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/gfp.h>
41 #include <linux/module.h>
43 /* People can turn this off for buggy TCP's found in printers etc. */
44 int sysctl_tcp_retrans_collapse __read_mostly = 1;
46 /* People can turn this on to work with those rare, broken TCPs that
47 * interpret the window field as a signed quantity.
49 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
51 /* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
55 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
57 int sysctl_tcp_mtu_probing __read_mostly = 0;
58 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
60 /* By default, RFC2861 behavior. */
61 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
63 int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
64 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
67 /* Account for new data that has been sent to the network. */
68 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
70 struct tcp_sock *tp = tcp_sk(sk);
71 unsigned int prior_packets = tp->packets_out;
73 tcp_advance_send_head(sk, skb);
74 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
76 /* Don't override Nagle indefinitely with F-RTO */
77 if (tp->frto_counter == 2)
80 tp->packets_out += tcp_skb_pcount(skb);
82 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
83 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
86 /* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
92 static inline __u32 tcp_acceptable_seq(struct sock *sk)
94 struct tcp_sock *tp = tcp_sk(sk);
96 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 return tcp_wnd_end(tp);
102 /* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
116 static __u16 tcp_advertise_mss(struct sock *sk)
118 struct tcp_sock *tp = tcp_sk(sk);
119 struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
123 unsigned int metric = dst_metric_advmss(dst);
134 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
136 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
138 struct tcp_sock *tp = tcp_sk(sk);
139 s32 delta = tcp_time_stamp - tp->lsndtime;
140 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
141 u32 cwnd = tp->snd_cwnd;
143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 restart_cwnd = min(restart_cwnd, cwnd);
148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_time_stamp;
152 tp->snd_cwnd_used = 0;
155 /* Congestion state accounting after a packet has been sent. */
156 static void tcp_event_data_sent(struct tcp_sock *tp,
157 struct sk_buff *skb, struct sock *sk)
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_time_stamp;
162 if (sysctl_tcp_slow_start_after_idle &&
163 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
164 tcp_cwnd_restart(sk, __sk_dst_get(sk));
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
171 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 icsk->icsk_ack.pingpong = 1;
175 /* Account for an ACK we sent. */
176 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
182 /* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
189 void tcp_select_initial_window(int __space, __u32 mss,
190 __u32 *rcv_wnd, __u32 *window_clamp,
191 int wscale_ok, __u8 *rcv_wscale,
194 unsigned int space = (__space < 0 ? 0 : __space);
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp == 0)
198 (*window_clamp) = (65535 << 14);
199 space = min(*window_clamp, space);
201 /* Quantize space offering to a multiple of mss if possible. */
203 space = (space / mss) * mss;
205 /* NOTE: offering an initial window larger than 32767
206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
213 if (sysctl_tcp_workaround_signed_windows)
214 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
220 /* Set window scaling on max possible window
221 * See RFC1323 for an explanation of the limit to 14
223 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
224 space = min_t(u32, space, *window_clamp);
225 while (space > 65535 && (*rcv_wscale) < 14) {
231 /* Set initial window to a value enough for senders starting with
232 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 * a limit on the initial window when mss is larger than 1460.
235 if (mss > (1 << *rcv_wscale)) {
236 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
239 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
240 /* when initializing use the value from init_rcv_wnd
241 * rather than the default from above
244 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
246 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 cur_win = tcp_receive_window(tp);
263 u32 new_win = __tcp_select_window(sk);
265 /* Never shrink the offered window */
266 if (new_win < cur_win) {
267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
272 * Relax Will Robinson.
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
279 /* Make sure we do not exceed the maximum possible
282 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
283 new_win = min(new_win, MAX_TCP_WINDOW);
285 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
287 /* RFC1323 scaling applied */
288 new_win >>= tp->rx_opt.rcv_wscale;
290 /* If we advertise zero window, disable fast path. */
297 /* Packet ECN state for a SYN-ACK */
298 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
300 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
301 if (!(tp->ecn_flags & TCP_ECN_OK))
302 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
305 /* Packet ECN state for a SYN. */
306 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
308 struct tcp_sock *tp = tcp_sk(sk);
311 if (sysctl_tcp_ecn == 1) {
312 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
313 tp->ecn_flags = TCP_ECN_OK;
317 static __inline__ void
318 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
320 if (inet_rsk(req)->ecn_ok)
324 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
327 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
330 struct tcp_sock *tp = tcp_sk(sk);
332 if (tp->ecn_flags & TCP_ECN_OK) {
333 /* Not-retransmitted data segment: set ECT and inject CWR. */
334 if (skb->len != tcp_header_len &&
335 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
337 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
338 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
339 tcp_hdr(skb)->cwr = 1;
340 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
343 /* ACK or retransmitted segment: clear ECT|CE */
344 INET_ECN_dontxmit(sk);
346 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
347 tcp_hdr(skb)->ece = 1;
351 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
352 * auto increment end seqno.
354 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
356 skb->ip_summed = CHECKSUM_PARTIAL;
359 TCP_SKB_CB(skb)->tcp_flags = flags;
360 TCP_SKB_CB(skb)->sacked = 0;
362 skb_shinfo(skb)->gso_segs = 1;
363 skb_shinfo(skb)->gso_size = 0;
364 skb_shinfo(skb)->gso_type = 0;
366 TCP_SKB_CB(skb)->seq = seq;
367 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
369 TCP_SKB_CB(skb)->end_seq = seq;
372 static inline int tcp_urg_mode(const struct tcp_sock *tp)
374 return tp->snd_una != tp->snd_up;
377 #define OPTION_SACK_ADVERTISE (1 << 0)
378 #define OPTION_TS (1 << 1)
379 #define OPTION_MD5 (1 << 2)
380 #define OPTION_WSCALE (1 << 3)
381 #define OPTION_COOKIE_EXTENSION (1 << 4)
383 struct tcp_out_options {
384 u8 options; /* bit field of OPTION_* */
385 u8 ws; /* window scale, 0 to disable */
386 u8 num_sack_blocks; /* number of SACK blocks to include */
387 u8 hash_size; /* bytes in hash_location */
388 u16 mss; /* 0 to disable */
389 __u32 tsval, tsecr; /* need to include OPTION_TS */
390 __u8 *hash_location; /* temporary pointer, overloaded */
393 /* The sysctl int routines are generic, so check consistency here.
395 static u8 tcp_cookie_size_check(u8 desired)
400 /* previously specified */
403 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
404 if (cookie_size <= 0)
405 /* no default specified */
408 if (cookie_size <= TCP_COOKIE_MIN)
409 /* value too small, specify minimum */
410 return TCP_COOKIE_MIN;
412 if (cookie_size >= TCP_COOKIE_MAX)
413 /* value too large, specify maximum */
414 return TCP_COOKIE_MAX;
417 /* 8-bit multiple, illegal, fix it */
420 return (u8)cookie_size;
423 /* Write previously computed TCP options to the packet.
425 * Beware: Something in the Internet is very sensitive to the ordering of
426 * TCP options, we learned this through the hard way, so be careful here.
427 * Luckily we can at least blame others for their non-compliance but from
428 * inter-operatibility perspective it seems that we're somewhat stuck with
429 * the ordering which we have been using if we want to keep working with
430 * those broken things (not that it currently hurts anybody as there isn't
431 * particular reason why the ordering would need to be changed).
433 * At least SACK_PERM as the first option is known to lead to a disaster
434 * (but it may well be that other scenarios fail similarly).
436 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
437 struct tcp_out_options *opts)
439 u8 options = opts->options; /* mungable copy */
441 /* Having both authentication and cookies for security is redundant,
442 * and there's certainly not enough room. Instead, the cookie-less
443 * extension variant is proposed.
445 * Consider the pessimal case with authentication. The options
447 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
449 if (unlikely(OPTION_MD5 & options)) {
450 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
451 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
452 (TCPOLEN_COOKIE_BASE << 16) |
453 (TCPOPT_MD5SIG << 8) |
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
458 (TCPOPT_MD5SIG << 8) |
461 options &= ~OPTION_COOKIE_EXTENSION;
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
479 options &= ~OPTION_SACK_ADVERTISE;
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_TIMESTAMP << 8) |
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
490 /* Specification requires after timestamp, so do it now.
492 * Consider the pessimal case without authentication. The options
494 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
496 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
497 __u8 *cookie_copy = opts->hash_location;
498 u8 cookie_size = opts->hash_size;
500 /* 8-bit multiple handled in tcp_cookie_size_check() above,
503 if (0x2 & cookie_size) {
504 __u8 *p = (__u8 *)ptr;
506 /* 16-bit multiple */
507 *p++ = TCPOPT_COOKIE;
508 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
509 *p++ = *cookie_copy++;
510 *p++ = *cookie_copy++;
514 /* 32-bit multiple */
515 *ptr++ = htonl(((TCPOPT_NOP << 24) |
517 (TCPOPT_COOKIE << 8) |
518 TCPOLEN_COOKIE_BASE) +
522 if (cookie_size > 0) {
523 memcpy(ptr, cookie_copy, cookie_size);
524 ptr += (cookie_size / 4);
528 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
529 *ptr++ = htonl((TCPOPT_NOP << 24) |
531 (TCPOPT_SACK_PERM << 8) |
535 if (unlikely(OPTION_WSCALE & options)) {
536 *ptr++ = htonl((TCPOPT_NOP << 24) |
537 (TCPOPT_WINDOW << 16) |
538 (TCPOLEN_WINDOW << 8) |
542 if (unlikely(opts->num_sack_blocks)) {
543 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
544 tp->duplicate_sack : tp->selective_acks;
547 *ptr++ = htonl((TCPOPT_NOP << 24) |
550 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
551 TCPOLEN_SACK_PERBLOCK)));
553 for (this_sack = 0; this_sack < opts->num_sack_blocks;
555 *ptr++ = htonl(sp[this_sack].start_seq);
556 *ptr++ = htonl(sp[this_sack].end_seq);
559 tp->rx_opt.dsack = 0;
563 /* Compute TCP options for SYN packets. This is not the final
564 * network wire format yet.
566 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
567 struct tcp_out_options *opts,
568 struct tcp_md5sig_key **md5) {
569 struct tcp_sock *tp = tcp_sk(sk);
570 struct tcp_cookie_values *cvp = tp->cookie_values;
571 unsigned remaining = MAX_TCP_OPTION_SPACE;
572 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
573 tcp_cookie_size_check(cvp->cookie_desired) :
576 #ifdef CONFIG_TCP_MD5SIG
577 *md5 = tp->af_specific->md5_lookup(sk, sk);
579 opts->options |= OPTION_MD5;
580 remaining -= TCPOLEN_MD5SIG_ALIGNED;
586 /* We always get an MSS option. The option bytes which will be seen in
587 * normal data packets should timestamps be used, must be in the MSS
588 * advertised. But we subtract them from tp->mss_cache so that
589 * calculations in tcp_sendmsg are simpler etc. So account for this
590 * fact here if necessary. If we don't do this correctly, as a
591 * receiver we won't recognize data packets as being full sized when we
592 * should, and thus we won't abide by the delayed ACK rules correctly.
593 * SACKs don't matter, we never delay an ACK when we have any of those
595 opts->mss = tcp_advertise_mss(sk);
596 remaining -= TCPOLEN_MSS_ALIGNED;
598 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
599 opts->options |= OPTION_TS;
600 opts->tsval = TCP_SKB_CB(skb)->when;
601 opts->tsecr = tp->rx_opt.ts_recent;
602 remaining -= TCPOLEN_TSTAMP_ALIGNED;
604 if (likely(sysctl_tcp_window_scaling)) {
605 opts->ws = tp->rx_opt.rcv_wscale;
606 opts->options |= OPTION_WSCALE;
607 remaining -= TCPOLEN_WSCALE_ALIGNED;
609 if (likely(sysctl_tcp_sack)) {
610 opts->options |= OPTION_SACK_ADVERTISE;
611 if (unlikely(!(OPTION_TS & opts->options)))
612 remaining -= TCPOLEN_SACKPERM_ALIGNED;
615 /* Note that timestamps are required by the specification.
617 * Odd numbers of bytes are prohibited by the specification, ensuring
618 * that the cookie is 16-bit aligned, and the resulting cookie pair is
622 (OPTION_TS & opts->options) &&
624 int need = TCPOLEN_COOKIE_BASE + cookie_size;
627 /* 32-bit multiple */
628 need += 2; /* NOPs */
630 if (need > remaining) {
631 /* try shrinking cookie to fit */
636 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
640 if (TCP_COOKIE_MIN <= cookie_size) {
641 opts->options |= OPTION_COOKIE_EXTENSION;
642 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
643 opts->hash_size = cookie_size;
645 /* Remember for future incarnations. */
646 cvp->cookie_desired = cookie_size;
648 if (cvp->cookie_desired != cvp->cookie_pair_size) {
649 /* Currently use random bytes as a nonce,
650 * assuming these are completely unpredictable
651 * by hostile users of the same system.
653 get_random_bytes(&cvp->cookie_pair[0],
655 cvp->cookie_pair_size = cookie_size;
661 return MAX_TCP_OPTION_SPACE - remaining;
664 /* Set up TCP options for SYN-ACKs. */
665 static unsigned tcp_synack_options(struct sock *sk,
666 struct request_sock *req,
667 unsigned mss, struct sk_buff *skb,
668 struct tcp_out_options *opts,
669 struct tcp_md5sig_key **md5,
670 struct tcp_extend_values *xvp)
672 struct inet_request_sock *ireq = inet_rsk(req);
673 unsigned remaining = MAX_TCP_OPTION_SPACE;
674 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
678 #ifdef CONFIG_TCP_MD5SIG
679 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
681 opts->options |= OPTION_MD5;
682 remaining -= TCPOLEN_MD5SIG_ALIGNED;
684 /* We can't fit any SACK blocks in a packet with MD5 + TS
685 * options. There was discussion about disabling SACK
686 * rather than TS in order to fit in better with old,
687 * buggy kernels, but that was deemed to be unnecessary.
689 ireq->tstamp_ok &= !ireq->sack_ok;
695 /* We always send an MSS option. */
697 remaining -= TCPOLEN_MSS_ALIGNED;
699 if (likely(ireq->wscale_ok)) {
700 opts->ws = ireq->rcv_wscale;
701 opts->options |= OPTION_WSCALE;
702 remaining -= TCPOLEN_WSCALE_ALIGNED;
704 if (likely(ireq->tstamp_ok)) {
705 opts->options |= OPTION_TS;
706 opts->tsval = TCP_SKB_CB(skb)->when;
707 opts->tsecr = req->ts_recent;
708 remaining -= TCPOLEN_TSTAMP_ALIGNED;
710 if (likely(ireq->sack_ok)) {
711 opts->options |= OPTION_SACK_ADVERTISE;
712 if (unlikely(!ireq->tstamp_ok))
713 remaining -= TCPOLEN_SACKPERM_ALIGNED;
716 /* Similar rationale to tcp_syn_options() applies here, too.
717 * If the <SYN> options fit, the same options should fit now!
721 cookie_plus > TCPOLEN_COOKIE_BASE) {
722 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
725 /* 32-bit multiple */
726 need += 2; /* NOPs */
728 if (need <= remaining) {
729 opts->options |= OPTION_COOKIE_EXTENSION;
730 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
733 /* There's no error return, so flag it. */
734 xvp->cookie_out_never = 1; /* true */
738 return MAX_TCP_OPTION_SPACE - remaining;
741 /* Compute TCP options for ESTABLISHED sockets. This is not the
742 * final wire format yet.
744 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
745 struct tcp_out_options *opts,
746 struct tcp_md5sig_key **md5) {
747 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
748 struct tcp_sock *tp = tcp_sk(sk);
750 unsigned int eff_sacks;
752 #ifdef CONFIG_TCP_MD5SIG
753 *md5 = tp->af_specific->md5_lookup(sk, sk);
754 if (unlikely(*md5)) {
755 opts->options |= OPTION_MD5;
756 size += TCPOLEN_MD5SIG_ALIGNED;
762 if (likely(tp->rx_opt.tstamp_ok)) {
763 opts->options |= OPTION_TS;
764 opts->tsval = tcb ? tcb->when : 0;
765 opts->tsecr = tp->rx_opt.ts_recent;
766 size += TCPOLEN_TSTAMP_ALIGNED;
769 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 if (unlikely(eff_sacks)) {
771 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
772 opts->num_sack_blocks =
773 min_t(unsigned, eff_sacks,
774 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 TCPOLEN_SACK_PERBLOCK);
776 size += TCPOLEN_SACK_BASE_ALIGNED +
777 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
783 /* This routine actually transmits TCP packets queued in by
784 * tcp_do_sendmsg(). This is used by both the initial
785 * transmission and possible later retransmissions.
786 * All SKB's seen here are completely headerless. It is our
787 * job to build the TCP header, and pass the packet down to
788 * IP so it can do the same plus pass the packet off to the
791 * We are working here with either a clone of the original
792 * SKB, or a fresh unique copy made by the retransmit engine.
794 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
797 const struct inet_connection_sock *icsk = inet_csk(sk);
798 struct inet_sock *inet;
800 struct tcp_skb_cb *tcb;
801 struct tcp_out_options opts;
802 unsigned tcp_options_size, tcp_header_size;
803 struct tcp_md5sig_key *md5;
807 BUG_ON(!skb || !tcp_skb_pcount(skb));
809 /* If congestion control is doing timestamping, we must
810 * take such a timestamp before we potentially clone/copy.
812 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
813 __net_timestamp(skb);
815 if (likely(clone_it)) {
816 if (unlikely(skb_cloned(skb)))
817 skb = pskb_copy(skb, gfp_mask);
819 skb = skb_clone(skb, gfp_mask);
826 tcb = TCP_SKB_CB(skb);
827 memset(&opts, 0, sizeof(opts));
829 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
830 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
832 tcp_options_size = tcp_established_options(sk, skb, &opts,
834 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
836 if (tcp_packets_in_flight(tp) == 0) {
837 tcp_ca_event(sk, CA_EVENT_TX_START);
842 skb_push(skb, tcp_header_size);
843 skb_reset_transport_header(skb);
844 skb_set_owner_w(skb, sk);
846 /* Build TCP header and checksum it. */
848 th->source = inet->inet_sport;
849 th->dest = inet->inet_dport;
850 th->seq = htonl(tcb->seq);
851 th->ack_seq = htonl(tp->rcv_nxt);
852 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
855 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
856 /* RFC1323: The window in SYN & SYN/ACK segments
859 th->window = htons(min(tp->rcv_wnd, 65535U));
861 th->window = htons(tcp_select_window(sk));
866 /* The urg_mode check is necessary during a below snd_una win probe */
867 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
868 if (before(tp->snd_up, tcb->seq + 0x10000)) {
869 th->urg_ptr = htons(tp->snd_up - tcb->seq);
871 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
872 th->urg_ptr = htons(0xFFFF);
877 tcp_options_write((__be32 *)(th + 1), tp, &opts);
878 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
879 TCP_ECN_send(sk, skb, tcp_header_size);
881 #ifdef CONFIG_TCP_MD5SIG
882 /* Calculate the MD5 hash, as we have all we need now */
884 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
885 tp->af_specific->calc_md5_hash(opts.hash_location,
890 icsk->icsk_af_ops->send_check(sk, skb);
892 if (likely(tcb->tcp_flags & TCPHDR_ACK))
893 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
895 if (skb->len != tcp_header_size)
896 tcp_event_data_sent(tp, skb, sk);
898 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
899 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
900 tcp_skb_pcount(skb));
902 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
903 if (likely(err <= 0))
906 tcp_enter_cwr(sk, 1);
908 return net_xmit_eval(err);
911 /* This routine just queues the buffer for sending.
913 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
914 * otherwise socket can stall.
916 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
918 struct tcp_sock *tp = tcp_sk(sk);
920 /* Advance write_seq and place onto the write_queue. */
921 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
922 skb_header_release(skb);
923 tcp_add_write_queue_tail(sk, skb);
924 sk->sk_wmem_queued += skb->truesize;
925 sk_mem_charge(sk, skb->truesize);
928 /* Initialize TSO segments for a packet. */
929 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
930 unsigned int mss_now)
932 if (skb->len <= mss_now || !sk_can_gso(sk) ||
933 skb->ip_summed == CHECKSUM_NONE) {
934 /* Avoid the costly divide in the normal
937 skb_shinfo(skb)->gso_segs = 1;
938 skb_shinfo(skb)->gso_size = 0;
939 skb_shinfo(skb)->gso_type = 0;
941 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
942 skb_shinfo(skb)->gso_size = mss_now;
943 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
947 /* When a modification to fackets out becomes necessary, we need to check
948 * skb is counted to fackets_out or not.
950 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
953 struct tcp_sock *tp = tcp_sk(sk);
955 if (!tp->sacked_out || tcp_is_reno(tp))
958 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
959 tp->fackets_out -= decr;
962 /* Pcount in the middle of the write queue got changed, we need to do various
963 * tweaks to fix counters
965 static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
967 struct tcp_sock *tp = tcp_sk(sk);
969 tp->packets_out -= decr;
971 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
972 tp->sacked_out -= decr;
973 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
974 tp->retrans_out -= decr;
975 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
976 tp->lost_out -= decr;
978 /* Reno case is special. Sigh... */
979 if (tcp_is_reno(tp) && decr > 0)
980 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
982 tcp_adjust_fackets_out(sk, skb, decr);
984 if (tp->lost_skb_hint &&
985 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
986 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
987 tp->lost_cnt_hint -= decr;
989 tcp_verify_left_out(tp);
992 /* Function to create two new TCP segments. Shrinks the given segment
993 * to the specified size and appends a new segment with the rest of the
994 * packet to the list. This won't be called frequently, I hope.
995 * Remember, these are still headerless SKBs at this point.
997 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
998 unsigned int mss_now)
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 struct sk_buff *buff;
1002 int nsize, old_factor;
1006 if (WARN_ON(len > skb->len))
1009 nsize = skb_headlen(skb) - len;
1013 if (skb_cloned(skb) &&
1014 skb_is_nonlinear(skb) &&
1015 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1018 /* Get a new skb... force flag on. */
1019 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1021 return -ENOMEM; /* We'll just try again later. */
1023 sk->sk_wmem_queued += buff->truesize;
1024 sk_mem_charge(sk, buff->truesize);
1025 nlen = skb->len - len - nsize;
1026 buff->truesize += nlen;
1027 skb->truesize -= nlen;
1029 /* Correct the sequence numbers. */
1030 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1031 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1032 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1034 /* PSH and FIN should only be set in the second packet. */
1035 flags = TCP_SKB_CB(skb)->tcp_flags;
1036 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1037 TCP_SKB_CB(buff)->tcp_flags = flags;
1038 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1040 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1041 /* Copy and checksum data tail into the new buffer. */
1042 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1043 skb_put(buff, nsize),
1048 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1050 skb->ip_summed = CHECKSUM_PARTIAL;
1051 skb_split(skb, buff, len);
1054 buff->ip_summed = skb->ip_summed;
1056 /* Looks stupid, but our code really uses when of
1057 * skbs, which it never sent before. --ANK
1059 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1060 buff->tstamp = skb->tstamp;
1062 old_factor = tcp_skb_pcount(skb);
1064 /* Fix up tso_factor for both original and new SKB. */
1065 tcp_set_skb_tso_segs(sk, skb, mss_now);
1066 tcp_set_skb_tso_segs(sk, buff, mss_now);
1068 /* If this packet has been sent out already, we must
1069 * adjust the various packet counters.
1071 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1072 int diff = old_factor - tcp_skb_pcount(skb) -
1073 tcp_skb_pcount(buff);
1076 tcp_adjust_pcount(sk, skb, diff);
1079 /* Link BUFF into the send queue. */
1080 skb_header_release(buff);
1081 tcp_insert_write_queue_after(skb, buff, sk);
1086 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1087 * eventually). The difference is that pulled data not copied, but
1088 * immediately discarded.
1090 static void __pskb_trim_head(struct sk_buff *skb, int len)
1096 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1097 if (skb_shinfo(skb)->frags[i].size <= eat) {
1098 skb_frag_unref(skb, i);
1099 eat -= skb_shinfo(skb)->frags[i].size;
1101 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1103 skb_shinfo(skb)->frags[k].page_offset += eat;
1104 skb_shinfo(skb)->frags[k].size -= eat;
1110 skb_shinfo(skb)->nr_frags = k;
1112 skb_reset_tail_pointer(skb);
1113 skb->data_len -= len;
1114 skb->len = skb->data_len;
1117 /* Remove acked data from a packet in the transmit queue. */
1118 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1120 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1123 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1124 if (unlikely(len < skb_headlen(skb)))
1125 __skb_pull(skb, len);
1127 __pskb_trim_head(skb, len - skb_headlen(skb));
1129 TCP_SKB_CB(skb)->seq += len;
1130 skb->ip_summed = CHECKSUM_PARTIAL;
1132 skb->truesize -= len;
1133 sk->sk_wmem_queued -= len;
1134 sk_mem_uncharge(sk, len);
1135 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1137 /* Any change of skb->len requires recalculation of tso
1140 if (tcp_skb_pcount(skb) > 1)
1141 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1146 /* Calculate MSS. Not accounting for SACKs here. */
1147 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1149 struct tcp_sock *tp = tcp_sk(sk);
1150 struct inet_connection_sock *icsk = inet_csk(sk);
1153 /* Calculate base mss without TCP options:
1154 It is MMS_S - sizeof(tcphdr) of rfc1122
1156 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1158 /* Clamp it (mss_clamp does not include tcp options) */
1159 if (mss_now > tp->rx_opt.mss_clamp)
1160 mss_now = tp->rx_opt.mss_clamp;
1162 /* Now subtract optional transport overhead */
1163 mss_now -= icsk->icsk_ext_hdr_len;
1165 /* Then reserve room for full set of TCP options and 8 bytes of data */
1169 /* Now subtract TCP options size, not including SACKs */
1170 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1175 /* Inverse of above */
1176 int tcp_mss_to_mtu(struct sock *sk, int mss)
1178 struct tcp_sock *tp = tcp_sk(sk);
1179 struct inet_connection_sock *icsk = inet_csk(sk);
1183 tp->tcp_header_len +
1184 icsk->icsk_ext_hdr_len +
1185 icsk->icsk_af_ops->net_header_len;
1190 /* MTU probing init per socket */
1191 void tcp_mtup_init(struct sock *sk)
1193 struct tcp_sock *tp = tcp_sk(sk);
1194 struct inet_connection_sock *icsk = inet_csk(sk);
1196 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1197 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1198 icsk->icsk_af_ops->net_header_len;
1199 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1200 icsk->icsk_mtup.probe_size = 0;
1202 EXPORT_SYMBOL(tcp_mtup_init);
1204 /* This function synchronize snd mss to current pmtu/exthdr set.
1206 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1207 for TCP options, but includes only bare TCP header.
1209 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1210 It is minimum of user_mss and mss received with SYN.
1211 It also does not include TCP options.
1213 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1215 tp->mss_cache is current effective sending mss, including
1216 all tcp options except for SACKs. It is evaluated,
1217 taking into account current pmtu, but never exceeds
1218 tp->rx_opt.mss_clamp.
1220 NOTE1. rfc1122 clearly states that advertised MSS
1221 DOES NOT include either tcp or ip options.
1223 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1224 are READ ONLY outside this function. --ANK (980731)
1226 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1228 struct tcp_sock *tp = tcp_sk(sk);
1229 struct inet_connection_sock *icsk = inet_csk(sk);
1232 if (icsk->icsk_mtup.search_high > pmtu)
1233 icsk->icsk_mtup.search_high = pmtu;
1235 mss_now = tcp_mtu_to_mss(sk, pmtu);
1236 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1238 /* And store cached results */
1239 icsk->icsk_pmtu_cookie = pmtu;
1240 if (icsk->icsk_mtup.enabled)
1241 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1242 tp->mss_cache = mss_now;
1246 EXPORT_SYMBOL(tcp_sync_mss);
1248 /* Compute the current effective MSS, taking SACKs and IP options,
1249 * and even PMTU discovery events into account.
1251 unsigned int tcp_current_mss(struct sock *sk)
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 struct dst_entry *dst = __sk_dst_get(sk);
1256 unsigned header_len;
1257 struct tcp_out_options opts;
1258 struct tcp_md5sig_key *md5;
1260 mss_now = tp->mss_cache;
1263 u32 mtu = dst_mtu(dst);
1264 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1265 mss_now = tcp_sync_mss(sk, mtu);
1268 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1269 sizeof(struct tcphdr);
1270 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1271 * some common options. If this is an odd packet (because we have SACK
1272 * blocks etc) then our calculated header_len will be different, and
1273 * we have to adjust mss_now correspondingly */
1274 if (header_len != tp->tcp_header_len) {
1275 int delta = (int) header_len - tp->tcp_header_len;
1282 /* Congestion window validation. (RFC2861) */
1283 static void tcp_cwnd_validate(struct sock *sk)
1285 struct tcp_sock *tp = tcp_sk(sk);
1287 if (tp->packets_out >= tp->snd_cwnd) {
1288 /* Network is feed fully. */
1289 tp->snd_cwnd_used = 0;
1290 tp->snd_cwnd_stamp = tcp_time_stamp;
1292 /* Network starves. */
1293 if (tp->packets_out > tp->snd_cwnd_used)
1294 tp->snd_cwnd_used = tp->packets_out;
1296 if (sysctl_tcp_slow_start_after_idle &&
1297 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1298 tcp_cwnd_application_limited(sk);
1302 /* Returns the portion of skb which can be sent right away without
1303 * introducing MSS oddities to segment boundaries. In rare cases where
1304 * mss_now != mss_cache, we will request caller to create a small skb
1305 * per input skb which could be mostly avoided here (if desired).
1307 * We explicitly want to create a request for splitting write queue tail
1308 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1309 * thus all the complexity (cwnd_len is always MSS multiple which we
1310 * return whenever allowed by the other factors). Basically we need the
1311 * modulo only when the receiver window alone is the limiting factor or
1312 * when we would be allowed to send the split-due-to-Nagle skb fully.
1314 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1315 unsigned int mss_now, unsigned int cwnd)
1317 struct tcp_sock *tp = tcp_sk(sk);
1318 u32 needed, window, cwnd_len;
1320 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1321 cwnd_len = mss_now * cwnd;
1323 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1326 needed = min(skb->len, window);
1328 if (cwnd_len <= needed)
1331 return needed - needed % mss_now;
1334 /* Can at least one segment of SKB be sent right now, according to the
1335 * congestion window rules? If so, return how many segments are allowed.
1337 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1338 struct sk_buff *skb)
1340 u32 in_flight, cwnd;
1342 /* Don't be strict about the congestion window for the final FIN. */
1343 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1344 tcp_skb_pcount(skb) == 1)
1347 in_flight = tcp_packets_in_flight(tp);
1348 cwnd = tp->snd_cwnd;
1349 if (in_flight < cwnd)
1350 return (cwnd - in_flight);
1355 /* Initialize TSO state of a skb.
1356 * This must be invoked the first time we consider transmitting
1357 * SKB onto the wire.
1359 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1360 unsigned int mss_now)
1362 int tso_segs = tcp_skb_pcount(skb);
1364 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1365 tcp_set_skb_tso_segs(sk, skb, mss_now);
1366 tso_segs = tcp_skb_pcount(skb);
1371 /* Minshall's variant of the Nagle send check. */
1372 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1374 return after(tp->snd_sml, tp->snd_una) &&
1375 !after(tp->snd_sml, tp->snd_nxt);
1378 /* Return 0, if packet can be sent now without violation Nagle's rules:
1379 * 1. It is full sized.
1380 * 2. Or it contains FIN. (already checked by caller)
1381 * 3. Or TCP_NODELAY was set.
1382 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1383 * With Minshall's modification: all sent small packets are ACKed.
1385 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1386 const struct sk_buff *skb,
1387 unsigned mss_now, int nonagle)
1389 return skb->len < mss_now &&
1390 ((nonagle & TCP_NAGLE_CORK) ||
1391 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1394 /* Return non-zero if the Nagle test allows this packet to be
1397 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1398 unsigned int cur_mss, int nonagle)
1400 /* Nagle rule does not apply to frames, which sit in the middle of the
1401 * write_queue (they have no chances to get new data).
1403 * This is implemented in the callers, where they modify the 'nonagle'
1404 * argument based upon the location of SKB in the send queue.
1406 if (nonagle & TCP_NAGLE_PUSH)
1409 /* Don't use the nagle rule for urgent data (or for the final FIN).
1410 * Nagle can be ignored during F-RTO too (see RFC4138).
1412 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1413 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1416 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1422 /* Does at least the first segment of SKB fit into the send window? */
1423 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1424 unsigned int cur_mss)
1426 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1428 if (skb->len > cur_mss)
1429 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1431 return !after(end_seq, tcp_wnd_end(tp));
1434 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1435 * should be put on the wire right now. If so, it returns the number of
1436 * packets allowed by the congestion window.
1438 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1439 unsigned int cur_mss, int nonagle)
1441 struct tcp_sock *tp = tcp_sk(sk);
1442 unsigned int cwnd_quota;
1444 tcp_init_tso_segs(sk, skb, cur_mss);
1446 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1449 cwnd_quota = tcp_cwnd_test(tp, skb);
1450 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1456 /* Test if sending is allowed right now. */
1457 int tcp_may_send_now(struct sock *sk)
1459 struct tcp_sock *tp = tcp_sk(sk);
1460 struct sk_buff *skb = tcp_send_head(sk);
1463 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1464 (tcp_skb_is_last(sk, skb) ?
1465 tp->nonagle : TCP_NAGLE_PUSH));
1468 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1469 * which is put after SKB on the list. It is very much like
1470 * tcp_fragment() except that it may make several kinds of assumptions
1471 * in order to speed up the splitting operation. In particular, we
1472 * know that all the data is in scatter-gather pages, and that the
1473 * packet has never been sent out before (and thus is not cloned).
1475 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1476 unsigned int mss_now, gfp_t gfp)
1478 struct sk_buff *buff;
1479 int nlen = skb->len - len;
1482 /* All of a TSO frame must be composed of paged data. */
1483 if (skb->len != skb->data_len)
1484 return tcp_fragment(sk, skb, len, mss_now);
1486 buff = sk_stream_alloc_skb(sk, 0, gfp);
1487 if (unlikely(buff == NULL))
1490 sk->sk_wmem_queued += buff->truesize;
1491 sk_mem_charge(sk, buff->truesize);
1492 buff->truesize += nlen;
1493 skb->truesize -= nlen;
1495 /* Correct the sequence numbers. */
1496 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1497 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1498 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1500 /* PSH and FIN should only be set in the second packet. */
1501 flags = TCP_SKB_CB(skb)->tcp_flags;
1502 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1503 TCP_SKB_CB(buff)->tcp_flags = flags;
1505 /* This packet was never sent out yet, so no SACK bits. */
1506 TCP_SKB_CB(buff)->sacked = 0;
1508 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1509 skb_split(skb, buff, len);
1511 /* Fix up tso_factor for both original and new SKB. */
1512 tcp_set_skb_tso_segs(sk, skb, mss_now);
1513 tcp_set_skb_tso_segs(sk, buff, mss_now);
1515 /* Link BUFF into the send queue. */
1516 skb_header_release(buff);
1517 tcp_insert_write_queue_after(skb, buff, sk);
1522 /* Try to defer sending, if possible, in order to minimize the amount
1523 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1525 * This algorithm is from John Heffner.
1527 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1529 struct tcp_sock *tp = tcp_sk(sk);
1530 const struct inet_connection_sock *icsk = inet_csk(sk);
1531 u32 send_win, cong_win, limit, in_flight;
1534 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1537 if (icsk->icsk_ca_state != TCP_CA_Open)
1540 /* Defer for less than two clock ticks. */
1541 if (tp->tso_deferred &&
1542 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1545 in_flight = tcp_packets_in_flight(tp);
1547 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1549 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1551 /* From in_flight test above, we know that cwnd > in_flight. */
1552 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1554 limit = min(send_win, cong_win);
1556 /* If a full-sized TSO skb can be sent, do it. */
1557 if (limit >= sk->sk_gso_max_size)
1560 /* Middle in queue won't get any more data, full sendable already? */
1561 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1564 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1566 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1568 /* If at least some fraction of a window is available,
1571 chunk /= win_divisor;
1575 /* Different approach, try not to defer past a single
1576 * ACK. Receiver should ACK every other full sized
1577 * frame, so if we have space for more than 3 frames
1580 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1584 /* Ok, it looks like it is advisable to defer. */
1585 tp->tso_deferred = 1 | (jiffies << 1);
1590 tp->tso_deferred = 0;
1594 /* Create a new MTU probe if we are ready.
1595 * MTU probe is regularly attempting to increase the path MTU by
1596 * deliberately sending larger packets. This discovers routing
1597 * changes resulting in larger path MTUs.
1599 * Returns 0 if we should wait to probe (no cwnd available),
1600 * 1 if a probe was sent,
1603 static int tcp_mtu_probe(struct sock *sk)
1605 struct tcp_sock *tp = tcp_sk(sk);
1606 struct inet_connection_sock *icsk = inet_csk(sk);
1607 struct sk_buff *skb, *nskb, *next;
1614 /* Not currently probing/verifying,
1616 * have enough cwnd, and
1617 * not SACKing (the variable headers throw things off) */
1618 if (!icsk->icsk_mtup.enabled ||
1619 icsk->icsk_mtup.probe_size ||
1620 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1621 tp->snd_cwnd < 11 ||
1622 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1625 /* Very simple search strategy: just double the MSS. */
1626 mss_now = tcp_current_mss(sk);
1627 probe_size = 2 * tp->mss_cache;
1628 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1629 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1630 /* TODO: set timer for probe_converge_event */
1634 /* Have enough data in the send queue to probe? */
1635 if (tp->write_seq - tp->snd_nxt < size_needed)
1638 if (tp->snd_wnd < size_needed)
1640 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1643 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1644 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1645 if (!tcp_packets_in_flight(tp))
1651 /* We're allowed to probe. Build it now. */
1652 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1654 sk->sk_wmem_queued += nskb->truesize;
1655 sk_mem_charge(sk, nskb->truesize);
1657 skb = tcp_send_head(sk);
1659 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1660 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1661 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1662 TCP_SKB_CB(nskb)->sacked = 0;
1664 nskb->ip_summed = skb->ip_summed;
1666 tcp_insert_write_queue_before(nskb, skb, sk);
1669 tcp_for_write_queue_from_safe(skb, next, sk) {
1670 copy = min_t(int, skb->len, probe_size - len);
1671 if (nskb->ip_summed)
1672 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1674 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1675 skb_put(nskb, copy),
1678 if (skb->len <= copy) {
1679 /* We've eaten all the data from this skb.
1681 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1682 tcp_unlink_write_queue(skb, sk);
1683 sk_wmem_free_skb(sk, skb);
1685 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1686 ~(TCPHDR_FIN|TCPHDR_PSH);
1687 if (!skb_shinfo(skb)->nr_frags) {
1688 skb_pull(skb, copy);
1689 if (skb->ip_summed != CHECKSUM_PARTIAL)
1690 skb->csum = csum_partial(skb->data,
1693 __pskb_trim_head(skb, copy);
1694 tcp_set_skb_tso_segs(sk, skb, mss_now);
1696 TCP_SKB_CB(skb)->seq += copy;
1701 if (len >= probe_size)
1704 tcp_init_tso_segs(sk, nskb, nskb->len);
1706 /* We're ready to send. If this fails, the probe will
1707 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1708 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1709 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1710 /* Decrement cwnd here because we are sending
1711 * effectively two packets. */
1713 tcp_event_new_data_sent(sk, nskb);
1715 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1716 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1717 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1725 /* This routine writes packets to the network. It advances the
1726 * send_head. This happens as incoming acks open up the remote
1729 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1730 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1731 * account rare use of URG, this is not a big flaw.
1733 * Returns 1, if no segments are in flight and we have queued segments, but
1734 * cannot send anything now because of SWS or another problem.
1736 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1737 int push_one, gfp_t gfp)
1739 struct tcp_sock *tp = tcp_sk(sk);
1740 struct sk_buff *skb;
1741 unsigned int tso_segs, sent_pkts;
1748 /* Do MTU probing. */
1749 result = tcp_mtu_probe(sk);
1752 } else if (result > 0) {
1757 while ((skb = tcp_send_head(sk))) {
1760 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1763 cwnd_quota = tcp_cwnd_test(tp, skb);
1767 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1770 if (tso_segs == 1) {
1771 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1772 (tcp_skb_is_last(sk, skb) ?
1773 nonagle : TCP_NAGLE_PUSH))))
1776 if (!push_one && tcp_tso_should_defer(sk, skb))
1781 if (tso_segs > 1 && !tcp_urg_mode(tp))
1782 limit = tcp_mss_split_point(sk, skb, mss_now,
1785 if (skb->len > limit &&
1786 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1789 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1791 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1794 /* Advance the send_head. This one is sent out.
1795 * This call will increment packets_out.
1797 tcp_event_new_data_sent(sk, skb);
1799 tcp_minshall_update(tp, mss_now, skb);
1800 sent_pkts += tcp_skb_pcount(skb);
1805 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
1806 tp->prr_out += sent_pkts;
1808 if (likely(sent_pkts)) {
1809 tcp_cwnd_validate(sk);
1812 return !tp->packets_out && tcp_send_head(sk);
1815 /* Push out any pending frames which were held back due to
1816 * TCP_CORK or attempt at coalescing tiny packets.
1817 * The socket must be locked by the caller.
1819 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1822 /* If we are closed, the bytes will have to remain here.
1823 * In time closedown will finish, we empty the write queue and
1824 * all will be happy.
1826 if (unlikely(sk->sk_state == TCP_CLOSE))
1829 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1830 tcp_check_probe_timer(sk);
1833 /* Send _single_ skb sitting at the send head. This function requires
1834 * true push pending frames to setup probe timer etc.
1836 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1838 struct sk_buff *skb = tcp_send_head(sk);
1840 BUG_ON(!skb || skb->len < mss_now);
1842 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1845 /* This function returns the amount that we can raise the
1846 * usable window based on the following constraints
1848 * 1. The window can never be shrunk once it is offered (RFC 793)
1849 * 2. We limit memory per socket
1852 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1853 * RECV.NEXT + RCV.WIN fixed until:
1854 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1856 * i.e. don't raise the right edge of the window until you can raise
1857 * it at least MSS bytes.
1859 * Unfortunately, the recommended algorithm breaks header prediction,
1860 * since header prediction assumes th->window stays fixed.
1862 * Strictly speaking, keeping th->window fixed violates the receiver
1863 * side SWS prevention criteria. The problem is that under this rule
1864 * a stream of single byte packets will cause the right side of the
1865 * window to always advance by a single byte.
1867 * Of course, if the sender implements sender side SWS prevention
1868 * then this will not be a problem.
1870 * BSD seems to make the following compromise:
1872 * If the free space is less than the 1/4 of the maximum
1873 * space available and the free space is less than 1/2 mss,
1874 * then set the window to 0.
1875 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1876 * Otherwise, just prevent the window from shrinking
1877 * and from being larger than the largest representable value.
1879 * This prevents incremental opening of the window in the regime
1880 * where TCP is limited by the speed of the reader side taking
1881 * data out of the TCP receive queue. It does nothing about
1882 * those cases where the window is constrained on the sender side
1883 * because the pipeline is full.
1885 * BSD also seems to "accidentally" limit itself to windows that are a
1886 * multiple of MSS, at least until the free space gets quite small.
1887 * This would appear to be a side effect of the mbuf implementation.
1888 * Combining these two algorithms results in the observed behavior
1889 * of having a fixed window size at almost all times.
1891 * Below we obtain similar behavior by forcing the offered window to
1892 * a multiple of the mss when it is feasible to do so.
1894 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1895 * Regular options like TIMESTAMP are taken into account.
1897 u32 __tcp_select_window(struct sock *sk)
1899 struct inet_connection_sock *icsk = inet_csk(sk);
1900 struct tcp_sock *tp = tcp_sk(sk);
1901 /* MSS for the peer's data. Previous versions used mss_clamp
1902 * here. I don't know if the value based on our guesses
1903 * of peer's MSS is better for the performance. It's more correct
1904 * but may be worse for the performance because of rcv_mss
1905 * fluctuations. --SAW 1998/11/1
1907 int mss = icsk->icsk_ack.rcv_mss;
1908 int free_space = tcp_space(sk);
1909 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1912 if (mss > full_space)
1915 if (free_space < (full_space >> 1)) {
1916 icsk->icsk_ack.quick = 0;
1918 if (tcp_memory_pressure)
1919 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1922 if (free_space < mss)
1926 if (free_space > tp->rcv_ssthresh)
1927 free_space = tp->rcv_ssthresh;
1929 /* Don't do rounding if we are using window scaling, since the
1930 * scaled window will not line up with the MSS boundary anyway.
1932 window = tp->rcv_wnd;
1933 if (tp->rx_opt.rcv_wscale) {
1934 window = free_space;
1936 /* Advertise enough space so that it won't get scaled away.
1937 * Import case: prevent zero window announcement if
1938 * 1<<rcv_wscale > mss.
1940 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1941 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1942 << tp->rx_opt.rcv_wscale);
1944 /* Get the largest window that is a nice multiple of mss.
1945 * Window clamp already applied above.
1946 * If our current window offering is within 1 mss of the
1947 * free space we just keep it. This prevents the divide
1948 * and multiply from happening most of the time.
1949 * We also don't do any window rounding when the free space
1952 if (window <= free_space - mss || window > free_space)
1953 window = (free_space / mss) * mss;
1954 else if (mss == full_space &&
1955 free_space > window + (full_space >> 1))
1956 window = free_space;
1962 /* Collapses two adjacent SKB's during retransmission. */
1963 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1965 struct tcp_sock *tp = tcp_sk(sk);
1966 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1967 int skb_size, next_skb_size;
1969 skb_size = skb->len;
1970 next_skb_size = next_skb->len;
1972 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1974 tcp_highest_sack_combine(sk, next_skb, skb);
1976 tcp_unlink_write_queue(next_skb, sk);
1978 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1981 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1982 skb->ip_summed = CHECKSUM_PARTIAL;
1984 if (skb->ip_summed != CHECKSUM_PARTIAL)
1985 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1987 /* Update sequence range on original skb. */
1988 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1990 /* Merge over control information. This moves PSH/FIN etc. over */
1991 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
1993 /* All done, get rid of second SKB and account for it so
1994 * packet counting does not break.
1996 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1998 /* changed transmit queue under us so clear hints */
1999 tcp_clear_retrans_hints_partial(tp);
2000 if (next_skb == tp->retransmit_skb_hint)
2001 tp->retransmit_skb_hint = skb;
2003 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2005 sk_wmem_free_skb(sk, next_skb);
2008 /* Check if coalescing SKBs is legal. */
2009 static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2011 if (tcp_skb_pcount(skb) > 1)
2013 /* TODO: SACK collapsing could be used to remove this condition */
2014 if (skb_shinfo(skb)->nr_frags != 0)
2016 if (skb_cloned(skb))
2018 if (skb == tcp_send_head(sk))
2020 /* Some heurestics for collapsing over SACK'd could be invented */
2021 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2027 /* Collapse packets in the retransmit queue to make to create
2028 * less packets on the wire. This is only done on retransmission.
2030 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2033 struct tcp_sock *tp = tcp_sk(sk);
2034 struct sk_buff *skb = to, *tmp;
2037 if (!sysctl_tcp_retrans_collapse)
2039 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2042 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2043 if (!tcp_can_collapse(sk, skb))
2055 /* Punt if not enough space exists in the first SKB for
2056 * the data in the second
2058 if (skb->len > skb_tailroom(to))
2061 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2064 tcp_collapse_retrans(sk, to);
2068 /* This retransmits one SKB. Policy decisions and retransmit queue
2069 * state updates are done by the caller. Returns non-zero if an
2070 * error occurred which prevented the send.
2072 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2074 struct tcp_sock *tp = tcp_sk(sk);
2075 struct inet_connection_sock *icsk = inet_csk(sk);
2076 unsigned int cur_mss;
2079 /* Inconslusive MTU probe */
2080 if (icsk->icsk_mtup.probe_size) {
2081 icsk->icsk_mtup.probe_size = 0;
2084 /* Do not sent more than we queued. 1/4 is reserved for possible
2085 * copying overhead: fragmentation, tunneling, mangling etc.
2087 if (atomic_read(&sk->sk_wmem_alloc) >
2088 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2091 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2092 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2094 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2098 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2099 return -EHOSTUNREACH; /* Routing failure or similar. */
2101 cur_mss = tcp_current_mss(sk);
2103 /* If receiver has shrunk his window, and skb is out of
2104 * new window, do not retransmit it. The exception is the
2105 * case, when window is shrunk to zero. In this case
2106 * our retransmit serves as a zero window probe.
2108 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2109 TCP_SKB_CB(skb)->seq != tp->snd_una)
2112 if (skb->len > cur_mss) {
2113 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2114 return -ENOMEM; /* We'll try again later. */
2116 int oldpcount = tcp_skb_pcount(skb);
2118 if (unlikely(oldpcount > 1)) {
2119 tcp_init_tso_segs(sk, skb, cur_mss);
2120 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2124 tcp_retrans_try_collapse(sk, skb, cur_mss);
2126 /* Some Solaris stacks overoptimize and ignore the FIN on a
2127 * retransmit when old data is attached. So strip it off
2128 * since it is cheap to do so and saves bytes on the network.
2131 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2132 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2133 if (!pskb_trim(skb, 0)) {
2134 /* Reuse, even though it does some unnecessary work */
2135 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2136 TCP_SKB_CB(skb)->tcp_flags);
2137 skb->ip_summed = CHECKSUM_NONE;
2141 /* Make a copy, if the first transmission SKB clone we made
2142 * is still in somebody's hands, else make a clone.
2144 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2146 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2149 /* Update global TCP statistics. */
2150 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2152 tp->total_retrans++;
2154 #if FASTRETRANS_DEBUG > 0
2155 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2156 if (net_ratelimit())
2157 printk(KERN_DEBUG "retrans_out leaked.\n");
2160 if (!tp->retrans_out)
2161 tp->lost_retrans_low = tp->snd_nxt;
2162 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2163 tp->retrans_out += tcp_skb_pcount(skb);
2165 /* Save stamp of the first retransmit. */
2166 if (!tp->retrans_stamp)
2167 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2169 tp->undo_retrans += tcp_skb_pcount(skb);
2171 /* snd_nxt is stored to detect loss of retransmitted segment,
2172 * see tcp_input.c tcp_sacktag_write_queue().
2174 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2179 /* Check if we forward retransmits are possible in the current
2180 * window/congestion state.
2182 static int tcp_can_forward_retransmit(struct sock *sk)
2184 const struct inet_connection_sock *icsk = inet_csk(sk);
2185 struct tcp_sock *tp = tcp_sk(sk);
2187 /* Forward retransmissions are possible only during Recovery. */
2188 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2191 /* No forward retransmissions in Reno are possible. */
2192 if (tcp_is_reno(tp))
2195 /* Yeah, we have to make difficult choice between forward transmission
2196 * and retransmission... Both ways have their merits...
2198 * For now we do not retransmit anything, while we have some new
2199 * segments to send. In the other cases, follow rule 3 for
2200 * NextSeg() specified in RFC3517.
2203 if (tcp_may_send_now(sk))
2209 /* This gets called after a retransmit timeout, and the initially
2210 * retransmitted data is acknowledged. It tries to continue
2211 * resending the rest of the retransmit queue, until either
2212 * we've sent it all or the congestion window limit is reached.
2213 * If doing SACK, the first ACK which comes back for a timeout
2214 * based retransmit packet might feed us FACK information again.
2215 * If so, we use it to avoid unnecessarily retransmissions.
2217 void tcp_xmit_retransmit_queue(struct sock *sk)
2219 const struct inet_connection_sock *icsk = inet_csk(sk);
2220 struct tcp_sock *tp = tcp_sk(sk);
2221 struct sk_buff *skb;
2222 struct sk_buff *hole = NULL;
2225 int fwd_rexmitting = 0;
2227 if (!tp->packets_out)
2231 tp->retransmit_high = tp->snd_una;
2233 if (tp->retransmit_skb_hint) {
2234 skb = tp->retransmit_skb_hint;
2235 last_lost = TCP_SKB_CB(skb)->end_seq;
2236 if (after(last_lost, tp->retransmit_high))
2237 last_lost = tp->retransmit_high;
2239 skb = tcp_write_queue_head(sk);
2240 last_lost = tp->snd_una;
2243 tcp_for_write_queue_from(skb, sk) {
2244 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2246 if (skb == tcp_send_head(sk))
2248 /* we could do better than to assign each time */
2250 tp->retransmit_skb_hint = skb;
2252 /* Assume this retransmit will generate
2253 * only one packet for congestion window
2254 * calculation purposes. This works because
2255 * tcp_retransmit_skb() will chop up the
2256 * packet to be MSS sized and all the
2257 * packet counting works out.
2259 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2262 if (fwd_rexmitting) {
2264 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2266 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2268 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2269 tp->retransmit_high = last_lost;
2270 if (!tcp_can_forward_retransmit(sk))
2272 /* Backtrack if necessary to non-L'ed skb */
2280 } else if (!(sacked & TCPCB_LOST)) {
2281 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2286 last_lost = TCP_SKB_CB(skb)->end_seq;
2287 if (icsk->icsk_ca_state != TCP_CA_Loss)
2288 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2290 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2293 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2296 if (tcp_retransmit_skb(sk, skb))
2298 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2300 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2301 tp->prr_out += tcp_skb_pcount(skb);
2303 if (skb == tcp_write_queue_head(sk))
2304 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2305 inet_csk(sk)->icsk_rto,
2310 /* Send a fin. The caller locks the socket for us. This cannot be
2311 * allowed to fail queueing a FIN frame under any circumstances.
2313 void tcp_send_fin(struct sock *sk)
2315 struct tcp_sock *tp = tcp_sk(sk);
2316 struct sk_buff *skb = tcp_write_queue_tail(sk);
2319 /* Optimization, tack on the FIN if we have a queue of
2320 * unsent frames. But be careful about outgoing SACKS
2323 mss_now = tcp_current_mss(sk);
2325 if (tcp_send_head(sk) != NULL) {
2326 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2327 TCP_SKB_CB(skb)->end_seq++;
2330 /* Socket is locked, keep trying until memory is available. */
2332 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2339 /* Reserve space for headers and prepare control bits. */
2340 skb_reserve(skb, MAX_TCP_HEADER);
2341 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2342 tcp_init_nondata_skb(skb, tp->write_seq,
2343 TCPHDR_ACK | TCPHDR_FIN);
2344 tcp_queue_skb(sk, skb);
2346 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2349 /* We get here when a process closes a file descriptor (either due to
2350 * an explicit close() or as a byproduct of exit()'ing) and there
2351 * was unread data in the receive queue. This behavior is recommended
2352 * by RFC 2525, section 2.17. -DaveM
2354 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2356 struct sk_buff *skb;
2358 /* NOTE: No TCP options attached and we never retransmit this. */
2359 skb = alloc_skb(MAX_TCP_HEADER, priority);
2361 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2365 /* Reserve space for headers and prepare control bits. */
2366 skb_reserve(skb, MAX_TCP_HEADER);
2367 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2368 TCPHDR_ACK | TCPHDR_RST);
2370 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2371 if (tcp_transmit_skb(sk, skb, 0, priority))
2372 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2374 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2377 /* Send a crossed SYN-ACK during socket establishment.
2378 * WARNING: This routine must only be called when we have already sent
2379 * a SYN packet that crossed the incoming SYN that caused this routine
2380 * to get called. If this assumption fails then the initial rcv_wnd
2381 * and rcv_wscale values will not be correct.
2383 int tcp_send_synack(struct sock *sk)
2385 struct sk_buff *skb;
2387 skb = tcp_write_queue_head(sk);
2388 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2389 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2392 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2393 if (skb_cloned(skb)) {
2394 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2397 tcp_unlink_write_queue(skb, sk);
2398 skb_header_release(nskb);
2399 __tcp_add_write_queue_head(sk, nskb);
2400 sk_wmem_free_skb(sk, skb);
2401 sk->sk_wmem_queued += nskb->truesize;
2402 sk_mem_charge(sk, nskb->truesize);
2406 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2407 TCP_ECN_send_synack(tcp_sk(sk), skb);
2409 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2410 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2413 /* Prepare a SYN-ACK. */
2414 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2415 struct request_sock *req,
2416 struct request_values *rvp)
2418 struct tcp_out_options opts;
2419 struct tcp_extend_values *xvp = tcp_xv(rvp);
2420 struct inet_request_sock *ireq = inet_rsk(req);
2421 struct tcp_sock *tp = tcp_sk(sk);
2422 const struct tcp_cookie_values *cvp = tp->cookie_values;
2424 struct sk_buff *skb;
2425 struct tcp_md5sig_key *md5;
2426 int tcp_header_size;
2428 int s_data_desired = 0;
2430 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2431 s_data_desired = cvp->s_data_desired;
2432 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2436 /* Reserve space for headers. */
2437 skb_reserve(skb, MAX_TCP_HEADER);
2439 skb_dst_set(skb, dst_clone(dst));
2441 mss = dst_metric_advmss(dst);
2442 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2443 mss = tp->rx_opt.user_mss;
2445 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2447 /* Set this up on the first call only */
2448 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2450 /* limit the window selection if the user enforce a smaller rx buffer */
2451 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2452 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2453 req->window_clamp = tcp_full_space(sk);
2455 /* tcp_full_space because it is guaranteed to be the first packet */
2456 tcp_select_initial_window(tcp_full_space(sk),
2457 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2462 dst_metric(dst, RTAX_INITRWND));
2463 ireq->rcv_wscale = rcv_wscale;
2466 memset(&opts, 0, sizeof(opts));
2467 #ifdef CONFIG_SYN_COOKIES
2468 if (unlikely(req->cookie_ts))
2469 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2472 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2473 tcp_header_size = tcp_synack_options(sk, req, mss,
2474 skb, &opts, &md5, xvp)
2477 skb_push(skb, tcp_header_size);
2478 skb_reset_transport_header(skb);
2481 memset(th, 0, sizeof(struct tcphdr));
2484 TCP_ECN_make_synack(req, th);
2485 th->source = ireq->loc_port;
2486 th->dest = ireq->rmt_port;
2487 /* Setting of flags are superfluous here for callers (and ECE is
2488 * not even correctly set)
2490 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2491 TCPHDR_SYN | TCPHDR_ACK);
2493 if (OPTION_COOKIE_EXTENSION & opts.options) {
2494 if (s_data_desired) {
2495 u8 *buf = skb_put(skb, s_data_desired);
2497 /* copy data directly from the listening socket. */
2498 memcpy(buf, cvp->s_data_payload, s_data_desired);
2499 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2502 if (opts.hash_size > 0) {
2503 __u32 workspace[SHA_WORKSPACE_WORDS];
2504 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2505 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2507 /* Secret recipe depends on the Timestamp, (future)
2508 * Sequence and Acknowledgment Numbers, Initiator
2509 * Cookie, and others handled by IP variant caller.
2511 *tail-- ^= opts.tsval;
2512 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2513 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2516 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2517 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2519 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2522 opts.hash_location =
2523 (__u8 *)&xvp->cookie_bakery[0];
2527 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2528 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2530 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2531 th->window = htons(min(req->rcv_wnd, 65535U));
2532 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2533 th->doff = (tcp_header_size >> 2);
2534 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2536 #ifdef CONFIG_TCP_MD5SIG
2537 /* Okay, we have all we need - do the md5 hash if needed */
2539 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2540 md5, NULL, req, skb);
2546 EXPORT_SYMBOL(tcp_make_synack);
2548 /* Do all connect socket setups that can be done AF independent. */
2549 static void tcp_connect_init(struct sock *sk)
2551 struct dst_entry *dst = __sk_dst_get(sk);
2552 struct tcp_sock *tp = tcp_sk(sk);
2555 /* We'll fix this up when we get a response from the other end.
2556 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2558 tp->tcp_header_len = sizeof(struct tcphdr) +
2559 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2561 #ifdef CONFIG_TCP_MD5SIG
2562 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2563 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2566 /* If user gave his TCP_MAXSEG, record it to clamp */
2567 if (tp->rx_opt.user_mss)
2568 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2571 tcp_sync_mss(sk, dst_mtu(dst));
2573 if (!tp->window_clamp)
2574 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2575 tp->advmss = dst_metric_advmss(dst);
2576 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2577 tp->advmss = tp->rx_opt.user_mss;
2579 tcp_initialize_rcv_mss(sk);
2581 /* limit the window selection if the user enforce a smaller rx buffer */
2582 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2583 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2584 tp->window_clamp = tcp_full_space(sk);
2586 tcp_select_initial_window(tcp_full_space(sk),
2587 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2590 sysctl_tcp_window_scaling,
2592 dst_metric(dst, RTAX_INITRWND));
2594 tp->rx_opt.rcv_wscale = rcv_wscale;
2595 tp->rcv_ssthresh = tp->rcv_wnd;
2598 sock_reset_flag(sk, SOCK_DONE);
2601 tp->snd_una = tp->write_seq;
2602 tp->snd_sml = tp->write_seq;
2603 tp->snd_up = tp->write_seq;
2608 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2609 inet_csk(sk)->icsk_retransmits = 0;
2610 tcp_clear_retrans(tp);
2613 /* Build a SYN and send it off. */
2614 int tcp_connect(struct sock *sk)
2616 struct tcp_sock *tp = tcp_sk(sk);
2617 struct sk_buff *buff;
2620 tcp_connect_init(sk);
2622 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2623 if (unlikely(buff == NULL))
2626 /* Reserve space for headers. */
2627 skb_reserve(buff, MAX_TCP_HEADER);
2629 tp->snd_nxt = tp->write_seq;
2630 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2631 TCP_ECN_send_syn(sk, buff);
2634 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2635 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2636 skb_header_release(buff);
2637 __tcp_add_write_queue_tail(sk, buff);
2638 sk->sk_wmem_queued += buff->truesize;
2639 sk_mem_charge(sk, buff->truesize);
2640 tp->packets_out += tcp_skb_pcount(buff);
2641 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2642 if (err == -ECONNREFUSED)
2645 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2646 * in order to make this packet get counted in tcpOutSegs.
2648 tp->snd_nxt = tp->write_seq;
2649 tp->pushed_seq = tp->write_seq;
2650 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2652 /* Timer for repeating the SYN until an answer. */
2653 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2654 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2657 EXPORT_SYMBOL(tcp_connect);
2659 /* Send out a delayed ack, the caller does the policy checking
2660 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2663 void tcp_send_delayed_ack(struct sock *sk)
2665 struct inet_connection_sock *icsk = inet_csk(sk);
2666 int ato = icsk->icsk_ack.ato;
2667 unsigned long timeout;
2669 if (ato > TCP_DELACK_MIN) {
2670 const struct tcp_sock *tp = tcp_sk(sk);
2671 int max_ato = HZ / 2;
2673 if (icsk->icsk_ack.pingpong ||
2674 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2675 max_ato = TCP_DELACK_MAX;
2677 /* Slow path, intersegment interval is "high". */
2679 /* If some rtt estimate is known, use it to bound delayed ack.
2680 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2684 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2690 ato = min(ato, max_ato);
2693 /* Stay within the limit we were given */
2694 timeout = jiffies + ato;
2696 /* Use new timeout only if there wasn't a older one earlier. */
2697 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2698 /* If delack timer was blocked or is about to expire,
2701 if (icsk->icsk_ack.blocked ||
2702 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2707 if (!time_before(timeout, icsk->icsk_ack.timeout))
2708 timeout = icsk->icsk_ack.timeout;
2710 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2711 icsk->icsk_ack.timeout = timeout;
2712 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2715 /* This routine sends an ack and also updates the window. */
2716 void tcp_send_ack(struct sock *sk)
2718 struct sk_buff *buff;
2720 /* If we have been reset, we may not send again. */
2721 if (sk->sk_state == TCP_CLOSE)
2724 /* We are not putting this on the write queue, so
2725 * tcp_transmit_skb() will set the ownership to this
2728 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2730 inet_csk_schedule_ack(sk);
2731 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2732 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2733 TCP_DELACK_MAX, TCP_RTO_MAX);
2737 /* Reserve space for headers and prepare control bits. */
2738 skb_reserve(buff, MAX_TCP_HEADER);
2739 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2741 /* Send it off, this clears delayed acks for us. */
2742 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2743 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2746 /* This routine sends a packet with an out of date sequence
2747 * number. It assumes the other end will try to ack it.
2749 * Question: what should we make while urgent mode?
2750 * 4.4BSD forces sending single byte of data. We cannot send
2751 * out of window data, because we have SND.NXT==SND.MAX...
2753 * Current solution: to send TWO zero-length segments in urgent mode:
2754 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2755 * out-of-date with SND.UNA-1 to probe window.
2757 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2759 struct tcp_sock *tp = tcp_sk(sk);
2760 struct sk_buff *skb;
2762 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2763 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2767 /* Reserve space for headers and set control bits. */
2768 skb_reserve(skb, MAX_TCP_HEADER);
2769 /* Use a previous sequence. This should cause the other
2770 * end to send an ack. Don't queue or clone SKB, just
2773 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2774 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2775 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2778 /* Initiate keepalive or window probe from timer. */
2779 int tcp_write_wakeup(struct sock *sk)
2781 struct tcp_sock *tp = tcp_sk(sk);
2782 struct sk_buff *skb;
2784 if (sk->sk_state == TCP_CLOSE)
2787 if ((skb = tcp_send_head(sk)) != NULL &&
2788 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2790 unsigned int mss = tcp_current_mss(sk);
2791 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2793 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2794 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2796 /* We are probing the opening of a window
2797 * but the window size is != 0
2798 * must have been a result SWS avoidance ( sender )
2800 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2802 seg_size = min(seg_size, mss);
2803 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2804 if (tcp_fragment(sk, skb, seg_size, mss))
2806 } else if (!tcp_skb_pcount(skb))
2807 tcp_set_skb_tso_segs(sk, skb, mss);
2809 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2810 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2811 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2813 tcp_event_new_data_sent(sk, skb);
2816 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2817 tcp_xmit_probe_skb(sk, 1);
2818 return tcp_xmit_probe_skb(sk, 0);
2822 /* A window probe timeout has occurred. If window is not closed send
2823 * a partial packet else a zero probe.
2825 void tcp_send_probe0(struct sock *sk)
2827 struct inet_connection_sock *icsk = inet_csk(sk);
2828 struct tcp_sock *tp = tcp_sk(sk);
2831 err = tcp_write_wakeup(sk);
2833 if (tp->packets_out || !tcp_send_head(sk)) {
2834 /* Cancel probe timer, if it is not required. */
2835 icsk->icsk_probes_out = 0;
2836 icsk->icsk_backoff = 0;
2841 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2842 icsk->icsk_backoff++;
2843 icsk->icsk_probes_out++;
2844 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2845 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2848 /* If packet was not sent due to local congestion,
2849 * do not backoff and do not remember icsk_probes_out.
2850 * Let local senders to fight for local resources.
2852 * Use accumulated backoff yet.
2854 if (!icsk->icsk_probes_out)
2855 icsk->icsk_probes_out = 1;
2856 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2857 min(icsk->icsk_rto << icsk->icsk_backoff,
2858 TCP_RESOURCE_PROBE_INTERVAL),