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 inet_connection_sock *icsk = inet_csk(sk);
78 struct tcp_sock *tp = tcp_sk(sk);
79 unsigned int prior_packets = tp->packets_out;
81 tcp_advance_send_head(sk, skb);
82 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
84 /* Don't override Nagle indefinitely with F-RTO */
85 if (tp->frto_counter == 2)
88 tp->packets_out += tcp_skb_pcount(skb);
89 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
90 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
94 /* SND.NXT, if window was not shrunk.
95 * If window has been shrunk, what should we make? It is not clear at all.
96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
98 * invalid. OK, let's make this for now:
100 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
102 const struct tcp_sock *tp = tcp_sk(sk);
104 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
107 return tcp_wnd_end(tp);
110 /* Calculate mss to advertise in SYN segment.
111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
113 * 1. It is independent of path mtu.
114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
116 * attached devices, because some buggy hosts are confused by
118 * 4. We do not make 3, we advertise MSS, calculated from first
119 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
120 * This may be overridden via information stored in routing table.
121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
122 * probably even Jumbo".
124 static __u16 tcp_advertise_mss(struct sock *sk)
126 struct tcp_sock *tp = tcp_sk(sk);
127 const struct dst_entry *dst = __sk_dst_get(sk);
128 int mss = tp->advmss;
131 unsigned int metric = dst_metric_advmss(dst);
142 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
143 * This is the first part of cwnd validation mechanism. */
144 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
146 struct tcp_sock *tp = tcp_sk(sk);
147 s32 delta = tcp_time_stamp - tp->lsndtime;
148 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
149 u32 cwnd = tp->snd_cwnd;
151 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
153 tp->snd_ssthresh = tcp_current_ssthresh(sk);
154 restart_cwnd = min(restart_cwnd, cwnd);
156 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
158 tp->snd_cwnd = max(cwnd, restart_cwnd);
159 tp->snd_cwnd_stamp = tcp_time_stamp;
160 tp->snd_cwnd_used = 0;
163 /* Congestion state accounting after a packet has been sent. */
164 static void tcp_event_data_sent(struct tcp_sock *tp,
167 struct inet_connection_sock *icsk = inet_csk(sk);
168 const u32 now = tcp_time_stamp;
170 if (sysctl_tcp_slow_start_after_idle &&
171 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
172 tcp_cwnd_restart(sk, __sk_dst_get(sk));
176 /* If it is a reply for ato after last received
177 * packet, enter pingpong mode.
179 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
180 icsk->icsk_ack.pingpong = 1;
183 /* Account for an ACK we sent. */
184 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
186 tcp_dec_quickack_mode(sk, pkts);
187 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 /* Determine a window scaling and initial window to offer.
191 * Based on the assumption that the given amount of space
192 * will be offered. Store the results in the tp structure.
193 * NOTE: for smooth operation initial space offering should
194 * be a multiple of mss if possible. We assume here that mss >= 1.
195 * This MUST be enforced by all callers.
197 void tcp_select_initial_window(int __space, __u32 mss,
198 __u32 *rcv_wnd, __u32 *window_clamp,
199 int wscale_ok, __u8 *rcv_wscale,
202 unsigned int space = (__space < 0 ? 0 : __space);
204 /* If no clamp set the clamp to the max possible scaled window */
205 if (*window_clamp == 0)
206 (*window_clamp) = (65535 << 14);
207 space = min(*window_clamp, space);
209 /* Quantize space offering to a multiple of mss if possible. */
211 space = (space / mss) * mss;
213 /* NOTE: offering an initial window larger than 32767
214 * will break some buggy TCP stacks. If the admin tells us
215 * it is likely we could be speaking with such a buggy stack
216 * we will truncate our initial window offering to 32K-1
217 * unless the remote has sent us a window scaling option,
218 * which we interpret as a sign the remote TCP is not
219 * misinterpreting the window field as a signed quantity.
221 if (sysctl_tcp_workaround_signed_windows)
222 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
228 /* Set window scaling on max possible window
229 * See RFC1323 for an explanation of the limit to 14
231 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
232 space = min_t(u32, space, *window_clamp);
233 while (space > 65535 && (*rcv_wscale) < 14) {
239 /* Set initial window to a value enough for senders starting with
240 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
241 * a limit on the initial window when mss is larger than 1460.
243 if (mss > (1 << *rcv_wscale)) {
244 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
247 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
248 /* when initializing use the value from init_rcv_wnd
249 * rather than the default from above
252 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
254 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
257 /* Set the clamp no higher than max representable value */
258 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
260 EXPORT_SYMBOL(tcp_select_initial_window);
262 /* Chose a new window to advertise, update state in tcp_sock for the
263 * socket, and return result with RFC1323 scaling applied. The return
264 * value can be stuffed directly into th->window for an outgoing
267 static u16 tcp_select_window(struct sock *sk)
269 struct tcp_sock *tp = tcp_sk(sk);
270 u32 cur_win = tcp_receive_window(tp);
271 u32 new_win = __tcp_select_window(sk);
273 /* Never shrink the offered window */
274 if (new_win < cur_win) {
275 /* Danger Will Robinson!
276 * Don't update rcv_wup/rcv_wnd here or else
277 * we will not be able to advertise a zero
278 * window in time. --DaveM
280 * Relax Will Robinson.
282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
287 /* Make sure we do not exceed the maximum possible
290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291 new_win = min(new_win, MAX_TCP_WINDOW);
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
298 /* If we advertise zero window, disable fast path. */
305 /* Packet ECN state for a SYN-ACK */
306 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
308 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
309 if (!(tp->ecn_flags & TCP_ECN_OK))
310 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
313 /* Packet ECN state for a SYN. */
314 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
316 struct tcp_sock *tp = tcp_sk(sk);
319 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
320 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
321 tp->ecn_flags = TCP_ECN_OK;
325 static __inline__ void
326 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
328 if (inet_rsk(req)->ecn_ok)
332 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
335 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
338 struct tcp_sock *tp = tcp_sk(sk);
340 if (tp->ecn_flags & TCP_ECN_OK) {
341 /* Not-retransmitted data segment: set ECT and inject CWR. */
342 if (skb->len != tcp_header_len &&
343 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
345 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
346 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
347 tcp_hdr(skb)->cwr = 1;
348 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
351 /* ACK or retransmitted segment: clear ECT|CE */
352 INET_ECN_dontxmit(sk);
354 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
355 tcp_hdr(skb)->ece = 1;
359 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
360 * auto increment end seqno.
362 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
364 skb->ip_summed = CHECKSUM_PARTIAL;
367 TCP_SKB_CB(skb)->tcp_flags = flags;
368 TCP_SKB_CB(skb)->sacked = 0;
370 skb_shinfo(skb)->gso_segs = 1;
371 skb_shinfo(skb)->gso_size = 0;
372 skb_shinfo(skb)->gso_type = 0;
374 TCP_SKB_CB(skb)->seq = seq;
375 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
377 TCP_SKB_CB(skb)->end_seq = seq;
380 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
382 return tp->snd_una != tp->snd_up;
385 #define OPTION_SACK_ADVERTISE (1 << 0)
386 #define OPTION_TS (1 << 1)
387 #define OPTION_MD5 (1 << 2)
388 #define OPTION_WSCALE (1 << 3)
389 #define OPTION_COOKIE_EXTENSION (1 << 4)
390 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
392 struct tcp_out_options {
393 u16 options; /* bit field of OPTION_* */
394 u16 mss; /* 0 to disable */
395 u8 ws; /* window scale, 0 to disable */
396 u8 num_sack_blocks; /* number of SACK blocks to include */
397 u8 hash_size; /* bytes in hash_location */
398 __u8 *hash_location; /* temporary pointer, overloaded */
399 __u32 tsval, tsecr; /* need to include OPTION_TS */
400 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
403 /* The sysctl int routines are generic, so check consistency here.
405 static u8 tcp_cookie_size_check(u8 desired)
410 /* previously specified */
413 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
414 if (cookie_size <= 0)
415 /* no default specified */
418 if (cookie_size <= TCP_COOKIE_MIN)
419 /* value too small, specify minimum */
420 return TCP_COOKIE_MIN;
422 if (cookie_size >= TCP_COOKIE_MAX)
423 /* value too large, specify maximum */
424 return TCP_COOKIE_MAX;
427 /* 8-bit multiple, illegal, fix it */
430 return (u8)cookie_size;
433 /* Write previously computed TCP options to the packet.
435 * Beware: Something in the Internet is very sensitive to the ordering of
436 * TCP options, we learned this through the hard way, so be careful here.
437 * Luckily we can at least blame others for their non-compliance but from
438 * inter-operatibility perspective it seems that we're somewhat stuck with
439 * the ordering which we have been using if we want to keep working with
440 * those broken things (not that it currently hurts anybody as there isn't
441 * particular reason why the ordering would need to be changed).
443 * At least SACK_PERM as the first option is known to lead to a disaster
444 * (but it may well be that other scenarios fail similarly).
446 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
447 struct tcp_out_options *opts)
449 u16 options = opts->options; /* mungable copy */
451 /* Having both authentication and cookies for security is redundant,
452 * and there's certainly not enough room. Instead, the cookie-less
453 * extension variant is proposed.
455 * Consider the pessimal case with authentication. The options
457 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
459 if (unlikely(OPTION_MD5 & options)) {
460 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
461 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
462 (TCPOLEN_COOKIE_BASE << 16) |
463 (TCPOPT_MD5SIG << 8) |
466 *ptr++ = htonl((TCPOPT_NOP << 24) |
468 (TCPOPT_MD5SIG << 8) |
471 options &= ~OPTION_COOKIE_EXTENSION;
472 /* overload cookie hash location */
473 opts->hash_location = (__u8 *)ptr;
477 if (unlikely(opts->mss)) {
478 *ptr++ = htonl((TCPOPT_MSS << 24) |
479 (TCPOLEN_MSS << 16) |
483 if (likely(OPTION_TS & options)) {
484 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
485 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
486 (TCPOLEN_SACK_PERM << 16) |
487 (TCPOPT_TIMESTAMP << 8) |
489 options &= ~OPTION_SACK_ADVERTISE;
491 *ptr++ = htonl((TCPOPT_NOP << 24) |
493 (TCPOPT_TIMESTAMP << 8) |
496 *ptr++ = htonl(opts->tsval);
497 *ptr++ = htonl(opts->tsecr);
500 /* Specification requires after timestamp, so do it now.
502 * Consider the pessimal case without authentication. The options
504 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
506 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
507 __u8 *cookie_copy = opts->hash_location;
508 u8 cookie_size = opts->hash_size;
510 /* 8-bit multiple handled in tcp_cookie_size_check() above,
513 if (0x2 & cookie_size) {
514 __u8 *p = (__u8 *)ptr;
516 /* 16-bit multiple */
517 *p++ = TCPOPT_COOKIE;
518 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
519 *p++ = *cookie_copy++;
520 *p++ = *cookie_copy++;
524 /* 32-bit multiple */
525 *ptr++ = htonl(((TCPOPT_NOP << 24) |
527 (TCPOPT_COOKIE << 8) |
528 TCPOLEN_COOKIE_BASE) +
532 if (cookie_size > 0) {
533 memcpy(ptr, cookie_copy, cookie_size);
534 ptr += (cookie_size / 4);
538 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
539 *ptr++ = htonl((TCPOPT_NOP << 24) |
541 (TCPOPT_SACK_PERM << 8) |
545 if (unlikely(OPTION_WSCALE & options)) {
546 *ptr++ = htonl((TCPOPT_NOP << 24) |
547 (TCPOPT_WINDOW << 16) |
548 (TCPOLEN_WINDOW << 8) |
552 if (unlikely(opts->num_sack_blocks)) {
553 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
554 tp->duplicate_sack : tp->selective_acks;
557 *ptr++ = htonl((TCPOPT_NOP << 24) |
560 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
561 TCPOLEN_SACK_PERBLOCK)));
563 for (this_sack = 0; this_sack < opts->num_sack_blocks;
565 *ptr++ = htonl(sp[this_sack].start_seq);
566 *ptr++ = htonl(sp[this_sack].end_seq);
569 tp->rx_opt.dsack = 0;
572 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
573 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
575 *ptr++ = htonl((TCPOPT_EXP << 24) |
576 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
577 TCPOPT_FASTOPEN_MAGIC);
579 memcpy(ptr, foc->val, foc->len);
580 if ((foc->len & 3) == 2) {
581 u8 *align = ((u8 *)ptr) + foc->len;
582 align[0] = align[1] = TCPOPT_NOP;
584 ptr += (foc->len + 3) >> 2;
588 /* Compute TCP options for SYN packets. This is not the final
589 * network wire format yet.
591 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
592 struct tcp_out_options *opts,
593 struct tcp_md5sig_key **md5)
595 struct tcp_sock *tp = tcp_sk(sk);
596 struct tcp_cookie_values *cvp = tp->cookie_values;
597 unsigned int remaining = MAX_TCP_OPTION_SPACE;
598 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
599 tcp_cookie_size_check(cvp->cookie_desired) :
601 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
603 #ifdef CONFIG_TCP_MD5SIG
604 *md5 = tp->af_specific->md5_lookup(sk, sk);
606 opts->options |= OPTION_MD5;
607 remaining -= TCPOLEN_MD5SIG_ALIGNED;
613 /* We always get an MSS option. The option bytes which will be seen in
614 * normal data packets should timestamps be used, must be in the MSS
615 * advertised. But we subtract them from tp->mss_cache so that
616 * calculations in tcp_sendmsg are simpler etc. So account for this
617 * fact here if necessary. If we don't do this correctly, as a
618 * receiver we won't recognize data packets as being full sized when we
619 * should, and thus we won't abide by the delayed ACK rules correctly.
620 * SACKs don't matter, we never delay an ACK when we have any of those
622 opts->mss = tcp_advertise_mss(sk);
623 remaining -= TCPOLEN_MSS_ALIGNED;
625 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
626 opts->options |= OPTION_TS;
627 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
628 opts->tsecr = tp->rx_opt.ts_recent;
629 remaining -= TCPOLEN_TSTAMP_ALIGNED;
631 if (likely(sysctl_tcp_window_scaling)) {
632 opts->ws = tp->rx_opt.rcv_wscale;
633 opts->options |= OPTION_WSCALE;
634 remaining -= TCPOLEN_WSCALE_ALIGNED;
636 if (likely(sysctl_tcp_sack)) {
637 opts->options |= OPTION_SACK_ADVERTISE;
638 if (unlikely(!(OPTION_TS & opts->options)))
639 remaining -= TCPOLEN_SACKPERM_ALIGNED;
642 if (fastopen && fastopen->cookie.len >= 0) {
643 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
644 need = (need + 3) & ~3U; /* Align to 32 bits */
645 if (remaining >= need) {
646 opts->options |= OPTION_FAST_OPEN_COOKIE;
647 opts->fastopen_cookie = &fastopen->cookie;
649 tp->syn_fastopen = 1;
652 /* Note that timestamps are required by the specification.
654 * Odd numbers of bytes are prohibited by the specification, ensuring
655 * that the cookie is 16-bit aligned, and the resulting cookie pair is
659 (OPTION_TS & opts->options) &&
661 int need = TCPOLEN_COOKIE_BASE + cookie_size;
664 /* 32-bit multiple */
665 need += 2; /* NOPs */
667 if (need > remaining) {
668 /* try shrinking cookie to fit */
673 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
677 if (TCP_COOKIE_MIN <= cookie_size) {
678 opts->options |= OPTION_COOKIE_EXTENSION;
679 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
680 opts->hash_size = cookie_size;
682 /* Remember for future incarnations. */
683 cvp->cookie_desired = cookie_size;
685 if (cvp->cookie_desired != cvp->cookie_pair_size) {
686 /* Currently use random bytes as a nonce,
687 * assuming these are completely unpredictable
688 * by hostile users of the same system.
690 get_random_bytes(&cvp->cookie_pair[0],
692 cvp->cookie_pair_size = cookie_size;
698 return MAX_TCP_OPTION_SPACE - remaining;
701 /* Set up TCP options for SYN-ACKs. */
702 static unsigned int tcp_synack_options(struct sock *sk,
703 struct request_sock *req,
704 unsigned int mss, struct sk_buff *skb,
705 struct tcp_out_options *opts,
706 struct tcp_md5sig_key **md5,
707 struct tcp_extend_values *xvp,
708 struct tcp_fastopen_cookie *foc)
710 struct inet_request_sock *ireq = inet_rsk(req);
711 unsigned int remaining = MAX_TCP_OPTION_SPACE;
712 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
716 #ifdef CONFIG_TCP_MD5SIG
717 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
719 opts->options |= OPTION_MD5;
720 remaining -= TCPOLEN_MD5SIG_ALIGNED;
722 /* We can't fit any SACK blocks in a packet with MD5 + TS
723 * options. There was discussion about disabling SACK
724 * rather than TS in order to fit in better with old,
725 * buggy kernels, but that was deemed to be unnecessary.
727 ireq->tstamp_ok &= !ireq->sack_ok;
733 /* We always send an MSS option. */
735 remaining -= TCPOLEN_MSS_ALIGNED;
737 if (likely(ireq->wscale_ok)) {
738 opts->ws = ireq->rcv_wscale;
739 opts->options |= OPTION_WSCALE;
740 remaining -= TCPOLEN_WSCALE_ALIGNED;
742 if (likely(ireq->tstamp_ok)) {
743 opts->options |= OPTION_TS;
744 opts->tsval = TCP_SKB_CB(skb)->when;
745 opts->tsecr = req->ts_recent;
746 remaining -= TCPOLEN_TSTAMP_ALIGNED;
748 if (likely(ireq->sack_ok)) {
749 opts->options |= OPTION_SACK_ADVERTISE;
750 if (unlikely(!ireq->tstamp_ok))
751 remaining -= TCPOLEN_SACKPERM_ALIGNED;
754 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
755 need = (need + 3) & ~3U; /* Align to 32 bits */
756 if (remaining >= need) {
757 opts->options |= OPTION_FAST_OPEN_COOKIE;
758 opts->fastopen_cookie = foc;
762 /* Similar rationale to tcp_syn_options() applies here, too.
763 * If the <SYN> options fit, the same options should fit now!
767 cookie_plus > TCPOLEN_COOKIE_BASE) {
768 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
771 /* 32-bit multiple */
772 need += 2; /* NOPs */
774 if (need <= remaining) {
775 opts->options |= OPTION_COOKIE_EXTENSION;
776 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
779 /* There's no error return, so flag it. */
780 xvp->cookie_out_never = 1; /* true */
784 return MAX_TCP_OPTION_SPACE - remaining;
787 /* Compute TCP options for ESTABLISHED sockets. This is not the
788 * final wire format yet.
790 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
791 struct tcp_out_options *opts,
792 struct tcp_md5sig_key **md5)
794 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
795 struct tcp_sock *tp = tcp_sk(sk);
796 unsigned int size = 0;
797 unsigned int eff_sacks;
799 #ifdef CONFIG_TCP_MD5SIG
800 *md5 = tp->af_specific->md5_lookup(sk, sk);
801 if (unlikely(*md5)) {
802 opts->options |= OPTION_MD5;
803 size += TCPOLEN_MD5SIG_ALIGNED;
809 if (likely(tp->rx_opt.tstamp_ok)) {
810 opts->options |= OPTION_TS;
811 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
812 opts->tsecr = tp->rx_opt.ts_recent;
813 size += TCPOLEN_TSTAMP_ALIGNED;
816 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
817 if (unlikely(eff_sacks)) {
818 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
819 opts->num_sack_blocks =
820 min_t(unsigned int, eff_sacks,
821 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
822 TCPOLEN_SACK_PERBLOCK);
823 size += TCPOLEN_SACK_BASE_ALIGNED +
824 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
831 /* TCP SMALL QUEUES (TSQ)
833 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
834 * to reduce RTT and bufferbloat.
835 * We do this using a special skb destructor (tcp_wfree).
837 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
838 * needs to be reallocated in a driver.
839 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
841 * Since transmit from skb destructor is forbidden, we use a tasklet
842 * to process all sockets that eventually need to send more skbs.
843 * We use one tasklet per cpu, with its own queue of sockets.
846 struct tasklet_struct tasklet;
847 struct list_head head; /* queue of tcp sockets */
849 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
851 static void tcp_tsq_handler(struct sock *sk)
853 if ((1 << sk->sk_state) &
854 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
855 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
856 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
859 * One tasklest per cpu tries to send more skbs.
860 * We run in tasklet context but need to disable irqs when
861 * transfering tsq->head because tcp_wfree() might
862 * interrupt us (non NAPI drivers)
864 static void tcp_tasklet_func(unsigned long data)
866 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
869 struct list_head *q, *n;
873 local_irq_save(flags);
874 list_splice_init(&tsq->head, &list);
875 local_irq_restore(flags);
877 list_for_each_safe(q, n, &list) {
878 tp = list_entry(q, struct tcp_sock, tsq_node);
879 list_del(&tp->tsq_node);
881 sk = (struct sock *)tp;
884 if (!sock_owned_by_user(sk)) {
887 /* defer the work to tcp_release_cb() */
888 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
892 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
897 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
898 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
899 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
900 (1UL << TCP_MTU_REDUCED_DEFERRED))
902 * tcp_release_cb - tcp release_sock() callback
905 * called from release_sock() to perform protocol dependent
906 * actions before socket release.
908 void tcp_release_cb(struct sock *sk)
910 struct tcp_sock *tp = tcp_sk(sk);
911 unsigned long flags, nflags;
913 /* perform an atomic operation only if at least one flag is set */
915 flags = tp->tsq_flags;
916 if (!(flags & TCP_DEFERRED_ALL))
918 nflags = flags & ~TCP_DEFERRED_ALL;
919 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
921 if (flags & (1UL << TCP_TSQ_DEFERRED))
924 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
925 tcp_write_timer_handler(sk);
928 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
929 tcp_delack_timer_handler(sk);
932 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
933 sk->sk_prot->mtu_reduced(sk);
937 EXPORT_SYMBOL(tcp_release_cb);
939 void __init tcp_tasklet_init(void)
943 for_each_possible_cpu(i) {
944 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
946 INIT_LIST_HEAD(&tsq->head);
947 tasklet_init(&tsq->tasklet,
954 * Write buffer destructor automatically called from kfree_skb.
955 * We cant xmit new skbs from this context, as we might already
958 static void tcp_wfree(struct sk_buff *skb)
960 struct sock *sk = skb->sk;
961 struct tcp_sock *tp = tcp_sk(sk);
963 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
964 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
966 struct tsq_tasklet *tsq;
968 /* Keep a ref on socket.
969 * This last ref will be released in tcp_tasklet_func()
971 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
973 /* queue this socket to tasklet queue */
974 local_irq_save(flags);
975 tsq = &__get_cpu_var(tsq_tasklet);
976 list_add(&tp->tsq_node, &tsq->head);
977 tasklet_schedule(&tsq->tasklet);
978 local_irq_restore(flags);
984 /* This routine actually transmits TCP packets queued in by
985 * tcp_do_sendmsg(). This is used by both the initial
986 * transmission and possible later retransmissions.
987 * All SKB's seen here are completely headerless. It is our
988 * job to build the TCP header, and pass the packet down to
989 * IP so it can do the same plus pass the packet off to the
992 * We are working here with either a clone of the original
993 * SKB, or a fresh unique copy made by the retransmit engine.
995 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
998 const struct inet_connection_sock *icsk = inet_csk(sk);
999 struct inet_sock *inet;
1000 struct tcp_sock *tp;
1001 struct tcp_skb_cb *tcb;
1002 struct tcp_out_options opts;
1003 unsigned int tcp_options_size, tcp_header_size;
1004 struct tcp_md5sig_key *md5;
1008 BUG_ON(!skb || !tcp_skb_pcount(skb));
1010 /* If congestion control is doing timestamping, we must
1011 * take such a timestamp before we potentially clone/copy.
1013 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
1014 __net_timestamp(skb);
1016 if (likely(clone_it)) {
1017 if (unlikely(skb_cloned(skb)))
1018 skb = pskb_copy(skb, gfp_mask);
1020 skb = skb_clone(skb, gfp_mask);
1027 tcb = TCP_SKB_CB(skb);
1028 memset(&opts, 0, sizeof(opts));
1030 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1031 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1033 tcp_options_size = tcp_established_options(sk, skb, &opts,
1035 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1037 if (tcp_packets_in_flight(tp) == 0) {
1038 tcp_ca_event(sk, CA_EVENT_TX_START);
1043 skb_push(skb, tcp_header_size);
1044 skb_reset_transport_header(skb);
1048 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
1049 tcp_wfree : sock_wfree;
1050 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
1052 /* Build TCP header and checksum it. */
1054 th->source = inet->inet_sport;
1055 th->dest = inet->inet_dport;
1056 th->seq = htonl(tcb->seq);
1057 th->ack_seq = htonl(tp->rcv_nxt);
1058 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1061 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1062 /* RFC1323: The window in SYN & SYN/ACK segments
1065 th->window = htons(min(tp->rcv_wnd, 65535U));
1067 th->window = htons(tcp_select_window(sk));
1072 /* The urg_mode check is necessary during a below snd_una win probe */
1073 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1074 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1075 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1077 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1078 th->urg_ptr = htons(0xFFFF);
1083 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1084 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1085 TCP_ECN_send(sk, skb, tcp_header_size);
1087 #ifdef CONFIG_TCP_MD5SIG
1088 /* Calculate the MD5 hash, as we have all we need now */
1090 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1091 tp->af_specific->calc_md5_hash(opts.hash_location,
1092 md5, sk, NULL, skb);
1096 icsk->icsk_af_ops->send_check(sk, skb);
1098 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1099 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1101 if (skb->len != tcp_header_size)
1102 tcp_event_data_sent(tp, sk);
1104 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1105 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1106 tcp_skb_pcount(skb));
1108 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1109 if (likely(err <= 0))
1112 tcp_enter_cwr(sk, 1);
1114 return net_xmit_eval(err);
1117 /* This routine just queues the buffer for sending.
1119 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1120 * otherwise socket can stall.
1122 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1124 struct tcp_sock *tp = tcp_sk(sk);
1126 /* Advance write_seq and place onto the write_queue. */
1127 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1128 skb_header_release(skb);
1129 tcp_add_write_queue_tail(sk, skb);
1130 sk->sk_wmem_queued += skb->truesize;
1131 sk_mem_charge(sk, skb->truesize);
1134 /* Initialize TSO segments for a packet. */
1135 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1136 unsigned int mss_now)
1138 if (skb->len <= mss_now || !sk_can_gso(sk) ||
1139 skb->ip_summed == CHECKSUM_NONE) {
1140 /* Avoid the costly divide in the normal
1143 skb_shinfo(skb)->gso_segs = 1;
1144 skb_shinfo(skb)->gso_size = 0;
1145 skb_shinfo(skb)->gso_type = 0;
1147 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1148 skb_shinfo(skb)->gso_size = mss_now;
1149 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1153 /* When a modification to fackets out becomes necessary, we need to check
1154 * skb is counted to fackets_out or not.
1156 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1159 struct tcp_sock *tp = tcp_sk(sk);
1161 if (!tp->sacked_out || tcp_is_reno(tp))
1164 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1165 tp->fackets_out -= decr;
1168 /* Pcount in the middle of the write queue got changed, we need to do various
1169 * tweaks to fix counters
1171 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1173 struct tcp_sock *tp = tcp_sk(sk);
1175 tp->packets_out -= decr;
1177 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1178 tp->sacked_out -= decr;
1179 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1180 tp->retrans_out -= decr;
1181 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1182 tp->lost_out -= decr;
1184 /* Reno case is special. Sigh... */
1185 if (tcp_is_reno(tp) && decr > 0)
1186 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1188 tcp_adjust_fackets_out(sk, skb, decr);
1190 if (tp->lost_skb_hint &&
1191 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1192 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1193 tp->lost_cnt_hint -= decr;
1195 tcp_verify_left_out(tp);
1198 /* Function to create two new TCP segments. Shrinks the given segment
1199 * to the specified size and appends a new segment with the rest of the
1200 * packet to the list. This won't be called frequently, I hope.
1201 * Remember, these are still headerless SKBs at this point.
1203 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1204 unsigned int mss_now)
1206 struct tcp_sock *tp = tcp_sk(sk);
1207 struct sk_buff *buff;
1208 int nsize, old_factor;
1212 if (WARN_ON(len > skb->len))
1215 nsize = skb_headlen(skb) - len;
1219 if (skb_cloned(skb) &&
1220 skb_is_nonlinear(skb) &&
1221 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1224 /* Get a new skb... force flag on. */
1225 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1227 return -ENOMEM; /* We'll just try again later. */
1229 sk->sk_wmem_queued += buff->truesize;
1230 sk_mem_charge(sk, buff->truesize);
1231 nlen = skb->len - len - nsize;
1232 buff->truesize += nlen;
1233 skb->truesize -= nlen;
1235 /* Correct the sequence numbers. */
1236 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1237 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1238 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1240 /* PSH and FIN should only be set in the second packet. */
1241 flags = TCP_SKB_CB(skb)->tcp_flags;
1242 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1243 TCP_SKB_CB(buff)->tcp_flags = flags;
1244 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1246 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1247 /* Copy and checksum data tail into the new buffer. */
1248 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1249 skb_put(buff, nsize),
1254 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1256 skb->ip_summed = CHECKSUM_PARTIAL;
1257 skb_split(skb, buff, len);
1260 buff->ip_summed = skb->ip_summed;
1262 /* Looks stupid, but our code really uses when of
1263 * skbs, which it never sent before. --ANK
1265 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1266 buff->tstamp = skb->tstamp;
1268 old_factor = tcp_skb_pcount(skb);
1270 /* Fix up tso_factor for both original and new SKB. */
1271 tcp_set_skb_tso_segs(sk, skb, mss_now);
1272 tcp_set_skb_tso_segs(sk, buff, mss_now);
1274 /* If this packet has been sent out already, we must
1275 * adjust the various packet counters.
1277 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1278 int diff = old_factor - tcp_skb_pcount(skb) -
1279 tcp_skb_pcount(buff);
1282 tcp_adjust_pcount(sk, skb, diff);
1285 /* Link BUFF into the send queue. */
1286 skb_header_release(buff);
1287 tcp_insert_write_queue_after(skb, buff, sk);
1292 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1293 * eventually). The difference is that pulled data not copied, but
1294 * immediately discarded.
1296 static void __pskb_trim_head(struct sk_buff *skb, int len)
1300 eat = min_t(int, len, skb_headlen(skb));
1302 __skb_pull(skb, eat);
1303 skb->avail_size -= eat;
1310 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1311 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1314 skb_frag_unref(skb, i);
1317 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1319 skb_shinfo(skb)->frags[k].page_offset += eat;
1320 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1326 skb_shinfo(skb)->nr_frags = k;
1328 skb_reset_tail_pointer(skb);
1329 skb->data_len -= len;
1330 skb->len = skb->data_len;
1333 /* Remove acked data from a packet in the transmit queue. */
1334 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1336 if (skb_unclone(skb, GFP_ATOMIC))
1339 __pskb_trim_head(skb, len);
1341 TCP_SKB_CB(skb)->seq += len;
1342 skb->ip_summed = CHECKSUM_PARTIAL;
1344 skb->truesize -= len;
1345 sk->sk_wmem_queued -= len;
1346 sk_mem_uncharge(sk, len);
1347 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1349 /* Any change of skb->len requires recalculation of tso factor. */
1350 if (tcp_skb_pcount(skb) > 1)
1351 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1356 /* Calculate MSS not accounting any TCP options. */
1357 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1359 const struct tcp_sock *tp = tcp_sk(sk);
1360 const struct inet_connection_sock *icsk = inet_csk(sk);
1363 /* Calculate base mss without TCP options:
1364 It is MMS_S - sizeof(tcphdr) of rfc1122
1366 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1368 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1369 if (icsk->icsk_af_ops->net_frag_header_len) {
1370 const struct dst_entry *dst = __sk_dst_get(sk);
1372 if (dst && dst_allfrag(dst))
1373 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1376 /* Clamp it (mss_clamp does not include tcp options) */
1377 if (mss_now > tp->rx_opt.mss_clamp)
1378 mss_now = tp->rx_opt.mss_clamp;
1380 /* Now subtract optional transport overhead */
1381 mss_now -= icsk->icsk_ext_hdr_len;
1383 /* Then reserve room for full set of TCP options and 8 bytes of data */
1389 /* Calculate MSS. Not accounting for SACKs here. */
1390 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1392 /* Subtract TCP options size, not including SACKs */
1393 return __tcp_mtu_to_mss(sk, pmtu) -
1394 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1397 /* Inverse of above */
1398 int tcp_mss_to_mtu(struct sock *sk, int mss)
1400 const struct tcp_sock *tp = tcp_sk(sk);
1401 const struct inet_connection_sock *icsk = inet_csk(sk);
1405 tp->tcp_header_len +
1406 icsk->icsk_ext_hdr_len +
1407 icsk->icsk_af_ops->net_header_len;
1409 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1410 if (icsk->icsk_af_ops->net_frag_header_len) {
1411 const struct dst_entry *dst = __sk_dst_get(sk);
1413 if (dst && dst_allfrag(dst))
1414 mtu += icsk->icsk_af_ops->net_frag_header_len;
1419 /* MTU probing init per socket */
1420 void tcp_mtup_init(struct sock *sk)
1422 struct tcp_sock *tp = tcp_sk(sk);
1423 struct inet_connection_sock *icsk = inet_csk(sk);
1425 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1426 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1427 icsk->icsk_af_ops->net_header_len;
1428 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1429 icsk->icsk_mtup.probe_size = 0;
1431 EXPORT_SYMBOL(tcp_mtup_init);
1433 /* This function synchronize snd mss to current pmtu/exthdr set.
1435 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1436 for TCP options, but includes only bare TCP header.
1438 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1439 It is minimum of user_mss and mss received with SYN.
1440 It also does not include TCP options.
1442 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1444 tp->mss_cache is current effective sending mss, including
1445 all tcp options except for SACKs. It is evaluated,
1446 taking into account current pmtu, but never exceeds
1447 tp->rx_opt.mss_clamp.
1449 NOTE1. rfc1122 clearly states that advertised MSS
1450 DOES NOT include either tcp or ip options.
1452 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1453 are READ ONLY outside this function. --ANK (980731)
1455 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1457 struct tcp_sock *tp = tcp_sk(sk);
1458 struct inet_connection_sock *icsk = inet_csk(sk);
1461 if (icsk->icsk_mtup.search_high > pmtu)
1462 icsk->icsk_mtup.search_high = pmtu;
1464 mss_now = tcp_mtu_to_mss(sk, pmtu);
1465 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1467 /* And store cached results */
1468 icsk->icsk_pmtu_cookie = pmtu;
1469 if (icsk->icsk_mtup.enabled)
1470 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1471 tp->mss_cache = mss_now;
1475 EXPORT_SYMBOL(tcp_sync_mss);
1477 /* Compute the current effective MSS, taking SACKs and IP options,
1478 * and even PMTU discovery events into account.
1480 unsigned int tcp_current_mss(struct sock *sk)
1482 const struct tcp_sock *tp = tcp_sk(sk);
1483 const struct dst_entry *dst = __sk_dst_get(sk);
1485 unsigned int header_len;
1486 struct tcp_out_options opts;
1487 struct tcp_md5sig_key *md5;
1489 mss_now = tp->mss_cache;
1492 u32 mtu = dst_mtu(dst);
1493 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1494 mss_now = tcp_sync_mss(sk, mtu);
1497 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1498 sizeof(struct tcphdr);
1499 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1500 * some common options. If this is an odd packet (because we have SACK
1501 * blocks etc) then our calculated header_len will be different, and
1502 * we have to adjust mss_now correspondingly */
1503 if (header_len != tp->tcp_header_len) {
1504 int delta = (int) header_len - tp->tcp_header_len;
1511 /* Congestion window validation. (RFC2861) */
1512 static void tcp_cwnd_validate(struct sock *sk)
1514 struct tcp_sock *tp = tcp_sk(sk);
1516 if (tp->packets_out >= tp->snd_cwnd) {
1517 /* Network is feed fully. */
1518 tp->snd_cwnd_used = 0;
1519 tp->snd_cwnd_stamp = tcp_time_stamp;
1521 /* Network starves. */
1522 if (tp->packets_out > tp->snd_cwnd_used)
1523 tp->snd_cwnd_used = tp->packets_out;
1525 if (sysctl_tcp_slow_start_after_idle &&
1526 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1527 tcp_cwnd_application_limited(sk);
1531 /* Returns the portion of skb which can be sent right away without
1532 * introducing MSS oddities to segment boundaries. In rare cases where
1533 * mss_now != mss_cache, we will request caller to create a small skb
1534 * per input skb which could be mostly avoided here (if desired).
1536 * We explicitly want to create a request for splitting write queue tail
1537 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1538 * thus all the complexity (cwnd_len is always MSS multiple which we
1539 * return whenever allowed by the other factors). Basically we need the
1540 * modulo only when the receiver window alone is the limiting factor or
1541 * when we would be allowed to send the split-due-to-Nagle skb fully.
1543 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1544 unsigned int mss_now, unsigned int max_segs)
1546 const struct tcp_sock *tp = tcp_sk(sk);
1547 u32 needed, window, max_len;
1549 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1550 max_len = mss_now * max_segs;
1552 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1555 needed = min(skb->len, window);
1557 if (max_len <= needed)
1560 return needed - needed % mss_now;
1563 /* Can at least one segment of SKB be sent right now, according to the
1564 * congestion window rules? If so, return how many segments are allowed.
1566 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1567 const struct sk_buff *skb)
1569 u32 in_flight, cwnd;
1571 /* Don't be strict about the congestion window for the final FIN. */
1572 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1573 tcp_skb_pcount(skb) == 1)
1576 in_flight = tcp_packets_in_flight(tp);
1577 cwnd = tp->snd_cwnd;
1578 if (in_flight < cwnd)
1579 return (cwnd - in_flight);
1584 /* Initialize TSO state of a skb.
1585 * This must be invoked the first time we consider transmitting
1586 * SKB onto the wire.
1588 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1589 unsigned int mss_now)
1591 int tso_segs = tcp_skb_pcount(skb);
1593 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1594 tcp_set_skb_tso_segs(sk, skb, mss_now);
1595 tso_segs = tcp_skb_pcount(skb);
1600 /* Minshall's variant of the Nagle send check. */
1601 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1603 return after(tp->snd_sml, tp->snd_una) &&
1604 !after(tp->snd_sml, tp->snd_nxt);
1607 /* Return false, if packet can be sent now without violation Nagle's rules:
1608 * 1. It is full sized.
1609 * 2. Or it contains FIN. (already checked by caller)
1610 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1611 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1612 * With Minshall's modification: all sent small packets are ACKed.
1614 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1615 const struct sk_buff *skb,
1616 unsigned int mss_now, int nonagle)
1618 return skb->len < mss_now &&
1619 ((nonagle & TCP_NAGLE_CORK) ||
1620 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1623 /* Return true if the Nagle test allows this packet to be
1626 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1627 unsigned int cur_mss, int nonagle)
1629 /* Nagle rule does not apply to frames, which sit in the middle of the
1630 * write_queue (they have no chances to get new data).
1632 * This is implemented in the callers, where they modify the 'nonagle'
1633 * argument based upon the location of SKB in the send queue.
1635 if (nonagle & TCP_NAGLE_PUSH)
1638 /* Don't use the nagle rule for urgent data (or for the final FIN).
1639 * Nagle can be ignored during F-RTO too (see RFC4138).
1641 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1642 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1645 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1651 /* Does at least the first segment of SKB fit into the send window? */
1652 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1653 const struct sk_buff *skb,
1654 unsigned int cur_mss)
1656 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1658 if (skb->len > cur_mss)
1659 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1661 return !after(end_seq, tcp_wnd_end(tp));
1664 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1665 * should be put on the wire right now. If so, it returns the number of
1666 * packets allowed by the congestion window.
1668 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1669 unsigned int cur_mss, int nonagle)
1671 const struct tcp_sock *tp = tcp_sk(sk);
1672 unsigned int cwnd_quota;
1674 tcp_init_tso_segs(sk, skb, cur_mss);
1676 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1679 cwnd_quota = tcp_cwnd_test(tp, skb);
1680 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1686 /* Test if sending is allowed right now. */
1687 bool tcp_may_send_now(struct sock *sk)
1689 const struct tcp_sock *tp = tcp_sk(sk);
1690 struct sk_buff *skb = tcp_send_head(sk);
1693 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1694 (tcp_skb_is_last(sk, skb) ?
1695 tp->nonagle : TCP_NAGLE_PUSH));
1698 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1699 * which is put after SKB on the list. It is very much like
1700 * tcp_fragment() except that it may make several kinds of assumptions
1701 * in order to speed up the splitting operation. In particular, we
1702 * know that all the data is in scatter-gather pages, and that the
1703 * packet has never been sent out before (and thus is not cloned).
1705 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1706 unsigned int mss_now, gfp_t gfp)
1708 struct sk_buff *buff;
1709 int nlen = skb->len - len;
1712 /* All of a TSO frame must be composed of paged data. */
1713 if (skb->len != skb->data_len)
1714 return tcp_fragment(sk, skb, len, mss_now);
1716 buff = sk_stream_alloc_skb(sk, 0, gfp);
1717 if (unlikely(buff == NULL))
1720 sk->sk_wmem_queued += buff->truesize;
1721 sk_mem_charge(sk, buff->truesize);
1722 buff->truesize += nlen;
1723 skb->truesize -= nlen;
1725 /* Correct the sequence numbers. */
1726 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1727 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1728 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1730 /* PSH and FIN should only be set in the second packet. */
1731 flags = TCP_SKB_CB(skb)->tcp_flags;
1732 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1733 TCP_SKB_CB(buff)->tcp_flags = flags;
1735 /* This packet was never sent out yet, so no SACK bits. */
1736 TCP_SKB_CB(buff)->sacked = 0;
1738 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1739 skb_split(skb, buff, len);
1741 /* Fix up tso_factor for both original and new SKB. */
1742 tcp_set_skb_tso_segs(sk, skb, mss_now);
1743 tcp_set_skb_tso_segs(sk, buff, mss_now);
1745 /* Link BUFF into the send queue. */
1746 skb_header_release(buff);
1747 tcp_insert_write_queue_after(skb, buff, sk);
1752 /* Try to defer sending, if possible, in order to minimize the amount
1753 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1755 * This algorithm is from John Heffner.
1757 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 const struct inet_connection_sock *icsk = inet_csk(sk);
1761 u32 send_win, cong_win, limit, in_flight;
1764 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1767 if (icsk->icsk_ca_state != TCP_CA_Open)
1770 /* Defer for less than two clock ticks. */
1771 if (tp->tso_deferred &&
1772 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1775 in_flight = tcp_packets_in_flight(tp);
1777 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1779 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1781 /* From in_flight test above, we know that cwnd > in_flight. */
1782 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1784 limit = min(send_win, cong_win);
1786 /* If a full-sized TSO skb can be sent, do it. */
1787 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1788 sk->sk_gso_max_segs * tp->mss_cache))
1791 /* Middle in queue won't get any more data, full sendable already? */
1792 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1795 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1797 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1799 /* If at least some fraction of a window is available,
1802 chunk /= win_divisor;
1806 /* Different approach, try not to defer past a single
1807 * ACK. Receiver should ACK every other full sized
1808 * frame, so if we have space for more than 3 frames
1811 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1815 /* Ok, it looks like it is advisable to defer. */
1816 tp->tso_deferred = 1 | (jiffies << 1);
1821 tp->tso_deferred = 0;
1825 /* Create a new MTU probe if we are ready.
1826 * MTU probe is regularly attempting to increase the path MTU by
1827 * deliberately sending larger packets. This discovers routing
1828 * changes resulting in larger path MTUs.
1830 * Returns 0 if we should wait to probe (no cwnd available),
1831 * 1 if a probe was sent,
1834 static int tcp_mtu_probe(struct sock *sk)
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct inet_connection_sock *icsk = inet_csk(sk);
1838 struct sk_buff *skb, *nskb, *next;
1845 /* Not currently probing/verifying,
1847 * have enough cwnd, and
1848 * not SACKing (the variable headers throw things off) */
1849 if (!icsk->icsk_mtup.enabled ||
1850 icsk->icsk_mtup.probe_size ||
1851 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1852 tp->snd_cwnd < 11 ||
1853 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1856 /* Very simple search strategy: just double the MSS. */
1857 mss_now = tcp_current_mss(sk);
1858 probe_size = 2 * tp->mss_cache;
1859 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1860 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1861 /* TODO: set timer for probe_converge_event */
1865 /* Have enough data in the send queue to probe? */
1866 if (tp->write_seq - tp->snd_nxt < size_needed)
1869 if (tp->snd_wnd < size_needed)
1871 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1874 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1875 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1876 if (!tcp_packets_in_flight(tp))
1882 /* We're allowed to probe. Build it now. */
1883 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1885 sk->sk_wmem_queued += nskb->truesize;
1886 sk_mem_charge(sk, nskb->truesize);
1888 skb = tcp_send_head(sk);
1890 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1891 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1892 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1893 TCP_SKB_CB(nskb)->sacked = 0;
1895 nskb->ip_summed = skb->ip_summed;
1897 tcp_insert_write_queue_before(nskb, skb, sk);
1900 tcp_for_write_queue_from_safe(skb, next, sk) {
1901 copy = min_t(int, skb->len, probe_size - len);
1902 if (nskb->ip_summed)
1903 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1905 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1906 skb_put(nskb, copy),
1909 if (skb->len <= copy) {
1910 /* We've eaten all the data from this skb.
1912 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1913 tcp_unlink_write_queue(skb, sk);
1914 sk_wmem_free_skb(sk, skb);
1916 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1917 ~(TCPHDR_FIN|TCPHDR_PSH);
1918 if (!skb_shinfo(skb)->nr_frags) {
1919 skb_pull(skb, copy);
1920 if (skb->ip_summed != CHECKSUM_PARTIAL)
1921 skb->csum = csum_partial(skb->data,
1924 __pskb_trim_head(skb, copy);
1925 tcp_set_skb_tso_segs(sk, skb, mss_now);
1927 TCP_SKB_CB(skb)->seq += copy;
1932 if (len >= probe_size)
1935 tcp_init_tso_segs(sk, nskb, nskb->len);
1937 /* We're ready to send. If this fails, the probe will
1938 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1939 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1940 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1941 /* Decrement cwnd here because we are sending
1942 * effectively two packets. */
1944 tcp_event_new_data_sent(sk, nskb);
1946 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1947 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1948 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1956 /* This routine writes packets to the network. It advances the
1957 * send_head. This happens as incoming acks open up the remote
1960 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1961 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1962 * account rare use of URG, this is not a big flaw.
1964 * Send at most one packet when push_one > 0. Temporarily ignore
1965 * cwnd limit to force at most one packet out when push_one == 2.
1967 * Returns true, if no segments are in flight and we have queued segments,
1968 * but cannot send anything now because of SWS or another problem.
1970 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1971 int push_one, gfp_t gfp)
1973 struct tcp_sock *tp = tcp_sk(sk);
1974 struct sk_buff *skb;
1975 unsigned int tso_segs, sent_pkts;
1982 /* Do MTU probing. */
1983 result = tcp_mtu_probe(sk);
1986 } else if (result > 0) {
1991 while ((skb = tcp_send_head(sk))) {
1995 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1998 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1999 goto repair; /* Skip network transmission */
2001 cwnd_quota = tcp_cwnd_test(tp, skb);
2004 /* Force out a loss probe pkt. */
2010 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2013 if (tso_segs == 1) {
2014 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2015 (tcp_skb_is_last(sk, skb) ?
2016 nonagle : TCP_NAGLE_PUSH))))
2019 if (!push_one && tcp_tso_should_defer(sk, skb))
2023 /* TSQ : sk_wmem_alloc accounts skb truesize,
2024 * including skb overhead. But thats OK.
2026 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
2027 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2031 if (tso_segs > 1 && !tcp_urg_mode(tp))
2032 limit = tcp_mss_split_point(sk, skb, mss_now,
2035 sk->sk_gso_max_segs));
2037 if (skb->len > limit &&
2038 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2041 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2043 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2047 /* Advance the send_head. This one is sent out.
2048 * This call will increment packets_out.
2050 tcp_event_new_data_sent(sk, skb);
2052 tcp_minshall_update(tp, mss_now, skb);
2053 sent_pkts += tcp_skb_pcount(skb);
2059 if (likely(sent_pkts)) {
2060 if (tcp_in_cwnd_reduction(sk))
2061 tp->prr_out += sent_pkts;
2063 /* Send one loss probe per tail loss episode. */
2065 tcp_schedule_loss_probe(sk);
2066 tcp_cwnd_validate(sk);
2069 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2072 bool tcp_schedule_loss_probe(struct sock *sk)
2074 struct inet_connection_sock *icsk = inet_csk(sk);
2075 struct tcp_sock *tp = tcp_sk(sk);
2076 u32 timeout, tlp_time_stamp, rto_time_stamp;
2077 u32 rtt = tp->srtt >> 3;
2079 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2081 /* No consecutive loss probes. */
2082 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2086 /* Don't do any loss probe on a Fast Open connection before 3WHS
2089 if (sk->sk_state == TCP_SYN_RECV)
2092 /* TLP is only scheduled when next timer event is RTO. */
2093 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2096 /* Schedule a loss probe in 2*RTT for SACK capable connections
2097 * in Open state, that are either limited by cwnd or application.
2099 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
2100 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2103 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2107 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2108 * for delayed ack when there's one outstanding packet.
2111 if (tp->packets_out == 1)
2112 timeout = max_t(u32, timeout,
2113 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2114 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2116 /* If RTO is shorter, just schedule TLP in its place. */
2117 tlp_time_stamp = tcp_time_stamp + timeout;
2118 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2119 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2120 s32 delta = rto_time_stamp - tcp_time_stamp;
2125 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2130 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2131 * retransmit the last segment.
2133 void tcp_send_loss_probe(struct sock *sk)
2135 struct tcp_sock *tp = tcp_sk(sk);
2136 struct sk_buff *skb;
2138 int mss = tcp_current_mss(sk);
2141 if (tcp_send_head(sk) != NULL) {
2142 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2146 /* At most one outstanding TLP retransmission. */
2147 if (tp->tlp_high_seq)
2150 /* Retransmit last segment. */
2151 skb = tcp_write_queue_tail(sk);
2155 pcount = tcp_skb_pcount(skb);
2156 if (WARN_ON(!pcount))
2159 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2160 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2162 skb = tcp_write_queue_tail(sk);
2165 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2168 /* Probe with zero data doesn't trigger fast recovery. */
2170 err = __tcp_retransmit_skb(sk, skb);
2172 /* Record snd_nxt for loss detection. */
2174 tp->tlp_high_seq = tp->snd_nxt;
2177 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2178 inet_csk(sk)->icsk_rto,
2182 NET_INC_STATS_BH(sock_net(sk),
2183 LINUX_MIB_TCPLOSSPROBES);
2187 /* Push out any pending frames which were held back due to
2188 * TCP_CORK or attempt at coalescing tiny packets.
2189 * The socket must be locked by the caller.
2191 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2194 /* If we are closed, the bytes will have to remain here.
2195 * In time closedown will finish, we empty the write queue and
2196 * all will be happy.
2198 if (unlikely(sk->sk_state == TCP_CLOSE))
2201 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2202 sk_gfp_atomic(sk, GFP_ATOMIC)))
2203 tcp_check_probe_timer(sk);
2206 /* Send _single_ skb sitting at the send head. This function requires
2207 * true push pending frames to setup probe timer etc.
2209 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2211 struct sk_buff *skb = tcp_send_head(sk);
2213 BUG_ON(!skb || skb->len < mss_now);
2215 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2218 /* This function returns the amount that we can raise the
2219 * usable window based on the following constraints
2221 * 1. The window can never be shrunk once it is offered (RFC 793)
2222 * 2. We limit memory per socket
2225 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2226 * RECV.NEXT + RCV.WIN fixed until:
2227 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2229 * i.e. don't raise the right edge of the window until you can raise
2230 * it at least MSS bytes.
2232 * Unfortunately, the recommended algorithm breaks header prediction,
2233 * since header prediction assumes th->window stays fixed.
2235 * Strictly speaking, keeping th->window fixed violates the receiver
2236 * side SWS prevention criteria. The problem is that under this rule
2237 * a stream of single byte packets will cause the right side of the
2238 * window to always advance by a single byte.
2240 * Of course, if the sender implements sender side SWS prevention
2241 * then this will not be a problem.
2243 * BSD seems to make the following compromise:
2245 * If the free space is less than the 1/4 of the maximum
2246 * space available and the free space is less than 1/2 mss,
2247 * then set the window to 0.
2248 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2249 * Otherwise, just prevent the window from shrinking
2250 * and from being larger than the largest representable value.
2252 * This prevents incremental opening of the window in the regime
2253 * where TCP is limited by the speed of the reader side taking
2254 * data out of the TCP receive queue. It does nothing about
2255 * those cases where the window is constrained on the sender side
2256 * because the pipeline is full.
2258 * BSD also seems to "accidentally" limit itself to windows that are a
2259 * multiple of MSS, at least until the free space gets quite small.
2260 * This would appear to be a side effect of the mbuf implementation.
2261 * Combining these two algorithms results in the observed behavior
2262 * of having a fixed window size at almost all times.
2264 * Below we obtain similar behavior by forcing the offered window to
2265 * a multiple of the mss when it is feasible to do so.
2267 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2268 * Regular options like TIMESTAMP are taken into account.
2270 u32 __tcp_select_window(struct sock *sk)
2272 struct inet_connection_sock *icsk = inet_csk(sk);
2273 struct tcp_sock *tp = tcp_sk(sk);
2274 /* MSS for the peer's data. Previous versions used mss_clamp
2275 * here. I don't know if the value based on our guesses
2276 * of peer's MSS is better for the performance. It's more correct
2277 * but may be worse for the performance because of rcv_mss
2278 * fluctuations. --SAW 1998/11/1
2280 int mss = icsk->icsk_ack.rcv_mss;
2281 int free_space = tcp_space(sk);
2282 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2285 if (mss > full_space)
2288 if (free_space < (full_space >> 1)) {
2289 icsk->icsk_ack.quick = 0;
2291 if (sk_under_memory_pressure(sk))
2292 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2295 if (free_space < mss)
2299 if (free_space > tp->rcv_ssthresh)
2300 free_space = tp->rcv_ssthresh;
2302 /* Don't do rounding if we are using window scaling, since the
2303 * scaled window will not line up with the MSS boundary anyway.
2305 window = tp->rcv_wnd;
2306 if (tp->rx_opt.rcv_wscale) {
2307 window = free_space;
2309 /* Advertise enough space so that it won't get scaled away.
2310 * Import case: prevent zero window announcement if
2311 * 1<<rcv_wscale > mss.
2313 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2314 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2315 << tp->rx_opt.rcv_wscale);
2317 /* Get the largest window that is a nice multiple of mss.
2318 * Window clamp already applied above.
2319 * If our current window offering is within 1 mss of the
2320 * free space we just keep it. This prevents the divide
2321 * and multiply from happening most of the time.
2322 * We also don't do any window rounding when the free space
2325 if (window <= free_space - mss || window > free_space)
2326 window = (free_space / mss) * mss;
2327 else if (mss == full_space &&
2328 free_space > window + (full_space >> 1))
2329 window = free_space;
2335 /* Collapses two adjacent SKB's during retransmission. */
2336 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2338 struct tcp_sock *tp = tcp_sk(sk);
2339 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2340 int skb_size, next_skb_size;
2342 skb_size = skb->len;
2343 next_skb_size = next_skb->len;
2345 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2347 tcp_highest_sack_combine(sk, next_skb, skb);
2349 tcp_unlink_write_queue(next_skb, sk);
2351 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2354 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2355 skb->ip_summed = CHECKSUM_PARTIAL;
2357 if (skb->ip_summed != CHECKSUM_PARTIAL)
2358 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2360 /* Update sequence range on original skb. */
2361 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2363 /* Merge over control information. This moves PSH/FIN etc. over */
2364 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2366 /* All done, get rid of second SKB and account for it so
2367 * packet counting does not break.
2369 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2371 /* changed transmit queue under us so clear hints */
2372 tcp_clear_retrans_hints_partial(tp);
2373 if (next_skb == tp->retransmit_skb_hint)
2374 tp->retransmit_skb_hint = skb;
2376 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2378 sk_wmem_free_skb(sk, next_skb);
2381 /* Check if coalescing SKBs is legal. */
2382 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2384 if (tcp_skb_pcount(skb) > 1)
2386 /* TODO: SACK collapsing could be used to remove this condition */
2387 if (skb_shinfo(skb)->nr_frags != 0)
2389 if (skb_cloned(skb))
2391 if (skb == tcp_send_head(sk))
2393 /* Some heurestics for collapsing over SACK'd could be invented */
2394 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2400 /* Collapse packets in the retransmit queue to make to create
2401 * less packets on the wire. This is only done on retransmission.
2403 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2406 struct tcp_sock *tp = tcp_sk(sk);
2407 struct sk_buff *skb = to, *tmp;
2410 if (!sysctl_tcp_retrans_collapse)
2412 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2415 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2416 if (!tcp_can_collapse(sk, skb))
2428 /* Punt if not enough space exists in the first SKB for
2429 * the data in the second
2431 if (skb->len > skb_availroom(to))
2434 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2437 tcp_collapse_retrans(sk, to);
2441 /* This retransmits one SKB. Policy decisions and retransmit queue
2442 * state updates are done by the caller. Returns non-zero if an
2443 * error occurred which prevented the send.
2445 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2447 struct tcp_sock *tp = tcp_sk(sk);
2448 struct inet_connection_sock *icsk = inet_csk(sk);
2449 unsigned int cur_mss;
2451 /* Inconslusive MTU probe */
2452 if (icsk->icsk_mtup.probe_size) {
2453 icsk->icsk_mtup.probe_size = 0;
2456 /* Do not sent more than we queued. 1/4 is reserved for possible
2457 * copying overhead: fragmentation, tunneling, mangling etc.
2459 if (atomic_read(&sk->sk_wmem_alloc) >
2460 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2463 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2464 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2466 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2470 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2471 return -EHOSTUNREACH; /* Routing failure or similar. */
2473 cur_mss = tcp_current_mss(sk);
2475 /* If receiver has shrunk his window, and skb is out of
2476 * new window, do not retransmit it. The exception is the
2477 * case, when window is shrunk to zero. In this case
2478 * our retransmit serves as a zero window probe.
2480 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2481 TCP_SKB_CB(skb)->seq != tp->snd_una)
2484 if (skb->len > cur_mss) {
2485 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2486 return -ENOMEM; /* We'll try again later. */
2488 int oldpcount = tcp_skb_pcount(skb);
2490 if (unlikely(oldpcount > 1)) {
2491 tcp_init_tso_segs(sk, skb, cur_mss);
2492 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2496 tcp_retrans_try_collapse(sk, skb, cur_mss);
2498 /* Some Solaris stacks overoptimize and ignore the FIN on a
2499 * retransmit when old data is attached. So strip it off
2500 * since it is cheap to do so and saves bytes on the network.
2503 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2504 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2505 if (!pskb_trim(skb, 0)) {
2506 /* Reuse, even though it does some unnecessary work */
2507 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2508 TCP_SKB_CB(skb)->tcp_flags);
2509 skb->ip_summed = CHECKSUM_NONE;
2513 /* Make a copy, if the first transmission SKB clone we made
2514 * is still in somebody's hands, else make a clone.
2516 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2518 /* make sure skb->data is aligned on arches that require it */
2519 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2520 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2522 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2525 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2529 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2531 struct tcp_sock *tp = tcp_sk(sk);
2532 int err = __tcp_retransmit_skb(sk, skb);
2535 /* Update global TCP statistics. */
2536 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2538 tp->total_retrans++;
2540 #if FASTRETRANS_DEBUG > 0
2541 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2542 net_dbg_ratelimited("retrans_out leaked\n");
2545 if (!tp->retrans_out)
2546 tp->lost_retrans_low = tp->snd_nxt;
2547 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2548 tp->retrans_out += tcp_skb_pcount(skb);
2550 /* Save stamp of the first retransmit. */
2551 if (!tp->retrans_stamp)
2552 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2554 tp->undo_retrans += tcp_skb_pcount(skb);
2556 /* snd_nxt is stored to detect loss of retransmitted segment,
2557 * see tcp_input.c tcp_sacktag_write_queue().
2559 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2564 /* Check if we forward retransmits are possible in the current
2565 * window/congestion state.
2567 static bool tcp_can_forward_retransmit(struct sock *sk)
2569 const struct inet_connection_sock *icsk = inet_csk(sk);
2570 const struct tcp_sock *tp = tcp_sk(sk);
2572 /* Forward retransmissions are possible only during Recovery. */
2573 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2576 /* No forward retransmissions in Reno are possible. */
2577 if (tcp_is_reno(tp))
2580 /* Yeah, we have to make difficult choice between forward transmission
2581 * and retransmission... Both ways have their merits...
2583 * For now we do not retransmit anything, while we have some new
2584 * segments to send. In the other cases, follow rule 3 for
2585 * NextSeg() specified in RFC3517.
2588 if (tcp_may_send_now(sk))
2594 /* This gets called after a retransmit timeout, and the initially
2595 * retransmitted data is acknowledged. It tries to continue
2596 * resending the rest of the retransmit queue, until either
2597 * we've sent it all or the congestion window limit is reached.
2598 * If doing SACK, the first ACK which comes back for a timeout
2599 * based retransmit packet might feed us FACK information again.
2600 * If so, we use it to avoid unnecessarily retransmissions.
2602 void tcp_xmit_retransmit_queue(struct sock *sk)
2604 const struct inet_connection_sock *icsk = inet_csk(sk);
2605 struct tcp_sock *tp = tcp_sk(sk);
2606 struct sk_buff *skb;
2607 struct sk_buff *hole = NULL;
2610 int fwd_rexmitting = 0;
2612 if (!tp->packets_out)
2616 tp->retransmit_high = tp->snd_una;
2618 if (tp->retransmit_skb_hint) {
2619 skb = tp->retransmit_skb_hint;
2620 last_lost = TCP_SKB_CB(skb)->end_seq;
2621 if (after(last_lost, tp->retransmit_high))
2622 last_lost = tp->retransmit_high;
2624 skb = tcp_write_queue_head(sk);
2625 last_lost = tp->snd_una;
2628 tcp_for_write_queue_from(skb, sk) {
2629 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2631 if (skb == tcp_send_head(sk))
2633 /* we could do better than to assign each time */
2635 tp->retransmit_skb_hint = skb;
2637 /* Assume this retransmit will generate
2638 * only one packet for congestion window
2639 * calculation purposes. This works because
2640 * tcp_retransmit_skb() will chop up the
2641 * packet to be MSS sized and all the
2642 * packet counting works out.
2644 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2647 if (fwd_rexmitting) {
2649 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2651 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2653 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2654 tp->retransmit_high = last_lost;
2655 if (!tcp_can_forward_retransmit(sk))
2657 /* Backtrack if necessary to non-L'ed skb */
2665 } else if (!(sacked & TCPCB_LOST)) {
2666 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2671 last_lost = TCP_SKB_CB(skb)->end_seq;
2672 if (icsk->icsk_ca_state != TCP_CA_Loss)
2673 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2675 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2678 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2681 if (tcp_retransmit_skb(sk, skb)) {
2682 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2685 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2687 if (tcp_in_cwnd_reduction(sk))
2688 tp->prr_out += tcp_skb_pcount(skb);
2690 if (skb == tcp_write_queue_head(sk))
2691 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2692 inet_csk(sk)->icsk_rto,
2697 /* Send a fin. The caller locks the socket for us. This cannot be
2698 * allowed to fail queueing a FIN frame under any circumstances.
2700 void tcp_send_fin(struct sock *sk)
2702 struct tcp_sock *tp = tcp_sk(sk);
2703 struct sk_buff *skb = tcp_write_queue_tail(sk);
2706 /* Optimization, tack on the FIN if we have a queue of
2707 * unsent frames. But be careful about outgoing SACKS
2710 mss_now = tcp_current_mss(sk);
2712 if (tcp_send_head(sk) != NULL) {
2713 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2714 TCP_SKB_CB(skb)->end_seq++;
2717 /* Socket is locked, keep trying until memory is available. */
2719 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2726 /* Reserve space for headers and prepare control bits. */
2727 skb_reserve(skb, MAX_TCP_HEADER);
2728 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2729 tcp_init_nondata_skb(skb, tp->write_seq,
2730 TCPHDR_ACK | TCPHDR_FIN);
2731 tcp_queue_skb(sk, skb);
2733 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2736 /* We get here when a process closes a file descriptor (either due to
2737 * an explicit close() or as a byproduct of exit()'ing) and there
2738 * was unread data in the receive queue. This behavior is recommended
2739 * by RFC 2525, section 2.17. -DaveM
2741 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2743 struct sk_buff *skb;
2745 /* NOTE: No TCP options attached and we never retransmit this. */
2746 skb = alloc_skb(MAX_TCP_HEADER, priority);
2748 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2752 /* Reserve space for headers and prepare control bits. */
2753 skb_reserve(skb, MAX_TCP_HEADER);
2754 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2755 TCPHDR_ACK | TCPHDR_RST);
2757 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2758 if (tcp_transmit_skb(sk, skb, 0, priority))
2759 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2761 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2764 /* Send a crossed SYN-ACK during socket establishment.
2765 * WARNING: This routine must only be called when we have already sent
2766 * a SYN packet that crossed the incoming SYN that caused this routine
2767 * to get called. If this assumption fails then the initial rcv_wnd
2768 * and rcv_wscale values will not be correct.
2770 int tcp_send_synack(struct sock *sk)
2772 struct sk_buff *skb;
2774 skb = tcp_write_queue_head(sk);
2775 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2776 pr_debug("%s: wrong queue state\n", __func__);
2779 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2780 if (skb_cloned(skb)) {
2781 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2784 tcp_unlink_write_queue(skb, sk);
2785 skb_header_release(nskb);
2786 __tcp_add_write_queue_head(sk, nskb);
2787 sk_wmem_free_skb(sk, skb);
2788 sk->sk_wmem_queued += nskb->truesize;
2789 sk_mem_charge(sk, nskb->truesize);
2793 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2794 TCP_ECN_send_synack(tcp_sk(sk), skb);
2796 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2797 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2801 * tcp_make_synack - Prepare a SYN-ACK.
2802 * sk: listener socket
2803 * dst: dst entry attached to the SYNACK
2804 * req: request_sock pointer
2805 * rvp: request_values pointer
2807 * Allocate one skb and build a SYNACK packet.
2808 * @dst is consumed : Caller should not use it again.
2810 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2811 struct request_sock *req,
2812 struct request_values *rvp,
2813 struct tcp_fastopen_cookie *foc)
2815 struct tcp_out_options opts;
2816 struct tcp_extend_values *xvp = tcp_xv(rvp);
2817 struct inet_request_sock *ireq = inet_rsk(req);
2818 struct tcp_sock *tp = tcp_sk(sk);
2819 const struct tcp_cookie_values *cvp = tp->cookie_values;
2821 struct sk_buff *skb;
2822 struct tcp_md5sig_key *md5;
2823 int tcp_header_size;
2825 int s_data_desired = 0;
2827 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2828 s_data_desired = cvp->s_data_desired;
2829 skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired,
2830 sk_gfp_atomic(sk, GFP_ATOMIC));
2831 if (unlikely(!skb)) {
2835 /* Reserve space for headers. */
2836 skb_reserve(skb, MAX_TCP_HEADER);
2838 skb_dst_set(skb, dst);
2840 mss = dst_metric_advmss(dst);
2841 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2842 mss = tp->rx_opt.user_mss;
2844 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2846 /* Set this up on the first call only */
2847 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2849 /* limit the window selection if the user enforce a smaller rx buffer */
2850 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2851 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2852 req->window_clamp = tcp_full_space(sk);
2854 /* tcp_full_space because it is guaranteed to be the first packet */
2855 tcp_select_initial_window(tcp_full_space(sk),
2856 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2861 dst_metric(dst, RTAX_INITRWND));
2862 ireq->rcv_wscale = rcv_wscale;
2865 memset(&opts, 0, sizeof(opts));
2866 #ifdef CONFIG_SYN_COOKIES
2867 if (unlikely(req->cookie_ts))
2868 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2871 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2872 tcp_header_size = tcp_synack_options(sk, req, mss,
2873 skb, &opts, &md5, xvp, foc)
2876 skb_push(skb, tcp_header_size);
2877 skb_reset_transport_header(skb);
2880 memset(th, 0, sizeof(struct tcphdr));
2883 TCP_ECN_make_synack(req, th);
2884 th->source = ireq->loc_port;
2885 th->dest = ireq->rmt_port;
2886 /* Setting of flags are superfluous here for callers (and ECE is
2887 * not even correctly set)
2889 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2890 TCPHDR_SYN | TCPHDR_ACK);
2892 if (OPTION_COOKIE_EXTENSION & opts.options) {
2893 if (s_data_desired) {
2894 u8 *buf = skb_put(skb, s_data_desired);
2896 /* copy data directly from the listening socket. */
2897 memcpy(buf, cvp->s_data_payload, s_data_desired);
2898 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2901 if (opts.hash_size > 0) {
2902 __u32 workspace[SHA_WORKSPACE_WORDS];
2903 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2904 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2906 /* Secret recipe depends on the Timestamp, (future)
2907 * Sequence and Acknowledgment Numbers, Initiator
2908 * Cookie, and others handled by IP variant caller.
2910 *tail-- ^= opts.tsval;
2911 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2912 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2915 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2916 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2918 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2921 opts.hash_location =
2922 (__u8 *)&xvp->cookie_bakery[0];
2926 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2927 /* XXX data is queued and acked as is. No buffer/window check */
2928 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2930 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2931 th->window = htons(min(req->rcv_wnd, 65535U));
2932 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2933 th->doff = (tcp_header_size >> 2);
2934 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2936 #ifdef CONFIG_TCP_MD5SIG
2937 /* Okay, we have all we need - do the md5 hash if needed */
2939 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2940 md5, NULL, req, skb);
2946 EXPORT_SYMBOL(tcp_make_synack);
2948 /* Do all connect socket setups that can be done AF independent. */
2949 void tcp_connect_init(struct sock *sk)
2951 const struct dst_entry *dst = __sk_dst_get(sk);
2952 struct tcp_sock *tp = tcp_sk(sk);
2955 /* We'll fix this up when we get a response from the other end.
2956 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2958 tp->tcp_header_len = sizeof(struct tcphdr) +
2959 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2961 #ifdef CONFIG_TCP_MD5SIG
2962 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2963 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2966 /* If user gave his TCP_MAXSEG, record it to clamp */
2967 if (tp->rx_opt.user_mss)
2968 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2971 tcp_sync_mss(sk, dst_mtu(dst));
2973 if (!tp->window_clamp)
2974 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2975 tp->advmss = dst_metric_advmss(dst);
2976 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2977 tp->advmss = tp->rx_opt.user_mss;
2979 tcp_initialize_rcv_mss(sk);
2981 /* limit the window selection if the user enforce a smaller rx buffer */
2982 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2983 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2984 tp->window_clamp = tcp_full_space(sk);
2986 tcp_select_initial_window(tcp_full_space(sk),
2987 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2990 sysctl_tcp_window_scaling,
2992 dst_metric(dst, RTAX_INITRWND));
2994 tp->rx_opt.rcv_wscale = rcv_wscale;
2995 tp->rcv_ssthresh = tp->rcv_wnd;
2998 sock_reset_flag(sk, SOCK_DONE);
3001 tp->snd_una = tp->write_seq;
3002 tp->snd_sml = tp->write_seq;
3003 tp->snd_up = tp->write_seq;
3004 tp->snd_nxt = tp->write_seq;
3006 if (likely(!tp->repair))
3008 tp->rcv_wup = tp->rcv_nxt;
3009 tp->copied_seq = tp->rcv_nxt;
3011 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3012 inet_csk(sk)->icsk_retransmits = 0;
3013 tcp_clear_retrans(tp);
3016 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3018 struct tcp_sock *tp = tcp_sk(sk);
3019 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3021 tcb->end_seq += skb->len;
3022 skb_header_release(skb);
3023 __tcp_add_write_queue_tail(sk, skb);
3024 sk->sk_wmem_queued += skb->truesize;
3025 sk_mem_charge(sk, skb->truesize);
3026 tp->write_seq = tcb->end_seq;
3027 tp->packets_out += tcp_skb_pcount(skb);
3030 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3031 * queue a data-only packet after the regular SYN, such that regular SYNs
3032 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3033 * only the SYN sequence, the data are retransmitted in the first ACK.
3034 * If cookie is not cached or other error occurs, falls back to send a
3035 * regular SYN with Fast Open cookie request option.
3037 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3039 struct tcp_sock *tp = tcp_sk(sk);
3040 struct tcp_fastopen_request *fo = tp->fastopen_req;
3041 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
3042 struct sk_buff *syn_data = NULL, *data;
3043 unsigned long last_syn_loss = 0;
3045 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3046 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3047 &syn_loss, &last_syn_loss);
3048 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3050 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3051 fo->cookie.len = -1;
3055 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3056 fo->cookie.len = -1;
3057 else if (fo->cookie.len <= 0)
3060 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3061 * user-MSS. Reserve maximum option space for middleboxes that add
3062 * private TCP options. The cost is reduced data space in SYN :(
3064 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3065 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3066 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3067 MAX_TCP_OPTION_SPACE;
3069 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
3071 if (syn_data == NULL)
3074 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3075 struct iovec *iov = &fo->data->msg_iov[i];
3076 unsigned char __user *from = iov->iov_base;
3077 int len = iov->iov_len;
3079 if (syn_data->len + len > space)
3080 len = space - syn_data->len;
3081 else if (i + 1 == iovlen)
3082 /* No more data pending in inet_wait_for_connect() */
3085 if (skb_add_data(syn_data, from, len))
3089 /* Queue a data-only packet after the regular SYN for retransmission */
3090 data = pskb_copy(syn_data, sk->sk_allocation);
3093 TCP_SKB_CB(data)->seq++;
3094 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3095 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3096 tcp_connect_queue_skb(sk, data);
3097 fo->copied = data->len;
3099 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
3100 tp->syn_data = (fo->copied > 0);
3101 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
3107 /* Send a regular SYN with Fast Open cookie request option */
3108 if (fo->cookie.len > 0)
3110 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3112 tp->syn_fastopen = 0;
3113 kfree_skb(syn_data);
3115 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3119 /* Build a SYN and send it off. */
3120 int tcp_connect(struct sock *sk)
3122 struct tcp_sock *tp = tcp_sk(sk);
3123 struct sk_buff *buff;
3126 tcp_connect_init(sk);
3128 if (unlikely(tp->repair)) {
3129 tcp_finish_connect(sk, NULL);
3133 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3134 if (unlikely(buff == NULL))
3137 /* Reserve space for headers. */
3138 skb_reserve(buff, MAX_TCP_HEADER);
3140 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3141 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
3142 tcp_connect_queue_skb(sk, buff);
3143 TCP_ECN_send_syn(sk, buff);
3145 /* Send off SYN; include data in Fast Open. */
3146 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3147 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3148 if (err == -ECONNREFUSED)
3151 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3152 * in order to make this packet get counted in tcpOutSegs.
3154 tp->snd_nxt = tp->write_seq;
3155 tp->pushed_seq = tp->write_seq;
3156 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3158 /* Timer for repeating the SYN until an answer. */
3159 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3160 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3163 EXPORT_SYMBOL(tcp_connect);
3165 /* Send out a delayed ack, the caller does the policy checking
3166 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3169 void tcp_send_delayed_ack(struct sock *sk)
3171 struct inet_connection_sock *icsk = inet_csk(sk);
3172 int ato = icsk->icsk_ack.ato;
3173 unsigned long timeout;
3175 if (ato > TCP_DELACK_MIN) {
3176 const struct tcp_sock *tp = tcp_sk(sk);
3177 int max_ato = HZ / 2;
3179 if (icsk->icsk_ack.pingpong ||
3180 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3181 max_ato = TCP_DELACK_MAX;
3183 /* Slow path, intersegment interval is "high". */
3185 /* If some rtt estimate is known, use it to bound delayed ack.
3186 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3190 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3196 ato = min(ato, max_ato);
3199 /* Stay within the limit we were given */
3200 timeout = jiffies + ato;
3202 /* Use new timeout only if there wasn't a older one earlier. */
3203 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3204 /* If delack timer was blocked or is about to expire,
3207 if (icsk->icsk_ack.blocked ||
3208 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3213 if (!time_before(timeout, icsk->icsk_ack.timeout))
3214 timeout = icsk->icsk_ack.timeout;
3216 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3217 icsk->icsk_ack.timeout = timeout;
3218 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3221 /* This routine sends an ack and also updates the window. */
3222 void tcp_send_ack(struct sock *sk)
3224 struct sk_buff *buff;
3226 /* If we have been reset, we may not send again. */
3227 if (sk->sk_state == TCP_CLOSE)
3230 /* We are not putting this on the write queue, so
3231 * tcp_transmit_skb() will set the ownership to this
3234 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3236 inet_csk_schedule_ack(sk);
3237 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3238 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3239 TCP_DELACK_MAX, TCP_RTO_MAX);
3243 /* Reserve space for headers and prepare control bits. */
3244 skb_reserve(buff, MAX_TCP_HEADER);
3245 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3247 /* Send it off, this clears delayed acks for us. */
3248 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3249 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3252 /* This routine sends a packet with an out of date sequence
3253 * number. It assumes the other end will try to ack it.
3255 * Question: what should we make while urgent mode?
3256 * 4.4BSD forces sending single byte of data. We cannot send
3257 * out of window data, because we have SND.NXT==SND.MAX...
3259 * Current solution: to send TWO zero-length segments in urgent mode:
3260 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3261 * out-of-date with SND.UNA-1 to probe window.
3263 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3265 struct tcp_sock *tp = tcp_sk(sk);
3266 struct sk_buff *skb;
3268 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3269 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3273 /* Reserve space for headers and set control bits. */
3274 skb_reserve(skb, MAX_TCP_HEADER);
3275 /* Use a previous sequence. This should cause the other
3276 * end to send an ack. Don't queue or clone SKB, just
3279 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3280 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3281 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3284 void tcp_send_window_probe(struct sock *sk)
3286 if (sk->sk_state == TCP_ESTABLISHED) {
3287 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3288 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3289 tcp_xmit_probe_skb(sk, 0);
3293 /* Initiate keepalive or window probe from timer. */
3294 int tcp_write_wakeup(struct sock *sk)
3296 struct tcp_sock *tp = tcp_sk(sk);
3297 struct sk_buff *skb;
3299 if (sk->sk_state == TCP_CLOSE)
3302 if ((skb = tcp_send_head(sk)) != NULL &&
3303 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3305 unsigned int mss = tcp_current_mss(sk);
3306 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3308 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3309 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3311 /* We are probing the opening of a window
3312 * but the window size is != 0
3313 * must have been a result SWS avoidance ( sender )
3315 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3317 seg_size = min(seg_size, mss);
3318 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3319 if (tcp_fragment(sk, skb, seg_size, mss))
3321 } else if (!tcp_skb_pcount(skb))
3322 tcp_set_skb_tso_segs(sk, skb, mss);
3324 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3325 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3326 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3328 tcp_event_new_data_sent(sk, skb);
3331 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3332 tcp_xmit_probe_skb(sk, 1);
3333 return tcp_xmit_probe_skb(sk, 0);
3337 /* A window probe timeout has occurred. If window is not closed send
3338 * a partial packet else a zero probe.
3340 void tcp_send_probe0(struct sock *sk)
3342 struct inet_connection_sock *icsk = inet_csk(sk);
3343 struct tcp_sock *tp = tcp_sk(sk);
3346 err = tcp_write_wakeup(sk);
3348 if (tp->packets_out || !tcp_send_head(sk)) {
3349 /* Cancel probe timer, if it is not required. */
3350 icsk->icsk_probes_out = 0;
3351 icsk->icsk_backoff = 0;
3356 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3357 icsk->icsk_backoff++;
3358 icsk->icsk_probes_out++;
3359 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3360 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3363 /* If packet was not sent due to local congestion,
3364 * do not backoff and do not remember icsk_probes_out.
3365 * Let local senders to fight for local resources.
3367 * Use accumulated backoff yet.
3369 if (!icsk->icsk_probes_out)
3370 icsk->icsk_probes_out = 1;
3371 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3372 min(icsk->icsk_rto << icsk->icsk_backoff,
3373 TCP_RESOURCE_PROBE_INTERVAL),