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 = 512;
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 indefinately 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;
122 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) {
123 mss = dst_metric(dst, RTAX_ADVMSS);
130 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
131 * This is the first part of cwnd validation mechanism. */
132 static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
134 struct tcp_sock *tp = tcp_sk(sk);
135 s32 delta = tcp_time_stamp - tp->lsndtime;
136 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
137 u32 cwnd = tp->snd_cwnd;
139 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
141 tp->snd_ssthresh = tcp_current_ssthresh(sk);
142 restart_cwnd = min(restart_cwnd, cwnd);
144 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
146 tp->snd_cwnd = max(cwnd, restart_cwnd);
147 tp->snd_cwnd_stamp = tcp_time_stamp;
148 tp->snd_cwnd_used = 0;
151 /* Congestion state accounting after a packet has been sent. */
152 static void tcp_event_data_sent(struct tcp_sock *tp,
153 struct sk_buff *skb, struct sock *sk)
155 struct inet_connection_sock *icsk = inet_csk(sk);
156 const u32 now = tcp_time_stamp;
158 if (sysctl_tcp_slow_start_after_idle &&
159 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
160 tcp_cwnd_restart(sk, __sk_dst_get(sk));
164 /* If it is a reply for ato after last received
165 * packet, enter pingpong mode.
167 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
168 icsk->icsk_ack.pingpong = 1;
171 /* Account for an ACK we sent. */
172 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
174 tcp_dec_quickack_mode(sk, pkts);
175 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
178 /* Determine a window scaling and initial window to offer.
179 * Based on the assumption that the given amount of space
180 * will be offered. Store the results in the tp structure.
181 * NOTE: for smooth operation initial space offering should
182 * be a multiple of mss if possible. We assume here that mss >= 1.
183 * This MUST be enforced by all callers.
185 void tcp_select_initial_window(int __space, __u32 mss,
186 __u32 *rcv_wnd, __u32 *window_clamp,
187 int wscale_ok, __u8 *rcv_wscale,
190 unsigned int space = (__space < 0 ? 0 : __space);
192 /* If no clamp set the clamp to the max possible scaled window */
193 if (*window_clamp == 0)
194 (*window_clamp) = (65535 << 14);
195 space = min(*window_clamp, space);
197 /* Quantize space offering to a multiple of mss if possible. */
199 space = (space / mss) * mss;
201 /* NOTE: offering an initial window larger than 32767
202 * will break some buggy TCP stacks. If the admin tells us
203 * it is likely we could be speaking with such a buggy stack
204 * we will truncate our initial window offering to 32K-1
205 * unless the remote has sent us a window scaling option,
206 * which we interpret as a sign the remote TCP is not
207 * misinterpreting the window field as a signed quantity.
209 if (sysctl_tcp_workaround_signed_windows)
210 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
216 /* Set window scaling on max possible window
217 * See RFC1323 for an explanation of the limit to 14
219 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
220 space = min_t(u32, space, *window_clamp);
221 while (space > 65535 && (*rcv_wscale) < 14) {
227 /* Set initial window to value enough for senders, following RFC5681. */
228 if (mss > (1 << *rcv_wscale)) {
229 int init_cwnd = rfc3390_bytes_to_packets(mss);
231 /* when initializing use the value from init_rcv_wnd
232 * rather than the default from above
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
237 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
240 /* Set the clamp no higher than max representable value */
241 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
243 EXPORT_SYMBOL(tcp_select_initial_window);
245 /* Chose a new window to advertise, update state in tcp_sock for the
246 * socket, and return result with RFC1323 scaling applied. The return
247 * value can be stuffed directly into th->window for an outgoing
250 static u16 tcp_select_window(struct sock *sk)
252 struct tcp_sock *tp = tcp_sk(sk);
253 u32 cur_win = tcp_receive_window(tp);
254 u32 new_win = __tcp_select_window(sk);
256 /* Never shrink the offered window */
257 if (new_win < cur_win) {
258 /* Danger Will Robinson!
259 * Don't update rcv_wup/rcv_wnd here or else
260 * we will not be able to advertise a zero
261 * window in time. --DaveM
263 * Relax Will Robinson.
265 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
267 tp->rcv_wnd = new_win;
268 tp->rcv_wup = tp->rcv_nxt;
270 /* Make sure we do not exceed the maximum possible
273 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
274 new_win = min(new_win, MAX_TCP_WINDOW);
276 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
278 /* RFC1323 scaling applied */
279 new_win >>= tp->rx_opt.rcv_wscale;
281 /* If we advertise zero window, disable fast path. */
288 /* Packet ECN state for a SYN-ACK */
289 static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
291 TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
292 if (!(tp->ecn_flags & TCP_ECN_OK))
293 TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
296 /* Packet ECN state for a SYN. */
297 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
299 struct tcp_sock *tp = tcp_sk(sk);
302 if (sysctl_tcp_ecn == 1) {
303 TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
304 tp->ecn_flags = TCP_ECN_OK;
308 static __inline__ void
309 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
311 if (inet_rsk(req)->ecn_ok)
315 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
318 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
321 struct tcp_sock *tp = tcp_sk(sk);
323 if (tp->ecn_flags & TCP_ECN_OK) {
324 /* Not-retransmitted data segment: set ECT and inject CWR. */
325 if (skb->len != tcp_header_len &&
326 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
328 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
329 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
330 tcp_hdr(skb)->cwr = 1;
331 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
334 /* ACK or retransmitted segment: clear ECT|CE */
335 INET_ECN_dontxmit(sk);
337 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
338 tcp_hdr(skb)->ece = 1;
342 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
343 * auto increment end seqno.
345 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
347 skb->ip_summed = CHECKSUM_PARTIAL;
350 TCP_SKB_CB(skb)->flags = flags;
351 TCP_SKB_CB(skb)->sacked = 0;
353 skb_shinfo(skb)->gso_segs = 1;
354 skb_shinfo(skb)->gso_size = 0;
355 skb_shinfo(skb)->gso_type = 0;
357 TCP_SKB_CB(skb)->seq = seq;
358 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
360 TCP_SKB_CB(skb)->end_seq = seq;
363 static inline int tcp_urg_mode(const struct tcp_sock *tp)
365 return tp->snd_una != tp->snd_up;
368 #define OPTION_SACK_ADVERTISE (1 << 0)
369 #define OPTION_TS (1 << 1)
370 #define OPTION_MD5 (1 << 2)
371 #define OPTION_WSCALE (1 << 3)
372 #define OPTION_COOKIE_EXTENSION (1 << 4)
374 struct tcp_out_options {
375 u8 options; /* bit field of OPTION_* */
376 u8 ws; /* window scale, 0 to disable */
377 u8 num_sack_blocks; /* number of SACK blocks to include */
378 u8 hash_size; /* bytes in hash_location */
379 u16 mss; /* 0 to disable */
380 __u32 tsval, tsecr; /* need to include OPTION_TS */
381 __u8 *hash_location; /* temporary pointer, overloaded */
384 /* The sysctl int routines are generic, so check consistency here.
386 static u8 tcp_cookie_size_check(u8 desired)
391 /* previously specified */
394 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
395 if (cookie_size <= 0)
396 /* no default specified */
399 if (cookie_size <= TCP_COOKIE_MIN)
400 /* value too small, specify minimum */
401 return TCP_COOKIE_MIN;
403 if (cookie_size >= TCP_COOKIE_MAX)
404 /* value too large, specify maximum */
405 return TCP_COOKIE_MAX;
408 /* 8-bit multiple, illegal, fix it */
411 return (u8)cookie_size;
414 /* Write previously computed TCP options to the packet.
416 * Beware: Something in the Internet is very sensitive to the ordering of
417 * TCP options, we learned this through the hard way, so be careful here.
418 * Luckily we can at least blame others for their non-compliance but from
419 * inter-operatibility perspective it seems that we're somewhat stuck with
420 * the ordering which we have been using if we want to keep working with
421 * those broken things (not that it currently hurts anybody as there isn't
422 * particular reason why the ordering would need to be changed).
424 * At least SACK_PERM as the first option is known to lead to a disaster
425 * (but it may well be that other scenarios fail similarly).
427 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
428 struct tcp_out_options *opts)
430 u8 options = opts->options; /* mungable copy */
432 /* Having both authentication and cookies for security is redundant,
433 * and there's certainly not enough room. Instead, the cookie-less
434 * extension variant is proposed.
436 * Consider the pessimal case with authentication. The options
438 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
440 if (unlikely(OPTION_MD5 & options)) {
441 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
442 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
443 (TCPOLEN_COOKIE_BASE << 16) |
444 (TCPOPT_MD5SIG << 8) |
447 *ptr++ = htonl((TCPOPT_NOP << 24) |
449 (TCPOPT_MD5SIG << 8) |
452 options &= ~OPTION_COOKIE_EXTENSION;
453 /* overload cookie hash location */
454 opts->hash_location = (__u8 *)ptr;
458 if (unlikely(opts->mss)) {
459 *ptr++ = htonl((TCPOPT_MSS << 24) |
460 (TCPOLEN_MSS << 16) |
464 if (likely(OPTION_TS & options)) {
465 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467 (TCPOLEN_SACK_PERM << 16) |
468 (TCPOPT_TIMESTAMP << 8) |
470 options &= ~OPTION_SACK_ADVERTISE;
472 *ptr++ = htonl((TCPOPT_NOP << 24) |
474 (TCPOPT_TIMESTAMP << 8) |
477 *ptr++ = htonl(opts->tsval);
478 *ptr++ = htonl(opts->tsecr);
481 /* Specification requires after timestamp, so do it now.
483 * Consider the pessimal case without authentication. The options
485 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
487 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
488 __u8 *cookie_copy = opts->hash_location;
489 u8 cookie_size = opts->hash_size;
491 /* 8-bit multiple handled in tcp_cookie_size_check() above,
494 if (0x2 & cookie_size) {
495 __u8 *p = (__u8 *)ptr;
497 /* 16-bit multiple */
498 *p++ = TCPOPT_COOKIE;
499 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
500 *p++ = *cookie_copy++;
501 *p++ = *cookie_copy++;
505 /* 32-bit multiple */
506 *ptr++ = htonl(((TCPOPT_NOP << 24) |
508 (TCPOPT_COOKIE << 8) |
509 TCPOLEN_COOKIE_BASE) +
513 if (cookie_size > 0) {
514 memcpy(ptr, cookie_copy, cookie_size);
515 ptr += (cookie_size / 4);
519 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
520 *ptr++ = htonl((TCPOPT_NOP << 24) |
522 (TCPOPT_SACK_PERM << 8) |
526 if (unlikely(OPTION_WSCALE & options)) {
527 *ptr++ = htonl((TCPOPT_NOP << 24) |
528 (TCPOPT_WINDOW << 16) |
529 (TCPOLEN_WINDOW << 8) |
533 if (unlikely(opts->num_sack_blocks)) {
534 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
535 tp->duplicate_sack : tp->selective_acks;
538 *ptr++ = htonl((TCPOPT_NOP << 24) |
541 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
542 TCPOLEN_SACK_PERBLOCK)));
544 for (this_sack = 0; this_sack < opts->num_sack_blocks;
546 *ptr++ = htonl(sp[this_sack].start_seq);
547 *ptr++ = htonl(sp[this_sack].end_seq);
550 tp->rx_opt.dsack = 0;
554 /* Compute TCP options for SYN packets. This is not the final
555 * network wire format yet.
557 static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
558 struct tcp_out_options *opts,
559 struct tcp_md5sig_key **md5) {
560 struct tcp_sock *tp = tcp_sk(sk);
561 struct tcp_cookie_values *cvp = tp->cookie_values;
562 unsigned remaining = MAX_TCP_OPTION_SPACE;
563 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
564 tcp_cookie_size_check(cvp->cookie_desired) :
567 #ifdef CONFIG_TCP_MD5SIG
568 *md5 = tp->af_specific->md5_lookup(sk, sk);
570 opts->options |= OPTION_MD5;
571 remaining -= TCPOLEN_MD5SIG_ALIGNED;
577 /* We always get an MSS option. The option bytes which will be seen in
578 * normal data packets should timestamps be used, must be in the MSS
579 * advertised. But we subtract them from tp->mss_cache so that
580 * calculations in tcp_sendmsg are simpler etc. So account for this
581 * fact here if necessary. If we don't do this correctly, as a
582 * receiver we won't recognize data packets as being full sized when we
583 * should, and thus we won't abide by the delayed ACK rules correctly.
584 * SACKs don't matter, we never delay an ACK when we have any of those
586 opts->mss = tcp_advertise_mss(sk);
587 remaining -= TCPOLEN_MSS_ALIGNED;
589 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
590 opts->options |= OPTION_TS;
591 opts->tsval = TCP_SKB_CB(skb)->when;
592 opts->tsecr = tp->rx_opt.ts_recent;
593 remaining -= TCPOLEN_TSTAMP_ALIGNED;
595 if (likely(sysctl_tcp_window_scaling)) {
596 opts->ws = tp->rx_opt.rcv_wscale;
597 opts->options |= OPTION_WSCALE;
598 remaining -= TCPOLEN_WSCALE_ALIGNED;
600 if (likely(sysctl_tcp_sack)) {
601 opts->options |= OPTION_SACK_ADVERTISE;
602 if (unlikely(!(OPTION_TS & opts->options)))
603 remaining -= TCPOLEN_SACKPERM_ALIGNED;
606 /* Note that timestamps are required by the specification.
608 * Odd numbers of bytes are prohibited by the specification, ensuring
609 * that the cookie is 16-bit aligned, and the resulting cookie pair is
613 (OPTION_TS & opts->options) &&
615 int need = TCPOLEN_COOKIE_BASE + cookie_size;
618 /* 32-bit multiple */
619 need += 2; /* NOPs */
621 if (need > remaining) {
622 /* try shrinking cookie to fit */
627 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
631 if (TCP_COOKIE_MIN <= cookie_size) {
632 opts->options |= OPTION_COOKIE_EXTENSION;
633 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
634 opts->hash_size = cookie_size;
636 /* Remember for future incarnations. */
637 cvp->cookie_desired = cookie_size;
639 if (cvp->cookie_desired != cvp->cookie_pair_size) {
640 /* Currently use random bytes as a nonce,
641 * assuming these are completely unpredictable
642 * by hostile users of the same system.
644 get_random_bytes(&cvp->cookie_pair[0],
646 cvp->cookie_pair_size = cookie_size;
652 return MAX_TCP_OPTION_SPACE - remaining;
655 /* Set up TCP options for SYN-ACKs. */
656 static unsigned tcp_synack_options(struct sock *sk,
657 struct request_sock *req,
658 unsigned mss, struct sk_buff *skb,
659 struct tcp_out_options *opts,
660 struct tcp_md5sig_key **md5,
661 struct tcp_extend_values *xvp)
663 struct inet_request_sock *ireq = inet_rsk(req);
664 unsigned remaining = MAX_TCP_OPTION_SPACE;
665 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
669 #ifdef CONFIG_TCP_MD5SIG
670 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
672 opts->options |= OPTION_MD5;
673 remaining -= TCPOLEN_MD5SIG_ALIGNED;
675 /* We can't fit any SACK blocks in a packet with MD5 + TS
676 * options. There was discussion about disabling SACK
677 * rather than TS in order to fit in better with old,
678 * buggy kernels, but that was deemed to be unnecessary.
680 ireq->tstamp_ok &= !ireq->sack_ok;
686 /* We always send an MSS option. */
688 remaining -= TCPOLEN_MSS_ALIGNED;
690 if (likely(ireq->wscale_ok)) {
691 opts->ws = ireq->rcv_wscale;
692 opts->options |= OPTION_WSCALE;
693 remaining -= TCPOLEN_WSCALE_ALIGNED;
695 if (likely(ireq->tstamp_ok)) {
696 opts->options |= OPTION_TS;
697 opts->tsval = TCP_SKB_CB(skb)->when;
698 opts->tsecr = req->ts_recent;
699 remaining -= TCPOLEN_TSTAMP_ALIGNED;
701 if (likely(ireq->sack_ok)) {
702 opts->options |= OPTION_SACK_ADVERTISE;
703 if (unlikely(!ireq->tstamp_ok))
704 remaining -= TCPOLEN_SACKPERM_ALIGNED;
707 /* Similar rationale to tcp_syn_options() applies here, too.
708 * If the <SYN> options fit, the same options should fit now!
712 cookie_plus > TCPOLEN_COOKIE_BASE) {
713 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
716 /* 32-bit multiple */
717 need += 2; /* NOPs */
719 if (need <= remaining) {
720 opts->options |= OPTION_COOKIE_EXTENSION;
721 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
724 /* There's no error return, so flag it. */
725 xvp->cookie_out_never = 1; /* true */
729 return MAX_TCP_OPTION_SPACE - remaining;
732 /* Compute TCP options for ESTABLISHED sockets. This is not the
733 * final wire format yet.
735 static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
736 struct tcp_out_options *opts,
737 struct tcp_md5sig_key **md5) {
738 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
739 struct tcp_sock *tp = tcp_sk(sk);
741 unsigned int eff_sacks;
743 #ifdef CONFIG_TCP_MD5SIG
744 *md5 = tp->af_specific->md5_lookup(sk, sk);
745 if (unlikely(*md5)) {
746 opts->options |= OPTION_MD5;
747 size += TCPOLEN_MD5SIG_ALIGNED;
753 if (likely(tp->rx_opt.tstamp_ok)) {
754 opts->options |= OPTION_TS;
755 opts->tsval = tcb ? tcb->when : 0;
756 opts->tsecr = tp->rx_opt.ts_recent;
757 size += TCPOLEN_TSTAMP_ALIGNED;
760 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
761 if (unlikely(eff_sacks)) {
762 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
763 opts->num_sack_blocks =
764 min_t(unsigned, eff_sacks,
765 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
766 TCPOLEN_SACK_PERBLOCK);
767 size += TCPOLEN_SACK_BASE_ALIGNED +
768 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
774 /* This routine actually transmits TCP packets queued in by
775 * tcp_do_sendmsg(). This is used by both the initial
776 * transmission and possible later retransmissions.
777 * All SKB's seen here are completely headerless. It is our
778 * job to build the TCP header, and pass the packet down to
779 * IP so it can do the same plus pass the packet off to the
782 * We are working here with either a clone of the original
783 * SKB, or a fresh unique copy made by the retransmit engine.
785 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
788 const struct inet_connection_sock *icsk = inet_csk(sk);
789 struct inet_sock *inet;
791 struct tcp_skb_cb *tcb;
792 struct tcp_out_options opts;
793 unsigned tcp_options_size, tcp_header_size;
794 struct tcp_md5sig_key *md5;
798 BUG_ON(!skb || !tcp_skb_pcount(skb));
800 /* If congestion control is doing timestamping, we must
801 * take such a timestamp before we potentially clone/copy.
803 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
804 __net_timestamp(skb);
806 if (likely(clone_it)) {
807 if (unlikely(skb_cloned(skb)))
808 skb = pskb_copy(skb, gfp_mask);
810 skb = skb_clone(skb, gfp_mask);
817 tcb = TCP_SKB_CB(skb);
818 memset(&opts, 0, sizeof(opts));
820 if (unlikely(tcb->flags & TCPHDR_SYN))
821 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
823 tcp_options_size = tcp_established_options(sk, skb, &opts,
825 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
827 if (tcp_packets_in_flight(tp) == 0)
828 tcp_ca_event(sk, CA_EVENT_TX_START);
830 skb_push(skb, tcp_header_size);
831 skb_reset_transport_header(skb);
832 skb_set_owner_w(skb, sk);
834 /* Build TCP header and checksum it. */
836 th->source = inet->inet_sport;
837 th->dest = inet->inet_dport;
838 th->seq = htonl(tcb->seq);
839 th->ack_seq = htonl(tp->rcv_nxt);
840 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
843 if (unlikely(tcb->flags & TCPHDR_SYN)) {
844 /* RFC1323: The window in SYN & SYN/ACK segments
847 th->window = htons(min(tp->rcv_wnd, 65535U));
849 th->window = htons(tcp_select_window(sk));
854 /* The urg_mode check is necessary during a below snd_una win probe */
855 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
856 if (before(tp->snd_up, tcb->seq + 0x10000)) {
857 th->urg_ptr = htons(tp->snd_up - tcb->seq);
859 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
860 th->urg_ptr = htons(0xFFFF);
865 tcp_options_write((__be32 *)(th + 1), tp, &opts);
866 if (likely((tcb->flags & TCPHDR_SYN) == 0))
867 TCP_ECN_send(sk, skb, tcp_header_size);
869 #ifdef CONFIG_TCP_MD5SIG
870 /* Calculate the MD5 hash, as we have all we need now */
872 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
873 tp->af_specific->calc_md5_hash(opts.hash_location,
878 icsk->icsk_af_ops->send_check(sk, skb);
880 if (likely(tcb->flags & TCPHDR_ACK))
881 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
883 if (skb->len != tcp_header_size)
884 tcp_event_data_sent(tp, skb, sk);
886 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
887 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
888 tcp_skb_pcount(skb));
890 err = icsk->icsk_af_ops->queue_xmit(skb);
891 if (likely(err <= 0))
894 tcp_enter_cwr(sk, 1);
896 return net_xmit_eval(err);
899 /* This routine just queues the buffer for sending.
901 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
902 * otherwise socket can stall.
904 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
906 struct tcp_sock *tp = tcp_sk(sk);
908 /* Advance write_seq and place onto the write_queue. */
909 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
910 skb_header_release(skb);
911 tcp_add_write_queue_tail(sk, skb);
912 sk->sk_wmem_queued += skb->truesize;
913 sk_mem_charge(sk, skb->truesize);
916 /* Initialize TSO segments for a packet. */
917 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
918 unsigned int mss_now)
920 if (skb->len <= mss_now || !sk_can_gso(sk) ||
921 skb->ip_summed == CHECKSUM_NONE) {
922 /* Avoid the costly divide in the normal
925 skb_shinfo(skb)->gso_segs = 1;
926 skb_shinfo(skb)->gso_size = 0;
927 skb_shinfo(skb)->gso_type = 0;
929 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
930 skb_shinfo(skb)->gso_size = mss_now;
931 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
935 /* When a modification to fackets out becomes necessary, we need to check
936 * skb is counted to fackets_out or not.
938 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
941 struct tcp_sock *tp = tcp_sk(sk);
943 if (!tp->sacked_out || tcp_is_reno(tp))
946 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
947 tp->fackets_out -= decr;
950 /* Pcount in the middle of the write queue got changed, we need to do various
951 * tweaks to fix counters
953 static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
955 struct tcp_sock *tp = tcp_sk(sk);
957 tp->packets_out -= decr;
959 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
960 tp->sacked_out -= decr;
961 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
962 tp->retrans_out -= decr;
963 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
964 tp->lost_out -= decr;
966 /* Reno case is special. Sigh... */
967 if (tcp_is_reno(tp) && decr > 0)
968 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
970 tcp_adjust_fackets_out(sk, skb, decr);
972 if (tp->lost_skb_hint &&
973 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
974 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
975 tp->lost_cnt_hint -= decr;
977 tcp_verify_left_out(tp);
980 /* Function to create two new TCP segments. Shrinks the given segment
981 * to the specified size and appends a new segment with the rest of the
982 * packet to the list. This won't be called frequently, I hope.
983 * Remember, these are still headerless SKBs at this point.
985 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
986 unsigned int mss_now)
988 struct tcp_sock *tp = tcp_sk(sk);
989 struct sk_buff *buff;
990 int nsize, old_factor;
994 BUG_ON(len > skb->len);
996 nsize = skb_headlen(skb) - len;
1000 if (skb_cloned(skb) &&
1001 skb_is_nonlinear(skb) &&
1002 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1005 /* Get a new skb... force flag on. */
1006 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1008 return -ENOMEM; /* We'll just try again later. */
1010 sk->sk_wmem_queued += buff->truesize;
1011 sk_mem_charge(sk, buff->truesize);
1012 nlen = skb->len - len - nsize;
1013 buff->truesize += nlen;
1014 skb->truesize -= nlen;
1016 /* Correct the sequence numbers. */
1017 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1018 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1019 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1021 /* PSH and FIN should only be set in the second packet. */
1022 flags = TCP_SKB_CB(skb)->flags;
1023 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1024 TCP_SKB_CB(buff)->flags = flags;
1025 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1027 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1028 /* Copy and checksum data tail into the new buffer. */
1029 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1030 skb_put(buff, nsize),
1035 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1037 skb->ip_summed = CHECKSUM_PARTIAL;
1038 skb_split(skb, buff, len);
1041 buff->ip_summed = skb->ip_summed;
1043 /* Looks stupid, but our code really uses when of
1044 * skbs, which it never sent before. --ANK
1046 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1047 buff->tstamp = skb->tstamp;
1049 old_factor = tcp_skb_pcount(skb);
1051 /* Fix up tso_factor for both original and new SKB. */
1052 tcp_set_skb_tso_segs(sk, skb, mss_now);
1053 tcp_set_skb_tso_segs(sk, buff, mss_now);
1055 /* If this packet has been sent out already, we must
1056 * adjust the various packet counters.
1058 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1059 int diff = old_factor - tcp_skb_pcount(skb) -
1060 tcp_skb_pcount(buff);
1063 tcp_adjust_pcount(sk, skb, diff);
1066 /* Link BUFF into the send queue. */
1067 skb_header_release(buff);
1068 tcp_insert_write_queue_after(skb, buff, sk);
1073 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1074 * eventually). The difference is that pulled data not copied, but
1075 * immediately discarded.
1077 static void __pskb_trim_head(struct sk_buff *skb, int len)
1083 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1084 if (skb_shinfo(skb)->frags[i].size <= eat) {
1085 put_page(skb_shinfo(skb)->frags[i].page);
1086 eat -= skb_shinfo(skb)->frags[i].size;
1088 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1090 skb_shinfo(skb)->frags[k].page_offset += eat;
1091 skb_shinfo(skb)->frags[k].size -= eat;
1097 skb_shinfo(skb)->nr_frags = k;
1099 skb_reset_tail_pointer(skb);
1100 skb->data_len -= len;
1101 skb->len = skb->data_len;
1104 /* Remove acked data from a packet in the transmit queue. */
1105 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1107 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1110 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1111 if (unlikely(len < skb_headlen(skb)))
1112 __skb_pull(skb, len);
1114 __pskb_trim_head(skb, len - skb_headlen(skb));
1116 TCP_SKB_CB(skb)->seq += len;
1117 skb->ip_summed = CHECKSUM_PARTIAL;
1119 skb->truesize -= len;
1120 sk->sk_wmem_queued -= len;
1121 sk_mem_uncharge(sk, len);
1122 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1124 /* Any change of skb->len requires recalculation of tso
1127 if (tcp_skb_pcount(skb) > 1)
1128 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1133 /* Calculate MSS. Not accounting for SACKs here. */
1134 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1136 struct tcp_sock *tp = tcp_sk(sk);
1137 struct inet_connection_sock *icsk = inet_csk(sk);
1140 /* Calculate base mss without TCP options:
1141 It is MMS_S - sizeof(tcphdr) of rfc1122
1143 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1145 /* Clamp it (mss_clamp does not include tcp options) */
1146 if (mss_now > tp->rx_opt.mss_clamp)
1147 mss_now = tp->rx_opt.mss_clamp;
1149 /* Now subtract optional transport overhead */
1150 mss_now -= icsk->icsk_ext_hdr_len;
1152 /* Then reserve room for full set of TCP options and 8 bytes of data */
1156 /* Now subtract TCP options size, not including SACKs */
1157 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1162 /* Inverse of above */
1163 int tcp_mss_to_mtu(struct sock *sk, int mss)
1165 struct tcp_sock *tp = tcp_sk(sk);
1166 struct inet_connection_sock *icsk = inet_csk(sk);
1170 tp->tcp_header_len +
1171 icsk->icsk_ext_hdr_len +
1172 icsk->icsk_af_ops->net_header_len;
1177 /* MTU probing init per socket */
1178 void tcp_mtup_init(struct sock *sk)
1180 struct tcp_sock *tp = tcp_sk(sk);
1181 struct inet_connection_sock *icsk = inet_csk(sk);
1183 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1184 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1185 icsk->icsk_af_ops->net_header_len;
1186 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1187 icsk->icsk_mtup.probe_size = 0;
1189 EXPORT_SYMBOL(tcp_mtup_init);
1191 /* This function synchronize snd mss to current pmtu/exthdr set.
1193 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1194 for TCP options, but includes only bare TCP header.
1196 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1197 It is minimum of user_mss and mss received with SYN.
1198 It also does not include TCP options.
1200 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1202 tp->mss_cache is current effective sending mss, including
1203 all tcp options except for SACKs. It is evaluated,
1204 taking into account current pmtu, but never exceeds
1205 tp->rx_opt.mss_clamp.
1207 NOTE1. rfc1122 clearly states that advertised MSS
1208 DOES NOT include either tcp or ip options.
1210 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1211 are READ ONLY outside this function. --ANK (980731)
1213 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1215 struct tcp_sock *tp = tcp_sk(sk);
1216 struct inet_connection_sock *icsk = inet_csk(sk);
1219 if (icsk->icsk_mtup.search_high > pmtu)
1220 icsk->icsk_mtup.search_high = pmtu;
1222 mss_now = tcp_mtu_to_mss(sk, pmtu);
1223 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1225 /* And store cached results */
1226 icsk->icsk_pmtu_cookie = pmtu;
1227 if (icsk->icsk_mtup.enabled)
1228 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1229 tp->mss_cache = mss_now;
1233 EXPORT_SYMBOL(tcp_sync_mss);
1235 /* Compute the current effective MSS, taking SACKs and IP options,
1236 * and even PMTU discovery events into account.
1238 unsigned int tcp_current_mss(struct sock *sk)
1240 struct tcp_sock *tp = tcp_sk(sk);
1241 struct dst_entry *dst = __sk_dst_get(sk);
1243 unsigned header_len;
1244 struct tcp_out_options opts;
1245 struct tcp_md5sig_key *md5;
1247 mss_now = tp->mss_cache;
1250 u32 mtu = dst_mtu(dst);
1251 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1252 mss_now = tcp_sync_mss(sk, mtu);
1255 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1256 sizeof(struct tcphdr);
1257 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1258 * some common options. If this is an odd packet (because we have SACK
1259 * blocks etc) then our calculated header_len will be different, and
1260 * we have to adjust mss_now correspondingly */
1261 if (header_len != tp->tcp_header_len) {
1262 int delta = (int) header_len - tp->tcp_header_len;
1269 /* Congestion window validation. (RFC2861) */
1270 static void tcp_cwnd_validate(struct sock *sk)
1272 struct tcp_sock *tp = tcp_sk(sk);
1274 if (tp->packets_out >= tp->snd_cwnd) {
1275 /* Network is feed fully. */
1276 tp->snd_cwnd_used = 0;
1277 tp->snd_cwnd_stamp = tcp_time_stamp;
1279 /* Network starves. */
1280 if (tp->packets_out > tp->snd_cwnd_used)
1281 tp->snd_cwnd_used = tp->packets_out;
1283 if (sysctl_tcp_slow_start_after_idle &&
1284 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1285 tcp_cwnd_application_limited(sk);
1289 /* Returns the portion of skb which can be sent right away without
1290 * introducing MSS oddities to segment boundaries. In rare cases where
1291 * mss_now != mss_cache, we will request caller to create a small skb
1292 * per input skb which could be mostly avoided here (if desired).
1294 * We explicitly want to create a request for splitting write queue tail
1295 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1296 * thus all the complexity (cwnd_len is always MSS multiple which we
1297 * return whenever allowed by the other factors). Basically we need the
1298 * modulo only when the receiver window alone is the limiting factor or
1299 * when we would be allowed to send the split-due-to-Nagle skb fully.
1301 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1302 unsigned int mss_now, unsigned int cwnd)
1304 struct tcp_sock *tp = tcp_sk(sk);
1305 u32 needed, window, cwnd_len;
1307 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1308 cwnd_len = mss_now * cwnd;
1310 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1313 needed = min(skb->len, window);
1315 if (cwnd_len <= needed)
1318 return needed - needed % mss_now;
1321 /* Can at least one segment of SKB be sent right now, according to the
1322 * congestion window rules? If so, return how many segments are allowed.
1324 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1325 struct sk_buff *skb)
1327 u32 in_flight, cwnd;
1329 /* Don't be strict about the congestion window for the final FIN. */
1330 if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
1333 in_flight = tcp_packets_in_flight(tp);
1334 cwnd = tp->snd_cwnd;
1335 if (in_flight < cwnd)
1336 return (cwnd - in_flight);
1341 /* Intialize TSO state of a skb.
1342 * This must be invoked the first time we consider transmitting
1343 * SKB onto the wire.
1345 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1346 unsigned int mss_now)
1348 int tso_segs = tcp_skb_pcount(skb);
1350 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1351 tcp_set_skb_tso_segs(sk, skb, mss_now);
1352 tso_segs = tcp_skb_pcount(skb);
1357 /* Minshall's variant of the Nagle send check. */
1358 static inline int tcp_minshall_check(const struct tcp_sock *tp)
1360 return after(tp->snd_sml, tp->snd_una) &&
1361 !after(tp->snd_sml, tp->snd_nxt);
1364 /* Return 0, if packet can be sent now without violation Nagle's rules:
1365 * 1. It is full sized.
1366 * 2. Or it contains FIN. (already checked by caller)
1367 * 3. Or TCP_NODELAY was set.
1368 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1369 * With Minshall's modification: all sent small packets are ACKed.
1371 static inline int tcp_nagle_check(const struct tcp_sock *tp,
1372 const struct sk_buff *skb,
1373 unsigned mss_now, int nonagle)
1375 return skb->len < mss_now &&
1376 ((nonagle & TCP_NAGLE_CORK) ||
1377 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1380 /* Return non-zero if the Nagle test allows this packet to be
1383 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1384 unsigned int cur_mss, int nonagle)
1386 /* Nagle rule does not apply to frames, which sit in the middle of the
1387 * write_queue (they have no chances to get new data).
1389 * This is implemented in the callers, where they modify the 'nonagle'
1390 * argument based upon the location of SKB in the send queue.
1392 if (nonagle & TCP_NAGLE_PUSH)
1395 /* Don't use the nagle rule for urgent data (or for the final FIN).
1396 * Nagle can be ignored during F-RTO too (see RFC4138).
1398 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1399 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
1402 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1408 /* Does at least the first segment of SKB fit into the send window? */
1409 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1410 unsigned int cur_mss)
1412 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1414 if (skb->len > cur_mss)
1415 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1417 return !after(end_seq, tcp_wnd_end(tp));
1420 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1421 * should be put on the wire right now. If so, it returns the number of
1422 * packets allowed by the congestion window.
1424 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1425 unsigned int cur_mss, int nonagle)
1427 struct tcp_sock *tp = tcp_sk(sk);
1428 unsigned int cwnd_quota;
1430 tcp_init_tso_segs(sk, skb, cur_mss);
1432 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1435 cwnd_quota = tcp_cwnd_test(tp, skb);
1436 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1442 /* Test if sending is allowed right now. */
1443 int tcp_may_send_now(struct sock *sk)
1445 struct tcp_sock *tp = tcp_sk(sk);
1446 struct sk_buff *skb = tcp_send_head(sk);
1449 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1450 (tcp_skb_is_last(sk, skb) ?
1451 tp->nonagle : TCP_NAGLE_PUSH));
1454 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1455 * which is put after SKB on the list. It is very much like
1456 * tcp_fragment() except that it may make several kinds of assumptions
1457 * in order to speed up the splitting operation. In particular, we
1458 * know that all the data is in scatter-gather pages, and that the
1459 * packet has never been sent out before (and thus is not cloned).
1461 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1462 unsigned int mss_now, gfp_t gfp)
1464 struct sk_buff *buff;
1465 int nlen = skb->len - len;
1468 /* All of a TSO frame must be composed of paged data. */
1469 if (skb->len != skb->data_len)
1470 return tcp_fragment(sk, skb, len, mss_now);
1472 buff = sk_stream_alloc_skb(sk, 0, gfp);
1473 if (unlikely(buff == NULL))
1476 sk->sk_wmem_queued += buff->truesize;
1477 sk_mem_charge(sk, buff->truesize);
1478 buff->truesize += nlen;
1479 skb->truesize -= nlen;
1481 /* Correct the sequence numbers. */
1482 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1483 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1484 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1486 /* PSH and FIN should only be set in the second packet. */
1487 flags = TCP_SKB_CB(skb)->flags;
1488 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1489 TCP_SKB_CB(buff)->flags = flags;
1491 /* This packet was never sent out yet, so no SACK bits. */
1492 TCP_SKB_CB(buff)->sacked = 0;
1494 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1495 skb_split(skb, buff, len);
1497 /* Fix up tso_factor for both original and new SKB. */
1498 tcp_set_skb_tso_segs(sk, skb, mss_now);
1499 tcp_set_skb_tso_segs(sk, buff, mss_now);
1501 /* Link BUFF into the send queue. */
1502 skb_header_release(buff);
1503 tcp_insert_write_queue_after(skb, buff, sk);
1508 /* Try to defer sending, if possible, in order to minimize the amount
1509 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1511 * This algorithm is from John Heffner.
1513 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1515 struct tcp_sock *tp = tcp_sk(sk);
1516 const struct inet_connection_sock *icsk = inet_csk(sk);
1517 u32 send_win, cong_win, limit, in_flight;
1520 if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
1523 if (icsk->icsk_ca_state != TCP_CA_Open)
1526 /* Defer for less than two clock ticks. */
1527 if (tp->tso_deferred &&
1528 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1531 in_flight = tcp_packets_in_flight(tp);
1533 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1535 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1537 /* From in_flight test above, we know that cwnd > in_flight. */
1538 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1540 limit = min(send_win, cong_win);
1542 /* If a full-sized TSO skb can be sent, do it. */
1543 if (limit >= sk->sk_gso_max_size)
1546 /* Middle in queue won't get any more data, full sendable already? */
1547 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1550 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1552 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1554 /* If at least some fraction of a window is available,
1557 chunk /= win_divisor;
1561 /* Different approach, try not to defer past a single
1562 * ACK. Receiver should ACK every other full sized
1563 * frame, so if we have space for more than 3 frames
1566 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1570 /* Ok, it looks like it is advisable to defer. */
1571 tp->tso_deferred = 1 | (jiffies << 1);
1576 tp->tso_deferred = 0;
1580 /* Create a new MTU probe if we are ready.
1581 * MTU probe is regularly attempting to increase the path MTU by
1582 * deliberately sending larger packets. This discovers routing
1583 * changes resulting in larger path MTUs.
1585 * Returns 0 if we should wait to probe (no cwnd available),
1586 * 1 if a probe was sent,
1589 static int tcp_mtu_probe(struct sock *sk)
1591 struct tcp_sock *tp = tcp_sk(sk);
1592 struct inet_connection_sock *icsk = inet_csk(sk);
1593 struct sk_buff *skb, *nskb, *next;
1600 /* Not currently probing/verifying,
1602 * have enough cwnd, and
1603 * not SACKing (the variable headers throw things off) */
1604 if (!icsk->icsk_mtup.enabled ||
1605 icsk->icsk_mtup.probe_size ||
1606 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1607 tp->snd_cwnd < 11 ||
1608 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1611 /* Very simple search strategy: just double the MSS. */
1612 mss_now = tcp_current_mss(sk);
1613 probe_size = 2 * tp->mss_cache;
1614 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1615 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1616 /* TODO: set timer for probe_converge_event */
1620 /* Have enough data in the send queue to probe? */
1621 if (tp->write_seq - tp->snd_nxt < size_needed)
1624 if (tp->snd_wnd < size_needed)
1626 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1629 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1630 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1631 if (!tcp_packets_in_flight(tp))
1637 /* We're allowed to probe. Build it now. */
1638 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1640 sk->sk_wmem_queued += nskb->truesize;
1641 sk_mem_charge(sk, nskb->truesize);
1643 skb = tcp_send_head(sk);
1645 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1646 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1647 TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
1648 TCP_SKB_CB(nskb)->sacked = 0;
1650 nskb->ip_summed = skb->ip_summed;
1652 tcp_insert_write_queue_before(nskb, skb, sk);
1655 tcp_for_write_queue_from_safe(skb, next, sk) {
1656 copy = min_t(int, skb->len, probe_size - len);
1657 if (nskb->ip_summed)
1658 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1660 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1661 skb_put(nskb, copy),
1664 if (skb->len <= copy) {
1665 /* We've eaten all the data from this skb.
1667 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1668 tcp_unlink_write_queue(skb, sk);
1669 sk_wmem_free_skb(sk, skb);
1671 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1672 ~(TCPHDR_FIN|TCPHDR_PSH);
1673 if (!skb_shinfo(skb)->nr_frags) {
1674 skb_pull(skb, copy);
1675 if (skb->ip_summed != CHECKSUM_PARTIAL)
1676 skb->csum = csum_partial(skb->data,
1679 __pskb_trim_head(skb, copy);
1680 tcp_set_skb_tso_segs(sk, skb, mss_now);
1682 TCP_SKB_CB(skb)->seq += copy;
1687 if (len >= probe_size)
1690 tcp_init_tso_segs(sk, nskb, nskb->len);
1692 /* We're ready to send. If this fails, the probe will
1693 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1694 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1695 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1696 /* Decrement cwnd here because we are sending
1697 * effectively two packets. */
1699 tcp_event_new_data_sent(sk, nskb);
1701 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1702 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1703 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1711 /* This routine writes packets to the network. It advances the
1712 * send_head. This happens as incoming acks open up the remote
1715 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1716 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1717 * account rare use of URG, this is not a big flaw.
1719 * Returns 1, if no segments are in flight and we have queued segments, but
1720 * cannot send anything now because of SWS or another problem.
1722 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1723 int push_one, gfp_t gfp)
1725 struct tcp_sock *tp = tcp_sk(sk);
1726 struct sk_buff *skb;
1727 unsigned int tso_segs, sent_pkts;
1734 /* Do MTU probing. */
1735 result = tcp_mtu_probe(sk);
1738 } else if (result > 0) {
1743 while ((skb = tcp_send_head(sk))) {
1746 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1749 cwnd_quota = tcp_cwnd_test(tp, skb);
1753 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1756 if (tso_segs == 1) {
1757 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1758 (tcp_skb_is_last(sk, skb) ?
1759 nonagle : TCP_NAGLE_PUSH))))
1762 if (!push_one && tcp_tso_should_defer(sk, skb))
1767 if (tso_segs > 1 && !tcp_urg_mode(tp))
1768 limit = tcp_mss_split_point(sk, skb, mss_now,
1771 if (skb->len > limit &&
1772 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1775 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1777 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1780 /* Advance the send_head. This one is sent out.
1781 * This call will increment packets_out.
1783 tcp_event_new_data_sent(sk, skb);
1785 tcp_minshall_update(tp, mss_now, skb);
1792 if (likely(sent_pkts)) {
1793 tcp_cwnd_validate(sk);
1796 return !tp->packets_out && tcp_send_head(sk);
1799 /* Push out any pending frames which were held back due to
1800 * TCP_CORK or attempt at coalescing tiny packets.
1801 * The socket must be locked by the caller.
1803 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1806 /* If we are closed, the bytes will have to remain here.
1807 * In time closedown will finish, we empty the write queue and
1808 * all will be happy.
1810 if (unlikely(sk->sk_state == TCP_CLOSE))
1813 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1814 tcp_check_probe_timer(sk);
1817 /* Send _single_ skb sitting at the send head. This function requires
1818 * true push pending frames to setup probe timer etc.
1820 void tcp_push_one(struct sock *sk, unsigned int mss_now)
1822 struct sk_buff *skb = tcp_send_head(sk);
1824 BUG_ON(!skb || skb->len < mss_now);
1826 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1829 /* This function returns the amount that we can raise the
1830 * usable window based on the following constraints
1832 * 1. The window can never be shrunk once it is offered (RFC 793)
1833 * 2. We limit memory per socket
1836 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1837 * RECV.NEXT + RCV.WIN fixed until:
1838 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1840 * i.e. don't raise the right edge of the window until you can raise
1841 * it at least MSS bytes.
1843 * Unfortunately, the recommended algorithm breaks header prediction,
1844 * since header prediction assumes th->window stays fixed.
1846 * Strictly speaking, keeping th->window fixed violates the receiver
1847 * side SWS prevention criteria. The problem is that under this rule
1848 * a stream of single byte packets will cause the right side of the
1849 * window to always advance by a single byte.
1851 * Of course, if the sender implements sender side SWS prevention
1852 * then this will not be a problem.
1854 * BSD seems to make the following compromise:
1856 * If the free space is less than the 1/4 of the maximum
1857 * space available and the free space is less than 1/2 mss,
1858 * then set the window to 0.
1859 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1860 * Otherwise, just prevent the window from shrinking
1861 * and from being larger than the largest representable value.
1863 * This prevents incremental opening of the window in the regime
1864 * where TCP is limited by the speed of the reader side taking
1865 * data out of the TCP receive queue. It does nothing about
1866 * those cases where the window is constrained on the sender side
1867 * because the pipeline is full.
1869 * BSD also seems to "accidentally" limit itself to windows that are a
1870 * multiple of MSS, at least until the free space gets quite small.
1871 * This would appear to be a side effect of the mbuf implementation.
1872 * Combining these two algorithms results in the observed behavior
1873 * of having a fixed window size at almost all times.
1875 * Below we obtain similar behavior by forcing the offered window to
1876 * a multiple of the mss when it is feasible to do so.
1878 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1879 * Regular options like TIMESTAMP are taken into account.
1881 u32 __tcp_select_window(struct sock *sk)
1883 struct inet_connection_sock *icsk = inet_csk(sk);
1884 struct tcp_sock *tp = tcp_sk(sk);
1885 /* MSS for the peer's data. Previous versions used mss_clamp
1886 * here. I don't know if the value based on our guesses
1887 * of peer's MSS is better for the performance. It's more correct
1888 * but may be worse for the performance because of rcv_mss
1889 * fluctuations. --SAW 1998/11/1
1891 int mss = icsk->icsk_ack.rcv_mss;
1892 int free_space = tcp_space(sk);
1893 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1896 if (mss > full_space)
1899 if (free_space < (full_space >> 1)) {
1900 icsk->icsk_ack.quick = 0;
1902 if (tcp_memory_pressure)
1903 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1906 if (free_space < mss)
1910 if (free_space > tp->rcv_ssthresh)
1911 free_space = tp->rcv_ssthresh;
1913 /* Don't do rounding if we are using window scaling, since the
1914 * scaled window will not line up with the MSS boundary anyway.
1916 window = tp->rcv_wnd;
1917 if (tp->rx_opt.rcv_wscale) {
1918 window = free_space;
1920 /* Advertise enough space so that it won't get scaled away.
1921 * Import case: prevent zero window announcement if
1922 * 1<<rcv_wscale > mss.
1924 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1925 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1926 << tp->rx_opt.rcv_wscale);
1928 /* Get the largest window that is a nice multiple of mss.
1929 * Window clamp already applied above.
1930 * If our current window offering is within 1 mss of the
1931 * free space we just keep it. This prevents the divide
1932 * and multiply from happening most of the time.
1933 * We also don't do any window rounding when the free space
1936 if (window <= free_space - mss || window > free_space)
1937 window = (free_space / mss) * mss;
1938 else if (mss == full_space &&
1939 free_space > window + (full_space >> 1))
1940 window = free_space;
1946 /* Collapses two adjacent SKB's during retransmission. */
1947 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1949 struct tcp_sock *tp = tcp_sk(sk);
1950 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1951 int skb_size, next_skb_size;
1953 skb_size = skb->len;
1954 next_skb_size = next_skb->len;
1956 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1958 tcp_highest_sack_combine(sk, next_skb, skb);
1960 tcp_unlink_write_queue(next_skb, sk);
1962 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1965 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1966 skb->ip_summed = CHECKSUM_PARTIAL;
1968 if (skb->ip_summed != CHECKSUM_PARTIAL)
1969 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1971 /* Update sequence range on original skb. */
1972 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1974 /* Merge over control information. This moves PSH/FIN etc. over */
1975 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1977 /* All done, get rid of second SKB and account for it so
1978 * packet counting does not break.
1980 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1982 /* changed transmit queue under us so clear hints */
1983 tcp_clear_retrans_hints_partial(tp);
1984 if (next_skb == tp->retransmit_skb_hint)
1985 tp->retransmit_skb_hint = skb;
1987 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
1989 sk_wmem_free_skb(sk, next_skb);
1992 /* Check if coalescing SKBs is legal. */
1993 static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
1995 if (tcp_skb_pcount(skb) > 1)
1997 /* TODO: SACK collapsing could be used to remove this condition */
1998 if (skb_shinfo(skb)->nr_frags != 0)
2000 if (skb_cloned(skb))
2002 if (skb == tcp_send_head(sk))
2004 /* Some heurestics for collapsing over SACK'd could be invented */
2005 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2011 /* Collapse packets in the retransmit queue to make to create
2012 * less packets on the wire. This is only done on retransmission.
2014 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 struct sk_buff *skb = to, *tmp;
2021 if (!sysctl_tcp_retrans_collapse)
2023 if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
2026 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2027 if (!tcp_can_collapse(sk, skb))
2039 /* Punt if not enough space exists in the first SKB for
2040 * the data in the second
2042 if (skb->len > skb_tailroom(to))
2045 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2048 tcp_collapse_retrans(sk, to);
2052 /* This retransmits one SKB. Policy decisions and retransmit queue
2053 * state updates are done by the caller. Returns non-zero if an
2054 * error occurred which prevented the send.
2056 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2058 struct tcp_sock *tp = tcp_sk(sk);
2059 struct inet_connection_sock *icsk = inet_csk(sk);
2060 unsigned int cur_mss;
2063 /* Inconslusive MTU probe */
2064 if (icsk->icsk_mtup.probe_size) {
2065 icsk->icsk_mtup.probe_size = 0;
2068 /* Do not sent more than we queued. 1/4 is reserved for possible
2069 * copying overhead: fragmentation, tunneling, mangling etc.
2071 if (atomic_read(&sk->sk_wmem_alloc) >
2072 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2075 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2076 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2078 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2082 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2083 return -EHOSTUNREACH; /* Routing failure or similar. */
2085 cur_mss = tcp_current_mss(sk);
2087 /* If receiver has shrunk his window, and skb is out of
2088 * new window, do not retransmit it. The exception is the
2089 * case, when window is shrunk to zero. In this case
2090 * our retransmit serves as a zero window probe.
2092 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2093 TCP_SKB_CB(skb)->seq != tp->snd_una)
2096 if (skb->len > cur_mss) {
2097 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2098 return -ENOMEM; /* We'll try again later. */
2100 int oldpcount = tcp_skb_pcount(skb);
2102 if (unlikely(oldpcount > 1)) {
2103 tcp_init_tso_segs(sk, skb, cur_mss);
2104 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2108 tcp_retrans_try_collapse(sk, skb, cur_mss);
2110 /* Some Solaris stacks overoptimize and ignore the FIN on a
2111 * retransmit when old data is attached. So strip it off
2112 * since it is cheap to do so and saves bytes on the network.
2115 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
2116 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2117 if (!pskb_trim(skb, 0)) {
2118 /* Reuse, even though it does some unnecessary work */
2119 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2120 TCP_SKB_CB(skb)->flags);
2121 skb->ip_summed = CHECKSUM_NONE;
2125 /* Make a copy, if the first transmission SKB clone we made
2126 * is still in somebody's hands, else make a clone.
2128 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2130 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2133 /* Update global TCP statistics. */
2134 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2136 tp->total_retrans++;
2138 #if FASTRETRANS_DEBUG > 0
2139 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2140 if (net_ratelimit())
2141 printk(KERN_DEBUG "retrans_out leaked.\n");
2144 if (!tp->retrans_out)
2145 tp->lost_retrans_low = tp->snd_nxt;
2146 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2147 tp->retrans_out += tcp_skb_pcount(skb);
2149 /* Save stamp of the first retransmit. */
2150 if (!tp->retrans_stamp)
2151 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2155 /* snd_nxt is stored to detect loss of retransmitted segment,
2156 * see tcp_input.c tcp_sacktag_write_queue().
2158 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2163 /* Check if we forward retransmits are possible in the current
2164 * window/congestion state.
2166 static int tcp_can_forward_retransmit(struct sock *sk)
2168 const struct inet_connection_sock *icsk = inet_csk(sk);
2169 struct tcp_sock *tp = tcp_sk(sk);
2171 /* Forward retransmissions are possible only during Recovery. */
2172 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2175 /* No forward retransmissions in Reno are possible. */
2176 if (tcp_is_reno(tp))
2179 /* Yeah, we have to make difficult choice between forward transmission
2180 * and retransmission... Both ways have their merits...
2182 * For now we do not retransmit anything, while we have some new
2183 * segments to send. In the other cases, follow rule 3 for
2184 * NextSeg() specified in RFC3517.
2187 if (tcp_may_send_now(sk))
2193 /* This gets called after a retransmit timeout, and the initially
2194 * retransmitted data is acknowledged. It tries to continue
2195 * resending the rest of the retransmit queue, until either
2196 * we've sent it all or the congestion window limit is reached.
2197 * If doing SACK, the first ACK which comes back for a timeout
2198 * based retransmit packet might feed us FACK information again.
2199 * If so, we use it to avoid unnecessarily retransmissions.
2201 void tcp_xmit_retransmit_queue(struct sock *sk)
2203 const struct inet_connection_sock *icsk = inet_csk(sk);
2204 struct tcp_sock *tp = tcp_sk(sk);
2205 struct sk_buff *skb;
2206 struct sk_buff *hole = NULL;
2209 int fwd_rexmitting = 0;
2211 if (!tp->packets_out)
2215 tp->retransmit_high = tp->snd_una;
2217 if (tp->retransmit_skb_hint) {
2218 skb = tp->retransmit_skb_hint;
2219 last_lost = TCP_SKB_CB(skb)->end_seq;
2220 if (after(last_lost, tp->retransmit_high))
2221 last_lost = tp->retransmit_high;
2223 skb = tcp_write_queue_head(sk);
2224 last_lost = tp->snd_una;
2227 tcp_for_write_queue_from(skb, sk) {
2228 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2230 if (skb == tcp_send_head(sk))
2232 /* we could do better than to assign each time */
2234 tp->retransmit_skb_hint = skb;
2236 /* Assume this retransmit will generate
2237 * only one packet for congestion window
2238 * calculation purposes. This works because
2239 * tcp_retransmit_skb() will chop up the
2240 * packet to be MSS sized and all the
2241 * packet counting works out.
2243 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2246 if (fwd_rexmitting) {
2248 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2250 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2252 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2253 tp->retransmit_high = last_lost;
2254 if (!tcp_can_forward_retransmit(sk))
2256 /* Backtrack if necessary to non-L'ed skb */
2264 } else if (!(sacked & TCPCB_LOST)) {
2265 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2270 last_lost = TCP_SKB_CB(skb)->end_seq;
2271 if (icsk->icsk_ca_state != TCP_CA_Loss)
2272 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2274 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2277 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2280 if (tcp_retransmit_skb(sk, skb))
2282 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2284 if (skb == tcp_write_queue_head(sk))
2285 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2286 inet_csk(sk)->icsk_rto,
2291 /* Send a fin. The caller locks the socket for us. This cannot be
2292 * allowed to fail queueing a FIN frame under any circumstances.
2294 void tcp_send_fin(struct sock *sk)
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb = tcp_write_queue_tail(sk);
2300 /* Optimization, tack on the FIN if we have a queue of
2301 * unsent frames. But be careful about outgoing SACKS
2304 mss_now = tcp_current_mss(sk);
2306 if (tcp_send_head(sk) != NULL) {
2307 TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
2308 TCP_SKB_CB(skb)->end_seq++;
2311 /* Socket is locked, keep trying until memory is available. */
2313 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2320 /* Reserve space for headers and prepare control bits. */
2321 skb_reserve(skb, MAX_TCP_HEADER);
2322 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2323 tcp_init_nondata_skb(skb, tp->write_seq,
2324 TCPHDR_ACK | TCPHDR_FIN);
2325 tcp_queue_skb(sk, skb);
2327 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2330 /* We get here when a process closes a file descriptor (either due to
2331 * an explicit close() or as a byproduct of exit()'ing) and there
2332 * was unread data in the receive queue. This behavior is recommended
2333 * by RFC 2525, section 2.17. -DaveM
2335 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2337 struct sk_buff *skb;
2339 /* NOTE: No TCP options attached and we never retransmit this. */
2340 skb = alloc_skb(MAX_TCP_HEADER, priority);
2342 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2346 /* Reserve space for headers and prepare control bits. */
2347 skb_reserve(skb, MAX_TCP_HEADER);
2348 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2349 TCPHDR_ACK | TCPHDR_RST);
2351 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2352 if (tcp_transmit_skb(sk, skb, 0, priority))
2353 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2355 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2358 /* Send a crossed SYN-ACK during socket establishment.
2359 * WARNING: This routine must only be called when we have already sent
2360 * a SYN packet that crossed the incoming SYN that caused this routine
2361 * to get called. If this assumption fails then the initial rcv_wnd
2362 * and rcv_wscale values will not be correct.
2364 int tcp_send_synack(struct sock *sk)
2366 struct sk_buff *skb;
2368 skb = tcp_write_queue_head(sk);
2369 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
2370 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2373 if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
2374 if (skb_cloned(skb)) {
2375 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2378 tcp_unlink_write_queue(skb, sk);
2379 skb_header_release(nskb);
2380 __tcp_add_write_queue_head(sk, nskb);
2381 sk_wmem_free_skb(sk, skb);
2382 sk->sk_wmem_queued += nskb->truesize;
2383 sk_mem_charge(sk, nskb->truesize);
2387 TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
2388 TCP_ECN_send_synack(tcp_sk(sk), skb);
2390 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2391 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2394 /* Prepare a SYN-ACK. */
2395 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2396 struct request_sock *req,
2397 struct request_values *rvp)
2399 struct tcp_out_options opts;
2400 struct tcp_extend_values *xvp = tcp_xv(rvp);
2401 struct inet_request_sock *ireq = inet_rsk(req);
2402 struct tcp_sock *tp = tcp_sk(sk);
2403 const struct tcp_cookie_values *cvp = tp->cookie_values;
2405 struct sk_buff *skb;
2406 struct tcp_md5sig_key *md5;
2407 int tcp_header_size;
2409 int s_data_desired = 0;
2411 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2412 s_data_desired = cvp->s_data_desired;
2413 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2417 /* Reserve space for headers. */
2418 skb_reserve(skb, MAX_TCP_HEADER);
2420 skb_dst_set(skb, dst_clone(dst));
2422 mss = dst_metric(dst, RTAX_ADVMSS);
2423 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2424 mss = tp->rx_opt.user_mss;
2426 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2428 /* Set this up on the first call only */
2429 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2431 /* limit the window selection if the user enforce a smaller rx buffer */
2432 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2433 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2434 req->window_clamp = tcp_full_space(sk);
2436 /* tcp_full_space because it is guaranteed to be the first packet */
2437 tcp_select_initial_window(tcp_full_space(sk),
2438 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2443 dst_metric(dst, RTAX_INITRWND));
2444 ireq->rcv_wscale = rcv_wscale;
2447 memset(&opts, 0, sizeof(opts));
2448 #ifdef CONFIG_SYN_COOKIES
2449 if (unlikely(req->cookie_ts))
2450 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2453 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2454 tcp_header_size = tcp_synack_options(sk, req, mss,
2455 skb, &opts, &md5, xvp)
2458 skb_push(skb, tcp_header_size);
2459 skb_reset_transport_header(skb);
2462 memset(th, 0, sizeof(struct tcphdr));
2465 TCP_ECN_make_synack(req, th);
2466 th->source = ireq->loc_port;
2467 th->dest = ireq->rmt_port;
2468 /* Setting of flags are superfluous here for callers (and ECE is
2469 * not even correctly set)
2471 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2472 TCPHDR_SYN | TCPHDR_ACK);
2474 if (OPTION_COOKIE_EXTENSION & opts.options) {
2475 if (s_data_desired) {
2476 u8 *buf = skb_put(skb, s_data_desired);
2478 /* copy data directly from the listening socket. */
2479 memcpy(buf, cvp->s_data_payload, s_data_desired);
2480 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2483 if (opts.hash_size > 0) {
2484 __u32 workspace[SHA_WORKSPACE_WORDS];
2485 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2486 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2488 /* Secret recipe depends on the Timestamp, (future)
2489 * Sequence and Acknowledgment Numbers, Initiator
2490 * Cookie, and others handled by IP variant caller.
2492 *tail-- ^= opts.tsval;
2493 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2494 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2497 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2498 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2500 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2503 opts.hash_location =
2504 (__u8 *)&xvp->cookie_bakery[0];
2508 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2509 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2511 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2512 th->window = htons(min(req->rcv_wnd, 65535U));
2513 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2514 th->doff = (tcp_header_size >> 2);
2515 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2517 #ifdef CONFIG_TCP_MD5SIG
2518 /* Okay, we have all we need - do the md5 hash if needed */
2520 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2521 md5, NULL, req, skb);
2527 EXPORT_SYMBOL(tcp_make_synack);
2529 /* Do all connect socket setups that can be done AF independent. */
2530 static void tcp_connect_init(struct sock *sk)
2532 struct dst_entry *dst = __sk_dst_get(sk);
2533 struct tcp_sock *tp = tcp_sk(sk);
2536 /* We'll fix this up when we get a response from the other end.
2537 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2539 tp->tcp_header_len = sizeof(struct tcphdr) +
2540 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2542 #ifdef CONFIG_TCP_MD5SIG
2543 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2544 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2547 /* If user gave his TCP_MAXSEG, record it to clamp */
2548 if (tp->rx_opt.user_mss)
2549 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2552 tcp_sync_mss(sk, dst_mtu(dst));
2554 if (!tp->window_clamp)
2555 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2556 tp->advmss = dst_metric(dst, RTAX_ADVMSS);
2557 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2558 tp->advmss = tp->rx_opt.user_mss;
2560 tcp_initialize_rcv_mss(sk);
2562 /* limit the window selection if the user enforce a smaller rx buffer */
2563 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2564 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2565 tp->window_clamp = tcp_full_space(sk);
2567 tcp_select_initial_window(tcp_full_space(sk),
2568 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2571 sysctl_tcp_window_scaling,
2573 dst_metric(dst, RTAX_INITRWND));
2575 tp->rx_opt.rcv_wscale = rcv_wscale;
2576 tp->rcv_ssthresh = tp->rcv_wnd;
2579 sock_reset_flag(sk, SOCK_DONE);
2582 tp->snd_una = tp->write_seq;
2583 tp->snd_sml = tp->write_seq;
2584 tp->snd_up = tp->write_seq;
2589 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2590 inet_csk(sk)->icsk_retransmits = 0;
2591 tcp_clear_retrans(tp);
2594 /* Build a SYN and send it off. */
2595 int tcp_connect(struct sock *sk)
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 struct sk_buff *buff;
2600 tcp_connect_init(sk);
2602 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2603 if (unlikely(buff == NULL))
2606 /* Reserve space for headers. */
2607 skb_reserve(buff, MAX_TCP_HEADER);
2609 tp->snd_nxt = tp->write_seq;
2610 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2611 TCP_ECN_send_syn(sk, buff);
2614 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2615 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2616 skb_header_release(buff);
2617 __tcp_add_write_queue_tail(sk, buff);
2618 sk->sk_wmem_queued += buff->truesize;
2619 sk_mem_charge(sk, buff->truesize);
2620 tp->packets_out += tcp_skb_pcount(buff);
2621 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2623 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2624 * in order to make this packet get counted in tcpOutSegs.
2626 tp->snd_nxt = tp->write_seq;
2627 tp->pushed_seq = tp->write_seq;
2628 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2630 /* Timer for repeating the SYN until an answer. */
2631 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2632 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2635 EXPORT_SYMBOL(tcp_connect);
2637 /* Send out a delayed ack, the caller does the policy checking
2638 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2641 void tcp_send_delayed_ack(struct sock *sk)
2643 struct inet_connection_sock *icsk = inet_csk(sk);
2644 int ato = icsk->icsk_ack.ato;
2645 unsigned long timeout;
2647 if (ato > TCP_DELACK_MIN) {
2648 const struct tcp_sock *tp = tcp_sk(sk);
2649 int max_ato = HZ / 2;
2651 if (icsk->icsk_ack.pingpong ||
2652 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2653 max_ato = TCP_DELACK_MAX;
2655 /* Slow path, intersegment interval is "high". */
2657 /* If some rtt estimate is known, use it to bound delayed ack.
2658 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2662 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2668 ato = min(ato, max_ato);
2671 /* Stay within the limit we were given */
2672 timeout = jiffies + ato;
2674 /* Use new timeout only if there wasn't a older one earlier. */
2675 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2676 /* If delack timer was blocked or is about to expire,
2679 if (icsk->icsk_ack.blocked ||
2680 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2685 if (!time_before(timeout, icsk->icsk_ack.timeout))
2686 timeout = icsk->icsk_ack.timeout;
2688 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2689 icsk->icsk_ack.timeout = timeout;
2690 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2693 /* This routine sends an ack and also updates the window. */
2694 void tcp_send_ack(struct sock *sk)
2696 struct sk_buff *buff;
2698 /* If we have been reset, we may not send again. */
2699 if (sk->sk_state == TCP_CLOSE)
2702 /* We are not putting this on the write queue, so
2703 * tcp_transmit_skb() will set the ownership to this
2706 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2708 inet_csk_schedule_ack(sk);
2709 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2710 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2711 TCP_DELACK_MAX, TCP_RTO_MAX);
2715 /* Reserve space for headers and prepare control bits. */
2716 skb_reserve(buff, MAX_TCP_HEADER);
2717 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2719 /* Send it off, this clears delayed acks for us. */
2720 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2721 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2724 /* This routine sends a packet with an out of date sequence
2725 * number. It assumes the other end will try to ack it.
2727 * Question: what should we make while urgent mode?
2728 * 4.4BSD forces sending single byte of data. We cannot send
2729 * out of window data, because we have SND.NXT==SND.MAX...
2731 * Current solution: to send TWO zero-length segments in urgent mode:
2732 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2733 * out-of-date with SND.UNA-1 to probe window.
2735 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2737 struct tcp_sock *tp = tcp_sk(sk);
2738 struct sk_buff *skb;
2740 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2741 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2745 /* Reserve space for headers and set control bits. */
2746 skb_reserve(skb, MAX_TCP_HEADER);
2747 /* Use a previous sequence. This should cause the other
2748 * end to send an ack. Don't queue or clone SKB, just
2751 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2752 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2753 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2756 /* Initiate keepalive or window probe from timer. */
2757 int tcp_write_wakeup(struct sock *sk)
2759 struct tcp_sock *tp = tcp_sk(sk);
2760 struct sk_buff *skb;
2762 if (sk->sk_state == TCP_CLOSE)
2765 if ((skb = tcp_send_head(sk)) != NULL &&
2766 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2768 unsigned int mss = tcp_current_mss(sk);
2769 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2771 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2772 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2774 /* We are probing the opening of a window
2775 * but the window size is != 0
2776 * must have been a result SWS avoidance ( sender )
2778 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2780 seg_size = min(seg_size, mss);
2781 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2782 if (tcp_fragment(sk, skb, seg_size, mss))
2784 } else if (!tcp_skb_pcount(skb))
2785 tcp_set_skb_tso_segs(sk, skb, mss);
2787 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2788 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2789 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2791 tcp_event_new_data_sent(sk, skb);
2794 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2795 tcp_xmit_probe_skb(sk, 1);
2796 return tcp_xmit_probe_skb(sk, 0);
2800 /* A window probe timeout has occurred. If window is not closed send
2801 * a partial packet else a zero probe.
2803 void tcp_send_probe0(struct sock *sk)
2805 struct inet_connection_sock *icsk = inet_csk(sk);
2806 struct tcp_sock *tp = tcp_sk(sk);
2809 err = tcp_write_wakeup(sk);
2811 if (tp->packets_out || !tcp_send_head(sk)) {
2812 /* Cancel probe timer, if it is not required. */
2813 icsk->icsk_probes_out = 0;
2814 icsk->icsk_backoff = 0;
2819 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2820 icsk->icsk_backoff++;
2821 icsk->icsk_probes_out++;
2822 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2823 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2826 /* If packet was not sent due to local congestion,
2827 * do not backoff and do not remember icsk_probes_out.
2828 * Let local senders to fight for local resources.
2830 * Use accumulated backoff yet.
2832 if (!icsk->icsk_probes_out)
2833 icsk->icsk_probes_out = 1;
2834 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2835 min(icsk->icsk_rto << icsk->icsk_backoff,
2836 TCP_RESOURCE_PROBE_INTERVAL),