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)
388 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
390 struct tcp_out_options {
391 u16 options; /* bit field of OPTION_* */
392 u16 mss; /* 0 to disable */
393 u8 ws; /* window scale, 0 to disable */
394 u8 num_sack_blocks; /* number of SACK blocks to include */
395 u8 hash_size; /* bytes in hash_location */
396 __u8 *hash_location; /* temporary pointer, overloaded */
397 __u32 tsval, tsecr; /* need to include OPTION_TS */
398 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
401 /* The sysctl int routines are generic, so check consistency here.
403 static u8 tcp_cookie_size_check(u8 desired)
408 /* previously specified */
411 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
412 if (cookie_size <= 0)
413 /* no default specified */
416 if (cookie_size <= TCP_COOKIE_MIN)
417 /* value too small, specify minimum */
418 return TCP_COOKIE_MIN;
420 if (cookie_size >= TCP_COOKIE_MAX)
421 /* value too large, specify maximum */
422 return TCP_COOKIE_MAX;
425 /* 8-bit multiple, illegal, fix it */
428 return (u8)cookie_size;
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operatibility perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
447 u16 options = opts->options; /* mungable copy */
449 /* Having both authentication and cookies for security is redundant,
450 * and there's certainly not enough room. Instead, the cookie-less
451 * extension variant is proposed.
453 * Consider the pessimal case with authentication. The options
455 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
457 if (unlikely(OPTION_MD5 & options)) {
458 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
459 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
460 (TCPOLEN_COOKIE_BASE << 16) |
461 (TCPOPT_MD5SIG << 8) |
464 *ptr++ = htonl((TCPOPT_NOP << 24) |
466 (TCPOPT_MD5SIG << 8) |
469 options &= ~OPTION_COOKIE_EXTENSION;
470 /* overload cookie hash location */
471 opts->hash_location = (__u8 *)ptr;
475 if (unlikely(opts->mss)) {
476 *ptr++ = htonl((TCPOPT_MSS << 24) |
477 (TCPOLEN_MSS << 16) |
481 if (likely(OPTION_TS & options)) {
482 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
483 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
484 (TCPOLEN_SACK_PERM << 16) |
485 (TCPOPT_TIMESTAMP << 8) |
487 options &= ~OPTION_SACK_ADVERTISE;
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_TIMESTAMP << 8) |
494 *ptr++ = htonl(opts->tsval);
495 *ptr++ = htonl(opts->tsecr);
498 /* Specification requires after timestamp, so do it now.
500 * Consider the pessimal case without authentication. The options
502 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
504 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
505 __u8 *cookie_copy = opts->hash_location;
506 u8 cookie_size = opts->hash_size;
508 /* 8-bit multiple handled in tcp_cookie_size_check() above,
511 if (0x2 & cookie_size) {
512 __u8 *p = (__u8 *)ptr;
514 /* 16-bit multiple */
515 *p++ = TCPOPT_COOKIE;
516 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
517 *p++ = *cookie_copy++;
518 *p++ = *cookie_copy++;
522 /* 32-bit multiple */
523 *ptr++ = htonl(((TCPOPT_NOP << 24) |
525 (TCPOPT_COOKIE << 8) |
526 TCPOLEN_COOKIE_BASE) +
530 if (cookie_size > 0) {
531 memcpy(ptr, cookie_copy, cookie_size);
532 ptr += (cookie_size / 4);
536 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
537 *ptr++ = htonl((TCPOPT_NOP << 24) |
539 (TCPOPT_SACK_PERM << 8) |
543 if (unlikely(OPTION_WSCALE & options)) {
544 *ptr++ = htonl((TCPOPT_NOP << 24) |
545 (TCPOPT_WINDOW << 16) |
546 (TCPOLEN_WINDOW << 8) |
550 if (unlikely(opts->num_sack_blocks)) {
551 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
552 tp->duplicate_sack : tp->selective_acks;
555 *ptr++ = htonl((TCPOPT_NOP << 24) |
558 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
559 TCPOLEN_SACK_PERBLOCK)));
561 for (this_sack = 0; this_sack < opts->num_sack_blocks;
563 *ptr++ = htonl(sp[this_sack].start_seq);
564 *ptr++ = htonl(sp[this_sack].end_seq);
567 tp->rx_opt.dsack = 0;
570 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
571 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
573 *ptr++ = htonl((TCPOPT_EXP << 24) |
574 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
575 TCPOPT_FASTOPEN_MAGIC);
577 memcpy(ptr, foc->val, foc->len);
578 if ((foc->len & 3) == 2) {
579 u8 *align = ((u8 *)ptr) + foc->len;
580 align[0] = align[1] = TCPOPT_NOP;
582 ptr += (foc->len + 3) >> 2;
586 /* Compute TCP options for SYN packets. This is not the final
587 * network wire format yet.
589 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
590 struct tcp_out_options *opts,
591 struct tcp_md5sig_key **md5)
593 struct tcp_sock *tp = tcp_sk(sk);
594 struct tcp_cookie_values *cvp = tp->cookie_values;
595 unsigned int remaining = MAX_TCP_OPTION_SPACE;
596 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
597 tcp_cookie_size_check(cvp->cookie_desired) :
599 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
601 #ifdef CONFIG_TCP_MD5SIG
602 *md5 = tp->af_specific->md5_lookup(sk, sk);
604 opts->options |= OPTION_MD5;
605 remaining -= TCPOLEN_MD5SIG_ALIGNED;
611 /* We always get an MSS option. The option bytes which will be seen in
612 * normal data packets should timestamps be used, must be in the MSS
613 * advertised. But we subtract them from tp->mss_cache so that
614 * calculations in tcp_sendmsg are simpler etc. So account for this
615 * fact here if necessary. If we don't do this correctly, as a
616 * receiver we won't recognize data packets as being full sized when we
617 * should, and thus we won't abide by the delayed ACK rules correctly.
618 * SACKs don't matter, we never delay an ACK when we have any of those
620 opts->mss = tcp_advertise_mss(sk);
621 remaining -= TCPOLEN_MSS_ALIGNED;
623 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
624 opts->options |= OPTION_TS;
625 opts->tsval = TCP_SKB_CB(skb)->when;
626 opts->tsecr = tp->rx_opt.ts_recent;
627 remaining -= TCPOLEN_TSTAMP_ALIGNED;
629 if (likely(sysctl_tcp_window_scaling)) {
630 opts->ws = tp->rx_opt.rcv_wscale;
631 opts->options |= OPTION_WSCALE;
632 remaining -= TCPOLEN_WSCALE_ALIGNED;
634 if (likely(sysctl_tcp_sack)) {
635 opts->options |= OPTION_SACK_ADVERTISE;
636 if (unlikely(!(OPTION_TS & opts->options)))
637 remaining -= TCPOLEN_SACKPERM_ALIGNED;
640 if (fastopen && fastopen->cookie.len >= 0) {
641 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
642 need = (need + 3) & ~3U; /* Align to 32 bits */
643 if (remaining >= need) {
644 opts->options |= OPTION_FAST_OPEN_COOKIE;
645 opts->fastopen_cookie = &fastopen->cookie;
647 tp->syn_fastopen = 1;
650 /* Note that timestamps are required by the specification.
652 * Odd numbers of bytes are prohibited by the specification, ensuring
653 * that the cookie is 16-bit aligned, and the resulting cookie pair is
657 (OPTION_TS & opts->options) &&
659 int need = TCPOLEN_COOKIE_BASE + cookie_size;
662 /* 32-bit multiple */
663 need += 2; /* NOPs */
665 if (need > remaining) {
666 /* try shrinking cookie to fit */
671 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
675 if (TCP_COOKIE_MIN <= cookie_size) {
676 opts->options |= OPTION_COOKIE_EXTENSION;
677 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
678 opts->hash_size = cookie_size;
680 /* Remember for future incarnations. */
681 cvp->cookie_desired = cookie_size;
683 if (cvp->cookie_desired != cvp->cookie_pair_size) {
684 /* Currently use random bytes as a nonce,
685 * assuming these are completely unpredictable
686 * by hostile users of the same system.
688 get_random_bytes(&cvp->cookie_pair[0],
690 cvp->cookie_pair_size = cookie_size;
696 return MAX_TCP_OPTION_SPACE - remaining;
699 /* Set up TCP options for SYN-ACKs. */
700 static unsigned int tcp_synack_options(struct sock *sk,
701 struct request_sock *req,
702 unsigned int mss, struct sk_buff *skb,
703 struct tcp_out_options *opts,
704 struct tcp_md5sig_key **md5,
705 struct tcp_extend_values *xvp)
707 struct inet_request_sock *ireq = inet_rsk(req);
708 unsigned int remaining = MAX_TCP_OPTION_SPACE;
709 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
713 #ifdef CONFIG_TCP_MD5SIG
714 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
716 opts->options |= OPTION_MD5;
717 remaining -= TCPOLEN_MD5SIG_ALIGNED;
719 /* We can't fit any SACK blocks in a packet with MD5 + TS
720 * options. There was discussion about disabling SACK
721 * rather than TS in order to fit in better with old,
722 * buggy kernels, but that was deemed to be unnecessary.
724 ireq->tstamp_ok &= !ireq->sack_ok;
730 /* We always send an MSS option. */
732 remaining -= TCPOLEN_MSS_ALIGNED;
734 if (likely(ireq->wscale_ok)) {
735 opts->ws = ireq->rcv_wscale;
736 opts->options |= OPTION_WSCALE;
737 remaining -= TCPOLEN_WSCALE_ALIGNED;
739 if (likely(ireq->tstamp_ok)) {
740 opts->options |= OPTION_TS;
741 opts->tsval = TCP_SKB_CB(skb)->when;
742 opts->tsecr = req->ts_recent;
743 remaining -= TCPOLEN_TSTAMP_ALIGNED;
745 if (likely(ireq->sack_ok)) {
746 opts->options |= OPTION_SACK_ADVERTISE;
747 if (unlikely(!ireq->tstamp_ok))
748 remaining -= TCPOLEN_SACKPERM_ALIGNED;
751 /* Similar rationale to tcp_syn_options() applies here, too.
752 * If the <SYN> options fit, the same options should fit now!
756 cookie_plus > TCPOLEN_COOKIE_BASE) {
757 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
760 /* 32-bit multiple */
761 need += 2; /* NOPs */
763 if (need <= remaining) {
764 opts->options |= OPTION_COOKIE_EXTENSION;
765 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
768 /* There's no error return, so flag it. */
769 xvp->cookie_out_never = 1; /* true */
773 return MAX_TCP_OPTION_SPACE - remaining;
776 /* Compute TCP options for ESTABLISHED sockets. This is not the
777 * final wire format yet.
779 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
780 struct tcp_out_options *opts,
781 struct tcp_md5sig_key **md5)
783 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
784 struct tcp_sock *tp = tcp_sk(sk);
785 unsigned int size = 0;
786 unsigned int eff_sacks;
788 #ifdef CONFIG_TCP_MD5SIG
789 *md5 = tp->af_specific->md5_lookup(sk, sk);
790 if (unlikely(*md5)) {
791 opts->options |= OPTION_MD5;
792 size += TCPOLEN_MD5SIG_ALIGNED;
798 if (likely(tp->rx_opt.tstamp_ok)) {
799 opts->options |= OPTION_TS;
800 opts->tsval = tcb ? tcb->when : 0;
801 opts->tsecr = tp->rx_opt.ts_recent;
802 size += TCPOLEN_TSTAMP_ALIGNED;
805 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
806 if (unlikely(eff_sacks)) {
807 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
808 opts->num_sack_blocks =
809 min_t(unsigned int, eff_sacks,
810 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
811 TCPOLEN_SACK_PERBLOCK);
812 size += TCPOLEN_SACK_BASE_ALIGNED +
813 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
820 /* TCP SMALL QUEUES (TSQ)
822 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
823 * to reduce RTT and bufferbloat.
824 * We do this using a special skb destructor (tcp_wfree).
826 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
827 * needs to be reallocated in a driver.
828 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
830 * Since transmit from skb destructor is forbidden, we use a tasklet
831 * to process all sockets that eventually need to send more skbs.
832 * We use one tasklet per cpu, with its own queue of sockets.
835 struct tasklet_struct tasklet;
836 struct list_head head; /* queue of tcp sockets */
838 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
841 * One tasklest per cpu tries to send more skbs.
842 * We run in tasklet context but need to disable irqs when
843 * transfering tsq->head because tcp_wfree() might
844 * interrupt us (non NAPI drivers)
846 static void tcp_tasklet_func(unsigned long data)
848 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
851 struct list_head *q, *n;
855 local_irq_save(flags);
856 list_splice_init(&tsq->head, &list);
857 local_irq_restore(flags);
859 list_for_each_safe(q, n, &list) {
860 tp = list_entry(q, struct tcp_sock, tsq_node);
861 list_del(&tp->tsq_node);
863 sk = (struct sock *)tp;
866 if (!sock_owned_by_user(sk)) {
867 if ((1 << sk->sk_state) &
868 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 |
869 TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
875 /* defer the work to tcp_release_cb() */
876 set_bit(TSQ_OWNED, &tp->tsq_flags);
880 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
886 * tcp_release_cb - tcp release_sock() callback
889 * called from release_sock() to perform protocol dependent
890 * actions before socket release.
892 void tcp_release_cb(struct sock *sk)
894 struct tcp_sock *tp = tcp_sk(sk);
896 if (test_and_clear_bit(TSQ_OWNED, &tp->tsq_flags)) {
897 if ((1 << sk->sk_state) &
898 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 |
899 TCPF_CLOSING | TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
906 EXPORT_SYMBOL(tcp_release_cb);
908 void __init tcp_tasklet_init(void)
912 for_each_possible_cpu(i) {
913 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
915 INIT_LIST_HEAD(&tsq->head);
916 tasklet_init(&tsq->tasklet,
923 * Write buffer destructor automatically called from kfree_skb.
924 * We cant xmit new skbs from this context, as we might already
927 void tcp_wfree(struct sk_buff *skb)
929 struct sock *sk = skb->sk;
930 struct tcp_sock *tp = tcp_sk(sk);
932 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
933 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
935 struct tsq_tasklet *tsq;
937 /* Keep a ref on socket.
938 * This last ref will be released in tcp_tasklet_func()
940 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
942 /* queue this socket to tasklet queue */
943 local_irq_save(flags);
944 tsq = &__get_cpu_var(tsq_tasklet);
945 list_add(&tp->tsq_node, &tsq->head);
946 tasklet_schedule(&tsq->tasklet);
947 local_irq_restore(flags);
953 /* This routine actually transmits TCP packets queued in by
954 * tcp_do_sendmsg(). This is used by both the initial
955 * transmission and possible later retransmissions.
956 * All SKB's seen here are completely headerless. It is our
957 * job to build the TCP header, and pass the packet down to
958 * IP so it can do the same plus pass the packet off to the
961 * We are working here with either a clone of the original
962 * SKB, or a fresh unique copy made by the retransmit engine.
964 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
967 const struct inet_connection_sock *icsk = inet_csk(sk);
968 struct inet_sock *inet;
970 struct tcp_skb_cb *tcb;
971 struct tcp_out_options opts;
972 unsigned int tcp_options_size, tcp_header_size;
973 struct tcp_md5sig_key *md5;
977 BUG_ON(!skb || !tcp_skb_pcount(skb));
979 /* If congestion control is doing timestamping, we must
980 * take such a timestamp before we potentially clone/copy.
982 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
983 __net_timestamp(skb);
985 if (likely(clone_it)) {
986 if (unlikely(skb_cloned(skb)))
987 skb = pskb_copy(skb, gfp_mask);
989 skb = skb_clone(skb, gfp_mask);
996 tcb = TCP_SKB_CB(skb);
997 memset(&opts, 0, sizeof(opts));
999 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1000 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1002 tcp_options_size = tcp_established_options(sk, skb, &opts,
1004 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1006 if (tcp_packets_in_flight(tp) == 0) {
1007 tcp_ca_event(sk, CA_EVENT_TX_START);
1012 skb_push(skb, tcp_header_size);
1013 skb_reset_transport_header(skb);
1017 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
1018 tcp_wfree : sock_wfree;
1019 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
1021 /* Build TCP header and checksum it. */
1023 th->source = inet->inet_sport;
1024 th->dest = inet->inet_dport;
1025 th->seq = htonl(tcb->seq);
1026 th->ack_seq = htonl(tp->rcv_nxt);
1027 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1030 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1031 /* RFC1323: The window in SYN & SYN/ACK segments
1034 th->window = htons(min(tp->rcv_wnd, 65535U));
1036 th->window = htons(tcp_select_window(sk));
1041 /* The urg_mode check is necessary during a below snd_una win probe */
1042 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1043 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1044 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1046 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1047 th->urg_ptr = htons(0xFFFF);
1052 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1053 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1054 TCP_ECN_send(sk, skb, tcp_header_size);
1056 #ifdef CONFIG_TCP_MD5SIG
1057 /* Calculate the MD5 hash, as we have all we need now */
1059 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1060 tp->af_specific->calc_md5_hash(opts.hash_location,
1061 md5, sk, NULL, skb);
1065 icsk->icsk_af_ops->send_check(sk, skb);
1067 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1068 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1070 if (skb->len != tcp_header_size)
1071 tcp_event_data_sent(tp, sk);
1073 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1074 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1075 tcp_skb_pcount(skb));
1077 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1078 if (likely(err <= 0))
1081 tcp_enter_cwr(sk, 1);
1083 return net_xmit_eval(err);
1086 /* This routine just queues the buffer for sending.
1088 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1089 * otherwise socket can stall.
1091 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1093 struct tcp_sock *tp = tcp_sk(sk);
1095 /* Advance write_seq and place onto the write_queue. */
1096 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1097 skb_header_release(skb);
1098 tcp_add_write_queue_tail(sk, skb);
1099 sk->sk_wmem_queued += skb->truesize;
1100 sk_mem_charge(sk, skb->truesize);
1103 /* Initialize TSO segments for a packet. */
1104 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1105 unsigned int mss_now)
1107 if (skb->len <= mss_now || !sk_can_gso(sk) ||
1108 skb->ip_summed == CHECKSUM_NONE) {
1109 /* Avoid the costly divide in the normal
1112 skb_shinfo(skb)->gso_segs = 1;
1113 skb_shinfo(skb)->gso_size = 0;
1114 skb_shinfo(skb)->gso_type = 0;
1116 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1117 skb_shinfo(skb)->gso_size = mss_now;
1118 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1122 /* When a modification to fackets out becomes necessary, we need to check
1123 * skb is counted to fackets_out or not.
1125 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1128 struct tcp_sock *tp = tcp_sk(sk);
1130 if (!tp->sacked_out || tcp_is_reno(tp))
1133 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1134 tp->fackets_out -= decr;
1137 /* Pcount in the middle of the write queue got changed, we need to do various
1138 * tweaks to fix counters
1140 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1142 struct tcp_sock *tp = tcp_sk(sk);
1144 tp->packets_out -= decr;
1146 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1147 tp->sacked_out -= decr;
1148 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1149 tp->retrans_out -= decr;
1150 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1151 tp->lost_out -= decr;
1153 /* Reno case is special. Sigh... */
1154 if (tcp_is_reno(tp) && decr > 0)
1155 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1157 tcp_adjust_fackets_out(sk, skb, decr);
1159 if (tp->lost_skb_hint &&
1160 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1161 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1162 tp->lost_cnt_hint -= decr;
1164 tcp_verify_left_out(tp);
1167 /* Function to create two new TCP segments. Shrinks the given segment
1168 * to the specified size and appends a new segment with the rest of the
1169 * packet to the list. This won't be called frequently, I hope.
1170 * Remember, these are still headerless SKBs at this point.
1172 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1173 unsigned int mss_now)
1175 struct tcp_sock *tp = tcp_sk(sk);
1176 struct sk_buff *buff;
1177 int nsize, old_factor;
1181 if (WARN_ON(len > skb->len))
1184 nsize = skb_headlen(skb) - len;
1188 if (skb_cloned(skb) &&
1189 skb_is_nonlinear(skb) &&
1190 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1193 /* Get a new skb... force flag on. */
1194 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1196 return -ENOMEM; /* We'll just try again later. */
1198 sk->sk_wmem_queued += buff->truesize;
1199 sk_mem_charge(sk, buff->truesize);
1200 nlen = skb->len - len - nsize;
1201 buff->truesize += nlen;
1202 skb->truesize -= nlen;
1204 /* Correct the sequence numbers. */
1205 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1206 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1207 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1209 /* PSH and FIN should only be set in the second packet. */
1210 flags = TCP_SKB_CB(skb)->tcp_flags;
1211 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1212 TCP_SKB_CB(buff)->tcp_flags = flags;
1213 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1215 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1216 /* Copy and checksum data tail into the new buffer. */
1217 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1218 skb_put(buff, nsize),
1223 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1225 skb->ip_summed = CHECKSUM_PARTIAL;
1226 skb_split(skb, buff, len);
1229 buff->ip_summed = skb->ip_summed;
1231 /* Looks stupid, but our code really uses when of
1232 * skbs, which it never sent before. --ANK
1234 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1235 buff->tstamp = skb->tstamp;
1237 old_factor = tcp_skb_pcount(skb);
1239 /* Fix up tso_factor for both original and new SKB. */
1240 tcp_set_skb_tso_segs(sk, skb, mss_now);
1241 tcp_set_skb_tso_segs(sk, buff, mss_now);
1243 /* If this packet has been sent out already, we must
1244 * adjust the various packet counters.
1246 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1247 int diff = old_factor - tcp_skb_pcount(skb) -
1248 tcp_skb_pcount(buff);
1251 tcp_adjust_pcount(sk, skb, diff);
1254 /* Link BUFF into the send queue. */
1255 skb_header_release(buff);
1256 tcp_insert_write_queue_after(skb, buff, sk);
1261 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1262 * eventually). The difference is that pulled data not copied, but
1263 * immediately discarded.
1265 static void __pskb_trim_head(struct sk_buff *skb, int len)
1269 eat = min_t(int, len, skb_headlen(skb));
1271 __skb_pull(skb, eat);
1272 skb->avail_size -= eat;
1279 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1280 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1283 skb_frag_unref(skb, i);
1286 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1288 skb_shinfo(skb)->frags[k].page_offset += eat;
1289 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1295 skb_shinfo(skb)->nr_frags = k;
1297 skb_reset_tail_pointer(skb);
1298 skb->data_len -= len;
1299 skb->len = skb->data_len;
1302 /* Remove acked data from a packet in the transmit queue. */
1303 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1305 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1308 __pskb_trim_head(skb, len);
1310 TCP_SKB_CB(skb)->seq += len;
1311 skb->ip_summed = CHECKSUM_PARTIAL;
1313 skb->truesize -= len;
1314 sk->sk_wmem_queued -= len;
1315 sk_mem_uncharge(sk, len);
1316 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1318 /* Any change of skb->len requires recalculation of tso factor. */
1319 if (tcp_skb_pcount(skb) > 1)
1320 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1325 /* Calculate MSS. Not accounting for SACKs here. */
1326 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1328 const struct tcp_sock *tp = tcp_sk(sk);
1329 const struct inet_connection_sock *icsk = inet_csk(sk);
1332 /* Calculate base mss without TCP options:
1333 It is MMS_S - sizeof(tcphdr) of rfc1122
1335 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1337 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1338 if (icsk->icsk_af_ops->net_frag_header_len) {
1339 const struct dst_entry *dst = __sk_dst_get(sk);
1341 if (dst && dst_allfrag(dst))
1342 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1345 /* Clamp it (mss_clamp does not include tcp options) */
1346 if (mss_now > tp->rx_opt.mss_clamp)
1347 mss_now = tp->rx_opt.mss_clamp;
1349 /* Now subtract optional transport overhead */
1350 mss_now -= icsk->icsk_ext_hdr_len;
1352 /* Then reserve room for full set of TCP options and 8 bytes of data */
1356 /* Now subtract TCP options size, not including SACKs */
1357 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1362 /* Inverse of above */
1363 int tcp_mss_to_mtu(struct sock *sk, int mss)
1365 const struct tcp_sock *tp = tcp_sk(sk);
1366 const struct inet_connection_sock *icsk = inet_csk(sk);
1370 tp->tcp_header_len +
1371 icsk->icsk_ext_hdr_len +
1372 icsk->icsk_af_ops->net_header_len;
1374 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1375 if (icsk->icsk_af_ops->net_frag_header_len) {
1376 const struct dst_entry *dst = __sk_dst_get(sk);
1378 if (dst && dst_allfrag(dst))
1379 mtu += icsk->icsk_af_ops->net_frag_header_len;
1384 /* MTU probing init per socket */
1385 void tcp_mtup_init(struct sock *sk)
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 struct inet_connection_sock *icsk = inet_csk(sk);
1390 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1391 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1392 icsk->icsk_af_ops->net_header_len;
1393 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1394 icsk->icsk_mtup.probe_size = 0;
1396 EXPORT_SYMBOL(tcp_mtup_init);
1398 /* This function synchronize snd mss to current pmtu/exthdr set.
1400 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1401 for TCP options, but includes only bare TCP header.
1403 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1404 It is minimum of user_mss and mss received with SYN.
1405 It also does not include TCP options.
1407 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1409 tp->mss_cache is current effective sending mss, including
1410 all tcp options except for SACKs. It is evaluated,
1411 taking into account current pmtu, but never exceeds
1412 tp->rx_opt.mss_clamp.
1414 NOTE1. rfc1122 clearly states that advertised MSS
1415 DOES NOT include either tcp or ip options.
1417 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1418 are READ ONLY outside this function. --ANK (980731)
1420 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1422 struct tcp_sock *tp = tcp_sk(sk);
1423 struct inet_connection_sock *icsk = inet_csk(sk);
1426 if (icsk->icsk_mtup.search_high > pmtu)
1427 icsk->icsk_mtup.search_high = pmtu;
1429 mss_now = tcp_mtu_to_mss(sk, pmtu);
1430 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1432 /* And store cached results */
1433 icsk->icsk_pmtu_cookie = pmtu;
1434 if (icsk->icsk_mtup.enabled)
1435 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1436 tp->mss_cache = mss_now;
1440 EXPORT_SYMBOL(tcp_sync_mss);
1442 /* Compute the current effective MSS, taking SACKs and IP options,
1443 * and even PMTU discovery events into account.
1445 unsigned int tcp_current_mss(struct sock *sk)
1447 const struct tcp_sock *tp = tcp_sk(sk);
1448 const struct dst_entry *dst = __sk_dst_get(sk);
1450 unsigned int header_len;
1451 struct tcp_out_options opts;
1452 struct tcp_md5sig_key *md5;
1454 mss_now = tp->mss_cache;
1457 u32 mtu = dst_mtu(dst);
1458 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1459 mss_now = tcp_sync_mss(sk, mtu);
1462 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1463 sizeof(struct tcphdr);
1464 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1465 * some common options. If this is an odd packet (because we have SACK
1466 * blocks etc) then our calculated header_len will be different, and
1467 * we have to adjust mss_now correspondingly */
1468 if (header_len != tp->tcp_header_len) {
1469 int delta = (int) header_len - tp->tcp_header_len;
1476 /* Congestion window validation. (RFC2861) */
1477 static void tcp_cwnd_validate(struct sock *sk)
1479 struct tcp_sock *tp = tcp_sk(sk);
1481 if (tp->packets_out >= tp->snd_cwnd) {
1482 /* Network is feed fully. */
1483 tp->snd_cwnd_used = 0;
1484 tp->snd_cwnd_stamp = tcp_time_stamp;
1486 /* Network starves. */
1487 if (tp->packets_out > tp->snd_cwnd_used)
1488 tp->snd_cwnd_used = tp->packets_out;
1490 if (sysctl_tcp_slow_start_after_idle &&
1491 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1492 tcp_cwnd_application_limited(sk);
1496 /* Returns the portion of skb which can be sent right away without
1497 * introducing MSS oddities to segment boundaries. In rare cases where
1498 * mss_now != mss_cache, we will request caller to create a small skb
1499 * per input skb which could be mostly avoided here (if desired).
1501 * We explicitly want to create a request for splitting write queue tail
1502 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1503 * thus all the complexity (cwnd_len is always MSS multiple which we
1504 * return whenever allowed by the other factors). Basically we need the
1505 * modulo only when the receiver window alone is the limiting factor or
1506 * when we would be allowed to send the split-due-to-Nagle skb fully.
1508 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1509 unsigned int mss_now, unsigned int cwnd)
1511 const struct tcp_sock *tp = tcp_sk(sk);
1512 u32 needed, window, cwnd_len;
1514 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1515 cwnd_len = mss_now * cwnd;
1517 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1520 needed = min(skb->len, window);
1522 if (cwnd_len <= needed)
1525 return needed - needed % mss_now;
1528 /* Can at least one segment of SKB be sent right now, according to the
1529 * congestion window rules? If so, return how many segments are allowed.
1531 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1532 const struct sk_buff *skb)
1534 u32 in_flight, cwnd;
1536 /* Don't be strict about the congestion window for the final FIN. */
1537 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1538 tcp_skb_pcount(skb) == 1)
1541 in_flight = tcp_packets_in_flight(tp);
1542 cwnd = tp->snd_cwnd;
1543 if (in_flight < cwnd)
1544 return (cwnd - in_flight);
1549 /* Initialize TSO state of a skb.
1550 * This must be invoked the first time we consider transmitting
1551 * SKB onto the wire.
1553 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1554 unsigned int mss_now)
1556 int tso_segs = tcp_skb_pcount(skb);
1558 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1559 tcp_set_skb_tso_segs(sk, skb, mss_now);
1560 tso_segs = tcp_skb_pcount(skb);
1565 /* Minshall's variant of the Nagle send check. */
1566 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1568 return after(tp->snd_sml, tp->snd_una) &&
1569 !after(tp->snd_sml, tp->snd_nxt);
1572 /* Return false, if packet can be sent now without violation Nagle's rules:
1573 * 1. It is full sized.
1574 * 2. Or it contains FIN. (already checked by caller)
1575 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1576 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1577 * With Minshall's modification: all sent small packets are ACKed.
1579 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1580 const struct sk_buff *skb,
1581 unsigned int mss_now, int nonagle)
1583 return skb->len < mss_now &&
1584 ((nonagle & TCP_NAGLE_CORK) ||
1585 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1588 /* Return true if the Nagle test allows this packet to be
1591 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1592 unsigned int cur_mss, int nonagle)
1594 /* Nagle rule does not apply to frames, which sit in the middle of the
1595 * write_queue (they have no chances to get new data).
1597 * This is implemented in the callers, where they modify the 'nonagle'
1598 * argument based upon the location of SKB in the send queue.
1600 if (nonagle & TCP_NAGLE_PUSH)
1603 /* Don't use the nagle rule for urgent data (or for the final FIN).
1604 * Nagle can be ignored during F-RTO too (see RFC4138).
1606 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1607 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1610 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1616 /* Does at least the first segment of SKB fit into the send window? */
1617 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1618 const struct sk_buff *skb,
1619 unsigned int cur_mss)
1621 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1623 if (skb->len > cur_mss)
1624 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1626 return !after(end_seq, tcp_wnd_end(tp));
1629 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1630 * should be put on the wire right now. If so, it returns the number of
1631 * packets allowed by the congestion window.
1633 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1634 unsigned int cur_mss, int nonagle)
1636 const struct tcp_sock *tp = tcp_sk(sk);
1637 unsigned int cwnd_quota;
1639 tcp_init_tso_segs(sk, skb, cur_mss);
1641 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1644 cwnd_quota = tcp_cwnd_test(tp, skb);
1645 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1651 /* Test if sending is allowed right now. */
1652 bool tcp_may_send_now(struct sock *sk)
1654 const struct tcp_sock *tp = tcp_sk(sk);
1655 struct sk_buff *skb = tcp_send_head(sk);
1658 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1659 (tcp_skb_is_last(sk, skb) ?
1660 tp->nonagle : TCP_NAGLE_PUSH));
1663 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1664 * which is put after SKB on the list. It is very much like
1665 * tcp_fragment() except that it may make several kinds of assumptions
1666 * in order to speed up the splitting operation. In particular, we
1667 * know that all the data is in scatter-gather pages, and that the
1668 * packet has never been sent out before (and thus is not cloned).
1670 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1671 unsigned int mss_now, gfp_t gfp)
1673 struct sk_buff *buff;
1674 int nlen = skb->len - len;
1677 /* All of a TSO frame must be composed of paged data. */
1678 if (skb->len != skb->data_len)
1679 return tcp_fragment(sk, skb, len, mss_now);
1681 buff = sk_stream_alloc_skb(sk, 0, gfp);
1682 if (unlikely(buff == NULL))
1685 sk->sk_wmem_queued += buff->truesize;
1686 sk_mem_charge(sk, buff->truesize);
1687 buff->truesize += nlen;
1688 skb->truesize -= nlen;
1690 /* Correct the sequence numbers. */
1691 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1692 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1693 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1695 /* PSH and FIN should only be set in the second packet. */
1696 flags = TCP_SKB_CB(skb)->tcp_flags;
1697 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1698 TCP_SKB_CB(buff)->tcp_flags = flags;
1700 /* This packet was never sent out yet, so no SACK bits. */
1701 TCP_SKB_CB(buff)->sacked = 0;
1703 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1704 skb_split(skb, buff, len);
1706 /* Fix up tso_factor for both original and new SKB. */
1707 tcp_set_skb_tso_segs(sk, skb, mss_now);
1708 tcp_set_skb_tso_segs(sk, buff, mss_now);
1710 /* Link BUFF into the send queue. */
1711 skb_header_release(buff);
1712 tcp_insert_write_queue_after(skb, buff, sk);
1717 /* Try to defer sending, if possible, in order to minimize the amount
1718 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1720 * This algorithm is from John Heffner.
1722 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1724 struct tcp_sock *tp = tcp_sk(sk);
1725 const struct inet_connection_sock *icsk = inet_csk(sk);
1726 u32 send_win, cong_win, limit, in_flight;
1729 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1732 if (icsk->icsk_ca_state != TCP_CA_Open)
1735 /* Defer for less than two clock ticks. */
1736 if (tp->tso_deferred &&
1737 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1740 in_flight = tcp_packets_in_flight(tp);
1742 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1744 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1746 /* From in_flight test above, we know that cwnd > in_flight. */
1747 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1749 limit = min(send_win, cong_win);
1751 /* If a full-sized TSO skb can be sent, do it. */
1752 if (limit >= sk->sk_gso_max_size)
1755 /* Middle in queue won't get any more data, full sendable already? */
1756 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1759 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1761 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1763 /* If at least some fraction of a window is available,
1766 chunk /= win_divisor;
1770 /* Different approach, try not to defer past a single
1771 * ACK. Receiver should ACK every other full sized
1772 * frame, so if we have space for more than 3 frames
1775 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1779 /* Ok, it looks like it is advisable to defer. */
1780 tp->tso_deferred = 1 | (jiffies << 1);
1785 tp->tso_deferred = 0;
1789 /* Create a new MTU probe if we are ready.
1790 * MTU probe is regularly attempting to increase the path MTU by
1791 * deliberately sending larger packets. This discovers routing
1792 * changes resulting in larger path MTUs.
1794 * Returns 0 if we should wait to probe (no cwnd available),
1795 * 1 if a probe was sent,
1798 static int tcp_mtu_probe(struct sock *sk)
1800 struct tcp_sock *tp = tcp_sk(sk);
1801 struct inet_connection_sock *icsk = inet_csk(sk);
1802 struct sk_buff *skb, *nskb, *next;
1809 /* Not currently probing/verifying,
1811 * have enough cwnd, and
1812 * not SACKing (the variable headers throw things off) */
1813 if (!icsk->icsk_mtup.enabled ||
1814 icsk->icsk_mtup.probe_size ||
1815 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1816 tp->snd_cwnd < 11 ||
1817 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1820 /* Very simple search strategy: just double the MSS. */
1821 mss_now = tcp_current_mss(sk);
1822 probe_size = 2 * tp->mss_cache;
1823 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1824 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1825 /* TODO: set timer for probe_converge_event */
1829 /* Have enough data in the send queue to probe? */
1830 if (tp->write_seq - tp->snd_nxt < size_needed)
1833 if (tp->snd_wnd < size_needed)
1835 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1838 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1839 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1840 if (!tcp_packets_in_flight(tp))
1846 /* We're allowed to probe. Build it now. */
1847 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1849 sk->sk_wmem_queued += nskb->truesize;
1850 sk_mem_charge(sk, nskb->truesize);
1852 skb = tcp_send_head(sk);
1854 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1855 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1856 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1857 TCP_SKB_CB(nskb)->sacked = 0;
1859 nskb->ip_summed = skb->ip_summed;
1861 tcp_insert_write_queue_before(nskb, skb, sk);
1864 tcp_for_write_queue_from_safe(skb, next, sk) {
1865 copy = min_t(int, skb->len, probe_size - len);
1866 if (nskb->ip_summed)
1867 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1869 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1870 skb_put(nskb, copy),
1873 if (skb->len <= copy) {
1874 /* We've eaten all the data from this skb.
1876 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1877 tcp_unlink_write_queue(skb, sk);
1878 sk_wmem_free_skb(sk, skb);
1880 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1881 ~(TCPHDR_FIN|TCPHDR_PSH);
1882 if (!skb_shinfo(skb)->nr_frags) {
1883 skb_pull(skb, copy);
1884 if (skb->ip_summed != CHECKSUM_PARTIAL)
1885 skb->csum = csum_partial(skb->data,
1888 __pskb_trim_head(skb, copy);
1889 tcp_set_skb_tso_segs(sk, skb, mss_now);
1891 TCP_SKB_CB(skb)->seq += copy;
1896 if (len >= probe_size)
1899 tcp_init_tso_segs(sk, nskb, nskb->len);
1901 /* We're ready to send. If this fails, the probe will
1902 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1903 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1904 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1905 /* Decrement cwnd here because we are sending
1906 * effectively two packets. */
1908 tcp_event_new_data_sent(sk, nskb);
1910 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1911 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1912 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1920 /* This routine writes packets to the network. It advances the
1921 * send_head. This happens as incoming acks open up the remote
1924 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1925 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1926 * account rare use of URG, this is not a big flaw.
1928 * Returns true, if no segments are in flight and we have queued segments,
1929 * but cannot send anything now because of SWS or another problem.
1931 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1932 int push_one, gfp_t gfp)
1934 struct tcp_sock *tp = tcp_sk(sk);
1935 struct sk_buff *skb;
1936 unsigned int tso_segs, sent_pkts;
1943 /* Do MTU probing. */
1944 result = tcp_mtu_probe(sk);
1947 } else if (result > 0) {
1952 while ((skb = tcp_send_head(sk))) {
1956 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1959 cwnd_quota = tcp_cwnd_test(tp, skb);
1963 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1966 if (tso_segs == 1) {
1967 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1968 (tcp_skb_is_last(sk, skb) ?
1969 nonagle : TCP_NAGLE_PUSH))))
1972 if (!push_one && tcp_tso_should_defer(sk, skb))
1976 /* TSQ : sk_wmem_alloc accounts skb truesize,
1977 * including skb overhead. But thats OK.
1979 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1980 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1984 if (tso_segs > 1 && !tcp_urg_mode(tp))
1985 limit = tcp_mss_split_point(sk, skb, mss_now,
1988 if (skb->len > limit &&
1989 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1992 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1994 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1997 /* Advance the send_head. This one is sent out.
1998 * This call will increment packets_out.
2000 tcp_event_new_data_sent(sk, skb);
2002 tcp_minshall_update(tp, mss_now, skb);
2003 sent_pkts += tcp_skb_pcount(skb);
2008 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2009 tp->prr_out += sent_pkts;
2011 if (likely(sent_pkts)) {
2012 tcp_cwnd_validate(sk);
2015 return !tp->packets_out && tcp_send_head(sk);
2018 /* Push out any pending frames which were held back due to
2019 * TCP_CORK or attempt at coalescing tiny packets.
2020 * The socket must be locked by the caller.
2022 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2025 /* If we are closed, the bytes will have to remain here.
2026 * In time closedown will finish, we empty the write queue and
2027 * all will be happy.
2029 if (unlikely(sk->sk_state == TCP_CLOSE))
2032 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
2033 tcp_check_probe_timer(sk);
2036 /* Send _single_ skb sitting at the send head. This function requires
2037 * true push pending frames to setup probe timer etc.
2039 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2041 struct sk_buff *skb = tcp_send_head(sk);
2043 BUG_ON(!skb || skb->len < mss_now);
2045 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2048 /* This function returns the amount that we can raise the
2049 * usable window based on the following constraints
2051 * 1. The window can never be shrunk once it is offered (RFC 793)
2052 * 2. We limit memory per socket
2055 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2056 * RECV.NEXT + RCV.WIN fixed until:
2057 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2059 * i.e. don't raise the right edge of the window until you can raise
2060 * it at least MSS bytes.
2062 * Unfortunately, the recommended algorithm breaks header prediction,
2063 * since header prediction assumes th->window stays fixed.
2065 * Strictly speaking, keeping th->window fixed violates the receiver
2066 * side SWS prevention criteria. The problem is that under this rule
2067 * a stream of single byte packets will cause the right side of the
2068 * window to always advance by a single byte.
2070 * Of course, if the sender implements sender side SWS prevention
2071 * then this will not be a problem.
2073 * BSD seems to make the following compromise:
2075 * If the free space is less than the 1/4 of the maximum
2076 * space available and the free space is less than 1/2 mss,
2077 * then set the window to 0.
2078 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2079 * Otherwise, just prevent the window from shrinking
2080 * and from being larger than the largest representable value.
2082 * This prevents incremental opening of the window in the regime
2083 * where TCP is limited by the speed of the reader side taking
2084 * data out of the TCP receive queue. It does nothing about
2085 * those cases where the window is constrained on the sender side
2086 * because the pipeline is full.
2088 * BSD also seems to "accidentally" limit itself to windows that are a
2089 * multiple of MSS, at least until the free space gets quite small.
2090 * This would appear to be a side effect of the mbuf implementation.
2091 * Combining these two algorithms results in the observed behavior
2092 * of having a fixed window size at almost all times.
2094 * Below we obtain similar behavior by forcing the offered window to
2095 * a multiple of the mss when it is feasible to do so.
2097 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2098 * Regular options like TIMESTAMP are taken into account.
2100 u32 __tcp_select_window(struct sock *sk)
2102 struct inet_connection_sock *icsk = inet_csk(sk);
2103 struct tcp_sock *tp = tcp_sk(sk);
2104 /* MSS for the peer's data. Previous versions used mss_clamp
2105 * here. I don't know if the value based on our guesses
2106 * of peer's MSS is better for the performance. It's more correct
2107 * but may be worse for the performance because of rcv_mss
2108 * fluctuations. --SAW 1998/11/1
2110 int mss = icsk->icsk_ack.rcv_mss;
2111 int free_space = tcp_space(sk);
2112 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2115 if (mss > full_space)
2118 if (free_space < (full_space >> 1)) {
2119 icsk->icsk_ack.quick = 0;
2121 if (sk_under_memory_pressure(sk))
2122 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2125 if (free_space < mss)
2129 if (free_space > tp->rcv_ssthresh)
2130 free_space = tp->rcv_ssthresh;
2132 /* Don't do rounding if we are using window scaling, since the
2133 * scaled window will not line up with the MSS boundary anyway.
2135 window = tp->rcv_wnd;
2136 if (tp->rx_opt.rcv_wscale) {
2137 window = free_space;
2139 /* Advertise enough space so that it won't get scaled away.
2140 * Import case: prevent zero window announcement if
2141 * 1<<rcv_wscale > mss.
2143 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2144 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2145 << tp->rx_opt.rcv_wscale);
2147 /* Get the largest window that is a nice multiple of mss.
2148 * Window clamp already applied above.
2149 * If our current window offering is within 1 mss of the
2150 * free space we just keep it. This prevents the divide
2151 * and multiply from happening most of the time.
2152 * We also don't do any window rounding when the free space
2155 if (window <= free_space - mss || window > free_space)
2156 window = (free_space / mss) * mss;
2157 else if (mss == full_space &&
2158 free_space > window + (full_space >> 1))
2159 window = free_space;
2165 /* Collapses two adjacent SKB's during retransmission. */
2166 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2168 struct tcp_sock *tp = tcp_sk(sk);
2169 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2170 int skb_size, next_skb_size;
2172 skb_size = skb->len;
2173 next_skb_size = next_skb->len;
2175 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2177 tcp_highest_sack_combine(sk, next_skb, skb);
2179 tcp_unlink_write_queue(next_skb, sk);
2181 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2184 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2185 skb->ip_summed = CHECKSUM_PARTIAL;
2187 if (skb->ip_summed != CHECKSUM_PARTIAL)
2188 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2190 /* Update sequence range on original skb. */
2191 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2193 /* Merge over control information. This moves PSH/FIN etc. over */
2194 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2196 /* All done, get rid of second SKB and account for it so
2197 * packet counting does not break.
2199 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2201 /* changed transmit queue under us so clear hints */
2202 tcp_clear_retrans_hints_partial(tp);
2203 if (next_skb == tp->retransmit_skb_hint)
2204 tp->retransmit_skb_hint = skb;
2206 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2208 sk_wmem_free_skb(sk, next_skb);
2211 /* Check if coalescing SKBs is legal. */
2212 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2214 if (tcp_skb_pcount(skb) > 1)
2216 /* TODO: SACK collapsing could be used to remove this condition */
2217 if (skb_shinfo(skb)->nr_frags != 0)
2219 if (skb_cloned(skb))
2221 if (skb == tcp_send_head(sk))
2223 /* Some heurestics for collapsing over SACK'd could be invented */
2224 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2230 /* Collapse packets in the retransmit queue to make to create
2231 * less packets on the wire. This is only done on retransmission.
2233 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2236 struct tcp_sock *tp = tcp_sk(sk);
2237 struct sk_buff *skb = to, *tmp;
2240 if (!sysctl_tcp_retrans_collapse)
2242 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2245 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2246 if (!tcp_can_collapse(sk, skb))
2258 /* Punt if not enough space exists in the first SKB for
2259 * the data in the second
2261 if (skb->len > skb_availroom(to))
2264 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2267 tcp_collapse_retrans(sk, to);
2271 /* This retransmits one SKB. Policy decisions and retransmit queue
2272 * state updates are done by the caller. Returns non-zero if an
2273 * error occurred which prevented the send.
2275 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2277 struct tcp_sock *tp = tcp_sk(sk);
2278 struct inet_connection_sock *icsk = inet_csk(sk);
2279 unsigned int cur_mss;
2282 /* Inconslusive MTU probe */
2283 if (icsk->icsk_mtup.probe_size) {
2284 icsk->icsk_mtup.probe_size = 0;
2287 /* Do not sent more than we queued. 1/4 is reserved for possible
2288 * copying overhead: fragmentation, tunneling, mangling etc.
2290 if (atomic_read(&sk->sk_wmem_alloc) >
2291 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2294 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2295 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2297 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2301 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2302 return -EHOSTUNREACH; /* Routing failure or similar. */
2304 cur_mss = tcp_current_mss(sk);
2306 /* If receiver has shrunk his window, and skb is out of
2307 * new window, do not retransmit it. The exception is the
2308 * case, when window is shrunk to zero. In this case
2309 * our retransmit serves as a zero window probe.
2311 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2312 TCP_SKB_CB(skb)->seq != tp->snd_una)
2315 if (skb->len > cur_mss) {
2316 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2317 return -ENOMEM; /* We'll try again later. */
2319 int oldpcount = tcp_skb_pcount(skb);
2321 if (unlikely(oldpcount > 1)) {
2322 tcp_init_tso_segs(sk, skb, cur_mss);
2323 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2327 tcp_retrans_try_collapse(sk, skb, cur_mss);
2329 /* Some Solaris stacks overoptimize and ignore the FIN on a
2330 * retransmit when old data is attached. So strip it off
2331 * since it is cheap to do so and saves bytes on the network.
2334 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2335 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2336 if (!pskb_trim(skb, 0)) {
2337 /* Reuse, even though it does some unnecessary work */
2338 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2339 TCP_SKB_CB(skb)->tcp_flags);
2340 skb->ip_summed = CHECKSUM_NONE;
2344 /* Make a copy, if the first transmission SKB clone we made
2345 * is still in somebody's hands, else make a clone.
2347 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2349 /* make sure skb->data is aligned on arches that require it */
2350 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2351 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2353 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2356 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2360 /* Update global TCP statistics. */
2361 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2363 tp->total_retrans++;
2365 #if FASTRETRANS_DEBUG > 0
2366 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2367 net_dbg_ratelimited("retrans_out leaked\n");
2370 if (!tp->retrans_out)
2371 tp->lost_retrans_low = tp->snd_nxt;
2372 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2373 tp->retrans_out += tcp_skb_pcount(skb);
2375 /* Save stamp of the first retransmit. */
2376 if (!tp->retrans_stamp)
2377 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2379 tp->undo_retrans += tcp_skb_pcount(skb);
2381 /* snd_nxt is stored to detect loss of retransmitted segment,
2382 * see tcp_input.c tcp_sacktag_write_queue().
2384 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2389 /* Check if we forward retransmits are possible in the current
2390 * window/congestion state.
2392 static bool tcp_can_forward_retransmit(struct sock *sk)
2394 const struct inet_connection_sock *icsk = inet_csk(sk);
2395 const struct tcp_sock *tp = tcp_sk(sk);
2397 /* Forward retransmissions are possible only during Recovery. */
2398 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2401 /* No forward retransmissions in Reno are possible. */
2402 if (tcp_is_reno(tp))
2405 /* Yeah, we have to make difficult choice between forward transmission
2406 * and retransmission... Both ways have their merits...
2408 * For now we do not retransmit anything, while we have some new
2409 * segments to send. In the other cases, follow rule 3 for
2410 * NextSeg() specified in RFC3517.
2413 if (tcp_may_send_now(sk))
2419 /* This gets called after a retransmit timeout, and the initially
2420 * retransmitted data is acknowledged. It tries to continue
2421 * resending the rest of the retransmit queue, until either
2422 * we've sent it all or the congestion window limit is reached.
2423 * If doing SACK, the first ACK which comes back for a timeout
2424 * based retransmit packet might feed us FACK information again.
2425 * If so, we use it to avoid unnecessarily retransmissions.
2427 void tcp_xmit_retransmit_queue(struct sock *sk)
2429 const struct inet_connection_sock *icsk = inet_csk(sk);
2430 struct tcp_sock *tp = tcp_sk(sk);
2431 struct sk_buff *skb;
2432 struct sk_buff *hole = NULL;
2435 int fwd_rexmitting = 0;
2437 if (!tp->packets_out)
2441 tp->retransmit_high = tp->snd_una;
2443 if (tp->retransmit_skb_hint) {
2444 skb = tp->retransmit_skb_hint;
2445 last_lost = TCP_SKB_CB(skb)->end_seq;
2446 if (after(last_lost, tp->retransmit_high))
2447 last_lost = tp->retransmit_high;
2449 skb = tcp_write_queue_head(sk);
2450 last_lost = tp->snd_una;
2453 tcp_for_write_queue_from(skb, sk) {
2454 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2456 if (skb == tcp_send_head(sk))
2458 /* we could do better than to assign each time */
2460 tp->retransmit_skb_hint = skb;
2462 /* Assume this retransmit will generate
2463 * only one packet for congestion window
2464 * calculation purposes. This works because
2465 * tcp_retransmit_skb() will chop up the
2466 * packet to be MSS sized and all the
2467 * packet counting works out.
2469 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2472 if (fwd_rexmitting) {
2474 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2476 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2478 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2479 tp->retransmit_high = last_lost;
2480 if (!tcp_can_forward_retransmit(sk))
2482 /* Backtrack if necessary to non-L'ed skb */
2490 } else if (!(sacked & TCPCB_LOST)) {
2491 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2496 last_lost = TCP_SKB_CB(skb)->end_seq;
2497 if (icsk->icsk_ca_state != TCP_CA_Loss)
2498 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2500 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2503 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2506 if (tcp_retransmit_skb(sk, skb)) {
2507 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2510 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2512 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2513 tp->prr_out += tcp_skb_pcount(skb);
2515 if (skb == tcp_write_queue_head(sk))
2516 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2517 inet_csk(sk)->icsk_rto,
2522 /* Send a fin. The caller locks the socket for us. This cannot be
2523 * allowed to fail queueing a FIN frame under any circumstances.
2525 void tcp_send_fin(struct sock *sk)
2527 struct tcp_sock *tp = tcp_sk(sk);
2528 struct sk_buff *skb = tcp_write_queue_tail(sk);
2531 /* Optimization, tack on the FIN if we have a queue of
2532 * unsent frames. But be careful about outgoing SACKS
2535 mss_now = tcp_current_mss(sk);
2537 if (tcp_send_head(sk) != NULL) {
2538 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2539 TCP_SKB_CB(skb)->end_seq++;
2542 /* Socket is locked, keep trying until memory is available. */
2544 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2551 /* Reserve space for headers and prepare control bits. */
2552 skb_reserve(skb, MAX_TCP_HEADER);
2553 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2554 tcp_init_nondata_skb(skb, tp->write_seq,
2555 TCPHDR_ACK | TCPHDR_FIN);
2556 tcp_queue_skb(sk, skb);
2558 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2561 /* We get here when a process closes a file descriptor (either due to
2562 * an explicit close() or as a byproduct of exit()'ing) and there
2563 * was unread data in the receive queue. This behavior is recommended
2564 * by RFC 2525, section 2.17. -DaveM
2566 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2568 struct sk_buff *skb;
2570 /* NOTE: No TCP options attached and we never retransmit this. */
2571 skb = alloc_skb(MAX_TCP_HEADER, priority);
2573 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2577 /* Reserve space for headers and prepare control bits. */
2578 skb_reserve(skb, MAX_TCP_HEADER);
2579 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2580 TCPHDR_ACK | TCPHDR_RST);
2582 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2583 if (tcp_transmit_skb(sk, skb, 0, priority))
2584 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2586 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2589 /* Send a crossed SYN-ACK during socket establishment.
2590 * WARNING: This routine must only be called when we have already sent
2591 * a SYN packet that crossed the incoming SYN that caused this routine
2592 * to get called. If this assumption fails then the initial rcv_wnd
2593 * and rcv_wscale values will not be correct.
2595 int tcp_send_synack(struct sock *sk)
2597 struct sk_buff *skb;
2599 skb = tcp_write_queue_head(sk);
2600 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2601 pr_debug("%s: wrong queue state\n", __func__);
2604 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2605 if (skb_cloned(skb)) {
2606 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2609 tcp_unlink_write_queue(skb, sk);
2610 skb_header_release(nskb);
2611 __tcp_add_write_queue_head(sk, nskb);
2612 sk_wmem_free_skb(sk, skb);
2613 sk->sk_wmem_queued += nskb->truesize;
2614 sk_mem_charge(sk, nskb->truesize);
2618 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2619 TCP_ECN_send_synack(tcp_sk(sk), skb);
2621 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2622 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2626 * tcp_make_synack - Prepare a SYN-ACK.
2627 * sk: listener socket
2628 * dst: dst entry attached to the SYNACK
2629 * req: request_sock pointer
2630 * rvp: request_values pointer
2632 * Allocate one skb and build a SYNACK packet.
2633 * @dst is consumed : Caller should not use it again.
2635 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2636 struct request_sock *req,
2637 struct request_values *rvp)
2639 struct tcp_out_options opts;
2640 struct tcp_extend_values *xvp = tcp_xv(rvp);
2641 struct inet_request_sock *ireq = inet_rsk(req);
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 const struct tcp_cookie_values *cvp = tp->cookie_values;
2645 struct sk_buff *skb;
2646 struct tcp_md5sig_key *md5;
2647 int tcp_header_size;
2649 int s_data_desired = 0;
2651 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2652 s_data_desired = cvp->s_data_desired;
2653 skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, GFP_ATOMIC);
2654 if (unlikely(!skb)) {
2658 /* Reserve space for headers. */
2659 skb_reserve(skb, MAX_TCP_HEADER);
2661 skb_dst_set(skb, dst);
2663 mss = dst_metric_advmss(dst);
2664 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2665 mss = tp->rx_opt.user_mss;
2667 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2669 /* Set this up on the first call only */
2670 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2672 /* limit the window selection if the user enforce a smaller rx buffer */
2673 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2674 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2675 req->window_clamp = tcp_full_space(sk);
2677 /* tcp_full_space because it is guaranteed to be the first packet */
2678 tcp_select_initial_window(tcp_full_space(sk),
2679 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2684 dst_metric(dst, RTAX_INITRWND));
2685 ireq->rcv_wscale = rcv_wscale;
2688 memset(&opts, 0, sizeof(opts));
2689 #ifdef CONFIG_SYN_COOKIES
2690 if (unlikely(req->cookie_ts))
2691 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2694 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2695 tcp_header_size = tcp_synack_options(sk, req, mss,
2696 skb, &opts, &md5, xvp)
2699 skb_push(skb, tcp_header_size);
2700 skb_reset_transport_header(skb);
2703 memset(th, 0, sizeof(struct tcphdr));
2706 TCP_ECN_make_synack(req, th);
2707 th->source = ireq->loc_port;
2708 th->dest = ireq->rmt_port;
2709 /* Setting of flags are superfluous here for callers (and ECE is
2710 * not even correctly set)
2712 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2713 TCPHDR_SYN | TCPHDR_ACK);
2715 if (OPTION_COOKIE_EXTENSION & opts.options) {
2716 if (s_data_desired) {
2717 u8 *buf = skb_put(skb, s_data_desired);
2719 /* copy data directly from the listening socket. */
2720 memcpy(buf, cvp->s_data_payload, s_data_desired);
2721 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2724 if (opts.hash_size > 0) {
2725 __u32 workspace[SHA_WORKSPACE_WORDS];
2726 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2727 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2729 /* Secret recipe depends on the Timestamp, (future)
2730 * Sequence and Acknowledgment Numbers, Initiator
2731 * Cookie, and others handled by IP variant caller.
2733 *tail-- ^= opts.tsval;
2734 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2735 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2738 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2739 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2741 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2744 opts.hash_location =
2745 (__u8 *)&xvp->cookie_bakery[0];
2749 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2750 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2752 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2753 th->window = htons(min(req->rcv_wnd, 65535U));
2754 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2755 th->doff = (tcp_header_size >> 2);
2756 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2758 #ifdef CONFIG_TCP_MD5SIG
2759 /* Okay, we have all we need - do the md5 hash if needed */
2761 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2762 md5, NULL, req, skb);
2768 EXPORT_SYMBOL(tcp_make_synack);
2770 /* Do all connect socket setups that can be done AF independent. */
2771 void tcp_connect_init(struct sock *sk)
2773 const struct dst_entry *dst = __sk_dst_get(sk);
2774 struct tcp_sock *tp = tcp_sk(sk);
2777 /* We'll fix this up when we get a response from the other end.
2778 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2780 tp->tcp_header_len = sizeof(struct tcphdr) +
2781 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2783 #ifdef CONFIG_TCP_MD5SIG
2784 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2785 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2788 /* If user gave his TCP_MAXSEG, record it to clamp */
2789 if (tp->rx_opt.user_mss)
2790 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2793 tcp_sync_mss(sk, dst_mtu(dst));
2795 if (!tp->window_clamp)
2796 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2797 tp->advmss = dst_metric_advmss(dst);
2798 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2799 tp->advmss = tp->rx_opt.user_mss;
2801 tcp_initialize_rcv_mss(sk);
2803 /* limit the window selection if the user enforce a smaller rx buffer */
2804 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2805 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2806 tp->window_clamp = tcp_full_space(sk);
2808 tcp_select_initial_window(tcp_full_space(sk),
2809 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2812 sysctl_tcp_window_scaling,
2814 dst_metric(dst, RTAX_INITRWND));
2816 tp->rx_opt.rcv_wscale = rcv_wscale;
2817 tp->rcv_ssthresh = tp->rcv_wnd;
2820 sock_reset_flag(sk, SOCK_DONE);
2823 tp->snd_una = tp->write_seq;
2824 tp->snd_sml = tp->write_seq;
2825 tp->snd_up = tp->write_seq;
2826 tp->snd_nxt = tp->write_seq;
2828 if (likely(!tp->repair))
2830 tp->rcv_wup = tp->rcv_nxt;
2831 tp->copied_seq = tp->rcv_nxt;
2833 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2834 inet_csk(sk)->icsk_retransmits = 0;
2835 tcp_clear_retrans(tp);
2838 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2840 struct tcp_sock *tp = tcp_sk(sk);
2841 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2843 tcb->end_seq += skb->len;
2844 skb_header_release(skb);
2845 __tcp_add_write_queue_tail(sk, skb);
2846 sk->sk_wmem_queued += skb->truesize;
2847 sk_mem_charge(sk, skb->truesize);
2848 tp->write_seq = tcb->end_seq;
2849 tp->packets_out += tcp_skb_pcount(skb);
2852 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2853 * queue a data-only packet after the regular SYN, such that regular SYNs
2854 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2855 * only the SYN sequence, the data are retransmitted in the first ACK.
2856 * If cookie is not cached or other error occurs, falls back to send a
2857 * regular SYN with Fast Open cookie request option.
2859 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2861 struct tcp_sock *tp = tcp_sk(sk);
2862 struct tcp_fastopen_request *fo = tp->fastopen_req;
2863 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2864 struct sk_buff *syn_data = NULL, *data;
2865 unsigned long last_syn_loss = 0;
2867 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2868 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2869 &syn_loss, &last_syn_loss);
2870 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2872 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2873 fo->cookie.len = -1;
2877 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2878 fo->cookie.len = -1;
2879 else if (fo->cookie.len <= 0)
2882 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2883 * user-MSS. Reserve maximum option space for middleboxes that add
2884 * private TCP options. The cost is reduced data space in SYN :(
2886 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2887 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2888 space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2889 MAX_TCP_OPTION_SPACE;
2891 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2893 if (syn_data == NULL)
2896 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2897 struct iovec *iov = &fo->data->msg_iov[i];
2898 unsigned char __user *from = iov->iov_base;
2899 int len = iov->iov_len;
2901 if (syn_data->len + len > space)
2902 len = space - syn_data->len;
2903 else if (i + 1 == iovlen)
2904 /* No more data pending in inet_wait_for_connect() */
2907 if (skb_add_data(syn_data, from, len))
2911 /* Queue a data-only packet after the regular SYN for retransmission */
2912 data = pskb_copy(syn_data, sk->sk_allocation);
2915 TCP_SKB_CB(data)->seq++;
2916 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2917 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2918 tcp_connect_queue_skb(sk, data);
2919 fo->copied = data->len;
2921 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2922 tp->syn_data = (fo->copied > 0);
2923 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2929 /* Send a regular SYN with Fast Open cookie request option */
2930 if (fo->cookie.len > 0)
2932 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2934 tp->syn_fastopen = 0;
2935 kfree_skb(syn_data);
2937 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2941 /* Build a SYN and send it off. */
2942 int tcp_connect(struct sock *sk)
2944 struct tcp_sock *tp = tcp_sk(sk);
2945 struct sk_buff *buff;
2948 tcp_connect_init(sk);
2950 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2951 if (unlikely(buff == NULL))
2954 /* Reserve space for headers. */
2955 skb_reserve(buff, MAX_TCP_HEADER);
2957 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2958 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2959 tcp_connect_queue_skb(sk, buff);
2960 TCP_ECN_send_syn(sk, buff);
2962 /* Send off SYN; include data in Fast Open. */
2963 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2964 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2965 if (err == -ECONNREFUSED)
2968 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2969 * in order to make this packet get counted in tcpOutSegs.
2971 tp->snd_nxt = tp->write_seq;
2972 tp->pushed_seq = tp->write_seq;
2973 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2975 /* Timer for repeating the SYN until an answer. */
2976 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2977 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2980 EXPORT_SYMBOL(tcp_connect);
2982 /* Send out a delayed ack, the caller does the policy checking
2983 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2986 void tcp_send_delayed_ack(struct sock *sk)
2988 struct inet_connection_sock *icsk = inet_csk(sk);
2989 int ato = icsk->icsk_ack.ato;
2990 unsigned long timeout;
2992 if (ato > TCP_DELACK_MIN) {
2993 const struct tcp_sock *tp = tcp_sk(sk);
2994 int max_ato = HZ / 2;
2996 if (icsk->icsk_ack.pingpong ||
2997 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2998 max_ato = TCP_DELACK_MAX;
3000 /* Slow path, intersegment interval is "high". */
3002 /* If some rtt estimate is known, use it to bound delayed ack.
3003 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3007 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3013 ato = min(ato, max_ato);
3016 /* Stay within the limit we were given */
3017 timeout = jiffies + ato;
3019 /* Use new timeout only if there wasn't a older one earlier. */
3020 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3021 /* If delack timer was blocked or is about to expire,
3024 if (icsk->icsk_ack.blocked ||
3025 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3030 if (!time_before(timeout, icsk->icsk_ack.timeout))
3031 timeout = icsk->icsk_ack.timeout;
3033 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3034 icsk->icsk_ack.timeout = timeout;
3035 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3038 /* This routine sends an ack and also updates the window. */
3039 void tcp_send_ack(struct sock *sk)
3041 struct sk_buff *buff;
3043 /* If we have been reset, we may not send again. */
3044 if (sk->sk_state == TCP_CLOSE)
3047 /* We are not putting this on the write queue, so
3048 * tcp_transmit_skb() will set the ownership to this
3051 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3053 inet_csk_schedule_ack(sk);
3054 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3055 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3056 TCP_DELACK_MAX, TCP_RTO_MAX);
3060 /* Reserve space for headers and prepare control bits. */
3061 skb_reserve(buff, MAX_TCP_HEADER);
3062 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3064 /* Send it off, this clears delayed acks for us. */
3065 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3066 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
3069 /* This routine sends a packet with an out of date sequence
3070 * number. It assumes the other end will try to ack it.
3072 * Question: what should we make while urgent mode?
3073 * 4.4BSD forces sending single byte of data. We cannot send
3074 * out of window data, because we have SND.NXT==SND.MAX...
3076 * Current solution: to send TWO zero-length segments in urgent mode:
3077 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3078 * out-of-date with SND.UNA-1 to probe window.
3080 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3082 struct tcp_sock *tp = tcp_sk(sk);
3083 struct sk_buff *skb;
3085 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3086 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3090 /* Reserve space for headers and set control bits. */
3091 skb_reserve(skb, MAX_TCP_HEADER);
3092 /* Use a previous sequence. This should cause the other
3093 * end to send an ack. Don't queue or clone SKB, just
3096 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3097 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3098 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3101 void tcp_send_window_probe(struct sock *sk)
3103 if (sk->sk_state == TCP_ESTABLISHED) {
3104 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3105 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3106 tcp_xmit_probe_skb(sk, 0);
3110 /* Initiate keepalive or window probe from timer. */
3111 int tcp_write_wakeup(struct sock *sk)
3113 struct tcp_sock *tp = tcp_sk(sk);
3114 struct sk_buff *skb;
3116 if (sk->sk_state == TCP_CLOSE)
3119 if ((skb = tcp_send_head(sk)) != NULL &&
3120 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3122 unsigned int mss = tcp_current_mss(sk);
3123 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3125 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3126 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3128 /* We are probing the opening of a window
3129 * but the window size is != 0
3130 * must have been a result SWS avoidance ( sender )
3132 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3134 seg_size = min(seg_size, mss);
3135 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3136 if (tcp_fragment(sk, skb, seg_size, mss))
3138 } else if (!tcp_skb_pcount(skb))
3139 tcp_set_skb_tso_segs(sk, skb, mss);
3141 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3142 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3143 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3145 tcp_event_new_data_sent(sk, skb);
3148 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3149 tcp_xmit_probe_skb(sk, 1);
3150 return tcp_xmit_probe_skb(sk, 0);
3154 /* A window probe timeout has occurred. If window is not closed send
3155 * a partial packet else a zero probe.
3157 void tcp_send_probe0(struct sock *sk)
3159 struct inet_connection_sock *icsk = inet_csk(sk);
3160 struct tcp_sock *tp = tcp_sk(sk);
3163 err = tcp_write_wakeup(sk);
3165 if (tp->packets_out || !tcp_send_head(sk)) {
3166 /* Cancel probe timer, if it is not required. */
3167 icsk->icsk_probes_out = 0;
3168 icsk->icsk_backoff = 0;
3173 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3174 icsk->icsk_backoff++;
3175 icsk->icsk_probes_out++;
3176 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3177 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3180 /* If packet was not sent due to local congestion,
3181 * do not backoff and do not remember icsk_probes_out.
3182 * Let local senders to fight for local resources.
3184 * Use accumulated backoff yet.
3186 if (!icsk->icsk_probes_out)
3187 icsk->icsk_probes_out = 1;
3188 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3189 min(icsk->icsk_rto << icsk->icsk_backoff,
3190 TCP_RESOURCE_PROBE_INTERVAL),