answer. */
2391 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2392 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2396 /* Send out a delayed ack, the caller does the policy checking
2397 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2400 void tcp_send_delayed_ack(struct sock *sk)
2402 struct inet_connection_sock *icsk = inet_csk(sk);
2403 int ato = icsk->icsk_ack.ato;
2404 unsigned long timeout;
2406 if (ato > TCP_DELACK_MIN) {
2407 const struct tcp_sock *tp = tcp_sk(sk);
2408 int max_ato = HZ / 2;
2410 if (icsk->icsk_ack.pingpong ||
2411 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2412 max_ato = TCP_DELACK_MAX;
2414 /* Slow path, intersegment interval is "high". */
2416 /* If some rtt estimate is known, use it to bound delayed ack.
2417 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2421 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2427 ato = min(ato, max_ato);
2430 /* Stay within the limit we were given */
2431 timeout = jiffies + ato;
2433 /* Use new timeout only if there wasn't a older one earlier. */
2434 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2435 /* If delack timer was blocked or is about to expire,
2438 if (icsk->icsk_ack.blocked ||
2439 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2444 if (!time_before(timeout, icsk->icsk_ack.timeout))
2445 timeout = icsk->icsk_ack.timeout;
2447 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2448 icsk->icsk_ack.timeout = timeout;
2449 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2452 /* This routine sends an ack and also updates the window. */
2453 void tcp_send_ack(struct sock *sk)
2455 struct sk_buff *buff;
2457 /* If we have been reset, we may not send again. */
2458 if (sk->sk_state == TCP_CLOSE)
2461 /* We are not putting this on the write queue, so
2462 * tcp_transmit_skb() will set the ownership to this
2465 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2467 inet_csk_schedule_ack(sk);
2468 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2469 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2470 TCP_DELACK_MAX, TCP_RTO_MAX);
2474 /* Reserve space for headers and prepare control bits. */
2475 skb_reserve(buff, MAX_TCP_HEADER);
2476 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2478 /* Send it off, this clears delayed acks for us. */
2479 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2480 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2483 /* This routine sends a packet with an out of date sequence
2484 * number. It assumes the other end will try to ack it.
2486 * Question: what should we make while urgent mode?
2487 * 4.4BSD forces sending single byte of data. We cannot send
2488 * out of window data, because we have SND.NXT==SND.MAX...
2490 * Current solution: to send TWO zero-length segments in urgent mode:
2491 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2492 * out-of-date with SND.UNA-1 to probe window.
2494 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2496 struct tcp_sock *tp = tcp_sk(sk);
2497 struct sk_buff *skb;
2499 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2500 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2504 /* Reserve space for headers and set control bits. */
2505 skb_reserve(skb, MAX_TCP_HEADER);
2506 /* Use a previous sequence. This should cause the other
2507 * end to send an ack. Don't queue or clone SKB, just
2510 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2511 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2512 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2515 int tcp_write_wakeup(struct sock *sk)
2517 struct tcp_sock *tp = tcp_sk(sk);
2518 struct sk_buff *skb;
2520 if (sk->sk_state == TCP_CLOSE)
2523 if ((skb = tcp_send_head(sk)) != NULL &&
2524 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2526 unsigned int mss = tcp_current_mss(sk, 0);
2527 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2529 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2530 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2532 /* We are probing the opening of a window
2533 * but the window size is != 0
2534 * must have been a result SWS avoidance ( sender )
2536 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2538 seg_size = min(seg_size, mss);
2539 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2540 if (tcp_fragment(sk, skb, seg_size, mss))
2542 } else if (!tcp_skb_pcount(skb))
2543 tcp_set_skb_tso_segs(sk, skb, mss);
2545 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2546 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2547 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2549 tcp_event_new_data_sent(sk, skb);
2552 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2553 tcp_xmit_probe_skb(sk, 1);
2554 return tcp_xmit_probe_skb(sk, 0);
2558 /* A window probe timeout has occurred. If window is not closed send
2559 * a partial packet else a zero probe.
2561 void tcp_send_probe0(struct sock *sk)
2563 struct inet_connection_sock *icsk = inet_csk(sk);
2564 struct tcp_sock *tp = tcp_sk(sk);
2567 err = tcp_write_wakeup(sk);
2569 if (tp->packets_out || !tcp_send_head(sk)) {
2570 /* Cancel probe timer, if it is not required. */
2571 icsk->icsk_probes_out = 0;
2572 icsk->icsk_backoff = 0;
2577 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2578 icsk->icsk_backoff++;
2579 icsk->icsk_probes_out++;
2580 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2581 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2584 /* If packet was not sent due to local congestion,
2585 * do not backoff and do not remember icsk_probes_out.
2586 * Let local senders to fight for local resources.
2588 * Use accumulated backoff yet.
2590 if (!icsk->icsk_probes_out)
2591 icsk->icsk_probes_out = 1;
2592 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2593 min(icsk->icsk_rto << icsk->icsk_backoff,
2594 TCP_RESOURCE_PROBE_INTERVAL),
2599 EXPORT_SYMBOL(tcp_select_initial_window);
2600 EXPORT_SYMBOL(tcp_connect);
2601 EXPORT_SYMBOL(tcp_make_synack);
2602 EXPORT_SYMBOL(tcp_simple_retransmit);
2603 EXPORT_SYMBOL(tcp_sync_mss);
2604 EXPORT_SYMBOL(tcp_mtup_init);