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
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 int sysctl_tcp_mtu_probing __read_mostly = 0;
63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
65 /* By default, RFC2861 behavior. */
66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
68 int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
69 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
71 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
72 int push_one, gfp_t gfp);
74 /* Account for new data that has been sent to the network. */
75 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
77 struct tcp_sock *tp = tcp_sk(sk);
78 unsigned int prior_packets = tp->packets_out;
80 tcp_advance_send_head(sk, skb);
81 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
83 /* Don't override Nagle indefinitely with F-RTO */
84 if (tp->frto_counter == 2)
87 tp->packets_out += tcp_skb_pcount(skb);
88 if (!prior_packets || tp->early_retrans_delayed)
92 /* SND.NXT, if window was not shrunk.
93 * If window has been shrunk, what should we make? It is not clear at all.
94 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
95 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
96 * invalid. OK, let's make this for now:
98 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
100 const struct tcp_sock *tp = tcp_sk(sk);
102 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
105 return tcp_wnd_end(tp);
108 /* Calculate mss to advertise in SYN segment.
109 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
111 * 1. It is independent of path mtu.
112 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
113 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
114 * attached devices, because some buggy hosts are confused by
116 * 4. We do not make 3, we advertise MSS, calculated from first
117 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
118 * This may be overridden via information stored in routing table.
119 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
120 * probably even Jumbo".
122 static __u16 tcp_advertise_mss(struct sock *sk)
124 struct tcp_sock *tp = tcp_sk(sk);
125 const struct dst_entry *dst = __sk_dst_get(sk);
126 int mss = tp->advmss;
129 unsigned int metric = dst_metric_advmss(dst);
140 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
141 * This is the first part of cwnd validation mechanism. */
142 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
144 struct tcp_sock *tp = tcp_sk(sk);
145 s32 delta = tcp_time_stamp - tp->lsndtime;
146 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
147 u32 cwnd = tp->snd_cwnd;
149 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
151 tp->snd_ssthresh = tcp_current_ssthresh(sk);
152 restart_cwnd = min(restart_cwnd, cwnd);
154 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
156 tp->snd_cwnd = max(cwnd, restart_cwnd);
157 tp->snd_cwnd_stamp = tcp_time_stamp;
158 tp->snd_cwnd_used = 0;
161 /* Congestion state accounting after a packet has been sent. */
162 static void tcp_event_data_sent(struct tcp_sock *tp,
165 struct inet_connection_sock *icsk = inet_csk(sk);
166 const u32 now = tcp_time_stamp;
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
178 icsk->icsk_ack.pingpong = 1;
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
184 tcp_dec_quickack_mode(sk, pkts);
185 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
188 /* Determine a window scaling and initial window to offer.
189 * Based on the assumption that the given amount of space
190 * will be offered. Store the results in the tp structure.
191 * NOTE: for smooth operation initial space offering should
192 * be a multiple of mss if possible. We assume here that mss >= 1.
193 * This MUST be enforced by all callers.
195 void tcp_select_initial_window(int __space, __u32 mss,
196 __u32 *rcv_wnd, __u32 *window_clamp,
197 int wscale_ok, __u8 *rcv_wscale,
200 unsigned int space = (__space < 0 ? 0 : __space);
202 /* If no clamp set the clamp to the max possible scaled window */
203 if (*window_clamp == 0)
204 (*window_clamp) = (65535 << 14);
205 space = min(*window_clamp, space);
207 /* Quantize space offering to a multiple of mss if possible. */
209 space = (space / mss) * mss;
211 /* NOTE: offering an initial window larger than 32767
212 * will break some buggy TCP stacks. If the admin tells us
213 * it is likely we could be speaking with such a buggy stack
214 * we will truncate our initial window offering to 32K-1
215 * unless the remote has sent us a window scaling option,
216 * which we interpret as a sign the remote TCP is not
217 * misinterpreting the window field as a signed quantity.
219 if (sysctl_tcp_workaround_signed_windows)
220 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
226 /* Set window scaling on max possible window
227 * See RFC1323 for an explanation of the limit to 14
229 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
230 space = min_t(u32, space, *window_clamp);
231 while (space > 65535 && (*rcv_wscale) < 14) {
237 /* Set initial window to a value enough for senders starting with
238 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
239 * a limit on the initial window when mss is larger than 1460.
241 if (mss > (1 << *rcv_wscale)) {
242 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
245 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
246 /* when initializing use the value from init_rcv_wnd
247 * rather than the default from above
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
252 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
255 /* Set the clamp no higher than max representable value */
256 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
258 EXPORT_SYMBOL(tcp_select_initial_window);
260 /* Chose a new window to advertise, update state in tcp_sock for the
261 * socket, and return result with RFC1323 scaling applied. The return
262 * value can be stuffed directly into th->window for an outgoing
265 static u16 tcp_select_window(struct sock *sk)
267 struct tcp_sock *tp = tcp_sk(sk);
268 u32 cur_win = tcp_receive_window(tp);
269 u32 new_win = __tcp_select_window(sk);
271 /* Never shrink the offered window */
272 if (new_win < cur_win) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
278 * Relax Will Robinson.
280 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
282 tp->rcv_wnd = new_win;
283 tp->rcv_wup = tp->rcv_nxt;
285 /* Make sure we do not exceed the maximum possible
288 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
289 new_win = min(new_win, MAX_TCP_WINDOW);
291 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
293 /* RFC1323 scaling applied */
294 new_win >>= tp->rx_opt.rcv_wscale;
296 /* If we advertise zero window, disable fast path. */
303 /* Packet ECN state for a SYN-ACK */
304 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
306 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
307 if (!(tp->ecn_flags & TCP_ECN_OK))
308 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
311 /* Packet ECN state for a SYN. */
312 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
314 struct tcp_sock *tp = tcp_sk(sk);
317 if (sysctl_tcp_ecn == 1) {
318 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
319 tp->ecn_flags = TCP_ECN_OK;
323 static __inline__ void
324 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
326 if (inet_rsk(req)->ecn_ok)
330 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
333 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
336 struct tcp_sock *tp = tcp_sk(sk);
338 if (tp->ecn_flags & TCP_ECN_OK) {
339 /* Not-retransmitted data segment: set ECT and inject CWR. */
340 if (skb->len != tcp_header_len &&
341 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
343 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
344 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
345 tcp_hdr(skb)->cwr = 1;
346 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
349 /* ACK or retransmitted segment: clear ECT|CE */
350 INET_ECN_dontxmit(sk);
352 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
353 tcp_hdr(skb)->ece = 1;
357 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
358 * auto increment end seqno.
360 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
362 skb->ip_summed = CHECKSUM_PARTIAL;
365 TCP_SKB_CB(skb)->tcp_flags = flags;
366 TCP_SKB_CB(skb)->sacked = 0;
368 skb_shinfo(skb)->gso_segs = 1;
369 skb_shinfo(skb)->gso_size = 0;
370 skb_shinfo(skb)->gso_type = 0;
372 TCP_SKB_CB(skb)->seq = seq;
373 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
375 TCP_SKB_CB(skb)->end_seq = seq;
378 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
380 return tp->snd_una != tp->snd_up;
383 #define OPTION_SACK_ADVERTISE (1 << 0)
384 #define OPTION_TS (1 << 1)
385 #define OPTION_MD5 (1 << 2)
386 #define OPTION_WSCALE (1 << 3)
387 #define OPTION_COOKIE_EXTENSION (1 << 4)
389 struct tcp_out_options {
390 u8 options; /* bit field of OPTION_* */
391 u8 ws; /* window scale, 0 to disable */
392 u8 num_sack_blocks; /* number of SACK blocks to include */
393 u8 hash_size; /* bytes in hash_location */
394 u16 mss; /* 0 to disable */
395 __u32 tsval, tsecr; /* need to include OPTION_TS */
396 __u8 *hash_location; /* temporary pointer, overloaded */
399 /* The sysctl int routines are generic, so check consistency here.
401 static u8 tcp_cookie_size_check(u8 desired)
406 /* previously specified */
409 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
410 if (cookie_size <= 0)
411 /* no default specified */
414 if (cookie_size <= TCP_COOKIE_MIN)
415 /* value too small, specify minimum */
416 return TCP_COOKIE_MIN;
418 if (cookie_size >= TCP_COOKIE_MAX)
419 /* value too large, specify maximum */
420 return TCP_COOKIE_MAX;
423 /* 8-bit multiple, illegal, fix it */
426 return (u8)cookie_size;
429 /* Write previously computed TCP options to the packet.
431 * Beware: Something in the Internet is very sensitive to the ordering of
432 * TCP options, we learned this through the hard way, so be careful here.
433 * Luckily we can at least blame others for their non-compliance but from
434 * inter-operatibility perspective it seems that we're somewhat stuck with
435 * the ordering which we have been using if we want to keep working with
436 * those broken things (not that it currently hurts anybody as there isn't
437 * particular reason why the ordering would need to be changed).
439 * At least SACK_PERM as the first option is known to lead to a disaster
440 * (but it may well be that other scenarios fail similarly).
442 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
443 struct tcp_out_options *opts)
445 u8 options = opts->options; /* mungable copy */
447 /* Having both authentication and cookies for security is redundant,
448 * and there's certainly not enough room. Instead, the cookie-less
449 * extension variant is proposed.
451 * Consider the pessimal case with authentication. The options
453 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
455 if (unlikely(OPTION_MD5 & options)) {
456 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
457 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
458 (TCPOLEN_COOKIE_BASE << 16) |
459 (TCPOPT_MD5SIG << 8) |
462 *ptr++ = htonl((TCPOPT_NOP << 24) |
464 (TCPOPT_MD5SIG << 8) |
467 options &= ~OPTION_COOKIE_EXTENSION;
468 /* overload cookie hash location */
469 opts->hash_location = (__u8 *)ptr;
473 if (unlikely(opts->mss)) {
474 *ptr++ = htonl((TCPOPT_MSS << 24) |
475 (TCPOLEN_MSS << 16) |
479 if (likely(OPTION_TS & options)) {
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
482 (TCPOLEN_SACK_PERM << 16) |
483 (TCPOPT_TIMESTAMP << 8) |
485 options &= ~OPTION_SACK_ADVERTISE;
487 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_TIMESTAMP << 8) |
492 *ptr++ = htonl(opts->tsval);
493 *ptr++ = htonl(opts->tsecr);
496 /* Specification requires after timestamp, so do it now.
498 * Consider the pessimal case without authentication. The options
500 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
502 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
503 __u8 *cookie_copy = opts->hash_location;
504 u8 cookie_size = opts->hash_size;
506 /* 8-bit multiple handled in tcp_cookie_size_check() above,
509 if (0x2 & cookie_size) {
510 __u8 *p = (__u8 *)ptr;
512 /* 16-bit multiple */
513 *p++ = TCPOPT_COOKIE;
514 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
515 *p++ = *cookie_copy++;
516 *p++ = *cookie_copy++;
520 /* 32-bit multiple */
521 *ptr++ = htonl(((TCPOPT_NOP << 24) |
523 (TCPOPT_COOKIE << 8) |
524 TCPOLEN_COOKIE_BASE) +
528 if (cookie_size > 0) {
529 memcpy(ptr, cookie_copy, cookie_size);
530 ptr += (cookie_size / 4);
534 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
535 *ptr++ = htonl((TCPOPT_NOP << 24) |
537 (TCPOPT_SACK_PERM << 8) |
541 if (unlikely(OPTION_WSCALE & options)) {
542 *ptr++ = htonl((TCPOPT_NOP << 24) |
543 (TCPOPT_WINDOW << 16) |
544 (TCPOLEN_WINDOW << 8) |
548 if (unlikely(opts->num_sack_blocks)) {
549 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
550 tp->duplicate_sack : tp->selective_acks;
553 *ptr++ = htonl((TCPOPT_NOP << 24) |
556 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
557 TCPOLEN_SACK_PERBLOCK)));
559 for (this_sack = 0; this_sack < opts->num_sack_blocks;
561 *ptr++ = htonl(sp[this_sack].start_seq);
562 *ptr++ = htonl(sp[this_sack].end_seq);
565 tp->rx_opt.dsack = 0;
569 /* Compute TCP options for SYN packets. This is not the final
570 * network wire format yet.
572 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
573 struct tcp_out_options *opts,
574 struct tcp_md5sig_key **md5)
576 struct tcp_sock *tp = tcp_sk(sk);
577 struct tcp_cookie_values *cvp = tp->cookie_values;
578 unsigned int remaining = MAX_TCP_OPTION_SPACE;
579 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
580 tcp_cookie_size_check(cvp->cookie_desired) :
583 #ifdef CONFIG_TCP_MD5SIG
584 *md5 = tp->af_specific->md5_lookup(sk, sk);
586 opts->options |= OPTION_MD5;
587 remaining -= TCPOLEN_MD5SIG_ALIGNED;
593 /* We always get an MSS option. The option bytes which will be seen in
594 * normal data packets should timestamps be used, must be in the MSS
595 * advertised. But we subtract them from tp->mss_cache so that
596 * calculations in tcp_sendmsg are simpler etc. So account for this
597 * fact here if necessary. If we don't do this correctly, as a
598 * receiver we won't recognize data packets as being full sized when we
599 * should, and thus we won't abide by the delayed ACK rules correctly.
600 * SACKs don't matter, we never delay an ACK when we have any of those
602 opts->mss = tcp_advertise_mss(sk);
603 remaining -= TCPOLEN_MSS_ALIGNED;
605 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
606 opts->options |= OPTION_TS;
607 opts->tsval = TCP_SKB_CB(skb)->when;
608 opts->tsecr = tp->rx_opt.ts_recent;
609 remaining -= TCPOLEN_TSTAMP_ALIGNED;
611 if (likely(sysctl_tcp_window_scaling)) {
612 opts->ws = tp->rx_opt.rcv_wscale;
613 opts->options |= OPTION_WSCALE;
614 remaining -= TCPOLEN_WSCALE_ALIGNED;
616 if (likely(sysctl_tcp_sack)) {
617 opts->options |= OPTION_SACK_ADVERTISE;
618 if (unlikely(!(OPTION_TS & opts->options)))
619 remaining -= TCPOLEN_SACKPERM_ALIGNED;
622 /* Note that timestamps are required by the specification.
624 * Odd numbers of bytes are prohibited by the specification, ensuring
625 * that the cookie is 16-bit aligned, and the resulting cookie pair is
629 (OPTION_TS & opts->options) &&
631 int need = TCPOLEN_COOKIE_BASE + cookie_size;
634 /* 32-bit multiple */
635 need += 2; /* NOPs */
637 if (need > remaining) {
638 /* try shrinking cookie to fit */
643 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
647 if (TCP_COOKIE_MIN <= cookie_size) {
648 opts->options |= OPTION_COOKIE_EXTENSION;
649 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
650 opts->hash_size = cookie_size;
652 /* Remember for future incarnations. */
653 cvp->cookie_desired = cookie_size;
655 if (cvp->cookie_desired != cvp->cookie_pair_size) {
656 /* Currently use random bytes as a nonce,
657 * assuming these are completely unpredictable
658 * by hostile users of the same system.
660 get_random_bytes(&cvp->cookie_pair[0],
662 cvp->cookie_pair_size = cookie_size;
668 return MAX_TCP_OPTION_SPACE - remaining;
671 /* Set up TCP options for SYN-ACKs. */
672 static unsigned int tcp_synack_options(struct sock *sk,
673 struct request_sock *req,
674 unsigned int mss, struct sk_buff *skb,
675 struct tcp_out_options *opts,
676 struct tcp_md5sig_key **md5,
677 struct tcp_extend_values *xvp)
679 struct inet_request_sock *ireq = inet_rsk(req);
680 unsigned int remaining = MAX_TCP_OPTION_SPACE;
681 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
685 #ifdef CONFIG_TCP_MD5SIG
686 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
688 opts->options |= OPTION_MD5;
689 remaining -= TCPOLEN_MD5SIG_ALIGNED;
691 /* We can't fit any SACK blocks in a packet with MD5 + TS
692 * options. There was discussion about disabling SACK
693 * rather than TS in order to fit in better with old,
694 * buggy kernels, but that was deemed to be unnecessary.
696 ireq->tstamp_ok &= !ireq->sack_ok;
702 /* We always send an MSS option. */
704 remaining -= TCPOLEN_MSS_ALIGNED;
706 if (likely(ireq->wscale_ok)) {
707 opts->ws = ireq->rcv_wscale;
708 opts->options |= OPTION_WSCALE;
709 remaining -= TCPOLEN_WSCALE_ALIGNED;
711 if (likely(ireq->tstamp_ok)) {
712 opts->options |= OPTION_TS;
713 opts->tsval = TCP_SKB_CB(skb)->when;
714 opts->tsecr = req->ts_recent;
715 remaining -= TCPOLEN_TSTAMP_ALIGNED;
717 if (likely(ireq->sack_ok)) {
718 opts->options |= OPTION_SACK_ADVERTISE;
719 if (unlikely(!ireq->tstamp_ok))
720 remaining -= TCPOLEN_SACKPERM_ALIGNED;
723 /* Similar rationale to tcp_syn_options() applies here, too.
724 * If the <SYN> options fit, the same options should fit now!
728 cookie_plus > TCPOLEN_COOKIE_BASE) {
729 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
732 /* 32-bit multiple */
733 need += 2; /* NOPs */
735 if (need <= remaining) {
736 opts->options |= OPTION_COOKIE_EXTENSION;
737 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
740 /* There's no error return, so flag it. */
741 xvp->cookie_out_never = 1; /* true */
745 return MAX_TCP_OPTION_SPACE - remaining;
748 /* Compute TCP options for ESTABLISHED sockets. This is not the
749 * final wire format yet.
751 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
752 struct tcp_out_options *opts,
753 struct tcp_md5sig_key **md5)
755 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
756 struct tcp_sock *tp = tcp_sk(sk);
757 unsigned int size = 0;
758 unsigned int eff_sacks;
760 #ifdef CONFIG_TCP_MD5SIG
761 *md5 = tp->af_specific->md5_lookup(sk, sk);
762 if (unlikely(*md5)) {
763 opts->options |= OPTION_MD5;
764 size += TCPOLEN_MD5SIG_ALIGNED;
770 if (likely(tp->rx_opt.tstamp_ok)) {
771 opts->options |= OPTION_TS;
772 opts->tsval = tcb ? tcb->when : 0;
773 opts->tsecr = tp->rx_opt.ts_recent;
774 size += TCPOLEN_TSTAMP_ALIGNED;
777 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
778 if (unlikely(eff_sacks)) {
779 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
780 opts->num_sack_blocks =
781 min_t(unsigned int, eff_sacks,
782 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
783 TCPOLEN_SACK_PERBLOCK);
784 size += TCPOLEN_SACK_BASE_ALIGNED +
785 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
792 /* TCP SMALL QUEUES (TSQ)
794 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
795 * to reduce RTT and bufferbloat.
796 * We do this using a special skb destructor (tcp_wfree).
798 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
799 * needs to be reallocated in a driver.
800 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
802 * Since transmit from skb destructor is forbidden, we use a tasklet
803 * to process all sockets that eventually need to send more skbs.
804 * We use one tasklet per cpu, with its own queue of sockets.
807 struct tasklet_struct tasklet;
808 struct list_head head; /* queue of tcp sockets */
810 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
813 * One tasklest per cpu tries to send more skbs.
814 * We run in tasklet context but need to disable irqs when
815 * transfering tsq->head because tcp_wfree() might
816 * interrupt us (non NAPI drivers)
818 static void tcp_tasklet_func(unsigned long data)
820 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
823 struct list_head *q, *n;
827 local_irq_save(flags);
828 list_splice_init(&tsq->head, &list);
829 local_irq_restore(flags);
831 list_for_each_safe(q, n, &list) {
832 tp = list_entry(q, struct tcp_sock, tsq_node);
833 list_del(&tp->tsq_node);
835 sk = (struct sock *)tp;
838 if (!sock_owned_by_user(sk)) {
839 if ((1 << sk->sk_state) &
840 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 |
841 TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
847 /* defer the work to tcp_release_cb() */
848 set_bit(TSQ_OWNED, &tp->tsq_flags);
852 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
858 * tcp_release_cb - tcp release_sock() callback
861 * called from release_sock() to perform protocol dependent
862 * actions before socket release.
864 void tcp_release_cb(struct sock *sk)
866 struct tcp_sock *tp = tcp_sk(sk);
868 if (test_and_clear_bit(TSQ_OWNED, &tp->tsq_flags)) {
869 if ((1 << sk->sk_state) &
870 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 |
871 TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
878 EXPORT_SYMBOL(tcp_release_cb);
880 void __init tcp_tasklet_init(void)
884 for_each_possible_cpu(i) {
885 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
887 INIT_LIST_HEAD(&tsq->head);
888 tasklet_init(&tsq->tasklet,
895 * Write buffer destructor automatically called from kfree_skb.
896 * We cant xmit new skbs from this context, as we might already
899 void tcp_wfree(struct sk_buff *skb)
901 struct sock *sk = skb->sk;
902 struct tcp_sock *tp = tcp_sk(sk);
904 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
905 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
907 struct tsq_tasklet *tsq;
909 /* Keep a ref on socket.
910 * This last ref will be released in tcp_tasklet_func()
912 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
914 /* queue this socket to tasklet queue */
915 local_irq_save(flags);
916 tsq = &__get_cpu_var(tsq_tasklet);
917 list_add(&tp->tsq_node, &tsq->head);
918 tasklet_schedule(&tsq->tasklet);
919 local_irq_restore(flags);
925 /* This routine actually transmits TCP packets queued in by
926 * tcp_do_sendmsg(). This is used by both the initial
927 * transmission and possible later retransmissions.
928 * All SKB's seen here are completely headerless. It is our
929 * job to build the TCP header, and pass the packet down to
930 * IP so it can do the same plus pass the packet off to the
933 * We are working here with either a clone of the original
934 * SKB, or a fresh unique copy made by the retransmit engine.
936 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
939 const struct inet_connection_sock *icsk = inet_csk(sk);
940 struct inet_sock *inet;
942 struct tcp_skb_cb *tcb;
943 struct tcp_out_options opts;
944 unsigned int tcp_options_size, tcp_header_size;
945 struct tcp_md5sig_key *md5;
949 BUG_ON(!skb || !tcp_skb_pcount(skb));
951 /* If congestion control is doing timestamping, we must
952 * take such a timestamp before we potentially clone/copy.
954 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
955 __net_timestamp(skb);
957 if (likely(clone_it)) {
958 if (unlikely(skb_cloned(skb)))
959 skb = pskb_copy(skb, gfp_mask);
961 skb = skb_clone(skb, gfp_mask);
968 tcb = TCP_SKB_CB(skb);
969 memset(&opts, 0, sizeof(opts));
971 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
972 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
974 tcp_options_size = tcp_established_options(sk, skb, &opts,
976 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
978 if (tcp_packets_in_flight(tp) == 0) {
979 tcp_ca_event(sk, CA_EVENT_TX_START);
984 skb_push(skb, tcp_header_size);
985 skb_reset_transport_header(skb);
989 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
990 tcp_wfree : sock_wfree;
991 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
993 /* Build TCP header and checksum it. */
995 th->source = inet->inet_sport;
996 th->dest = inet->inet_dport;
997 th->seq = htonl(tcb->seq);
998 th->ack_seq = htonl(tp->rcv_nxt);
999 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1002 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1003 /* RFC1323: The window in SYN & SYN/ACK segments
1006 th->window = htons(min(tp->rcv_wnd, 65535U));
1008 th->window = htons(tcp_select_window(sk));
1013 /* The urg_mode check is necessary during a below snd_una win probe */
1014 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1015 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1016 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1018 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1019 th->urg_ptr = htons(0xFFFF);
1024 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1025 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1026 TCP_ECN_send(sk, skb, tcp_header_size);
1028 #ifdef CONFIG_TCP_MD5SIG
1029 /* Calculate the MD5 hash, as we have all we need now */
1031 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1032 tp->af_specific->calc_md5_hash(opts.hash_location,
1033 md5, sk, NULL, skb);
1037 icsk->icsk_af_ops->send_check(sk, skb);
1039 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1040 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1042 if (skb->len != tcp_header_size)
1043 tcp_event_data_sent(tp, sk);
1045 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1046 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1047 tcp_skb_pcount(skb));
1049 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1050 if (likely(err <= 0))
1053 tcp_enter_cwr(sk, 1);
1055 return net_xmit_eval(err);
1058 /* This routine just queues the buffer for sending.
1060 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1061 * otherwise socket can stall.
1063 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1065 struct tcp_sock *tp = tcp_sk(sk);
1067 /* Advance write_seq and place onto the write_queue. */
1068 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1069 skb_header_release(skb);
1070 tcp_add_write_queue_tail(sk, skb);
1071 sk->sk_wmem_queued += skb->truesize;
1072 sk_mem_charge(sk, skb->truesize);
1075 /* Initialize TSO segments for a packet. */
1076 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1077 unsigned int mss_now)
1079 if (skb->len <= mss_now || !sk_can_gso(sk) ||
1080 skb->ip_summed == CHECKSUM_NONE) {
1081 /* Avoid the costly divide in the normal
1084 skb_shinfo(skb)->gso_segs = 1;
1085 skb_shinfo(skb)->gso_size = 0;
1086 skb_shinfo(skb)->gso_type = 0;
1088 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1089 skb_shinfo(skb)->gso_size = mss_now;
1090 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1094 /* When a modification to fackets out becomes necessary, we need to check
1095 * skb is counted to fackets_out or not.
1097 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1100 struct tcp_sock *tp = tcp_sk(sk);
1102 if (!tp->sacked_out || tcp_is_reno(tp))
1105 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1106 tp->fackets_out -= decr;
1109 /* Pcount in the middle of the write queue got changed, we need to do various
1110 * tweaks to fix counters
1112 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1114 struct tcp_sock *tp = tcp_sk(sk);
1116 tp->packets_out -= decr;
1118 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1119 tp->sacked_out -= decr;
1120 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1121 tp->retrans_out -= decr;
1122 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1123 tp->lost_out -= decr;
1125 /* Reno case is special. Sigh... */
1126 if (tcp_is_reno(tp) && decr > 0)
1127 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1129 tcp_adjust_fackets_out(sk, skb, decr);
1131 if (tp->lost_skb_hint &&
1132 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1133 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1134 tp->lost_cnt_hint -= decr;
1136 tcp_verify_left_out(tp);
1139 /* Function to create two new TCP segments. Shrinks the given segment
1140 * to the specified size and appends a new segment with the rest of the
1141 * packet to the list. This won't be called frequently, I hope.
1142 * Remember, these are still headerless SKBs at this point.
1144 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1145 unsigned int mss_now)
1147 struct tcp_sock *tp = tcp_sk(sk);
1148 struct sk_buff *buff;
1149 int nsize, old_factor;
1153 if (WARN_ON(len > skb->len))
1156 nsize = skb_headlen(skb) - len;
1160 if (skb_cloned(skb) &&
1161 skb_is_nonlinear(skb) &&
1162 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1165 /* Get a new skb... force flag on. */
1166 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1168 return -ENOMEM; /* We'll just try again later. */
1170 sk->sk_wmem_queued += buff->truesize;
1171 sk_mem_charge(sk, buff->truesize);
1172 nlen = skb->len - len - nsize;
1173 buff->truesize += nlen;
1174 skb->truesize -= nlen;
1176 /* Correct the sequence numbers. */
1177 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1178 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1179 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1181 /* PSH and FIN should only be set in the second packet. */
1182 flags = TCP_SKB_CB(skb)->tcp_flags;
1183 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1184 TCP_SKB_CB(buff)->tcp_flags = flags;
1185 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1187 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1188 /* Copy and checksum data tail into the new buffer. */
1189 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1190 skb_put(buff, nsize),
1195 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1197 skb->ip_summed = CHECKSUM_PARTIAL;
1198 skb_split(skb, buff, len);
1201 buff->ip_summed = skb->ip_summed;
1203 /* Looks stupid, but our code really uses when of
1204 * skbs, which it never sent before. --ANK
1206 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1207 buff->tstamp = skb->tstamp;
1209 old_factor = tcp_skb_pcount(skb);
1211 /* Fix up tso_factor for both original and new SKB. */
1212 tcp_set_skb_tso_segs(sk, skb, mss_now);
1213 tcp_set_skb_tso_segs(sk, buff, mss_now);
1215 /* If this packet has been sent out already, we must
1216 * adjust the various packet counters.
1218 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1219 int diff = old_factor - tcp_skb_pcount(skb) -
1220 tcp_skb_pcount(buff);
1223 tcp_adjust_pcount(sk, skb, diff);
1226 /* Link BUFF into the send queue. */
1227 skb_header_release(buff);
1228 tcp_insert_write_queue_after(skb, buff, sk);
1233 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1234 * eventually). The difference is that pulled data not copied, but
1235 * immediately discarded.
1237 static void __pskb_trim_head(struct sk_buff *skb, int len)
1241 eat = min_t(int, len, skb_headlen(skb));
1243 __skb_pull(skb, eat);
1244 skb->avail_size -= eat;
1251 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1252 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1255 skb_frag_unref(skb, i);
1258 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1260 skb_shinfo(skb)->frags[k].page_offset += eat;
1261 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1267 skb_shinfo(skb)->nr_frags = k;
1269 skb_reset_tail_pointer(skb);
1270 skb->data_len -= len;
1271 skb->len = skb->data_len;
1274 /* Remove acked data from a packet in the transmit queue. */
1275 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1277 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1280 __pskb_trim_head(skb, len);
1282 TCP_SKB_CB(skb)->seq += len;
1283 skb->ip_summed = CHECKSUM_PARTIAL;
1285 skb->truesize -= len;
1286 sk->sk_wmem_queued -= len;
1287 sk_mem_uncharge(sk, len);
1288 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1290 /* Any change of skb->len requires recalculation of tso factor. */
1291 if (tcp_skb_pcount(skb) > 1)
1292 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1297 /* Calculate MSS. Not accounting for SACKs here. */
1298 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1300 const struct tcp_sock *tp = tcp_sk(sk);
1301 const struct inet_connection_sock *icsk = inet_csk(sk);
1304 /* Calculate base mss without TCP options:
1305 It is MMS_S - sizeof(tcphdr) of rfc1122
1307 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1309 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1310 if (icsk->icsk_af_ops->net_frag_header_len) {
1311 const struct dst_entry *dst = __sk_dst_get(sk);
1313 if (dst && dst_allfrag(dst))
1314 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1317 /* Clamp it (mss_clamp does not include tcp options) */
1318 if (mss_now > tp->rx_opt.mss_clamp)
1319 mss_now = tp->rx_opt.mss_clamp;
1321 /* Now subtract optional transport overhead */
1322 mss_now -= icsk->icsk_ext_hdr_len;
1324 /* Then reserve room for full set of TCP options and 8 bytes of data */
1328 /* Now subtract TCP options size, not including SACKs */
1329 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1334 /* Inverse of above */
1335 int tcp_mss_to_mtu(struct sock *sk, int mss)
1337 const struct tcp_sock *tp = tcp_sk(sk);
1338 const struct inet_connection_sock *icsk = inet_csk(sk);
1342 tp->tcp_header_len +
1343 icsk->icsk_ext_hdr_len +
1344 icsk->icsk_af_ops->net_header_len;
1346 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1347 if (icsk->icsk_af_ops->net_frag_header_len) {
1348 const struct dst_entry *dst = __sk_dst_get(sk);
1350 if (dst && dst_allfrag(dst))
1351 mtu += icsk->icsk_af_ops->net_frag_header_len;
1356 /* MTU probing init per socket */
1357 void tcp_mtup_init(struct sock *sk)
1359 struct tcp_sock *tp = tcp_sk(sk);
1360 struct inet_connection_sock *icsk = inet_csk(sk);
1362 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1363 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1364 icsk->icsk_af_ops->net_header_len;
1365 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1366 icsk->icsk_mtup.probe_size = 0;
1368 EXPORT_SYMBOL(tcp_mtup_init);
1370 /* This function synchronize snd mss to current pmtu/exthdr set.
1372 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1373 for TCP options, but includes only bare TCP header.
1375 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1376 It is minimum of user_mss and mss received with SYN.
1377 It also does not include TCP options.
1379 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1381 tp->mss_cache is current effective sending mss, including
1382 all tcp options except for SACKs. It is evaluated,
1383 taking into account current pmtu, but never exceeds
1384 tp->rx_opt.mss_clamp.
1386 NOTE1. rfc1122 clearly states that advertised MSS
1387 DOES NOT include either tcp or ip options.
1389 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1390 are READ ONLY outside this function. --ANK (980731)
1392 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1394 struct tcp_sock *tp = tcp_sk(sk);
1395 struct inet_connection_sock *icsk = inet_csk(sk);
1398 if (icsk->icsk_mtup.search_high > pmtu)
1399 icsk->icsk_mtup.search_high = pmtu;
1401 mss_now = tcp_mtu_to_mss(sk, pmtu);
1402 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1404 /* And store cached results */
1405 icsk->icsk_pmtu_cookie = pmtu;
1406 if (icsk->icsk_mtup.enabled)
1407 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1408 tp->mss_cache = mss_now;
1412 EXPORT_SYMBOL(tcp_sync_mss);
1414 /* Compute the current effective MSS, taking SACKs and IP options,
1415 * and even PMTU discovery events into account.
1417 unsigned int tcp_current_mss(struct sock *sk)
1419 const struct tcp_sock *tp = tcp_sk(sk);
1420 const struct dst_entry *dst = __sk_dst_get(sk);
1422 unsigned int header_len;
1423 struct tcp_out_options opts;
1424 struct tcp_md5sig_key *md5;
1426 mss_now = tp->mss_cache;
1429 u32 mtu = dst_mtu(dst);
1430 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1431 mss_now = tcp_sync_mss(sk, mtu);
1434 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1435 sizeof(struct tcphdr);
1436 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1437 * some common options. If this is an odd packet (because we have SACK
1438 * blocks etc) then our calculated header_len will be different, and
1439 * we have to adjust mss_now correspondingly */
1440 if (header_len != tp->tcp_header_len) {
1441 int delta = (int) header_len - tp->tcp_header_len;
1448 /* Congestion window validation. (RFC2861) */
1449 static void tcp_cwnd_validate(struct sock *sk)
1451 struct tcp_sock *tp = tcp_sk(sk);
1453 if (tp->packets_out >= tp->snd_cwnd) {
1454 /* Network is feed fully. */
1455 tp->snd_cwnd_used = 0;
1456 tp->snd_cwnd_stamp = tcp_time_stamp;
1458 /* Network starves. */
1459 if (tp->packets_out > tp->snd_cwnd_used)
1460 tp->snd_cwnd_used = tp->packets_out;
1462 if (sysctl_tcp_slow_start_after_idle &&
1463 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1464 tcp_cwnd_application_limited(sk);
1468 /* Returns the portion of skb which can be sent right away without
1469 * introducing MSS oddities to segment boundaries. In rare cases where
1470 * mss_now != mss_cache, we will request caller to create a small skb
1471 * per input skb which could be mostly avoided here (if desired).
1473 * We explicitly want to create a request for splitting write queue tail
1474 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1475 * thus all the complexity (cwnd_len is always MSS multiple which we
1476 * return whenever allowed by the other factors). Basically we need the
1477 * modulo only when the receiver window alone is the limiting factor or
1478 * when we would be allowed to send the split-due-to-Nagle skb fully.
1480 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1481 unsigned int mss_now, unsigned int cwnd)
1483 const struct tcp_sock *tp = tcp_sk(sk);
1484 u32 needed, window, cwnd_len;
1486 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1487 cwnd_len = mss_now * cwnd;
1489 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1492 needed = min(skb->len, window);
1494 if (cwnd_len <= needed)
1497 return needed - needed % mss_now;
1500 /* Can at least one segment of SKB be sent right now, according to the
1501 * congestion window rules? If so, return how many segments are allowed.
1503 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1504 const struct sk_buff *skb)
1506 u32 in_flight, cwnd;
1508 /* Don't be strict about the congestion window for the final FIN. */
1509 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1510 tcp_skb_pcount(skb) == 1)
1513 in_flight = tcp_packets_in_flight(tp);
1514 cwnd = tp->snd_cwnd;
1515 if (in_flight < cwnd)
1516 return (cwnd - in_flight);
1521 /* Initialize TSO state of a skb.
1522 * This must be invoked the first time we consider transmitting
1523 * SKB onto the wire.
1525 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1526 unsigned int mss_now)
1528 int tso_segs = tcp_skb_pcount(skb);
1530 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1531 tcp_set_skb_tso_segs(sk, skb, mss_now);
1532 tso_segs = tcp_skb_pcount(skb);
1537 /* Minshall's variant of the Nagle send check. */
1538 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1540 return after(tp->snd_sml, tp->snd_una) &&
1541 !after(tp->snd_sml, tp->snd_nxt);
1544 /* Return false, if packet can be sent now without violation Nagle's rules:
1545 * 1. It is full sized.
1546 * 2. Or it contains FIN. (already checked by caller)
1547 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1548 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1549 * With Minshall's modification: all sent small packets are ACKed.
1551 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1552 const struct sk_buff *skb,
1553 unsigned int mss_now, int nonagle)
1555 return skb->len < mss_now &&
1556 ((nonagle & TCP_NAGLE_CORK) ||
1557 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1560 /* Return true if the Nagle test allows this packet to be
1563 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1564 unsigned int cur_mss, int nonagle)
1566 /* Nagle rule does not apply to frames, which sit in the middle of the
1567 * write_queue (they have no chances to get new data).
1569 * This is implemented in the callers, where they modify the 'nonagle'
1570 * argument based upon the location of SKB in the send queue.
1572 if (nonagle & TCP_NAGLE_PUSH)
1575 /* Don't use the nagle rule for urgent data (or for the final FIN).
1576 * Nagle can be ignored during F-RTO too (see RFC4138).
1578 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1579 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1582 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1588 /* Does at least the first segment of SKB fit into the send window? */
1589 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1590 const struct sk_buff *skb,
1591 unsigned int cur_mss)
1593 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1595 if (skb->len > cur_mss)
1596 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1598 return !after(end_seq, tcp_wnd_end(tp));
1601 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1602 * should be put on the wire right now. If so, it returns the number of
1603 * packets allowed by the congestion window.
1605 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1606 unsigned int cur_mss, int nonagle)
1608 const struct tcp_sock *tp = tcp_sk(sk);
1609 unsigned int cwnd_quota;
1611 tcp_init_tso_segs(sk, skb, cur_mss);
1613 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1616 cwnd_quota = tcp_cwnd_test(tp, skb);
1617 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1623 /* Test if sending is allowed right now. */
1624 bool tcp_may_send_now(struct sock *sk)
1626 const struct tcp_sock *tp = tcp_sk(sk);
1627 struct sk_buff *skb = tcp_send_head(sk);
1630 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1631 (tcp_skb_is_last(sk, skb) ?
1632 tp->nonagle : TCP_NAGLE_PUSH));
1635 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1636 * which is put after SKB on the list. It is very much like
1637 * tcp_fragment() except that it may make several kinds of assumptions
1638 * in order to speed up the splitting operation. In particular, we
1639 * know that all the data is in scatter-gather pages, and that the
1640 * packet has never been sent out before (and thus is not cloned).
1642 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1643 unsigned int mss_now, gfp_t gfp)
1645 struct sk_buff *buff;
1646 int nlen = skb->len - len;
1649 /* All of a TSO frame must be composed of paged data. */
1650 if (skb->len != skb->data_len)
1651 return tcp_fragment(sk, skb, len, mss_now);
1653 buff = sk_stream_alloc_skb(sk, 0, gfp);
1654 if (unlikely(buff == NULL))
1657 sk->sk_wmem_queued += buff->truesize;
1658 sk_mem_charge(sk, buff->truesize);
1659 buff->truesize += nlen;
1660 skb->truesize -= nlen;
1662 /* Correct the sequence numbers. */
1663 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1664 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1665 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1667 /* PSH and FIN should only be set in the second packet. */
1668 flags = TCP_SKB_CB(skb)->tcp_flags;
1669 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1670 TCP_SKB_CB(buff)->tcp_flags = flags;
1672 /* This packet was never sent out yet, so no SACK bits. */
1673 TCP_SKB_CB(buff)->sacked = 0;
1675 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1676 skb_split(skb, buff, len);
1678 /* Fix up tso_factor for both original and new SKB. */
1679 tcp_set_skb_tso_segs(sk, skb, mss_now);
1680 tcp_set_skb_tso_segs(sk, buff, mss_now);
1682 /* Link BUFF into the send queue. */
1683 skb_header_release(buff);
1684 tcp_insert_write_queue_after(skb, buff, sk);
1689 /* Try to defer sending, if possible, in order to minimize the amount
1690 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1692 * This algorithm is from John Heffner.
1694 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1696 struct tcp_sock *tp = tcp_sk(sk);
1697 const struct inet_connection_sock *icsk = inet_csk(sk);
1698 u32 send_win, cong_win, limit, in_flight;
1701 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1704 if (icsk->icsk_ca_state != TCP_CA_Open)
1707 /* Defer for less than two clock ticks. */
1708 if (tp->tso_deferred &&
1709 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1712 in_flight = tcp_packets_in_flight(tp);
1714 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1716 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1718 /* From in_flight test above, we know that cwnd > in_flight. */
1719 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1721 limit = min(send_win, cong_win);
1723 /* If a full-sized TSO skb can be sent, do it. */
1724 if (limit >= sk->sk_gso_max_size)
1727 /* Middle in queue won't get any more data, full sendable already? */
1728 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1731 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1733 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1735 /* If at least some fraction of a window is available,
1738 chunk /= win_divisor;
1742 /* Different approach, try not to defer past a single
1743 * ACK. Receiver should ACK every other full sized
1744 * frame, so if we have space for more than 3 frames
1747 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1751 /* Ok, it looks like it is advisable to defer. */
1752 tp->tso_deferred = 1 | (jiffies << 1);
1757 tp->tso_deferred = 0;
1761 /* Create a new MTU probe if we are ready.
1762 * MTU probe is regularly attempting to increase the path MTU by
1763 * deliberately sending larger packets. This discovers routing
1764 * changes resulting in larger path MTUs.
1766 * Returns 0 if we should wait to probe (no cwnd available),
1767 * 1 if a probe was sent,
1770 static int tcp_mtu_probe(struct sock *sk)
1772 struct tcp_sock *tp = tcp_sk(sk);
1773 struct inet_connection_sock *icsk = inet_csk(sk);
1774 struct sk_buff *skb, *nskb, *next;
1781 /* Not currently probing/verifying,
1783 * have enough cwnd, and
1784 * not SACKing (the variable headers throw things off) */
1785 if (!icsk->icsk_mtup.enabled ||
1786 icsk->icsk_mtup.probe_size ||
1787 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1788 tp->snd_cwnd < 11 ||
1789 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1792 /* Very simple search strategy: just double the MSS. */
1793 mss_now = tcp_current_mss(sk);
1794 probe_size = 2 * tp->mss_cache;
1795 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1796 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1797 /* TODO: set timer for probe_converge_event */
1801 /* Have enough data in the send queue to probe? */
1802 if (tp->write_seq - tp->snd_nxt < size_needed)
1805 if (tp->snd_wnd < size_needed)
1807 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1810 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1811 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1812 if (!tcp_packets_in_flight(tp))
1818 /* We're allowed to probe. Build it now. */
1819 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1821 sk->sk_wmem_queued += nskb->truesize;
1822 sk_mem_charge(sk, nskb->truesize);
1824 skb = tcp_send_head(sk);
1826 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1827 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1828 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1829 TCP_SKB_CB(nskb)->sacked = 0;
1831 nskb->ip_summed = skb->ip_summed;
1833 tcp_insert_write_queue_before(nskb, skb, sk);
1836 tcp_for_write_queue_from_safe(skb, next, sk) {
1837 copy = min_t(int, skb->len, probe_size - len);
1838 if (nskb->ip_summed)
1839 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1841 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1842 skb_put(nskb, copy),
1845 if (skb->len <= copy) {
1846 /* We've eaten all the data from this skb.
1848 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1849 tcp_unlink_write_queue(skb, sk);
1850 sk_wmem_free_skb(sk, skb);
1852 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1853 ~(TCPHDR_FIN|TCPHDR_PSH);
1854 if (!skb_shinfo(skb)->nr_frags) {
1855 skb_pull(skb, copy);
1856 if (skb->ip_summed != CHECKSUM_PARTIAL)
1857 skb->csum = csum_partial(skb->data,
1860 __pskb_trim_head(skb, copy);
1861 tcp_set_skb_tso_segs(sk, skb, mss_now);
1863 TCP_SKB_CB(skb)->seq += copy;
1868 if (len >= probe_size)
1871 tcp_init_tso_segs(sk, nskb, nskb->len);
1873 /* We're ready to send. If this fails, the probe will
1874 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1875 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1876 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1877 /* Decrement cwnd here because we are sending
1878 * effectively two packets. */
1880 tcp_event_new_data_sent(sk, nskb);
1882 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1883 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1884 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1892 /* This routine writes packets to the network. It advances the
1893 * send_head. This happens as incoming acks open up the remote
1896 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1897 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1898 * account rare use of URG, this is not a big flaw.
1900 * Returns true, if no segments are in flight and we have queued segments,
1901 * but cannot send anything now because of SWS or another problem.
1903 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1904 int push_one, gfp_t gfp)
1906 struct tcp_sock *tp = tcp_sk(sk);
1907 struct sk_buff *skb;
1908 unsigned int tso_segs, sent_pkts;
1915 /* Do MTU probing. */
1916 result = tcp_mtu_probe(sk);
1919 } else if (result > 0) {
1924 while ((skb = tcp_send_head(sk))) {
1928 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1931 cwnd_quota = tcp_cwnd_test(tp, skb);
1935 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1938 if (tso_segs == 1) {
1939 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1940 (tcp_skb_is_last(sk, skb) ?
1941 nonagle : TCP_NAGLE_PUSH))))
1944 if (!push_one && tcp_tso_should_defer(sk, skb))
1948 /* TSQ : sk_wmem_alloc accounts skb truesize,
1949 * including skb overhead. But thats OK.
1951 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1952 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1956 if (tso_segs > 1 && !tcp_urg_mode(tp))
1957 limit = tcp_mss_split_point(sk, skb, mss_now,
1960 if (skb->len > limit &&
1961 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1964 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1966 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1969 /* Advance the send_head. This one is sent out.
1970 * This call will increment packets_out.
1972 tcp_event_new_data_sent(sk, skb);
1974 tcp_minshall_update(tp, mss_now, skb);
1975 sent_pkts += tcp_skb_pcount(skb);
1980 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
1981 tp->prr_out += sent_pkts;
1983 if (likely(sent_pkts)) {
1984 tcp_cwnd_validate(sk);
1987 return !tp->packets_out && tcp_send_head(sk);
1990 /* Push out any pending frames which were held back due to
1991 * TCP_CORK or attempt at coalescing tiny packets.
1992 * The socket must be locked by the caller.
1994 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1997 /* If we are closed, the bytes will have to remain here.
1998 * In time closedown will finish, we empty the write queue and
1999 * all will be happy.
2001 if (unlikely(sk->sk_state == TCP_CLOSE))
2004 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
2005 tcp_check_probe_timer(sk);
2008 /* Send _single_ skb sitting at the send head. This function requires
2009 * true push pending frames to setup probe timer etc.
2011 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2013 struct sk_buff *skb = tcp_send_head(sk);
2015 BUG_ON(!skb || skb->len < mss_now);
2017 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2020 /* This function returns the amount that we can raise the
2021 * usable window based on the following constraints
2023 * 1. The window can never be shrunk once it is offered (RFC 793)
2024 * 2. We limit memory per socket
2027 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2028 * RECV.NEXT + RCV.WIN fixed until:
2029 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2031 * i.e. don't raise the right edge of the window until you can raise
2032 * it at least MSS bytes.
2034 * Unfortunately, the recommended algorithm breaks header prediction,
2035 * since header prediction assumes th->window stays fixed.
2037 * Strictly speaking, keeping th->window fixed violates the receiver
2038 * side SWS prevention criteria. The problem is that under this rule
2039 * a stream of single byte packets will cause the right side of the
2040 * window to always advance by a single byte.
2042 * Of course, if the sender implements sender side SWS prevention
2043 * then this will not be a problem.
2045 * BSD seems to make the following compromise:
2047 * If the free space is less than the 1/4 of the maximum
2048 * space available and the free space is less than 1/2 mss,
2049 * then set the window to 0.
2050 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2051 * Otherwise, just prevent the window from shrinking
2052 * and from being larger than the largest representable value.
2054 * This prevents incremental opening of the window in the regime
2055 * where TCP is limited by the speed of the reader side taking
2056 * data out of the TCP receive queue. It does nothing about
2057 * those cases where the window is constrained on the sender side
2058 * because the pipeline is full.
2060 * BSD also seems to "accidentally" limit itself to windows that are a
2061 * multiple of MSS, at least until the free space gets quite small.
2062 * This would appear to be a side effect of the mbuf implementation.
2063 * Combining these two algorithms results in the observed behavior
2064 * of having a fixed window size at almost all times.
2066 * Below we obtain similar behavior by forcing the offered window to
2067 * a multiple of the mss when it is feasible to do so.
2069 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2070 * Regular options like TIMESTAMP are taken into account.
2072 u32 __tcp_select_window(struct sock *sk)
2074 struct inet_connection_sock *icsk = inet_csk(sk);
2075 struct tcp_sock *tp = tcp_sk(sk);
2076 /* MSS for the peer's data. Previous versions used mss_clamp
2077 * here. I don't know if the value based on our guesses
2078 * of peer's MSS is better for the performance. It's more correct
2079 * but may be worse for the performance because of rcv_mss
2080 * fluctuations. --SAW 1998/11/1
2082 int mss = icsk->icsk_ack.rcv_mss;
2083 int free_space = tcp_space(sk);
2084 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2087 if (mss > full_space)
2090 if (free_space < (full_space >> 1)) {
2091 icsk->icsk_ack.quick = 0;
2093 if (sk_under_memory_pressure(sk))
2094 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2097 if (free_space < mss)
2101 if (free_space > tp->rcv_ssthresh)
2102 free_space = tp->rcv_ssthresh;
2104 /* Don't do rounding if we are using window scaling, since the
2105 * scaled window will not line up with the MSS boundary anyway.
2107 window = tp->rcv_wnd;
2108 if (tp->rx_opt.rcv_wscale) {
2109 window = free_space;
2111 /* Advertise enough space so that it won't get scaled away.
2112 * Import case: prevent zero window announcement if
2113 * 1<<rcv_wscale > mss.
2115 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2116 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2117 << tp->rx_opt.rcv_wscale);
2119 /* Get the largest window that is a nice multiple of mss.
2120 * Window clamp already applied above.
2121 * If our current window offering is within 1 mss of the
2122 * free space we just keep it. This prevents the divide
2123 * and multiply from happening most of the time.
2124 * We also don't do any window rounding when the free space
2127 if (window <= free_space - mss || window > free_space)
2128 window = (free_space / mss) * mss;
2129 else if (mss == full_space &&
2130 free_space > window + (full_space >> 1))
2131 window = free_space;
2137 /* Collapses two adjacent SKB's during retransmission. */
2138 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2140 struct tcp_sock *tp = tcp_sk(sk);
2141 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2142 int skb_size, next_skb_size;
2144 skb_size = skb->len;
2145 next_skb_size = next_skb->len;
2147 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2149 tcp_highest_sack_combine(sk, next_skb, skb);
2151 tcp_unlink_write_queue(next_skb, sk);
2153 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2156 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2157 skb->ip_summed = CHECKSUM_PARTIAL;
2159 if (skb->ip_summed != CHECKSUM_PARTIAL)
2160 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2162 /* Update sequence range on original skb. */
2163 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2165 /* Merge over control information. This moves PSH/FIN etc. over */
2166 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2168 /* All done, get rid of second SKB and account for it so
2169 * packet counting does not break.
2171 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2173 /* changed transmit queue under us so clear hints */
2174 tcp_clear_retrans_hints_partial(tp);
2175 if (next_skb == tp->retransmit_skb_hint)
2176 tp->retransmit_skb_hint = skb;
2178 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2180 sk_wmem_free_skb(sk, next_skb);
2183 /* Check if coalescing SKBs is legal. */
2184 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2186 if (tcp_skb_pcount(skb) > 1)
2188 /* TODO: SACK collapsing could be used to remove this condition */
2189 if (skb_shinfo(skb)->nr_frags != 0)
2191 if (skb_cloned(skb))
2193 if (skb == tcp_send_head(sk))
2195 /* Some heurestics for collapsing over SACK'd could be invented */
2196 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2202 /* Collapse packets in the retransmit queue to make to create
2203 * less packets on the wire. This is only done on retransmission.
2205 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2208 struct tcp_sock *tp = tcp_sk(sk);
2209 struct sk_buff *skb = to, *tmp;
2212 if (!sysctl_tcp_retrans_collapse)
2214 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2217 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2218 if (!tcp_can_collapse(sk, skb))
2230 /* Punt if not enough space exists in the first SKB for
2231 * the data in the second
2233 if (skb->len > skb_availroom(to))
2236 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2239 tcp_collapse_retrans(sk, to);
2243 /* This retransmits one SKB. Policy decisions and retransmit queue
2244 * state updates are done by the caller. Returns non-zero if an
2245 * error occurred which prevented the send.
2247 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2249 struct tcp_sock *tp = tcp_sk(sk);
2250 struct inet_connection_sock *icsk = inet_csk(sk);
2251 unsigned int cur_mss;
2254 /* Inconslusive MTU probe */
2255 if (icsk->icsk_mtup.probe_size) {
2256 icsk->icsk_mtup.probe_size = 0;
2259 /* Do not sent more than we queued. 1/4 is reserved for possible
2260 * copying overhead: fragmentation, tunneling, mangling etc.
2262 if (atomic_read(&sk->sk_wmem_alloc) >
2263 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2266 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2267 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2269 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2273 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2274 return -EHOSTUNREACH; /* Routing failure or similar. */
2276 cur_mss = tcp_current_mss(sk);
2278 /* If receiver has shrunk his window, and skb is out of
2279 * new window, do not retransmit it. The exception is the
2280 * case, when window is shrunk to zero. In this case
2281 * our retransmit serves as a zero window probe.
2283 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2284 TCP_SKB_CB(skb)->seq != tp->snd_una)
2287 if (skb->len > cur_mss) {
2288 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2289 return -ENOMEM; /* We'll try again later. */
2291 int oldpcount = tcp_skb_pcount(skb);
2293 if (unlikely(oldpcount > 1)) {
2294 tcp_init_tso_segs(sk, skb, cur_mss);
2295 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2299 tcp_retrans_try_collapse(sk, skb, cur_mss);
2301 /* Some Solaris stacks overoptimize and ignore the FIN on a
2302 * retransmit when old data is attached. So strip it off
2303 * since it is cheap to do so and saves bytes on the network.
2306 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2307 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2308 if (!pskb_trim(skb, 0)) {
2309 /* Reuse, even though it does some unnecessary work */
2310 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2311 TCP_SKB_CB(skb)->tcp_flags);
2312 skb->ip_summed = CHECKSUM_NONE;
2316 /* Make a copy, if the first transmission SKB clone we made
2317 * is still in somebody's hands, else make a clone.
2319 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2321 /* make sure skb->data is aligned on arches that require it */
2322 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2323 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2325 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2328 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2332 /* Update global TCP statistics. */
2333 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2335 tp->total_retrans++;
2337 #if FASTRETRANS_DEBUG > 0
2338 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2339 net_dbg_ratelimited("retrans_out leaked\n");
2342 if (!tp->retrans_out)
2343 tp->lost_retrans_low = tp->snd_nxt;
2344 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2345 tp->retrans_out += tcp_skb_pcount(skb);
2347 /* Save stamp of the first retransmit. */
2348 if (!tp->retrans_stamp)
2349 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2351 tp->undo_retrans += tcp_skb_pcount(skb);
2353 /* snd_nxt is stored to detect loss of retransmitted segment,
2354 * see tcp_input.c tcp_sacktag_write_queue().
2356 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2361 /* Check if we forward retransmits are possible in the current
2362 * window/congestion state.
2364 static bool tcp_can_forward_retransmit(struct sock *sk)
2366 const struct inet_connection_sock *icsk = inet_csk(sk);
2367 const struct tcp_sock *tp = tcp_sk(sk);
2369 /* Forward retransmissions are possible only during Recovery. */
2370 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2373 /* No forward retransmissions in Reno are possible. */
2374 if (tcp_is_reno(tp))
2377 /* Yeah, we have to make difficult choice between forward transmission
2378 * and retransmission... Both ways have their merits...
2380 * For now we do not retransmit anything, while we have some new
2381 * segments to send. In the other cases, follow rule 3 for
2382 * NextSeg() specified in RFC3517.
2385 if (tcp_may_send_now(sk))
2391 /* This gets called after a retransmit timeout, and the initially
2392 * retransmitted data is acknowledged. It tries to continue
2393 * resending the rest of the retransmit queue, until either
2394 * we've sent it all or the congestion window limit is reached.
2395 * If doing SACK, the first ACK which comes back for a timeout
2396 * based retransmit packet might feed us FACK information again.
2397 * If so, we use it to avoid unnecessarily retransmissions.
2399 void tcp_xmit_retransmit_queue(struct sock *sk)
2401 const struct inet_connection_sock *icsk = inet_csk(sk);
2402 struct tcp_sock *tp = tcp_sk(sk);
2403 struct sk_buff *skb;
2404 struct sk_buff *hole = NULL;
2407 int fwd_rexmitting = 0;
2409 if (!tp->packets_out)
2413 tp->retransmit_high = tp->snd_una;
2415 if (tp->retransmit_skb_hint) {
2416 skb = tp->retransmit_skb_hint;
2417 last_lost = TCP_SKB_CB(skb)->end_seq;
2418 if (after(last_lost, tp->retransmit_high))
2419 last_lost = tp->retransmit_high;
2421 skb = tcp_write_queue_head(sk);
2422 last_lost = tp->snd_una;
2425 tcp_for_write_queue_from(skb, sk) {
2426 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2428 if (skb == tcp_send_head(sk))
2430 /* we could do better than to assign each time */
2432 tp->retransmit_skb_hint = skb;
2434 /* Assume this retransmit will generate
2435 * only one packet for congestion window
2436 * calculation purposes. This works because
2437 * tcp_retransmit_skb() will chop up the
2438 * packet to be MSS sized and all the
2439 * packet counting works out.
2441 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2444 if (fwd_rexmitting) {
2446 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2448 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2450 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2451 tp->retransmit_high = last_lost;
2452 if (!tcp_can_forward_retransmit(sk))
2454 /* Backtrack if necessary to non-L'ed skb */
2462 } else if (!(sacked & TCPCB_LOST)) {
2463 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2468 last_lost = TCP_SKB_CB(skb)->end_seq;
2469 if (icsk->icsk_ca_state != TCP_CA_Loss)
2470 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2472 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2475 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2478 if (tcp_retransmit_skb(sk, skb)) {
2479 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2482 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2484 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2485 tp->prr_out += tcp_skb_pcount(skb);
2487 if (skb == tcp_write_queue_head(sk))
2488 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2489 inet_csk(sk)->icsk_rto,
2494 /* Send a fin. The caller locks the socket for us. This cannot be
2495 * allowed to fail queueing a FIN frame under any circumstances.
2497 void tcp_send_fin(struct sock *sk)
2499 struct tcp_sock *tp = tcp_sk(sk);
2500 struct sk_buff *skb = tcp_write_queue_tail(sk);
2503 /* Optimization, tack on the FIN if we have a queue of
2504 * unsent frames. But be careful about outgoing SACKS
2507 mss_now = tcp_current_mss(sk);
2509 if (tcp_send_head(sk) != NULL) {
2510 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2511 TCP_SKB_CB(skb)->end_seq++;
2514 /* Socket is locked, keep trying until memory is available. */
2516 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2523 /* Reserve space for headers and prepare control bits. */
2524 skb_reserve(skb, MAX_TCP_HEADER);
2525 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2526 tcp_init_nondata_skb(skb, tp->write_seq,
2527 TCPHDR_ACK | TCPHDR_FIN);
2528 tcp_queue_skb(sk, skb);
2530 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2533 /* We get here when a process closes a file descriptor (either due to
2534 * an explicit close() or as a byproduct of exit()'ing) and there
2535 * was unread data in the receive queue. This behavior is recommended
2536 * by RFC 2525, section 2.17. -DaveM
2538 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2540 struct sk_buff *skb;
2542 /* NOTE: No TCP options attached and we never retransmit this. */
2543 skb = alloc_skb(MAX_TCP_HEADER, priority);
2545 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2549 /* Reserve space for headers and prepare control bits. */
2550 skb_reserve(skb, MAX_TCP_HEADER);
2551 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2552 TCPHDR_ACK | TCPHDR_RST);
2554 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2555 if (tcp_transmit_skb(sk, skb, 0, priority))
2556 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2558 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2561 /* Send a crossed SYN-ACK during socket establishment.
2562 * WARNING: This routine must only be called when we have already sent
2563 * a SYN packet that crossed the incoming SYN that caused this routine
2564 * to get called. If this assumption fails then the initial rcv_wnd
2565 * and rcv_wscale values will not be correct.
2567 int tcp_send_synack(struct sock *sk)
2569 struct sk_buff *skb;
2571 skb = tcp_write_queue_head(sk);
2572 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2573 pr_debug("%s: wrong queue state\n", __func__);
2576 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2577 if (skb_cloned(skb)) {
2578 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2581 tcp_unlink_write_queue(skb, sk);
2582 skb_header_release(nskb);
2583 __tcp_add_write_queue_head(sk, nskb);
2584 sk_wmem_free_skb(sk, skb);
2585 sk->sk_wmem_queued += nskb->truesize;
2586 sk_mem_charge(sk, nskb->truesize);
2590 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2591 TCP_ECN_send_synack(tcp_sk(sk), skb);
2593 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2594 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2598 * tcp_make_synack - Prepare a SYN-ACK.
2599 * sk: listener socket
2600 * dst: dst entry attached to the SYNACK
2601 * req: request_sock pointer
2602 * rvp: request_values pointer
2604 * Allocate one skb and build a SYNACK packet.
2605 * @dst is consumed : Caller should not use it again.
2607 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2608 struct request_sock *req,
2609 struct request_values *rvp)
2611 struct tcp_out_options opts;
2612 struct tcp_extend_values *xvp = tcp_xv(rvp);
2613 struct inet_request_sock *ireq = inet_rsk(req);
2614 struct tcp_sock *tp = tcp_sk(sk);
2615 const struct tcp_cookie_values *cvp = tp->cookie_values;
2617 struct sk_buff *skb;
2618 struct tcp_md5sig_key *md5;
2619 int tcp_header_size;
2621 int s_data_desired = 0;
2623 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2624 s_data_desired = cvp->s_data_desired;
2625 skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, GFP_ATOMIC);
2626 if (unlikely(!skb)) {
2630 /* Reserve space for headers. */
2631 skb_reserve(skb, MAX_TCP_HEADER);
2633 skb_dst_set(skb, dst);
2635 mss = dst_metric_advmss(dst);
2636 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2637 mss = tp->rx_opt.user_mss;
2639 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2641 /* Set this up on the first call only */
2642 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2644 /* limit the window selection if the user enforce a smaller rx buffer */
2645 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2646 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2647 req->window_clamp = tcp_full_space(sk);
2649 /* tcp_full_space because it is guaranteed to be the first packet */
2650 tcp_select_initial_window(tcp_full_space(sk),
2651 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2656 dst_metric(dst, RTAX_INITRWND));
2657 ireq->rcv_wscale = rcv_wscale;
2660 memset(&opts, 0, sizeof(opts));
2661 #ifdef CONFIG_SYN_COOKIES
2662 if (unlikely(req->cookie_ts))
2663 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2666 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2667 tcp_header_size = tcp_synack_options(sk, req, mss,
2668 skb, &opts, &md5, xvp)
2671 skb_push(skb, tcp_header_size);
2672 skb_reset_transport_header(skb);
2675 memset(th, 0, sizeof(struct tcphdr));
2678 TCP_ECN_make_synack(req, th);
2679 th->source = ireq->loc_port;
2680 th->dest = ireq->rmt_port;
2681 /* Setting of flags are superfluous here for callers (and ECE is
2682 * not even correctly set)
2684 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2685 TCPHDR_SYN | TCPHDR_ACK);
2687 if (OPTION_COOKIE_EXTENSION & opts.options) {
2688 if (s_data_desired) {
2689 u8 *buf = skb_put(skb, s_data_desired);
2691 /* copy data directly from the listening socket. */
2692 memcpy(buf, cvp->s_data_payload, s_data_desired);
2693 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2696 if (opts.hash_size > 0) {
2697 __u32 workspace[SHA_WORKSPACE_WORDS];
2698 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2699 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2701 /* Secret recipe depends on the Timestamp, (future)
2702 * Sequence and Acknowledgment Numbers, Initiator
2703 * Cookie, and others handled by IP variant caller.
2705 *tail-- ^= opts.tsval;
2706 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2707 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2710 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2711 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2713 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2716 opts.hash_location =
2717 (__u8 *)&xvp->cookie_bakery[0];
2721 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2722 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2724 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2725 th->window = htons(min(req->rcv_wnd, 65535U));
2726 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2727 th->doff = (tcp_header_size >> 2);
2728 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2730 #ifdef CONFIG_TCP_MD5SIG
2731 /* Okay, we have all we need - do the md5 hash if needed */
2733 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2734 md5, NULL, req, skb);
2740 EXPORT_SYMBOL(tcp_make_synack);
2742 /* Do all connect socket setups that can be done AF independent. */
2743 void tcp_connect_init(struct sock *sk)
2745 const struct dst_entry *dst = __sk_dst_get(sk);
2746 struct tcp_sock *tp = tcp_sk(sk);
2749 /* We'll fix this up when we get a response from the other end.
2750 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2752 tp->tcp_header_len = sizeof(struct tcphdr) +
2753 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2755 #ifdef CONFIG_TCP_MD5SIG
2756 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2757 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2760 /* If user gave his TCP_MAXSEG, record it to clamp */
2761 if (tp->rx_opt.user_mss)
2762 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2765 tcp_sync_mss(sk, dst_mtu(dst));
2767 if (!tp->window_clamp)
2768 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2769 tp->advmss = dst_metric_advmss(dst);
2770 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2771 tp->advmss = tp->rx_opt.user_mss;
2773 tcp_initialize_rcv_mss(sk);
2775 /* limit the window selection if the user enforce a smaller rx buffer */
2776 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2777 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2778 tp->window_clamp = tcp_full_space(sk);
2780 tcp_select_initial_window(tcp_full_space(sk),
2781 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2784 sysctl_tcp_window_scaling,
2786 dst_metric(dst, RTAX_INITRWND));
2788 tp->rx_opt.rcv_wscale = rcv_wscale;
2789 tp->rcv_ssthresh = tp->rcv_wnd;
2792 sock_reset_flag(sk, SOCK_DONE);
2795 tp->snd_una = tp->write_seq;
2796 tp->snd_sml = tp->write_seq;
2797 tp->snd_up = tp->write_seq;
2798 tp->snd_nxt = tp->write_seq;
2800 if (likely(!tp->repair))
2802 tp->rcv_wup = tp->rcv_nxt;
2803 tp->copied_seq = tp->rcv_nxt;
2805 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2806 inet_csk(sk)->icsk_retransmits = 0;
2807 tcp_clear_retrans(tp);
2810 /* Build a SYN and send it off. */
2811 int tcp_connect(struct sock *sk)
2813 struct tcp_sock *tp = tcp_sk(sk);
2814 struct sk_buff *buff;
2817 tcp_connect_init(sk);
2819 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2820 if (unlikely(buff == NULL))
2823 /* Reserve space for headers. */
2824 skb_reserve(buff, MAX_TCP_HEADER);
2826 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2827 TCP_ECN_send_syn(sk, buff);
2830 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2831 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2832 skb_header_release(buff);
2833 __tcp_add_write_queue_tail(sk, buff);
2834 sk->sk_wmem_queued += buff->truesize;
2835 sk_mem_charge(sk, buff->truesize);
2836 tp->packets_out += tcp_skb_pcount(buff);
2837 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2838 if (err == -ECONNREFUSED)
2841 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2842 * in order to make this packet get counted in tcpOutSegs.
2844 tp->snd_nxt = tp->write_seq;
2845 tp->pushed_seq = tp->write_seq;
2846 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2848 /* Timer for repeating the SYN until an answer. */
2849 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2850 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2853 EXPORT_SYMBOL(tcp_connect);
2855 /* Send out a delayed ack, the caller does the policy checking
2856 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2859 void tcp_send_delayed_ack(struct sock *sk)
2861 struct inet_connection_sock *icsk = inet_csk(sk);
2862 int ato = icsk->icsk_ack.ato;
2863 unsigned long timeout;
2865 if (ato > TCP_DELACK_MIN) {
2866 const struct tcp_sock *tp = tcp_sk(sk);
2867 int max_ato = HZ / 2;
2869 if (icsk->icsk_ack.pingpong ||
2870 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2871 max_ato = TCP_DELACK_MAX;
2873 /* Slow path, intersegment interval is "high". */
2875 /* If some rtt estimate is known, use it to bound delayed ack.
2876 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2880 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2886 ato = min(ato, max_ato);
2889 /* Stay within the limit we were given */
2890 timeout = jiffies + ato;
2892 /* Use new timeout only if there wasn't a older one earlier. */
2893 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2894 /* If delack timer was blocked or is about to expire,
2897 if (icsk->icsk_ack.blocked ||
2898 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2903 if (!time_before(timeout, icsk->icsk_ack.timeout))
2904 timeout = icsk->icsk_ack.timeout;
2906 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2907 icsk->icsk_ack.timeout = timeout;
2908 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2911 /* This routine sends an ack and also updates the window. */
2912 void tcp_send_ack(struct sock *sk)
2914 struct sk_buff *buff;
2916 /* If we have been reset, we may not send again. */
2917 if (sk->sk_state == TCP_CLOSE)
2920 /* We are not putting this on the write queue, so
2921 * tcp_transmit_skb() will set the ownership to this
2924 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2926 inet_csk_schedule_ack(sk);
2927 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2928 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2929 TCP_DELACK_MAX, TCP_RTO_MAX);
2933 /* Reserve space for headers and prepare control bits. */
2934 skb_reserve(buff, MAX_TCP_HEADER);
2935 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2937 /* Send it off, this clears delayed acks for us. */
2938 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2939 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2942 /* This routine sends a packet with an out of date sequence
2943 * number. It assumes the other end will try to ack it.
2945 * Question: what should we make while urgent mode?
2946 * 4.4BSD forces sending single byte of data. We cannot send
2947 * out of window data, because we have SND.NXT==SND.MAX...
2949 * Current solution: to send TWO zero-length segments in urgent mode:
2950 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2951 * out-of-date with SND.UNA-1 to probe window.
2953 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2955 struct tcp_sock *tp = tcp_sk(sk);
2956 struct sk_buff *skb;
2958 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2959 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2963 /* Reserve space for headers and set control bits. */
2964 skb_reserve(skb, MAX_TCP_HEADER);
2965 /* Use a previous sequence. This should cause the other
2966 * end to send an ack. Don't queue or clone SKB, just
2969 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2970 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2971 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2974 void tcp_send_window_probe(struct sock *sk)
2976 if (sk->sk_state == TCP_ESTABLISHED) {
2977 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
2978 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
2979 tcp_xmit_probe_skb(sk, 0);
2983 /* Initiate keepalive or window probe from timer. */
2984 int tcp_write_wakeup(struct sock *sk)
2986 struct tcp_sock *tp = tcp_sk(sk);
2987 struct sk_buff *skb;
2989 if (sk->sk_state == TCP_CLOSE)
2992 if ((skb = tcp_send_head(sk)) != NULL &&
2993 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2995 unsigned int mss = tcp_current_mss(sk);
2996 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2998 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2999 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3001 /* We are probing the opening of a window
3002 * but the window size is != 0
3003 * must have been a result SWS avoidance ( sender )
3005 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3007 seg_size = min(seg_size, mss);
3008 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3009 if (tcp_fragment(sk, skb, seg_size, mss))
3011 } else if (!tcp_skb_pcount(skb))
3012 tcp_set_skb_tso_segs(sk, skb, mss);
3014 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3015 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3016 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3018 tcp_event_new_data_sent(sk, skb);
3021 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3022 tcp_xmit_probe_skb(sk, 1);
3023 return tcp_xmit_probe_skb(sk, 0);
3027 /* A window probe timeout has occurred. If window is not closed send
3028 * a partial packet else a zero probe.
3030 void tcp_send_probe0(struct sock *sk)
3032 struct inet_connection_sock *icsk = inet_csk(sk);
3033 struct tcp_sock *tp = tcp_sk(sk);
3036 err = tcp_write_wakeup(sk);
3038 if (tp->packets_out || !tcp_send_head(sk)) {
3039 /* Cancel probe timer, if it is not required. */
3040 icsk->icsk_probes_out = 0;
3041 icsk->icsk_backoff = 0;
3046 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3047 icsk->icsk_backoff++;
3048 icsk->icsk_probes_out++;
3049 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3050 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3053 /* If packet was not sent due to local congestion,
3054 * do not backoff and do not remember icsk_probes_out.
3055 * Let local senders to fight for local resources.
3057 * Use accumulated backoff yet.
3059 if (!icsk->icsk_probes_out)
3060 icsk->icsk_probes_out = 1;
3061 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3062 min(icsk->icsk_rto << icsk->icsk_backoff,
3063 TCP_RESOURCE_PROBE_INTERVAL),