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 if (tp->rx_opt.dsack) {
445 tp->rx_opt.dsack = 0;
446 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks;
451 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
452 struct tcp_out_options *opts,
453 struct tcp_md5sig_key **md5) {
454 struct tcp_sock *tp = tcp_sk(sk);
457 #ifdef CONFIG_TCP_MD5SIG
458 *md5 = tp->af_specific->md5_lookup(sk, sk);
460 opts->options |= OPTION_MD5;
461 size += TCPOLEN_MD5SIG_ALIGNED;
467 /* We always get an MSS option. The option bytes which will be seen in
468 * normal data packets should timestamps be used, must be in the MSS
469 * advertised. But we subtract them from tp->mss_cache so that
470 * calculations in tcp_sendmsg are simpler etc. So account for this
471 * fact here if necessary. If we don't do this correctly, as a
472 * receiver we won't recognize data packets as being full sized when we
473 * should, and thus we won't abide by the delayed ACK rules correctly.
474 * SACKs don't matter, we never delay an ACK when we have any of those
476 opts->mss = tcp_advertise_mss(sk);
477 size += TCPOLEN_MSS_ALIGNED;
479 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
480 opts->options |= OPTION_TS;
481 opts->tsval = TCP_SKB_CB(skb)->when;
482 opts->tsecr = tp->rx_opt.ts_recent;
483 size += TCPOLEN_TSTAMP_ALIGNED;
485 if (likely(sysctl_tcp_window_scaling)) {
486 opts->ws = tp->rx_opt.rcv_wscale;
488 size += TCPOLEN_WSCALE_ALIGNED;
490 if (likely(sysctl_tcp_sack)) {
491 opts->options |= OPTION_SACK_ADVERTISE;
492 if (unlikely(!(OPTION_TS & opts->options)))
493 size += TCPOLEN_SACKPERM_ALIGNED;
499 static unsigned tcp_synack_options(struct sock *sk,
500 struct request_sock *req,
501 unsigned mss, struct sk_buff *skb,
502 struct tcp_out_options *opts,
503 struct tcp_md5sig_key **md5) {
505 struct inet_request_sock *ireq = inet_rsk(req);
508 #ifdef CONFIG_TCP_MD5SIG
509 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
511 opts->options |= OPTION_MD5;
512 size += TCPOLEN_MD5SIG_ALIGNED;
518 /* we can't fit any SACK blocks in a packet with MD5 + TS
519 options. There was discussion about disabling SACK rather than TS in
520 order to fit in better with old, buggy kernels, but that was deemed
521 to be unnecessary. */
522 doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok);
525 size += TCPOLEN_MSS_ALIGNED;
527 if (likely(ireq->wscale_ok)) {
528 opts->ws = ireq->rcv_wscale;
530 size += TCPOLEN_WSCALE_ALIGNED;
532 if (likely(doing_ts)) {
533 opts->options |= OPTION_TS;
534 opts->tsval = TCP_SKB_CB(skb)->when;
535 opts->tsecr = req->ts_recent;
536 size += TCPOLEN_TSTAMP_ALIGNED;
538 if (likely(ireq->sack_ok)) {
539 opts->options |= OPTION_SACK_ADVERTISE;
540 if (unlikely(!doing_ts))
541 size += TCPOLEN_SACKPERM_ALIGNED;
547 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
548 struct tcp_out_options *opts,
549 struct tcp_md5sig_key **md5) {
550 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
551 struct tcp_sock *tp = tcp_sk(sk);
554 #ifdef CONFIG_TCP_MD5SIG
555 *md5 = tp->af_specific->md5_lookup(sk, sk);
556 if (unlikely(*md5)) {
557 opts->options |= OPTION_MD5;
558 size += TCPOLEN_MD5SIG_ALIGNED;
564 if (likely(tp->rx_opt.tstamp_ok)) {
565 opts->options |= OPTION_TS;
566 opts->tsval = tcb ? tcb->when : 0;
567 opts->tsecr = tp->rx_opt.ts_recent;
568 size += TCPOLEN_TSTAMP_ALIGNED;
571 if (unlikely(tp->rx_opt.eff_sacks)) {
572 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
573 opts->num_sack_blocks =
574 min_t(unsigned, tp->rx_opt.eff_sacks,
575 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
576 TCPOLEN_SACK_PERBLOCK);
577 size += TCPOLEN_SACK_BASE_ALIGNED +
578 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
584 /* This routine actually transmits TCP packets queued in by
585 * tcp_do_sendmsg(). This is used by both the initial
586 * transmission and possible later retransmissions.
587 * All SKB's seen here are completely headerless. It is our
588 * job to build the TCP header, and pass the packet down to
589 * IP so it can do the same plus pass the packet off to the
592 * We are working here with either a clone of the original
593 * SKB, or a fresh unique copy made by the retransmit engine.
595 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
598 const struct inet_connection_sock *icsk = inet_csk(sk);
599 struct inet_sock *inet;
601 struct tcp_skb_cb *tcb;
602 struct tcp_out_options opts;
603 unsigned tcp_options_size, tcp_header_size;
604 struct tcp_md5sig_key *md5;
605 __u8 *md5_hash_location;
609 BUG_ON(!skb || !tcp_skb_pcount(skb));
611 /* If congestion control is doing timestamping, we must
612 * take such a timestamp before we potentially clone/copy.
614 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
615 __net_timestamp(skb);
617 if (likely(clone_it)) {
618 if (unlikely(skb_cloned(skb)))
619 skb = pskb_copy(skb, gfp_mask);
621 skb = skb_clone(skb, gfp_mask);
628 tcb = TCP_SKB_CB(skb);
629 memset(&opts, 0, sizeof(opts));
631 if (unlikely(tcb->flags & TCPCB_FLAG_SYN))
632 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
634 tcp_options_size = tcp_established_options(sk, skb, &opts,
636 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
638 if (tcp_packets_in_flight(tp) == 0)
639 tcp_ca_event(sk, CA_EVENT_TX_START);
641 skb_push(skb, tcp_header_size);
642 skb_reset_transport_header(skb);
643 skb_set_owner_w(skb, sk);
645 /* Build TCP header and checksum it. */
647 th->source = inet->sport;
648 th->dest = inet->dport;
649 th->seq = htonl(tcb->seq);
650 th->ack_seq = htonl(tp->rcv_nxt);
651 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
654 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) {
655 /* RFC1323: The window in SYN & SYN/ACK segments
658 th->window = htons(min(tp->rcv_wnd, 65535U));
660 th->window = htons(tcp_select_window(sk));
665 /* The urg_mode check is necessary during a below snd_una win probe */
666 if (unlikely(tcp_urg_mode(tp) &&
667 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) {
668 th->urg_ptr = htons(tp->snd_up - tcb->seq);
672 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
673 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0))
674 TCP_ECN_send(sk, skb, tcp_header_size);
676 #ifdef CONFIG_TCP_MD5SIG
677 /* Calculate the MD5 hash, as we have all we need now */
679 sk->sk_route_caps &= ~NETIF_F_GSO_MASK;
680 tp->af_specific->calc_md5_hash(md5_hash_location,
685 icsk->icsk_af_ops->send_check(sk, skb->len, skb);
687 if (likely(tcb->flags & TCPCB_FLAG_ACK))
688 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
690 if (skb->len != tcp_header_size)
691 tcp_event_data_sent(tp, skb, sk);
693 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
694 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
696 err = icsk->icsk_af_ops->queue_xmit(skb, 0);
697 if (likely(err <= 0))
700 tcp_enter_cwr(sk, 1);
702 return net_xmit_eval(err);
705 /* This routine just queue's the buffer
707 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
708 * otherwise socket can stall.
710 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
712 struct tcp_sock *tp = tcp_sk(sk);
714 /* Advance write_seq and place onto the write_queue. */
715 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
716 skb_header_release(skb);
717 tcp_add_write_queue_tail(sk, skb);
718 sk->sk_wmem_queued += skb->truesize;
719 sk_mem_charge(sk, skb->truesize);
722 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
723 unsigned int mss_now)
725 if (skb->len <= mss_now || !sk_can_gso(sk) ||
726 tcp_urg_mode(tcp_sk(sk))) {
727 /* Avoid the costly divide in the normal
730 skb_shinfo(skb)->gso_segs = 1;
731 skb_shinfo(skb)->gso_size = 0;
732 skb_shinfo(skb)->gso_type = 0;
734 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
735 skb_shinfo(skb)->gso_size = mss_now;
736 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
740 /* When a modification to fackets out becomes necessary, we need to check
741 * skb is counted to fackets_out or not.
743 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
746 struct tcp_sock *tp = tcp_sk(sk);
748 if (!tp->sacked_out || tcp_is_reno(tp))
751 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
752 tp->fackets_out -= decr;
755 /* Function to create two new TCP segments. Shrinks the given segment
756 * to the specified size and appends a new segment with the rest of the
757 * packet to the list. This won't be called frequently, I hope.
758 * Remember, these are still headerless SKBs at this point.
760 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
761 unsigned int mss_now)
763 struct tcp_sock *tp = tcp_sk(sk);
764 struct sk_buff *buff;
765 int nsize, old_factor;
769 BUG_ON(len > skb->len);
771 tcp_clear_retrans_hints_partial(tp);
772 nsize = skb_headlen(skb) - len;
776 if (skb_cloned(skb) &&
777 skb_is_nonlinear(skb) &&
778 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
781 /* Get a new skb... force flag on. */
782 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
784 return -ENOMEM; /* We'll just try again later. */
786 sk->sk_wmem_queued += buff->truesize;
787 sk_mem_charge(sk, buff->truesize);
788 nlen = skb->len - len - nsize;
789 buff->truesize += nlen;
790 skb->truesize -= nlen;
792 /* Correct the sequence numbers. */
793 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
794 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
795 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
797 /* PSH and FIN should only be set in the second packet. */
798 flags = TCP_SKB_CB(skb)->flags;
799 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
800 TCP_SKB_CB(buff)->flags = flags;
801 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
803 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
804 /* Copy and checksum data tail into the new buffer. */
805 buff->csum = csum_partial_copy_nocheck(skb->data + len,
806 skb_put(buff, nsize),
811 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
813 skb->ip_summed = CHECKSUM_PARTIAL;
814 skb_split(skb, buff, len);
817 buff->ip_summed = skb->ip_summed;
819 /* Looks stupid, but our code really uses when of
820 * skbs, which it never sent before. --ANK
822 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
823 buff->tstamp = skb->tstamp;
825 old_factor = tcp_skb_pcount(skb);
827 /* Fix up tso_factor for both original and new SKB. */
828 tcp_set_skb_tso_segs(sk, skb, mss_now);
829 tcp_set_skb_tso_segs(sk, buff, mss_now);
831 /* If this packet has been sent out already, we must
832 * adjust the various packet counters.
834 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
835 int diff = old_factor - tcp_skb_pcount(skb) -
836 tcp_skb_pcount(buff);
838 tp->packets_out -= diff;
840 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
841 tp->sacked_out -= diff;
842 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
843 tp->retrans_out -= diff;
845 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
846 tp->lost_out -= diff;
848 /* Adjust Reno SACK estimate. */
849 if (tcp_is_reno(tp) && diff > 0) {
850 tcp_dec_pcount_approx_int(&tp->sacked_out, diff);
851 tcp_verify_left_out(tp);
853 tcp_adjust_fackets_out(sk, skb, diff);
856 /* Link BUFF into the send queue. */
857 skb_header_release(buff);
858 tcp_insert_write_queue_after(skb, buff, sk);
863 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
864 * eventually). The difference is that pulled data not copied, but
865 * immediately discarded.
867 static void __pskb_trim_head(struct sk_buff *skb, int len)
873 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
874 if (skb_shinfo(skb)->frags[i].size <= eat) {
875 put_page(skb_shinfo(skb)->frags[i].page);
876 eat -= skb_shinfo(skb)->frags[i].size;
878 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
880 skb_shinfo(skb)->frags[k].page_offset += eat;
881 skb_shinfo(skb)->frags[k].size -= eat;
887 skb_shinfo(skb)->nr_frags = k;
889 skb_reset_tail_pointer(skb);
890 skb->data_len -= len;
891 skb->len = skb->data_len;
894 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
896 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
899 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
900 if (unlikely(len < skb_headlen(skb)))
901 __skb_pull(skb, len);
903 __pskb_trim_head(skb, len - skb_headlen(skb));
905 TCP_SKB_CB(skb)->seq += len;
906 skb->ip_summed = CHECKSUM_PARTIAL;
908 skb->truesize -= len;
909 sk->sk_wmem_queued -= len;
910 sk_mem_uncharge(sk, len);
911 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
913 /* Any change of skb->len requires recalculation of tso
916 if (tcp_skb_pcount(skb) > 1)
917 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1));
922 /* Not accounting for SACKs here. */
923 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
925 struct tcp_sock *tp = tcp_sk(sk);
926 struct inet_connection_sock *icsk = inet_csk(sk);
929 /* Calculate base mss without TCP options:
930 It is MMS_S - sizeof(tcphdr) of rfc1122
932 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
934 /* Clamp it (mss_clamp does not include tcp options) */
935 if (mss_now > tp->rx_opt.mss_clamp)
936 mss_now = tp->rx_opt.mss_clamp;
938 /* Now subtract optional transport overhead */
939 mss_now -= icsk->icsk_ext_hdr_len;
941 /* Then reserve room for full set of TCP options and 8 bytes of data */
945 /* Now subtract TCP options size, not including SACKs */
946 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
951 /* Inverse of above */
952 int tcp_mss_to_mtu(struct sock *sk, int mss)
954 struct tcp_sock *tp = tcp_sk(sk);
955 struct inet_connection_sock *icsk = inet_csk(sk);
960 icsk->icsk_ext_hdr_len +
961 icsk->icsk_af_ops->net_header_len;
966 void tcp_mtup_init(struct sock *sk)
968 struct tcp_sock *tp = tcp_sk(sk);
969 struct inet_connection_sock *icsk = inet_csk(sk);
971 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
972 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
973 icsk->icsk_af_ops->net_header_len;
974 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
975 icsk->icsk_mtup.probe_size = 0;
978 /* Bound MSS / TSO packet size with the half of the window */
979 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize)
981 if (tp->max_window && pktsize > (tp->max_window >> 1))
982 return max(tp->max_window >> 1, 68U - tp->tcp_header_len);
987 /* This function synchronize snd mss to current pmtu/exthdr set.
989 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
990 for TCP options, but includes only bare TCP header.
992 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
993 It is minimum of user_mss and mss received with SYN.
994 It also does not include TCP options.
996 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
998 tp->mss_cache is current effective sending mss, including
999 all tcp options except for SACKs. It is evaluated,
1000 taking into account current pmtu, but never exceeds
1001 tp->rx_opt.mss_clamp.
1003 NOTE1. rfc1122 clearly states that advertised MSS
1004 DOES NOT include either tcp or ip options.
1006 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1007 are READ ONLY outside this function. --ANK (980731)
1009 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1011 struct tcp_sock *tp = tcp_sk(sk);
1012 struct inet_connection_sock *icsk = inet_csk(sk);
1015 if (icsk->icsk_mtup.search_high > pmtu)
1016 icsk->icsk_mtup.search_high = pmtu;
1018 mss_now = tcp_mtu_to_mss(sk, pmtu);
1019 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1021 /* And store cached results */
1022 icsk->icsk_pmtu_cookie = pmtu;
1023 if (icsk->icsk_mtup.enabled)
1024 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1025 tp->mss_cache = mss_now;
1030 /* Compute the current effective MSS, taking SACKs and IP options,
1031 * and even PMTU discovery events into account.
1033 * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up
1034 * cannot be large. However, taking into account rare use of URG, this
1035 * is not a big flaw.
1037 unsigned int tcp_current_mss(struct sock *sk, int large_allowed)
1039 struct tcp_sock *tp = tcp_sk(sk);
1040 struct dst_entry *dst = __sk_dst_get(sk);
1044 unsigned header_len;
1045 struct tcp_out_options opts;
1046 struct tcp_md5sig_key *md5;
1048 mss_now = tp->mss_cache;
1050 if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp))
1054 u32 mtu = dst_mtu(dst);
1055 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1056 mss_now = tcp_sync_mss(sk, mtu);
1059 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1060 sizeof(struct tcphdr);
1061 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1062 * some common options. If this is an odd packet (because we have SACK
1063 * blocks etc) then our calculated header_len will be different, and
1064 * we have to adjust mss_now correspondingly */
1065 if (header_len != tp->tcp_header_len) {
1066 int delta = (int) header_len - tp->tcp_header_len;
1070 xmit_size_goal = mss_now;
1073 xmit_size_goal = ((sk->sk_gso_max_size - 1) -
1074 inet_csk(sk)->icsk_af_ops->net_header_len -
1075 inet_csk(sk)->icsk_ext_hdr_len -
1076 tp->tcp_header_len);
1078 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal);
1079 xmit_size_goal -= (xmit_size_goal % mss_now);
1081 tp->xmit_size_goal = xmit_size_goal;
1086 /* Congestion window validation. (RFC2861) */
1087 static void tcp_cwnd_validate(struct sock *sk)
1089 struct tcp_sock *tp = tcp_sk(sk);
1091 if (tp->packets_out >= tp->snd_cwnd) {
1092 /* Network is feed fully. */
1093 tp->snd_cwnd_used = 0;
1094 tp->snd_cwnd_stamp = tcp_time_stamp;
1096 /* Network starves. */
1097 if (tp->packets_out > tp->snd_cwnd_used)
1098 tp->snd_cwnd_used = tp->packets_out;
1100 if (sysctl_tcp_slow_start_after_idle &&
1101 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1102 tcp_cwnd_application_limited(sk);
1106 /* Returns the portion of skb which can be sent right away without
1107 * introducing MSS oddities to segment boundaries. In rare cases where
1108 * mss_now != mss_cache, we will request caller to create a small skb
1109 * per input skb which could be mostly avoided here (if desired).
1111 * We explicitly want to create a request for splitting write queue tail
1112 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1113 * thus all the complexity (cwnd_len is always MSS multiple which we
1114 * return whenever allowed by the other factors). Basically we need the
1115 * modulo only when the receiver window alone is the limiting factor or
1116 * when we would be allowed to send the split-due-to-Nagle skb fully.
1118 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1119 unsigned int mss_now, unsigned int cwnd)
1121 struct tcp_sock *tp = tcp_sk(sk);
1122 u32 needed, window, cwnd_len;
1124 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1125 cwnd_len = mss_now * cwnd;
1127 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1130 needed = min(skb->len, window);
1132 if (cwnd_len <= needed)
1135 return needed - needed % mss_now;
1138 /* Can at least one segment of SKB be sent right now, according to the
1139 * congestion window rules? If so, return how many segments are allowed.
1141 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1142 struct sk_buff *skb)
1144 u32 in_flight, cwnd;
1146 /* Don't be strict about the congestion window for the final FIN. */
1147 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1148 tcp_skb_pcount(skb) == 1)
1151 in_flight = tcp_packets_in_flight(tp);
1152 cwnd = tp->snd_cwnd;
1153 if (in_flight < cwnd)
1154 return (cwnd - in_flight);
1159 /* This must be invoked the first time we consider transmitting
1160 * SKB onto the wire.
1162 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1163 unsigned int mss_now)
1165 int tso_segs = tcp_skb_pcount(skb);
1168 (tso_segs > 1 && (tcp_skb_mss(skb) != mss_now ||
1169 tcp_urg_mode(tcp_sk(sk))))) {
1170 tcp_set_skb_tso_segs(sk, skb, mss_now);
1171 tso_segs = tcp_skb_pcount(skb);
1176 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1178 return after(tp->snd_sml,tp->snd_una) &&
1179 !after(tp->snd_sml, tp->snd_nxt);
1182 /* Return 0, if packet can be sent now without violation Nagle's rules:
1183 * 1. It is full sized.
1184 * 2. Or it contains FIN. (already checked by caller)
1185 * 3. Or TCP_NODELAY was set.
1186 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1187 * With Minshall's modification: all sent small packets are ACKed.
1189 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1190 const struct sk_buff *skb,
1191 unsigned mss_now, int nonagle)
1193 return (skb->len < mss_now &&
1194 ((nonagle & TCP_NAGLE_CORK) ||
1195 (!nonagle && tp->packets_out && tcp_minshall_check(tp))));
1198 /* Return non-zero if the Nagle test allows this packet to be
1201 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1202 unsigned int cur_mss, int nonagle)
1204 /* Nagle rule does not apply to frames, which sit in the middle of the
1205 * write_queue (they have no chances to get new data).
1207 * This is implemented in the callers, where they modify the 'nonagle'
1208 * argument based upon the location of SKB in the send queue.
1210 if (nonagle & TCP_NAGLE_PUSH)
1213 /* Don't use the nagle rule for urgent data (or for the final FIN).
1214 * Nagle can be ignored during F-RTO too (see RFC4138).
1216 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1217 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN))
1220 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1226 /* Does at least the first segment of SKB fit into the send window? */
1227 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1228 unsigned int cur_mss)
1230 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1232 if (skb->len > cur_mss)
1233 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1235 return !after(end_seq, tcp_wnd_end(tp));
1238 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1239 * should be put on the wire right now. If so, it returns the number of
1240 * packets allowed by the congestion window.
1242 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1243 unsigned int cur_mss, int nonagle)
1245 struct tcp_sock *tp = tcp_sk(sk);
1246 unsigned int cwnd_quota;
1248 tcp_init_tso_segs(sk, skb, cur_mss);
1250 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1253 cwnd_quota = tcp_cwnd_test(tp, skb);
1254 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1260 int tcp_may_send_now(struct sock *sk)
1262 struct tcp_sock *tp = tcp_sk(sk);
1263 struct sk_buff *skb = tcp_send_head(sk);
1266 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1),
1267 (tcp_skb_is_last(sk, skb) ?
1268 tp->nonagle : TCP_NAGLE_PUSH)));
1271 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1272 * which is put after SKB on the list. It is very much like
1273 * tcp_fragment() except that it may make several kinds of assumptions
1274 * in order to speed up the splitting operation. In particular, we
1275 * know that all the data is in scatter-gather pages, and that the
1276 * packet has never been sent out before (and thus is not cloned).
1278 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1279 unsigned int mss_now)
1281 struct sk_buff *buff;
1282 int nlen = skb->len - len;
1285 /* All of a TSO frame must be composed of paged data. */
1286 if (skb->len != skb->data_len)
1287 return tcp_fragment(sk, skb, len, mss_now);
1289 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC);
1290 if (unlikely(buff == NULL))
1293 sk->sk_wmem_queued += buff->truesize;
1294 sk_mem_charge(sk, buff->truesize);
1295 buff->truesize += nlen;
1296 skb->truesize -= nlen;
1298 /* Correct the sequence numbers. */
1299 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1300 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1301 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1303 /* PSH and FIN should only be set in the second packet. */
1304 flags = TCP_SKB_CB(skb)->flags;
1305 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH);
1306 TCP_SKB_CB(buff)->flags = flags;
1308 /* This packet was never sent out yet, so no SACK bits. */
1309 TCP_SKB_CB(buff)->sacked = 0;
1311 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1312 skb_split(skb, buff, len);
1314 /* Fix up tso_factor for both original and new SKB. */
1315 tcp_set_skb_tso_segs(sk, skb, mss_now);
1316 tcp_set_skb_tso_segs(sk, buff, mss_now);
1318 /* Link BUFF into the send queue. */
1319 skb_header_release(buff);
1320 tcp_insert_write_queue_after(skb, buff, sk);
1325 /* Try to defer sending, if possible, in order to minimize the amount
1326 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1328 * This algorithm is from John Heffner.
1330 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1332 struct tcp_sock *tp = tcp_sk(sk);
1333 const struct inet_connection_sock *icsk = inet_csk(sk);
1334 u32 send_win, cong_win, limit, in_flight;
1336 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)
1339 if (icsk->icsk_ca_state != TCP_CA_Open)
1342 /* Defer for less than two clock ticks. */
1343 if (tp->tso_deferred &&
1344 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1347 in_flight = tcp_packets_in_flight(tp);
1349 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1351 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1353 /* From in_flight test above, we know that cwnd > in_flight. */
1354 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1356 limit = min(send_win, cong_win);
1358 /* If a full-sized TSO skb can be sent, do it. */
1359 if (limit >= sk->sk_gso_max_size)
1362 if (sysctl_tcp_tso_win_divisor) {
1363 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1365 /* If at least some fraction of a window is available,
1368 chunk /= sysctl_tcp_tso_win_divisor;
1372 /* Different approach, try not to defer past a single
1373 * ACK. Receiver should ACK every other full sized
1374 * frame, so if we have space for more than 3 frames
1377 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1381 /* Ok, it looks like it is advisable to defer. */
1382 tp->tso_deferred = 1 | (jiffies << 1);
1387 tp->tso_deferred = 0;
1391 /* Create a new MTU probe if we are ready.
1392 * Returns 0 if we should wait to probe (no cwnd available),
1393 * 1 if a probe was sent,
1396 static int tcp_mtu_probe(struct sock *sk)
1398 struct tcp_sock *tp = tcp_sk(sk);
1399 struct inet_connection_sock *icsk = inet_csk(sk);
1400 struct sk_buff *skb, *nskb, *next;
1407 /* Not currently probing/verifying,
1409 * have enough cwnd, and
1410 * not SACKing (the variable headers throw things off) */
1411 if (!icsk->icsk_mtup.enabled ||
1412 icsk->icsk_mtup.probe_size ||
1413 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1414 tp->snd_cwnd < 11 ||
1415 tp->rx_opt.eff_sacks)
1418 /* Very simple search strategy: just double the MSS. */
1419 mss_now = tcp_current_mss(sk, 0);
1420 probe_size = 2 * tp->mss_cache;
1421 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1422 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1423 /* TODO: set timer for probe_converge_event */
1427 /* Have enough data in the send queue to probe? */
1428 if (tp->write_seq - tp->snd_nxt < size_needed)
1431 if (tp->snd_wnd < size_needed)
1433 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1436 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1437 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1438 if (!tcp_packets_in_flight(tp))
1444 /* We're allowed to probe. Build it now. */
1445 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1447 sk->sk_wmem_queued += nskb->truesize;
1448 sk_mem_charge(sk, nskb->truesize);
1450 skb = tcp_send_head(sk);
1452 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1453 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1454 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK;
1455 TCP_SKB_CB(nskb)->sacked = 0;
1457 nskb->ip_summed = skb->ip_summed;
1459 tcp_insert_write_queue_before(nskb, skb, sk);
1462 tcp_for_write_queue_from_safe(skb, next, sk) {
1463 copy = min_t(int, skb->len, probe_size - len);
1464 if (nskb->ip_summed)
1465 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1467 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1468 skb_put(nskb, copy),
1471 if (skb->len <= copy) {
1472 /* We've eaten all the data from this skb.
1474 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1475 tcp_unlink_write_queue(skb, sk);
1476 sk_wmem_free_skb(sk, skb);
1478 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1479 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH);
1480 if (!skb_shinfo(skb)->nr_frags) {
1481 skb_pull(skb, copy);
1482 if (skb->ip_summed != CHECKSUM_PARTIAL)
1483 skb->csum = csum_partial(skb->data,
1486 __pskb_trim_head(skb, copy);
1487 tcp_set_skb_tso_segs(sk, skb, mss_now);
1489 TCP_SKB_CB(skb)->seq += copy;
1494 if (len >= probe_size)
1497 tcp_init_tso_segs(sk, nskb, nskb->len);
1499 /* We're ready to send. If this fails, the probe will
1500 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1501 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1502 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1503 /* Decrement cwnd here because we are sending
1504 * effectively two packets. */
1506 tcp_event_new_data_sent(sk, nskb);
1508 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1509 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1510 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1518 /* This routine writes packets to the network. It advances the
1519 * send_head. This happens as incoming acks open up the remote
1522 * Returns 1, if no segments are in flight and we have queued segments, but
1523 * cannot send anything now because of SWS or another problem.
1525 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle)
1527 struct tcp_sock *tp = tcp_sk(sk);
1528 struct sk_buff *skb;
1529 unsigned int tso_segs, sent_pkts;
1533 /* If we are closed, the bytes will have to remain here.
1534 * In time closedown will finish, we empty the write queue and all
1537 if (unlikely(sk->sk_state == TCP_CLOSE))
1542 /* Do MTU probing. */
1543 if ((result = tcp_mtu_probe(sk)) == 0) {
1545 } else if (result > 0) {
1549 while ((skb = tcp_send_head(sk))) {
1552 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1555 cwnd_quota = tcp_cwnd_test(tp, skb);
1559 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1562 if (tso_segs == 1) {
1563 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1564 (tcp_skb_is_last(sk, skb) ?
1565 nonagle : TCP_NAGLE_PUSH))))
1568 if (tcp_tso_should_defer(sk, skb))
1574 limit = tcp_mss_split_point(sk, skb, mss_now,
1577 if (skb->len > limit &&
1578 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1581 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1583 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC)))
1586 /* Advance the send_head. This one is sent out.
1587 * This call will increment packets_out.
1589 tcp_event_new_data_sent(sk, skb);
1591 tcp_minshall_update(tp, mss_now, skb);
1595 if (likely(sent_pkts)) {
1596 tcp_cwnd_validate(sk);
1599 return !tp->packets_out && tcp_send_head(sk);
1602 /* Push out any pending frames which were held back due to
1603 * TCP_CORK or attempt at coalescing tiny packets.
1604 * The socket must be locked by the caller.
1606 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1609 struct sk_buff *skb = tcp_send_head(sk);
1612 if (tcp_write_xmit(sk, cur_mss, nonagle))
1613 tcp_check_probe_timer(sk);
1617 /* Send _single_ skb sitting at the send head. This function requires
1618 * true push pending frames to setup probe timer etc.
1620 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1622 struct sk_buff *skb = tcp_send_head(sk);
1623 unsigned int tso_segs, cwnd_quota;
1625 BUG_ON(!skb || skb->len < mss_now);
1627 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1628 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH);
1630 if (likely(cwnd_quota)) {
1637 limit = tcp_mss_split_point(sk, skb, mss_now,
1640 if (skb->len > limit &&
1641 unlikely(tso_fragment(sk, skb, limit, mss_now)))
1644 /* Send it out now. */
1645 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1647 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) {
1648 tcp_event_new_data_sent(sk, skb);
1649 tcp_cwnd_validate(sk);
1655 /* This function returns the amount that we can raise the
1656 * usable window based on the following constraints
1658 * 1. The window can never be shrunk once it is offered (RFC 793)
1659 * 2. We limit memory per socket
1662 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1663 * RECV.NEXT + RCV.WIN fixed until:
1664 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1666 * i.e. don't raise the right edge of the window until you can raise
1667 * it at least MSS bytes.
1669 * Unfortunately, the recommended algorithm breaks header prediction,
1670 * since header prediction assumes th->window stays fixed.
1672 * Strictly speaking, keeping th->window fixed violates the receiver
1673 * side SWS prevention criteria. The problem is that under this rule
1674 * a stream of single byte packets will cause the right side of the
1675 * window to always advance by a single byte.
1677 * Of course, if the sender implements sender side SWS prevention
1678 * then this will not be a problem.
1680 * BSD seems to make the following compromise:
1682 * If the free space is less than the 1/4 of the maximum
1683 * space available and the free space is less than 1/2 mss,
1684 * then set the window to 0.
1685 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1686 * Otherwise, just prevent the window from shrinking
1687 * and from being larger than the largest representable value.
1689 * This prevents incremental opening of the window in the regime
1690 * where TCP is limited by the speed of the reader side taking
1691 * data out of the TCP receive queue. It does nothing about
1692 * those cases where the window is constrained on the sender side
1693 * because the pipeline is full.
1695 * BSD also seems to "accidentally" limit itself to windows that are a
1696 * multiple of MSS, at least until the free space gets quite small.
1697 * This would appear to be a side effect of the mbuf implementation.
1698 * Combining these two algorithms results in the observed behavior
1699 * of having a fixed window size at almost all times.
1701 * Below we obtain similar behavior by forcing the offered window to
1702 * a multiple of the mss when it is feasible to do so.
1704 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1705 * Regular options like TIMESTAMP are taken into account.
1707 u32 __tcp_select_window(struct sock *sk)
1709 struct inet_connection_sock *icsk = inet_csk(sk);
1710 struct tcp_sock *tp = tcp_sk(sk);
1711 /* MSS for the peer's data. Previous versions used mss_clamp
1712 * here. I don't know if the value based on our guesses
1713 * of peer's MSS is better for the performance. It's more correct
1714 * but may be worse for the performance because of rcv_mss
1715 * fluctuations. --SAW 1998/11/1
1717 int mss = icsk->icsk_ack.rcv_mss;
1718 int free_space = tcp_space(sk);
1719 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1722 if (mss > full_space)
1725 if (free_space < (full_space >> 1)) {
1726 icsk->icsk_ack.quick = 0;
1728 if (tcp_memory_pressure)
1729 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1732 if (free_space < mss)
1736 if (free_space > tp->rcv_ssthresh)
1737 free_space = tp->rcv_ssthresh;
1739 /* Don't do rounding if we are using window scaling, since the
1740 * scaled window will not line up with the MSS boundary anyway.
1742 window = tp->rcv_wnd;
1743 if (tp->rx_opt.rcv_wscale) {
1744 window = free_space;
1746 /* Advertise enough space so that it won't get scaled away.
1747 * Import case: prevent zero window announcement if
1748 * 1<<rcv_wscale > mss.
1750 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1751 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1752 << tp->rx_opt.rcv_wscale);
1754 /* Get the largest window that is a nice multiple of mss.
1755 * Window clamp already applied above.
1756 * If our current window offering is within 1 mss of the
1757 * free space we just keep it. This prevents the divide
1758 * and multiply from happening most of the time.
1759 * We also don't do any window rounding when the free space
1762 if (window <= free_space - mss || window > free_space)
1763 window = (free_space / mss) * mss;
1764 else if (mss == full_space &&
1765 free_space > window + (full_space >> 1))
1766 window = free_space;
1772 /* Attempt to collapse two adjacent SKB's during retransmission. */
1773 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb,
1776 struct tcp_sock *tp = tcp_sk(sk);
1777 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1778 int skb_size, next_skb_size;
1781 /* The first test we must make is that neither of these two
1782 * SKB's are still referenced by someone else.
1784 if (skb_cloned(skb) || skb_cloned(next_skb))
1787 skb_size = skb->len;
1788 next_skb_size = next_skb->len;
1789 flags = TCP_SKB_CB(skb)->flags;
1791 /* Also punt if next skb has been SACK'd. */
1792 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED)
1795 /* Next skb is out of window. */
1796 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp)))
1799 /* Punt if not enough space exists in the first SKB for
1800 * the data in the second, or the total combined payload
1801 * would exceed the MSS.
1803 if ((next_skb_size > skb_tailroom(skb)) ||
1804 ((skb_size + next_skb_size) > mss_now))
1807 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1809 tcp_highest_sack_combine(sk, next_skb, skb);
1811 /* Ok. We will be able to collapse the packet. */
1812 tcp_unlink_write_queue(next_skb, sk);
1814 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1817 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1818 skb->ip_summed = CHECKSUM_PARTIAL;
1820 if (skb->ip_summed != CHECKSUM_PARTIAL)
1821 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1823 /* Update sequence range on original skb. */
1824 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1826 /* Merge over control information. */
1827 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */
1828 TCP_SKB_CB(skb)->flags = flags;
1830 /* All done, get rid of second SKB and account for it so
1831 * packet counting does not break.
1833 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1834 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS)
1835 tp->retrans_out -= tcp_skb_pcount(next_skb);
1836 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST)
1837 tp->lost_out -= tcp_skb_pcount(next_skb);
1838 /* Reno case is special. Sigh... */
1839 if (tcp_is_reno(tp) && tp->sacked_out)
1840 tcp_dec_pcount_approx(&tp->sacked_out, next_skb);
1842 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb));
1843 tp->packets_out -= tcp_skb_pcount(next_skb);
1845 /* changed transmit queue under us so clear hints */
1846 tcp_clear_retrans_hints_partial(tp);
1847 if (next_skb == tp->retransmit_skb_hint)
1848 tp->retransmit_skb_hint = skb;
1850 sk_wmem_free_skb(sk, next_skb);
1853 /* Do a simple retransmit without using the backoff mechanisms in
1854 * tcp_timer. This is used for path mtu discovery.
1855 * The socket is already locked here.
1857 void tcp_simple_retransmit(struct sock *sk)
1859 const struct inet_connection_sock *icsk = inet_csk(sk);
1860 struct tcp_sock *tp = tcp_sk(sk);
1861 struct sk_buff *skb;
1862 unsigned int mss = tcp_current_mss(sk, 0);
1863 u32 prior_lost = tp->lost_out;
1865 tcp_for_write_queue(skb, sk) {
1866 if (skb == tcp_send_head(sk))
1868 if (skb->len > mss &&
1869 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
1870 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1871 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1872 tp->retrans_out -= tcp_skb_pcount(skb);
1874 tcp_skb_mark_lost_uncond_verify(tp, skb);
1878 tcp_clear_retrans_hints_partial(tp);
1880 if (prior_lost == tp->lost_out)
1883 if (tcp_is_reno(tp))
1884 tcp_limit_reno_sacked(tp);
1886 tcp_verify_left_out(tp);
1888 /* Don't muck with the congestion window here.
1889 * Reason is that we do not increase amount of _data_
1890 * in network, but units changed and effective
1891 * cwnd/ssthresh really reduced now.
1893 if (icsk->icsk_ca_state != TCP_CA_Loss) {
1894 tp->high_seq = tp->snd_nxt;
1895 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1896 tp->prior_ssthresh = 0;
1897 tp->undo_marker = 0;
1898 tcp_set_ca_state(sk, TCP_CA_Loss);
1900 tcp_xmit_retransmit_queue(sk);
1903 /* This retransmits one SKB. Policy decisions and retransmit queue
1904 * state updates are done by the caller. Returns non-zero if an
1905 * error occurred which prevented the send.
1907 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
1909 struct tcp_sock *tp = tcp_sk(sk);
1910 struct inet_connection_sock *icsk = inet_csk(sk);
1911 unsigned int cur_mss;
1914 /* Inconslusive MTU probe */
1915 if (icsk->icsk_mtup.probe_size) {
1916 icsk->icsk_mtup.probe_size = 0;
1919 /* Do not sent more than we queued. 1/4 is reserved for possible
1920 * copying overhead: fragmentation, tunneling, mangling etc.
1922 if (atomic_read(&sk->sk_wmem_alloc) >
1923 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
1926 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
1927 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1929 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
1933 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
1934 return -EHOSTUNREACH; /* Routing failure or similar. */
1936 cur_mss = tcp_current_mss(sk, 0);
1938 /* If receiver has shrunk his window, and skb is out of
1939 * new window, do not retransmit it. The exception is the
1940 * case, when window is shrunk to zero. In this case
1941 * our retransmit serves as a zero window probe.
1943 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))
1944 && TCP_SKB_CB(skb)->seq != tp->snd_una)
1947 if (skb->len > cur_mss) {
1948 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
1949 return -ENOMEM; /* We'll try again later. */
1952 /* Collapse two adjacent packets if worthwhile and we can. */
1953 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) &&
1954 (skb->len < (cur_mss >> 1)) &&
1955 (!tcp_skb_is_last(sk, skb)) &&
1956 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) &&
1957 (skb_shinfo(skb)->nr_frags == 0 &&
1958 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) &&
1959 (tcp_skb_pcount(skb) == 1 &&
1960 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) &&
1961 (sysctl_tcp_retrans_collapse != 0))
1962 tcp_retrans_try_collapse(sk, skb, cur_mss);
1964 /* Some Solaris stacks overoptimize and ignore the FIN on a
1965 * retransmit when old data is attached. So strip it off
1966 * since it is cheap to do so and saves bytes on the network.
1969 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) &&
1970 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
1971 if (!pskb_trim(skb, 0)) {
1972 /* Reuse, even though it does some unnecessary work */
1973 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
1974 TCP_SKB_CB(skb)->flags);
1975 skb->ip_summed = CHECKSUM_NONE;
1979 /* Make a copy, if the first transmission SKB clone we made
1980 * is still in somebody's hands, else make a clone.
1982 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1984 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1987 /* Update global TCP statistics. */
1988 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
1990 tp->total_retrans++;
1992 #if FASTRETRANS_DEBUG > 0
1993 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
1994 if (net_ratelimit())
1995 printk(KERN_DEBUG "retrans_out leaked.\n");
1998 if (!tp->retrans_out)
1999 tp->lost_retrans_low = tp->snd_nxt;
2000 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2001 tp->retrans_out += tcp_skb_pcount(skb);
2003 /* Save stamp of the first retransmit. */
2004 if (!tp->retrans_stamp)
2005 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2009 /* snd_nxt is stored to detect loss of retransmitted segment,
2010 * see tcp_input.c tcp_sacktag_write_queue().
2012 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2017 static int tcp_can_forward_retransmit(struct sock *sk)
2019 const struct inet_connection_sock *icsk = inet_csk(sk);
2020 struct tcp_sock *tp = tcp_sk(sk);
2022 /* Forward retransmissions are possible only during Recovery. */
2023 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2026 /* No forward retransmissions in Reno are possible. */
2027 if (tcp_is_reno(tp))
2030 /* Yeah, we have to make difficult choice between forward transmission
2031 * and retransmission... Both ways have their merits...
2033 * For now we do not retransmit anything, while we have some new
2034 * segments to send. In the other cases, follow rule 3 for
2035 * NextSeg() specified in RFC3517.
2038 if (tcp_may_send_now(sk))
2044 /* This gets called after a retransmit timeout, and the initially
2045 * retransmitted data is acknowledged. It tries to continue
2046 * resending the rest of the retransmit queue, until either
2047 * we've sent it all or the congestion window limit is reached.
2048 * If doing SACK, the first ACK which comes back for a timeout
2049 * based retransmit packet might feed us FACK information again.
2050 * If so, we use it to avoid unnecessarily retransmissions.
2052 void tcp_xmit_retransmit_queue(struct sock *sk)
2054 const struct inet_connection_sock *icsk = inet_csk(sk);
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct sk_buff *skb;
2057 struct sk_buff *hole = NULL;
2060 int fwd_rexmitting = 0;
2063 tp->retransmit_high = tp->snd_una;
2065 if (tp->retransmit_skb_hint) {
2066 skb = tp->retransmit_skb_hint;
2067 last_lost = TCP_SKB_CB(skb)->end_seq;
2068 if (after(last_lost, tp->retransmit_high))
2069 last_lost = tp->retransmit_high;
2071 skb = tcp_write_queue_head(sk);
2072 last_lost = tp->snd_una;
2075 /* First pass: retransmit lost packets. */
2076 tcp_for_write_queue_from(skb, sk) {
2077 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2079 if (skb == tcp_send_head(sk))
2081 /* we could do better than to assign each time */
2083 tp->retransmit_skb_hint = skb;
2085 /* Assume this retransmit will generate
2086 * only one packet for congestion window
2087 * calculation purposes. This works because
2088 * tcp_retransmit_skb() will chop up the
2089 * packet to be MSS sized and all the
2090 * packet counting works out.
2092 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2095 if (fwd_rexmitting) {
2097 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2099 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2101 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2102 tp->retransmit_high = last_lost;
2103 if (!tcp_can_forward_retransmit(sk))
2105 /* Backtrack if necessary to non-L'ed skb */
2113 } else if (!(sacked & TCPCB_LOST)) {
2114 if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS))
2119 last_lost = TCP_SKB_CB(skb)->end_seq;
2120 if (icsk->icsk_ca_state != TCP_CA_Loss)
2121 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2123 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2126 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2129 if (tcp_retransmit_skb(sk, skb))
2131 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2133 if (skb == tcp_write_queue_head(sk))
2134 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2135 inet_csk(sk)->icsk_rto,
2140 /* Send a fin. The caller locks the socket for us. This cannot be
2141 * allowed to fail queueing a FIN frame under any circumstances.
2143 void tcp_send_fin(struct sock *sk)
2145 struct tcp_sock *tp = tcp_sk(sk);
2146 struct sk_buff *skb = tcp_write_queue_tail(sk);
2149 /* Optimization, tack on the FIN if we have a queue of
2150 * unsent frames. But be careful about outgoing SACKS
2153 mss_now = tcp_current_mss(sk, 1);
2155 if (tcp_send_head(sk) != NULL) {
2156 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN;
2157 TCP_SKB_CB(skb)->end_seq++;
2160 /* Socket is locked, keep trying until memory is available. */
2162 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL);
2168 /* Reserve space for headers and prepare control bits. */
2169 skb_reserve(skb, MAX_TCP_HEADER);
2170 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2171 tcp_init_nondata_skb(skb, tp->write_seq,
2172 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN);
2173 tcp_queue_skb(sk, skb);
2175 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2178 /* We get here when a process closes a file descriptor (either due to
2179 * an explicit close() or as a byproduct of exit()'ing) and there
2180 * was unread data in the receive queue. This behavior is recommended
2181 * by RFC 2525, section 2.17. -DaveM
2183 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2185 struct sk_buff *skb;
2187 /* NOTE: No TCP options attached and we never retransmit this. */
2188 skb = alloc_skb(MAX_TCP_HEADER, priority);
2190 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2194 /* Reserve space for headers and prepare control bits. */
2195 skb_reserve(skb, MAX_TCP_HEADER);
2196 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2197 TCPCB_FLAG_ACK | TCPCB_FLAG_RST);
2199 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2200 if (tcp_transmit_skb(sk, skb, 0, priority))
2201 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2203 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2206 /* WARNING: This routine must only be called when we have already sent
2207 * a SYN packet that crossed the incoming SYN that caused this routine
2208 * to get called. If this assumption fails then the initial rcv_wnd
2209 * and rcv_wscale values will not be correct.
2211 int tcp_send_synack(struct sock *sk)
2213 struct sk_buff *skb;
2215 skb = tcp_write_queue_head(sk);
2216 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) {
2217 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2220 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) {
2221 if (skb_cloned(skb)) {
2222 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2225 tcp_unlink_write_queue(skb, sk);
2226 skb_header_release(nskb);
2227 __tcp_add_write_queue_head(sk, nskb);
2228 sk_wmem_free_skb(sk, skb);
2229 sk->sk_wmem_queued += nskb->truesize;
2230 sk_mem_charge(sk, nskb->truesize);
2234 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK;
2235 TCP_ECN_send_synack(tcp_sk(sk), skb);
2237 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2238 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2242 * Prepare a SYN-ACK.
2244 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2245 struct request_sock *req)
2247 struct inet_request_sock *ireq = inet_rsk(req);
2248 struct tcp_sock *tp = tcp_sk(sk);
2250 int tcp_header_size;
2251 struct tcp_out_options opts;
2252 struct sk_buff *skb;
2253 struct tcp_md5sig_key *md5;
2254 __u8 *md5_hash_location;
2257 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2261 /* Reserve space for headers. */
2262 skb_reserve(skb, MAX_TCP_HEADER);
2264 skb->dst = dst_clone(dst);
2266 mss = dst_metric(dst, RTAX_ADVMSS);
2267 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2268 mss = tp->rx_opt.user_mss;
2270 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2272 /* Set this up on the first call only */
2273 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2274 /* tcp_full_space because it is guaranteed to be the first packet */
2275 tcp_select_initial_window(tcp_full_space(sk),
2276 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2281 ireq->rcv_wscale = rcv_wscale;
2284 memset(&opts, 0, sizeof(opts));
2285 #ifdef CONFIG_SYN_COOKIES
2286 if (unlikely(req->cookie_ts))
2287 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2290 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2291 tcp_header_size = tcp_synack_options(sk, req, mss,
2293 sizeof(struct tcphdr);
2295 skb_push(skb, tcp_header_size);
2296 skb_reset_transport_header(skb);
2299 memset(th, 0, sizeof(struct tcphdr));
2302 TCP_ECN_make_synack(req, th);
2303 th->source = ireq->loc_port;
2304 th->dest = ireq->rmt_port;
2305 /* Setting of flags are superfluous here for callers (and ECE is
2306 * not even correctly set)
2308 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2309 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK);
2310 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2311 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2313 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2314 th->window = htons(min(req->rcv_wnd, 65535U));
2315 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location);
2316 th->doff = (tcp_header_size >> 2);
2317 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
2319 #ifdef CONFIG_TCP_MD5SIG
2320 /* Okay, we have all we need - do the md5 hash if needed */
2322 tp->af_specific->calc_md5_hash(md5_hash_location,
2323 md5, NULL, req, skb);
2331 * Do all connect socket setups that can be done AF independent.
2333 static void tcp_connect_init(struct sock *sk)
2335 struct dst_entry *dst = __sk_dst_get(sk);
2336 struct tcp_sock *tp = tcp_sk(sk);
2339 /* We'll fix this up when we get a response from the other end.
2340 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2342 tp->tcp_header_len = sizeof(struct tcphdr) +
2343 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2345 #ifdef CONFIG_TCP_MD5SIG
2346 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2347 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2350 /* If user gave his TCP_MAXSEG, record it to clamp */
2351 if (tp->rx_opt.user_mss)
2352 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2355 tcp_sync_mss(sk, dst_mtu(dst));
2357 if (!tp->window_clamp)
2358 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2359 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2360 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2361 tp->advmss = tp->rx_opt.user_mss;
2363 tcp_initialize_rcv_mss(sk);
2365 tcp_select_initial_window(tcp_full_space(sk),
2366 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2369 sysctl_tcp_window_scaling,
2372 tp->rx_opt.rcv_wscale = rcv_wscale;
2373 tp->rcv_ssthresh = tp->rcv_wnd;
2376 sock_reset_flag(sk, SOCK_DONE);
2378 tcp_init_wl(tp, tp->write_seq, 0);
2379 tp->snd_una = tp->write_seq;
2380 tp->snd_sml = tp->write_seq;
2381 tp->snd_up = tp->write_seq;
2386 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2387 inet_csk(sk)->icsk_retransmits = 0;
2388 tcp_clear_retrans(tp);
2392 * Build a SYN and send it off.
2394 int tcp_connect(struct sock *sk)
2396 struct tcp_sock *tp = tcp_sk(sk);
2397 struct sk_buff *buff;
2399 tcp_connect_init(sk);
2401 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2402 if (unlikely(buff == NULL))
2405 /* Reserve space for headers. */
2406 skb_reserve(buff, MAX_TCP_HEADER);
2408 tp->snd_nxt = tp->write_seq;
2409 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN);
2410 TCP_ECN_send_syn(sk, buff);
2413 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2414 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2415 skb_header_release(buff);
2416 __tcp_add_write_queue_tail(sk, buff);
2417 sk->sk_wmem_queued += buff->truesize;
2418 sk_mem_charge(sk, buff->truesize);
2419 tp->packets_out += tcp_skb_pcount(buff);
2420 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL);
2422 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2423 * in order to make this packet get counted in tcpOutSegs.
2425 tp->snd_nxt = tp->write_seq;
2426 tp->pushed_seq = tp->write_seq;
2427 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2429 /* Timer for repeating the SYN until an answer. */
2430 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2431 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2435 /* Send out a delayed ack, the caller does the policy checking
2436 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2439 void tcp_send_delayed_ack(struct sock *sk)
2441 struct inet_connection_sock *icsk = inet_csk(sk);
2442 int ato = icsk->icsk_ack.ato;
2443 unsigned long timeout;
2445 if (ato > TCP_DELACK_MIN) {
2446 const struct tcp_sock *tp = tcp_sk(sk);
2447 int max_ato = HZ / 2;
2449 if (icsk->icsk_ack.pingpong ||
2450 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2451 max_ato = TCP_DELACK_MAX;
2453 /* Slow path, intersegment interval is "high". */
2455 /* If some rtt estimate is known, use it to bound delayed ack.
2456 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2460 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2466 ato = min(ato, max_ato);
2469 /* Stay within the limit we were given */
2470 timeout = jiffies + ato;
2472 /* Use new timeout only if there wasn't a older one earlier. */
2473 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2474 /* If delack timer was blocked or is about to expire,
2477 if (icsk->icsk_ack.blocked ||
2478 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2483 if (!time_before(timeout, icsk->icsk_ack.timeout))
2484 timeout = icsk->icsk_ack.timeout;
2486 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2487 icsk->icsk_ack.timeout = timeout;
2488 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2491 /* This routine sends an ack and also updates the window. */
2492 void tcp_send_ack(struct sock *sk)
2494 struct sk_buff *buff;
2496 /* If we have been reset, we may not send again. */
2497 if (sk->sk_state == TCP_CLOSE)
2500 /* We are not putting this on the write queue, so
2501 * tcp_transmit_skb() will set the ownership to this
2504 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2506 inet_csk_schedule_ack(sk);
2507 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2508 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2509 TCP_DELACK_MAX, TCP_RTO_MAX);
2513 /* Reserve space for headers and prepare control bits. */
2514 skb_reserve(buff, MAX_TCP_HEADER);
2515 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK);
2517 /* Send it off, this clears delayed acks for us. */
2518 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2519 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2522 /* This routine sends a packet with an out of date sequence
2523 * number. It assumes the other end will try to ack it.
2525 * Question: what should we make while urgent mode?
2526 * 4.4BSD forces sending single byte of data. We cannot send
2527 * out of window data, because we have SND.NXT==SND.MAX...
2529 * Current solution: to send TWO zero-length segments in urgent mode:
2530 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2531 * out-of-date with SND.UNA-1 to probe window.
2533 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2535 struct tcp_sock *tp = tcp_sk(sk);
2536 struct sk_buff *skb;
2538 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2539 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2543 /* Reserve space for headers and set control bits. */
2544 skb_reserve(skb, MAX_TCP_HEADER);
2545 /* Use a previous sequence. This should cause the other
2546 * end to send an ack. Don't queue or clone SKB, just
2549 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK);
2550 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2551 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2554 int tcp_write_wakeup(struct sock *sk)
2556 struct tcp_sock *tp = tcp_sk(sk);
2557 struct sk_buff *skb;
2559 if (sk->sk_state == TCP_CLOSE)
2562 if ((skb = tcp_send_head(sk)) != NULL &&
2563 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2565 unsigned int mss = tcp_current_mss(sk, 0);
2566 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2568 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2569 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2571 /* We are probing the opening of a window
2572 * but the window size is != 0
2573 * must have been a result SWS avoidance ( sender )
2575 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2577 seg_size = min(seg_size, mss);
2578 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2579 if (tcp_fragment(sk, skb, seg_size, mss))
2581 } else if (!tcp_skb_pcount(skb))
2582 tcp_set_skb_tso_segs(sk, skb, mss);
2584 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH;
2585 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2586 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2588 tcp_event_new_data_sent(sk, skb);
2591 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2592 tcp_xmit_probe_skb(sk, 1);
2593 return tcp_xmit_probe_skb(sk, 0);
2597 /* A window probe timeout has occurred. If window is not closed send
2598 * a partial packet else a zero probe.
2600 void tcp_send_probe0(struct sock *sk)
2602 struct inet_connection_sock *icsk = inet_csk(sk);
2603 struct tcp_sock *tp = tcp_sk(sk);
2606 err = tcp_write_wakeup(sk);
2608 if (tp->packets_out || !tcp_send_head(sk)) {
2609 /* Cancel probe timer, if it is not required. */
2610 icsk->icsk_probes_out = 0;
2611 icsk->icsk_backoff = 0;
2616 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2617 icsk->icsk_backoff++;
2618 icsk->icsk_probes_out++;
2619 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2620 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2623 /* If packet was not sent due to local congestion,
2624 * do not backoff and do not remember icsk_probes_out.
2625 * Let local senders to fight for local resources.
2627 * Use accumulated backoff yet.
2629 if (!icsk->icsk_probes_out)
2630 icsk->icsk_probes_out = 1;
2631 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2632 min(icsk->icsk_rto << icsk->icsk_backoff,
2633 TCP_RESOURCE_PROBE_INTERVAL),
2638 EXPORT_SYMBOL(tcp_select_initial_window);
2639 EXPORT_SYMBOL(tcp_connect);
2640 EXPORT_SYMBOL(tcp_make_synack);
2641 EXPORT_SYMBOL(tcp_simple_retransmit);
2642 EXPORT_SYMBOL(tcp_sync_mss);
2643 EXPORT_SYMBOL(tcp_mtup_init);