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);
840 static void tcp_tsq_handler(struct sock *sk)
842 if ((1 << sk->sk_state) &
843 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
844 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
845 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
848 * One tasklest per cpu tries to send more skbs.
849 * We run in tasklet context but need to disable irqs when
850 * transfering tsq->head because tcp_wfree() might
851 * interrupt us (non NAPI drivers)
853 static void tcp_tasklet_func(unsigned long data)
855 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
858 struct list_head *q, *n;
862 local_irq_save(flags);
863 list_splice_init(&tsq->head, &list);
864 local_irq_restore(flags);
866 list_for_each_safe(q, n, &list) {
867 tp = list_entry(q, struct tcp_sock, tsq_node);
868 list_del(&tp->tsq_node);
870 sk = (struct sock *)tp;
873 if (!sock_owned_by_user(sk)) {
876 /* defer the work to tcp_release_cb() */
877 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
881 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
886 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
887 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
888 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
889 (1UL << TCP_MTU_REDUCED_DEFERRED))
891 * tcp_release_cb - tcp release_sock() callback
894 * called from release_sock() to perform protocol dependent
895 * actions before socket release.
897 void tcp_release_cb(struct sock *sk)
899 struct tcp_sock *tp = tcp_sk(sk);
900 unsigned long flags, nflags;
902 /* perform an atomic operation only if at least one flag is set */
904 flags = tp->tsq_flags;
905 if (!(flags & TCP_DEFERRED_ALL))
907 nflags = flags & ~TCP_DEFERRED_ALL;
908 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
910 if (flags & (1UL << TCP_TSQ_DEFERRED))
913 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
914 tcp_write_timer_handler(sk);
917 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
918 tcp_delack_timer_handler(sk);
921 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
922 sk->sk_prot->mtu_reduced(sk);
926 EXPORT_SYMBOL(tcp_release_cb);
928 void __init tcp_tasklet_init(void)
932 for_each_possible_cpu(i) {
933 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
935 INIT_LIST_HEAD(&tsq->head);
936 tasklet_init(&tsq->tasklet,
943 * Write buffer destructor automatically called from kfree_skb.
944 * We cant xmit new skbs from this context, as we might already
947 static void tcp_wfree(struct sk_buff *skb)
949 struct sock *sk = skb->sk;
950 struct tcp_sock *tp = tcp_sk(sk);
952 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
953 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
955 struct tsq_tasklet *tsq;
957 /* Keep a ref on socket.
958 * This last ref will be released in tcp_tasklet_func()
960 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
962 /* queue this socket to tasklet queue */
963 local_irq_save(flags);
964 tsq = &__get_cpu_var(tsq_tasklet);
965 list_add(&tp->tsq_node, &tsq->head);
966 tasklet_schedule(&tsq->tasklet);
967 local_irq_restore(flags);
973 /* This routine actually transmits TCP packets queued in by
974 * tcp_do_sendmsg(). This is used by both the initial
975 * transmission and possible later retransmissions.
976 * All SKB's seen here are completely headerless. It is our
977 * job to build the TCP header, and pass the packet down to
978 * IP so it can do the same plus pass the packet off to the
981 * We are working here with either a clone of the original
982 * SKB, or a fresh unique copy made by the retransmit engine.
984 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
987 const struct inet_connection_sock *icsk = inet_csk(sk);
988 struct inet_sock *inet;
990 struct tcp_skb_cb *tcb;
991 struct tcp_out_options opts;
992 unsigned int tcp_options_size, tcp_header_size;
993 struct tcp_md5sig_key *md5;
997 BUG_ON(!skb || !tcp_skb_pcount(skb));
999 /* If congestion control is doing timestamping, we must
1000 * take such a timestamp before we potentially clone/copy.
1002 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
1003 __net_timestamp(skb);
1005 if (likely(clone_it)) {
1006 if (unlikely(skb_cloned(skb)))
1007 skb = pskb_copy(skb, gfp_mask);
1009 skb = skb_clone(skb, gfp_mask);
1016 tcb = TCP_SKB_CB(skb);
1017 memset(&opts, 0, sizeof(opts));
1019 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1020 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1022 tcp_options_size = tcp_established_options(sk, skb, &opts,
1024 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1026 if (tcp_packets_in_flight(tp) == 0) {
1027 tcp_ca_event(sk, CA_EVENT_TX_START);
1032 skb_push(skb, tcp_header_size);
1033 skb_reset_transport_header(skb);
1037 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
1038 tcp_wfree : sock_wfree;
1039 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
1041 /* Build TCP header and checksum it. */
1043 th->source = inet->inet_sport;
1044 th->dest = inet->inet_dport;
1045 th->seq = htonl(tcb->seq);
1046 th->ack_seq = htonl(tp->rcv_nxt);
1047 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1050 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1051 /* RFC1323: The window in SYN & SYN/ACK segments
1054 th->window = htons(min(tp->rcv_wnd, 65535U));
1056 th->window = htons(tcp_select_window(sk));
1061 /* The urg_mode check is necessary during a below snd_una win probe */
1062 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1063 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1064 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1066 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1067 th->urg_ptr = htons(0xFFFF);
1072 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1073 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1074 TCP_ECN_send(sk, skb, tcp_header_size);
1076 #ifdef CONFIG_TCP_MD5SIG
1077 /* Calculate the MD5 hash, as we have all we need now */
1079 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1080 tp->af_specific->calc_md5_hash(opts.hash_location,
1081 md5, sk, NULL, skb);
1085 icsk->icsk_af_ops->send_check(sk, skb);
1087 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1088 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1090 if (skb->len != tcp_header_size)
1091 tcp_event_data_sent(tp, sk);
1093 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1094 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1095 tcp_skb_pcount(skb));
1097 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1098 if (likely(err <= 0))
1101 tcp_enter_cwr(sk, 1);
1103 return net_xmit_eval(err);
1106 /* This routine just queues the buffer for sending.
1108 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1109 * otherwise socket can stall.
1111 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1113 struct tcp_sock *tp = tcp_sk(sk);
1115 /* Advance write_seq and place onto the write_queue. */
1116 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1117 skb_header_release(skb);
1118 tcp_add_write_queue_tail(sk, skb);
1119 sk->sk_wmem_queued += skb->truesize;
1120 sk_mem_charge(sk, skb->truesize);
1123 /* Initialize TSO segments for a packet. */
1124 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1125 unsigned int mss_now)
1127 if (skb->len <= mss_now || !sk_can_gso(sk) ||
1128 skb->ip_summed == CHECKSUM_NONE) {
1129 /* Avoid the costly divide in the normal
1132 skb_shinfo(skb)->gso_segs = 1;
1133 skb_shinfo(skb)->gso_size = 0;
1134 skb_shinfo(skb)->gso_type = 0;
1136 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1137 skb_shinfo(skb)->gso_size = mss_now;
1138 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1142 /* When a modification to fackets out becomes necessary, we need to check
1143 * skb is counted to fackets_out or not.
1145 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1148 struct tcp_sock *tp = tcp_sk(sk);
1150 if (!tp->sacked_out || tcp_is_reno(tp))
1153 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1154 tp->fackets_out -= decr;
1157 /* Pcount in the middle of the write queue got changed, we need to do various
1158 * tweaks to fix counters
1160 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1162 struct tcp_sock *tp = tcp_sk(sk);
1164 tp->packets_out -= decr;
1166 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1167 tp->sacked_out -= decr;
1168 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1169 tp->retrans_out -= decr;
1170 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1171 tp->lost_out -= decr;
1173 /* Reno case is special. Sigh... */
1174 if (tcp_is_reno(tp) && decr > 0)
1175 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1177 tcp_adjust_fackets_out(sk, skb, decr);
1179 if (tp->lost_skb_hint &&
1180 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1181 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1182 tp->lost_cnt_hint -= decr;
1184 tcp_verify_left_out(tp);
1187 /* Function to create two new TCP segments. Shrinks the given segment
1188 * to the specified size and appends a new segment with the rest of the
1189 * packet to the list. This won't be called frequently, I hope.
1190 * Remember, these are still headerless SKBs at this point.
1192 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1193 unsigned int mss_now)
1195 struct tcp_sock *tp = tcp_sk(sk);
1196 struct sk_buff *buff;
1197 int nsize, old_factor;
1201 if (WARN_ON(len > skb->len))
1204 nsize = skb_headlen(skb) - len;
1208 if (skb_cloned(skb) &&
1209 skb_is_nonlinear(skb) &&
1210 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1213 /* Get a new skb... force flag on. */
1214 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1216 return -ENOMEM; /* We'll just try again later. */
1218 sk->sk_wmem_queued += buff->truesize;
1219 sk_mem_charge(sk, buff->truesize);
1220 nlen = skb->len - len - nsize;
1221 buff->truesize += nlen;
1222 skb->truesize -= nlen;
1224 /* Correct the sequence numbers. */
1225 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1226 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1227 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1229 /* PSH and FIN should only be set in the second packet. */
1230 flags = TCP_SKB_CB(skb)->tcp_flags;
1231 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1232 TCP_SKB_CB(buff)->tcp_flags = flags;
1233 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1235 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1236 /* Copy and checksum data tail into the new buffer. */
1237 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1238 skb_put(buff, nsize),
1243 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1245 skb->ip_summed = CHECKSUM_PARTIAL;
1246 skb_split(skb, buff, len);
1249 buff->ip_summed = skb->ip_summed;
1251 /* Looks stupid, but our code really uses when of
1252 * skbs, which it never sent before. --ANK
1254 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1255 buff->tstamp = skb->tstamp;
1257 old_factor = tcp_skb_pcount(skb);
1259 /* Fix up tso_factor for both original and new SKB. */
1260 tcp_set_skb_tso_segs(sk, skb, mss_now);
1261 tcp_set_skb_tso_segs(sk, buff, mss_now);
1263 /* If this packet has been sent out already, we must
1264 * adjust the various packet counters.
1266 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1267 int diff = old_factor - tcp_skb_pcount(skb) -
1268 tcp_skb_pcount(buff);
1271 tcp_adjust_pcount(sk, skb, diff);
1274 /* Link BUFF into the send queue. */
1275 skb_header_release(buff);
1276 tcp_insert_write_queue_after(skb, buff, sk);
1281 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1282 * eventually). The difference is that pulled data not copied, but
1283 * immediately discarded.
1285 static void __pskb_trim_head(struct sk_buff *skb, int len)
1289 eat = min_t(int, len, skb_headlen(skb));
1291 __skb_pull(skb, eat);
1292 skb->avail_size -= eat;
1299 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1300 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1303 skb_frag_unref(skb, i);
1306 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1308 skb_shinfo(skb)->frags[k].page_offset += eat;
1309 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1315 skb_shinfo(skb)->nr_frags = k;
1317 skb_reset_tail_pointer(skb);
1318 skb->data_len -= len;
1319 skb->len = skb->data_len;
1322 /* Remove acked data from a packet in the transmit queue. */
1323 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1325 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1328 __pskb_trim_head(skb, len);
1330 TCP_SKB_CB(skb)->seq += len;
1331 skb->ip_summed = CHECKSUM_PARTIAL;
1333 skb->truesize -= len;
1334 sk->sk_wmem_queued -= len;
1335 sk_mem_uncharge(sk, len);
1336 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1338 /* Any change of skb->len requires recalculation of tso factor. */
1339 if (tcp_skb_pcount(skb) > 1)
1340 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1345 /* Calculate MSS. Not accounting for SACKs here. */
1346 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1348 const struct tcp_sock *tp = tcp_sk(sk);
1349 const struct inet_connection_sock *icsk = inet_csk(sk);
1352 /* Calculate base mss without TCP options:
1353 It is MMS_S - sizeof(tcphdr) of rfc1122
1355 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1357 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1358 if (icsk->icsk_af_ops->net_frag_header_len) {
1359 const struct dst_entry *dst = __sk_dst_get(sk);
1361 if (dst && dst_allfrag(dst))
1362 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1365 /* Clamp it (mss_clamp does not include tcp options) */
1366 if (mss_now > tp->rx_opt.mss_clamp)
1367 mss_now = tp->rx_opt.mss_clamp;
1369 /* Now subtract optional transport overhead */
1370 mss_now -= icsk->icsk_ext_hdr_len;
1372 /* Then reserve room for full set of TCP options and 8 bytes of data */
1376 /* Now subtract TCP options size, not including SACKs */
1377 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1382 /* Inverse of above */
1383 int tcp_mss_to_mtu(struct sock *sk, int mss)
1385 const struct tcp_sock *tp = tcp_sk(sk);
1386 const struct inet_connection_sock *icsk = inet_csk(sk);
1390 tp->tcp_header_len +
1391 icsk->icsk_ext_hdr_len +
1392 icsk->icsk_af_ops->net_header_len;
1394 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1395 if (icsk->icsk_af_ops->net_frag_header_len) {
1396 const struct dst_entry *dst = __sk_dst_get(sk);
1398 if (dst && dst_allfrag(dst))
1399 mtu += icsk->icsk_af_ops->net_frag_header_len;
1404 /* MTU probing init per socket */
1405 void tcp_mtup_init(struct sock *sk)
1407 struct tcp_sock *tp = tcp_sk(sk);
1408 struct inet_connection_sock *icsk = inet_csk(sk);
1410 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1411 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1412 icsk->icsk_af_ops->net_header_len;
1413 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1414 icsk->icsk_mtup.probe_size = 0;
1416 EXPORT_SYMBOL(tcp_mtup_init);
1418 /* This function synchronize snd mss to current pmtu/exthdr set.
1420 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1421 for TCP options, but includes only bare TCP header.
1423 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1424 It is minimum of user_mss and mss received with SYN.
1425 It also does not include TCP options.
1427 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1429 tp->mss_cache is current effective sending mss, including
1430 all tcp options except for SACKs. It is evaluated,
1431 taking into account current pmtu, but never exceeds
1432 tp->rx_opt.mss_clamp.
1434 NOTE1. rfc1122 clearly states that advertised MSS
1435 DOES NOT include either tcp or ip options.
1437 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1438 are READ ONLY outside this function. --ANK (980731)
1440 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1442 struct tcp_sock *tp = tcp_sk(sk);
1443 struct inet_connection_sock *icsk = inet_csk(sk);
1446 if (icsk->icsk_mtup.search_high > pmtu)
1447 icsk->icsk_mtup.search_high = pmtu;
1449 mss_now = tcp_mtu_to_mss(sk, pmtu);
1450 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1452 /* And store cached results */
1453 icsk->icsk_pmtu_cookie = pmtu;
1454 if (icsk->icsk_mtup.enabled)
1455 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1456 tp->mss_cache = mss_now;
1460 EXPORT_SYMBOL(tcp_sync_mss);
1462 /* Compute the current effective MSS, taking SACKs and IP options,
1463 * and even PMTU discovery events into account.
1465 unsigned int tcp_current_mss(struct sock *sk)
1467 const struct tcp_sock *tp = tcp_sk(sk);
1468 const struct dst_entry *dst = __sk_dst_get(sk);
1470 unsigned int header_len;
1471 struct tcp_out_options opts;
1472 struct tcp_md5sig_key *md5;
1474 mss_now = tp->mss_cache;
1477 u32 mtu = dst_mtu(dst);
1478 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1479 mss_now = tcp_sync_mss(sk, mtu);
1482 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1483 sizeof(struct tcphdr);
1484 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1485 * some common options. If this is an odd packet (because we have SACK
1486 * blocks etc) then our calculated header_len will be different, and
1487 * we have to adjust mss_now correspondingly */
1488 if (header_len != tp->tcp_header_len) {
1489 int delta = (int) header_len - tp->tcp_header_len;
1496 /* Congestion window validation. (RFC2861) */
1497 static void tcp_cwnd_validate(struct sock *sk)
1499 struct tcp_sock *tp = tcp_sk(sk);
1501 if (tp->packets_out >= tp->snd_cwnd) {
1502 /* Network is feed fully. */
1503 tp->snd_cwnd_used = 0;
1504 tp->snd_cwnd_stamp = tcp_time_stamp;
1506 /* Network starves. */
1507 if (tp->packets_out > tp->snd_cwnd_used)
1508 tp->snd_cwnd_used = tp->packets_out;
1510 if (sysctl_tcp_slow_start_after_idle &&
1511 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1512 tcp_cwnd_application_limited(sk);
1516 /* Returns the portion of skb which can be sent right away without
1517 * introducing MSS oddities to segment boundaries. In rare cases where
1518 * mss_now != mss_cache, we will request caller to create a small skb
1519 * per input skb which could be mostly avoided here (if desired).
1521 * We explicitly want to create a request for splitting write queue tail
1522 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1523 * thus all the complexity (cwnd_len is always MSS multiple which we
1524 * return whenever allowed by the other factors). Basically we need the
1525 * modulo only when the receiver window alone is the limiting factor or
1526 * when we would be allowed to send the split-due-to-Nagle skb fully.
1528 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1529 unsigned int mss_now, unsigned int max_segs)
1531 const struct tcp_sock *tp = tcp_sk(sk);
1532 u32 needed, window, max_len;
1534 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1535 max_len = mss_now * max_segs;
1537 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1540 needed = min(skb->len, window);
1542 if (max_len <= needed)
1545 return needed - needed % mss_now;
1548 /* Can at least one segment of SKB be sent right now, according to the
1549 * congestion window rules? If so, return how many segments are allowed.
1551 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1552 const struct sk_buff *skb)
1554 u32 in_flight, cwnd;
1556 /* Don't be strict about the congestion window for the final FIN. */
1557 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1558 tcp_skb_pcount(skb) == 1)
1561 in_flight = tcp_packets_in_flight(tp);
1562 cwnd = tp->snd_cwnd;
1563 if (in_flight < cwnd)
1564 return (cwnd - in_flight);
1569 /* Initialize TSO state of a skb.
1570 * This must be invoked the first time we consider transmitting
1571 * SKB onto the wire.
1573 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1574 unsigned int mss_now)
1576 int tso_segs = tcp_skb_pcount(skb);
1578 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1579 tcp_set_skb_tso_segs(sk, skb, mss_now);
1580 tso_segs = tcp_skb_pcount(skb);
1585 /* Minshall's variant of the Nagle send check. */
1586 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1588 return after(tp->snd_sml, tp->snd_una) &&
1589 !after(tp->snd_sml, tp->snd_nxt);
1592 /* Return false, if packet can be sent now without violation Nagle's rules:
1593 * 1. It is full sized.
1594 * 2. Or it contains FIN. (already checked by caller)
1595 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1596 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1597 * With Minshall's modification: all sent small packets are ACKed.
1599 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1600 const struct sk_buff *skb,
1601 unsigned int mss_now, int nonagle)
1603 return skb->len < mss_now &&
1604 ((nonagle & TCP_NAGLE_CORK) ||
1605 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1608 /* Return true if the Nagle test allows this packet to be
1611 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1612 unsigned int cur_mss, int nonagle)
1614 /* Nagle rule does not apply to frames, which sit in the middle of the
1615 * write_queue (they have no chances to get new data).
1617 * This is implemented in the callers, where they modify the 'nonagle'
1618 * argument based upon the location of SKB in the send queue.
1620 if (nonagle & TCP_NAGLE_PUSH)
1623 /* Don't use the nagle rule for urgent data (or for the final FIN).
1624 * Nagle can be ignored during F-RTO too (see RFC4138).
1626 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1627 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1630 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1636 /* Does at least the first segment of SKB fit into the send window? */
1637 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1638 const struct sk_buff *skb,
1639 unsigned int cur_mss)
1641 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1643 if (skb->len > cur_mss)
1644 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1646 return !after(end_seq, tcp_wnd_end(tp));
1649 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1650 * should be put on the wire right now. If so, it returns the number of
1651 * packets allowed by the congestion window.
1653 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1654 unsigned int cur_mss, int nonagle)
1656 const struct tcp_sock *tp = tcp_sk(sk);
1657 unsigned int cwnd_quota;
1659 tcp_init_tso_segs(sk, skb, cur_mss);
1661 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1664 cwnd_quota = tcp_cwnd_test(tp, skb);
1665 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1671 /* Test if sending is allowed right now. */
1672 bool tcp_may_send_now(struct sock *sk)
1674 const struct tcp_sock *tp = tcp_sk(sk);
1675 struct sk_buff *skb = tcp_send_head(sk);
1678 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1679 (tcp_skb_is_last(sk, skb) ?
1680 tp->nonagle : TCP_NAGLE_PUSH));
1683 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1684 * which is put after SKB on the list. It is very much like
1685 * tcp_fragment() except that it may make several kinds of assumptions
1686 * in order to speed up the splitting operation. In particular, we
1687 * know that all the data is in scatter-gather pages, and that the
1688 * packet has never been sent out before (and thus is not cloned).
1690 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1691 unsigned int mss_now, gfp_t gfp)
1693 struct sk_buff *buff;
1694 int nlen = skb->len - len;
1697 /* All of a TSO frame must be composed of paged data. */
1698 if (skb->len != skb->data_len)
1699 return tcp_fragment(sk, skb, len, mss_now);
1701 buff = sk_stream_alloc_skb(sk, 0, gfp);
1702 if (unlikely(buff == NULL))
1705 sk->sk_wmem_queued += buff->truesize;
1706 sk_mem_charge(sk, buff->truesize);
1707 buff->truesize += nlen;
1708 skb->truesize -= nlen;
1710 /* Correct the sequence numbers. */
1711 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1712 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1713 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1715 /* PSH and FIN should only be set in the second packet. */
1716 flags = TCP_SKB_CB(skb)->tcp_flags;
1717 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1718 TCP_SKB_CB(buff)->tcp_flags = flags;
1720 /* This packet was never sent out yet, so no SACK bits. */
1721 TCP_SKB_CB(buff)->sacked = 0;
1723 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1724 skb_split(skb, buff, len);
1726 /* Fix up tso_factor for both original and new SKB. */
1727 tcp_set_skb_tso_segs(sk, skb, mss_now);
1728 tcp_set_skb_tso_segs(sk, buff, mss_now);
1730 /* Link BUFF into the send queue. */
1731 skb_header_release(buff);
1732 tcp_insert_write_queue_after(skb, buff, sk);
1737 /* Try to defer sending, if possible, in order to minimize the amount
1738 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1740 * This algorithm is from John Heffner.
1742 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1744 struct tcp_sock *tp = tcp_sk(sk);
1745 const struct inet_connection_sock *icsk = inet_csk(sk);
1746 u32 send_win, cong_win, limit, in_flight;
1749 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1752 if (icsk->icsk_ca_state != TCP_CA_Open)
1755 /* Defer for less than two clock ticks. */
1756 if (tp->tso_deferred &&
1757 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1760 in_flight = tcp_packets_in_flight(tp);
1762 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1764 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1766 /* From in_flight test above, we know that cwnd > in_flight. */
1767 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1769 limit = min(send_win, cong_win);
1771 /* If a full-sized TSO skb can be sent, do it. */
1772 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1773 sk->sk_gso_max_segs * tp->mss_cache))
1776 /* Middle in queue won't get any more data, full sendable already? */
1777 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1780 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1782 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1784 /* If at least some fraction of a window is available,
1787 chunk /= win_divisor;
1791 /* Different approach, try not to defer past a single
1792 * ACK. Receiver should ACK every other full sized
1793 * frame, so if we have space for more than 3 frames
1796 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1800 /* Ok, it looks like it is advisable to defer. */
1801 tp->tso_deferred = 1 | (jiffies << 1);
1806 tp->tso_deferred = 0;
1810 /* Create a new MTU probe if we are ready.
1811 * MTU probe is regularly attempting to increase the path MTU by
1812 * deliberately sending larger packets. This discovers routing
1813 * changes resulting in larger path MTUs.
1815 * Returns 0 if we should wait to probe (no cwnd available),
1816 * 1 if a probe was sent,
1819 static int tcp_mtu_probe(struct sock *sk)
1821 struct tcp_sock *tp = tcp_sk(sk);
1822 struct inet_connection_sock *icsk = inet_csk(sk);
1823 struct sk_buff *skb, *nskb, *next;
1830 /* Not currently probing/verifying,
1832 * have enough cwnd, and
1833 * not SACKing (the variable headers throw things off) */
1834 if (!icsk->icsk_mtup.enabled ||
1835 icsk->icsk_mtup.probe_size ||
1836 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1837 tp->snd_cwnd < 11 ||
1838 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1841 /* Very simple search strategy: just double the MSS. */
1842 mss_now = tcp_current_mss(sk);
1843 probe_size = 2 * tp->mss_cache;
1844 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1845 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1846 /* TODO: set timer for probe_converge_event */
1850 /* Have enough data in the send queue to probe? */
1851 if (tp->write_seq - tp->snd_nxt < size_needed)
1854 if (tp->snd_wnd < size_needed)
1856 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1859 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1860 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1861 if (!tcp_packets_in_flight(tp))
1867 /* We're allowed to probe. Build it now. */
1868 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1870 sk->sk_wmem_queued += nskb->truesize;
1871 sk_mem_charge(sk, nskb->truesize);
1873 skb = tcp_send_head(sk);
1875 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1876 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1877 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1878 TCP_SKB_CB(nskb)->sacked = 0;
1880 nskb->ip_summed = skb->ip_summed;
1882 tcp_insert_write_queue_before(nskb, skb, sk);
1885 tcp_for_write_queue_from_safe(skb, next, sk) {
1886 copy = min_t(int, skb->len, probe_size - len);
1887 if (nskb->ip_summed)
1888 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1890 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1891 skb_put(nskb, copy),
1894 if (skb->len <= copy) {
1895 /* We've eaten all the data from this skb.
1897 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1898 tcp_unlink_write_queue(skb, sk);
1899 sk_wmem_free_skb(sk, skb);
1901 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1902 ~(TCPHDR_FIN|TCPHDR_PSH);
1903 if (!skb_shinfo(skb)->nr_frags) {
1904 skb_pull(skb, copy);
1905 if (skb->ip_summed != CHECKSUM_PARTIAL)
1906 skb->csum = csum_partial(skb->data,
1909 __pskb_trim_head(skb, copy);
1910 tcp_set_skb_tso_segs(sk, skb, mss_now);
1912 TCP_SKB_CB(skb)->seq += copy;
1917 if (len >= probe_size)
1920 tcp_init_tso_segs(sk, nskb, nskb->len);
1922 /* We're ready to send. If this fails, the probe will
1923 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1924 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1925 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1926 /* Decrement cwnd here because we are sending
1927 * effectively two packets. */
1929 tcp_event_new_data_sent(sk, nskb);
1931 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1932 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1933 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1941 /* This routine writes packets to the network. It advances the
1942 * send_head. This happens as incoming acks open up the remote
1945 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1946 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1947 * account rare use of URG, this is not a big flaw.
1949 * Returns true, if no segments are in flight and we have queued segments,
1950 * but cannot send anything now because of SWS or another problem.
1952 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1953 int push_one, gfp_t gfp)
1955 struct tcp_sock *tp = tcp_sk(sk);
1956 struct sk_buff *skb;
1957 unsigned int tso_segs, sent_pkts;
1964 /* Do MTU probing. */
1965 result = tcp_mtu_probe(sk);
1968 } else if (result > 0) {
1973 while ((skb = tcp_send_head(sk))) {
1977 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1980 cwnd_quota = tcp_cwnd_test(tp, skb);
1984 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1987 if (tso_segs == 1) {
1988 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1989 (tcp_skb_is_last(sk, skb) ?
1990 nonagle : TCP_NAGLE_PUSH))))
1993 if (!push_one && tcp_tso_should_defer(sk, skb))
1997 /* TSQ : sk_wmem_alloc accounts skb truesize,
1998 * including skb overhead. But thats OK.
2000 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
2001 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2005 if (tso_segs > 1 && !tcp_urg_mode(tp))
2006 limit = tcp_mss_split_point(sk, skb, mss_now,
2009 sk->sk_gso_max_segs));
2011 if (skb->len > limit &&
2012 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2015 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2017 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2020 /* Advance the send_head. This one is sent out.
2021 * This call will increment packets_out.
2023 tcp_event_new_data_sent(sk, skb);
2025 tcp_minshall_update(tp, mss_now, skb);
2026 sent_pkts += tcp_skb_pcount(skb);
2031 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2032 tp->prr_out += sent_pkts;
2034 if (likely(sent_pkts)) {
2035 tcp_cwnd_validate(sk);
2038 return !tp->packets_out && tcp_send_head(sk);
2041 /* Push out any pending frames which were held back due to
2042 * TCP_CORK or attempt at coalescing tiny packets.
2043 * The socket must be locked by the caller.
2045 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2048 /* If we are closed, the bytes will have to remain here.
2049 * In time closedown will finish, we empty the write queue and
2050 * all will be happy.
2052 if (unlikely(sk->sk_state == TCP_CLOSE))
2055 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2056 sk_gfp_atomic(sk, GFP_ATOMIC)))
2057 tcp_check_probe_timer(sk);
2060 /* Send _single_ skb sitting at the send head. This function requires
2061 * true push pending frames to setup probe timer etc.
2063 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2065 struct sk_buff *skb = tcp_send_head(sk);
2067 BUG_ON(!skb || skb->len < mss_now);
2069 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2072 /* This function returns the amount that we can raise the
2073 * usable window based on the following constraints
2075 * 1. The window can never be shrunk once it is offered (RFC 793)
2076 * 2. We limit memory per socket
2079 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2080 * RECV.NEXT + RCV.WIN fixed until:
2081 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2083 * i.e. don't raise the right edge of the window until you can raise
2084 * it at least MSS bytes.
2086 * Unfortunately, the recommended algorithm breaks header prediction,
2087 * since header prediction assumes th->window stays fixed.
2089 * Strictly speaking, keeping th->window fixed violates the receiver
2090 * side SWS prevention criteria. The problem is that under this rule
2091 * a stream of single byte packets will cause the right side of the
2092 * window to always advance by a single byte.
2094 * Of course, if the sender implements sender side SWS prevention
2095 * then this will not be a problem.
2097 * BSD seems to make the following compromise:
2099 * If the free space is less than the 1/4 of the maximum
2100 * space available and the free space is less than 1/2 mss,
2101 * then set the window to 0.
2102 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2103 * Otherwise, just prevent the window from shrinking
2104 * and from being larger than the largest representable value.
2106 * This prevents incremental opening of the window in the regime
2107 * where TCP is limited by the speed of the reader side taking
2108 * data out of the TCP receive queue. It does nothing about
2109 * those cases where the window is constrained on the sender side
2110 * because the pipeline is full.
2112 * BSD also seems to "accidentally" limit itself to windows that are a
2113 * multiple of MSS, at least until the free space gets quite small.
2114 * This would appear to be a side effect of the mbuf implementation.
2115 * Combining these two algorithms results in the observed behavior
2116 * of having a fixed window size at almost all times.
2118 * Below we obtain similar behavior by forcing the offered window to
2119 * a multiple of the mss when it is feasible to do so.
2121 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2122 * Regular options like TIMESTAMP are taken into account.
2124 u32 __tcp_select_window(struct sock *sk)
2126 struct inet_connection_sock *icsk = inet_csk(sk);
2127 struct tcp_sock *tp = tcp_sk(sk);
2128 /* MSS for the peer's data. Previous versions used mss_clamp
2129 * here. I don't know if the value based on our guesses
2130 * of peer's MSS is better for the performance. It's more correct
2131 * but may be worse for the performance because of rcv_mss
2132 * fluctuations. --SAW 1998/11/1
2134 int mss = icsk->icsk_ack.rcv_mss;
2135 int free_space = tcp_space(sk);
2136 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2139 if (mss > full_space)
2142 if (free_space < (full_space >> 1)) {
2143 icsk->icsk_ack.quick = 0;
2145 if (sk_under_memory_pressure(sk))
2146 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2149 if (free_space < mss)
2153 if (free_space > tp->rcv_ssthresh)
2154 free_space = tp->rcv_ssthresh;
2156 /* Don't do rounding if we are using window scaling, since the
2157 * scaled window will not line up with the MSS boundary anyway.
2159 window = tp->rcv_wnd;
2160 if (tp->rx_opt.rcv_wscale) {
2161 window = free_space;
2163 /* Advertise enough space so that it won't get scaled away.
2164 * Import case: prevent zero window announcement if
2165 * 1<<rcv_wscale > mss.
2167 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2168 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2169 << tp->rx_opt.rcv_wscale);
2171 /* Get the largest window that is a nice multiple of mss.
2172 * Window clamp already applied above.
2173 * If our current window offering is within 1 mss of the
2174 * free space we just keep it. This prevents the divide
2175 * and multiply from happening most of the time.
2176 * We also don't do any window rounding when the free space
2179 if (window <= free_space - mss || window > free_space)
2180 window = (free_space / mss) * mss;
2181 else if (mss == full_space &&
2182 free_space > window + (full_space >> 1))
2183 window = free_space;
2189 /* Collapses two adjacent SKB's during retransmission. */
2190 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2194 int skb_size, next_skb_size;
2196 skb_size = skb->len;
2197 next_skb_size = next_skb->len;
2199 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2201 tcp_highest_sack_combine(sk, next_skb, skb);
2203 tcp_unlink_write_queue(next_skb, sk);
2205 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2208 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2209 skb->ip_summed = CHECKSUM_PARTIAL;
2211 if (skb->ip_summed != CHECKSUM_PARTIAL)
2212 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2214 /* Update sequence range on original skb. */
2215 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2217 /* Merge over control information. This moves PSH/FIN etc. over */
2218 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2220 /* All done, get rid of second SKB and account for it so
2221 * packet counting does not break.
2223 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2225 /* changed transmit queue under us so clear hints */
2226 tcp_clear_retrans_hints_partial(tp);
2227 if (next_skb == tp->retransmit_skb_hint)
2228 tp->retransmit_skb_hint = skb;
2230 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2232 sk_wmem_free_skb(sk, next_skb);
2235 /* Check if coalescing SKBs is legal. */
2236 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2238 if (tcp_skb_pcount(skb) > 1)
2240 /* TODO: SACK collapsing could be used to remove this condition */
2241 if (skb_shinfo(skb)->nr_frags != 0)
2243 if (skb_cloned(skb))
2245 if (skb == tcp_send_head(sk))
2247 /* Some heurestics for collapsing over SACK'd could be invented */
2248 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2254 /* Collapse packets in the retransmit queue to make to create
2255 * less packets on the wire. This is only done on retransmission.
2257 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2260 struct tcp_sock *tp = tcp_sk(sk);
2261 struct sk_buff *skb = to, *tmp;
2264 if (!sysctl_tcp_retrans_collapse)
2266 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2269 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2270 if (!tcp_can_collapse(sk, skb))
2282 /* Punt if not enough space exists in the first SKB for
2283 * the data in the second
2285 if (skb->len > skb_availroom(to))
2288 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2291 tcp_collapse_retrans(sk, to);
2295 /* This retransmits one SKB. Policy decisions and retransmit queue
2296 * state updates are done by the caller. Returns non-zero if an
2297 * error occurred which prevented the send.
2299 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2301 struct tcp_sock *tp = tcp_sk(sk);
2302 struct inet_connection_sock *icsk = inet_csk(sk);
2303 unsigned int cur_mss;
2306 /* Inconslusive MTU probe */
2307 if (icsk->icsk_mtup.probe_size) {
2308 icsk->icsk_mtup.probe_size = 0;
2311 /* Do not sent more than we queued. 1/4 is reserved for possible
2312 * copying overhead: fragmentation, tunneling, mangling etc.
2314 if (atomic_read(&sk->sk_wmem_alloc) >
2315 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2318 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2319 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2321 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2325 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2326 return -EHOSTUNREACH; /* Routing failure or similar. */
2328 cur_mss = tcp_current_mss(sk);
2330 /* If receiver has shrunk his window, and skb is out of
2331 * new window, do not retransmit it. The exception is the
2332 * case, when window is shrunk to zero. In this case
2333 * our retransmit serves as a zero window probe.
2335 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2336 TCP_SKB_CB(skb)->seq != tp->snd_una)
2339 if (skb->len > cur_mss) {
2340 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2341 return -ENOMEM; /* We'll try again later. */
2343 int oldpcount = tcp_skb_pcount(skb);
2345 if (unlikely(oldpcount > 1)) {
2346 tcp_init_tso_segs(sk, skb, cur_mss);
2347 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2351 tcp_retrans_try_collapse(sk, skb, cur_mss);
2353 /* Some Solaris stacks overoptimize and ignore the FIN on a
2354 * retransmit when old data is attached. So strip it off
2355 * since it is cheap to do so and saves bytes on the network.
2358 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2359 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2360 if (!pskb_trim(skb, 0)) {
2361 /* Reuse, even though it does some unnecessary work */
2362 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2363 TCP_SKB_CB(skb)->tcp_flags);
2364 skb->ip_summed = CHECKSUM_NONE;
2368 /* Make a copy, if the first transmission SKB clone we made
2369 * is still in somebody's hands, else make a clone.
2371 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2373 /* make sure skb->data is aligned on arches that require it */
2374 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2375 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2377 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2380 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2384 /* Update global TCP statistics. */
2385 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2387 tp->total_retrans++;
2389 #if FASTRETRANS_DEBUG > 0
2390 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2391 net_dbg_ratelimited("retrans_out leaked\n");
2394 if (!tp->retrans_out)
2395 tp->lost_retrans_low = tp->snd_nxt;
2396 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2397 tp->retrans_out += tcp_skb_pcount(skb);
2399 /* Save stamp of the first retransmit. */
2400 if (!tp->retrans_stamp)
2401 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2403 tp->undo_retrans += tcp_skb_pcount(skb);
2405 /* snd_nxt is stored to detect loss of retransmitted segment,
2406 * see tcp_input.c tcp_sacktag_write_queue().
2408 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2413 /* Check if we forward retransmits are possible in the current
2414 * window/congestion state.
2416 static bool tcp_can_forward_retransmit(struct sock *sk)
2418 const struct inet_connection_sock *icsk = inet_csk(sk);
2419 const struct tcp_sock *tp = tcp_sk(sk);
2421 /* Forward retransmissions are possible only during Recovery. */
2422 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2425 /* No forward retransmissions in Reno are possible. */
2426 if (tcp_is_reno(tp))
2429 /* Yeah, we have to make difficult choice between forward transmission
2430 * and retransmission... Both ways have their merits...
2432 * For now we do not retransmit anything, while we have some new
2433 * segments to send. In the other cases, follow rule 3 for
2434 * NextSeg() specified in RFC3517.
2437 if (tcp_may_send_now(sk))
2443 /* This gets called after a retransmit timeout, and the initially
2444 * retransmitted data is acknowledged. It tries to continue
2445 * resending the rest of the retransmit queue, until either
2446 * we've sent it all or the congestion window limit is reached.
2447 * If doing SACK, the first ACK which comes back for a timeout
2448 * based retransmit packet might feed us FACK information again.
2449 * If so, we use it to avoid unnecessarily retransmissions.
2451 void tcp_xmit_retransmit_queue(struct sock *sk)
2453 const struct inet_connection_sock *icsk = inet_csk(sk);
2454 struct tcp_sock *tp = tcp_sk(sk);
2455 struct sk_buff *skb;
2456 struct sk_buff *hole = NULL;
2459 int fwd_rexmitting = 0;
2461 if (!tp->packets_out)
2465 tp->retransmit_high = tp->snd_una;
2467 if (tp->retransmit_skb_hint) {
2468 skb = tp->retransmit_skb_hint;
2469 last_lost = TCP_SKB_CB(skb)->end_seq;
2470 if (after(last_lost, tp->retransmit_high))
2471 last_lost = tp->retransmit_high;
2473 skb = tcp_write_queue_head(sk);
2474 last_lost = tp->snd_una;
2477 tcp_for_write_queue_from(skb, sk) {
2478 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2480 if (skb == tcp_send_head(sk))
2482 /* we could do better than to assign each time */
2484 tp->retransmit_skb_hint = skb;
2486 /* Assume this retransmit will generate
2487 * only one packet for congestion window
2488 * calculation purposes. This works because
2489 * tcp_retransmit_skb() will chop up the
2490 * packet to be MSS sized and all the
2491 * packet counting works out.
2493 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2496 if (fwd_rexmitting) {
2498 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2500 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2502 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2503 tp->retransmit_high = last_lost;
2504 if (!tcp_can_forward_retransmit(sk))
2506 /* Backtrack if necessary to non-L'ed skb */
2514 } else if (!(sacked & TCPCB_LOST)) {
2515 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2520 last_lost = TCP_SKB_CB(skb)->end_seq;
2521 if (icsk->icsk_ca_state != TCP_CA_Loss)
2522 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2524 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2527 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2530 if (tcp_retransmit_skb(sk, skb)) {
2531 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2534 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2536 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2537 tp->prr_out += tcp_skb_pcount(skb);
2539 if (skb == tcp_write_queue_head(sk))
2540 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2541 inet_csk(sk)->icsk_rto,
2546 /* Send a fin. The caller locks the socket for us. This cannot be
2547 * allowed to fail queueing a FIN frame under any circumstances.
2549 void tcp_send_fin(struct sock *sk)
2551 struct tcp_sock *tp = tcp_sk(sk);
2552 struct sk_buff *skb = tcp_write_queue_tail(sk);
2555 /* Optimization, tack on the FIN if we have a queue of
2556 * unsent frames. But be careful about outgoing SACKS
2559 mss_now = tcp_current_mss(sk);
2561 if (tcp_send_head(sk) != NULL) {
2562 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2563 TCP_SKB_CB(skb)->end_seq++;
2566 /* Socket is locked, keep trying until memory is available. */
2568 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2575 /* Reserve space for headers and prepare control bits. */
2576 skb_reserve(skb, MAX_TCP_HEADER);
2577 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2578 tcp_init_nondata_skb(skb, tp->write_seq,
2579 TCPHDR_ACK | TCPHDR_FIN);
2580 tcp_queue_skb(sk, skb);
2582 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2585 /* We get here when a process closes a file descriptor (either due to
2586 * an explicit close() or as a byproduct of exit()'ing) and there
2587 * was unread data in the receive queue. This behavior is recommended
2588 * by RFC 2525, section 2.17. -DaveM
2590 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2592 struct sk_buff *skb;
2594 /* NOTE: No TCP options attached and we never retransmit this. */
2595 skb = alloc_skb(MAX_TCP_HEADER, priority);
2597 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2601 /* Reserve space for headers and prepare control bits. */
2602 skb_reserve(skb, MAX_TCP_HEADER);
2603 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2604 TCPHDR_ACK | TCPHDR_RST);
2606 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2607 if (tcp_transmit_skb(sk, skb, 0, priority))
2608 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2610 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2613 /* Send a crossed SYN-ACK during socket establishment.
2614 * WARNING: This routine must only be called when we have already sent
2615 * a SYN packet that crossed the incoming SYN that caused this routine
2616 * to get called. If this assumption fails then the initial rcv_wnd
2617 * and rcv_wscale values will not be correct.
2619 int tcp_send_synack(struct sock *sk)
2621 struct sk_buff *skb;
2623 skb = tcp_write_queue_head(sk);
2624 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2625 pr_debug("%s: wrong queue state\n", __func__);
2628 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2629 if (skb_cloned(skb)) {
2630 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2633 tcp_unlink_write_queue(skb, sk);
2634 skb_header_release(nskb);
2635 __tcp_add_write_queue_head(sk, nskb);
2636 sk_wmem_free_skb(sk, skb);
2637 sk->sk_wmem_queued += nskb->truesize;
2638 sk_mem_charge(sk, nskb->truesize);
2642 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2643 TCP_ECN_send_synack(tcp_sk(sk), skb);
2645 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2646 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2650 * tcp_make_synack - Prepare a SYN-ACK.
2651 * sk: listener socket
2652 * dst: dst entry attached to the SYNACK
2653 * req: request_sock pointer
2654 * rvp: request_values pointer
2656 * Allocate one skb and build a SYNACK packet.
2657 * @dst is consumed : Caller should not use it again.
2659 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2660 struct request_sock *req,
2661 struct request_values *rvp)
2663 struct tcp_out_options opts;
2664 struct tcp_extend_values *xvp = tcp_xv(rvp);
2665 struct inet_request_sock *ireq = inet_rsk(req);
2666 struct tcp_sock *tp = tcp_sk(sk);
2667 const struct tcp_cookie_values *cvp = tp->cookie_values;
2669 struct sk_buff *skb;
2670 struct tcp_md5sig_key *md5;
2671 int tcp_header_size;
2673 int s_data_desired = 0;
2675 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2676 s_data_desired = cvp->s_data_desired;
2677 skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired,
2678 sk_gfp_atomic(sk, GFP_ATOMIC));
2679 if (unlikely(!skb)) {
2683 /* Reserve space for headers. */
2684 skb_reserve(skb, MAX_TCP_HEADER);
2686 skb_dst_set(skb, dst);
2688 mss = dst_metric_advmss(dst);
2689 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2690 mss = tp->rx_opt.user_mss;
2692 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2694 /* Set this up on the first call only */
2695 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2697 /* limit the window selection if the user enforce a smaller rx buffer */
2698 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2699 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2700 req->window_clamp = tcp_full_space(sk);
2702 /* tcp_full_space because it is guaranteed to be the first packet */
2703 tcp_select_initial_window(tcp_full_space(sk),
2704 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2709 dst_metric(dst, RTAX_INITRWND));
2710 ireq->rcv_wscale = rcv_wscale;
2713 memset(&opts, 0, sizeof(opts));
2714 #ifdef CONFIG_SYN_COOKIES
2715 if (unlikely(req->cookie_ts))
2716 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2719 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2720 tcp_header_size = tcp_synack_options(sk, req, mss,
2721 skb, &opts, &md5, xvp)
2724 skb_push(skb, tcp_header_size);
2725 skb_reset_transport_header(skb);
2728 memset(th, 0, sizeof(struct tcphdr));
2731 TCP_ECN_make_synack(req, th);
2732 th->source = ireq->loc_port;
2733 th->dest = ireq->rmt_port;
2734 /* Setting of flags are superfluous here for callers (and ECE is
2735 * not even correctly set)
2737 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2738 TCPHDR_SYN | TCPHDR_ACK);
2740 if (OPTION_COOKIE_EXTENSION & opts.options) {
2741 if (s_data_desired) {
2742 u8 *buf = skb_put(skb, s_data_desired);
2744 /* copy data directly from the listening socket. */
2745 memcpy(buf, cvp->s_data_payload, s_data_desired);
2746 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2749 if (opts.hash_size > 0) {
2750 __u32 workspace[SHA_WORKSPACE_WORDS];
2751 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2752 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2754 /* Secret recipe depends on the Timestamp, (future)
2755 * Sequence and Acknowledgment Numbers, Initiator
2756 * Cookie, and others handled by IP variant caller.
2758 *tail-- ^= opts.tsval;
2759 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2760 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2763 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2764 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2766 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2769 opts.hash_location =
2770 (__u8 *)&xvp->cookie_bakery[0];
2774 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2775 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2777 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2778 th->window = htons(min(req->rcv_wnd, 65535U));
2779 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2780 th->doff = (tcp_header_size >> 2);
2781 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2783 #ifdef CONFIG_TCP_MD5SIG
2784 /* Okay, we have all we need - do the md5 hash if needed */
2786 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2787 md5, NULL, req, skb);
2793 EXPORT_SYMBOL(tcp_make_synack);
2795 /* Do all connect socket setups that can be done AF independent. */
2796 void tcp_connect_init(struct sock *sk)
2798 const struct dst_entry *dst = __sk_dst_get(sk);
2799 struct tcp_sock *tp = tcp_sk(sk);
2802 /* We'll fix this up when we get a response from the other end.
2803 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2805 tp->tcp_header_len = sizeof(struct tcphdr) +
2806 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2808 #ifdef CONFIG_TCP_MD5SIG
2809 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2810 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2813 /* If user gave his TCP_MAXSEG, record it to clamp */
2814 if (tp->rx_opt.user_mss)
2815 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2818 tcp_sync_mss(sk, dst_mtu(dst));
2820 if (!tp->window_clamp)
2821 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2822 tp->advmss = dst_metric_advmss(dst);
2823 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2824 tp->advmss = tp->rx_opt.user_mss;
2826 tcp_initialize_rcv_mss(sk);
2828 /* limit the window selection if the user enforce a smaller rx buffer */
2829 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2830 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2831 tp->window_clamp = tcp_full_space(sk);
2833 tcp_select_initial_window(tcp_full_space(sk),
2834 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2837 sysctl_tcp_window_scaling,
2839 dst_metric(dst, RTAX_INITRWND));
2841 tp->rx_opt.rcv_wscale = rcv_wscale;
2842 tp->rcv_ssthresh = tp->rcv_wnd;
2845 sock_reset_flag(sk, SOCK_DONE);
2848 tp->snd_una = tp->write_seq;
2849 tp->snd_sml = tp->write_seq;
2850 tp->snd_up = tp->write_seq;
2851 tp->snd_nxt = tp->write_seq;
2853 if (likely(!tp->repair))
2855 tp->rcv_wup = tp->rcv_nxt;
2856 tp->copied_seq = tp->rcv_nxt;
2858 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2859 inet_csk(sk)->icsk_retransmits = 0;
2860 tcp_clear_retrans(tp);
2863 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2865 struct tcp_sock *tp = tcp_sk(sk);
2866 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2868 tcb->end_seq += skb->len;
2869 skb_header_release(skb);
2870 __tcp_add_write_queue_tail(sk, skb);
2871 sk->sk_wmem_queued += skb->truesize;
2872 sk_mem_charge(sk, skb->truesize);
2873 tp->write_seq = tcb->end_seq;
2874 tp->packets_out += tcp_skb_pcount(skb);
2877 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2878 * queue a data-only packet after the regular SYN, such that regular SYNs
2879 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2880 * only the SYN sequence, the data are retransmitted in the first ACK.
2881 * If cookie is not cached or other error occurs, falls back to send a
2882 * regular SYN with Fast Open cookie request option.
2884 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2886 struct tcp_sock *tp = tcp_sk(sk);
2887 struct tcp_fastopen_request *fo = tp->fastopen_req;
2888 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2889 struct sk_buff *syn_data = NULL, *data;
2890 unsigned long last_syn_loss = 0;
2892 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2893 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2894 &syn_loss, &last_syn_loss);
2895 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2897 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2898 fo->cookie.len = -1;
2902 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2903 fo->cookie.len = -1;
2904 else if (fo->cookie.len <= 0)
2907 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2908 * user-MSS. Reserve maximum option space for middleboxes that add
2909 * private TCP options. The cost is reduced data space in SYN :(
2911 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2912 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2913 space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2914 MAX_TCP_OPTION_SPACE;
2916 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2918 if (syn_data == NULL)
2921 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2922 struct iovec *iov = &fo->data->msg_iov[i];
2923 unsigned char __user *from = iov->iov_base;
2924 int len = iov->iov_len;
2926 if (syn_data->len + len > space)
2927 len = space - syn_data->len;
2928 else if (i + 1 == iovlen)
2929 /* No more data pending in inet_wait_for_connect() */
2932 if (skb_add_data(syn_data, from, len))
2936 /* Queue a data-only packet after the regular SYN for retransmission */
2937 data = pskb_copy(syn_data, sk->sk_allocation);
2940 TCP_SKB_CB(data)->seq++;
2941 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2942 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2943 tcp_connect_queue_skb(sk, data);
2944 fo->copied = data->len;
2946 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2947 tp->syn_data = (fo->copied > 0);
2948 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2954 /* Send a regular SYN with Fast Open cookie request option */
2955 if (fo->cookie.len > 0)
2957 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2959 tp->syn_fastopen = 0;
2960 kfree_skb(syn_data);
2962 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2966 /* Build a SYN and send it off. */
2967 int tcp_connect(struct sock *sk)
2969 struct tcp_sock *tp = tcp_sk(sk);
2970 struct sk_buff *buff;
2973 tcp_connect_init(sk);
2975 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2976 if (unlikely(buff == NULL))
2979 /* Reserve space for headers. */
2980 skb_reserve(buff, MAX_TCP_HEADER);
2982 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2983 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2984 tcp_connect_queue_skb(sk, buff);
2985 TCP_ECN_send_syn(sk, buff);
2987 /* Send off SYN; include data in Fast Open. */
2988 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2989 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2990 if (err == -ECONNREFUSED)
2993 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2994 * in order to make this packet get counted in tcpOutSegs.
2996 tp->snd_nxt = tp->write_seq;
2997 tp->pushed_seq = tp->write_seq;
2998 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3000 /* Timer for repeating the SYN until an answer. */
3001 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3002 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3005 EXPORT_SYMBOL(tcp_connect);
3007 /* Send out a delayed ack, the caller does the policy checking
3008 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3011 void tcp_send_delayed_ack(struct sock *sk)
3013 struct inet_connection_sock *icsk = inet_csk(sk);
3014 int ato = icsk->icsk_ack.ato;
3015 unsigned long timeout;
3017 if (ato > TCP_DELACK_MIN) {
3018 const struct tcp_sock *tp = tcp_sk(sk);
3019 int max_ato = HZ / 2;
3021 if (icsk->icsk_ack.pingpong ||
3022 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3023 max_ato = TCP_DELACK_MAX;
3025 /* Slow path, intersegment interval is "high". */
3027 /* If some rtt estimate is known, use it to bound delayed ack.
3028 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3032 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3038 ato = min(ato, max_ato);
3041 /* Stay within the limit we were given */
3042 timeout = jiffies + ato;
3044 /* Use new timeout only if there wasn't a older one earlier. */
3045 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3046 /* If delack timer was blocked or is about to expire,
3049 if (icsk->icsk_ack.blocked ||
3050 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3055 if (!time_before(timeout, icsk->icsk_ack.timeout))
3056 timeout = icsk->icsk_ack.timeout;
3058 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3059 icsk->icsk_ack.timeout = timeout;
3060 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3063 /* This routine sends an ack and also updates the window. */
3064 void tcp_send_ack(struct sock *sk)
3066 struct sk_buff *buff;
3068 /* If we have been reset, we may not send again. */
3069 if (sk->sk_state == TCP_CLOSE)
3072 /* We are not putting this on the write queue, so
3073 * tcp_transmit_skb() will set the ownership to this
3076 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3078 inet_csk_schedule_ack(sk);
3079 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3080 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3081 TCP_DELACK_MAX, TCP_RTO_MAX);
3085 /* Reserve space for headers and prepare control bits. */
3086 skb_reserve(buff, MAX_TCP_HEADER);
3087 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3089 /* Send it off, this clears delayed acks for us. */
3090 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3091 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3094 /* This routine sends a packet with an out of date sequence
3095 * number. It assumes the other end will try to ack it.
3097 * Question: what should we make while urgent mode?
3098 * 4.4BSD forces sending single byte of data. We cannot send
3099 * out of window data, because we have SND.NXT==SND.MAX...
3101 * Current solution: to send TWO zero-length segments in urgent mode:
3102 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3103 * out-of-date with SND.UNA-1 to probe window.
3105 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3107 struct tcp_sock *tp = tcp_sk(sk);
3108 struct sk_buff *skb;
3110 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3111 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3115 /* Reserve space for headers and set control bits. */
3116 skb_reserve(skb, MAX_TCP_HEADER);
3117 /* Use a previous sequence. This should cause the other
3118 * end to send an ack. Don't queue or clone SKB, just
3121 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3122 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3123 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3126 void tcp_send_window_probe(struct sock *sk)
3128 if (sk->sk_state == TCP_ESTABLISHED) {
3129 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3130 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3131 tcp_xmit_probe_skb(sk, 0);
3135 /* Initiate keepalive or window probe from timer. */
3136 int tcp_write_wakeup(struct sock *sk)
3138 struct tcp_sock *tp = tcp_sk(sk);
3139 struct sk_buff *skb;
3141 if (sk->sk_state == TCP_CLOSE)
3144 if ((skb = tcp_send_head(sk)) != NULL &&
3145 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3147 unsigned int mss = tcp_current_mss(sk);
3148 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3150 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3151 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3153 /* We are probing the opening of a window
3154 * but the window size is != 0
3155 * must have been a result SWS avoidance ( sender )
3157 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3159 seg_size = min(seg_size, mss);
3160 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3161 if (tcp_fragment(sk, skb, seg_size, mss))
3163 } else if (!tcp_skb_pcount(skb))
3164 tcp_set_skb_tso_segs(sk, skb, mss);
3166 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3167 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3168 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3170 tcp_event_new_data_sent(sk, skb);
3173 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3174 tcp_xmit_probe_skb(sk, 1);
3175 return tcp_xmit_probe_skb(sk, 0);
3179 /* A window probe timeout has occurred. If window is not closed send
3180 * a partial packet else a zero probe.
3182 void tcp_send_probe0(struct sock *sk)
3184 struct inet_connection_sock *icsk = inet_csk(sk);
3185 struct tcp_sock *tp = tcp_sk(sk);
3188 err = tcp_write_wakeup(sk);
3190 if (tp->packets_out || !tcp_send_head(sk)) {
3191 /* Cancel probe timer, if it is not required. */
3192 icsk->icsk_probes_out = 0;
3193 icsk->icsk_backoff = 0;
3198 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3199 icsk->icsk_backoff++;
3200 icsk->icsk_probes_out++;
3201 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3202 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3205 /* If packet was not sent due to local congestion,
3206 * do not backoff and do not remember icsk_probes_out.
3207 * Let local senders to fight for local resources.
3209 * Use accumulated backoff yet.
3211 if (!icsk->icsk_probes_out)
3212 icsk->icsk_probes_out = 1;
3213 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3214 min(icsk->icsk_rto << icsk->icsk_backoff,
3215 TCP_RESOURCE_PROBE_INTERVAL),