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 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
74 struct inet_connection_sock *icsk = inet_csk(sk);
75 struct tcp_sock *tp = tcp_sk(sk);
76 unsigned int prior_packets = tp->packets_out;
78 tcp_advance_send_head(sk, skb);
79 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
81 /* Don't override Nagle indefinitely with F-RTO */
82 if (tp->frto_counter == 2)
85 tp->packets_out += tcp_skb_pcount(skb);
86 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
87 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;
167 if (sysctl_tcp_slow_start_after_idle &&
168 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
169 tcp_cwnd_restart(sk, __sk_dst_get(sk));
173 /* If it is a reply for ato after last received
174 * packet, enter pingpong mode.
176 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
177 icsk->icsk_ack.pingpong = 1;
180 /* Account for an ACK we sent. */
181 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
183 tcp_dec_quickack_mode(sk, pkts);
184 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
187 /* Determine a window scaling and initial window to offer.
188 * Based on the assumption that the given amount of space
189 * will be offered. Store the results in the tp structure.
190 * NOTE: for smooth operation initial space offering should
191 * be a multiple of mss if possible. We assume here that mss >= 1.
192 * This MUST be enforced by all callers.
194 void tcp_select_initial_window(int __space, __u32 mss,
195 __u32 *rcv_wnd, __u32 *window_clamp,
196 int wscale_ok, __u8 *rcv_wscale,
199 unsigned int space = (__space < 0 ? 0 : __space);
201 /* If no clamp set the clamp to the max possible scaled window */
202 if (*window_clamp == 0)
203 (*window_clamp) = (65535 << 14);
204 space = min(*window_clamp, space);
206 /* Quantize space offering to a multiple of mss if possible. */
208 space = (space / mss) * mss;
210 /* NOTE: offering an initial window larger than 32767
211 * will break some buggy TCP stacks. If the admin tells us
212 * it is likely we could be speaking with such a buggy stack
213 * we will truncate our initial window offering to 32K-1
214 * unless the remote has sent us a window scaling option,
215 * which we interpret as a sign the remote TCP is not
216 * misinterpreting the window field as a signed quantity.
218 if (sysctl_tcp_workaround_signed_windows)
219 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
225 /* Set window scaling on max possible window
226 * See RFC1323 for an explanation of the limit to 14
228 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
229 space = min_t(u32, space, *window_clamp);
230 while (space > 65535 && (*rcv_wscale) < 14) {
236 /* Set initial window to a value enough for senders starting with
237 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
238 * a limit on the initial window when mss is larger than 1460.
240 if (mss > (1 << *rcv_wscale)) {
241 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
244 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
245 /* when initializing use the value from init_rcv_wnd
246 * rather than the default from above
249 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
251 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
257 EXPORT_SYMBOL(tcp_select_initial_window);
259 /* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
264 static u16 tcp_select_window(struct sock *sk)
266 struct tcp_sock *tp = tcp_sk(sk);
267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
270 /* Never shrink the offered window */
271 if (new_win < cur_win) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
277 * Relax Will Robinson.
279 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
281 tp->rcv_wnd = new_win;
282 tp->rcv_wup = tp->rcv_nxt;
284 /* Make sure we do not exceed the maximum possible
287 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
302 /* Packet ECN state for a SYN-ACK */
303 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
305 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
306 if (!(tp->ecn_flags & TCP_ECN_OK))
307 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
310 /* Packet ECN state for a SYN. */
311 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
313 struct tcp_sock *tp = tcp_sk(sk);
316 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
317 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
318 tp->ecn_flags = TCP_ECN_OK;
322 static __inline__ void
323 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
325 if (inet_rsk(req)->ecn_ok)
329 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
332 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
335 struct tcp_sock *tp = tcp_sk(sk);
337 if (tp->ecn_flags & TCP_ECN_OK) {
338 /* Not-retransmitted data segment: set ECT and inject CWR. */
339 if (skb->len != tcp_header_len &&
340 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
342 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
343 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
344 tcp_hdr(skb)->cwr = 1;
345 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
348 /* ACK or retransmitted segment: clear ECT|CE */
349 INET_ECN_dontxmit(sk);
351 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
352 tcp_hdr(skb)->ece = 1;
356 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
357 * auto increment end seqno.
359 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
361 skb->ip_summed = CHECKSUM_PARTIAL;
364 TCP_SKB_CB(skb)->tcp_flags = flags;
365 TCP_SKB_CB(skb)->sacked = 0;
367 skb_shinfo(skb)->gso_segs = 1;
368 skb_shinfo(skb)->gso_size = 0;
369 skb_shinfo(skb)->gso_type = 0;
371 TCP_SKB_CB(skb)->seq = seq;
372 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
374 TCP_SKB_CB(skb)->end_seq = seq;
377 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
379 return tp->snd_una != tp->snd_up;
382 #define OPTION_SACK_ADVERTISE (1 << 0)
383 #define OPTION_TS (1 << 1)
384 #define OPTION_MD5 (1 << 2)
385 #define OPTION_WSCALE (1 << 3)
386 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
388 struct tcp_out_options {
389 u16 options; /* bit field of OPTION_* */
390 u16 mss; /* 0 to disable */
391 u8 ws; /* window scale, 0 to disable */
392 u8 num_sack_blocks; /* number of SACK blocks to include */
393 u8 hash_size; /* bytes in hash_location */
394 __u8 *hash_location; /* temporary pointer, overloaded */
395 __u32 tsval, tsecr; /* need to include OPTION_TS */
396 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
399 /* Write previously computed TCP options to the packet.
401 * Beware: Something in the Internet is very sensitive to the ordering of
402 * TCP options, we learned this through the hard way, so be careful here.
403 * Luckily we can at least blame others for their non-compliance but from
404 * inter-operatibility perspective it seems that we're somewhat stuck with
405 * the ordering which we have been using if we want to keep working with
406 * those broken things (not that it currently hurts anybody as there isn't
407 * particular reason why the ordering would need to be changed).
409 * At least SACK_PERM as the first option is known to lead to a disaster
410 * (but it may well be that other scenarios fail similarly).
412 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
413 struct tcp_out_options *opts)
415 u16 options = opts->options; /* mungable copy */
417 if (unlikely(OPTION_MD5 & options)) {
418 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
419 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
420 /* overload cookie hash location */
421 opts->hash_location = (__u8 *)ptr;
425 if (unlikely(opts->mss)) {
426 *ptr++ = htonl((TCPOPT_MSS << 24) |
427 (TCPOLEN_MSS << 16) |
431 if (likely(OPTION_TS & options)) {
432 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
433 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
434 (TCPOLEN_SACK_PERM << 16) |
435 (TCPOPT_TIMESTAMP << 8) |
437 options &= ~OPTION_SACK_ADVERTISE;
439 *ptr++ = htonl((TCPOPT_NOP << 24) |
441 (TCPOPT_TIMESTAMP << 8) |
444 *ptr++ = htonl(opts->tsval);
445 *ptr++ = htonl(opts->tsecr);
448 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
449 *ptr++ = htonl((TCPOPT_NOP << 24) |
451 (TCPOPT_SACK_PERM << 8) |
455 if (unlikely(OPTION_WSCALE & options)) {
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
457 (TCPOPT_WINDOW << 16) |
458 (TCPOLEN_WINDOW << 8) |
462 if (unlikely(opts->num_sack_blocks)) {
463 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
464 tp->duplicate_sack : tp->selective_acks;
467 *ptr++ = htonl((TCPOPT_NOP << 24) |
470 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
471 TCPOLEN_SACK_PERBLOCK)));
473 for (this_sack = 0; this_sack < opts->num_sack_blocks;
475 *ptr++ = htonl(sp[this_sack].start_seq);
476 *ptr++ = htonl(sp[this_sack].end_seq);
479 tp->rx_opt.dsack = 0;
482 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
483 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
485 *ptr++ = htonl((TCPOPT_EXP << 24) |
486 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
487 TCPOPT_FASTOPEN_MAGIC);
489 memcpy(ptr, foc->val, foc->len);
490 if ((foc->len & 3) == 2) {
491 u8 *align = ((u8 *)ptr) + foc->len;
492 align[0] = align[1] = TCPOPT_NOP;
494 ptr += (foc->len + 3) >> 2;
498 /* Compute TCP options for SYN packets. This is not the final
499 * network wire format yet.
501 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
502 struct tcp_out_options *opts,
503 struct tcp_md5sig_key **md5)
505 struct tcp_sock *tp = tcp_sk(sk);
506 unsigned int remaining = MAX_TCP_OPTION_SPACE;
507 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
509 #ifdef CONFIG_TCP_MD5SIG
510 *md5 = tp->af_specific->md5_lookup(sk, sk);
512 opts->options |= OPTION_MD5;
513 remaining -= TCPOLEN_MD5SIG_ALIGNED;
519 /* We always get an MSS option. The option bytes which will be seen in
520 * normal data packets should timestamps be used, must be in the MSS
521 * advertised. But we subtract them from tp->mss_cache so that
522 * calculations in tcp_sendmsg are simpler etc. So account for this
523 * fact here if necessary. If we don't do this correctly, as a
524 * receiver we won't recognize data packets as being full sized when we
525 * should, and thus we won't abide by the delayed ACK rules correctly.
526 * SACKs don't matter, we never delay an ACK when we have any of those
528 opts->mss = tcp_advertise_mss(sk);
529 remaining -= TCPOLEN_MSS_ALIGNED;
531 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
532 opts->options |= OPTION_TS;
533 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
534 opts->tsecr = tp->rx_opt.ts_recent;
535 remaining -= TCPOLEN_TSTAMP_ALIGNED;
537 if (likely(sysctl_tcp_window_scaling)) {
538 opts->ws = tp->rx_opt.rcv_wscale;
539 opts->options |= OPTION_WSCALE;
540 remaining -= TCPOLEN_WSCALE_ALIGNED;
542 if (likely(sysctl_tcp_sack)) {
543 opts->options |= OPTION_SACK_ADVERTISE;
544 if (unlikely(!(OPTION_TS & opts->options)))
545 remaining -= TCPOLEN_SACKPERM_ALIGNED;
548 if (fastopen && fastopen->cookie.len >= 0) {
549 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
550 need = (need + 3) & ~3U; /* Align to 32 bits */
551 if (remaining >= need) {
552 opts->options |= OPTION_FAST_OPEN_COOKIE;
553 opts->fastopen_cookie = &fastopen->cookie;
555 tp->syn_fastopen = 1;
559 return MAX_TCP_OPTION_SPACE - remaining;
562 /* Set up TCP options for SYN-ACKs. */
563 static unsigned int tcp_synack_options(struct sock *sk,
564 struct request_sock *req,
565 unsigned int mss, struct sk_buff *skb,
566 struct tcp_out_options *opts,
567 struct tcp_md5sig_key **md5,
568 struct tcp_fastopen_cookie *foc)
570 struct inet_request_sock *ireq = inet_rsk(req);
571 unsigned int remaining = MAX_TCP_OPTION_SPACE;
573 #ifdef CONFIG_TCP_MD5SIG
574 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
576 opts->options |= OPTION_MD5;
577 remaining -= TCPOLEN_MD5SIG_ALIGNED;
579 /* We can't fit any SACK blocks in a packet with MD5 + TS
580 * options. There was discussion about disabling SACK
581 * rather than TS in order to fit in better with old,
582 * buggy kernels, but that was deemed to be unnecessary.
584 ireq->tstamp_ok &= !ireq->sack_ok;
590 /* We always send an MSS option. */
592 remaining -= TCPOLEN_MSS_ALIGNED;
594 if (likely(ireq->wscale_ok)) {
595 opts->ws = ireq->rcv_wscale;
596 opts->options |= OPTION_WSCALE;
597 remaining -= TCPOLEN_WSCALE_ALIGNED;
599 if (likely(ireq->tstamp_ok)) {
600 opts->options |= OPTION_TS;
601 opts->tsval = TCP_SKB_CB(skb)->when;
602 opts->tsecr = req->ts_recent;
603 remaining -= TCPOLEN_TSTAMP_ALIGNED;
605 if (likely(ireq->sack_ok)) {
606 opts->options |= OPTION_SACK_ADVERTISE;
607 if (unlikely(!ireq->tstamp_ok))
608 remaining -= TCPOLEN_SACKPERM_ALIGNED;
611 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
612 need = (need + 3) & ~3U; /* Align to 32 bits */
613 if (remaining >= need) {
614 opts->options |= OPTION_FAST_OPEN_COOKIE;
615 opts->fastopen_cookie = foc;
620 return MAX_TCP_OPTION_SPACE - remaining;
623 /* Compute TCP options for ESTABLISHED sockets. This is not the
624 * final wire format yet.
626 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
627 struct tcp_out_options *opts,
628 struct tcp_md5sig_key **md5)
630 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
631 struct tcp_sock *tp = tcp_sk(sk);
632 unsigned int size = 0;
633 unsigned int eff_sacks;
635 #ifdef CONFIG_TCP_MD5SIG
636 *md5 = tp->af_specific->md5_lookup(sk, sk);
637 if (unlikely(*md5)) {
638 opts->options |= OPTION_MD5;
639 size += TCPOLEN_MD5SIG_ALIGNED;
645 if (likely(tp->rx_opt.tstamp_ok)) {
646 opts->options |= OPTION_TS;
647 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
648 opts->tsecr = tp->rx_opt.ts_recent;
649 size += TCPOLEN_TSTAMP_ALIGNED;
652 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
653 if (unlikely(eff_sacks)) {
654 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
655 opts->num_sack_blocks =
656 min_t(unsigned int, eff_sacks,
657 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
658 TCPOLEN_SACK_PERBLOCK);
659 size += TCPOLEN_SACK_BASE_ALIGNED +
660 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
667 /* TCP SMALL QUEUES (TSQ)
669 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
670 * to reduce RTT and bufferbloat.
671 * We do this using a special skb destructor (tcp_wfree).
673 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
674 * needs to be reallocated in a driver.
675 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
677 * Since transmit from skb destructor is forbidden, we use a tasklet
678 * to process all sockets that eventually need to send more skbs.
679 * We use one tasklet per cpu, with its own queue of sockets.
682 struct tasklet_struct tasklet;
683 struct list_head head; /* queue of tcp sockets */
685 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
687 static void tcp_tsq_handler(struct sock *sk)
689 if ((1 << sk->sk_state) &
690 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
691 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
692 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
695 * One tasklest per cpu tries to send more skbs.
696 * We run in tasklet context but need to disable irqs when
697 * transfering tsq->head because tcp_wfree() might
698 * interrupt us (non NAPI drivers)
700 static void tcp_tasklet_func(unsigned long data)
702 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
705 struct list_head *q, *n;
709 local_irq_save(flags);
710 list_splice_init(&tsq->head, &list);
711 local_irq_restore(flags);
713 list_for_each_safe(q, n, &list) {
714 tp = list_entry(q, struct tcp_sock, tsq_node);
715 list_del(&tp->tsq_node);
717 sk = (struct sock *)tp;
720 if (!sock_owned_by_user(sk)) {
723 /* defer the work to tcp_release_cb() */
724 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
728 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
733 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
734 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
735 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
736 (1UL << TCP_MTU_REDUCED_DEFERRED))
738 * tcp_release_cb - tcp release_sock() callback
741 * called from release_sock() to perform protocol dependent
742 * actions before socket release.
744 void tcp_release_cb(struct sock *sk)
746 struct tcp_sock *tp = tcp_sk(sk);
747 unsigned long flags, nflags;
749 /* perform an atomic operation only if at least one flag is set */
751 flags = tp->tsq_flags;
752 if (!(flags & TCP_DEFERRED_ALL))
754 nflags = flags & ~TCP_DEFERRED_ALL;
755 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
757 if (flags & (1UL << TCP_TSQ_DEFERRED))
760 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
761 tcp_write_timer_handler(sk);
764 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
765 tcp_delack_timer_handler(sk);
768 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
769 sk->sk_prot->mtu_reduced(sk);
773 EXPORT_SYMBOL(tcp_release_cb);
775 void __init tcp_tasklet_init(void)
779 for_each_possible_cpu(i) {
780 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
782 INIT_LIST_HEAD(&tsq->head);
783 tasklet_init(&tsq->tasklet,
790 * Write buffer destructor automatically called from kfree_skb.
791 * We cant xmit new skbs from this context, as we might already
794 static void tcp_wfree(struct sk_buff *skb)
796 struct sock *sk = skb->sk;
797 struct tcp_sock *tp = tcp_sk(sk);
799 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
800 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
802 struct tsq_tasklet *tsq;
804 /* Keep a ref on socket.
805 * This last ref will be released in tcp_tasklet_func()
807 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
809 /* queue this socket to tasklet queue */
810 local_irq_save(flags);
811 tsq = &__get_cpu_var(tsq_tasklet);
812 list_add(&tp->tsq_node, &tsq->head);
813 tasklet_schedule(&tsq->tasklet);
814 local_irq_restore(flags);
820 /* This routine actually transmits TCP packets queued in by
821 * tcp_do_sendmsg(). This is used by both the initial
822 * transmission and possible later retransmissions.
823 * All SKB's seen here are completely headerless. It is our
824 * job to build the TCP header, and pass the packet down to
825 * IP so it can do the same plus pass the packet off to the
828 * We are working here with either a clone of the original
829 * SKB, or a fresh unique copy made by the retransmit engine.
831 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
834 const struct inet_connection_sock *icsk = inet_csk(sk);
835 struct inet_sock *inet;
837 struct tcp_skb_cb *tcb;
838 struct tcp_out_options opts;
839 unsigned int tcp_options_size, tcp_header_size;
840 struct tcp_md5sig_key *md5;
844 BUG_ON(!skb || !tcp_skb_pcount(skb));
846 /* If congestion control is doing timestamping, we must
847 * take such a timestamp before we potentially clone/copy.
849 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
850 __net_timestamp(skb);
852 if (likely(clone_it)) {
853 if (unlikely(skb_cloned(skb)))
854 skb = pskb_copy(skb, gfp_mask);
856 skb = skb_clone(skb, gfp_mask);
863 tcb = TCP_SKB_CB(skb);
864 memset(&opts, 0, sizeof(opts));
866 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
867 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
869 tcp_options_size = tcp_established_options(sk, skb, &opts,
871 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
873 if (tcp_packets_in_flight(tp) == 0) {
874 tcp_ca_event(sk, CA_EVENT_TX_START);
879 skb_push(skb, tcp_header_size);
880 skb_reset_transport_header(skb);
884 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
885 tcp_wfree : sock_wfree;
886 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
888 /* Build TCP header and checksum it. */
890 th->source = inet->inet_sport;
891 th->dest = inet->inet_dport;
892 th->seq = htonl(tcb->seq);
893 th->ack_seq = htonl(tp->rcv_nxt);
894 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
897 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
898 /* RFC1323: The window in SYN & SYN/ACK segments
901 th->window = htons(min(tp->rcv_wnd, 65535U));
903 th->window = htons(tcp_select_window(sk));
908 /* The urg_mode check is necessary during a below snd_una win probe */
909 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
910 if (before(tp->snd_up, tcb->seq + 0x10000)) {
911 th->urg_ptr = htons(tp->snd_up - tcb->seq);
913 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
914 th->urg_ptr = htons(0xFFFF);
919 tcp_options_write((__be32 *)(th + 1), tp, &opts);
920 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
921 TCP_ECN_send(sk, skb, tcp_header_size);
923 #ifdef CONFIG_TCP_MD5SIG
924 /* Calculate the MD5 hash, as we have all we need now */
926 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
927 tp->af_specific->calc_md5_hash(opts.hash_location,
932 icsk->icsk_af_ops->send_check(sk, skb);
934 if (likely(tcb->tcp_flags & TCPHDR_ACK))
935 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
937 if (skb->len != tcp_header_size)
938 tcp_event_data_sent(tp, sk);
940 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
941 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
942 tcp_skb_pcount(skb));
944 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
945 if (likely(err <= 0))
948 tcp_enter_cwr(sk, 1);
950 return net_xmit_eval(err);
953 /* This routine just queues the buffer for sending.
955 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
956 * otherwise socket can stall.
958 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
960 struct tcp_sock *tp = tcp_sk(sk);
962 /* Advance write_seq and place onto the write_queue. */
963 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
964 skb_header_release(skb);
965 tcp_add_write_queue_tail(sk, skb);
966 sk->sk_wmem_queued += skb->truesize;
967 sk_mem_charge(sk, skb->truesize);
970 /* Initialize TSO segments for a packet. */
971 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
972 unsigned int mss_now)
974 if (skb->len <= mss_now || !sk_can_gso(sk) ||
975 skb->ip_summed == CHECKSUM_NONE) {
976 /* Avoid the costly divide in the normal
979 skb_shinfo(skb)->gso_segs = 1;
980 skb_shinfo(skb)->gso_size = 0;
981 skb_shinfo(skb)->gso_type = 0;
983 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
984 skb_shinfo(skb)->gso_size = mss_now;
985 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
989 /* When a modification to fackets out becomes necessary, we need to check
990 * skb is counted to fackets_out or not.
992 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
995 struct tcp_sock *tp = tcp_sk(sk);
997 if (!tp->sacked_out || tcp_is_reno(tp))
1000 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1001 tp->fackets_out -= decr;
1004 /* Pcount in the middle of the write queue got changed, we need to do various
1005 * tweaks to fix counters
1007 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1009 struct tcp_sock *tp = tcp_sk(sk);
1011 tp->packets_out -= decr;
1013 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1014 tp->sacked_out -= decr;
1015 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1016 tp->retrans_out -= decr;
1017 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1018 tp->lost_out -= decr;
1020 /* Reno case is special. Sigh... */
1021 if (tcp_is_reno(tp) && decr > 0)
1022 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1024 tcp_adjust_fackets_out(sk, skb, decr);
1026 if (tp->lost_skb_hint &&
1027 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1028 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1029 tp->lost_cnt_hint -= decr;
1031 tcp_verify_left_out(tp);
1034 /* Function to create two new TCP segments. Shrinks the given segment
1035 * to the specified size and appends a new segment with the rest of the
1036 * packet to the list. This won't be called frequently, I hope.
1037 * Remember, these are still headerless SKBs at this point.
1039 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1040 unsigned int mss_now)
1042 struct tcp_sock *tp = tcp_sk(sk);
1043 struct sk_buff *buff;
1044 int nsize, old_factor;
1048 if (WARN_ON(len > skb->len))
1051 nsize = skb_headlen(skb) - len;
1055 if (skb_cloned(skb) &&
1056 skb_is_nonlinear(skb) &&
1057 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1060 /* Get a new skb... force flag on. */
1061 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1063 return -ENOMEM; /* We'll just try again later. */
1065 sk->sk_wmem_queued += buff->truesize;
1066 sk_mem_charge(sk, buff->truesize);
1067 nlen = skb->len - len - nsize;
1068 buff->truesize += nlen;
1069 skb->truesize -= nlen;
1071 /* Correct the sequence numbers. */
1072 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1073 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1074 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1076 /* PSH and FIN should only be set in the second packet. */
1077 flags = TCP_SKB_CB(skb)->tcp_flags;
1078 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1079 TCP_SKB_CB(buff)->tcp_flags = flags;
1080 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1082 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1083 /* Copy and checksum data tail into the new buffer. */
1084 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1085 skb_put(buff, nsize),
1090 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1092 skb->ip_summed = CHECKSUM_PARTIAL;
1093 skb_split(skb, buff, len);
1096 buff->ip_summed = skb->ip_summed;
1098 /* Looks stupid, but our code really uses when of
1099 * skbs, which it never sent before. --ANK
1101 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1102 buff->tstamp = skb->tstamp;
1104 old_factor = tcp_skb_pcount(skb);
1106 /* Fix up tso_factor for both original and new SKB. */
1107 tcp_set_skb_tso_segs(sk, skb, mss_now);
1108 tcp_set_skb_tso_segs(sk, buff, mss_now);
1110 /* If this packet has been sent out already, we must
1111 * adjust the various packet counters.
1113 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1114 int diff = old_factor - tcp_skb_pcount(skb) -
1115 tcp_skb_pcount(buff);
1118 tcp_adjust_pcount(sk, skb, diff);
1121 /* Link BUFF into the send queue. */
1122 skb_header_release(buff);
1123 tcp_insert_write_queue_after(skb, buff, sk);
1128 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1129 * eventually). The difference is that pulled data not copied, but
1130 * immediately discarded.
1132 static void __pskb_trim_head(struct sk_buff *skb, int len)
1136 eat = min_t(int, len, skb_headlen(skb));
1138 __skb_pull(skb, eat);
1145 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1146 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1149 skb_frag_unref(skb, i);
1152 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1154 skb_shinfo(skb)->frags[k].page_offset += eat;
1155 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1161 skb_shinfo(skb)->nr_frags = k;
1163 skb_reset_tail_pointer(skb);
1164 skb->data_len -= len;
1165 skb->len = skb->data_len;
1168 /* Remove acked data from a packet in the transmit queue. */
1169 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1171 if (skb_unclone(skb, GFP_ATOMIC))
1174 __pskb_trim_head(skb, len);
1176 TCP_SKB_CB(skb)->seq += len;
1177 skb->ip_summed = CHECKSUM_PARTIAL;
1179 skb->truesize -= len;
1180 sk->sk_wmem_queued -= len;
1181 sk_mem_uncharge(sk, len);
1182 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1184 /* Any change of skb->len requires recalculation of tso factor. */
1185 if (tcp_skb_pcount(skb) > 1)
1186 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1191 /* Calculate MSS not accounting any TCP options. */
1192 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1194 const struct tcp_sock *tp = tcp_sk(sk);
1195 const struct inet_connection_sock *icsk = inet_csk(sk);
1198 /* Calculate base mss without TCP options:
1199 It is MMS_S - sizeof(tcphdr) of rfc1122
1201 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1203 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1204 if (icsk->icsk_af_ops->net_frag_header_len) {
1205 const struct dst_entry *dst = __sk_dst_get(sk);
1207 if (dst && dst_allfrag(dst))
1208 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1211 /* Clamp it (mss_clamp does not include tcp options) */
1212 if (mss_now > tp->rx_opt.mss_clamp)
1213 mss_now = tp->rx_opt.mss_clamp;
1215 /* Now subtract optional transport overhead */
1216 mss_now -= icsk->icsk_ext_hdr_len;
1218 /* Then reserve room for full set of TCP options and 8 bytes of data */
1224 /* Calculate MSS. Not accounting for SACKs here. */
1225 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1227 /* Subtract TCP options size, not including SACKs */
1228 return __tcp_mtu_to_mss(sk, pmtu) -
1229 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1232 /* Inverse of above */
1233 int tcp_mss_to_mtu(struct sock *sk, int mss)
1235 const struct tcp_sock *tp = tcp_sk(sk);
1236 const struct inet_connection_sock *icsk = inet_csk(sk);
1240 tp->tcp_header_len +
1241 icsk->icsk_ext_hdr_len +
1242 icsk->icsk_af_ops->net_header_len;
1244 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1245 if (icsk->icsk_af_ops->net_frag_header_len) {
1246 const struct dst_entry *dst = __sk_dst_get(sk);
1248 if (dst && dst_allfrag(dst))
1249 mtu += icsk->icsk_af_ops->net_frag_header_len;
1254 /* MTU probing init per socket */
1255 void tcp_mtup_init(struct sock *sk)
1257 struct tcp_sock *tp = tcp_sk(sk);
1258 struct inet_connection_sock *icsk = inet_csk(sk);
1260 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1261 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1262 icsk->icsk_af_ops->net_header_len;
1263 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1264 icsk->icsk_mtup.probe_size = 0;
1266 EXPORT_SYMBOL(tcp_mtup_init);
1268 /* This function synchronize snd mss to current pmtu/exthdr set.
1270 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1271 for TCP options, but includes only bare TCP header.
1273 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1274 It is minimum of user_mss and mss received with SYN.
1275 It also does not include TCP options.
1277 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1279 tp->mss_cache is current effective sending mss, including
1280 all tcp options except for SACKs. It is evaluated,
1281 taking into account current pmtu, but never exceeds
1282 tp->rx_opt.mss_clamp.
1284 NOTE1. rfc1122 clearly states that advertised MSS
1285 DOES NOT include either tcp or ip options.
1287 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1288 are READ ONLY outside this function. --ANK (980731)
1290 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1292 struct tcp_sock *tp = tcp_sk(sk);
1293 struct inet_connection_sock *icsk = inet_csk(sk);
1296 if (icsk->icsk_mtup.search_high > pmtu)
1297 icsk->icsk_mtup.search_high = pmtu;
1299 mss_now = tcp_mtu_to_mss(sk, pmtu);
1300 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1302 /* And store cached results */
1303 icsk->icsk_pmtu_cookie = pmtu;
1304 if (icsk->icsk_mtup.enabled)
1305 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1306 tp->mss_cache = mss_now;
1310 EXPORT_SYMBOL(tcp_sync_mss);
1312 /* Compute the current effective MSS, taking SACKs and IP options,
1313 * and even PMTU discovery events into account.
1315 unsigned int tcp_current_mss(struct sock *sk)
1317 const struct tcp_sock *tp = tcp_sk(sk);
1318 const struct dst_entry *dst = __sk_dst_get(sk);
1320 unsigned int header_len;
1321 struct tcp_out_options opts;
1322 struct tcp_md5sig_key *md5;
1324 mss_now = tp->mss_cache;
1327 u32 mtu = dst_mtu(dst);
1328 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1329 mss_now = tcp_sync_mss(sk, mtu);
1332 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1333 sizeof(struct tcphdr);
1334 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1335 * some common options. If this is an odd packet (because we have SACK
1336 * blocks etc) then our calculated header_len will be different, and
1337 * we have to adjust mss_now correspondingly */
1338 if (header_len != tp->tcp_header_len) {
1339 int delta = (int) header_len - tp->tcp_header_len;
1346 /* Congestion window validation. (RFC2861) */
1347 static void tcp_cwnd_validate(struct sock *sk)
1349 struct tcp_sock *tp = tcp_sk(sk);
1351 if (tp->packets_out >= tp->snd_cwnd) {
1352 /* Network is feed fully. */
1353 tp->snd_cwnd_used = 0;
1354 tp->snd_cwnd_stamp = tcp_time_stamp;
1356 /* Network starves. */
1357 if (tp->packets_out > tp->snd_cwnd_used)
1358 tp->snd_cwnd_used = tp->packets_out;
1360 if (sysctl_tcp_slow_start_after_idle &&
1361 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1362 tcp_cwnd_application_limited(sk);
1366 /* Returns the portion of skb which can be sent right away without
1367 * introducing MSS oddities to segment boundaries. In rare cases where
1368 * mss_now != mss_cache, we will request caller to create a small skb
1369 * per input skb which could be mostly avoided here (if desired).
1371 * We explicitly want to create a request for splitting write queue tail
1372 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1373 * thus all the complexity (cwnd_len is always MSS multiple which we
1374 * return whenever allowed by the other factors). Basically we need the
1375 * modulo only when the receiver window alone is the limiting factor or
1376 * when we would be allowed to send the split-due-to-Nagle skb fully.
1378 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1379 unsigned int mss_now, unsigned int max_segs)
1381 const struct tcp_sock *tp = tcp_sk(sk);
1382 u32 needed, window, max_len;
1384 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1385 max_len = mss_now * max_segs;
1387 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1390 needed = min(skb->len, window);
1392 if (max_len <= needed)
1395 return needed - needed % mss_now;
1398 /* Can at least one segment of SKB be sent right now, according to the
1399 * congestion window rules? If so, return how many segments are allowed.
1401 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1402 const struct sk_buff *skb)
1404 u32 in_flight, cwnd;
1406 /* Don't be strict about the congestion window for the final FIN. */
1407 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1408 tcp_skb_pcount(skb) == 1)
1411 in_flight = tcp_packets_in_flight(tp);
1412 cwnd = tp->snd_cwnd;
1413 if (in_flight < cwnd)
1414 return (cwnd - in_flight);
1419 /* Initialize TSO state of a skb.
1420 * This must be invoked the first time we consider transmitting
1421 * SKB onto the wire.
1423 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1424 unsigned int mss_now)
1426 int tso_segs = tcp_skb_pcount(skb);
1428 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1429 tcp_set_skb_tso_segs(sk, skb, mss_now);
1430 tso_segs = tcp_skb_pcount(skb);
1435 /* Minshall's variant of the Nagle send check. */
1436 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1438 return after(tp->snd_sml, tp->snd_una) &&
1439 !after(tp->snd_sml, tp->snd_nxt);
1442 /* Return false, if packet can be sent now without violation Nagle's rules:
1443 * 1. It is full sized.
1444 * 2. Or it contains FIN. (already checked by caller)
1445 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1446 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1447 * With Minshall's modification: all sent small packets are ACKed.
1449 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1450 const struct sk_buff *skb,
1451 unsigned int mss_now, int nonagle)
1453 return skb->len < mss_now &&
1454 ((nonagle & TCP_NAGLE_CORK) ||
1455 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1458 /* Return true if the Nagle test allows this packet to be
1461 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1462 unsigned int cur_mss, int nonagle)
1464 /* Nagle rule does not apply to frames, which sit in the middle of the
1465 * write_queue (they have no chances to get new data).
1467 * This is implemented in the callers, where they modify the 'nonagle'
1468 * argument based upon the location of SKB in the send queue.
1470 if (nonagle & TCP_NAGLE_PUSH)
1473 /* Don't use the nagle rule for urgent data (or for the final FIN).
1474 * Nagle can be ignored during F-RTO too (see RFC4138).
1476 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1477 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1480 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1486 /* Does at least the first segment of SKB fit into the send window? */
1487 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1488 const struct sk_buff *skb,
1489 unsigned int cur_mss)
1491 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1493 if (skb->len > cur_mss)
1494 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1496 return !after(end_seq, tcp_wnd_end(tp));
1499 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1500 * should be put on the wire right now. If so, it returns the number of
1501 * packets allowed by the congestion window.
1503 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1504 unsigned int cur_mss, int nonagle)
1506 const struct tcp_sock *tp = tcp_sk(sk);
1507 unsigned int cwnd_quota;
1509 tcp_init_tso_segs(sk, skb, cur_mss);
1511 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1514 cwnd_quota = tcp_cwnd_test(tp, skb);
1515 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1521 /* Test if sending is allowed right now. */
1522 bool tcp_may_send_now(struct sock *sk)
1524 const struct tcp_sock *tp = tcp_sk(sk);
1525 struct sk_buff *skb = tcp_send_head(sk);
1528 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1529 (tcp_skb_is_last(sk, skb) ?
1530 tp->nonagle : TCP_NAGLE_PUSH));
1533 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1534 * which is put after SKB on the list. It is very much like
1535 * tcp_fragment() except that it may make several kinds of assumptions
1536 * in order to speed up the splitting operation. In particular, we
1537 * know that all the data is in scatter-gather pages, and that the
1538 * packet has never been sent out before (and thus is not cloned).
1540 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1541 unsigned int mss_now, gfp_t gfp)
1543 struct sk_buff *buff;
1544 int nlen = skb->len - len;
1547 /* All of a TSO frame must be composed of paged data. */
1548 if (skb->len != skb->data_len)
1549 return tcp_fragment(sk, skb, len, mss_now);
1551 buff = sk_stream_alloc_skb(sk, 0, gfp);
1552 if (unlikely(buff == NULL))
1555 sk->sk_wmem_queued += buff->truesize;
1556 sk_mem_charge(sk, buff->truesize);
1557 buff->truesize += nlen;
1558 skb->truesize -= nlen;
1560 /* Correct the sequence numbers. */
1561 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1562 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1563 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1565 /* PSH and FIN should only be set in the second packet. */
1566 flags = TCP_SKB_CB(skb)->tcp_flags;
1567 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1568 TCP_SKB_CB(buff)->tcp_flags = flags;
1570 /* This packet was never sent out yet, so no SACK bits. */
1571 TCP_SKB_CB(buff)->sacked = 0;
1573 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1574 skb_split(skb, buff, len);
1576 /* Fix up tso_factor for both original and new SKB. */
1577 tcp_set_skb_tso_segs(sk, skb, mss_now);
1578 tcp_set_skb_tso_segs(sk, buff, mss_now);
1580 /* Link BUFF into the send queue. */
1581 skb_header_release(buff);
1582 tcp_insert_write_queue_after(skb, buff, sk);
1587 /* Try to defer sending, if possible, in order to minimize the amount
1588 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1590 * This algorithm is from John Heffner.
1592 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1594 struct tcp_sock *tp = tcp_sk(sk);
1595 const struct inet_connection_sock *icsk = inet_csk(sk);
1596 u32 send_win, cong_win, limit, in_flight;
1599 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1602 if (icsk->icsk_ca_state != TCP_CA_Open)
1605 /* Defer for less than two clock ticks. */
1606 if (tp->tso_deferred &&
1607 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1610 in_flight = tcp_packets_in_flight(tp);
1612 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1614 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1616 /* From in_flight test above, we know that cwnd > in_flight. */
1617 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1619 limit = min(send_win, cong_win);
1621 /* If a full-sized TSO skb can be sent, do it. */
1622 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1623 sk->sk_gso_max_segs * tp->mss_cache))
1626 /* Middle in queue won't get any more data, full sendable already? */
1627 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1630 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1632 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1634 /* If at least some fraction of a window is available,
1637 chunk /= win_divisor;
1641 /* Different approach, try not to defer past a single
1642 * ACK. Receiver should ACK every other full sized
1643 * frame, so if we have space for more than 3 frames
1646 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1650 /* Ok, it looks like it is advisable to defer. */
1651 tp->tso_deferred = 1 | (jiffies << 1);
1656 tp->tso_deferred = 0;
1660 /* Create a new MTU probe if we are ready.
1661 * MTU probe is regularly attempting to increase the path MTU by
1662 * deliberately sending larger packets. This discovers routing
1663 * changes resulting in larger path MTUs.
1665 * Returns 0 if we should wait to probe (no cwnd available),
1666 * 1 if a probe was sent,
1669 static int tcp_mtu_probe(struct sock *sk)
1671 struct tcp_sock *tp = tcp_sk(sk);
1672 struct inet_connection_sock *icsk = inet_csk(sk);
1673 struct sk_buff *skb, *nskb, *next;
1680 /* Not currently probing/verifying,
1682 * have enough cwnd, and
1683 * not SACKing (the variable headers throw things off) */
1684 if (!icsk->icsk_mtup.enabled ||
1685 icsk->icsk_mtup.probe_size ||
1686 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1687 tp->snd_cwnd < 11 ||
1688 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1691 /* Very simple search strategy: just double the MSS. */
1692 mss_now = tcp_current_mss(sk);
1693 probe_size = 2 * tp->mss_cache;
1694 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1695 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1696 /* TODO: set timer for probe_converge_event */
1700 /* Have enough data in the send queue to probe? */
1701 if (tp->write_seq - tp->snd_nxt < size_needed)
1704 if (tp->snd_wnd < size_needed)
1706 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1709 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1710 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1711 if (!tcp_packets_in_flight(tp))
1717 /* We're allowed to probe. Build it now. */
1718 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1720 sk->sk_wmem_queued += nskb->truesize;
1721 sk_mem_charge(sk, nskb->truesize);
1723 skb = tcp_send_head(sk);
1725 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1726 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1727 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1728 TCP_SKB_CB(nskb)->sacked = 0;
1730 nskb->ip_summed = skb->ip_summed;
1732 tcp_insert_write_queue_before(nskb, skb, sk);
1735 tcp_for_write_queue_from_safe(skb, next, sk) {
1736 copy = min_t(int, skb->len, probe_size - len);
1737 if (nskb->ip_summed)
1738 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1740 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1741 skb_put(nskb, copy),
1744 if (skb->len <= copy) {
1745 /* We've eaten all the data from this skb.
1747 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1748 tcp_unlink_write_queue(skb, sk);
1749 sk_wmem_free_skb(sk, skb);
1751 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1752 ~(TCPHDR_FIN|TCPHDR_PSH);
1753 if (!skb_shinfo(skb)->nr_frags) {
1754 skb_pull(skb, copy);
1755 if (skb->ip_summed != CHECKSUM_PARTIAL)
1756 skb->csum = csum_partial(skb->data,
1759 __pskb_trim_head(skb, copy);
1760 tcp_set_skb_tso_segs(sk, skb, mss_now);
1762 TCP_SKB_CB(skb)->seq += copy;
1767 if (len >= probe_size)
1770 tcp_init_tso_segs(sk, nskb, nskb->len);
1772 /* We're ready to send. If this fails, the probe will
1773 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1774 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1775 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1776 /* Decrement cwnd here because we are sending
1777 * effectively two packets. */
1779 tcp_event_new_data_sent(sk, nskb);
1781 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1782 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1783 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1791 /* This routine writes packets to the network. It advances the
1792 * send_head. This happens as incoming acks open up the remote
1795 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1796 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1797 * account rare use of URG, this is not a big flaw.
1799 * Send at most one packet when push_one > 0. Temporarily ignore
1800 * cwnd limit to force at most one packet out when push_one == 2.
1802 * Returns true, if no segments are in flight and we have queued segments,
1803 * but cannot send anything now because of SWS or another problem.
1805 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1806 int push_one, gfp_t gfp)
1808 struct tcp_sock *tp = tcp_sk(sk);
1809 struct sk_buff *skb;
1810 unsigned int tso_segs, sent_pkts;
1817 /* Do MTU probing. */
1818 result = tcp_mtu_probe(sk);
1821 } else if (result > 0) {
1826 while ((skb = tcp_send_head(sk))) {
1830 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1833 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1834 goto repair; /* Skip network transmission */
1836 cwnd_quota = tcp_cwnd_test(tp, skb);
1839 /* Force out a loss probe pkt. */
1845 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1848 if (tso_segs == 1) {
1849 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1850 (tcp_skb_is_last(sk, skb) ?
1851 nonagle : TCP_NAGLE_PUSH))))
1854 if (!push_one && tcp_tso_should_defer(sk, skb))
1858 /* TSQ : sk_wmem_alloc accounts skb truesize,
1859 * including skb overhead. But thats OK.
1861 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1862 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1866 if (tso_segs > 1 && !tcp_urg_mode(tp))
1867 limit = tcp_mss_split_point(sk, skb, mss_now,
1870 sk->sk_gso_max_segs));
1872 if (skb->len > limit &&
1873 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1876 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1878 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1882 /* Advance the send_head. This one is sent out.
1883 * This call will increment packets_out.
1885 tcp_event_new_data_sent(sk, skb);
1887 tcp_minshall_update(tp, mss_now, skb);
1888 sent_pkts += tcp_skb_pcount(skb);
1894 if (likely(sent_pkts)) {
1895 if (tcp_in_cwnd_reduction(sk))
1896 tp->prr_out += sent_pkts;
1898 /* Send one loss probe per tail loss episode. */
1900 tcp_schedule_loss_probe(sk);
1901 tcp_cwnd_validate(sk);
1904 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1907 bool tcp_schedule_loss_probe(struct sock *sk)
1909 struct inet_connection_sock *icsk = inet_csk(sk);
1910 struct tcp_sock *tp = tcp_sk(sk);
1911 u32 timeout, tlp_time_stamp, rto_time_stamp;
1912 u32 rtt = tp->srtt >> 3;
1914 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1916 /* No consecutive loss probes. */
1917 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1921 /* Don't do any loss probe on a Fast Open connection before 3WHS
1924 if (sk->sk_state == TCP_SYN_RECV)
1927 /* TLP is only scheduled when next timer event is RTO. */
1928 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1931 /* Schedule a loss probe in 2*RTT for SACK capable connections
1932 * in Open state, that are either limited by cwnd or application.
1934 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1935 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1938 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1942 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1943 * for delayed ack when there's one outstanding packet.
1946 if (tp->packets_out == 1)
1947 timeout = max_t(u32, timeout,
1948 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1949 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1951 /* If RTO is shorter, just schedule TLP in its place. */
1952 tlp_time_stamp = tcp_time_stamp + timeout;
1953 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1954 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1955 s32 delta = rto_time_stamp - tcp_time_stamp;
1960 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1965 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1966 * retransmit the last segment.
1968 void tcp_send_loss_probe(struct sock *sk)
1970 struct tcp_sock *tp = tcp_sk(sk);
1971 struct sk_buff *skb;
1973 int mss = tcp_current_mss(sk);
1976 if (tcp_send_head(sk) != NULL) {
1977 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1981 /* At most one outstanding TLP retransmission. */
1982 if (tp->tlp_high_seq)
1985 /* Retransmit last segment. */
1986 skb = tcp_write_queue_tail(sk);
1990 pcount = tcp_skb_pcount(skb);
1991 if (WARN_ON(!pcount))
1994 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
1995 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
1997 skb = tcp_write_queue_tail(sk);
2000 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2003 /* Probe with zero data doesn't trigger fast recovery. */
2005 err = __tcp_retransmit_skb(sk, skb);
2007 /* Record snd_nxt for loss detection. */
2009 tp->tlp_high_seq = tp->snd_nxt;
2012 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2013 inet_csk(sk)->icsk_rto,
2017 NET_INC_STATS_BH(sock_net(sk),
2018 LINUX_MIB_TCPLOSSPROBES);
2022 /* Push out any pending frames which were held back due to
2023 * TCP_CORK or attempt at coalescing tiny packets.
2024 * The socket must be locked by the caller.
2026 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2029 /* If we are closed, the bytes will have to remain here.
2030 * In time closedown will finish, we empty the write queue and
2031 * all will be happy.
2033 if (unlikely(sk->sk_state == TCP_CLOSE))
2036 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2037 sk_gfp_atomic(sk, GFP_ATOMIC)))
2038 tcp_check_probe_timer(sk);
2041 /* Send _single_ skb sitting at the send head. This function requires
2042 * true push pending frames to setup probe timer etc.
2044 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2046 struct sk_buff *skb = tcp_send_head(sk);
2048 BUG_ON(!skb || skb->len < mss_now);
2050 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2053 /* This function returns the amount that we can raise the
2054 * usable window based on the following constraints
2056 * 1. The window can never be shrunk once it is offered (RFC 793)
2057 * 2. We limit memory per socket
2060 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2061 * RECV.NEXT + RCV.WIN fixed until:
2062 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2064 * i.e. don't raise the right edge of the window until you can raise
2065 * it at least MSS bytes.
2067 * Unfortunately, the recommended algorithm breaks header prediction,
2068 * since header prediction assumes th->window stays fixed.
2070 * Strictly speaking, keeping th->window fixed violates the receiver
2071 * side SWS prevention criteria. The problem is that under this rule
2072 * a stream of single byte packets will cause the right side of the
2073 * window to always advance by a single byte.
2075 * Of course, if the sender implements sender side SWS prevention
2076 * then this will not be a problem.
2078 * BSD seems to make the following compromise:
2080 * If the free space is less than the 1/4 of the maximum
2081 * space available and the free space is less than 1/2 mss,
2082 * then set the window to 0.
2083 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2084 * Otherwise, just prevent the window from shrinking
2085 * and from being larger than the largest representable value.
2087 * This prevents incremental opening of the window in the regime
2088 * where TCP is limited by the speed of the reader side taking
2089 * data out of the TCP receive queue. It does nothing about
2090 * those cases where the window is constrained on the sender side
2091 * because the pipeline is full.
2093 * BSD also seems to "accidentally" limit itself to windows that are a
2094 * multiple of MSS, at least until the free space gets quite small.
2095 * This would appear to be a side effect of the mbuf implementation.
2096 * Combining these two algorithms results in the observed behavior
2097 * of having a fixed window size at almost all times.
2099 * Below we obtain similar behavior by forcing the offered window to
2100 * a multiple of the mss when it is feasible to do so.
2102 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2103 * Regular options like TIMESTAMP are taken into account.
2105 u32 __tcp_select_window(struct sock *sk)
2107 struct inet_connection_sock *icsk = inet_csk(sk);
2108 struct tcp_sock *tp = tcp_sk(sk);
2109 /* MSS for the peer's data. Previous versions used mss_clamp
2110 * here. I don't know if the value based on our guesses
2111 * of peer's MSS is better for the performance. It's more correct
2112 * but may be worse for the performance because of rcv_mss
2113 * fluctuations. --SAW 1998/11/1
2115 int mss = icsk->icsk_ack.rcv_mss;
2116 int free_space = tcp_space(sk);
2117 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2120 if (mss > full_space)
2123 if (free_space < (full_space >> 1)) {
2124 icsk->icsk_ack.quick = 0;
2126 if (sk_under_memory_pressure(sk))
2127 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2130 if (free_space < mss)
2134 if (free_space > tp->rcv_ssthresh)
2135 free_space = tp->rcv_ssthresh;
2137 /* Don't do rounding if we are using window scaling, since the
2138 * scaled window will not line up with the MSS boundary anyway.
2140 window = tp->rcv_wnd;
2141 if (tp->rx_opt.rcv_wscale) {
2142 window = free_space;
2144 /* Advertise enough space so that it won't get scaled away.
2145 * Import case: prevent zero window announcement if
2146 * 1<<rcv_wscale > mss.
2148 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2149 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2150 << tp->rx_opt.rcv_wscale);
2152 /* Get the largest window that is a nice multiple of mss.
2153 * Window clamp already applied above.
2154 * If our current window offering is within 1 mss of the
2155 * free space we just keep it. This prevents the divide
2156 * and multiply from happening most of the time.
2157 * We also don't do any window rounding when the free space
2160 if (window <= free_space - mss || window > free_space)
2161 window = (free_space / mss) * mss;
2162 else if (mss == full_space &&
2163 free_space > window + (full_space >> 1))
2164 window = free_space;
2170 /* Collapses two adjacent SKB's during retransmission. */
2171 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2173 struct tcp_sock *tp = tcp_sk(sk);
2174 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2175 int skb_size, next_skb_size;
2177 skb_size = skb->len;
2178 next_skb_size = next_skb->len;
2180 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2182 tcp_highest_sack_combine(sk, next_skb, skb);
2184 tcp_unlink_write_queue(next_skb, sk);
2186 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2189 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2190 skb->ip_summed = CHECKSUM_PARTIAL;
2192 if (skb->ip_summed != CHECKSUM_PARTIAL)
2193 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2195 /* Update sequence range on original skb. */
2196 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2198 /* Merge over control information. This moves PSH/FIN etc. over */
2199 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2201 /* All done, get rid of second SKB and account for it so
2202 * packet counting does not break.
2204 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2206 /* changed transmit queue under us so clear hints */
2207 tcp_clear_retrans_hints_partial(tp);
2208 if (next_skb == tp->retransmit_skb_hint)
2209 tp->retransmit_skb_hint = skb;
2211 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2213 sk_wmem_free_skb(sk, next_skb);
2216 /* Check if coalescing SKBs is legal. */
2217 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2219 if (tcp_skb_pcount(skb) > 1)
2221 /* TODO: SACK collapsing could be used to remove this condition */
2222 if (skb_shinfo(skb)->nr_frags != 0)
2224 if (skb_cloned(skb))
2226 if (skb == tcp_send_head(sk))
2228 /* Some heurestics for collapsing over SACK'd could be invented */
2229 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2235 /* Collapse packets in the retransmit queue to make to create
2236 * less packets on the wire. This is only done on retransmission.
2238 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2241 struct tcp_sock *tp = tcp_sk(sk);
2242 struct sk_buff *skb = to, *tmp;
2245 if (!sysctl_tcp_retrans_collapse)
2247 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2250 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2251 if (!tcp_can_collapse(sk, skb))
2263 /* Punt if not enough space exists in the first SKB for
2264 * the data in the second
2266 if (skb->len > skb_availroom(to))
2269 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2272 tcp_collapse_retrans(sk, to);
2276 /* This retransmits one SKB. Policy decisions and retransmit queue
2277 * state updates are done by the caller. Returns non-zero if an
2278 * error occurred which prevented the send.
2280 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2282 struct tcp_sock *tp = tcp_sk(sk);
2283 struct inet_connection_sock *icsk = inet_csk(sk);
2284 unsigned int cur_mss;
2286 /* Inconslusive MTU probe */
2287 if (icsk->icsk_mtup.probe_size) {
2288 icsk->icsk_mtup.probe_size = 0;
2291 /* Do not sent more than we queued. 1/4 is reserved for possible
2292 * copying overhead: fragmentation, tunneling, mangling etc.
2294 if (atomic_read(&sk->sk_wmem_alloc) >
2295 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2298 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2299 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2301 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2305 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2306 return -EHOSTUNREACH; /* Routing failure or similar. */
2308 cur_mss = tcp_current_mss(sk);
2310 /* If receiver has shrunk his window, and skb is out of
2311 * new window, do not retransmit it. The exception is the
2312 * case, when window is shrunk to zero. In this case
2313 * our retransmit serves as a zero window probe.
2315 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2316 TCP_SKB_CB(skb)->seq != tp->snd_una)
2319 if (skb->len > cur_mss) {
2320 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2321 return -ENOMEM; /* We'll try again later. */
2323 int oldpcount = tcp_skb_pcount(skb);
2325 if (unlikely(oldpcount > 1)) {
2326 tcp_init_tso_segs(sk, skb, cur_mss);
2327 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2331 tcp_retrans_try_collapse(sk, skb, cur_mss);
2333 /* Some Solaris stacks overoptimize and ignore the FIN on a
2334 * retransmit when old data is attached. So strip it off
2335 * since it is cheap to do so and saves bytes on the network.
2338 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2339 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2340 if (!pskb_trim(skb, 0)) {
2341 /* Reuse, even though it does some unnecessary work */
2342 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2343 TCP_SKB_CB(skb)->tcp_flags);
2344 skb->ip_summed = CHECKSUM_NONE;
2348 /* Make a copy, if the first transmission SKB clone we made
2349 * is still in somebody's hands, else make a clone.
2351 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2353 /* make sure skb->data is aligned on arches that require it */
2354 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2355 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2357 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2360 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2364 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2366 struct tcp_sock *tp = tcp_sk(sk);
2367 int err = __tcp_retransmit_skb(sk, skb);
2370 /* Update global TCP statistics. */
2371 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2373 tp->total_retrans++;
2375 #if FASTRETRANS_DEBUG > 0
2376 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2377 net_dbg_ratelimited("retrans_out leaked\n");
2380 if (!tp->retrans_out)
2381 tp->lost_retrans_low = tp->snd_nxt;
2382 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2383 tp->retrans_out += tcp_skb_pcount(skb);
2385 /* Save stamp of the first retransmit. */
2386 if (!tp->retrans_stamp)
2387 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2389 tp->undo_retrans += tcp_skb_pcount(skb);
2391 /* snd_nxt is stored to detect loss of retransmitted segment,
2392 * see tcp_input.c tcp_sacktag_write_queue().
2394 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2399 /* Check if we forward retransmits are possible in the current
2400 * window/congestion state.
2402 static bool tcp_can_forward_retransmit(struct sock *sk)
2404 const struct inet_connection_sock *icsk = inet_csk(sk);
2405 const struct tcp_sock *tp = tcp_sk(sk);
2407 /* Forward retransmissions are possible only during Recovery. */
2408 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2411 /* No forward retransmissions in Reno are possible. */
2412 if (tcp_is_reno(tp))
2415 /* Yeah, we have to make difficult choice between forward transmission
2416 * and retransmission... Both ways have their merits...
2418 * For now we do not retransmit anything, while we have some new
2419 * segments to send. In the other cases, follow rule 3 for
2420 * NextSeg() specified in RFC3517.
2423 if (tcp_may_send_now(sk))
2429 /* This gets called after a retransmit timeout, and the initially
2430 * retransmitted data is acknowledged. It tries to continue
2431 * resending the rest of the retransmit queue, until either
2432 * we've sent it all or the congestion window limit is reached.
2433 * If doing SACK, the first ACK which comes back for a timeout
2434 * based retransmit packet might feed us FACK information again.
2435 * If so, we use it to avoid unnecessarily retransmissions.
2437 void tcp_xmit_retransmit_queue(struct sock *sk)
2439 const struct inet_connection_sock *icsk = inet_csk(sk);
2440 struct tcp_sock *tp = tcp_sk(sk);
2441 struct sk_buff *skb;
2442 struct sk_buff *hole = NULL;
2445 int fwd_rexmitting = 0;
2447 if (!tp->packets_out)
2451 tp->retransmit_high = tp->snd_una;
2453 if (tp->retransmit_skb_hint) {
2454 skb = tp->retransmit_skb_hint;
2455 last_lost = TCP_SKB_CB(skb)->end_seq;
2456 if (after(last_lost, tp->retransmit_high))
2457 last_lost = tp->retransmit_high;
2459 skb = tcp_write_queue_head(sk);
2460 last_lost = tp->snd_una;
2463 tcp_for_write_queue_from(skb, sk) {
2464 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2466 if (skb == tcp_send_head(sk))
2468 /* we could do better than to assign each time */
2470 tp->retransmit_skb_hint = skb;
2472 /* Assume this retransmit will generate
2473 * only one packet for congestion window
2474 * calculation purposes. This works because
2475 * tcp_retransmit_skb() will chop up the
2476 * packet to be MSS sized and all the
2477 * packet counting works out.
2479 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2482 if (fwd_rexmitting) {
2484 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2486 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2488 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2489 tp->retransmit_high = last_lost;
2490 if (!tcp_can_forward_retransmit(sk))
2492 /* Backtrack if necessary to non-L'ed skb */
2500 } else if (!(sacked & TCPCB_LOST)) {
2501 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2506 last_lost = TCP_SKB_CB(skb)->end_seq;
2507 if (icsk->icsk_ca_state != TCP_CA_Loss)
2508 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2510 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2513 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2516 if (tcp_retransmit_skb(sk, skb)) {
2517 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2520 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2522 if (tcp_in_cwnd_reduction(sk))
2523 tp->prr_out += tcp_skb_pcount(skb);
2525 if (skb == tcp_write_queue_head(sk))
2526 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2527 inet_csk(sk)->icsk_rto,
2532 /* Send a fin. The caller locks the socket for us. This cannot be
2533 * allowed to fail queueing a FIN frame under any circumstances.
2535 void tcp_send_fin(struct sock *sk)
2537 struct tcp_sock *tp = tcp_sk(sk);
2538 struct sk_buff *skb = tcp_write_queue_tail(sk);
2541 /* Optimization, tack on the FIN if we have a queue of
2542 * unsent frames. But be careful about outgoing SACKS
2545 mss_now = tcp_current_mss(sk);
2547 if (tcp_send_head(sk) != NULL) {
2548 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2549 TCP_SKB_CB(skb)->end_seq++;
2552 /* Socket is locked, keep trying until memory is available. */
2554 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2561 /* Reserve space for headers and prepare control bits. */
2562 skb_reserve(skb, MAX_TCP_HEADER);
2563 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2564 tcp_init_nondata_skb(skb, tp->write_seq,
2565 TCPHDR_ACK | TCPHDR_FIN);
2566 tcp_queue_skb(sk, skb);
2568 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2571 /* We get here when a process closes a file descriptor (either due to
2572 * an explicit close() or as a byproduct of exit()'ing) and there
2573 * was unread data in the receive queue. This behavior is recommended
2574 * by RFC 2525, section 2.17. -DaveM
2576 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2578 struct sk_buff *skb;
2580 /* NOTE: No TCP options attached and we never retransmit this. */
2581 skb = alloc_skb(MAX_TCP_HEADER, priority);
2583 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2587 /* Reserve space for headers and prepare control bits. */
2588 skb_reserve(skb, MAX_TCP_HEADER);
2589 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2590 TCPHDR_ACK | TCPHDR_RST);
2592 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2593 if (tcp_transmit_skb(sk, skb, 0, priority))
2594 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2596 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2599 /* Send a crossed SYN-ACK during socket establishment.
2600 * WARNING: This routine must only be called when we have already sent
2601 * a SYN packet that crossed the incoming SYN that caused this routine
2602 * to get called. If this assumption fails then the initial rcv_wnd
2603 * and rcv_wscale values will not be correct.
2605 int tcp_send_synack(struct sock *sk)
2607 struct sk_buff *skb;
2609 skb = tcp_write_queue_head(sk);
2610 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2611 pr_debug("%s: wrong queue state\n", __func__);
2614 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2615 if (skb_cloned(skb)) {
2616 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2619 tcp_unlink_write_queue(skb, sk);
2620 skb_header_release(nskb);
2621 __tcp_add_write_queue_head(sk, nskb);
2622 sk_wmem_free_skb(sk, skb);
2623 sk->sk_wmem_queued += nskb->truesize;
2624 sk_mem_charge(sk, nskb->truesize);
2628 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2629 TCP_ECN_send_synack(tcp_sk(sk), skb);
2631 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2632 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2636 * tcp_make_synack - Prepare a SYN-ACK.
2637 * sk: listener socket
2638 * dst: dst entry attached to the SYNACK
2639 * req: request_sock pointer
2641 * Allocate one skb and build a SYNACK packet.
2642 * @dst is consumed : Caller should not use it again.
2644 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2645 struct request_sock *req,
2646 struct tcp_fastopen_cookie *foc)
2648 struct tcp_out_options opts;
2649 struct inet_request_sock *ireq = inet_rsk(req);
2650 struct tcp_sock *tp = tcp_sk(sk);
2652 struct sk_buff *skb;
2653 struct tcp_md5sig_key *md5;
2654 int tcp_header_size;
2657 skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC));
2658 if (unlikely(!skb)) {
2662 /* Reserve space for headers. */
2663 skb_reserve(skb, MAX_TCP_HEADER);
2665 skb_dst_set(skb, dst);
2667 mss = dst_metric_advmss(dst);
2668 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2669 mss = tp->rx_opt.user_mss;
2671 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2673 /* Set this up on the first call only */
2674 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2676 /* limit the window selection if the user enforce a smaller rx buffer */
2677 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2678 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2679 req->window_clamp = tcp_full_space(sk);
2681 /* tcp_full_space because it is guaranteed to be the first packet */
2682 tcp_select_initial_window(tcp_full_space(sk),
2683 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2688 dst_metric(dst, RTAX_INITRWND));
2689 ireq->rcv_wscale = rcv_wscale;
2692 memset(&opts, 0, sizeof(opts));
2693 #ifdef CONFIG_SYN_COOKIES
2694 if (unlikely(req->cookie_ts))
2695 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2698 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2699 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2702 skb_push(skb, tcp_header_size);
2703 skb_reset_transport_header(skb);
2706 memset(th, 0, sizeof(struct tcphdr));
2709 TCP_ECN_make_synack(req, th);
2710 th->source = ireq->loc_port;
2711 th->dest = ireq->rmt_port;
2712 /* Setting of flags are superfluous here for callers (and ECE is
2713 * not even correctly set)
2715 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2716 TCPHDR_SYN | TCPHDR_ACK);
2718 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2719 /* XXX data is queued and acked as is. No buffer/window check */
2720 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2722 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2723 th->window = htons(min(req->rcv_wnd, 65535U));
2724 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2725 th->doff = (tcp_header_size >> 2);
2726 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2728 #ifdef CONFIG_TCP_MD5SIG
2729 /* Okay, we have all we need - do the md5 hash if needed */
2731 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2732 md5, NULL, req, skb);
2738 EXPORT_SYMBOL(tcp_make_synack);
2740 /* Do all connect socket setups that can be done AF independent. */
2741 void tcp_connect_init(struct sock *sk)
2743 const struct dst_entry *dst = __sk_dst_get(sk);
2744 struct tcp_sock *tp = tcp_sk(sk);
2747 /* We'll fix this up when we get a response from the other end.
2748 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2750 tp->tcp_header_len = sizeof(struct tcphdr) +
2751 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2753 #ifdef CONFIG_TCP_MD5SIG
2754 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2755 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2758 /* If user gave his TCP_MAXSEG, record it to clamp */
2759 if (tp->rx_opt.user_mss)
2760 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2763 tcp_sync_mss(sk, dst_mtu(dst));
2765 if (!tp->window_clamp)
2766 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2767 tp->advmss = dst_metric_advmss(dst);
2768 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2769 tp->advmss = tp->rx_opt.user_mss;
2771 tcp_initialize_rcv_mss(sk);
2773 /* limit the window selection if the user enforce a smaller rx buffer */
2774 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2775 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2776 tp->window_clamp = tcp_full_space(sk);
2778 tcp_select_initial_window(tcp_full_space(sk),
2779 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2782 sysctl_tcp_window_scaling,
2784 dst_metric(dst, RTAX_INITRWND));
2786 tp->rx_opt.rcv_wscale = rcv_wscale;
2787 tp->rcv_ssthresh = tp->rcv_wnd;
2790 sock_reset_flag(sk, SOCK_DONE);
2793 tp->snd_una = tp->write_seq;
2794 tp->snd_sml = tp->write_seq;
2795 tp->snd_up = tp->write_seq;
2796 tp->snd_nxt = tp->write_seq;
2798 if (likely(!tp->repair))
2800 tp->rcv_wup = tp->rcv_nxt;
2801 tp->copied_seq = tp->rcv_nxt;
2803 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2804 inet_csk(sk)->icsk_retransmits = 0;
2805 tcp_clear_retrans(tp);
2808 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2810 struct tcp_sock *tp = tcp_sk(sk);
2811 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2813 tcb->end_seq += skb->len;
2814 skb_header_release(skb);
2815 __tcp_add_write_queue_tail(sk, skb);
2816 sk->sk_wmem_queued += skb->truesize;
2817 sk_mem_charge(sk, skb->truesize);
2818 tp->write_seq = tcb->end_seq;
2819 tp->packets_out += tcp_skb_pcount(skb);
2822 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2823 * queue a data-only packet after the regular SYN, such that regular SYNs
2824 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2825 * only the SYN sequence, the data are retransmitted in the first ACK.
2826 * If cookie is not cached or other error occurs, falls back to send a
2827 * regular SYN with Fast Open cookie request option.
2829 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2831 struct tcp_sock *tp = tcp_sk(sk);
2832 struct tcp_fastopen_request *fo = tp->fastopen_req;
2833 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2834 struct sk_buff *syn_data = NULL, *data;
2835 unsigned long last_syn_loss = 0;
2837 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2838 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2839 &syn_loss, &last_syn_loss);
2840 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2842 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2843 fo->cookie.len = -1;
2847 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2848 fo->cookie.len = -1;
2849 else if (fo->cookie.len <= 0)
2852 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2853 * user-MSS. Reserve maximum option space for middleboxes that add
2854 * private TCP options. The cost is reduced data space in SYN :(
2856 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2857 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2858 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2859 MAX_TCP_OPTION_SPACE;
2861 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2863 if (syn_data == NULL)
2866 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2867 struct iovec *iov = &fo->data->msg_iov[i];
2868 unsigned char __user *from = iov->iov_base;
2869 int len = iov->iov_len;
2871 if (syn_data->len + len > space)
2872 len = space - syn_data->len;
2873 else if (i + 1 == iovlen)
2874 /* No more data pending in inet_wait_for_connect() */
2877 if (skb_add_data(syn_data, from, len))
2881 /* Queue a data-only packet after the regular SYN for retransmission */
2882 data = pskb_copy(syn_data, sk->sk_allocation);
2885 TCP_SKB_CB(data)->seq++;
2886 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2887 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2888 tcp_connect_queue_skb(sk, data);
2889 fo->copied = data->len;
2891 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2892 tp->syn_data = (fo->copied > 0);
2893 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2899 /* Send a regular SYN with Fast Open cookie request option */
2900 if (fo->cookie.len > 0)
2902 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2904 tp->syn_fastopen = 0;
2905 kfree_skb(syn_data);
2907 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2911 /* Build a SYN and send it off. */
2912 int tcp_connect(struct sock *sk)
2914 struct tcp_sock *tp = tcp_sk(sk);
2915 struct sk_buff *buff;
2918 tcp_connect_init(sk);
2920 if (unlikely(tp->repair)) {
2921 tcp_finish_connect(sk, NULL);
2925 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2926 if (unlikely(buff == NULL))
2929 /* Reserve space for headers. */
2930 skb_reserve(buff, MAX_TCP_HEADER);
2932 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2933 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2934 tcp_connect_queue_skb(sk, buff);
2935 TCP_ECN_send_syn(sk, buff);
2937 /* Send off SYN; include data in Fast Open. */
2938 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2939 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2940 if (err == -ECONNREFUSED)
2943 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2944 * in order to make this packet get counted in tcpOutSegs.
2946 tp->snd_nxt = tp->write_seq;
2947 tp->pushed_seq = tp->write_seq;
2948 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2950 /* Timer for repeating the SYN until an answer. */
2951 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2952 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2955 EXPORT_SYMBOL(tcp_connect);
2957 /* Send out a delayed ack, the caller does the policy checking
2958 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2961 void tcp_send_delayed_ack(struct sock *sk)
2963 struct inet_connection_sock *icsk = inet_csk(sk);
2964 int ato = icsk->icsk_ack.ato;
2965 unsigned long timeout;
2967 if (ato > TCP_DELACK_MIN) {
2968 const struct tcp_sock *tp = tcp_sk(sk);
2969 int max_ato = HZ / 2;
2971 if (icsk->icsk_ack.pingpong ||
2972 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2973 max_ato = TCP_DELACK_MAX;
2975 /* Slow path, intersegment interval is "high". */
2977 /* If some rtt estimate is known, use it to bound delayed ack.
2978 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2982 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2988 ato = min(ato, max_ato);
2991 /* Stay within the limit we were given */
2992 timeout = jiffies + ato;
2994 /* Use new timeout only if there wasn't a older one earlier. */
2995 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2996 /* If delack timer was blocked or is about to expire,
2999 if (icsk->icsk_ack.blocked ||
3000 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3005 if (!time_before(timeout, icsk->icsk_ack.timeout))
3006 timeout = icsk->icsk_ack.timeout;
3008 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3009 icsk->icsk_ack.timeout = timeout;
3010 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3013 /* This routine sends an ack and also updates the window. */
3014 void tcp_send_ack(struct sock *sk)
3016 struct sk_buff *buff;
3018 /* If we have been reset, we may not send again. */
3019 if (sk->sk_state == TCP_CLOSE)
3022 /* We are not putting this on the write queue, so
3023 * tcp_transmit_skb() will set the ownership to this
3026 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3028 inet_csk_schedule_ack(sk);
3029 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3030 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3031 TCP_DELACK_MAX, TCP_RTO_MAX);
3035 /* Reserve space for headers and prepare control bits. */
3036 skb_reserve(buff, MAX_TCP_HEADER);
3037 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3039 /* Send it off, this clears delayed acks for us. */
3040 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3041 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3044 /* This routine sends a packet with an out of date sequence
3045 * number. It assumes the other end will try to ack it.
3047 * Question: what should we make while urgent mode?
3048 * 4.4BSD forces sending single byte of data. We cannot send
3049 * out of window data, because we have SND.NXT==SND.MAX...
3051 * Current solution: to send TWO zero-length segments in urgent mode:
3052 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3053 * out-of-date with SND.UNA-1 to probe window.
3055 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3057 struct tcp_sock *tp = tcp_sk(sk);
3058 struct sk_buff *skb;
3060 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3061 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3065 /* Reserve space for headers and set control bits. */
3066 skb_reserve(skb, MAX_TCP_HEADER);
3067 /* Use a previous sequence. This should cause the other
3068 * end to send an ack. Don't queue or clone SKB, just
3071 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3072 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3073 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3076 void tcp_send_window_probe(struct sock *sk)
3078 if (sk->sk_state == TCP_ESTABLISHED) {
3079 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3080 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3081 tcp_xmit_probe_skb(sk, 0);
3085 /* Initiate keepalive or window probe from timer. */
3086 int tcp_write_wakeup(struct sock *sk)
3088 struct tcp_sock *tp = tcp_sk(sk);
3089 struct sk_buff *skb;
3091 if (sk->sk_state == TCP_CLOSE)
3094 if ((skb = tcp_send_head(sk)) != NULL &&
3095 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3097 unsigned int mss = tcp_current_mss(sk);
3098 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3100 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3101 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3103 /* We are probing the opening of a window
3104 * but the window size is != 0
3105 * must have been a result SWS avoidance ( sender )
3107 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3109 seg_size = min(seg_size, mss);
3110 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3111 if (tcp_fragment(sk, skb, seg_size, mss))
3113 } else if (!tcp_skb_pcount(skb))
3114 tcp_set_skb_tso_segs(sk, skb, mss);
3116 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3117 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3118 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3120 tcp_event_new_data_sent(sk, skb);
3123 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3124 tcp_xmit_probe_skb(sk, 1);
3125 return tcp_xmit_probe_skb(sk, 0);
3129 /* A window probe timeout has occurred. If window is not closed send
3130 * a partial packet else a zero probe.
3132 void tcp_send_probe0(struct sock *sk)
3134 struct inet_connection_sock *icsk = inet_csk(sk);
3135 struct tcp_sock *tp = tcp_sk(sk);
3138 err = tcp_write_wakeup(sk);
3140 if (tp->packets_out || !tcp_send_head(sk)) {
3141 /* Cancel probe timer, if it is not required. */
3142 icsk->icsk_probes_out = 0;
3143 icsk->icsk_backoff = 0;
3148 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3149 icsk->icsk_backoff++;
3150 icsk->icsk_probes_out++;
3151 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3152 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3155 /* If packet was not sent due to local congestion,
3156 * do not backoff and do not remember icsk_probes_out.
3157 * Let local senders to fight for local resources.
3159 * Use accumulated backoff yet.
3161 if (!icsk->icsk_probes_out)
3162 icsk->icsk_probes_out = 1;
3163 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3164 min(icsk->icsk_rto << icsk->icsk_backoff,
3165 TCP_RESOURCE_PROBE_INTERVAL),