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, const 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(const struct sock *sk)
94 const 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 const 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, const 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,
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, space, sysctl_tcp_rmem[2]);
224 space = max_t(u32, space, sysctl_rmem_max);
225 space = min_t(u32, space, *window_clamp);
226 while (space > 65535 && (*rcv_wscale) < 14) {
232 /* Set initial window to a value enough for senders starting with
233 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
234 * a limit on the initial window when mss is larger than 1460.
236 if (mss > (1 << *rcv_wscale)) {
237 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
240 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
241 /* when initializing use the value from init_rcv_wnd
242 * rather than the default from above
245 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
247 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
253 EXPORT_SYMBOL(tcp_select_initial_window);
255 /* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
260 static u16 tcp_select_window(struct sock *sk)
262 struct tcp_sock *tp = tcp_sk(sk);
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
291 /* If we advertise zero window, disable fast path. */
298 /* Packet ECN state for a SYN-ACK */
299 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
301 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
302 if (!(tp->ecn_flags & TCP_ECN_OK))
303 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
306 /* Packet ECN state for a SYN. */
307 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
309 struct tcp_sock *tp = tcp_sk(sk);
312 if (sysctl_tcp_ecn == 1) {
313 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
314 tp->ecn_flags = TCP_ECN_OK;
318 static __inline__ void
319 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
321 if (inet_rsk(req)->ecn_ok)
325 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
328 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
331 struct tcp_sock *tp = tcp_sk(sk);
333 if (tp->ecn_flags & TCP_ECN_OK) {
334 /* Not-retransmitted data segment: set ECT and inject CWR. */
335 if (skb->len != tcp_header_len &&
336 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
338 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
339 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
340 tcp_hdr(skb)->cwr = 1;
341 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
344 /* ACK or retransmitted segment: clear ECT|CE */
345 INET_ECN_dontxmit(sk);
347 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
348 tcp_hdr(skb)->ece = 1;
352 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
353 * auto increment end seqno.
355 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
357 skb->ip_summed = CHECKSUM_PARTIAL;
360 TCP_SKB_CB(skb)->tcp_flags = flags;
361 TCP_SKB_CB(skb)->sacked = 0;
363 skb_shinfo(skb)->gso_segs = 1;
364 skb_shinfo(skb)->gso_size = 0;
365 skb_shinfo(skb)->gso_type = 0;
367 TCP_SKB_CB(skb)->seq = seq;
368 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
370 TCP_SKB_CB(skb)->end_seq = seq;
373 static inline int tcp_urg_mode(const struct tcp_sock *tp)
375 return tp->snd_una != tp->snd_up;
378 #define OPTION_SACK_ADVERTISE (1 << 0)
379 #define OPTION_TS (1 << 1)
380 #define OPTION_MD5 (1 << 2)
381 #define OPTION_WSCALE (1 << 3)
382 #define OPTION_COOKIE_EXTENSION (1 << 4)
384 struct tcp_out_options {
385 u8 options; /* bit field of OPTION_* */
386 u8 ws; /* window scale, 0 to disable */
387 u8 num_sack_blocks; /* number of SACK blocks to include */
388 u8 hash_size; /* bytes in hash_location */
389 u16 mss; /* 0 to disable */
390 __u32 tsval, tsecr; /* need to include OPTION_TS */
391 __u8 *hash_location; /* temporary pointer, overloaded */
394 /* The sysctl int routines are generic, so check consistency here.
396 static u8 tcp_cookie_size_check(u8 desired)
401 /* previously specified */
404 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
405 if (cookie_size <= 0)
406 /* no default specified */
409 if (cookie_size <= TCP_COOKIE_MIN)
410 /* value too small, specify minimum */
411 return TCP_COOKIE_MIN;
413 if (cookie_size >= TCP_COOKIE_MAX)
414 /* value too large, specify maximum */
415 return TCP_COOKIE_MAX;
418 /* 8-bit multiple, illegal, fix it */
421 return (u8)cookie_size;
424 /* Write previously computed TCP options to the packet.
426 * Beware: Something in the Internet is very sensitive to the ordering of
427 * TCP options, we learned this through the hard way, so be careful here.
428 * Luckily we can at least blame others for their non-compliance but from
429 * inter-operatibility perspective it seems that we're somewhat stuck with
430 * the ordering which we have been using if we want to keep working with
431 * those broken things (not that it currently hurts anybody as there isn't
432 * particular reason why the ordering would need to be changed).
434 * At least SACK_PERM as the first option is known to lead to a disaster
435 * (but it may well be that other scenarios fail similarly).
437 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
438 struct tcp_out_options *opts)
440 u8 options = opts->options; /* mungable copy */
442 /* Having both authentication and cookies for security is redundant,
443 * and there's certainly not enough room. Instead, the cookie-less
444 * extension variant is proposed.
446 * Consider the pessimal case with authentication. The options
448 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
450 if (unlikely(OPTION_MD5 & options)) {
451 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
452 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
453 (TCPOLEN_COOKIE_BASE << 16) |
454 (TCPOPT_MD5SIG << 8) |
457 *ptr++ = htonl((TCPOPT_NOP << 24) |
459 (TCPOPT_MD5SIG << 8) |
462 options &= ~OPTION_COOKIE_EXTENSION;
463 /* overload cookie hash location */
464 opts->hash_location = (__u8 *)ptr;
468 if (unlikely(opts->mss)) {
469 *ptr++ = htonl((TCPOPT_MSS << 24) |
470 (TCPOLEN_MSS << 16) |
474 if (likely(OPTION_TS & options)) {
475 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
476 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
477 (TCPOLEN_SACK_PERM << 16) |
478 (TCPOPT_TIMESTAMP << 8) |
480 options &= ~OPTION_SACK_ADVERTISE;
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
484 (TCPOPT_TIMESTAMP << 8) |
487 *ptr++ = htonl(opts->tsval);
488 *ptr++ = htonl(opts->tsecr);
491 /* Specification requires after timestamp, so do it now.
493 * Consider the pessimal case without authentication. The options
495 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
497 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
498 __u8 *cookie_copy = opts->hash_location;
499 u8 cookie_size = opts->hash_size;
501 /* 8-bit multiple handled in tcp_cookie_size_check() above,
504 if (0x2 & cookie_size) {
505 __u8 *p = (__u8 *)ptr;
507 /* 16-bit multiple */
508 *p++ = TCPOPT_COOKIE;
509 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
510 *p++ = *cookie_copy++;
511 *p++ = *cookie_copy++;
515 /* 32-bit multiple */
516 *ptr++ = htonl(((TCPOPT_NOP << 24) |
518 (TCPOPT_COOKIE << 8) |
519 TCPOLEN_COOKIE_BASE) +
523 if (cookie_size > 0) {
524 memcpy(ptr, cookie_copy, cookie_size);
525 ptr += (cookie_size / 4);
529 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
530 *ptr++ = htonl((TCPOPT_NOP << 24) |
532 (TCPOPT_SACK_PERM << 8) |
536 if (unlikely(OPTION_WSCALE & options)) {
537 *ptr++ = htonl((TCPOPT_NOP << 24) |
538 (TCPOPT_WINDOW << 16) |
539 (TCPOLEN_WINDOW << 8) |
543 if (unlikely(opts->num_sack_blocks)) {
544 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
545 tp->duplicate_sack : tp->selective_acks;
548 *ptr++ = htonl((TCPOPT_NOP << 24) |
551 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
552 TCPOLEN_SACK_PERBLOCK)));
554 for (this_sack = 0; this_sack < opts->num_sack_blocks;
556 *ptr++ = htonl(sp[this_sack].start_seq);
557 *ptr++ = htonl(sp[this_sack].end_seq);
560 tp->rx_opt.dsack = 0;
564 /* Compute TCP options for SYN packets. This is not the final
565 * network wire format yet.
567 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
568 struct tcp_out_options *opts,
569 struct tcp_md5sig_key **md5)
571 struct tcp_sock *tp = tcp_sk(sk);
572 struct tcp_cookie_values *cvp = tp->cookie_values;
573 unsigned remaining = MAX_TCP_OPTION_SPACE;
574 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
575 tcp_cookie_size_check(cvp->cookie_desired) :
578 #ifdef CONFIG_TCP_MD5SIG
579 *md5 = tp->af_specific->md5_lookup(sk, sk);
581 opts->options |= OPTION_MD5;
582 remaining -= TCPOLEN_MD5SIG_ALIGNED;
588 /* We always get an MSS option. The option bytes which will be seen in
589 * normal data packets should timestamps be used, must be in the MSS
590 * advertised. But we subtract them from tp->mss_cache so that
591 * calculations in tcp_sendmsg are simpler etc. So account for this
592 * fact here if necessary. If we don't do this correctly, as a
593 * receiver we won't recognize data packets as being full sized when we
594 * should, and thus we won't abide by the delayed ACK rules correctly.
595 * SACKs don't matter, we never delay an ACK when we have any of those
597 opts->mss = tcp_advertise_mss(sk);
598 remaining -= TCPOLEN_MSS_ALIGNED;
600 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
601 opts->options |= OPTION_TS;
602 opts->tsval = TCP_SKB_CB(skb)->when;
603 opts->tsecr = tp->rx_opt.ts_recent;
604 remaining -= TCPOLEN_TSTAMP_ALIGNED;
606 if (likely(sysctl_tcp_window_scaling)) {
607 opts->ws = tp->rx_opt.rcv_wscale;
608 opts->options |= OPTION_WSCALE;
609 remaining -= TCPOLEN_WSCALE_ALIGNED;
611 if (likely(sysctl_tcp_sack)) {
612 opts->options |= OPTION_SACK_ADVERTISE;
613 if (unlikely(!(OPTION_TS & opts->options)))
614 remaining -= TCPOLEN_SACKPERM_ALIGNED;
617 /* Note that timestamps are required by the specification.
619 * Odd numbers of bytes are prohibited by the specification, ensuring
620 * that the cookie is 16-bit aligned, and the resulting cookie pair is
624 (OPTION_TS & opts->options) &&
626 int need = TCPOLEN_COOKIE_BASE + cookie_size;
629 /* 32-bit multiple */
630 need += 2; /* NOPs */
632 if (need > remaining) {
633 /* try shrinking cookie to fit */
638 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
642 if (TCP_COOKIE_MIN <= cookie_size) {
643 opts->options |= OPTION_COOKIE_EXTENSION;
644 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
645 opts->hash_size = cookie_size;
647 /* Remember for future incarnations. */
648 cvp->cookie_desired = cookie_size;
650 if (cvp->cookie_desired != cvp->cookie_pair_size) {
651 /* Currently use random bytes as a nonce,
652 * assuming these are completely unpredictable
653 * by hostile users of the same system.
655 get_random_bytes(&cvp->cookie_pair[0],
657 cvp->cookie_pair_size = cookie_size;
663 return MAX_TCP_OPTION_SPACE - remaining;
666 /* Set up TCP options for SYN-ACKs. */
667 static unsigned tcp_synack_options(struct sock *sk,
668 struct request_sock *req,
669 unsigned mss, struct sk_buff *skb,
670 struct tcp_out_options *opts,
671 struct tcp_md5sig_key **md5,
672 struct tcp_extend_values *xvp)
674 struct inet_request_sock *ireq = inet_rsk(req);
675 unsigned remaining = MAX_TCP_OPTION_SPACE;
676 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
680 #ifdef CONFIG_TCP_MD5SIG
681 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
683 opts->options |= OPTION_MD5;
684 remaining -= TCPOLEN_MD5SIG_ALIGNED;
686 /* We can't fit any SACK blocks in a packet with MD5 + TS
687 * options. There was discussion about disabling SACK
688 * rather than TS in order to fit in better with old,
689 * buggy kernels, but that was deemed to be unnecessary.
691 ireq->tstamp_ok &= !ireq->sack_ok;
697 /* We always send an MSS option. */
699 remaining -= TCPOLEN_MSS_ALIGNED;
701 if (likely(ireq->wscale_ok)) {
702 opts->ws = ireq->rcv_wscale;
703 opts->options |= OPTION_WSCALE;
704 remaining -= TCPOLEN_WSCALE_ALIGNED;
706 if (likely(ireq->tstamp_ok)) {
707 opts->options |= OPTION_TS;
708 opts->tsval = TCP_SKB_CB(skb)->when;
709 opts->tsecr = req->ts_recent;
710 remaining -= TCPOLEN_TSTAMP_ALIGNED;
712 if (likely(ireq->sack_ok)) {
713 opts->options |= OPTION_SACK_ADVERTISE;
714 if (unlikely(!ireq->tstamp_ok))
715 remaining -= TCPOLEN_SACKPERM_ALIGNED;
718 /* Similar rationale to tcp_syn_options() applies here, too.
719 * If the <SYN> options fit, the same options should fit now!
723 cookie_plus > TCPOLEN_COOKIE_BASE) {
724 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
727 /* 32-bit multiple */
728 need += 2; /* NOPs */
730 if (need <= remaining) {
731 opts->options |= OPTION_COOKIE_EXTENSION;
732 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
735 /* There's no error return, so flag it. */
736 xvp->cookie_out_never = 1; /* true */
740 return MAX_TCP_OPTION_SPACE - remaining;
743 /* Compute TCP options for ESTABLISHED sockets. This is not the
744 * final wire format yet.
746 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
747 struct tcp_out_options *opts,
748 struct tcp_md5sig_key **md5)
750 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
751 struct tcp_sock *tp = tcp_sk(sk);
753 unsigned int eff_sacks;
755 #ifdef CONFIG_TCP_MD5SIG
756 *md5 = tp->af_specific->md5_lookup(sk, sk);
757 if (unlikely(*md5)) {
758 opts->options |= OPTION_MD5;
759 size += TCPOLEN_MD5SIG_ALIGNED;
765 if (likely(tp->rx_opt.tstamp_ok)) {
766 opts->options |= OPTION_TS;
767 opts->tsval = tcb ? tcb->when : 0;
768 opts->tsecr = tp->rx_opt.ts_recent;
769 size += TCPOLEN_TSTAMP_ALIGNED;
772 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
773 if (unlikely(eff_sacks)) {
774 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
775 opts->num_sack_blocks =
776 min_t(unsigned, eff_sacks,
777 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
778 TCPOLEN_SACK_PERBLOCK);
779 size += TCPOLEN_SACK_BASE_ALIGNED +
780 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
786 /* This routine actually transmits TCP packets queued in by
787 * tcp_do_sendmsg(). This is used by both the initial
788 * transmission and possible later retransmissions.
789 * All SKB's seen here are completely headerless. It is our
790 * job to build the TCP header, and pass the packet down to
791 * IP so it can do the same plus pass the packet off to the
794 * We are working here with either a clone of the original
795 * SKB, or a fresh unique copy made by the retransmit engine.
797 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
800 const struct inet_connection_sock *icsk = inet_csk(sk);
801 struct inet_sock *inet;
803 struct tcp_skb_cb *tcb;
804 struct tcp_out_options opts;
805 unsigned tcp_options_size, tcp_header_size;
806 struct tcp_md5sig_key *md5;
810 BUG_ON(!skb || !tcp_skb_pcount(skb));
812 /* If congestion control is doing timestamping, we must
813 * take such a timestamp before we potentially clone/copy.
815 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
816 __net_timestamp(skb);
818 if (likely(clone_it)) {
819 if (unlikely(skb_cloned(skb)))
820 skb = pskb_copy(skb, gfp_mask);
822 skb = skb_clone(skb, gfp_mask);
829 tcb = TCP_SKB_CB(skb);
830 memset(&opts, 0, sizeof(opts));
832 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
833 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
835 tcp_options_size = tcp_established_options(sk, skb, &opts,
837 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
839 if (tcp_packets_in_flight(tp) == 0)
840 tcp_ca_event(sk, CA_EVENT_TX_START);
842 /* if no packet is in qdisc/device queue, then allow XPS to select
845 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
847 skb_push(skb, tcp_header_size);
848 skb_reset_transport_header(skb);
849 skb_set_owner_w(skb, sk);
851 /* Build TCP header and checksum it. */
853 th->source = inet->inet_sport;
854 th->dest = inet->inet_dport;
855 th->seq = htonl(tcb->seq);
856 th->ack_seq = htonl(tp->rcv_nxt);
857 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
860 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
861 /* RFC1323: The window in SYN & SYN/ACK segments
864 th->window = htons(min(tp->rcv_wnd, 65535U));
866 th->window = htons(tcp_select_window(sk));
871 /* The urg_mode check is necessary during a below snd_una win probe */
872 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
873 if (before(tp->snd_up, tcb->seq + 0x10000)) {
874 th->urg_ptr = htons(tp->snd_up - tcb->seq);
876 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
877 th->urg_ptr = htons(0xFFFF);
882 tcp_options_write((__be32 *)(th + 1), tp, &opts);
883 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
884 TCP_ECN_send(sk, skb, tcp_header_size);
886 #ifdef CONFIG_TCP_MD5SIG
887 /* Calculate the MD5 hash, as we have all we need now */
889 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
890 tp->af_specific->calc_md5_hash(opts.hash_location,
895 icsk->icsk_af_ops->send_check(sk, skb);
897 if (likely(tcb->tcp_flags & TCPHDR_ACK))
898 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
900 if (skb->len != tcp_header_size)
901 tcp_event_data_sent(tp, sk);
903 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
904 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
905 tcp_skb_pcount(skb));
907 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
908 if (likely(err <= 0))
911 tcp_enter_cwr(sk, 1);
913 return net_xmit_eval(err);
916 /* This routine just queues the buffer for sending.
918 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
919 * otherwise socket can stall.
921 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
923 struct tcp_sock *tp = tcp_sk(sk);
925 /* Advance write_seq and place onto the write_queue. */
926 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
927 skb_header_release(skb);
928 tcp_add_write_queue_tail(sk, skb);
929 sk->sk_wmem_queued += skb->truesize;
930 sk_mem_charge(sk, skb->truesize);
933 /* Initialize TSO segments for a packet. */
934 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
935 unsigned int mss_now)
937 /* Make sure we own this skb before messing gso_size/gso_segs */
938 WARN_ON_ONCE(skb_cloned(skb));
940 if (skb->len <= mss_now || !sk_can_gso(sk) ||
941 skb->ip_summed == CHECKSUM_NONE) {
942 /* Avoid the costly divide in the normal
945 skb_shinfo(skb)->gso_segs = 1;
946 skb_shinfo(skb)->gso_size = 0;
947 skb_shinfo(skb)->gso_type = 0;
949 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
950 skb_shinfo(skb)->gso_size = mss_now;
951 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
955 /* When a modification to fackets out becomes necessary, we need to check
956 * skb is counted to fackets_out or not.
958 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
961 struct tcp_sock *tp = tcp_sk(sk);
963 if (!tp->sacked_out || tcp_is_reno(tp))
966 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
967 tp->fackets_out -= decr;
970 /* Pcount in the middle of the write queue got changed, we need to do various
971 * tweaks to fix counters
973 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
975 struct tcp_sock *tp = tcp_sk(sk);
977 tp->packets_out -= decr;
979 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
980 tp->sacked_out -= decr;
981 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
982 tp->retrans_out -= decr;
983 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
984 tp->lost_out -= decr;
986 /* Reno case is special. Sigh... */
987 if (tcp_is_reno(tp) && decr > 0)
988 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
990 tcp_adjust_fackets_out(sk, skb, decr);
992 if (tp->lost_skb_hint &&
993 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
994 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
995 tp->lost_cnt_hint -= decr;
997 tcp_verify_left_out(tp);
1000 /* Function to create two new TCP segments. Shrinks the given segment
1001 * to the specified size and appends a new segment with the rest of the
1002 * packet to the list. This won't be called frequently, I hope.
1003 * Remember, these are still headerless SKBs at this point.
1005 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1006 unsigned int mss_now)
1008 struct tcp_sock *tp = tcp_sk(sk);
1009 struct sk_buff *buff;
1010 int nsize, old_factor;
1014 if (WARN_ON(len > skb->len))
1017 nsize = skb_headlen(skb) - len;
1021 if (skb_unclone(skb, GFP_ATOMIC))
1024 /* Get a new skb... force flag on. */
1025 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1027 return -ENOMEM; /* We'll just try again later. */
1029 sk->sk_wmem_queued += buff->truesize;
1030 sk_mem_charge(sk, buff->truesize);
1031 nlen = skb->len - len - nsize;
1032 buff->truesize += nlen;
1033 skb->truesize -= nlen;
1035 /* Correct the sequence numbers. */
1036 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1037 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1038 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1040 /* PSH and FIN should only be set in the second packet. */
1041 flags = TCP_SKB_CB(skb)->tcp_flags;
1042 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1043 TCP_SKB_CB(buff)->tcp_flags = flags;
1044 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1046 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1047 /* Copy and checksum data tail into the new buffer. */
1048 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1049 skb_put(buff, nsize),
1054 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1056 skb->ip_summed = CHECKSUM_PARTIAL;
1057 skb_split(skb, buff, len);
1060 buff->ip_summed = skb->ip_summed;
1062 /* Looks stupid, but our code really uses when of
1063 * skbs, which it never sent before. --ANK
1065 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1066 buff->tstamp = skb->tstamp;
1068 old_factor = tcp_skb_pcount(skb);
1070 /* Fix up tso_factor for both original and new SKB. */
1071 tcp_set_skb_tso_segs(sk, skb, mss_now);
1072 tcp_set_skb_tso_segs(sk, buff, mss_now);
1074 /* If this packet has been sent out already, we must
1075 * adjust the various packet counters.
1077 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1078 int diff = old_factor - tcp_skb_pcount(skb) -
1079 tcp_skb_pcount(buff);
1082 tcp_adjust_pcount(sk, skb, diff);
1085 /* Link BUFF into the send queue. */
1086 skb_header_release(buff);
1087 tcp_insert_write_queue_after(skb, buff, sk);
1092 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1093 * eventually). The difference is that pulled data not copied, but
1094 * immediately discarded.
1096 static void __pskb_trim_head(struct sk_buff *skb, int len)
1100 eat = min_t(int, len, skb_headlen(skb));
1102 __skb_pull(skb, eat);
1109 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1110 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1113 skb_frag_unref(skb, i);
1116 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1118 skb_shinfo(skb)->frags[k].page_offset += eat;
1119 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1125 skb_shinfo(skb)->nr_frags = k;
1127 skb_reset_tail_pointer(skb);
1128 skb->data_len -= len;
1129 skb->len = skb->data_len;
1132 /* Remove acked data from a packet in the transmit queue. */
1133 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1135 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1138 __pskb_trim_head(skb, len);
1140 TCP_SKB_CB(skb)->seq += len;
1141 skb->ip_summed = CHECKSUM_PARTIAL;
1143 skb->truesize -= len;
1144 sk->sk_wmem_queued -= len;
1145 sk_mem_uncharge(sk, len);
1146 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1148 /* Any change of skb->len requires recalculation of tso factor. */
1149 if (tcp_skb_pcount(skb) > 1)
1150 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1155 /* Calculate MSS. Not accounting for SACKs here. */
1156 int tcp_mtu_to_mss(const struct sock *sk, int pmtu)
1158 const struct tcp_sock *tp = tcp_sk(sk);
1159 const struct inet_connection_sock *icsk = inet_csk(sk);
1162 /* Calculate base mss without TCP options:
1163 It is MMS_S - sizeof(tcphdr) of rfc1122
1165 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1167 /* Clamp it (mss_clamp does not include tcp options) */
1168 if (mss_now > tp->rx_opt.mss_clamp)
1169 mss_now = tp->rx_opt.mss_clamp;
1171 /* Now subtract optional transport overhead */
1172 mss_now -= icsk->icsk_ext_hdr_len;
1174 /* Then reserve room for full set of TCP options and 8 bytes of data */
1178 /* Now subtract TCP options size, not including SACKs */
1179 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1184 /* Inverse of above */
1185 int tcp_mss_to_mtu(const struct sock *sk, int mss)
1187 const struct tcp_sock *tp = tcp_sk(sk);
1188 const struct inet_connection_sock *icsk = inet_csk(sk);
1192 tp->tcp_header_len +
1193 icsk->icsk_ext_hdr_len +
1194 icsk->icsk_af_ops->net_header_len;
1199 /* MTU probing init per socket */
1200 void tcp_mtup_init(struct sock *sk)
1202 struct tcp_sock *tp = tcp_sk(sk);
1203 struct inet_connection_sock *icsk = inet_csk(sk);
1205 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1206 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1207 icsk->icsk_af_ops->net_header_len;
1208 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1209 icsk->icsk_mtup.probe_size = 0;
1211 EXPORT_SYMBOL(tcp_mtup_init);
1213 /* This function synchronize snd mss to current pmtu/exthdr set.
1215 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1216 for TCP options, but includes only bare TCP header.
1218 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1219 It is minimum of user_mss and mss received with SYN.
1220 It also does not include TCP options.
1222 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1224 tp->mss_cache is current effective sending mss, including
1225 all tcp options except for SACKs. It is evaluated,
1226 taking into account current pmtu, but never exceeds
1227 tp->rx_opt.mss_clamp.
1229 NOTE1. rfc1122 clearly states that advertised MSS
1230 DOES NOT include either tcp or ip options.
1232 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1233 are READ ONLY outside this function. --ANK (980731)
1235 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1237 struct tcp_sock *tp = tcp_sk(sk);
1238 struct inet_connection_sock *icsk = inet_csk(sk);
1241 if (icsk->icsk_mtup.search_high > pmtu)
1242 icsk->icsk_mtup.search_high = pmtu;
1244 mss_now = tcp_mtu_to_mss(sk, pmtu);
1245 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1247 /* And store cached results */
1248 icsk->icsk_pmtu_cookie = pmtu;
1249 if (icsk->icsk_mtup.enabled)
1250 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1251 tp->mss_cache = mss_now;
1255 EXPORT_SYMBOL(tcp_sync_mss);
1257 /* Compute the current effective MSS, taking SACKs and IP options,
1258 * and even PMTU discovery events into account.
1260 unsigned int tcp_current_mss(struct sock *sk)
1262 const struct tcp_sock *tp = tcp_sk(sk);
1263 const struct dst_entry *dst = __sk_dst_get(sk);
1265 unsigned header_len;
1266 struct tcp_out_options opts;
1267 struct tcp_md5sig_key *md5;
1269 mss_now = tp->mss_cache;
1272 u32 mtu = dst_mtu(dst);
1273 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1274 mss_now = tcp_sync_mss(sk, mtu);
1277 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1278 sizeof(struct tcphdr);
1279 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1280 * some common options. If this is an odd packet (because we have SACK
1281 * blocks etc) then our calculated header_len will be different, and
1282 * we have to adjust mss_now correspondingly */
1283 if (header_len != tp->tcp_header_len) {
1284 int delta = (int) header_len - tp->tcp_header_len;
1291 /* Congestion window validation. (RFC2861) */
1292 static void tcp_cwnd_validate(struct sock *sk)
1294 struct tcp_sock *tp = tcp_sk(sk);
1296 if (tp->packets_out >= tp->snd_cwnd) {
1297 /* Network is feed fully. */
1298 tp->snd_cwnd_used = 0;
1299 tp->snd_cwnd_stamp = tcp_time_stamp;
1301 /* Network starves. */
1302 if (tp->packets_out > tp->snd_cwnd_used)
1303 tp->snd_cwnd_used = tp->packets_out;
1305 if (sysctl_tcp_slow_start_after_idle &&
1306 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1307 tcp_cwnd_application_limited(sk);
1311 /* Returns the portion of skb which can be sent right away without
1312 * introducing MSS oddities to segment boundaries. In rare cases where
1313 * mss_now != mss_cache, we will request caller to create a small skb
1314 * per input skb which could be mostly avoided here (if desired).
1316 * We explicitly want to create a request for splitting write queue tail
1317 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1318 * thus all the complexity (cwnd_len is always MSS multiple which we
1319 * return whenever allowed by the other factors). Basically we need the
1320 * modulo only when the receiver window alone is the limiting factor or
1321 * when we would be allowed to send the split-due-to-Nagle skb fully.
1323 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1324 unsigned int mss_now, unsigned int cwnd)
1326 const struct tcp_sock *tp = tcp_sk(sk);
1327 u32 needed, window, cwnd_len;
1329 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1330 cwnd_len = mss_now * cwnd;
1332 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1335 needed = min(skb->len, window);
1337 if (cwnd_len <= needed)
1340 return needed - needed % mss_now;
1343 /* Can at least one segment of SKB be sent right now, according to the
1344 * congestion window rules? If so, return how many segments are allowed.
1346 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1347 const struct sk_buff *skb)
1349 u32 in_flight, cwnd;
1351 /* Don't be strict about the congestion window for the final FIN. */
1352 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1353 tcp_skb_pcount(skb) == 1)
1356 in_flight = tcp_packets_in_flight(tp);
1357 cwnd = tp->snd_cwnd;
1358 if (in_flight < cwnd)
1359 return (cwnd - in_flight);
1364 /* Initialize TSO state of a skb.
1365 * This must be invoked the first time we consider transmitting
1366 * SKB onto the wire.
1368 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1369 unsigned int mss_now)
1371 int tso_segs = tcp_skb_pcount(skb);
1373 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1374 tcp_set_skb_tso_segs(sk, skb, mss_now);
1375 tso_segs = tcp_skb_pcount(skb);
1380 /* Minshall's variant of the Nagle send check. */
1381 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1383 return after(tp->snd_sml, tp->snd_una) &&
1384 !after(tp->snd_sml, tp->snd_nxt);
1387 /* Return 0, if packet can be sent now without violation Nagle's rules:
1388 * 1. It is full sized.
1389 * 2. Or it contains FIN. (already checked by caller)
1390 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1391 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1392 * With Minshall's modification: all sent small packets are ACKed.
1394 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1395 const struct sk_buff *skb,
1396 unsigned mss_now, int nonagle)
1398 return skb->len < mss_now &&
1399 ((nonagle & TCP_NAGLE_CORK) ||
1400 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1403 /* Return non-zero if the Nagle test allows this packet to be
1406 static inline int tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1407 unsigned int cur_mss, int nonagle)
1409 /* Nagle rule does not apply to frames, which sit in the middle of the
1410 * write_queue (they have no chances to get new data).
1412 * This is implemented in the callers, where they modify the 'nonagle'
1413 * argument based upon the location of SKB in the send queue.
1415 if (nonagle & TCP_NAGLE_PUSH)
1418 /* Don't use the nagle rule for urgent data (or for the final FIN).
1419 * Nagle can be ignored during F-RTO too (see RFC4138).
1421 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1422 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1425 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1431 /* Does at least the first segment of SKB fit into the send window? */
1432 static inline int tcp_snd_wnd_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1433 unsigned int cur_mss)
1435 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1437 if (skb->len > cur_mss)
1438 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1440 return !after(end_seq, tcp_wnd_end(tp));
1443 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1444 * should be put on the wire right now. If so, it returns the number of
1445 * packets allowed by the congestion window.
1447 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1448 unsigned int cur_mss, int nonagle)
1450 const struct tcp_sock *tp = tcp_sk(sk);
1451 unsigned int cwnd_quota;
1453 tcp_init_tso_segs(sk, skb, cur_mss);
1455 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1458 cwnd_quota = tcp_cwnd_test(tp, skb);
1459 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1465 /* Test if sending is allowed right now. */
1466 int tcp_may_send_now(struct sock *sk)
1468 const struct tcp_sock *tp = tcp_sk(sk);
1469 struct sk_buff *skb = tcp_send_head(sk);
1472 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1473 (tcp_skb_is_last(sk, skb) ?
1474 tp->nonagle : TCP_NAGLE_PUSH));
1477 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1478 * which is put after SKB on the list. It is very much like
1479 * tcp_fragment() except that it may make several kinds of assumptions
1480 * in order to speed up the splitting operation. In particular, we
1481 * know that all the data is in scatter-gather pages, and that the
1482 * packet has never been sent out before (and thus is not cloned).
1484 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1485 unsigned int mss_now, gfp_t gfp)
1487 struct sk_buff *buff;
1488 int nlen = skb->len - len;
1491 /* All of a TSO frame must be composed of paged data. */
1492 if (skb->len != skb->data_len)
1493 return tcp_fragment(sk, skb, len, mss_now);
1495 buff = sk_stream_alloc_skb(sk, 0, gfp);
1496 if (unlikely(buff == NULL))
1499 sk->sk_wmem_queued += buff->truesize;
1500 sk_mem_charge(sk, buff->truesize);
1501 buff->truesize += nlen;
1502 skb->truesize -= nlen;
1504 /* Correct the sequence numbers. */
1505 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1506 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1507 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1509 /* PSH and FIN should only be set in the second packet. */
1510 flags = TCP_SKB_CB(skb)->tcp_flags;
1511 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1512 TCP_SKB_CB(buff)->tcp_flags = flags;
1514 /* This packet was never sent out yet, so no SACK bits. */
1515 TCP_SKB_CB(buff)->sacked = 0;
1517 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1518 skb_split(skb, buff, len);
1520 /* Fix up tso_factor for both original and new SKB. */
1521 tcp_set_skb_tso_segs(sk, skb, mss_now);
1522 tcp_set_skb_tso_segs(sk, buff, mss_now);
1524 /* Link BUFF into the send queue. */
1525 skb_header_release(buff);
1526 tcp_insert_write_queue_after(skb, buff, sk);
1531 /* Try to defer sending, if possible, in order to minimize the amount
1532 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1534 * This algorithm is from John Heffner.
1536 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1538 struct tcp_sock *tp = tcp_sk(sk);
1539 const struct inet_connection_sock *icsk = inet_csk(sk);
1540 u32 send_win, cong_win, limit, in_flight;
1543 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1546 if (icsk->icsk_ca_state != TCP_CA_Open)
1549 /* Defer for less than two clock ticks. */
1550 if (tp->tso_deferred &&
1551 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1554 in_flight = tcp_packets_in_flight(tp);
1556 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1558 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1560 /* From in_flight test above, we know that cwnd > in_flight. */
1561 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1563 limit = min(send_win, cong_win);
1565 /* If a full-sized TSO skb can be sent, do it. */
1566 if (limit >= sk->sk_gso_max_size)
1569 /* Middle in queue won't get any more data, full sendable already? */
1570 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1573 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1575 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1577 /* If at least some fraction of a window is available,
1580 chunk /= win_divisor;
1584 /* Different approach, try not to defer past a single
1585 * ACK. Receiver should ACK every other full sized
1586 * frame, so if we have space for more than 3 frames
1589 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1593 /* Ok, it looks like it is advisable to defer.
1594 * Do not rearm the timer if already set to not break TCP ACK clocking.
1596 if (!tp->tso_deferred)
1597 tp->tso_deferred = 1 | (jiffies << 1);
1602 tp->tso_deferred = 0;
1606 /* Create a new MTU probe if we are ready.
1607 * MTU probe is regularly attempting to increase the path MTU by
1608 * deliberately sending larger packets. This discovers routing
1609 * changes resulting in larger path MTUs.
1611 * Returns 0 if we should wait to probe (no cwnd available),
1612 * 1 if a probe was sent,
1615 static int tcp_mtu_probe(struct sock *sk)
1617 struct tcp_sock *tp = tcp_sk(sk);
1618 struct inet_connection_sock *icsk = inet_csk(sk);
1619 struct sk_buff *skb, *nskb, *next;
1626 /* Not currently probing/verifying,
1628 * have enough cwnd, and
1629 * not SACKing (the variable headers throw things off) */
1630 if (!icsk->icsk_mtup.enabled ||
1631 icsk->icsk_mtup.probe_size ||
1632 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1633 tp->snd_cwnd < 11 ||
1634 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1637 /* Very simple search strategy: just double the MSS. */
1638 mss_now = tcp_current_mss(sk);
1639 probe_size = 2 * tp->mss_cache;
1640 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1641 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1642 /* TODO: set timer for probe_converge_event */
1646 /* Have enough data in the send queue to probe? */
1647 if (tp->write_seq - tp->snd_nxt < size_needed)
1650 if (tp->snd_wnd < size_needed)
1652 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1655 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1656 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1657 if (!tcp_packets_in_flight(tp))
1663 /* We're allowed to probe. Build it now. */
1664 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1666 sk->sk_wmem_queued += nskb->truesize;
1667 sk_mem_charge(sk, nskb->truesize);
1669 skb = tcp_send_head(sk);
1671 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1672 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1673 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1674 TCP_SKB_CB(nskb)->sacked = 0;
1676 nskb->ip_summed = skb->ip_summed;
1678 tcp_insert_write_queue_before(nskb, skb, sk);
1681 tcp_for_write_queue_from_safe(skb, next, sk) {
1682 copy = min_t(int, skb->len, probe_size - len);
1683 if (nskb->ip_summed) {
1684 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1686 __wsum csum = skb_copy_and_csum_bits(skb, 0,
1687 skb_put(nskb, copy),
1689 nskb->csum = csum_block_add(nskb->csum, csum, len);
1692 if (skb->len <= copy) {
1693 /* We've eaten all the data from this skb.
1695 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1696 tcp_unlink_write_queue(skb, sk);
1697 sk_wmem_free_skb(sk, skb);
1699 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1700 ~(TCPHDR_FIN|TCPHDR_PSH);
1701 if (!skb_shinfo(skb)->nr_frags) {
1702 skb_pull(skb, copy);
1703 if (skb->ip_summed != CHECKSUM_PARTIAL)
1704 skb->csum = csum_partial(skb->data,
1707 __pskb_trim_head(skb, copy);
1708 tcp_set_skb_tso_segs(sk, skb, mss_now);
1710 TCP_SKB_CB(skb)->seq += copy;
1715 if (len >= probe_size)
1718 tcp_init_tso_segs(sk, nskb, nskb->len);
1720 /* We're ready to send. If this fails, the probe will
1721 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1722 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1723 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1724 /* Decrement cwnd here because we are sending
1725 * effectively two packets. */
1727 tcp_event_new_data_sent(sk, nskb);
1729 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1730 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1731 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1739 /* This routine writes packets to the network. It advances the
1740 * send_head. This happens as incoming acks open up the remote
1743 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1744 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1745 * account rare use of URG, this is not a big flaw.
1747 * Returns 1, if no segments are in flight and we have queued segments, but
1748 * cannot send anything now because of SWS or another problem.
1750 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1751 int push_one, gfp_t gfp)
1753 struct tcp_sock *tp = tcp_sk(sk);
1754 struct sk_buff *skb;
1755 unsigned int tso_segs, sent_pkts;
1762 /* Do MTU probing. */
1763 result = tcp_mtu_probe(sk);
1766 } else if (result > 0) {
1771 while ((skb = tcp_send_head(sk))) {
1774 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1777 cwnd_quota = tcp_cwnd_test(tp, skb);
1781 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1784 if (tso_segs == 1) {
1785 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1786 (tcp_skb_is_last(sk, skb) ?
1787 nonagle : TCP_NAGLE_PUSH))))
1790 if (!push_one && tcp_tso_should_defer(sk, skb))
1795 if (tso_segs > 1 && !tcp_urg_mode(tp))
1796 limit = tcp_mss_split_point(sk, skb, mss_now,
1799 if (skb->len > limit &&
1800 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1803 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1805 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1808 /* Advance the send_head. This one is sent out.
1809 * This call will increment packets_out.
1811 tcp_event_new_data_sent(sk, skb);
1813 tcp_minshall_update(tp, mss_now, skb);
1814 sent_pkts += tcp_skb_pcount(skb);
1819 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
1820 tp->prr_out += sent_pkts;
1822 if (likely(sent_pkts)) {
1823 tcp_cwnd_validate(sk);
1826 return !tp->packets_out && tcp_send_head(sk);
1829 /* Push out any pending frames which were held back due to
1830 * TCP_CORK or attempt at coalescing tiny packets.
1831 * The socket must be locked by the caller.
1833 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1836 /* If we are closed, the bytes will have to remain here.
1837 * In time closedown will finish, we empty the write queue and
1838 * all will be happy.
1840 if (unlikely(sk->sk_state == TCP_CLOSE))
1843 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1844 tcp_check_probe_timer(sk);
1847 /* Send _single_ skb sitting at the send head. This function requires
1848 * true push pending frames to setup probe timer etc.
1850 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1852 struct sk_buff *skb = tcp_send_head(sk);
1854 BUG_ON(!skb || skb->len < mss_now);
1856 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1859 /* This function returns the amount that we can raise the
1860 * usable window based on the following constraints
1862 * 1. The window can never be shrunk once it is offered (RFC 793)
1863 * 2. We limit memory per socket
1866 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1867 * RECV.NEXT + RCV.WIN fixed until:
1868 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1870 * i.e. don't raise the right edge of the window until you can raise
1871 * it at least MSS bytes.
1873 * Unfortunately, the recommended algorithm breaks header prediction,
1874 * since header prediction assumes th->window stays fixed.
1876 * Strictly speaking, keeping th->window fixed violates the receiver
1877 * side SWS prevention criteria. The problem is that under this rule
1878 * a stream of single byte packets will cause the right side of the
1879 * window to always advance by a single byte.
1881 * Of course, if the sender implements sender side SWS prevention
1882 * then this will not be a problem.
1884 * BSD seems to make the following compromise:
1886 * If the free space is less than the 1/4 of the maximum
1887 * space available and the free space is less than 1/2 mss,
1888 * then set the window to 0.
1889 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1890 * Otherwise, just prevent the window from shrinking
1891 * and from being larger than the largest representable value.
1893 * This prevents incremental opening of the window in the regime
1894 * where TCP is limited by the speed of the reader side taking
1895 * data out of the TCP receive queue. It does nothing about
1896 * those cases where the window is constrained on the sender side
1897 * because the pipeline is full.
1899 * BSD also seems to "accidentally" limit itself to windows that are a
1900 * multiple of MSS, at least until the free space gets quite small.
1901 * This would appear to be a side effect of the mbuf implementation.
1902 * Combining these two algorithms results in the observed behavior
1903 * of having a fixed window size at almost all times.
1905 * Below we obtain similar behavior by forcing the offered window to
1906 * a multiple of the mss when it is feasible to do so.
1908 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1909 * Regular options like TIMESTAMP are taken into account.
1911 u32 __tcp_select_window(struct sock *sk)
1913 struct inet_connection_sock *icsk = inet_csk(sk);
1914 struct tcp_sock *tp = tcp_sk(sk);
1915 /* MSS for the peer's data. Previous versions used mss_clamp
1916 * here. I don't know if the value based on our guesses
1917 * of peer's MSS is better for the performance. It's more correct
1918 * but may be worse for the performance because of rcv_mss
1919 * fluctuations. --SAW 1998/11/1
1921 int mss = icsk->icsk_ack.rcv_mss;
1922 int free_space = tcp_space(sk);
1923 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1926 if (unlikely(mss > full_space)) {
1931 if (free_space < (full_space >> 1)) {
1932 icsk->icsk_ack.quick = 0;
1934 if (tcp_memory_pressure)
1935 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1938 if (free_space < mss)
1942 if (free_space > tp->rcv_ssthresh)
1943 free_space = tp->rcv_ssthresh;
1945 /* Don't do rounding if we are using window scaling, since the
1946 * scaled window will not line up with the MSS boundary anyway.
1948 window = tp->rcv_wnd;
1949 if (tp->rx_opt.rcv_wscale) {
1950 window = free_space;
1952 /* Advertise enough space so that it won't get scaled away.
1953 * Import case: prevent zero window announcement if
1954 * 1<<rcv_wscale > mss.
1956 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1957 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1958 << tp->rx_opt.rcv_wscale);
1960 /* Get the largest window that is a nice multiple of mss.
1961 * Window clamp already applied above.
1962 * If our current window offering is within 1 mss of the
1963 * free space we just keep it. This prevents the divide
1964 * and multiply from happening most of the time.
1965 * We also don't do any window rounding when the free space
1968 if (window <= free_space - mss || window > free_space)
1969 window = (free_space / mss) * mss;
1970 else if (mss == full_space &&
1971 free_space > window + (full_space >> 1))
1972 window = free_space;
1978 /* Collapses two adjacent SKB's during retransmission. */
1979 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1983 int skb_size, next_skb_size;
1985 skb_size = skb->len;
1986 next_skb_size = next_skb->len;
1988 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1990 tcp_highest_sack_combine(sk, next_skb, skb);
1992 tcp_unlink_write_queue(next_skb, sk);
1994 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1997 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1998 skb->ip_summed = CHECKSUM_PARTIAL;
2000 if (skb->ip_summed != CHECKSUM_PARTIAL)
2001 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2003 /* Update sequence range on original skb. */
2004 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2006 /* Merge over control information. This moves PSH/FIN etc. over */
2007 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2009 /* All done, get rid of second SKB and account for it so
2010 * packet counting does not break.
2012 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2014 /* changed transmit queue under us so clear hints */
2015 tcp_clear_retrans_hints_partial(tp);
2016 if (next_skb == tp->retransmit_skb_hint)
2017 tp->retransmit_skb_hint = skb;
2019 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2021 sk_wmem_free_skb(sk, next_skb);
2024 /* Check if coalescing SKBs is legal. */
2025 static int tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2027 if (tcp_skb_pcount(skb) > 1)
2029 /* TODO: SACK collapsing could be used to remove this condition */
2030 if (skb_shinfo(skb)->nr_frags != 0)
2032 if (skb_cloned(skb))
2034 if (skb == tcp_send_head(sk))
2036 /* Some heurestics for collapsing over SACK'd could be invented */
2037 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2043 /* Collapse packets in the retransmit queue to make to create
2044 * less packets on the wire. This is only done on retransmission.
2046 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2049 struct tcp_sock *tp = tcp_sk(sk);
2050 struct sk_buff *skb = to, *tmp;
2053 if (!sysctl_tcp_retrans_collapse)
2055 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2058 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2059 if (!tcp_can_collapse(sk, skb))
2071 /* Punt if not enough space exists in the first SKB for
2072 * the data in the second
2074 if (skb->len > skb_availroom(to))
2077 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2080 tcp_collapse_retrans(sk, to);
2084 /* This retransmits one SKB. Policy decisions and retransmit queue
2085 * state updates are done by the caller. Returns non-zero if an
2086 * error occurred which prevented the send.
2088 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2090 struct tcp_sock *tp = tcp_sk(sk);
2091 struct inet_connection_sock *icsk = inet_csk(sk);
2092 unsigned int cur_mss;
2095 /* Inconslusive MTU probe */
2096 if (icsk->icsk_mtup.probe_size) {
2097 icsk->icsk_mtup.probe_size = 0;
2100 /* Do not sent more than we queued. 1/4 is reserved for possible
2101 * copying overhead: fragmentation, tunneling, mangling etc.
2103 if (atomic_read(&sk->sk_wmem_alloc) >
2104 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2108 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2109 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2111 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2115 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2116 return -EHOSTUNREACH; /* Routing failure or similar. */
2118 cur_mss = tcp_current_mss(sk);
2120 /* If receiver has shrunk his window, and skb is out of
2121 * new window, do not retransmit it. The exception is the
2122 * case, when window is shrunk to zero. In this case
2123 * our retransmit serves as a zero window probe.
2125 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2126 TCP_SKB_CB(skb)->seq != tp->snd_una)
2129 if (skb->len > cur_mss) {
2130 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2131 return -ENOMEM; /* We'll try again later. */
2133 int oldpcount = tcp_skb_pcount(skb);
2135 if (unlikely(oldpcount > 1)) {
2136 if (skb_unclone(skb, GFP_ATOMIC))
2138 tcp_init_tso_segs(sk, skb, cur_mss);
2139 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2143 tcp_retrans_try_collapse(sk, skb, cur_mss);
2145 /* Some Solaris stacks overoptimize and ignore the FIN on a
2146 * retransmit when old data is attached. So strip it off
2147 * since it is cheap to do so and saves bytes on the network.
2150 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2151 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2152 if (!pskb_trim(skb, 0)) {
2153 /* Reuse, even though it does some unnecessary work */
2154 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2155 TCP_SKB_CB(skb)->tcp_flags);
2156 skb->ip_summed = CHECKSUM_NONE;
2160 /* Make a copy, if the first transmission SKB clone we made
2161 * is still in somebody's hands, else make a clone.
2163 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2165 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2168 /* Update global TCP statistics. */
2169 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2171 tp->total_retrans++;
2173 #if FASTRETRANS_DEBUG > 0
2174 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2175 if (net_ratelimit())
2176 printk(KERN_DEBUG "retrans_out leaked.\n");
2179 if (!tp->retrans_out)
2180 tp->lost_retrans_low = tp->snd_nxt;
2181 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2182 tp->retrans_out += tcp_skb_pcount(skb);
2184 /* Save stamp of the first retransmit. */
2185 if (!tp->retrans_stamp)
2186 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2188 tp->undo_retrans += tcp_skb_pcount(skb);
2190 /* snd_nxt is stored to detect loss of retransmitted segment,
2191 * see tcp_input.c tcp_sacktag_write_queue().
2193 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2198 /* Check if we forward retransmits are possible in the current
2199 * window/congestion state.
2201 static int tcp_can_forward_retransmit(struct sock *sk)
2203 const struct inet_connection_sock *icsk = inet_csk(sk);
2204 const struct tcp_sock *tp = tcp_sk(sk);
2206 /* Forward retransmissions are possible only during Recovery. */
2207 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2210 /* No forward retransmissions in Reno are possible. */
2211 if (tcp_is_reno(tp))
2214 /* Yeah, we have to make difficult choice between forward transmission
2215 * and retransmission... Both ways have their merits...
2217 * For now we do not retransmit anything, while we have some new
2218 * segments to send. In the other cases, follow rule 3 for
2219 * NextSeg() specified in RFC3517.
2222 if (tcp_may_send_now(sk))
2228 /* This gets called after a retransmit timeout, and the initially
2229 * retransmitted data is acknowledged. It tries to continue
2230 * resending the rest of the retransmit queue, until either
2231 * we've sent it all or the congestion window limit is reached.
2232 * If doing SACK, the first ACK which comes back for a timeout
2233 * based retransmit packet might feed us FACK information again.
2234 * If so, we use it to avoid unnecessarily retransmissions.
2236 void tcp_xmit_retransmit_queue(struct sock *sk)
2238 const struct inet_connection_sock *icsk = inet_csk(sk);
2239 struct tcp_sock *tp = tcp_sk(sk);
2240 struct sk_buff *skb;
2241 struct sk_buff *hole = NULL;
2244 int fwd_rexmitting = 0;
2246 if (!tp->packets_out)
2250 tp->retransmit_high = tp->snd_una;
2252 if (tp->retransmit_skb_hint) {
2253 skb = tp->retransmit_skb_hint;
2254 last_lost = TCP_SKB_CB(skb)->end_seq;
2255 if (after(last_lost, tp->retransmit_high))
2256 last_lost = tp->retransmit_high;
2258 skb = tcp_write_queue_head(sk);
2259 last_lost = tp->snd_una;
2262 tcp_for_write_queue_from(skb, sk) {
2263 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2265 if (skb == tcp_send_head(sk))
2267 /* we could do better than to assign each time */
2269 tp->retransmit_skb_hint = skb;
2271 /* Assume this retransmit will generate
2272 * only one packet for congestion window
2273 * calculation purposes. This works because
2274 * tcp_retransmit_skb() will chop up the
2275 * packet to be MSS sized and all the
2276 * packet counting works out.
2278 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2281 if (fwd_rexmitting) {
2283 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2285 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2287 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2288 tp->retransmit_high = last_lost;
2289 if (!tcp_can_forward_retransmit(sk))
2291 /* Backtrack if necessary to non-L'ed skb */
2299 } else if (!(sacked & TCPCB_LOST)) {
2300 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2305 last_lost = TCP_SKB_CB(skb)->end_seq;
2306 if (icsk->icsk_ca_state != TCP_CA_Loss)
2307 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2309 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2312 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2315 if (tcp_retransmit_skb(sk, skb))
2317 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2319 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2320 tp->prr_out += tcp_skb_pcount(skb);
2322 if (skb == tcp_write_queue_head(sk))
2323 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2324 inet_csk(sk)->icsk_rto,
2329 /* Send a FIN. The caller locks the socket for us.
2330 * We should try to send a FIN packet really hard, but eventually give up.
2332 void tcp_send_fin(struct sock *sk)
2334 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2335 struct tcp_sock *tp = tcp_sk(sk);
2337 /* Optimization, tack on the FIN if we have one skb in write queue and
2338 * this skb was not yet sent, or we are under memory pressure.
2339 * Note: in the latter case, FIN packet will be sent after a timeout,
2340 * as TCP stack thinks it has already been transmitted.
2342 if (tskb && (tcp_send_head(sk) || tcp_memory_pressure)) {
2344 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2345 TCP_SKB_CB(tskb)->end_seq++;
2347 if (!tcp_send_head(sk)) {
2348 /* This means tskb was already sent.
2349 * Pretend we included the FIN on previous transmit.
2350 * We need to set tp->snd_nxt to the value it would have
2351 * if FIN had been sent. This is because retransmit path
2352 * does not change tp->snd_nxt.
2358 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2359 if (unlikely(!skb)) {
2365 /* Reserve space for headers and prepare control bits. */
2366 skb_reserve(skb, MAX_TCP_HEADER);
2367 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2368 tcp_init_nondata_skb(skb, tp->write_seq,
2369 TCPHDR_ACK | TCPHDR_FIN);
2370 tcp_queue_skb(sk, skb);
2372 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2375 /* We get here when a process closes a file descriptor (either due to
2376 * an explicit close() or as a byproduct of exit()'ing) and there
2377 * was unread data in the receive queue. This behavior is recommended
2378 * by RFC 2525, section 2.17. -DaveM
2380 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2382 struct sk_buff *skb;
2384 /* NOTE: No TCP options attached and we never retransmit this. */
2385 skb = alloc_skb(MAX_TCP_HEADER, priority);
2387 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2391 /* Reserve space for headers and prepare control bits. */
2392 skb_reserve(skb, MAX_TCP_HEADER);
2393 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2394 TCPHDR_ACK | TCPHDR_RST);
2396 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2397 if (tcp_transmit_skb(sk, skb, 0, priority))
2398 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2400 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2403 /* Send a crossed SYN-ACK during socket establishment.
2404 * WARNING: This routine must only be called when we have already sent
2405 * a SYN packet that crossed the incoming SYN that caused this routine
2406 * to get called. If this assumption fails then the initial rcv_wnd
2407 * and rcv_wscale values will not be correct.
2409 int tcp_send_synack(struct sock *sk)
2411 struct sk_buff *skb;
2413 skb = tcp_write_queue_head(sk);
2414 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2415 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2418 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2419 if (skb_cloned(skb)) {
2420 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2423 tcp_unlink_write_queue(skb, sk);
2424 skb_header_release(nskb);
2425 __tcp_add_write_queue_head(sk, nskb);
2426 sk_wmem_free_skb(sk, skb);
2427 sk->sk_wmem_queued += nskb->truesize;
2428 sk_mem_charge(sk, nskb->truesize);
2432 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2433 TCP_ECN_send_synack(tcp_sk(sk), skb);
2435 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2436 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2439 /* Prepare a SYN-ACK. */
2440 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2441 struct request_sock *req,
2442 struct request_values *rvp)
2444 struct tcp_out_options opts;
2445 struct tcp_extend_values *xvp = tcp_xv(rvp);
2446 struct inet_request_sock *ireq = inet_rsk(req);
2447 struct tcp_sock *tp = tcp_sk(sk);
2448 const struct tcp_cookie_values *cvp = tp->cookie_values;
2450 struct sk_buff *skb;
2451 struct tcp_md5sig_key *md5;
2452 int tcp_header_size;
2454 int s_data_desired = 0;
2456 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2457 s_data_desired = cvp->s_data_desired;
2458 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2462 /* Reserve space for headers. */
2463 skb_reserve(skb, MAX_TCP_HEADER);
2465 skb_dst_set(skb, dst_clone(dst));
2467 mss = dst_metric_advmss(dst);
2468 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2469 mss = tp->rx_opt.user_mss;
2471 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2473 /* Set this up on the first call only */
2474 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2476 /* limit the window selection if the user enforce a smaller rx buffer */
2477 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2478 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2479 req->window_clamp = tcp_full_space(sk);
2481 /* tcp_full_space because it is guaranteed to be the first packet */
2482 tcp_select_initial_window(tcp_full_space(sk),
2483 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2488 dst_metric(dst, RTAX_INITRWND));
2489 ireq->rcv_wscale = rcv_wscale;
2492 memset(&opts, 0, sizeof(opts));
2493 #ifdef CONFIG_SYN_COOKIES
2494 if (unlikely(req->cookie_ts))
2495 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2498 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2499 tcp_header_size = tcp_synack_options(sk, req, mss,
2500 skb, &opts, &md5, xvp)
2503 skb_push(skb, tcp_header_size);
2504 skb_reset_transport_header(skb);
2507 memset(th, 0, sizeof(struct tcphdr));
2510 TCP_ECN_make_synack(req, th);
2511 th->source = ireq->loc_port;
2512 th->dest = ireq->rmt_port;
2513 /* Setting of flags are superfluous here for callers (and ECE is
2514 * not even correctly set)
2516 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2517 TCPHDR_SYN | TCPHDR_ACK);
2519 if (OPTION_COOKIE_EXTENSION & opts.options) {
2520 if (s_data_desired) {
2521 u8 *buf = skb_put(skb, s_data_desired);
2523 /* copy data directly from the listening socket. */
2524 memcpy(buf, cvp->s_data_payload, s_data_desired);
2525 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2528 if (opts.hash_size > 0) {
2529 __u32 workspace[SHA_WORKSPACE_WORDS];
2530 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2531 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2533 /* Secret recipe depends on the Timestamp, (future)
2534 * Sequence and Acknowledgment Numbers, Initiator
2535 * Cookie, and others handled by IP variant caller.
2537 *tail-- ^= opts.tsval;
2538 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2539 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2542 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2543 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2545 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2548 opts.hash_location =
2549 (__u8 *)&xvp->cookie_bakery[0];
2553 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2554 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2556 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2557 th->window = htons(min(req->rcv_wnd, 65535U));
2558 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2559 th->doff = (tcp_header_size >> 2);
2560 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2562 #ifdef CONFIG_TCP_MD5SIG
2563 /* Okay, we have all we need - do the md5 hash if needed */
2565 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2566 md5, NULL, req, skb);
2572 EXPORT_SYMBOL(tcp_make_synack);
2574 /* Do all connect socket setups that can be done AF independent. */
2575 static void tcp_connect_init(struct sock *sk)
2577 const struct dst_entry *dst = __sk_dst_get(sk);
2578 struct tcp_sock *tp = tcp_sk(sk);
2581 /* We'll fix this up when we get a response from the other end.
2582 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2584 tp->tcp_header_len = sizeof(struct tcphdr) +
2585 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2587 #ifdef CONFIG_TCP_MD5SIG
2588 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2589 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2592 /* If user gave his TCP_MAXSEG, record it to clamp */
2593 if (tp->rx_opt.user_mss)
2594 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2597 tcp_sync_mss(sk, dst_mtu(dst));
2599 if (!tp->window_clamp)
2600 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2601 tp->advmss = dst_metric_advmss(dst);
2602 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2603 tp->advmss = tp->rx_opt.user_mss;
2605 tcp_initialize_rcv_mss(sk);
2607 /* limit the window selection if the user enforce a smaller rx buffer */
2608 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2609 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2610 tp->window_clamp = tcp_full_space(sk);
2612 tcp_select_initial_window(tcp_full_space(sk),
2613 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2616 sysctl_tcp_window_scaling,
2618 dst_metric(dst, RTAX_INITRWND));
2620 tp->rx_opt.rcv_wscale = rcv_wscale;
2621 tp->rcv_ssthresh = tp->rcv_wnd;
2624 sock_reset_flag(sk, SOCK_DONE);
2627 tp->snd_una = tp->write_seq;
2628 tp->snd_sml = tp->write_seq;
2629 tp->snd_up = tp->write_seq;
2634 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2635 inet_csk(sk)->icsk_retransmits = 0;
2636 tcp_clear_retrans(tp);
2639 /* Build a SYN and send it off. */
2640 int tcp_connect(struct sock *sk)
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 struct sk_buff *buff;
2646 tcp_connect_init(sk);
2648 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
2649 if (unlikely(!buff))
2652 tp->snd_nxt = tp->write_seq;
2653 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2654 TCP_ECN_send_syn(sk, buff);
2657 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2658 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2659 skb_header_release(buff);
2660 __tcp_add_write_queue_tail(sk, buff);
2661 sk->sk_wmem_queued += buff->truesize;
2662 sk_mem_charge(sk, buff->truesize);
2663 tp->packets_out += tcp_skb_pcount(buff);
2664 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2665 if (err == -ECONNREFUSED)
2668 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2669 * in order to make this packet get counted in tcpOutSegs.
2671 tp->snd_nxt = tp->write_seq;
2672 tp->pushed_seq = tp->write_seq;
2673 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2675 /* Timer for repeating the SYN until an answer. */
2676 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2677 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2680 EXPORT_SYMBOL(tcp_connect);
2682 /* Send out a delayed ack, the caller does the policy checking
2683 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2686 void tcp_send_delayed_ack(struct sock *sk)
2688 struct inet_connection_sock *icsk = inet_csk(sk);
2689 int ato = icsk->icsk_ack.ato;
2690 unsigned long timeout;
2692 if (ato > TCP_DELACK_MIN) {
2693 const struct tcp_sock *tp = tcp_sk(sk);
2694 int max_ato = HZ / 2;
2696 if (icsk->icsk_ack.pingpong ||
2697 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2698 max_ato = TCP_DELACK_MAX;
2700 /* Slow path, intersegment interval is "high". */
2702 /* If some rtt estimate is known, use it to bound delayed ack.
2703 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2707 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2713 ato = min(ato, max_ato);
2716 /* Stay within the limit we were given */
2717 timeout = jiffies + ato;
2719 /* Use new timeout only if there wasn't a older one earlier. */
2720 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2721 /* If delack timer was blocked or is about to expire,
2724 if (icsk->icsk_ack.blocked ||
2725 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2730 if (!time_before(timeout, icsk->icsk_ack.timeout))
2731 timeout = icsk->icsk_ack.timeout;
2733 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2734 icsk->icsk_ack.timeout = timeout;
2735 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2738 /* This routine sends an ack and also updates the window. */
2739 void tcp_send_ack(struct sock *sk)
2741 struct sk_buff *buff;
2743 /* If we have been reset, we may not send again. */
2744 if (sk->sk_state == TCP_CLOSE)
2747 /* We are not putting this on the write queue, so
2748 * tcp_transmit_skb() will set the ownership to this
2751 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2753 inet_csk_schedule_ack(sk);
2754 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2755 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2756 TCP_DELACK_MAX, TCP_RTO_MAX);
2760 /* Reserve space for headers and prepare control bits. */
2761 skb_reserve(buff, MAX_TCP_HEADER);
2762 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2764 /* Send it off, this clears delayed acks for us. */
2765 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2766 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2769 /* This routine sends a packet with an out of date sequence
2770 * number. It assumes the other end will try to ack it.
2772 * Question: what should we make while urgent mode?
2773 * 4.4BSD forces sending single byte of data. We cannot send
2774 * out of window data, because we have SND.NXT==SND.MAX...
2776 * Current solution: to send TWO zero-length segments in urgent mode:
2777 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2778 * out-of-date with SND.UNA-1 to probe window.
2780 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2782 struct tcp_sock *tp = tcp_sk(sk);
2783 struct sk_buff *skb;
2785 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2786 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2790 /* Reserve space for headers and set control bits. */
2791 skb_reserve(skb, MAX_TCP_HEADER);
2792 /* Use a previous sequence. This should cause the other
2793 * end to send an ack. Don't queue or clone SKB, just
2796 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2797 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2798 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2801 /* Initiate keepalive or window probe from timer. */
2802 int tcp_write_wakeup(struct sock *sk)
2804 struct tcp_sock *tp = tcp_sk(sk);
2805 struct sk_buff *skb;
2807 if (sk->sk_state == TCP_CLOSE)
2810 if ((skb = tcp_send_head(sk)) != NULL &&
2811 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2813 unsigned int mss = tcp_current_mss(sk);
2814 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2816 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2817 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2819 /* We are probing the opening of a window
2820 * but the window size is != 0
2821 * must have been a result SWS avoidance ( sender )
2823 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2825 seg_size = min(seg_size, mss);
2826 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2827 if (tcp_fragment(sk, skb, seg_size, mss))
2829 } else if (!tcp_skb_pcount(skb))
2830 tcp_set_skb_tso_segs(sk, skb, mss);
2832 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2833 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2834 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2836 tcp_event_new_data_sent(sk, skb);
2839 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2840 tcp_xmit_probe_skb(sk, 1);
2841 return tcp_xmit_probe_skb(sk, 0);
2845 /* A window probe timeout has occurred. If window is not closed send
2846 * a partial packet else a zero probe.
2848 void tcp_send_probe0(struct sock *sk)
2850 struct inet_connection_sock *icsk = inet_csk(sk);
2851 struct tcp_sock *tp = tcp_sk(sk);
2854 err = tcp_write_wakeup(sk);
2856 if (tp->packets_out || !tcp_send_head(sk)) {
2857 /* Cancel probe timer, if it is not required. */
2858 icsk->icsk_probes_out = 0;
2859 icsk->icsk_backoff = 0;
2864 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2865 icsk->icsk_backoff++;
2866 icsk->icsk_probes_out++;
2867 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2868 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2871 /* If packet was not sent due to local congestion,
2872 * do not backoff and do not remember icsk_probes_out.
2873 * Let local senders to fight for local resources.
2875 * Use accumulated backoff yet.
2877 if (!icsk->icsk_probes_out)
2878 icsk->icsk_probes_out = 1;
2879 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2880 min(icsk->icsk_rto << icsk->icsk_backoff,
2881 TCP_RESOURCE_PROBE_INTERVAL),