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 tp->packets_out += tcp_skb_pcount(skb);
82 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
83 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
87 /* SND.NXT, if window was not shrunk.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 const struct tcp_sock *tp = tcp_sk(sk);
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
100 return tcp_wnd_end(tp);
103 /* Calculate mss to advertise in SYN segment.
104 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 * 1. It is independent of path mtu.
107 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
108 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
109 * attached devices, because some buggy hosts are confused by
111 * 4. We do not make 3, we advertise MSS, calculated from first
112 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
113 * This may be overridden via information stored in routing table.
114 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
115 * probably even Jumbo".
117 static __u16 tcp_advertise_mss(struct sock *sk)
119 struct tcp_sock *tp = tcp_sk(sk);
120 const struct dst_entry *dst = __sk_dst_get(sk);
121 int mss = tp->advmss;
124 unsigned int metric = dst_metric_advmss(dst);
135 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
136 * This is the first part of cwnd validation mechanism. */
137 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
139 struct tcp_sock *tp = tcp_sk(sk);
140 s32 delta = tcp_time_stamp - tp->lsndtime;
141 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
142 u32 cwnd = tp->snd_cwnd;
144 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146 tp->snd_ssthresh = tcp_current_ssthresh(sk);
147 restart_cwnd = min(restart_cwnd, cwnd);
149 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 tp->snd_cwnd = max(cwnd, restart_cwnd);
152 tp->snd_cwnd_stamp = tcp_time_stamp;
153 tp->snd_cwnd_used = 0;
156 /* Congestion state accounting after a packet has been sent. */
157 static void tcp_event_data_sent(struct tcp_sock *tp,
160 struct inet_connection_sock *icsk = inet_csk(sk);
161 const u32 now = tcp_time_stamp;
163 if (sysctl_tcp_slow_start_after_idle &&
164 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
165 tcp_cwnd_restart(sk, __sk_dst_get(sk));
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
183 /* Determine a window scaling and initial window to offer.
184 * Based on the assumption that the given amount of space
185 * will be offered. Store the results in the tp structure.
186 * NOTE: for smooth operation initial space offering should
187 * be a multiple of mss if possible. We assume here that mss >= 1.
188 * This MUST be enforced by all callers.
190 void tcp_select_initial_window(int __space, __u32 mss,
191 __u32 *rcv_wnd, __u32 *window_clamp,
192 int wscale_ok, __u8 *rcv_wscale,
195 unsigned int space = (__space < 0 ? 0 : __space);
197 /* If no clamp set the clamp to the max possible scaled window */
198 if (*window_clamp == 0)
199 (*window_clamp) = (65535 << 14);
200 space = min(*window_clamp, space);
202 /* Quantize space offering to a multiple of mss if possible. */
204 space = (space / mss) * mss;
206 /* NOTE: offering an initial window larger than 32767
207 * will break some buggy TCP stacks. If the admin tells us
208 * it is likely we could be speaking with such a buggy stack
209 * we will truncate our initial window offering to 32K-1
210 * unless the remote has sent us a window scaling option,
211 * which we interpret as a sign the remote TCP is not
212 * misinterpreting the window field as a signed quantity.
214 if (sysctl_tcp_workaround_signed_windows)
215 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
221 /* Set window scaling on max possible window
222 * See RFC1323 for an explanation of the limit to 14
224 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
225 space = min_t(u32, space, *window_clamp);
226 while (space > 65535 && (*rcv_wscale) < 14) {
232 /* Set initial window to a value enough for senders starting with
233 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
234 * a limit on the initial window when mss is larger than 1460.
236 if (mss > (1 << *rcv_wscale)) {
237 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
240 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
241 /* when initializing use the value from init_rcv_wnd
242 * rather than the default from above
245 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
247 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
253 EXPORT_SYMBOL(tcp_select_initial_window);
255 /* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
260 static u16 tcp_select_window(struct sock *sk)
262 struct tcp_sock *tp = tcp_sk(sk);
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
291 /* If we advertise zero window, disable fast path. */
298 /* Packet ECN state for a SYN-ACK */
299 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
301 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
302 if (!(tp->ecn_flags & TCP_ECN_OK))
303 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
306 /* Packet ECN state for a SYN. */
307 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
309 struct tcp_sock *tp = tcp_sk(sk);
312 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
313 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
314 tp->ecn_flags = TCP_ECN_OK;
318 static __inline__ void
319 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
321 if (inet_rsk(req)->ecn_ok)
325 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
328 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
331 struct tcp_sock *tp = tcp_sk(sk);
333 if (tp->ecn_flags & TCP_ECN_OK) {
334 /* Not-retransmitted data segment: set ECT and inject CWR. */
335 if (skb->len != tcp_header_len &&
336 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
338 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
339 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
340 tcp_hdr(skb)->cwr = 1;
341 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
344 /* ACK or retransmitted segment: clear ECT|CE */
345 INET_ECN_dontxmit(sk);
347 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
348 tcp_hdr(skb)->ece = 1;
352 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
353 * auto increment end seqno.
355 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
357 skb->ip_summed = CHECKSUM_PARTIAL;
360 TCP_SKB_CB(skb)->tcp_flags = flags;
361 TCP_SKB_CB(skb)->sacked = 0;
363 skb_shinfo(skb)->gso_segs = 1;
364 skb_shinfo(skb)->gso_size = 0;
365 skb_shinfo(skb)->gso_type = 0;
367 TCP_SKB_CB(skb)->seq = seq;
368 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
370 TCP_SKB_CB(skb)->end_seq = seq;
373 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
375 return tp->snd_una != tp->snd_up;
378 #define OPTION_SACK_ADVERTISE (1 << 0)
379 #define OPTION_TS (1 << 1)
380 #define OPTION_MD5 (1 << 2)
381 #define OPTION_WSCALE (1 << 3)
382 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
384 struct tcp_out_options {
385 u16 options; /* bit field of OPTION_* */
386 u16 mss; /* 0 to disable */
387 u8 ws; /* window scale, 0 to disable */
388 u8 num_sack_blocks; /* number of SACK blocks to include */
389 u8 hash_size; /* bytes in hash_location */
390 __u8 *hash_location; /* temporary pointer, overloaded */
391 __u32 tsval, tsecr; /* need to include OPTION_TS */
392 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
395 /* Write previously computed TCP options to the packet.
397 * Beware: Something in the Internet is very sensitive to the ordering of
398 * TCP options, we learned this through the hard way, so be careful here.
399 * Luckily we can at least blame others for their non-compliance but from
400 * inter-operatibility perspective it seems that we're somewhat stuck with
401 * the ordering which we have been using if we want to keep working with
402 * those broken things (not that it currently hurts anybody as there isn't
403 * particular reason why the ordering would need to be changed).
405 * At least SACK_PERM as the first option is known to lead to a disaster
406 * (but it may well be that other scenarios fail similarly).
408 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
409 struct tcp_out_options *opts)
411 u16 options = opts->options; /* mungable copy */
413 if (unlikely(OPTION_MD5 & options)) {
414 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
415 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
416 /* overload cookie hash location */
417 opts->hash_location = (__u8 *)ptr;
421 if (unlikely(opts->mss)) {
422 *ptr++ = htonl((TCPOPT_MSS << 24) |
423 (TCPOLEN_MSS << 16) |
427 if (likely(OPTION_TS & options)) {
428 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
429 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
430 (TCPOLEN_SACK_PERM << 16) |
431 (TCPOPT_TIMESTAMP << 8) |
433 options &= ~OPTION_SACK_ADVERTISE;
435 *ptr++ = htonl((TCPOPT_NOP << 24) |
437 (TCPOPT_TIMESTAMP << 8) |
440 *ptr++ = htonl(opts->tsval);
441 *ptr++ = htonl(opts->tsecr);
444 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
445 *ptr++ = htonl((TCPOPT_NOP << 24) |
447 (TCPOPT_SACK_PERM << 8) |
451 if (unlikely(OPTION_WSCALE & options)) {
452 *ptr++ = htonl((TCPOPT_NOP << 24) |
453 (TCPOPT_WINDOW << 16) |
454 (TCPOLEN_WINDOW << 8) |
458 if (unlikely(opts->num_sack_blocks)) {
459 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
460 tp->duplicate_sack : tp->selective_acks;
463 *ptr++ = htonl((TCPOPT_NOP << 24) |
466 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
467 TCPOLEN_SACK_PERBLOCK)));
469 for (this_sack = 0; this_sack < opts->num_sack_blocks;
471 *ptr++ = htonl(sp[this_sack].start_seq);
472 *ptr++ = htonl(sp[this_sack].end_seq);
475 tp->rx_opt.dsack = 0;
478 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
479 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
481 *ptr++ = htonl((TCPOPT_EXP << 24) |
482 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
483 TCPOPT_FASTOPEN_MAGIC);
485 memcpy(ptr, foc->val, foc->len);
486 if ((foc->len & 3) == 2) {
487 u8 *align = ((u8 *)ptr) + foc->len;
488 align[0] = align[1] = TCPOPT_NOP;
490 ptr += (foc->len + 3) >> 2;
494 /* Compute TCP options for SYN packets. This is not the final
495 * network wire format yet.
497 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
498 struct tcp_out_options *opts,
499 struct tcp_md5sig_key **md5)
501 struct tcp_sock *tp = tcp_sk(sk);
502 unsigned int remaining = MAX_TCP_OPTION_SPACE;
503 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
505 #ifdef CONFIG_TCP_MD5SIG
506 *md5 = tp->af_specific->md5_lookup(sk, sk);
508 opts->options |= OPTION_MD5;
509 remaining -= TCPOLEN_MD5SIG_ALIGNED;
515 /* We always get an MSS option. The option bytes which will be seen in
516 * normal data packets should timestamps be used, must be in the MSS
517 * advertised. But we subtract them from tp->mss_cache so that
518 * calculations in tcp_sendmsg are simpler etc. So account for this
519 * fact here if necessary. If we don't do this correctly, as a
520 * receiver we won't recognize data packets as being full sized when we
521 * should, and thus we won't abide by the delayed ACK rules correctly.
522 * SACKs don't matter, we never delay an ACK when we have any of those
524 opts->mss = tcp_advertise_mss(sk);
525 remaining -= TCPOLEN_MSS_ALIGNED;
527 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
528 opts->options |= OPTION_TS;
529 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
530 opts->tsecr = tp->rx_opt.ts_recent;
531 remaining -= TCPOLEN_TSTAMP_ALIGNED;
533 if (likely(sysctl_tcp_window_scaling)) {
534 opts->ws = tp->rx_opt.rcv_wscale;
535 opts->options |= OPTION_WSCALE;
536 remaining -= TCPOLEN_WSCALE_ALIGNED;
538 if (likely(sysctl_tcp_sack)) {
539 opts->options |= OPTION_SACK_ADVERTISE;
540 if (unlikely(!(OPTION_TS & opts->options)))
541 remaining -= TCPOLEN_SACKPERM_ALIGNED;
544 if (fastopen && fastopen->cookie.len >= 0) {
545 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
546 need = (need + 3) & ~3U; /* Align to 32 bits */
547 if (remaining >= need) {
548 opts->options |= OPTION_FAST_OPEN_COOKIE;
549 opts->fastopen_cookie = &fastopen->cookie;
551 tp->syn_fastopen = 1;
555 return MAX_TCP_OPTION_SPACE - remaining;
558 /* Set up TCP options for SYN-ACKs. */
559 static unsigned int tcp_synack_options(struct sock *sk,
560 struct request_sock *req,
561 unsigned int mss, struct sk_buff *skb,
562 struct tcp_out_options *opts,
563 struct tcp_md5sig_key **md5,
564 struct tcp_fastopen_cookie *foc)
566 struct inet_request_sock *ireq = inet_rsk(req);
567 unsigned int remaining = MAX_TCP_OPTION_SPACE;
569 #ifdef CONFIG_TCP_MD5SIG
570 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
572 opts->options |= OPTION_MD5;
573 remaining -= TCPOLEN_MD5SIG_ALIGNED;
575 /* We can't fit any SACK blocks in a packet with MD5 + TS
576 * options. There was discussion about disabling SACK
577 * rather than TS in order to fit in better with old,
578 * buggy kernels, but that was deemed to be unnecessary.
580 ireq->tstamp_ok &= !ireq->sack_ok;
586 /* We always send an MSS option. */
588 remaining -= TCPOLEN_MSS_ALIGNED;
590 if (likely(ireq->wscale_ok)) {
591 opts->ws = ireq->rcv_wscale;
592 opts->options |= OPTION_WSCALE;
593 remaining -= TCPOLEN_WSCALE_ALIGNED;
595 if (likely(ireq->tstamp_ok)) {
596 opts->options |= OPTION_TS;
597 opts->tsval = TCP_SKB_CB(skb)->when;
598 opts->tsecr = req->ts_recent;
599 remaining -= TCPOLEN_TSTAMP_ALIGNED;
601 if (likely(ireq->sack_ok)) {
602 opts->options |= OPTION_SACK_ADVERTISE;
603 if (unlikely(!ireq->tstamp_ok))
604 remaining -= TCPOLEN_SACKPERM_ALIGNED;
607 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
608 need = (need + 3) & ~3U; /* Align to 32 bits */
609 if (remaining >= need) {
610 opts->options |= OPTION_FAST_OPEN_COOKIE;
611 opts->fastopen_cookie = foc;
616 return MAX_TCP_OPTION_SPACE - remaining;
619 /* Compute TCP options for ESTABLISHED sockets. This is not the
620 * final wire format yet.
622 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
623 struct tcp_out_options *opts,
624 struct tcp_md5sig_key **md5)
626 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
627 struct tcp_sock *tp = tcp_sk(sk);
628 unsigned int size = 0;
629 unsigned int eff_sacks;
631 #ifdef CONFIG_TCP_MD5SIG
632 *md5 = tp->af_specific->md5_lookup(sk, sk);
633 if (unlikely(*md5)) {
634 opts->options |= OPTION_MD5;
635 size += TCPOLEN_MD5SIG_ALIGNED;
641 if (likely(tp->rx_opt.tstamp_ok)) {
642 opts->options |= OPTION_TS;
643 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
644 opts->tsecr = tp->rx_opt.ts_recent;
645 size += TCPOLEN_TSTAMP_ALIGNED;
648 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
649 if (unlikely(eff_sacks)) {
650 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
651 opts->num_sack_blocks =
652 min_t(unsigned int, eff_sacks,
653 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
654 TCPOLEN_SACK_PERBLOCK);
655 size += TCPOLEN_SACK_BASE_ALIGNED +
656 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
663 /* TCP SMALL QUEUES (TSQ)
665 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
666 * to reduce RTT and bufferbloat.
667 * We do this using a special skb destructor (tcp_wfree).
669 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
670 * needs to be reallocated in a driver.
671 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
673 * Since transmit from skb destructor is forbidden, we use a tasklet
674 * to process all sockets that eventually need to send more skbs.
675 * We use one tasklet per cpu, with its own queue of sockets.
678 struct tasklet_struct tasklet;
679 struct list_head head; /* queue of tcp sockets */
681 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
683 static void tcp_tsq_handler(struct sock *sk)
685 if ((1 << sk->sk_state) &
686 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
687 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
688 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
691 * One tasklest per cpu tries to send more skbs.
692 * We run in tasklet context but need to disable irqs when
693 * transfering tsq->head because tcp_wfree() might
694 * interrupt us (non NAPI drivers)
696 static void tcp_tasklet_func(unsigned long data)
698 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
701 struct list_head *q, *n;
705 local_irq_save(flags);
706 list_splice_init(&tsq->head, &list);
707 local_irq_restore(flags);
709 list_for_each_safe(q, n, &list) {
710 tp = list_entry(q, struct tcp_sock, tsq_node);
711 list_del(&tp->tsq_node);
713 sk = (struct sock *)tp;
716 if (!sock_owned_by_user(sk)) {
719 /* defer the work to tcp_release_cb() */
720 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
724 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
729 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
730 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
731 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
732 (1UL << TCP_MTU_REDUCED_DEFERRED))
734 * tcp_release_cb - tcp release_sock() callback
737 * called from release_sock() to perform protocol dependent
738 * actions before socket release.
740 void tcp_release_cb(struct sock *sk)
742 struct tcp_sock *tp = tcp_sk(sk);
743 unsigned long flags, nflags;
745 /* perform an atomic operation only if at least one flag is set */
747 flags = tp->tsq_flags;
748 if (!(flags & TCP_DEFERRED_ALL))
750 nflags = flags & ~TCP_DEFERRED_ALL;
751 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
753 if (flags & (1UL << TCP_TSQ_DEFERRED))
756 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
757 tcp_write_timer_handler(sk);
760 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
761 tcp_delack_timer_handler(sk);
764 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
765 sk->sk_prot->mtu_reduced(sk);
769 EXPORT_SYMBOL(tcp_release_cb);
771 void __init tcp_tasklet_init(void)
775 for_each_possible_cpu(i) {
776 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
778 INIT_LIST_HEAD(&tsq->head);
779 tasklet_init(&tsq->tasklet,
786 * Write buffer destructor automatically called from kfree_skb.
787 * We cant xmit new skbs from this context, as we might already
790 void tcp_wfree(struct sk_buff *skb)
792 struct sock *sk = skb->sk;
793 struct tcp_sock *tp = tcp_sk(sk);
795 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
796 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
798 struct tsq_tasklet *tsq;
800 /* Keep a ref on socket.
801 * This last ref will be released in tcp_tasklet_func()
803 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
805 /* queue this socket to tasklet queue */
806 local_irq_save(flags);
807 tsq = &__get_cpu_var(tsq_tasklet);
808 list_add(&tp->tsq_node, &tsq->head);
809 tasklet_schedule(&tsq->tasklet);
810 local_irq_restore(flags);
816 /* This routine actually transmits TCP packets queued in by
817 * tcp_do_sendmsg(). This is used by both the initial
818 * transmission and possible later retransmissions.
819 * All SKB's seen here are completely headerless. It is our
820 * job to build the TCP header, and pass the packet down to
821 * IP so it can do the same plus pass the packet off to the
824 * We are working here with either a clone of the original
825 * SKB, or a fresh unique copy made by the retransmit engine.
827 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
830 const struct inet_connection_sock *icsk = inet_csk(sk);
831 struct inet_sock *inet;
833 struct tcp_skb_cb *tcb;
834 struct tcp_out_options opts;
835 unsigned int tcp_options_size, tcp_header_size;
836 struct tcp_md5sig_key *md5;
840 BUG_ON(!skb || !tcp_skb_pcount(skb));
842 /* If congestion control is doing timestamping, we must
843 * take such a timestamp before we potentially clone/copy.
845 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
846 __net_timestamp(skb);
848 if (likely(clone_it)) {
849 const struct sk_buff *fclone = skb + 1;
851 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
852 fclone->fclone == SKB_FCLONE_CLONE))
853 NET_INC_STATS_BH(sock_net(sk),
854 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
856 if (unlikely(skb_cloned(skb)))
857 skb = pskb_copy(skb, gfp_mask);
859 skb = skb_clone(skb, gfp_mask);
866 tcb = TCP_SKB_CB(skb);
867 memset(&opts, 0, sizeof(opts));
869 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
870 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
872 tcp_options_size = tcp_established_options(sk, skb, &opts,
874 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
876 if (tcp_packets_in_flight(tp) == 0) {
877 tcp_ca_event(sk, CA_EVENT_TX_START);
882 skb_push(skb, tcp_header_size);
883 skb_reset_transport_header(skb);
887 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
888 tcp_wfree : sock_wfree;
889 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
891 /* Build TCP header and checksum it. */
893 th->source = inet->inet_sport;
894 th->dest = inet->inet_dport;
895 th->seq = htonl(tcb->seq);
896 th->ack_seq = htonl(tp->rcv_nxt);
897 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
900 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
901 /* RFC1323: The window in SYN & SYN/ACK segments
904 th->window = htons(min(tp->rcv_wnd, 65535U));
906 th->window = htons(tcp_select_window(sk));
911 /* The urg_mode check is necessary during a below snd_una win probe */
912 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
913 if (before(tp->snd_up, tcb->seq + 0x10000)) {
914 th->urg_ptr = htons(tp->snd_up - tcb->seq);
916 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
917 th->urg_ptr = htons(0xFFFF);
922 tcp_options_write((__be32 *)(th + 1), tp, &opts);
923 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
924 TCP_ECN_send(sk, skb, tcp_header_size);
926 #ifdef CONFIG_TCP_MD5SIG
927 /* Calculate the MD5 hash, as we have all we need now */
929 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
930 tp->af_specific->calc_md5_hash(opts.hash_location,
935 icsk->icsk_af_ops->send_check(sk, skb);
937 if (likely(tcb->tcp_flags & TCPHDR_ACK))
938 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
940 if (skb->len != tcp_header_size)
941 tcp_event_data_sent(tp, sk);
943 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
944 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
945 tcp_skb_pcount(skb));
947 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
948 if (likely(err <= 0))
951 tcp_enter_cwr(sk, 1);
953 return net_xmit_eval(err);
956 /* This routine just queues the buffer for sending.
958 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
959 * otherwise socket can stall.
961 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
963 struct tcp_sock *tp = tcp_sk(sk);
965 /* Advance write_seq and place onto the write_queue. */
966 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
967 skb_header_release(skb);
968 tcp_add_write_queue_tail(sk, skb);
969 sk->sk_wmem_queued += skb->truesize;
970 sk_mem_charge(sk, skb->truesize);
973 /* Initialize TSO segments for a packet. */
974 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
975 unsigned int mss_now)
977 if (skb->len <= mss_now || !sk_can_gso(sk) ||
978 skb->ip_summed == CHECKSUM_NONE) {
979 /* Avoid the costly divide in the normal
982 skb_shinfo(skb)->gso_segs = 1;
983 skb_shinfo(skb)->gso_size = 0;
984 skb_shinfo(skb)->gso_type = 0;
986 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
987 skb_shinfo(skb)->gso_size = mss_now;
988 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
992 /* When a modification to fackets out becomes necessary, we need to check
993 * skb is counted to fackets_out or not.
995 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
998 struct tcp_sock *tp = tcp_sk(sk);
1000 if (!tp->sacked_out || tcp_is_reno(tp))
1003 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1004 tp->fackets_out -= decr;
1007 /* Pcount in the middle of the write queue got changed, we need to do various
1008 * tweaks to fix counters
1010 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1012 struct tcp_sock *tp = tcp_sk(sk);
1014 tp->packets_out -= decr;
1016 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1017 tp->sacked_out -= decr;
1018 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1019 tp->retrans_out -= decr;
1020 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1021 tp->lost_out -= decr;
1023 /* Reno case is special. Sigh... */
1024 if (tcp_is_reno(tp) && decr > 0)
1025 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1027 tcp_adjust_fackets_out(sk, skb, decr);
1029 if (tp->lost_skb_hint &&
1030 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1031 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1032 tp->lost_cnt_hint -= decr;
1034 tcp_verify_left_out(tp);
1037 /* Function to create two new TCP segments. Shrinks the given segment
1038 * to the specified size and appends a new segment with the rest of the
1039 * packet to the list. This won't be called frequently, I hope.
1040 * Remember, these are still headerless SKBs at this point.
1042 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1043 unsigned int mss_now)
1045 struct tcp_sock *tp = tcp_sk(sk);
1046 struct sk_buff *buff;
1047 int nsize, old_factor;
1051 if (WARN_ON(len > skb->len))
1054 nsize = skb_headlen(skb) - len;
1058 if (skb_cloned(skb) &&
1059 skb_is_nonlinear(skb) &&
1060 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1063 /* Get a new skb... force flag on. */
1064 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1066 return -ENOMEM; /* We'll just try again later. */
1068 sk->sk_wmem_queued += buff->truesize;
1069 sk_mem_charge(sk, buff->truesize);
1070 nlen = skb->len - len - nsize;
1071 buff->truesize += nlen;
1072 skb->truesize -= nlen;
1074 /* Correct the sequence numbers. */
1075 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1076 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1077 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1079 /* PSH and FIN should only be set in the second packet. */
1080 flags = TCP_SKB_CB(skb)->tcp_flags;
1081 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1082 TCP_SKB_CB(buff)->tcp_flags = flags;
1083 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1085 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1086 /* Copy and checksum data tail into the new buffer. */
1087 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1088 skb_put(buff, nsize),
1093 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1095 skb->ip_summed = CHECKSUM_PARTIAL;
1096 skb_split(skb, buff, len);
1099 buff->ip_summed = skb->ip_summed;
1101 /* Looks stupid, but our code really uses when of
1102 * skbs, which it never sent before. --ANK
1104 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1105 buff->tstamp = skb->tstamp;
1107 old_factor = tcp_skb_pcount(skb);
1109 /* Fix up tso_factor for both original and new SKB. */
1110 tcp_set_skb_tso_segs(sk, skb, mss_now);
1111 tcp_set_skb_tso_segs(sk, buff, mss_now);
1113 /* If this packet has been sent out already, we must
1114 * adjust the various packet counters.
1116 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1117 int diff = old_factor - tcp_skb_pcount(skb) -
1118 tcp_skb_pcount(buff);
1121 tcp_adjust_pcount(sk, skb, diff);
1124 /* Link BUFF into the send queue. */
1125 skb_header_release(buff);
1126 tcp_insert_write_queue_after(skb, buff, sk);
1131 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1132 * eventually). The difference is that pulled data not copied, but
1133 * immediately discarded.
1135 static void __pskb_trim_head(struct sk_buff *skb, int len)
1139 eat = min_t(int, len, skb_headlen(skb));
1141 __skb_pull(skb, eat);
1148 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1149 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1152 skb_frag_unref(skb, i);
1155 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1157 skb_shinfo(skb)->frags[k].page_offset += eat;
1158 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1164 skb_shinfo(skb)->nr_frags = k;
1166 skb_reset_tail_pointer(skb);
1167 skb->data_len -= len;
1168 skb->len = skb->data_len;
1171 /* Remove acked data from a packet in the transmit queue. */
1172 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1174 if (skb_unclone(skb, GFP_ATOMIC))
1177 __pskb_trim_head(skb, len);
1179 TCP_SKB_CB(skb)->seq += len;
1180 skb->ip_summed = CHECKSUM_PARTIAL;
1182 skb->truesize -= len;
1183 sk->sk_wmem_queued -= len;
1184 sk_mem_uncharge(sk, len);
1185 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1187 /* Any change of skb->len requires recalculation of tso factor. */
1188 if (tcp_skb_pcount(skb) > 1)
1189 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1194 /* Calculate MSS not accounting any TCP options. */
1195 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1197 const struct tcp_sock *tp = tcp_sk(sk);
1198 const struct inet_connection_sock *icsk = inet_csk(sk);
1201 /* Calculate base mss without TCP options:
1202 It is MMS_S - sizeof(tcphdr) of rfc1122
1204 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1206 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1207 if (icsk->icsk_af_ops->net_frag_header_len) {
1208 const struct dst_entry *dst = __sk_dst_get(sk);
1210 if (dst && dst_allfrag(dst))
1211 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1214 /* Clamp it (mss_clamp does not include tcp options) */
1215 if (mss_now > tp->rx_opt.mss_clamp)
1216 mss_now = tp->rx_opt.mss_clamp;
1218 /* Now subtract optional transport overhead */
1219 mss_now -= icsk->icsk_ext_hdr_len;
1221 /* Then reserve room for full set of TCP options and 8 bytes of data */
1227 /* Calculate MSS. Not accounting for SACKs here. */
1228 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1230 /* Subtract TCP options size, not including SACKs */
1231 return __tcp_mtu_to_mss(sk, pmtu) -
1232 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1235 /* Inverse of above */
1236 int tcp_mss_to_mtu(struct sock *sk, int mss)
1238 const struct tcp_sock *tp = tcp_sk(sk);
1239 const struct inet_connection_sock *icsk = inet_csk(sk);
1243 tp->tcp_header_len +
1244 icsk->icsk_ext_hdr_len +
1245 icsk->icsk_af_ops->net_header_len;
1247 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1248 if (icsk->icsk_af_ops->net_frag_header_len) {
1249 const struct dst_entry *dst = __sk_dst_get(sk);
1251 if (dst && dst_allfrag(dst))
1252 mtu += icsk->icsk_af_ops->net_frag_header_len;
1257 /* MTU probing init per socket */
1258 void tcp_mtup_init(struct sock *sk)
1260 struct tcp_sock *tp = tcp_sk(sk);
1261 struct inet_connection_sock *icsk = inet_csk(sk);
1263 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1264 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1265 icsk->icsk_af_ops->net_header_len;
1266 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1267 icsk->icsk_mtup.probe_size = 0;
1269 EXPORT_SYMBOL(tcp_mtup_init);
1271 /* This function synchronize snd mss to current pmtu/exthdr set.
1273 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1274 for TCP options, but includes only bare TCP header.
1276 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1277 It is minimum of user_mss and mss received with SYN.
1278 It also does not include TCP options.
1280 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1282 tp->mss_cache is current effective sending mss, including
1283 all tcp options except for SACKs. It is evaluated,
1284 taking into account current pmtu, but never exceeds
1285 tp->rx_opt.mss_clamp.
1287 NOTE1. rfc1122 clearly states that advertised MSS
1288 DOES NOT include either tcp or ip options.
1290 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1291 are READ ONLY outside this function. --ANK (980731)
1293 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1295 struct tcp_sock *tp = tcp_sk(sk);
1296 struct inet_connection_sock *icsk = inet_csk(sk);
1299 if (icsk->icsk_mtup.search_high > pmtu)
1300 icsk->icsk_mtup.search_high = pmtu;
1302 mss_now = tcp_mtu_to_mss(sk, pmtu);
1303 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1305 /* And store cached results */
1306 icsk->icsk_pmtu_cookie = pmtu;
1307 if (icsk->icsk_mtup.enabled)
1308 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1309 tp->mss_cache = mss_now;
1313 EXPORT_SYMBOL(tcp_sync_mss);
1315 /* Compute the current effective MSS, taking SACKs and IP options,
1316 * and even PMTU discovery events into account.
1318 unsigned int tcp_current_mss(struct sock *sk)
1320 const struct tcp_sock *tp = tcp_sk(sk);
1321 const struct dst_entry *dst = __sk_dst_get(sk);
1323 unsigned int header_len;
1324 struct tcp_out_options opts;
1325 struct tcp_md5sig_key *md5;
1327 mss_now = tp->mss_cache;
1330 u32 mtu = dst_mtu(dst);
1331 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1332 mss_now = tcp_sync_mss(sk, mtu);
1335 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1336 sizeof(struct tcphdr);
1337 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1338 * some common options. If this is an odd packet (because we have SACK
1339 * blocks etc) then our calculated header_len will be different, and
1340 * we have to adjust mss_now correspondingly */
1341 if (header_len != tp->tcp_header_len) {
1342 int delta = (int) header_len - tp->tcp_header_len;
1349 /* Congestion window validation. (RFC2861) */
1350 static void tcp_cwnd_validate(struct sock *sk)
1352 struct tcp_sock *tp = tcp_sk(sk);
1354 if (tp->packets_out >= tp->snd_cwnd) {
1355 /* Network is feed fully. */
1356 tp->snd_cwnd_used = 0;
1357 tp->snd_cwnd_stamp = tcp_time_stamp;
1359 /* Network starves. */
1360 if (tp->packets_out > tp->snd_cwnd_used)
1361 tp->snd_cwnd_used = tp->packets_out;
1363 if (sysctl_tcp_slow_start_after_idle &&
1364 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1365 tcp_cwnd_application_limited(sk);
1369 /* Returns the portion of skb which can be sent right away without
1370 * introducing MSS oddities to segment boundaries. In rare cases where
1371 * mss_now != mss_cache, we will request caller to create a small skb
1372 * per input skb which could be mostly avoided here (if desired).
1374 * We explicitly want to create a request for splitting write queue tail
1375 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1376 * thus all the complexity (cwnd_len is always MSS multiple which we
1377 * return whenever allowed by the other factors). Basically we need the
1378 * modulo only when the receiver window alone is the limiting factor or
1379 * when we would be allowed to send the split-due-to-Nagle skb fully.
1381 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1382 unsigned int mss_now, unsigned int max_segs)
1384 const struct tcp_sock *tp = tcp_sk(sk);
1385 u32 needed, window, max_len;
1387 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1388 max_len = mss_now * max_segs;
1390 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1393 needed = min(skb->len, window);
1395 if (max_len <= needed)
1398 return needed - needed % mss_now;
1401 /* Can at least one segment of SKB be sent right now, according to the
1402 * congestion window rules? If so, return how many segments are allowed.
1404 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1405 const struct sk_buff *skb)
1407 u32 in_flight, cwnd;
1409 /* Don't be strict about the congestion window for the final FIN. */
1410 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1411 tcp_skb_pcount(skb) == 1)
1414 in_flight = tcp_packets_in_flight(tp);
1415 cwnd = tp->snd_cwnd;
1416 if (in_flight < cwnd)
1417 return (cwnd - in_flight);
1422 /* Initialize TSO state of a skb.
1423 * This must be invoked the first time we consider transmitting
1424 * SKB onto the wire.
1426 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1427 unsigned int mss_now)
1429 int tso_segs = tcp_skb_pcount(skb);
1431 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1432 tcp_set_skb_tso_segs(sk, skb, mss_now);
1433 tso_segs = tcp_skb_pcount(skb);
1438 /* Minshall's variant of the Nagle send check. */
1439 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1441 return after(tp->snd_sml, tp->snd_una) &&
1442 !after(tp->snd_sml, tp->snd_nxt);
1445 /* Return false, if packet can be sent now without violation Nagle's rules:
1446 * 1. It is full sized.
1447 * 2. Or it contains FIN. (already checked by caller)
1448 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1449 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1450 * With Minshall's modification: all sent small packets are ACKed.
1452 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1453 const struct sk_buff *skb,
1454 unsigned int mss_now, int nonagle)
1456 return skb->len < mss_now &&
1457 ((nonagle & TCP_NAGLE_CORK) ||
1458 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1461 /* Return true if the Nagle test allows this packet to be
1464 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1465 unsigned int cur_mss, int nonagle)
1467 /* Nagle rule does not apply to frames, which sit in the middle of the
1468 * write_queue (they have no chances to get new data).
1470 * This is implemented in the callers, where they modify the 'nonagle'
1471 * argument based upon the location of SKB in the send queue.
1473 if (nonagle & TCP_NAGLE_PUSH)
1476 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1477 if (tcp_urg_mode(tp) || (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 * Do not rearm the timer if already set to not break TCP ACK clocking.
1653 if (!tp->tso_deferred)
1654 tp->tso_deferred = 1 | (jiffies << 1);
1659 tp->tso_deferred = 0;
1663 /* Create a new MTU probe if we are ready.
1664 * MTU probe is regularly attempting to increase the path MTU by
1665 * deliberately sending larger packets. This discovers routing
1666 * changes resulting in larger path MTUs.
1668 * Returns 0 if we should wait to probe (no cwnd available),
1669 * 1 if a probe was sent,
1672 static int tcp_mtu_probe(struct sock *sk)
1674 struct tcp_sock *tp = tcp_sk(sk);
1675 struct inet_connection_sock *icsk = inet_csk(sk);
1676 struct sk_buff *skb, *nskb, *next;
1683 /* Not currently probing/verifying,
1685 * have enough cwnd, and
1686 * not SACKing (the variable headers throw things off) */
1687 if (!icsk->icsk_mtup.enabled ||
1688 icsk->icsk_mtup.probe_size ||
1689 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1690 tp->snd_cwnd < 11 ||
1691 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1694 /* Very simple search strategy: just double the MSS. */
1695 mss_now = tcp_current_mss(sk);
1696 probe_size = 2 * tp->mss_cache;
1697 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1698 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1699 /* TODO: set timer for probe_converge_event */
1703 /* Have enough data in the send queue to probe? */
1704 if (tp->write_seq - tp->snd_nxt < size_needed)
1707 if (tp->snd_wnd < size_needed)
1709 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1712 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1713 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1714 if (!tcp_packets_in_flight(tp))
1720 /* We're allowed to probe. Build it now. */
1721 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1723 sk->sk_wmem_queued += nskb->truesize;
1724 sk_mem_charge(sk, nskb->truesize);
1726 skb = tcp_send_head(sk);
1728 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1729 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1730 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1731 TCP_SKB_CB(nskb)->sacked = 0;
1733 nskb->ip_summed = skb->ip_summed;
1735 tcp_insert_write_queue_before(nskb, skb, sk);
1738 tcp_for_write_queue_from_safe(skb, next, sk) {
1739 copy = min_t(int, skb->len, probe_size - len);
1740 if (nskb->ip_summed)
1741 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1743 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1744 skb_put(nskb, copy),
1747 if (skb->len <= copy) {
1748 /* We've eaten all the data from this skb.
1750 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1751 tcp_unlink_write_queue(skb, sk);
1752 sk_wmem_free_skb(sk, skb);
1754 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1755 ~(TCPHDR_FIN|TCPHDR_PSH);
1756 if (!skb_shinfo(skb)->nr_frags) {
1757 skb_pull(skb, copy);
1758 if (skb->ip_summed != CHECKSUM_PARTIAL)
1759 skb->csum = csum_partial(skb->data,
1762 __pskb_trim_head(skb, copy);
1763 tcp_set_skb_tso_segs(sk, skb, mss_now);
1765 TCP_SKB_CB(skb)->seq += copy;
1770 if (len >= probe_size)
1773 tcp_init_tso_segs(sk, nskb, nskb->len);
1775 /* We're ready to send. If this fails, the probe will
1776 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1777 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1778 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1779 /* Decrement cwnd here because we are sending
1780 * effectively two packets. */
1782 tcp_event_new_data_sent(sk, nskb);
1784 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1785 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1786 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1794 /* This routine writes packets to the network. It advances the
1795 * send_head. This happens as incoming acks open up the remote
1798 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1799 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1800 * account rare use of URG, this is not a big flaw.
1802 * Send at most one packet when push_one > 0. Temporarily ignore
1803 * cwnd limit to force at most one packet out when push_one == 2.
1805 * Returns true, if no segments are in flight and we have queued segments,
1806 * but cannot send anything now because of SWS or another problem.
1808 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1809 int push_one, gfp_t gfp)
1811 struct tcp_sock *tp = tcp_sk(sk);
1812 struct sk_buff *skb;
1813 unsigned int tso_segs, sent_pkts;
1820 /* Do MTU probing. */
1821 result = tcp_mtu_probe(sk);
1824 } else if (result > 0) {
1829 while ((skb = tcp_send_head(sk))) {
1833 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1836 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1837 goto repair; /* Skip network transmission */
1839 cwnd_quota = tcp_cwnd_test(tp, skb);
1842 /* Force out a loss probe pkt. */
1848 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1851 if (tso_segs == 1) {
1852 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1853 (tcp_skb_is_last(sk, skb) ?
1854 nonagle : TCP_NAGLE_PUSH))))
1857 if (!push_one && tcp_tso_should_defer(sk, skb))
1861 /* TSQ : sk_wmem_alloc accounts skb truesize,
1862 * including skb overhead. But thats OK.
1864 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1865 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1869 if (tso_segs > 1 && !tcp_urg_mode(tp))
1870 limit = tcp_mss_split_point(sk, skb, mss_now,
1873 sk->sk_gso_max_segs));
1875 if (skb->len > limit &&
1876 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1879 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1881 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1885 /* Advance the send_head. This one is sent out.
1886 * This call will increment packets_out.
1888 tcp_event_new_data_sent(sk, skb);
1890 tcp_minshall_update(tp, mss_now, skb);
1891 sent_pkts += tcp_skb_pcount(skb);
1897 if (likely(sent_pkts)) {
1898 if (tcp_in_cwnd_reduction(sk))
1899 tp->prr_out += sent_pkts;
1901 /* Send one loss probe per tail loss episode. */
1903 tcp_schedule_loss_probe(sk);
1904 tcp_cwnd_validate(sk);
1907 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1910 bool tcp_schedule_loss_probe(struct sock *sk)
1912 struct inet_connection_sock *icsk = inet_csk(sk);
1913 struct tcp_sock *tp = tcp_sk(sk);
1914 u32 timeout, tlp_time_stamp, rto_time_stamp;
1915 u32 rtt = tp->srtt >> 3;
1917 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1919 /* No consecutive loss probes. */
1920 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1924 /* Don't do any loss probe on a Fast Open connection before 3WHS
1927 if (sk->sk_state == TCP_SYN_RECV)
1930 /* TLP is only scheduled when next timer event is RTO. */
1931 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1934 /* Schedule a loss probe in 2*RTT for SACK capable connections
1935 * in Open state, that are either limited by cwnd or application.
1937 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1938 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1941 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1945 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1946 * for delayed ack when there's one outstanding packet.
1949 if (tp->packets_out == 1)
1950 timeout = max_t(u32, timeout,
1951 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1952 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1954 /* If RTO is shorter, just schedule TLP in its place. */
1955 tlp_time_stamp = tcp_time_stamp + timeout;
1956 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1957 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1958 s32 delta = rto_time_stamp - tcp_time_stamp;
1963 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1968 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1969 * retransmit the last segment.
1971 void tcp_send_loss_probe(struct sock *sk)
1973 struct tcp_sock *tp = tcp_sk(sk);
1974 struct sk_buff *skb;
1976 int mss = tcp_current_mss(sk);
1979 if (tcp_send_head(sk) != NULL) {
1980 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1984 /* At most one outstanding TLP retransmission. */
1985 if (tp->tlp_high_seq)
1988 /* Retransmit last segment. */
1989 skb = tcp_write_queue_tail(sk);
1993 pcount = tcp_skb_pcount(skb);
1994 if (WARN_ON(!pcount))
1997 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
1998 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2000 skb = tcp_write_queue_tail(sk);
2003 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2006 /* Probe with zero data doesn't trigger fast recovery. */
2008 err = __tcp_retransmit_skb(sk, skb);
2010 /* Record snd_nxt for loss detection. */
2012 tp->tlp_high_seq = tp->snd_nxt;
2015 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2016 inet_csk(sk)->icsk_rto,
2020 NET_INC_STATS_BH(sock_net(sk),
2021 LINUX_MIB_TCPLOSSPROBES);
2025 /* Push out any pending frames which were held back due to
2026 * TCP_CORK or attempt at coalescing tiny packets.
2027 * The socket must be locked by the caller.
2029 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2032 /* If we are closed, the bytes will have to remain here.
2033 * In time closedown will finish, we empty the write queue and
2034 * all will be happy.
2036 if (unlikely(sk->sk_state == TCP_CLOSE))
2039 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2040 sk_gfp_atomic(sk, GFP_ATOMIC)))
2041 tcp_check_probe_timer(sk);
2044 /* Send _single_ skb sitting at the send head. This function requires
2045 * true push pending frames to setup probe timer etc.
2047 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2049 struct sk_buff *skb = tcp_send_head(sk);
2051 BUG_ON(!skb || skb->len < mss_now);
2053 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2056 /* This function returns the amount that we can raise the
2057 * usable window based on the following constraints
2059 * 1. The window can never be shrunk once it is offered (RFC 793)
2060 * 2. We limit memory per socket
2063 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2064 * RECV.NEXT + RCV.WIN fixed until:
2065 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2067 * i.e. don't raise the right edge of the window until you can raise
2068 * it at least MSS bytes.
2070 * Unfortunately, the recommended algorithm breaks header prediction,
2071 * since header prediction assumes th->window stays fixed.
2073 * Strictly speaking, keeping th->window fixed violates the receiver
2074 * side SWS prevention criteria. The problem is that under this rule
2075 * a stream of single byte packets will cause the right side of the
2076 * window to always advance by a single byte.
2078 * Of course, if the sender implements sender side SWS prevention
2079 * then this will not be a problem.
2081 * BSD seems to make the following compromise:
2083 * If the free space is less than the 1/4 of the maximum
2084 * space available and the free space is less than 1/2 mss,
2085 * then set the window to 0.
2086 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2087 * Otherwise, just prevent the window from shrinking
2088 * and from being larger than the largest representable value.
2090 * This prevents incremental opening of the window in the regime
2091 * where TCP is limited by the speed of the reader side taking
2092 * data out of the TCP receive queue. It does nothing about
2093 * those cases where the window is constrained on the sender side
2094 * because the pipeline is full.
2096 * BSD also seems to "accidentally" limit itself to windows that are a
2097 * multiple of MSS, at least until the free space gets quite small.
2098 * This would appear to be a side effect of the mbuf implementation.
2099 * Combining these two algorithms results in the observed behavior
2100 * of having a fixed window size at almost all times.
2102 * Below we obtain similar behavior by forcing the offered window to
2103 * a multiple of the mss when it is feasible to do so.
2105 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2106 * Regular options like TIMESTAMP are taken into account.
2108 u32 __tcp_select_window(struct sock *sk)
2110 struct inet_connection_sock *icsk = inet_csk(sk);
2111 struct tcp_sock *tp = tcp_sk(sk);
2112 /* MSS for the peer's data. Previous versions used mss_clamp
2113 * here. I don't know if the value based on our guesses
2114 * of peer's MSS is better for the performance. It's more correct
2115 * but may be worse for the performance because of rcv_mss
2116 * fluctuations. --SAW 1998/11/1
2118 int mss = icsk->icsk_ack.rcv_mss;
2119 int free_space = tcp_space(sk);
2120 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2123 if (mss > full_space)
2126 if (free_space < (full_space >> 1)) {
2127 icsk->icsk_ack.quick = 0;
2129 if (sk_under_memory_pressure(sk))
2130 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2133 if (free_space < mss)
2137 if (free_space > tp->rcv_ssthresh)
2138 free_space = tp->rcv_ssthresh;
2140 /* Don't do rounding if we are using window scaling, since the
2141 * scaled window will not line up with the MSS boundary anyway.
2143 window = tp->rcv_wnd;
2144 if (tp->rx_opt.rcv_wscale) {
2145 window = free_space;
2147 /* Advertise enough space so that it won't get scaled away.
2148 * Import case: prevent zero window announcement if
2149 * 1<<rcv_wscale > mss.
2151 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2152 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2153 << tp->rx_opt.rcv_wscale);
2155 /* Get the largest window that is a nice multiple of mss.
2156 * Window clamp already applied above.
2157 * If our current window offering is within 1 mss of the
2158 * free space we just keep it. This prevents the divide
2159 * and multiply from happening most of the time.
2160 * We also don't do any window rounding when the free space
2163 if (window <= free_space - mss || window > free_space)
2164 window = (free_space / mss) * mss;
2165 else if (mss == full_space &&
2166 free_space > window + (full_space >> 1))
2167 window = free_space;
2173 /* Collapses two adjacent SKB's during retransmission. */
2174 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2176 struct tcp_sock *tp = tcp_sk(sk);
2177 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2178 int skb_size, next_skb_size;
2180 skb_size = skb->len;
2181 next_skb_size = next_skb->len;
2183 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2185 tcp_highest_sack_combine(sk, next_skb, skb);
2187 tcp_unlink_write_queue(next_skb, sk);
2189 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2192 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2193 skb->ip_summed = CHECKSUM_PARTIAL;
2195 if (skb->ip_summed != CHECKSUM_PARTIAL)
2196 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2198 /* Update sequence range on original skb. */
2199 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2201 /* Merge over control information. This moves PSH/FIN etc. over */
2202 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2204 /* All done, get rid of second SKB and account for it so
2205 * packet counting does not break.
2207 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2209 /* changed transmit queue under us so clear hints */
2210 tcp_clear_retrans_hints_partial(tp);
2211 if (next_skb == tp->retransmit_skb_hint)
2212 tp->retransmit_skb_hint = skb;
2214 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2216 sk_wmem_free_skb(sk, next_skb);
2219 /* Check if coalescing SKBs is legal. */
2220 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2222 if (tcp_skb_pcount(skb) > 1)
2224 /* TODO: SACK collapsing could be used to remove this condition */
2225 if (skb_shinfo(skb)->nr_frags != 0)
2227 if (skb_cloned(skb))
2229 if (skb == tcp_send_head(sk))
2231 /* Some heurestics for collapsing over SACK'd could be invented */
2232 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2238 /* Collapse packets in the retransmit queue to make to create
2239 * less packets on the wire. This is only done on retransmission.
2241 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2244 struct tcp_sock *tp = tcp_sk(sk);
2245 struct sk_buff *skb = to, *tmp;
2248 if (!sysctl_tcp_retrans_collapse)
2250 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2253 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2254 if (!tcp_can_collapse(sk, skb))
2266 /* Punt if not enough space exists in the first SKB for
2267 * the data in the second
2269 if (skb->len > skb_availroom(to))
2272 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2275 tcp_collapse_retrans(sk, to);
2279 /* This retransmits one SKB. Policy decisions and retransmit queue
2280 * state updates are done by the caller. Returns non-zero if an
2281 * error occurred which prevented the send.
2283 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2285 struct tcp_sock *tp = tcp_sk(sk);
2286 struct inet_connection_sock *icsk = inet_csk(sk);
2287 unsigned int cur_mss;
2289 /* Inconslusive MTU probe */
2290 if (icsk->icsk_mtup.probe_size) {
2291 icsk->icsk_mtup.probe_size = 0;
2294 /* Do not sent more than we queued. 1/4 is reserved for possible
2295 * copying overhead: fragmentation, tunneling, mangling etc.
2297 if (atomic_read(&sk->sk_wmem_alloc) >
2298 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2301 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2302 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2304 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2308 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2309 return -EHOSTUNREACH; /* Routing failure or similar. */
2311 cur_mss = tcp_current_mss(sk);
2313 /* If receiver has shrunk his window, and skb is out of
2314 * new window, do not retransmit it. The exception is the
2315 * case, when window is shrunk to zero. In this case
2316 * our retransmit serves as a zero window probe.
2318 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2319 TCP_SKB_CB(skb)->seq != tp->snd_una)
2322 if (skb->len > cur_mss) {
2323 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2324 return -ENOMEM; /* We'll try again later. */
2326 int oldpcount = tcp_skb_pcount(skb);
2328 if (unlikely(oldpcount > 1)) {
2329 tcp_init_tso_segs(sk, skb, cur_mss);
2330 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2334 tcp_retrans_try_collapse(sk, skb, cur_mss);
2336 /* Some Solaris stacks overoptimize and ignore the FIN on a
2337 * retransmit when old data is attached. So strip it off
2338 * since it is cheap to do so and saves bytes on the network.
2341 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2342 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2343 if (!pskb_trim(skb, 0)) {
2344 /* Reuse, even though it does some unnecessary work */
2345 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2346 TCP_SKB_CB(skb)->tcp_flags);
2347 skb->ip_summed = CHECKSUM_NONE;
2351 /* Make a copy, if the first transmission SKB clone we made
2352 * is still in somebody's hands, else make a clone.
2354 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2356 /* make sure skb->data is aligned on arches that require it
2357 * and check if ack-trimming & collapsing extended the headroom
2358 * beyond what csum_start can cover.
2360 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2361 skb_headroom(skb) >= 0xFFFF)) {
2362 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2364 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2367 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2371 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2373 struct tcp_sock *tp = tcp_sk(sk);
2374 int err = __tcp_retransmit_skb(sk, skb);
2377 /* Update global TCP statistics. */
2378 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2380 tp->total_retrans++;
2382 #if FASTRETRANS_DEBUG > 0
2383 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2384 net_dbg_ratelimited("retrans_out leaked\n");
2387 if (!tp->retrans_out)
2388 tp->lost_retrans_low = tp->snd_nxt;
2389 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2390 tp->retrans_out += tcp_skb_pcount(skb);
2392 /* Save stamp of the first retransmit. */
2393 if (!tp->retrans_stamp)
2394 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2396 tp->undo_retrans += tcp_skb_pcount(skb);
2398 /* snd_nxt is stored to detect loss of retransmitted segment,
2399 * see tcp_input.c tcp_sacktag_write_queue().
2401 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2406 /* Check if we forward retransmits are possible in the current
2407 * window/congestion state.
2409 static bool tcp_can_forward_retransmit(struct sock *sk)
2411 const struct inet_connection_sock *icsk = inet_csk(sk);
2412 const struct tcp_sock *tp = tcp_sk(sk);
2414 /* Forward retransmissions are possible only during Recovery. */
2415 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2418 /* No forward retransmissions in Reno are possible. */
2419 if (tcp_is_reno(tp))
2422 /* Yeah, we have to make difficult choice between forward transmission
2423 * and retransmission... Both ways have their merits...
2425 * For now we do not retransmit anything, while we have some new
2426 * segments to send. In the other cases, follow rule 3 for
2427 * NextSeg() specified in RFC3517.
2430 if (tcp_may_send_now(sk))
2436 /* This gets called after a retransmit timeout, and the initially
2437 * retransmitted data is acknowledged. It tries to continue
2438 * resending the rest of the retransmit queue, until either
2439 * we've sent it all or the congestion window limit is reached.
2440 * If doing SACK, the first ACK which comes back for a timeout
2441 * based retransmit packet might feed us FACK information again.
2442 * If so, we use it to avoid unnecessarily retransmissions.
2444 void tcp_xmit_retransmit_queue(struct sock *sk)
2446 const struct inet_connection_sock *icsk = inet_csk(sk);
2447 struct tcp_sock *tp = tcp_sk(sk);
2448 struct sk_buff *skb;
2449 struct sk_buff *hole = NULL;
2452 int fwd_rexmitting = 0;
2454 if (!tp->packets_out)
2458 tp->retransmit_high = tp->snd_una;
2460 if (tp->retransmit_skb_hint) {
2461 skb = tp->retransmit_skb_hint;
2462 last_lost = TCP_SKB_CB(skb)->end_seq;
2463 if (after(last_lost, tp->retransmit_high))
2464 last_lost = tp->retransmit_high;
2466 skb = tcp_write_queue_head(sk);
2467 last_lost = tp->snd_una;
2470 tcp_for_write_queue_from(skb, sk) {
2471 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2473 if (skb == tcp_send_head(sk))
2475 /* we could do better than to assign each time */
2477 tp->retransmit_skb_hint = skb;
2479 /* Assume this retransmit will generate
2480 * only one packet for congestion window
2481 * calculation purposes. This works because
2482 * tcp_retransmit_skb() will chop up the
2483 * packet to be MSS sized and all the
2484 * packet counting works out.
2486 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2489 if (fwd_rexmitting) {
2491 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2493 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2495 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2496 tp->retransmit_high = last_lost;
2497 if (!tcp_can_forward_retransmit(sk))
2499 /* Backtrack if necessary to non-L'ed skb */
2507 } else if (!(sacked & TCPCB_LOST)) {
2508 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2513 last_lost = TCP_SKB_CB(skb)->end_seq;
2514 if (icsk->icsk_ca_state != TCP_CA_Loss)
2515 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2517 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2520 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2523 if (tcp_retransmit_skb(sk, skb)) {
2524 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2527 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2529 if (tcp_in_cwnd_reduction(sk))
2530 tp->prr_out += tcp_skb_pcount(skb);
2532 if (skb == tcp_write_queue_head(sk))
2533 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2534 inet_csk(sk)->icsk_rto,
2539 /* Send a fin. The caller locks the socket for us. This cannot be
2540 * allowed to fail queueing a FIN frame under any circumstances.
2542 void tcp_send_fin(struct sock *sk)
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 struct sk_buff *skb = tcp_write_queue_tail(sk);
2548 /* Optimization, tack on the FIN if we have a queue of
2549 * unsent frames. But be careful about outgoing SACKS
2552 mss_now = tcp_current_mss(sk);
2554 if (tcp_send_head(sk) != NULL) {
2555 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2556 TCP_SKB_CB(skb)->end_seq++;
2559 /* Socket is locked, keep trying until memory is available. */
2561 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2568 /* Reserve space for headers and prepare control bits. */
2569 skb_reserve(skb, MAX_TCP_HEADER);
2570 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2571 tcp_init_nondata_skb(skb, tp->write_seq,
2572 TCPHDR_ACK | TCPHDR_FIN);
2573 tcp_queue_skb(sk, skb);
2575 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2578 /* We get here when a process closes a file descriptor (either due to
2579 * an explicit close() or as a byproduct of exit()'ing) and there
2580 * was unread data in the receive queue. This behavior is recommended
2581 * by RFC 2525, section 2.17. -DaveM
2583 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2585 struct sk_buff *skb;
2587 /* NOTE: No TCP options attached and we never retransmit this. */
2588 skb = alloc_skb(MAX_TCP_HEADER, priority);
2590 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2594 /* Reserve space for headers and prepare control bits. */
2595 skb_reserve(skb, MAX_TCP_HEADER);
2596 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2597 TCPHDR_ACK | TCPHDR_RST);
2599 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2600 if (tcp_transmit_skb(sk, skb, 0, priority))
2601 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2603 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2606 /* Send a crossed SYN-ACK during socket establishment.
2607 * WARNING: This routine must only be called when we have already sent
2608 * a SYN packet that crossed the incoming SYN that caused this routine
2609 * to get called. If this assumption fails then the initial rcv_wnd
2610 * and rcv_wscale values will not be correct.
2612 int tcp_send_synack(struct sock *sk)
2614 struct sk_buff *skb;
2616 skb = tcp_write_queue_head(sk);
2617 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2618 pr_debug("%s: wrong queue state\n", __func__);
2621 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2622 if (skb_cloned(skb)) {
2623 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2626 tcp_unlink_write_queue(skb, sk);
2627 skb_header_release(nskb);
2628 __tcp_add_write_queue_head(sk, nskb);
2629 sk_wmem_free_skb(sk, skb);
2630 sk->sk_wmem_queued += nskb->truesize;
2631 sk_mem_charge(sk, nskb->truesize);
2635 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2636 TCP_ECN_send_synack(tcp_sk(sk), skb);
2638 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2639 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2643 * tcp_make_synack - Prepare a SYN-ACK.
2644 * sk: listener socket
2645 * dst: dst entry attached to the SYNACK
2646 * req: request_sock pointer
2648 * Allocate one skb and build a SYNACK packet.
2649 * @dst is consumed : Caller should not use it again.
2651 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2652 struct request_sock *req,
2653 struct tcp_fastopen_cookie *foc)
2655 struct tcp_out_options opts;
2656 struct inet_request_sock *ireq = inet_rsk(req);
2657 struct tcp_sock *tp = tcp_sk(sk);
2659 struct sk_buff *skb;
2660 struct tcp_md5sig_key *md5;
2661 int tcp_header_size;
2664 skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC));
2665 if (unlikely(!skb)) {
2669 /* Reserve space for headers. */
2670 skb_reserve(skb, MAX_TCP_HEADER);
2672 skb_dst_set(skb, dst);
2673 security_skb_owned_by(skb, sk);
2675 mss = dst_metric_advmss(dst);
2676 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2677 mss = tp->rx_opt.user_mss;
2679 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2681 /* Set this up on the first call only */
2682 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2684 /* limit the window selection if the user enforce a smaller rx buffer */
2685 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2686 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2687 req->window_clamp = tcp_full_space(sk);
2689 /* tcp_full_space because it is guaranteed to be the first packet */
2690 tcp_select_initial_window(tcp_full_space(sk),
2691 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2696 dst_metric(dst, RTAX_INITRWND));
2697 ireq->rcv_wscale = rcv_wscale;
2700 memset(&opts, 0, sizeof(opts));
2701 #ifdef CONFIG_SYN_COOKIES
2702 if (unlikely(req->cookie_ts))
2703 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2706 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2707 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2710 skb_push(skb, tcp_header_size);
2711 skb_reset_transport_header(skb);
2714 memset(th, 0, sizeof(struct tcphdr));
2717 TCP_ECN_make_synack(req, th);
2718 th->source = ireq->loc_port;
2719 th->dest = ireq->rmt_port;
2720 /* Setting of flags are superfluous here for callers (and ECE is
2721 * not even correctly set)
2723 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2724 TCPHDR_SYN | TCPHDR_ACK);
2726 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2727 /* XXX data is queued and acked as is. No buffer/window check */
2728 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2730 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2731 th->window = htons(min(req->rcv_wnd, 65535U));
2732 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2733 th->doff = (tcp_header_size >> 2);
2734 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2736 #ifdef CONFIG_TCP_MD5SIG
2737 /* Okay, we have all we need - do the md5 hash if needed */
2739 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2740 md5, NULL, req, skb);
2746 EXPORT_SYMBOL(tcp_make_synack);
2748 /* Do all connect socket setups that can be done AF independent. */
2749 void tcp_connect_init(struct sock *sk)
2751 const struct dst_entry *dst = __sk_dst_get(sk);
2752 struct tcp_sock *tp = tcp_sk(sk);
2755 /* We'll fix this up when we get a response from the other end.
2756 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2758 tp->tcp_header_len = sizeof(struct tcphdr) +
2759 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2761 #ifdef CONFIG_TCP_MD5SIG
2762 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2763 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2766 /* If user gave his TCP_MAXSEG, record it to clamp */
2767 if (tp->rx_opt.user_mss)
2768 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2771 tcp_sync_mss(sk, dst_mtu(dst));
2773 if (!tp->window_clamp)
2774 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2775 tp->advmss = dst_metric_advmss(dst);
2776 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2777 tp->advmss = tp->rx_opt.user_mss;
2779 tcp_initialize_rcv_mss(sk);
2781 /* limit the window selection if the user enforce a smaller rx buffer */
2782 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2783 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2784 tp->window_clamp = tcp_full_space(sk);
2786 tcp_select_initial_window(tcp_full_space(sk),
2787 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2790 sysctl_tcp_window_scaling,
2792 dst_metric(dst, RTAX_INITRWND));
2794 tp->rx_opt.rcv_wscale = rcv_wscale;
2795 tp->rcv_ssthresh = tp->rcv_wnd;
2798 sock_reset_flag(sk, SOCK_DONE);
2801 tp->snd_una = tp->write_seq;
2802 tp->snd_sml = tp->write_seq;
2803 tp->snd_up = tp->write_seq;
2804 tp->snd_nxt = tp->write_seq;
2806 if (likely(!tp->repair))
2808 tp->rcv_wup = tp->rcv_nxt;
2809 tp->copied_seq = tp->rcv_nxt;
2811 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2812 inet_csk(sk)->icsk_retransmits = 0;
2813 tcp_clear_retrans(tp);
2816 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2818 struct tcp_sock *tp = tcp_sk(sk);
2819 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2821 tcb->end_seq += skb->len;
2822 skb_header_release(skb);
2823 __tcp_add_write_queue_tail(sk, skb);
2824 sk->sk_wmem_queued += skb->truesize;
2825 sk_mem_charge(sk, skb->truesize);
2826 tp->write_seq = tcb->end_seq;
2827 tp->packets_out += tcp_skb_pcount(skb);
2830 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2831 * queue a data-only packet after the regular SYN, such that regular SYNs
2832 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2833 * only the SYN sequence, the data are retransmitted in the first ACK.
2834 * If cookie is not cached or other error occurs, falls back to send a
2835 * regular SYN with Fast Open cookie request option.
2837 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2839 struct tcp_sock *tp = tcp_sk(sk);
2840 struct tcp_fastopen_request *fo = tp->fastopen_req;
2841 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2842 struct sk_buff *syn_data = NULL, *data;
2843 unsigned long last_syn_loss = 0;
2845 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2846 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2847 &syn_loss, &last_syn_loss);
2848 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2850 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2851 fo->cookie.len = -1;
2855 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2856 fo->cookie.len = -1;
2857 else if (fo->cookie.len <= 0)
2860 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2861 * user-MSS. Reserve maximum option space for middleboxes that add
2862 * private TCP options. The cost is reduced data space in SYN :(
2864 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2865 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2866 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2867 MAX_TCP_OPTION_SPACE;
2869 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2871 if (syn_data == NULL)
2874 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2875 struct iovec *iov = &fo->data->msg_iov[i];
2876 unsigned char __user *from = iov->iov_base;
2877 int len = iov->iov_len;
2879 if (syn_data->len + len > space)
2880 len = space - syn_data->len;
2881 else if (i + 1 == iovlen)
2882 /* No more data pending in inet_wait_for_connect() */
2885 if (skb_add_data(syn_data, from, len))
2889 /* Queue a data-only packet after the regular SYN for retransmission */
2890 data = pskb_copy(syn_data, sk->sk_allocation);
2893 TCP_SKB_CB(data)->seq++;
2894 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2895 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2896 tcp_connect_queue_skb(sk, data);
2897 fo->copied = data->len;
2899 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2900 tp->syn_data = (fo->copied > 0);
2901 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2907 /* Send a regular SYN with Fast Open cookie request option */
2908 if (fo->cookie.len > 0)
2910 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2912 tp->syn_fastopen = 0;
2913 kfree_skb(syn_data);
2915 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2919 /* Build a SYN and send it off. */
2920 int tcp_connect(struct sock *sk)
2922 struct tcp_sock *tp = tcp_sk(sk);
2923 struct sk_buff *buff;
2926 tcp_connect_init(sk);
2928 if (unlikely(tp->repair)) {
2929 tcp_finish_connect(sk, NULL);
2933 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2934 if (unlikely(buff == NULL))
2937 /* Reserve space for headers. */
2938 skb_reserve(buff, MAX_TCP_HEADER);
2940 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2941 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2942 tcp_connect_queue_skb(sk, buff);
2943 TCP_ECN_send_syn(sk, buff);
2945 /* Send off SYN; include data in Fast Open. */
2946 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2947 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2948 if (err == -ECONNREFUSED)
2951 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2952 * in order to make this packet get counted in tcpOutSegs.
2954 tp->snd_nxt = tp->write_seq;
2955 tp->pushed_seq = tp->write_seq;
2956 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2958 /* Timer for repeating the SYN until an answer. */
2959 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2960 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2963 EXPORT_SYMBOL(tcp_connect);
2965 /* Send out a delayed ack, the caller does the policy checking
2966 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2969 void tcp_send_delayed_ack(struct sock *sk)
2971 struct inet_connection_sock *icsk = inet_csk(sk);
2972 int ato = icsk->icsk_ack.ato;
2973 unsigned long timeout;
2975 if (ato > TCP_DELACK_MIN) {
2976 const struct tcp_sock *tp = tcp_sk(sk);
2977 int max_ato = HZ / 2;
2979 if (icsk->icsk_ack.pingpong ||
2980 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2981 max_ato = TCP_DELACK_MAX;
2983 /* Slow path, intersegment interval is "high". */
2985 /* If some rtt estimate is known, use it to bound delayed ack.
2986 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2990 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2996 ato = min(ato, max_ato);
2999 /* Stay within the limit we were given */
3000 timeout = jiffies + ato;
3002 /* Use new timeout only if there wasn't a older one earlier. */
3003 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3004 /* If delack timer was blocked or is about to expire,
3007 if (icsk->icsk_ack.blocked ||
3008 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3013 if (!time_before(timeout, icsk->icsk_ack.timeout))
3014 timeout = icsk->icsk_ack.timeout;
3016 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3017 icsk->icsk_ack.timeout = timeout;
3018 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3021 /* This routine sends an ack and also updates the window. */
3022 void tcp_send_ack(struct sock *sk)
3024 struct sk_buff *buff;
3026 /* If we have been reset, we may not send again. */
3027 if (sk->sk_state == TCP_CLOSE)
3030 /* We are not putting this on the write queue, so
3031 * tcp_transmit_skb() will set the ownership to this
3034 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3036 inet_csk_schedule_ack(sk);
3037 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3038 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3039 TCP_DELACK_MAX, TCP_RTO_MAX);
3043 /* Reserve space for headers and prepare control bits. */
3044 skb_reserve(buff, MAX_TCP_HEADER);
3045 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3047 /* Send it off, this clears delayed acks for us. */
3048 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3049 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3052 /* This routine sends a packet with an out of date sequence
3053 * number. It assumes the other end will try to ack it.
3055 * Question: what should we make while urgent mode?
3056 * 4.4BSD forces sending single byte of data. We cannot send
3057 * out of window data, because we have SND.NXT==SND.MAX...
3059 * Current solution: to send TWO zero-length segments in urgent mode:
3060 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3061 * out-of-date with SND.UNA-1 to probe window.
3063 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3065 struct tcp_sock *tp = tcp_sk(sk);
3066 struct sk_buff *skb;
3068 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3069 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3073 /* Reserve space for headers and set control bits. */
3074 skb_reserve(skb, MAX_TCP_HEADER);
3075 /* Use a previous sequence. This should cause the other
3076 * end to send an ack. Don't queue or clone SKB, just
3079 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3080 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3081 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3084 void tcp_send_window_probe(struct sock *sk)
3086 if (sk->sk_state == TCP_ESTABLISHED) {
3087 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3088 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3089 tcp_xmit_probe_skb(sk, 0);
3093 /* Initiate keepalive or window probe from timer. */
3094 int tcp_write_wakeup(struct sock *sk)
3096 struct tcp_sock *tp = tcp_sk(sk);
3097 struct sk_buff *skb;
3099 if (sk->sk_state == TCP_CLOSE)
3102 if ((skb = tcp_send_head(sk)) != NULL &&
3103 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3105 unsigned int mss = tcp_current_mss(sk);
3106 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3108 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3109 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3111 /* We are probing the opening of a window
3112 * but the window size is != 0
3113 * must have been a result SWS avoidance ( sender )
3115 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3117 seg_size = min(seg_size, mss);
3118 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3119 if (tcp_fragment(sk, skb, seg_size, mss))
3121 } else if (!tcp_skb_pcount(skb))
3122 tcp_set_skb_tso_segs(sk, skb, mss);
3124 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3125 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3126 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3128 tcp_event_new_data_sent(sk, skb);
3131 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3132 tcp_xmit_probe_skb(sk, 1);
3133 return tcp_xmit_probe_skb(sk, 0);
3137 /* A window probe timeout has occurred. If window is not closed send
3138 * a partial packet else a zero probe.
3140 void tcp_send_probe0(struct sock *sk)
3142 struct inet_connection_sock *icsk = inet_csk(sk);
3143 struct tcp_sock *tp = tcp_sk(sk);
3146 err = tcp_write_wakeup(sk);
3148 if (tp->packets_out || !tcp_send_head(sk)) {
3149 /* Cancel probe timer, if it is not required. */
3150 icsk->icsk_probes_out = 0;
3151 icsk->icsk_backoff = 0;
3156 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3157 icsk->icsk_backoff++;
3158 icsk->icsk_probes_out++;
3159 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3160 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3163 /* If packet was not sent due to local congestion,
3164 * do not backoff and do not remember icsk_probes_out.
3165 * Let local senders to fight for local resources.
3167 * Use accumulated backoff yet.
3169 if (!icsk->icsk_probes_out)
3170 icsk->icsk_probes_out = 1;
3171 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3172 min(icsk->icsk_rto << icsk->icsk_backoff,
3173 TCP_RESOURCE_PROBE_INTERVAL),