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 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
69 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
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 tp->packets_out += tcp_skb_pcount(skb);
85 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
86 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
99 const struct tcp_sock *tp = tcp_sk(sk);
101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
143 struct tcp_sock *tp = tcp_sk(sk);
144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
166 const struct dst_entry *dst = __sk_dst_get(sk);
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 icsk->icsk_ack.pingpong = 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 u32 tcp_default_init_rwnd(u32 mss)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd = TCP_INIT_CWND * 2;
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
213 int wscale_ok, __u8 *rcv_wscale,
216 unsigned int space = (__space < 0 ? 0 : __space);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (65535 << 14);
221 space = min(*window_clamp, space);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space = (space / mss) * mss;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 while (space > 65535 && (*rcv_wscale) < 14) {
253 if (mss > (1 << *rcv_wscale)) {
254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
262 EXPORT_SYMBOL(tcp_select_initial_window);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16 tcp_select_window(struct sock *sk)
271 struct tcp_sock *tp = tcp_sk(sk);
272 u32 cur_win = tcp_receive_window(tp);
273 u32 new_win = __tcp_select_window(sk);
275 /* Never shrink the offered window */
276 if (new_win < cur_win) {
277 /* Danger Will Robinson!
278 * Don't update rcv_wup/rcv_wnd here or else
279 * we will not be able to advertise a zero
280 * window in time. --DaveM
282 * Relax Will Robinson.
284 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
286 tp->rcv_wnd = new_win;
287 tp->rcv_wup = tp->rcv_nxt;
289 /* Make sure we do not exceed the maximum possible
292 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
293 new_win = min(new_win, MAX_TCP_WINDOW);
295 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
297 /* RFC1323 scaling applied */
298 new_win >>= tp->rx_opt.rcv_wscale;
300 /* If we advertise zero window, disable fast path. */
307 /* Packet ECN state for a SYN-ACK */
308 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
310 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
311 if (!(tp->ecn_flags & TCP_ECN_OK))
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
315 /* Packet ECN state for a SYN. */
316 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
318 struct tcp_sock *tp = tcp_sk(sk);
321 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
322 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
323 tp->ecn_flags = TCP_ECN_OK;
327 static __inline__ void
328 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
330 if (inet_rsk(req)->ecn_ok)
334 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
337 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
340 struct tcp_sock *tp = tcp_sk(sk);
342 if (tp->ecn_flags & TCP_ECN_OK) {
343 /* Not-retransmitted data segment: set ECT and inject CWR. */
344 if (skb->len != tcp_header_len &&
345 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
347 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
348 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
349 tcp_hdr(skb)->cwr = 1;
350 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
353 /* ACK or retransmitted segment: clear ECT|CE */
354 INET_ECN_dontxmit(sk);
356 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
357 tcp_hdr(skb)->ece = 1;
361 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
362 * auto increment end seqno.
364 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
366 struct skb_shared_info *shinfo = skb_shinfo(skb);
368 skb->ip_summed = CHECKSUM_PARTIAL;
371 TCP_SKB_CB(skb)->tcp_flags = flags;
372 TCP_SKB_CB(skb)->sacked = 0;
374 shinfo->gso_segs = 1;
375 shinfo->gso_size = 0;
376 shinfo->gso_type = 0;
378 TCP_SKB_CB(skb)->seq = seq;
379 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
381 TCP_SKB_CB(skb)->end_seq = seq;
384 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
386 return tp->snd_una != tp->snd_up;
389 #define OPTION_SACK_ADVERTISE (1 << 0)
390 #define OPTION_TS (1 << 1)
391 #define OPTION_MD5 (1 << 2)
392 #define OPTION_WSCALE (1 << 3)
393 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
395 struct tcp_out_options {
396 u16 options; /* bit field of OPTION_* */
397 u16 mss; /* 0 to disable */
398 u8 ws; /* window scale, 0 to disable */
399 u8 num_sack_blocks; /* number of SACK blocks to include */
400 u8 hash_size; /* bytes in hash_location */
401 __u8 *hash_location; /* temporary pointer, overloaded */
402 __u32 tsval, tsecr; /* need to include OPTION_TS */
403 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
406 /* Write previously computed TCP options to the packet.
408 * Beware: Something in the Internet is very sensitive to the ordering of
409 * TCP options, we learned this through the hard way, so be careful here.
410 * Luckily we can at least blame others for their non-compliance but from
411 * inter-operability perspective it seems that we're somewhat stuck with
412 * the ordering which we have been using if we want to keep working with
413 * those broken things (not that it currently hurts anybody as there isn't
414 * particular reason why the ordering would need to be changed).
416 * At least SACK_PERM as the first option is known to lead to a disaster
417 * (but it may well be that other scenarios fail similarly).
419 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
420 struct tcp_out_options *opts)
422 u16 options = opts->options; /* mungable copy */
424 if (unlikely(OPTION_MD5 & options)) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
426 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
427 /* overload cookie hash location */
428 opts->hash_location = (__u8 *)ptr;
432 if (unlikely(opts->mss)) {
433 *ptr++ = htonl((TCPOPT_MSS << 24) |
434 (TCPOLEN_MSS << 16) |
438 if (likely(OPTION_TS & options)) {
439 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
440 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
441 (TCPOLEN_SACK_PERM << 16) |
442 (TCPOPT_TIMESTAMP << 8) |
444 options &= ~OPTION_SACK_ADVERTISE;
446 *ptr++ = htonl((TCPOPT_NOP << 24) |
448 (TCPOPT_TIMESTAMP << 8) |
451 *ptr++ = htonl(opts->tsval);
452 *ptr++ = htonl(opts->tsecr);
455 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
458 (TCPOPT_SACK_PERM << 8) |
462 if (unlikely(OPTION_WSCALE & options)) {
463 *ptr++ = htonl((TCPOPT_NOP << 24) |
464 (TCPOPT_WINDOW << 16) |
465 (TCPOLEN_WINDOW << 8) |
469 if (unlikely(opts->num_sack_blocks)) {
470 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
471 tp->duplicate_sack : tp->selective_acks;
474 *ptr++ = htonl((TCPOPT_NOP << 24) |
477 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
478 TCPOLEN_SACK_PERBLOCK)));
480 for (this_sack = 0; this_sack < opts->num_sack_blocks;
482 *ptr++ = htonl(sp[this_sack].start_seq);
483 *ptr++ = htonl(sp[this_sack].end_seq);
486 tp->rx_opt.dsack = 0;
489 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
490 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
492 *ptr++ = htonl((TCPOPT_EXP << 24) |
493 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
494 TCPOPT_FASTOPEN_MAGIC);
496 memcpy(ptr, foc->val, foc->len);
497 if ((foc->len & 3) == 2) {
498 u8 *align = ((u8 *)ptr) + foc->len;
499 align[0] = align[1] = TCPOPT_NOP;
501 ptr += (foc->len + 3) >> 2;
505 /* Compute TCP options for SYN packets. This is not the final
506 * network wire format yet.
508 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
509 struct tcp_out_options *opts,
510 struct tcp_md5sig_key **md5)
512 struct tcp_sock *tp = tcp_sk(sk);
513 unsigned int remaining = MAX_TCP_OPTION_SPACE;
514 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
516 #ifdef CONFIG_TCP_MD5SIG
517 *md5 = tp->af_specific->md5_lookup(sk, sk);
519 opts->options |= OPTION_MD5;
520 remaining -= TCPOLEN_MD5SIG_ALIGNED;
526 /* We always get an MSS option. The option bytes which will be seen in
527 * normal data packets should timestamps be used, must be in the MSS
528 * advertised. But we subtract them from tp->mss_cache so that
529 * calculations in tcp_sendmsg are simpler etc. So account for this
530 * fact here if necessary. If we don't do this correctly, as a
531 * receiver we won't recognize data packets as being full sized when we
532 * should, and thus we won't abide by the delayed ACK rules correctly.
533 * SACKs don't matter, we never delay an ACK when we have any of those
535 opts->mss = tcp_advertise_mss(sk);
536 remaining -= TCPOLEN_MSS_ALIGNED;
538 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
539 opts->options |= OPTION_TS;
540 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
541 opts->tsecr = tp->rx_opt.ts_recent;
542 remaining -= TCPOLEN_TSTAMP_ALIGNED;
544 if (likely(sysctl_tcp_window_scaling)) {
545 opts->ws = tp->rx_opt.rcv_wscale;
546 opts->options |= OPTION_WSCALE;
547 remaining -= TCPOLEN_WSCALE_ALIGNED;
549 if (likely(sysctl_tcp_sack)) {
550 opts->options |= OPTION_SACK_ADVERTISE;
551 if (unlikely(!(OPTION_TS & opts->options)))
552 remaining -= TCPOLEN_SACKPERM_ALIGNED;
555 if (fastopen && fastopen->cookie.len >= 0) {
556 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
557 need = (need + 3) & ~3U; /* Align to 32 bits */
558 if (remaining >= need) {
559 opts->options |= OPTION_FAST_OPEN_COOKIE;
560 opts->fastopen_cookie = &fastopen->cookie;
562 tp->syn_fastopen = 1;
566 return MAX_TCP_OPTION_SPACE - remaining;
569 /* Set up TCP options for SYN-ACKs. */
570 static unsigned int tcp_synack_options(struct sock *sk,
571 struct request_sock *req,
572 unsigned int mss, struct sk_buff *skb,
573 struct tcp_out_options *opts,
574 struct tcp_md5sig_key **md5,
575 struct tcp_fastopen_cookie *foc)
577 struct inet_request_sock *ireq = inet_rsk(req);
578 unsigned int remaining = MAX_TCP_OPTION_SPACE;
580 #ifdef CONFIG_TCP_MD5SIG
581 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
583 opts->options |= OPTION_MD5;
584 remaining -= TCPOLEN_MD5SIG_ALIGNED;
586 /* We can't fit any SACK blocks in a packet with MD5 + TS
587 * options. There was discussion about disabling SACK
588 * rather than TS in order to fit in better with old,
589 * buggy kernels, but that was deemed to be unnecessary.
591 ireq->tstamp_ok &= !ireq->sack_ok;
597 /* We always send an MSS option. */
599 remaining -= TCPOLEN_MSS_ALIGNED;
601 if (likely(ireq->wscale_ok)) {
602 opts->ws = ireq->rcv_wscale;
603 opts->options |= OPTION_WSCALE;
604 remaining -= TCPOLEN_WSCALE_ALIGNED;
606 if (likely(ireq->tstamp_ok)) {
607 opts->options |= OPTION_TS;
608 opts->tsval = TCP_SKB_CB(skb)->when;
609 opts->tsecr = req->ts_recent;
610 remaining -= TCPOLEN_TSTAMP_ALIGNED;
612 if (likely(ireq->sack_ok)) {
613 opts->options |= OPTION_SACK_ADVERTISE;
614 if (unlikely(!ireq->tstamp_ok))
615 remaining -= TCPOLEN_SACKPERM_ALIGNED;
618 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
619 need = (need + 3) & ~3U; /* Align to 32 bits */
620 if (remaining >= need) {
621 opts->options |= OPTION_FAST_OPEN_COOKIE;
622 opts->fastopen_cookie = foc;
627 return MAX_TCP_OPTION_SPACE - remaining;
630 /* Compute TCP options for ESTABLISHED sockets. This is not the
631 * final wire format yet.
633 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
634 struct tcp_out_options *opts,
635 struct tcp_md5sig_key **md5)
637 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
638 struct tcp_sock *tp = tcp_sk(sk);
639 unsigned int size = 0;
640 unsigned int eff_sacks;
644 #ifdef CONFIG_TCP_MD5SIG
645 *md5 = tp->af_specific->md5_lookup(sk, sk);
646 if (unlikely(*md5)) {
647 opts->options |= OPTION_MD5;
648 size += TCPOLEN_MD5SIG_ALIGNED;
654 if (likely(tp->rx_opt.tstamp_ok)) {
655 opts->options |= OPTION_TS;
656 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
657 opts->tsecr = tp->rx_opt.ts_recent;
658 size += TCPOLEN_TSTAMP_ALIGNED;
661 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
662 if (unlikely(eff_sacks)) {
663 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
664 opts->num_sack_blocks =
665 min_t(unsigned int, eff_sacks,
666 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
667 TCPOLEN_SACK_PERBLOCK);
668 size += TCPOLEN_SACK_BASE_ALIGNED +
669 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
676 /* TCP SMALL QUEUES (TSQ)
678 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
679 * to reduce RTT and bufferbloat.
680 * We do this using a special skb destructor (tcp_wfree).
682 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
683 * needs to be reallocated in a driver.
684 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
686 * Since transmit from skb destructor is forbidden, we use a tasklet
687 * to process all sockets that eventually need to send more skbs.
688 * We use one tasklet per cpu, with its own queue of sockets.
691 struct tasklet_struct tasklet;
692 struct list_head head; /* queue of tcp sockets */
694 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
696 static void tcp_tsq_handler(struct sock *sk)
698 if ((1 << sk->sk_state) &
699 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
700 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
701 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
705 * One tasklet per cpu tries to send more skbs.
706 * We run in tasklet context but need to disable irqs when
707 * transferring tsq->head because tcp_wfree() might
708 * interrupt us (non NAPI drivers)
710 static void tcp_tasklet_func(unsigned long data)
712 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
715 struct list_head *q, *n;
719 local_irq_save(flags);
720 list_splice_init(&tsq->head, &list);
721 local_irq_restore(flags);
723 list_for_each_safe(q, n, &list) {
724 tp = list_entry(q, struct tcp_sock, tsq_node);
725 list_del(&tp->tsq_node);
727 sk = (struct sock *)tp;
730 if (!sock_owned_by_user(sk)) {
733 /* defer the work to tcp_release_cb() */
734 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
738 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
743 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
744 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
745 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
746 (1UL << TCP_MTU_REDUCED_DEFERRED))
748 * tcp_release_cb - tcp release_sock() callback
751 * called from release_sock() to perform protocol dependent
752 * actions before socket release.
754 void tcp_release_cb(struct sock *sk)
756 struct tcp_sock *tp = tcp_sk(sk);
757 unsigned long flags, nflags;
759 /* perform an atomic operation only if at least one flag is set */
761 flags = tp->tsq_flags;
762 if (!(flags & TCP_DEFERRED_ALL))
764 nflags = flags & ~TCP_DEFERRED_ALL;
765 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
767 if (flags & (1UL << TCP_TSQ_DEFERRED))
770 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
771 tcp_write_timer_handler(sk);
774 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
775 tcp_delack_timer_handler(sk);
778 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
779 sk->sk_prot->mtu_reduced(sk);
783 EXPORT_SYMBOL(tcp_release_cb);
785 void __init tcp_tasklet_init(void)
789 for_each_possible_cpu(i) {
790 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
792 INIT_LIST_HEAD(&tsq->head);
793 tasklet_init(&tsq->tasklet,
800 * Write buffer destructor automatically called from kfree_skb.
801 * We can't xmit new skbs from this context, as we might already
804 void tcp_wfree(struct sk_buff *skb)
806 struct sock *sk = skb->sk;
807 struct tcp_sock *tp = tcp_sk(sk);
809 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
810 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
812 struct tsq_tasklet *tsq;
814 /* Keep a ref on socket.
815 * This last ref will be released in tcp_tasklet_func()
817 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
819 /* queue this socket to tasklet queue */
820 local_irq_save(flags);
821 tsq = &__get_cpu_var(tsq_tasklet);
822 list_add(&tp->tsq_node, &tsq->head);
823 tasklet_schedule(&tsq->tasklet);
824 local_irq_restore(flags);
830 /* This routine actually transmits TCP packets queued in by
831 * tcp_do_sendmsg(). This is used by both the initial
832 * transmission and possible later retransmissions.
833 * All SKB's seen here are completely headerless. It is our
834 * job to build the TCP header, and pass the packet down to
835 * IP so it can do the same plus pass the packet off to the
838 * We are working here with either a clone of the original
839 * SKB, or a fresh unique copy made by the retransmit engine.
841 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
844 const struct inet_connection_sock *icsk = inet_csk(sk);
845 struct inet_sock *inet;
847 struct tcp_skb_cb *tcb;
848 struct tcp_out_options opts;
849 unsigned int tcp_options_size, tcp_header_size;
850 struct tcp_md5sig_key *md5;
854 BUG_ON(!skb || !tcp_skb_pcount(skb));
857 const struct sk_buff *fclone = skb + 1;
859 /* If congestion control is doing timestamping, we must
860 * take such a timestamp before we potentially clone/copy.
862 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
863 __net_timestamp(skb);
865 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
866 fclone->fclone == SKB_FCLONE_CLONE))
867 NET_INC_STATS(sock_net(sk),
868 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
870 if (unlikely(skb_cloned(skb)))
871 skb = pskb_copy(skb, gfp_mask);
873 skb = skb_clone(skb, gfp_mask);
880 tcb = TCP_SKB_CB(skb);
881 memset(&opts, 0, sizeof(opts));
883 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
884 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
886 tcp_options_size = tcp_established_options(sk, skb, &opts,
888 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
890 if (tcp_packets_in_flight(tp) == 0)
891 tcp_ca_event(sk, CA_EVENT_TX_START);
893 /* if no packet is in qdisc/device queue, then allow XPS to select
896 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
898 skb_push(skb, tcp_header_size);
899 skb_reset_transport_header(skb);
903 skb->destructor = tcp_wfree;
904 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
906 /* Build TCP header and checksum it. */
908 th->source = inet->inet_sport;
909 th->dest = inet->inet_dport;
910 th->seq = htonl(tcb->seq);
911 th->ack_seq = htonl(tp->rcv_nxt);
912 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
915 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
916 /* RFC1323: The window in SYN & SYN/ACK segments
919 th->window = htons(min(tp->rcv_wnd, 65535U));
921 th->window = htons(tcp_select_window(sk));
926 /* The urg_mode check is necessary during a below snd_una win probe */
927 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
928 if (before(tp->snd_up, tcb->seq + 0x10000)) {
929 th->urg_ptr = htons(tp->snd_up - tcb->seq);
931 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
932 th->urg_ptr = htons(0xFFFF);
937 tcp_options_write((__be32 *)(th + 1), tp, &opts);
938 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
939 TCP_ECN_send(sk, skb, tcp_header_size);
941 #ifdef CONFIG_TCP_MD5SIG
942 /* Calculate the MD5 hash, as we have all we need now */
944 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
945 tp->af_specific->calc_md5_hash(opts.hash_location,
950 icsk->icsk_af_ops->send_check(sk, skb);
952 if (likely(tcb->tcp_flags & TCPHDR_ACK))
953 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
955 if (skb->len != tcp_header_size)
956 tcp_event_data_sent(tp, sk);
958 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
959 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
960 tcp_skb_pcount(skb));
962 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
963 if (likely(err <= 0))
966 tcp_enter_cwr(sk, 1);
968 return net_xmit_eval(err);
971 /* This routine just queues the buffer for sending.
973 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
974 * otherwise socket can stall.
976 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
978 struct tcp_sock *tp = tcp_sk(sk);
980 /* Advance write_seq and place onto the write_queue. */
981 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
982 skb_header_release(skb);
983 tcp_add_write_queue_tail(sk, skb);
984 sk->sk_wmem_queued += skb->truesize;
985 sk_mem_charge(sk, skb->truesize);
988 /* Initialize TSO segments for a packet. */
989 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
990 unsigned int mss_now)
992 struct skb_shared_info *shinfo = skb_shinfo(skb);
994 /* Make sure we own this skb before messing gso_size/gso_segs */
995 WARN_ON_ONCE(skb_cloned(skb));
997 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
998 /* Avoid the costly divide in the normal
1001 shinfo->gso_segs = 1;
1002 shinfo->gso_size = 0;
1003 shinfo->gso_type = 0;
1005 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1006 shinfo->gso_size = mss_now;
1007 shinfo->gso_type = sk->sk_gso_type;
1011 /* When a modification to fackets out becomes necessary, we need to check
1012 * skb is counted to fackets_out or not.
1014 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1017 struct tcp_sock *tp = tcp_sk(sk);
1019 if (!tp->sacked_out || tcp_is_reno(tp))
1022 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1023 tp->fackets_out -= decr;
1026 /* Pcount in the middle of the write queue got changed, we need to do various
1027 * tweaks to fix counters
1029 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1031 struct tcp_sock *tp = tcp_sk(sk);
1033 tp->packets_out -= decr;
1035 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1036 tp->sacked_out -= decr;
1037 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1038 tp->retrans_out -= decr;
1039 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1040 tp->lost_out -= decr;
1042 /* Reno case is special. Sigh... */
1043 if (tcp_is_reno(tp) && decr > 0)
1044 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1046 tcp_adjust_fackets_out(sk, skb, decr);
1048 if (tp->lost_skb_hint &&
1049 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1050 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1051 tp->lost_cnt_hint -= decr;
1053 tcp_verify_left_out(tp);
1056 /* Function to create two new TCP segments. Shrinks the given segment
1057 * to the specified size and appends a new segment with the rest of the
1058 * packet to the list. This won't be called frequently, I hope.
1059 * Remember, these are still headerless SKBs at this point.
1061 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1062 unsigned int mss_now)
1064 struct tcp_sock *tp = tcp_sk(sk);
1065 struct sk_buff *buff;
1066 int nsize, old_factor;
1070 if (WARN_ON(len > skb->len))
1073 nsize = skb_headlen(skb) - len;
1077 if (skb_unclone(skb, GFP_ATOMIC))
1080 /* Get a new skb... force flag on. */
1081 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1083 return -ENOMEM; /* We'll just try again later. */
1085 sk->sk_wmem_queued += buff->truesize;
1086 sk_mem_charge(sk, buff->truesize);
1087 nlen = skb->len - len - nsize;
1088 buff->truesize += nlen;
1089 skb->truesize -= nlen;
1091 /* Correct the sequence numbers. */
1092 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1093 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1094 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1096 /* PSH and FIN should only be set in the second packet. */
1097 flags = TCP_SKB_CB(skb)->tcp_flags;
1098 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1099 TCP_SKB_CB(buff)->tcp_flags = flags;
1100 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1102 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1103 /* Copy and checksum data tail into the new buffer. */
1104 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1105 skb_put(buff, nsize),
1110 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1112 skb->ip_summed = CHECKSUM_PARTIAL;
1113 skb_split(skb, buff, len);
1116 buff->ip_summed = skb->ip_summed;
1118 /* Looks stupid, but our code really uses when of
1119 * skbs, which it never sent before. --ANK
1121 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1122 buff->tstamp = skb->tstamp;
1124 old_factor = tcp_skb_pcount(skb);
1126 /* Fix up tso_factor for both original and new SKB. */
1127 tcp_set_skb_tso_segs(sk, skb, mss_now);
1128 tcp_set_skb_tso_segs(sk, buff, mss_now);
1130 /* If this packet has been sent out already, we must
1131 * adjust the various packet counters.
1133 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1134 int diff = old_factor - tcp_skb_pcount(skb) -
1135 tcp_skb_pcount(buff);
1138 tcp_adjust_pcount(sk, skb, diff);
1141 /* Link BUFF into the send queue. */
1142 skb_header_release(buff);
1143 tcp_insert_write_queue_after(skb, buff, sk);
1148 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1149 * eventually). The difference is that pulled data not copied, but
1150 * immediately discarded.
1152 static void __pskb_trim_head(struct sk_buff *skb, int len)
1154 struct skb_shared_info *shinfo;
1157 eat = min_t(int, len, skb_headlen(skb));
1159 __skb_pull(skb, eat);
1166 shinfo = skb_shinfo(skb);
1167 for (i = 0; i < shinfo->nr_frags; i++) {
1168 int size = skb_frag_size(&shinfo->frags[i]);
1171 skb_frag_unref(skb, i);
1174 shinfo->frags[k] = shinfo->frags[i];
1176 shinfo->frags[k].page_offset += eat;
1177 skb_frag_size_sub(&shinfo->frags[k], eat);
1183 shinfo->nr_frags = k;
1185 skb_reset_tail_pointer(skb);
1186 skb->data_len -= len;
1187 skb->len = skb->data_len;
1190 /* Remove acked data from a packet in the transmit queue. */
1191 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1193 if (skb_unclone(skb, GFP_ATOMIC))
1196 __pskb_trim_head(skb, len);
1198 TCP_SKB_CB(skb)->seq += len;
1199 skb->ip_summed = CHECKSUM_PARTIAL;
1201 skb->truesize -= len;
1202 sk->sk_wmem_queued -= len;
1203 sk_mem_uncharge(sk, len);
1204 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1206 /* Any change of skb->len requires recalculation of tso factor. */
1207 if (tcp_skb_pcount(skb) > 1)
1208 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1213 /* Calculate MSS not accounting any TCP options. */
1214 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1216 const struct tcp_sock *tp = tcp_sk(sk);
1217 const struct inet_connection_sock *icsk = inet_csk(sk);
1220 /* Calculate base mss without TCP options:
1221 It is MMS_S - sizeof(tcphdr) of rfc1122
1223 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1225 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1226 if (icsk->icsk_af_ops->net_frag_header_len) {
1227 const struct dst_entry *dst = __sk_dst_get(sk);
1229 if (dst && dst_allfrag(dst))
1230 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1233 /* Clamp it (mss_clamp does not include tcp options) */
1234 if (mss_now > tp->rx_opt.mss_clamp)
1235 mss_now = tp->rx_opt.mss_clamp;
1237 /* Now subtract optional transport overhead */
1238 mss_now -= icsk->icsk_ext_hdr_len;
1240 /* Then reserve room for full set of TCP options and 8 bytes of data */
1246 /* Calculate MSS. Not accounting for SACKs here. */
1247 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1249 /* Subtract TCP options size, not including SACKs */
1250 return __tcp_mtu_to_mss(sk, pmtu) -
1251 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1254 /* Inverse of above */
1255 int tcp_mss_to_mtu(struct sock *sk, int mss)
1257 const struct tcp_sock *tp = tcp_sk(sk);
1258 const struct inet_connection_sock *icsk = inet_csk(sk);
1262 tp->tcp_header_len +
1263 icsk->icsk_ext_hdr_len +
1264 icsk->icsk_af_ops->net_header_len;
1266 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1267 if (icsk->icsk_af_ops->net_frag_header_len) {
1268 const struct dst_entry *dst = __sk_dst_get(sk);
1270 if (dst && dst_allfrag(dst))
1271 mtu += icsk->icsk_af_ops->net_frag_header_len;
1276 /* MTU probing init per socket */
1277 void tcp_mtup_init(struct sock *sk)
1279 struct tcp_sock *tp = tcp_sk(sk);
1280 struct inet_connection_sock *icsk = inet_csk(sk);
1282 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1283 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1284 icsk->icsk_af_ops->net_header_len;
1285 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1286 icsk->icsk_mtup.probe_size = 0;
1288 EXPORT_SYMBOL(tcp_mtup_init);
1290 /* This function synchronize snd mss to current pmtu/exthdr set.
1292 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1293 for TCP options, but includes only bare TCP header.
1295 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1296 It is minimum of user_mss and mss received with SYN.
1297 It also does not include TCP options.
1299 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1301 tp->mss_cache is current effective sending mss, including
1302 all tcp options except for SACKs. It is evaluated,
1303 taking into account current pmtu, but never exceeds
1304 tp->rx_opt.mss_clamp.
1306 NOTE1. rfc1122 clearly states that advertised MSS
1307 DOES NOT include either tcp or ip options.
1309 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1310 are READ ONLY outside this function. --ANK (980731)
1312 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1314 struct tcp_sock *tp = tcp_sk(sk);
1315 struct inet_connection_sock *icsk = inet_csk(sk);
1318 if (icsk->icsk_mtup.search_high > pmtu)
1319 icsk->icsk_mtup.search_high = pmtu;
1321 mss_now = tcp_mtu_to_mss(sk, pmtu);
1322 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1324 /* And store cached results */
1325 icsk->icsk_pmtu_cookie = pmtu;
1326 if (icsk->icsk_mtup.enabled)
1327 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1328 tp->mss_cache = mss_now;
1332 EXPORT_SYMBOL(tcp_sync_mss);
1334 /* Compute the current effective MSS, taking SACKs and IP options,
1335 * and even PMTU discovery events into account.
1337 unsigned int tcp_current_mss(struct sock *sk)
1339 const struct tcp_sock *tp = tcp_sk(sk);
1340 const struct dst_entry *dst = __sk_dst_get(sk);
1342 unsigned int header_len;
1343 struct tcp_out_options opts;
1344 struct tcp_md5sig_key *md5;
1346 mss_now = tp->mss_cache;
1349 u32 mtu = dst_mtu(dst);
1350 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1351 mss_now = tcp_sync_mss(sk, mtu);
1354 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1355 sizeof(struct tcphdr);
1356 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1357 * some common options. If this is an odd packet (because we have SACK
1358 * blocks etc) then our calculated header_len will be different, and
1359 * we have to adjust mss_now correspondingly */
1360 if (header_len != tp->tcp_header_len) {
1361 int delta = (int) header_len - tp->tcp_header_len;
1368 /* Congestion window validation. (RFC2861) */
1369 static void tcp_cwnd_validate(struct sock *sk)
1371 struct tcp_sock *tp = tcp_sk(sk);
1373 if (tp->packets_out >= tp->snd_cwnd) {
1374 /* Network is feed fully. */
1375 tp->snd_cwnd_used = 0;
1376 tp->snd_cwnd_stamp = tcp_time_stamp;
1378 /* Network starves. */
1379 if (tp->packets_out > tp->snd_cwnd_used)
1380 tp->snd_cwnd_used = tp->packets_out;
1382 if (sysctl_tcp_slow_start_after_idle &&
1383 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1384 tcp_cwnd_application_limited(sk);
1388 /* Minshall's variant of the Nagle send check. */
1389 static bool tcp_minshall_check(const struct tcp_sock *tp)
1391 return after(tp->snd_sml, tp->snd_una) &&
1392 !after(tp->snd_sml, tp->snd_nxt);
1395 /* Update snd_sml if this skb is under mss
1396 * Note that a TSO packet might end with a sub-mss segment
1397 * The test is really :
1398 * if ((skb->len % mss) != 0)
1399 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1400 * But we can avoid doing the divide again given we already have
1401 * skb_pcount = skb->len / mss_now
1403 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1404 const struct sk_buff *skb)
1406 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1407 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1410 /* Return false, if packet can be sent now without violation Nagle's rules:
1411 * 1. It is full sized. (provided by caller in %partial bool)
1412 * 2. Or it contains FIN. (already checked by caller)
1413 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1414 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1415 * With Minshall's modification: all sent small packets are ACKed.
1417 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1418 unsigned int mss_now, int nonagle)
1421 ((nonagle & TCP_NAGLE_CORK) ||
1422 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1424 /* Returns the portion of skb which can be sent right away */
1425 static unsigned int tcp_mss_split_point(const struct sock *sk,
1426 const struct sk_buff *skb,
1427 unsigned int mss_now,
1428 unsigned int max_segs,
1431 const struct tcp_sock *tp = tcp_sk(sk);
1432 u32 partial, needed, window, max_len;
1434 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1435 max_len = mss_now * max_segs;
1437 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1440 needed = min(skb->len, window);
1442 if (max_len <= needed)
1445 partial = needed % mss_now;
1446 /* If last segment is not a full MSS, check if Nagle rules allow us
1447 * to include this last segment in this skb.
1448 * Otherwise, we'll split the skb at last MSS boundary
1450 if (tcp_nagle_check(partial != 0, tp, mss_now, nonagle))
1451 return needed - partial;
1456 /* Can at least one segment of SKB be sent right now, according to the
1457 * congestion window rules? If so, return how many segments are allowed.
1459 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1460 const struct sk_buff *skb)
1462 u32 in_flight, cwnd;
1464 /* Don't be strict about the congestion window for the final FIN. */
1465 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1466 tcp_skb_pcount(skb) == 1)
1469 in_flight = tcp_packets_in_flight(tp);
1470 cwnd = tp->snd_cwnd;
1471 if (in_flight < cwnd)
1472 return (cwnd - in_flight);
1477 /* Initialize TSO state of a skb.
1478 * This must be invoked the first time we consider transmitting
1479 * SKB onto the wire.
1481 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1482 unsigned int mss_now)
1484 int tso_segs = tcp_skb_pcount(skb);
1486 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1487 tcp_set_skb_tso_segs(sk, skb, mss_now);
1488 tso_segs = tcp_skb_pcount(skb);
1494 /* Return true if the Nagle test allows this packet to be
1497 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1498 unsigned int cur_mss, int nonagle)
1500 /* Nagle rule does not apply to frames, which sit in the middle of the
1501 * write_queue (they have no chances to get new data).
1503 * This is implemented in the callers, where they modify the 'nonagle'
1504 * argument based upon the location of SKB in the send queue.
1506 if (nonagle & TCP_NAGLE_PUSH)
1509 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1510 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1513 if (!tcp_nagle_check(skb->len < cur_mss, tp, cur_mss, nonagle))
1519 /* Does at least the first segment of SKB fit into the send window? */
1520 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1521 const struct sk_buff *skb,
1522 unsigned int cur_mss)
1524 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1526 if (skb->len > cur_mss)
1527 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1529 return !after(end_seq, tcp_wnd_end(tp));
1532 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1533 * should be put on the wire right now. If so, it returns the number of
1534 * packets allowed by the congestion window.
1536 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1537 unsigned int cur_mss, int nonagle)
1539 const struct tcp_sock *tp = tcp_sk(sk);
1540 unsigned int cwnd_quota;
1542 tcp_init_tso_segs(sk, skb, cur_mss);
1544 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1547 cwnd_quota = tcp_cwnd_test(tp, skb);
1548 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1554 /* Test if sending is allowed right now. */
1555 bool tcp_may_send_now(struct sock *sk)
1557 const struct tcp_sock *tp = tcp_sk(sk);
1558 struct sk_buff *skb = tcp_send_head(sk);
1561 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1562 (tcp_skb_is_last(sk, skb) ?
1563 tp->nonagle : TCP_NAGLE_PUSH));
1566 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1567 * which is put after SKB on the list. It is very much like
1568 * tcp_fragment() except that it may make several kinds of assumptions
1569 * in order to speed up the splitting operation. In particular, we
1570 * know that all the data is in scatter-gather pages, and that the
1571 * packet has never been sent out before (and thus is not cloned).
1573 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1574 unsigned int mss_now, gfp_t gfp)
1576 struct sk_buff *buff;
1577 int nlen = skb->len - len;
1580 /* All of a TSO frame must be composed of paged data. */
1581 if (skb->len != skb->data_len)
1582 return tcp_fragment(sk, skb, len, mss_now);
1584 buff = sk_stream_alloc_skb(sk, 0, gfp);
1585 if (unlikely(buff == NULL))
1588 sk->sk_wmem_queued += buff->truesize;
1589 sk_mem_charge(sk, buff->truesize);
1590 buff->truesize += nlen;
1591 skb->truesize -= nlen;
1593 /* Correct the sequence numbers. */
1594 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1595 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1596 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1598 /* PSH and FIN should only be set in the second packet. */
1599 flags = TCP_SKB_CB(skb)->tcp_flags;
1600 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1601 TCP_SKB_CB(buff)->tcp_flags = flags;
1603 /* This packet was never sent out yet, so no SACK bits. */
1604 TCP_SKB_CB(buff)->sacked = 0;
1606 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1607 skb_split(skb, buff, len);
1609 /* Fix up tso_factor for both original and new SKB. */
1610 tcp_set_skb_tso_segs(sk, skb, mss_now);
1611 tcp_set_skb_tso_segs(sk, buff, mss_now);
1613 /* Link BUFF into the send queue. */
1614 skb_header_release(buff);
1615 tcp_insert_write_queue_after(skb, buff, sk);
1620 /* Try to defer sending, if possible, in order to minimize the amount
1621 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1623 * This algorithm is from John Heffner.
1625 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1627 struct tcp_sock *tp = tcp_sk(sk);
1628 const struct inet_connection_sock *icsk = inet_csk(sk);
1629 u32 send_win, cong_win, limit, in_flight;
1632 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1635 if (icsk->icsk_ca_state != TCP_CA_Open)
1638 /* Defer for less than two clock ticks. */
1639 if (tp->tso_deferred &&
1640 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1643 in_flight = tcp_packets_in_flight(tp);
1645 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1647 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1649 /* From in_flight test above, we know that cwnd > in_flight. */
1650 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1652 limit = min(send_win, cong_win);
1654 /* If a full-sized TSO skb can be sent, do it. */
1655 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1656 tp->xmit_size_goal_segs * tp->mss_cache))
1659 /* Middle in queue won't get any more data, full sendable already? */
1660 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1663 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1665 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1667 /* If at least some fraction of a window is available,
1670 chunk /= win_divisor;
1674 /* Different approach, try not to defer past a single
1675 * ACK. Receiver should ACK every other full sized
1676 * frame, so if we have space for more than 3 frames
1679 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1683 /* Ok, it looks like it is advisable to defer.
1684 * Do not rearm the timer if already set to not break TCP ACK clocking.
1686 if (!tp->tso_deferred)
1687 tp->tso_deferred = 1 | (jiffies << 1);
1692 tp->tso_deferred = 0;
1696 /* Create a new MTU probe if we are ready.
1697 * MTU probe is regularly attempting to increase the path MTU by
1698 * deliberately sending larger packets. This discovers routing
1699 * changes resulting in larger path MTUs.
1701 * Returns 0 if we should wait to probe (no cwnd available),
1702 * 1 if a probe was sent,
1705 static int tcp_mtu_probe(struct sock *sk)
1707 struct tcp_sock *tp = tcp_sk(sk);
1708 struct inet_connection_sock *icsk = inet_csk(sk);
1709 struct sk_buff *skb, *nskb, *next;
1716 /* Not currently probing/verifying,
1718 * have enough cwnd, and
1719 * not SACKing (the variable headers throw things off) */
1720 if (!icsk->icsk_mtup.enabled ||
1721 icsk->icsk_mtup.probe_size ||
1722 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1723 tp->snd_cwnd < 11 ||
1724 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1727 /* Very simple search strategy: just double the MSS. */
1728 mss_now = tcp_current_mss(sk);
1729 probe_size = 2 * tp->mss_cache;
1730 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1731 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1732 /* TODO: set timer for probe_converge_event */
1736 /* Have enough data in the send queue to probe? */
1737 if (tp->write_seq - tp->snd_nxt < size_needed)
1740 if (tp->snd_wnd < size_needed)
1742 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1745 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1746 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1747 if (!tcp_packets_in_flight(tp))
1753 /* We're allowed to probe. Build it now. */
1754 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1756 sk->sk_wmem_queued += nskb->truesize;
1757 sk_mem_charge(sk, nskb->truesize);
1759 skb = tcp_send_head(sk);
1761 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1762 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1763 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1764 TCP_SKB_CB(nskb)->sacked = 0;
1766 nskb->ip_summed = skb->ip_summed;
1768 tcp_insert_write_queue_before(nskb, skb, sk);
1771 tcp_for_write_queue_from_safe(skb, next, sk) {
1772 copy = min_t(int, skb->len, probe_size - len);
1773 if (nskb->ip_summed)
1774 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1776 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1777 skb_put(nskb, copy),
1780 if (skb->len <= copy) {
1781 /* We've eaten all the data from this skb.
1783 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1784 tcp_unlink_write_queue(skb, sk);
1785 sk_wmem_free_skb(sk, skb);
1787 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1788 ~(TCPHDR_FIN|TCPHDR_PSH);
1789 if (!skb_shinfo(skb)->nr_frags) {
1790 skb_pull(skb, copy);
1791 if (skb->ip_summed != CHECKSUM_PARTIAL)
1792 skb->csum = csum_partial(skb->data,
1795 __pskb_trim_head(skb, copy);
1796 tcp_set_skb_tso_segs(sk, skb, mss_now);
1798 TCP_SKB_CB(skb)->seq += copy;
1803 if (len >= probe_size)
1806 tcp_init_tso_segs(sk, nskb, nskb->len);
1808 /* We're ready to send. If this fails, the probe will
1809 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1810 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1811 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1812 /* Decrement cwnd here because we are sending
1813 * effectively two packets. */
1815 tcp_event_new_data_sent(sk, nskb);
1817 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1818 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1819 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1827 /* This routine writes packets to the network. It advances the
1828 * send_head. This happens as incoming acks open up the remote
1831 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1832 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1833 * account rare use of URG, this is not a big flaw.
1835 * Send at most one packet when push_one > 0. Temporarily ignore
1836 * cwnd limit to force at most one packet out when push_one == 2.
1838 * Returns true, if no segments are in flight and we have queued segments,
1839 * but cannot send anything now because of SWS or another problem.
1841 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1842 int push_one, gfp_t gfp)
1844 struct tcp_sock *tp = tcp_sk(sk);
1845 struct sk_buff *skb;
1846 unsigned int tso_segs, sent_pkts;
1853 /* Do MTU probing. */
1854 result = tcp_mtu_probe(sk);
1857 } else if (result > 0) {
1862 while ((skb = tcp_send_head(sk))) {
1865 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1868 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1869 goto repair; /* Skip network transmission */
1871 cwnd_quota = tcp_cwnd_test(tp, skb);
1874 /* Force out a loss probe pkt. */
1880 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1883 if (tso_segs == 1) {
1884 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1885 (tcp_skb_is_last(sk, skb) ?
1886 nonagle : TCP_NAGLE_PUSH))))
1889 if (!push_one && tcp_tso_should_defer(sk, skb))
1893 /* TCP Small Queues :
1894 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1896 * - better RTT estimation and ACK scheduling
1899 * Alas, some drivers / subsystems require a fair amount
1900 * of queued bytes to ensure line rate.
1901 * One example is wifi aggregation (802.11 AMPDU)
1903 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1904 sk->sk_pacing_rate >> 10);
1906 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1907 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1908 /* It is possible TX completion already happened
1909 * before we set TSQ_THROTTLED, so we must
1910 * test again the condition.
1911 * We abuse smp_mb__after_clear_bit() because
1912 * there is no smp_mb__after_set_bit() yet
1914 smp_mb__after_clear_bit();
1915 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1920 if (tso_segs > 1 && !tcp_urg_mode(tp))
1921 limit = tcp_mss_split_point(sk, skb, mss_now,
1924 sk->sk_gso_max_segs),
1927 if (skb->len > limit &&
1928 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1931 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1933 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1937 /* Advance the send_head. This one is sent out.
1938 * This call will increment packets_out.
1940 tcp_event_new_data_sent(sk, skb);
1942 tcp_minshall_update(tp, mss_now, skb);
1943 sent_pkts += tcp_skb_pcount(skb);
1949 if (likely(sent_pkts)) {
1950 if (tcp_in_cwnd_reduction(sk))
1951 tp->prr_out += sent_pkts;
1953 /* Send one loss probe per tail loss episode. */
1955 tcp_schedule_loss_probe(sk);
1956 tcp_cwnd_validate(sk);
1959 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1962 bool tcp_schedule_loss_probe(struct sock *sk)
1964 struct inet_connection_sock *icsk = inet_csk(sk);
1965 struct tcp_sock *tp = tcp_sk(sk);
1966 u32 timeout, tlp_time_stamp, rto_time_stamp;
1967 u32 rtt = tp->srtt >> 3;
1969 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1971 /* No consecutive loss probes. */
1972 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1976 /* Don't do any loss probe on a Fast Open connection before 3WHS
1979 if (sk->sk_state == TCP_SYN_RECV)
1982 /* TLP is only scheduled when next timer event is RTO. */
1983 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1986 /* Schedule a loss probe in 2*RTT for SACK capable connections
1987 * in Open state, that are either limited by cwnd or application.
1989 if (sysctl_tcp_early_retrans < 3 || !tp->srtt || !tp->packets_out ||
1990 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1993 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1997 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1998 * for delayed ack when there's one outstanding packet.
2001 if (tp->packets_out == 1)
2002 timeout = max_t(u32, timeout,
2003 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2004 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2006 /* If RTO is shorter, just schedule TLP in its place. */
2007 tlp_time_stamp = tcp_time_stamp + timeout;
2008 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2009 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2010 s32 delta = rto_time_stamp - tcp_time_stamp;
2015 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2020 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2021 * retransmit the last segment.
2023 void tcp_send_loss_probe(struct sock *sk)
2025 struct tcp_sock *tp = tcp_sk(sk);
2026 struct sk_buff *skb;
2028 int mss = tcp_current_mss(sk);
2031 if (tcp_send_head(sk) != NULL) {
2032 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2036 /* At most one outstanding TLP retransmission. */
2037 if (tp->tlp_high_seq)
2040 /* Retransmit last segment. */
2041 skb = tcp_write_queue_tail(sk);
2045 pcount = tcp_skb_pcount(skb);
2046 if (WARN_ON(!pcount))
2049 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2050 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2052 skb = tcp_write_queue_tail(sk);
2055 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2058 /* Probe with zero data doesn't trigger fast recovery. */
2060 err = __tcp_retransmit_skb(sk, skb);
2062 /* Record snd_nxt for loss detection. */
2064 tp->tlp_high_seq = tp->snd_nxt;
2067 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2068 inet_csk(sk)->icsk_rto,
2072 NET_INC_STATS_BH(sock_net(sk),
2073 LINUX_MIB_TCPLOSSPROBES);
2077 /* Push out any pending frames which were held back due to
2078 * TCP_CORK or attempt at coalescing tiny packets.
2079 * The socket must be locked by the caller.
2081 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2084 /* If we are closed, the bytes will have to remain here.
2085 * In time closedown will finish, we empty the write queue and
2086 * all will be happy.
2088 if (unlikely(sk->sk_state == TCP_CLOSE))
2091 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2092 sk_gfp_atomic(sk, GFP_ATOMIC)))
2093 tcp_check_probe_timer(sk);
2096 /* Send _single_ skb sitting at the send head. This function requires
2097 * true push pending frames to setup probe timer etc.
2099 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2101 struct sk_buff *skb = tcp_send_head(sk);
2103 BUG_ON(!skb || skb->len < mss_now);
2105 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2108 /* This function returns the amount that we can raise the
2109 * usable window based on the following constraints
2111 * 1. The window can never be shrunk once it is offered (RFC 793)
2112 * 2. We limit memory per socket
2115 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2116 * RECV.NEXT + RCV.WIN fixed until:
2117 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2119 * i.e. don't raise the right edge of the window until you can raise
2120 * it at least MSS bytes.
2122 * Unfortunately, the recommended algorithm breaks header prediction,
2123 * since header prediction assumes th->window stays fixed.
2125 * Strictly speaking, keeping th->window fixed violates the receiver
2126 * side SWS prevention criteria. The problem is that under this rule
2127 * a stream of single byte packets will cause the right side of the
2128 * window to always advance by a single byte.
2130 * Of course, if the sender implements sender side SWS prevention
2131 * then this will not be a problem.
2133 * BSD seems to make the following compromise:
2135 * If the free space is less than the 1/4 of the maximum
2136 * space available and the free space is less than 1/2 mss,
2137 * then set the window to 0.
2138 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2139 * Otherwise, just prevent the window from shrinking
2140 * and from being larger than the largest representable value.
2142 * This prevents incremental opening of the window in the regime
2143 * where TCP is limited by the speed of the reader side taking
2144 * data out of the TCP receive queue. It does nothing about
2145 * those cases where the window is constrained on the sender side
2146 * because the pipeline is full.
2148 * BSD also seems to "accidentally" limit itself to windows that are a
2149 * multiple of MSS, at least until the free space gets quite small.
2150 * This would appear to be a side effect of the mbuf implementation.
2151 * Combining these two algorithms results in the observed behavior
2152 * of having a fixed window size at almost all times.
2154 * Below we obtain similar behavior by forcing the offered window to
2155 * a multiple of the mss when it is feasible to do so.
2157 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2158 * Regular options like TIMESTAMP are taken into account.
2160 u32 __tcp_select_window(struct sock *sk)
2162 struct inet_connection_sock *icsk = inet_csk(sk);
2163 struct tcp_sock *tp = tcp_sk(sk);
2164 /* MSS for the peer's data. Previous versions used mss_clamp
2165 * here. I don't know if the value based on our guesses
2166 * of peer's MSS is better for the performance. It's more correct
2167 * but may be worse for the performance because of rcv_mss
2168 * fluctuations. --SAW 1998/11/1
2170 int mss = icsk->icsk_ack.rcv_mss;
2171 int free_space = tcp_space(sk);
2172 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2175 if (mss > full_space)
2178 if (free_space < (full_space >> 1)) {
2179 icsk->icsk_ack.quick = 0;
2181 if (sk_under_memory_pressure(sk))
2182 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2185 if (free_space < mss)
2189 if (free_space > tp->rcv_ssthresh)
2190 free_space = tp->rcv_ssthresh;
2192 /* Don't do rounding if we are using window scaling, since the
2193 * scaled window will not line up with the MSS boundary anyway.
2195 window = tp->rcv_wnd;
2196 if (tp->rx_opt.rcv_wscale) {
2197 window = free_space;
2199 /* Advertise enough space so that it won't get scaled away.
2200 * Import case: prevent zero window announcement if
2201 * 1<<rcv_wscale > mss.
2203 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2204 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2205 << tp->rx_opt.rcv_wscale);
2207 /* Get the largest window that is a nice multiple of mss.
2208 * Window clamp already applied above.
2209 * If our current window offering is within 1 mss of the
2210 * free space we just keep it. This prevents the divide
2211 * and multiply from happening most of the time.
2212 * We also don't do any window rounding when the free space
2215 if (window <= free_space - mss || window > free_space)
2216 window = (free_space / mss) * mss;
2217 else if (mss == full_space &&
2218 free_space > window + (full_space >> 1))
2219 window = free_space;
2225 /* Collapses two adjacent SKB's during retransmission. */
2226 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2228 struct tcp_sock *tp = tcp_sk(sk);
2229 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2230 int skb_size, next_skb_size;
2232 skb_size = skb->len;
2233 next_skb_size = next_skb->len;
2235 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2237 tcp_highest_sack_combine(sk, next_skb, skb);
2239 tcp_unlink_write_queue(next_skb, sk);
2241 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2244 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2245 skb->ip_summed = CHECKSUM_PARTIAL;
2247 if (skb->ip_summed != CHECKSUM_PARTIAL)
2248 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2250 /* Update sequence range on original skb. */
2251 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2253 /* Merge over control information. This moves PSH/FIN etc. over */
2254 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2256 /* All done, get rid of second SKB and account for it so
2257 * packet counting does not break.
2259 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2261 /* changed transmit queue under us so clear hints */
2262 tcp_clear_retrans_hints_partial(tp);
2263 if (next_skb == tp->retransmit_skb_hint)
2264 tp->retransmit_skb_hint = skb;
2266 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2268 sk_wmem_free_skb(sk, next_skb);
2271 /* Check if coalescing SKBs is legal. */
2272 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2274 if (tcp_skb_pcount(skb) > 1)
2276 /* TODO: SACK collapsing could be used to remove this condition */
2277 if (skb_shinfo(skb)->nr_frags != 0)
2279 if (skb_cloned(skb))
2281 if (skb == tcp_send_head(sk))
2283 /* Some heurestics for collapsing over SACK'd could be invented */
2284 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2290 /* Collapse packets in the retransmit queue to make to create
2291 * less packets on the wire. This is only done on retransmission.
2293 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2296 struct tcp_sock *tp = tcp_sk(sk);
2297 struct sk_buff *skb = to, *tmp;
2300 if (!sysctl_tcp_retrans_collapse)
2302 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2305 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2306 if (!tcp_can_collapse(sk, skb))
2318 /* Punt if not enough space exists in the first SKB for
2319 * the data in the second
2321 if (skb->len > skb_availroom(to))
2324 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2327 tcp_collapse_retrans(sk, to);
2331 /* This retransmits one SKB. Policy decisions and retransmit queue
2332 * state updates are done by the caller. Returns non-zero if an
2333 * error occurred which prevented the send.
2335 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2337 struct tcp_sock *tp = tcp_sk(sk);
2338 struct inet_connection_sock *icsk = inet_csk(sk);
2339 unsigned int cur_mss;
2341 /* Inconslusive MTU probe */
2342 if (icsk->icsk_mtup.probe_size) {
2343 icsk->icsk_mtup.probe_size = 0;
2346 /* Do not sent more than we queued. 1/4 is reserved for possible
2347 * copying overhead: fragmentation, tunneling, mangling etc.
2349 if (atomic_read(&sk->sk_wmem_alloc) >
2350 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2353 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2354 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2356 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2360 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2361 return -EHOSTUNREACH; /* Routing failure or similar. */
2363 cur_mss = tcp_current_mss(sk);
2365 /* If receiver has shrunk his window, and skb is out of
2366 * new window, do not retransmit it. The exception is the
2367 * case, when window is shrunk to zero. In this case
2368 * our retransmit serves as a zero window probe.
2370 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2371 TCP_SKB_CB(skb)->seq != tp->snd_una)
2374 if (skb->len > cur_mss) {
2375 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2376 return -ENOMEM; /* We'll try again later. */
2378 int oldpcount = tcp_skb_pcount(skb);
2380 if (unlikely(oldpcount > 1)) {
2381 if (skb_unclone(skb, GFP_ATOMIC))
2383 tcp_init_tso_segs(sk, skb, cur_mss);
2384 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2388 tcp_retrans_try_collapse(sk, skb, cur_mss);
2390 /* Make a copy, if the first transmission SKB clone we made
2391 * is still in somebody's hands, else make a clone.
2393 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2395 /* make sure skb->data is aligned on arches that require it
2396 * and check if ack-trimming & collapsing extended the headroom
2397 * beyond what csum_start can cover.
2399 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2400 skb_headroom(skb) >= 0xFFFF)) {
2401 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2403 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2406 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2410 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2412 struct tcp_sock *tp = tcp_sk(sk);
2413 int err = __tcp_retransmit_skb(sk, skb);
2416 /* Update global TCP statistics. */
2417 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2419 tp->total_retrans++;
2421 #if FASTRETRANS_DEBUG > 0
2422 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2423 net_dbg_ratelimited("retrans_out leaked\n");
2426 if (!tp->retrans_out)
2427 tp->lost_retrans_low = tp->snd_nxt;
2428 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2429 tp->retrans_out += tcp_skb_pcount(skb);
2431 /* Save stamp of the first retransmit. */
2432 if (!tp->retrans_stamp)
2433 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2435 tp->undo_retrans += tcp_skb_pcount(skb);
2437 /* snd_nxt is stored to detect loss of retransmitted segment,
2438 * see tcp_input.c tcp_sacktag_write_queue().
2440 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2442 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2447 /* Check if we forward retransmits are possible in the current
2448 * window/congestion state.
2450 static bool tcp_can_forward_retransmit(struct sock *sk)
2452 const struct inet_connection_sock *icsk = inet_csk(sk);
2453 const struct tcp_sock *tp = tcp_sk(sk);
2455 /* Forward retransmissions are possible only during Recovery. */
2456 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2459 /* No forward retransmissions in Reno are possible. */
2460 if (tcp_is_reno(tp))
2463 /* Yeah, we have to make difficult choice between forward transmission
2464 * and retransmission... Both ways have their merits...
2466 * For now we do not retransmit anything, while we have some new
2467 * segments to send. In the other cases, follow rule 3 for
2468 * NextSeg() specified in RFC3517.
2471 if (tcp_may_send_now(sk))
2477 /* This gets called after a retransmit timeout, and the initially
2478 * retransmitted data is acknowledged. It tries to continue
2479 * resending the rest of the retransmit queue, until either
2480 * we've sent it all or the congestion window limit is reached.
2481 * If doing SACK, the first ACK which comes back for a timeout
2482 * based retransmit packet might feed us FACK information again.
2483 * If so, we use it to avoid unnecessarily retransmissions.
2485 void tcp_xmit_retransmit_queue(struct sock *sk)
2487 const struct inet_connection_sock *icsk = inet_csk(sk);
2488 struct tcp_sock *tp = tcp_sk(sk);
2489 struct sk_buff *skb;
2490 struct sk_buff *hole = NULL;
2493 int fwd_rexmitting = 0;
2495 if (!tp->packets_out)
2499 tp->retransmit_high = tp->snd_una;
2501 if (tp->retransmit_skb_hint) {
2502 skb = tp->retransmit_skb_hint;
2503 last_lost = TCP_SKB_CB(skb)->end_seq;
2504 if (after(last_lost, tp->retransmit_high))
2505 last_lost = tp->retransmit_high;
2507 skb = tcp_write_queue_head(sk);
2508 last_lost = tp->snd_una;
2511 tcp_for_write_queue_from(skb, sk) {
2512 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2514 if (skb == tcp_send_head(sk))
2516 /* we could do better than to assign each time */
2518 tp->retransmit_skb_hint = skb;
2520 /* Assume this retransmit will generate
2521 * only one packet for congestion window
2522 * calculation purposes. This works because
2523 * tcp_retransmit_skb() will chop up the
2524 * packet to be MSS sized and all the
2525 * packet counting works out.
2527 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2530 if (fwd_rexmitting) {
2532 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2534 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2536 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2537 tp->retransmit_high = last_lost;
2538 if (!tcp_can_forward_retransmit(sk))
2540 /* Backtrack if necessary to non-L'ed skb */
2548 } else if (!(sacked & TCPCB_LOST)) {
2549 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2554 last_lost = TCP_SKB_CB(skb)->end_seq;
2555 if (icsk->icsk_ca_state != TCP_CA_Loss)
2556 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2558 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2561 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2564 if (tcp_retransmit_skb(sk, skb))
2567 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2569 if (tcp_in_cwnd_reduction(sk))
2570 tp->prr_out += tcp_skb_pcount(skb);
2572 if (skb == tcp_write_queue_head(sk))
2573 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2574 inet_csk(sk)->icsk_rto,
2579 /* Send a fin. The caller locks the socket for us. This cannot be
2580 * allowed to fail queueing a FIN frame under any circumstances.
2582 void tcp_send_fin(struct sock *sk)
2584 struct tcp_sock *tp = tcp_sk(sk);
2585 struct sk_buff *skb = tcp_write_queue_tail(sk);
2588 /* Optimization, tack on the FIN if we have a queue of
2589 * unsent frames. But be careful about outgoing SACKS
2592 mss_now = tcp_current_mss(sk);
2594 if (tcp_send_head(sk) != NULL) {
2595 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2596 TCP_SKB_CB(skb)->end_seq++;
2599 /* Socket is locked, keep trying until memory is available. */
2601 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2608 /* Reserve space for headers and prepare control bits. */
2609 skb_reserve(skb, MAX_TCP_HEADER);
2610 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2611 tcp_init_nondata_skb(skb, tp->write_seq,
2612 TCPHDR_ACK | TCPHDR_FIN);
2613 tcp_queue_skb(sk, skb);
2615 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2618 /* We get here when a process closes a file descriptor (either due to
2619 * an explicit close() or as a byproduct of exit()'ing) and there
2620 * was unread data in the receive queue. This behavior is recommended
2621 * by RFC 2525, section 2.17. -DaveM
2623 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2625 struct sk_buff *skb;
2627 /* NOTE: No TCP options attached and we never retransmit this. */
2628 skb = alloc_skb(MAX_TCP_HEADER, priority);
2630 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2634 /* Reserve space for headers and prepare control bits. */
2635 skb_reserve(skb, MAX_TCP_HEADER);
2636 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2637 TCPHDR_ACK | TCPHDR_RST);
2639 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2640 if (tcp_transmit_skb(sk, skb, 0, priority))
2641 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2643 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2646 /* Send a crossed SYN-ACK during socket establishment.
2647 * WARNING: This routine must only be called when we have already sent
2648 * a SYN packet that crossed the incoming SYN that caused this routine
2649 * to get called. If this assumption fails then the initial rcv_wnd
2650 * and rcv_wscale values will not be correct.
2652 int tcp_send_synack(struct sock *sk)
2654 struct sk_buff *skb;
2656 skb = tcp_write_queue_head(sk);
2657 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2658 pr_debug("%s: wrong queue state\n", __func__);
2661 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2662 if (skb_cloned(skb)) {
2663 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2666 tcp_unlink_write_queue(skb, sk);
2667 skb_header_release(nskb);
2668 __tcp_add_write_queue_head(sk, nskb);
2669 sk_wmem_free_skb(sk, skb);
2670 sk->sk_wmem_queued += nskb->truesize;
2671 sk_mem_charge(sk, nskb->truesize);
2675 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2676 TCP_ECN_send_synack(tcp_sk(sk), skb);
2678 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2679 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2683 * tcp_make_synack - Prepare a SYN-ACK.
2684 * sk: listener socket
2685 * dst: dst entry attached to the SYNACK
2686 * req: request_sock pointer
2688 * Allocate one skb and build a SYNACK packet.
2689 * @dst is consumed : Caller should not use it again.
2691 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2692 struct request_sock *req,
2693 struct tcp_fastopen_cookie *foc)
2695 struct tcp_out_options opts;
2696 struct inet_request_sock *ireq = inet_rsk(req);
2697 struct tcp_sock *tp = tcp_sk(sk);
2699 struct sk_buff *skb;
2700 struct tcp_md5sig_key *md5;
2701 int tcp_header_size;
2704 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2705 if (unlikely(!skb)) {
2709 /* Reserve space for headers. */
2710 skb_reserve(skb, MAX_TCP_HEADER);
2712 skb_dst_set(skb, dst);
2713 security_skb_owned_by(skb, sk);
2715 mss = dst_metric_advmss(dst);
2716 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2717 mss = tp->rx_opt.user_mss;
2719 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2721 /* Set this up on the first call only */
2722 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2724 /* limit the window selection if the user enforce a smaller rx buffer */
2725 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2726 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2727 req->window_clamp = tcp_full_space(sk);
2729 /* tcp_full_space because it is guaranteed to be the first packet */
2730 tcp_select_initial_window(tcp_full_space(sk),
2731 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2736 dst_metric(dst, RTAX_INITRWND));
2737 ireq->rcv_wscale = rcv_wscale;
2740 memset(&opts, 0, sizeof(opts));
2741 #ifdef CONFIG_SYN_COOKIES
2742 if (unlikely(req->cookie_ts))
2743 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2746 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2747 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2750 skb_push(skb, tcp_header_size);
2751 skb_reset_transport_header(skb);
2754 memset(th, 0, sizeof(struct tcphdr));
2757 TCP_ECN_make_synack(req, th);
2758 th->source = htons(ireq->ir_num);
2759 th->dest = ireq->ir_rmt_port;
2760 /* Setting of flags are superfluous here for callers (and ECE is
2761 * not even correctly set)
2763 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2764 TCPHDR_SYN | TCPHDR_ACK);
2766 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2767 /* XXX data is queued and acked as is. No buffer/window check */
2768 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2770 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2771 th->window = htons(min(req->rcv_wnd, 65535U));
2772 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2773 th->doff = (tcp_header_size >> 2);
2774 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2776 #ifdef CONFIG_TCP_MD5SIG
2777 /* Okay, we have all we need - do the md5 hash if needed */
2779 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2780 md5, NULL, req, skb);
2786 EXPORT_SYMBOL(tcp_make_synack);
2788 /* Do all connect socket setups that can be done AF independent. */
2789 static void tcp_connect_init(struct sock *sk)
2791 const struct dst_entry *dst = __sk_dst_get(sk);
2792 struct tcp_sock *tp = tcp_sk(sk);
2795 /* We'll fix this up when we get a response from the other end.
2796 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2798 tp->tcp_header_len = sizeof(struct tcphdr) +
2799 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2801 #ifdef CONFIG_TCP_MD5SIG
2802 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2803 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2806 /* If user gave his TCP_MAXSEG, record it to clamp */
2807 if (tp->rx_opt.user_mss)
2808 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2811 tcp_sync_mss(sk, dst_mtu(dst));
2813 if (!tp->window_clamp)
2814 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2815 tp->advmss = dst_metric_advmss(dst);
2816 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2817 tp->advmss = tp->rx_opt.user_mss;
2819 tcp_initialize_rcv_mss(sk);
2821 /* limit the window selection if the user enforce a smaller rx buffer */
2822 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2823 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2824 tp->window_clamp = tcp_full_space(sk);
2826 tcp_select_initial_window(tcp_full_space(sk),
2827 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2830 sysctl_tcp_window_scaling,
2832 dst_metric(dst, RTAX_INITRWND));
2834 tp->rx_opt.rcv_wscale = rcv_wscale;
2835 tp->rcv_ssthresh = tp->rcv_wnd;
2838 sock_reset_flag(sk, SOCK_DONE);
2841 tp->snd_una = tp->write_seq;
2842 tp->snd_sml = tp->write_seq;
2843 tp->snd_up = tp->write_seq;
2844 tp->snd_nxt = tp->write_seq;
2846 if (likely(!tp->repair))
2849 tp->rcv_tstamp = tcp_time_stamp;
2850 tp->rcv_wup = tp->rcv_nxt;
2851 tp->copied_seq = tp->rcv_nxt;
2853 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2854 inet_csk(sk)->icsk_retransmits = 0;
2855 tcp_clear_retrans(tp);
2858 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2860 struct tcp_sock *tp = tcp_sk(sk);
2861 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2863 tcb->end_seq += skb->len;
2864 skb_header_release(skb);
2865 __tcp_add_write_queue_tail(sk, skb);
2866 sk->sk_wmem_queued += skb->truesize;
2867 sk_mem_charge(sk, skb->truesize);
2868 tp->write_seq = tcb->end_seq;
2869 tp->packets_out += tcp_skb_pcount(skb);
2872 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2873 * queue a data-only packet after the regular SYN, such that regular SYNs
2874 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2875 * only the SYN sequence, the data are retransmitted in the first ACK.
2876 * If cookie is not cached or other error occurs, falls back to send a
2877 * regular SYN with Fast Open cookie request option.
2879 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2881 struct tcp_sock *tp = tcp_sk(sk);
2882 struct tcp_fastopen_request *fo = tp->fastopen_req;
2883 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2884 struct sk_buff *syn_data = NULL, *data;
2885 unsigned long last_syn_loss = 0;
2887 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2888 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2889 &syn_loss, &last_syn_loss);
2890 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2892 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2893 fo->cookie.len = -1;
2897 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2898 fo->cookie.len = -1;
2899 else if (fo->cookie.len <= 0)
2902 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2903 * user-MSS. Reserve maximum option space for middleboxes that add
2904 * private TCP options. The cost is reduced data space in SYN :(
2906 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2907 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2908 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2909 MAX_TCP_OPTION_SPACE;
2911 space = min_t(size_t, space, fo->size);
2913 /* limit to order-0 allocations */
2914 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
2916 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
2918 if (syn_data == NULL)
2921 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2922 struct iovec *iov = &fo->data->msg_iov[i];
2923 unsigned char __user *from = iov->iov_base;
2924 int len = iov->iov_len;
2926 if (syn_data->len + len > space)
2927 len = space - syn_data->len;
2928 else if (i + 1 == iovlen)
2929 /* No more data pending in inet_wait_for_connect() */
2932 if (skb_add_data(syn_data, from, len))
2936 /* Queue a data-only packet after the regular SYN for retransmission */
2937 data = pskb_copy(syn_data, sk->sk_allocation);
2940 TCP_SKB_CB(data)->seq++;
2941 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2942 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2943 tcp_connect_queue_skb(sk, data);
2944 fo->copied = data->len;
2946 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2947 tp->syn_data = (fo->copied > 0);
2948 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2954 /* Send a regular SYN with Fast Open cookie request option */
2955 if (fo->cookie.len > 0)
2957 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2959 tp->syn_fastopen = 0;
2960 kfree_skb(syn_data);
2962 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2966 /* Build a SYN and send it off. */
2967 int tcp_connect(struct sock *sk)
2969 struct tcp_sock *tp = tcp_sk(sk);
2970 struct sk_buff *buff;
2973 tcp_connect_init(sk);
2975 if (unlikely(tp->repair)) {
2976 tcp_finish_connect(sk, NULL);
2980 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2981 if (unlikely(buff == NULL))
2984 /* Reserve space for headers. */
2985 skb_reserve(buff, MAX_TCP_HEADER);
2987 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2988 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2989 tcp_connect_queue_skb(sk, buff);
2990 TCP_ECN_send_syn(sk, buff);
2992 /* Send off SYN; include data in Fast Open. */
2993 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2994 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2995 if (err == -ECONNREFUSED)
2998 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2999 * in order to make this packet get counted in tcpOutSegs.
3001 tp->snd_nxt = tp->write_seq;
3002 tp->pushed_seq = tp->write_seq;
3003 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3005 /* Timer for repeating the SYN until an answer. */
3006 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3007 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3010 EXPORT_SYMBOL(tcp_connect);
3012 /* Send out a delayed ack, the caller does the policy checking
3013 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3016 void tcp_send_delayed_ack(struct sock *sk)
3018 struct inet_connection_sock *icsk = inet_csk(sk);
3019 int ato = icsk->icsk_ack.ato;
3020 unsigned long timeout;
3022 if (ato > TCP_DELACK_MIN) {
3023 const struct tcp_sock *tp = tcp_sk(sk);
3024 int max_ato = HZ / 2;
3026 if (icsk->icsk_ack.pingpong ||
3027 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3028 max_ato = TCP_DELACK_MAX;
3030 /* Slow path, intersegment interval is "high". */
3032 /* If some rtt estimate is known, use it to bound delayed ack.
3033 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3037 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3043 ato = min(ato, max_ato);
3046 /* Stay within the limit we were given */
3047 timeout = jiffies + ato;
3049 /* Use new timeout only if there wasn't a older one earlier. */
3050 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3051 /* If delack timer was blocked or is about to expire,
3054 if (icsk->icsk_ack.blocked ||
3055 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3060 if (!time_before(timeout, icsk->icsk_ack.timeout))
3061 timeout = icsk->icsk_ack.timeout;
3063 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3064 icsk->icsk_ack.timeout = timeout;
3065 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3068 /* This routine sends an ack and also updates the window. */
3069 void tcp_send_ack(struct sock *sk)
3071 struct sk_buff *buff;
3073 /* If we have been reset, we may not send again. */
3074 if (sk->sk_state == TCP_CLOSE)
3077 /* We are not putting this on the write queue, so
3078 * tcp_transmit_skb() will set the ownership to this
3081 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3083 inet_csk_schedule_ack(sk);
3084 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3085 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3086 TCP_DELACK_MAX, TCP_RTO_MAX);
3090 /* Reserve space for headers and prepare control bits. */
3091 skb_reserve(buff, MAX_TCP_HEADER);
3092 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3094 /* Send it off, this clears delayed acks for us. */
3095 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3096 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3099 /* This routine sends a packet with an out of date sequence
3100 * number. It assumes the other end will try to ack it.
3102 * Question: what should we make while urgent mode?
3103 * 4.4BSD forces sending single byte of data. We cannot send
3104 * out of window data, because we have SND.NXT==SND.MAX...
3106 * Current solution: to send TWO zero-length segments in urgent mode:
3107 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3108 * out-of-date with SND.UNA-1 to probe window.
3110 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3112 struct tcp_sock *tp = tcp_sk(sk);
3113 struct sk_buff *skb;
3115 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3116 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3120 /* Reserve space for headers and set control bits. */
3121 skb_reserve(skb, MAX_TCP_HEADER);
3122 /* Use a previous sequence. This should cause the other
3123 * end to send an ack. Don't queue or clone SKB, just
3126 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3127 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3128 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3131 void tcp_send_window_probe(struct sock *sk)
3133 if (sk->sk_state == TCP_ESTABLISHED) {
3134 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3135 tcp_xmit_probe_skb(sk, 0);
3139 /* Initiate keepalive or window probe from timer. */
3140 int tcp_write_wakeup(struct sock *sk)
3142 struct tcp_sock *tp = tcp_sk(sk);
3143 struct sk_buff *skb;
3145 if (sk->sk_state == TCP_CLOSE)
3148 if ((skb = tcp_send_head(sk)) != NULL &&
3149 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3151 unsigned int mss = tcp_current_mss(sk);
3152 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3154 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3155 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3157 /* We are probing the opening of a window
3158 * but the window size is != 0
3159 * must have been a result SWS avoidance ( sender )
3161 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3163 seg_size = min(seg_size, mss);
3164 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3165 if (tcp_fragment(sk, skb, seg_size, mss))
3167 } else if (!tcp_skb_pcount(skb))
3168 tcp_set_skb_tso_segs(sk, skb, mss);
3170 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3171 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3172 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3174 tcp_event_new_data_sent(sk, skb);
3177 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3178 tcp_xmit_probe_skb(sk, 1);
3179 return tcp_xmit_probe_skb(sk, 0);
3183 /* A window probe timeout has occurred. If window is not closed send
3184 * a partial packet else a zero probe.
3186 void tcp_send_probe0(struct sock *sk)
3188 struct inet_connection_sock *icsk = inet_csk(sk);
3189 struct tcp_sock *tp = tcp_sk(sk);
3192 err = tcp_write_wakeup(sk);
3194 if (tp->packets_out || !tcp_send_head(sk)) {
3195 /* Cancel probe timer, if it is not required. */
3196 icsk->icsk_probes_out = 0;
3197 icsk->icsk_backoff = 0;
3202 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3203 icsk->icsk_backoff++;
3204 icsk->icsk_probes_out++;
3205 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3206 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3209 /* If packet was not sent due to local congestion,
3210 * do not backoff and do not remember icsk_probes_out.
3211 * Let local senders to fight for local resources.
3213 * Use accumulated backoff yet.
3215 if (!icsk->icsk_probes_out)
3216 icsk->icsk_probes_out = 1;
3217 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3218 min(icsk->icsk_rto << icsk->icsk_backoff,
3219 TCP_RESOURCE_PROBE_INTERVAL),