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 static 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 if (unlikely(skb_cloned(skb)))
850 skb = pskb_copy(skb, gfp_mask);
852 skb = skb_clone(skb, gfp_mask);
859 tcb = TCP_SKB_CB(skb);
860 memset(&opts, 0, sizeof(opts));
862 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
863 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
865 tcp_options_size = tcp_established_options(sk, skb, &opts,
867 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
869 if (tcp_packets_in_flight(tp) == 0) {
870 tcp_ca_event(sk, CA_EVENT_TX_START);
875 skb_push(skb, tcp_header_size);
876 skb_reset_transport_header(skb);
880 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
881 tcp_wfree : sock_wfree;
882 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
884 /* Build TCP header and checksum it. */
886 th->source = inet->inet_sport;
887 th->dest = inet->inet_dport;
888 th->seq = htonl(tcb->seq);
889 th->ack_seq = htonl(tp->rcv_nxt);
890 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
893 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
894 /* RFC1323: The window in SYN & SYN/ACK segments
897 th->window = htons(min(tp->rcv_wnd, 65535U));
899 th->window = htons(tcp_select_window(sk));
904 /* The urg_mode check is necessary during a below snd_una win probe */
905 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
906 if (before(tp->snd_up, tcb->seq + 0x10000)) {
907 th->urg_ptr = htons(tp->snd_up - tcb->seq);
909 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
910 th->urg_ptr = htons(0xFFFF);
915 tcp_options_write((__be32 *)(th + 1), tp, &opts);
916 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
917 TCP_ECN_send(sk, skb, tcp_header_size);
919 #ifdef CONFIG_TCP_MD5SIG
920 /* Calculate the MD5 hash, as we have all we need now */
922 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
923 tp->af_specific->calc_md5_hash(opts.hash_location,
928 icsk->icsk_af_ops->send_check(sk, skb);
930 if (likely(tcb->tcp_flags & TCPHDR_ACK))
931 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
933 if (skb->len != tcp_header_size)
934 tcp_event_data_sent(tp, sk);
936 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
937 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
938 tcp_skb_pcount(skb));
940 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
941 if (likely(err <= 0))
944 tcp_enter_cwr(sk, 1);
946 return net_xmit_eval(err);
949 /* This routine just queues the buffer for sending.
951 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
952 * otherwise socket can stall.
954 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
956 struct tcp_sock *tp = tcp_sk(sk);
958 /* Advance write_seq and place onto the write_queue. */
959 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
960 skb_header_release(skb);
961 tcp_add_write_queue_tail(sk, skb);
962 sk->sk_wmem_queued += skb->truesize;
963 sk_mem_charge(sk, skb->truesize);
966 /* Initialize TSO segments for a packet. */
967 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
968 unsigned int mss_now)
970 if (skb->len <= mss_now || !sk_can_gso(sk) ||
971 skb->ip_summed == CHECKSUM_NONE) {
972 /* Avoid the costly divide in the normal
975 skb_shinfo(skb)->gso_segs = 1;
976 skb_shinfo(skb)->gso_size = 0;
977 skb_shinfo(skb)->gso_type = 0;
979 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
980 skb_shinfo(skb)->gso_size = mss_now;
981 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
985 /* When a modification to fackets out becomes necessary, we need to check
986 * skb is counted to fackets_out or not.
988 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
991 struct tcp_sock *tp = tcp_sk(sk);
993 if (!tp->sacked_out || tcp_is_reno(tp))
996 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
997 tp->fackets_out -= decr;
1000 /* Pcount in the middle of the write queue got changed, we need to do various
1001 * tweaks to fix counters
1003 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1005 struct tcp_sock *tp = tcp_sk(sk);
1007 tp->packets_out -= decr;
1009 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1010 tp->sacked_out -= decr;
1011 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1012 tp->retrans_out -= decr;
1013 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1014 tp->lost_out -= decr;
1016 /* Reno case is special. Sigh... */
1017 if (tcp_is_reno(tp) && decr > 0)
1018 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1020 tcp_adjust_fackets_out(sk, skb, decr);
1022 if (tp->lost_skb_hint &&
1023 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1024 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1025 tp->lost_cnt_hint -= decr;
1027 tcp_verify_left_out(tp);
1030 /* Function to create two new TCP segments. Shrinks the given segment
1031 * to the specified size and appends a new segment with the rest of the
1032 * packet to the list. This won't be called frequently, I hope.
1033 * Remember, these are still headerless SKBs at this point.
1035 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1036 unsigned int mss_now)
1038 struct tcp_sock *tp = tcp_sk(sk);
1039 struct sk_buff *buff;
1040 int nsize, old_factor;
1044 if (WARN_ON(len > skb->len))
1047 nsize = skb_headlen(skb) - len;
1051 if (skb_cloned(skb) &&
1052 skb_is_nonlinear(skb) &&
1053 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1056 /* Get a new skb... force flag on. */
1057 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1059 return -ENOMEM; /* We'll just try again later. */
1061 sk->sk_wmem_queued += buff->truesize;
1062 sk_mem_charge(sk, buff->truesize);
1063 nlen = skb->len - len - nsize;
1064 buff->truesize += nlen;
1065 skb->truesize -= nlen;
1067 /* Correct the sequence numbers. */
1068 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1069 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1070 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1072 /* PSH and FIN should only be set in the second packet. */
1073 flags = TCP_SKB_CB(skb)->tcp_flags;
1074 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1075 TCP_SKB_CB(buff)->tcp_flags = flags;
1076 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1078 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1079 /* Copy and checksum data tail into the new buffer. */
1080 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1081 skb_put(buff, nsize),
1086 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1088 skb->ip_summed = CHECKSUM_PARTIAL;
1089 skb_split(skb, buff, len);
1092 buff->ip_summed = skb->ip_summed;
1094 /* Looks stupid, but our code really uses when of
1095 * skbs, which it never sent before. --ANK
1097 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1098 buff->tstamp = skb->tstamp;
1100 old_factor = tcp_skb_pcount(skb);
1102 /* Fix up tso_factor for both original and new SKB. */
1103 tcp_set_skb_tso_segs(sk, skb, mss_now);
1104 tcp_set_skb_tso_segs(sk, buff, mss_now);
1106 /* If this packet has been sent out already, we must
1107 * adjust the various packet counters.
1109 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1110 int diff = old_factor - tcp_skb_pcount(skb) -
1111 tcp_skb_pcount(buff);
1114 tcp_adjust_pcount(sk, skb, diff);
1117 /* Link BUFF into the send queue. */
1118 skb_header_release(buff);
1119 tcp_insert_write_queue_after(skb, buff, sk);
1124 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1125 * eventually). The difference is that pulled data not copied, but
1126 * immediately discarded.
1128 static void __pskb_trim_head(struct sk_buff *skb, int len)
1132 eat = min_t(int, len, skb_headlen(skb));
1134 __skb_pull(skb, eat);
1141 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1142 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1145 skb_frag_unref(skb, i);
1148 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1150 skb_shinfo(skb)->frags[k].page_offset += eat;
1151 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1157 skb_shinfo(skb)->nr_frags = k;
1159 skb_reset_tail_pointer(skb);
1160 skb->data_len -= len;
1161 skb->len = skb->data_len;
1164 /* Remove acked data from a packet in the transmit queue. */
1165 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1167 if (skb_unclone(skb, GFP_ATOMIC))
1170 __pskb_trim_head(skb, len);
1172 TCP_SKB_CB(skb)->seq += len;
1173 skb->ip_summed = CHECKSUM_PARTIAL;
1175 skb->truesize -= len;
1176 sk->sk_wmem_queued -= len;
1177 sk_mem_uncharge(sk, len);
1178 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1180 /* Any change of skb->len requires recalculation of tso factor. */
1181 if (tcp_skb_pcount(skb) > 1)
1182 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1187 /* Calculate MSS not accounting any TCP options. */
1188 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1190 const struct tcp_sock *tp = tcp_sk(sk);
1191 const struct inet_connection_sock *icsk = inet_csk(sk);
1194 /* Calculate base mss without TCP options:
1195 It is MMS_S - sizeof(tcphdr) of rfc1122
1197 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1199 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1200 if (icsk->icsk_af_ops->net_frag_header_len) {
1201 const struct dst_entry *dst = __sk_dst_get(sk);
1203 if (dst && dst_allfrag(dst))
1204 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1207 /* Clamp it (mss_clamp does not include tcp options) */
1208 if (mss_now > tp->rx_opt.mss_clamp)
1209 mss_now = tp->rx_opt.mss_clamp;
1211 /* Now subtract optional transport overhead */
1212 mss_now -= icsk->icsk_ext_hdr_len;
1214 /* Then reserve room for full set of TCP options and 8 bytes of data */
1220 /* Calculate MSS. Not accounting for SACKs here. */
1221 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1223 /* Subtract TCP options size, not including SACKs */
1224 return __tcp_mtu_to_mss(sk, pmtu) -
1225 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1228 /* Inverse of above */
1229 int tcp_mss_to_mtu(struct sock *sk, int mss)
1231 const struct tcp_sock *tp = tcp_sk(sk);
1232 const struct inet_connection_sock *icsk = inet_csk(sk);
1236 tp->tcp_header_len +
1237 icsk->icsk_ext_hdr_len +
1238 icsk->icsk_af_ops->net_header_len;
1240 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1241 if (icsk->icsk_af_ops->net_frag_header_len) {
1242 const struct dst_entry *dst = __sk_dst_get(sk);
1244 if (dst && dst_allfrag(dst))
1245 mtu += icsk->icsk_af_ops->net_frag_header_len;
1250 /* MTU probing init per socket */
1251 void tcp_mtup_init(struct sock *sk)
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 struct inet_connection_sock *icsk = inet_csk(sk);
1256 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1257 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1258 icsk->icsk_af_ops->net_header_len;
1259 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1260 icsk->icsk_mtup.probe_size = 0;
1262 EXPORT_SYMBOL(tcp_mtup_init);
1264 /* This function synchronize snd mss to current pmtu/exthdr set.
1266 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1267 for TCP options, but includes only bare TCP header.
1269 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1270 It is minimum of user_mss and mss received with SYN.
1271 It also does not include TCP options.
1273 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1275 tp->mss_cache is current effective sending mss, including
1276 all tcp options except for SACKs. It is evaluated,
1277 taking into account current pmtu, but never exceeds
1278 tp->rx_opt.mss_clamp.
1280 NOTE1. rfc1122 clearly states that advertised MSS
1281 DOES NOT include either tcp or ip options.
1283 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1284 are READ ONLY outside this function. --ANK (980731)
1286 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1288 struct tcp_sock *tp = tcp_sk(sk);
1289 struct inet_connection_sock *icsk = inet_csk(sk);
1292 if (icsk->icsk_mtup.search_high > pmtu)
1293 icsk->icsk_mtup.search_high = pmtu;
1295 mss_now = tcp_mtu_to_mss(sk, pmtu);
1296 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1298 /* And store cached results */
1299 icsk->icsk_pmtu_cookie = pmtu;
1300 if (icsk->icsk_mtup.enabled)
1301 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1302 tp->mss_cache = mss_now;
1306 EXPORT_SYMBOL(tcp_sync_mss);
1308 /* Compute the current effective MSS, taking SACKs and IP options,
1309 * and even PMTU discovery events into account.
1311 unsigned int tcp_current_mss(struct sock *sk)
1313 const struct tcp_sock *tp = tcp_sk(sk);
1314 const struct dst_entry *dst = __sk_dst_get(sk);
1316 unsigned int header_len;
1317 struct tcp_out_options opts;
1318 struct tcp_md5sig_key *md5;
1320 mss_now = tp->mss_cache;
1323 u32 mtu = dst_mtu(dst);
1324 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1325 mss_now = tcp_sync_mss(sk, mtu);
1328 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1329 sizeof(struct tcphdr);
1330 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1331 * some common options. If this is an odd packet (because we have SACK
1332 * blocks etc) then our calculated header_len will be different, and
1333 * we have to adjust mss_now correspondingly */
1334 if (header_len != tp->tcp_header_len) {
1335 int delta = (int) header_len - tp->tcp_header_len;
1342 /* Congestion window validation. (RFC2861) */
1343 static void tcp_cwnd_validate(struct sock *sk)
1345 struct tcp_sock *tp = tcp_sk(sk);
1347 if (tp->packets_out >= tp->snd_cwnd) {
1348 /* Network is feed fully. */
1349 tp->snd_cwnd_used = 0;
1350 tp->snd_cwnd_stamp = tcp_time_stamp;
1352 /* Network starves. */
1353 if (tp->packets_out > tp->snd_cwnd_used)
1354 tp->snd_cwnd_used = tp->packets_out;
1356 if (sysctl_tcp_slow_start_after_idle &&
1357 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1358 tcp_cwnd_application_limited(sk);
1362 /* Returns the portion of skb which can be sent right away without
1363 * introducing MSS oddities to segment boundaries. In rare cases where
1364 * mss_now != mss_cache, we will request caller to create a small skb
1365 * per input skb which could be mostly avoided here (if desired).
1367 * We explicitly want to create a request for splitting write queue tail
1368 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1369 * thus all the complexity (cwnd_len is always MSS multiple which we
1370 * return whenever allowed by the other factors). Basically we need the
1371 * modulo only when the receiver window alone is the limiting factor or
1372 * when we would be allowed to send the split-due-to-Nagle skb fully.
1374 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1375 unsigned int mss_now, unsigned int max_segs)
1377 const struct tcp_sock *tp = tcp_sk(sk);
1378 u32 needed, window, max_len;
1380 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1381 max_len = mss_now * max_segs;
1383 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1386 needed = min(skb->len, window);
1388 if (max_len <= needed)
1391 return needed - needed % mss_now;
1394 /* Can at least one segment of SKB be sent right now, according to the
1395 * congestion window rules? If so, return how many segments are allowed.
1397 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1398 const struct sk_buff *skb)
1400 u32 in_flight, cwnd;
1402 /* Don't be strict about the congestion window for the final FIN. */
1403 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1404 tcp_skb_pcount(skb) == 1)
1407 in_flight = tcp_packets_in_flight(tp);
1408 cwnd = tp->snd_cwnd;
1409 if (in_flight < cwnd)
1410 return (cwnd - in_flight);
1415 /* Initialize TSO state of a skb.
1416 * This must be invoked the first time we consider transmitting
1417 * SKB onto the wire.
1419 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1420 unsigned int mss_now)
1422 int tso_segs = tcp_skb_pcount(skb);
1424 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1425 tcp_set_skb_tso_segs(sk, skb, mss_now);
1426 tso_segs = tcp_skb_pcount(skb);
1431 /* Minshall's variant of the Nagle send check. */
1432 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1434 return after(tp->snd_sml, tp->snd_una) &&
1435 !after(tp->snd_sml, tp->snd_nxt);
1438 /* Return false, if packet can be sent now without violation Nagle's rules:
1439 * 1. It is full sized.
1440 * 2. Or it contains FIN. (already checked by caller)
1441 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1442 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1443 * With Minshall's modification: all sent small packets are ACKed.
1445 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1446 const struct sk_buff *skb,
1447 unsigned int mss_now, int nonagle)
1449 return skb->len < mss_now &&
1450 ((nonagle & TCP_NAGLE_CORK) ||
1451 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1454 /* Return true if the Nagle test allows this packet to be
1457 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1458 unsigned int cur_mss, int nonagle)
1460 /* Nagle rule does not apply to frames, which sit in the middle of the
1461 * write_queue (they have no chances to get new data).
1463 * This is implemented in the callers, where they modify the 'nonagle'
1464 * argument based upon the location of SKB in the send queue.
1466 if (nonagle & TCP_NAGLE_PUSH)
1469 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1470 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1473 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1479 /* Does at least the first segment of SKB fit into the send window? */
1480 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1481 const struct sk_buff *skb,
1482 unsigned int cur_mss)
1484 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1486 if (skb->len > cur_mss)
1487 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1489 return !after(end_seq, tcp_wnd_end(tp));
1492 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1493 * should be put on the wire right now. If so, it returns the number of
1494 * packets allowed by the congestion window.
1496 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1497 unsigned int cur_mss, int nonagle)
1499 const struct tcp_sock *tp = tcp_sk(sk);
1500 unsigned int cwnd_quota;
1502 tcp_init_tso_segs(sk, skb, cur_mss);
1504 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1507 cwnd_quota = tcp_cwnd_test(tp, skb);
1508 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1514 /* Test if sending is allowed right now. */
1515 bool tcp_may_send_now(struct sock *sk)
1517 const struct tcp_sock *tp = tcp_sk(sk);
1518 struct sk_buff *skb = tcp_send_head(sk);
1521 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1522 (tcp_skb_is_last(sk, skb) ?
1523 tp->nonagle : TCP_NAGLE_PUSH));
1526 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1527 * which is put after SKB on the list. It is very much like
1528 * tcp_fragment() except that it may make several kinds of assumptions
1529 * in order to speed up the splitting operation. In particular, we
1530 * know that all the data is in scatter-gather pages, and that the
1531 * packet has never been sent out before (and thus is not cloned).
1533 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1534 unsigned int mss_now, gfp_t gfp)
1536 struct sk_buff *buff;
1537 int nlen = skb->len - len;
1540 /* All of a TSO frame must be composed of paged data. */
1541 if (skb->len != skb->data_len)
1542 return tcp_fragment(sk, skb, len, mss_now);
1544 buff = sk_stream_alloc_skb(sk, 0, gfp);
1545 if (unlikely(buff == NULL))
1548 sk->sk_wmem_queued += buff->truesize;
1549 sk_mem_charge(sk, buff->truesize);
1550 buff->truesize += nlen;
1551 skb->truesize -= nlen;
1553 /* Correct the sequence numbers. */
1554 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1555 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1556 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1558 /* PSH and FIN should only be set in the second packet. */
1559 flags = TCP_SKB_CB(skb)->tcp_flags;
1560 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1561 TCP_SKB_CB(buff)->tcp_flags = flags;
1563 /* This packet was never sent out yet, so no SACK bits. */
1564 TCP_SKB_CB(buff)->sacked = 0;
1566 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1567 skb_split(skb, buff, len);
1569 /* Fix up tso_factor for both original and new SKB. */
1570 tcp_set_skb_tso_segs(sk, skb, mss_now);
1571 tcp_set_skb_tso_segs(sk, buff, mss_now);
1573 /* Link BUFF into the send queue. */
1574 skb_header_release(buff);
1575 tcp_insert_write_queue_after(skb, buff, sk);
1580 /* Try to defer sending, if possible, in order to minimize the amount
1581 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1583 * This algorithm is from John Heffner.
1585 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1587 struct tcp_sock *tp = tcp_sk(sk);
1588 const struct inet_connection_sock *icsk = inet_csk(sk);
1589 u32 send_win, cong_win, limit, in_flight;
1592 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1595 if (icsk->icsk_ca_state != TCP_CA_Open)
1598 /* Defer for less than two clock ticks. */
1599 if (tp->tso_deferred &&
1600 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1603 in_flight = tcp_packets_in_flight(tp);
1605 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1607 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1609 /* From in_flight test above, we know that cwnd > in_flight. */
1610 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1612 limit = min(send_win, cong_win);
1614 /* If a full-sized TSO skb can be sent, do it. */
1615 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1616 sk->sk_gso_max_segs * tp->mss_cache))
1619 /* Middle in queue won't get any more data, full sendable already? */
1620 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1623 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1625 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1627 /* If at least some fraction of a window is available,
1630 chunk /= win_divisor;
1634 /* Different approach, try not to defer past a single
1635 * ACK. Receiver should ACK every other full sized
1636 * frame, so if we have space for more than 3 frames
1639 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1643 /* Ok, it looks like it is advisable to defer. */
1644 tp->tso_deferred = 1 | (jiffies << 1);
1649 tp->tso_deferred = 0;
1653 /* Create a new MTU probe if we are ready.
1654 * MTU probe is regularly attempting to increase the path MTU by
1655 * deliberately sending larger packets. This discovers routing
1656 * changes resulting in larger path MTUs.
1658 * Returns 0 if we should wait to probe (no cwnd available),
1659 * 1 if a probe was sent,
1662 static int tcp_mtu_probe(struct sock *sk)
1664 struct tcp_sock *tp = tcp_sk(sk);
1665 struct inet_connection_sock *icsk = inet_csk(sk);
1666 struct sk_buff *skb, *nskb, *next;
1673 /* Not currently probing/verifying,
1675 * have enough cwnd, and
1676 * not SACKing (the variable headers throw things off) */
1677 if (!icsk->icsk_mtup.enabled ||
1678 icsk->icsk_mtup.probe_size ||
1679 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1680 tp->snd_cwnd < 11 ||
1681 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1684 /* Very simple search strategy: just double the MSS. */
1685 mss_now = tcp_current_mss(sk);
1686 probe_size = 2 * tp->mss_cache;
1687 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1688 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1689 /* TODO: set timer for probe_converge_event */
1693 /* Have enough data in the send queue to probe? */
1694 if (tp->write_seq - tp->snd_nxt < size_needed)
1697 if (tp->snd_wnd < size_needed)
1699 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1702 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1703 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1704 if (!tcp_packets_in_flight(tp))
1710 /* We're allowed to probe. Build it now. */
1711 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1713 sk->sk_wmem_queued += nskb->truesize;
1714 sk_mem_charge(sk, nskb->truesize);
1716 skb = tcp_send_head(sk);
1718 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1719 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1720 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1721 TCP_SKB_CB(nskb)->sacked = 0;
1723 nskb->ip_summed = skb->ip_summed;
1725 tcp_insert_write_queue_before(nskb, skb, sk);
1728 tcp_for_write_queue_from_safe(skb, next, sk) {
1729 copy = min_t(int, skb->len, probe_size - len);
1730 if (nskb->ip_summed)
1731 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1733 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1734 skb_put(nskb, copy),
1737 if (skb->len <= copy) {
1738 /* We've eaten all the data from this skb.
1740 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1741 tcp_unlink_write_queue(skb, sk);
1742 sk_wmem_free_skb(sk, skb);
1744 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1745 ~(TCPHDR_FIN|TCPHDR_PSH);
1746 if (!skb_shinfo(skb)->nr_frags) {
1747 skb_pull(skb, copy);
1748 if (skb->ip_summed != CHECKSUM_PARTIAL)
1749 skb->csum = csum_partial(skb->data,
1752 __pskb_trim_head(skb, copy);
1753 tcp_set_skb_tso_segs(sk, skb, mss_now);
1755 TCP_SKB_CB(skb)->seq += copy;
1760 if (len >= probe_size)
1763 tcp_init_tso_segs(sk, nskb, nskb->len);
1765 /* We're ready to send. If this fails, the probe will
1766 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1767 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1768 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1769 /* Decrement cwnd here because we are sending
1770 * effectively two packets. */
1772 tcp_event_new_data_sent(sk, nskb);
1774 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1775 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1776 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1784 /* This routine writes packets to the network. It advances the
1785 * send_head. This happens as incoming acks open up the remote
1788 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1789 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1790 * account rare use of URG, this is not a big flaw.
1792 * Send at most one packet when push_one > 0. Temporarily ignore
1793 * cwnd limit to force at most one packet out when push_one == 2.
1795 * Returns true, if no segments are in flight and we have queued segments,
1796 * but cannot send anything now because of SWS or another problem.
1798 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1799 int push_one, gfp_t gfp)
1801 struct tcp_sock *tp = tcp_sk(sk);
1802 struct sk_buff *skb;
1803 unsigned int tso_segs, sent_pkts;
1810 /* Do MTU probing. */
1811 result = tcp_mtu_probe(sk);
1814 } else if (result > 0) {
1819 while ((skb = tcp_send_head(sk))) {
1823 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1826 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1827 goto repair; /* Skip network transmission */
1829 cwnd_quota = tcp_cwnd_test(tp, skb);
1832 /* Force out a loss probe pkt. */
1838 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1841 if (tso_segs == 1) {
1842 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1843 (tcp_skb_is_last(sk, skb) ?
1844 nonagle : TCP_NAGLE_PUSH))))
1847 if (!push_one && tcp_tso_should_defer(sk, skb))
1851 /* TSQ : sk_wmem_alloc accounts skb truesize,
1852 * including skb overhead. But thats OK.
1854 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
1855 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1859 if (tso_segs > 1 && !tcp_urg_mode(tp))
1860 limit = tcp_mss_split_point(sk, skb, mss_now,
1863 sk->sk_gso_max_segs));
1865 if (skb->len > limit &&
1866 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1869 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1871 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1875 /* Advance the send_head. This one is sent out.
1876 * This call will increment packets_out.
1878 tcp_event_new_data_sent(sk, skb);
1880 tcp_minshall_update(tp, mss_now, skb);
1881 sent_pkts += tcp_skb_pcount(skb);
1887 if (likely(sent_pkts)) {
1888 if (tcp_in_cwnd_reduction(sk))
1889 tp->prr_out += sent_pkts;
1891 /* Send one loss probe per tail loss episode. */
1893 tcp_schedule_loss_probe(sk);
1894 tcp_cwnd_validate(sk);
1897 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1900 bool tcp_schedule_loss_probe(struct sock *sk)
1902 struct inet_connection_sock *icsk = inet_csk(sk);
1903 struct tcp_sock *tp = tcp_sk(sk);
1904 u32 timeout, tlp_time_stamp, rto_time_stamp;
1905 u32 rtt = tp->srtt >> 3;
1907 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1909 /* No consecutive loss probes. */
1910 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1914 /* Don't do any loss probe on a Fast Open connection before 3WHS
1917 if (sk->sk_state == TCP_SYN_RECV)
1920 /* TLP is only scheduled when next timer event is RTO. */
1921 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1924 /* Schedule a loss probe in 2*RTT for SACK capable connections
1925 * in Open state, that are either limited by cwnd or application.
1927 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1928 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1931 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1935 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1936 * for delayed ack when there's one outstanding packet.
1939 if (tp->packets_out == 1)
1940 timeout = max_t(u32, timeout,
1941 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1942 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1944 /* If RTO is shorter, just schedule TLP in its place. */
1945 tlp_time_stamp = tcp_time_stamp + timeout;
1946 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1947 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1948 s32 delta = rto_time_stamp - tcp_time_stamp;
1953 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1958 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1959 * retransmit the last segment.
1961 void tcp_send_loss_probe(struct sock *sk)
1963 struct tcp_sock *tp = tcp_sk(sk);
1964 struct sk_buff *skb;
1966 int mss = tcp_current_mss(sk);
1969 if (tcp_send_head(sk) != NULL) {
1970 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1974 /* At most one outstanding TLP retransmission. */
1975 if (tp->tlp_high_seq)
1978 /* Retransmit last segment. */
1979 skb = tcp_write_queue_tail(sk);
1983 pcount = tcp_skb_pcount(skb);
1984 if (WARN_ON(!pcount))
1987 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
1988 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
1990 skb = tcp_write_queue_tail(sk);
1993 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
1996 /* Probe with zero data doesn't trigger fast recovery. */
1998 err = __tcp_retransmit_skb(sk, skb);
2000 /* Record snd_nxt for loss detection. */
2002 tp->tlp_high_seq = tp->snd_nxt;
2005 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2006 inet_csk(sk)->icsk_rto,
2010 NET_INC_STATS_BH(sock_net(sk),
2011 LINUX_MIB_TCPLOSSPROBES);
2015 /* Push out any pending frames which were held back due to
2016 * TCP_CORK or attempt at coalescing tiny packets.
2017 * The socket must be locked by the caller.
2019 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2022 /* If we are closed, the bytes will have to remain here.
2023 * In time closedown will finish, we empty the write queue and
2024 * all will be happy.
2026 if (unlikely(sk->sk_state == TCP_CLOSE))
2029 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2030 sk_gfp_atomic(sk, GFP_ATOMIC)))
2031 tcp_check_probe_timer(sk);
2034 /* Send _single_ skb sitting at the send head. This function requires
2035 * true push pending frames to setup probe timer etc.
2037 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2039 struct sk_buff *skb = tcp_send_head(sk);
2041 BUG_ON(!skb || skb->len < mss_now);
2043 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2046 /* This function returns the amount that we can raise the
2047 * usable window based on the following constraints
2049 * 1. The window can never be shrunk once it is offered (RFC 793)
2050 * 2. We limit memory per socket
2053 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2054 * RECV.NEXT + RCV.WIN fixed until:
2055 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2057 * i.e. don't raise the right edge of the window until you can raise
2058 * it at least MSS bytes.
2060 * Unfortunately, the recommended algorithm breaks header prediction,
2061 * since header prediction assumes th->window stays fixed.
2063 * Strictly speaking, keeping th->window fixed violates the receiver
2064 * side SWS prevention criteria. The problem is that under this rule
2065 * a stream of single byte packets will cause the right side of the
2066 * window to always advance by a single byte.
2068 * Of course, if the sender implements sender side SWS prevention
2069 * then this will not be a problem.
2071 * BSD seems to make the following compromise:
2073 * If the free space is less than the 1/4 of the maximum
2074 * space available and the free space is less than 1/2 mss,
2075 * then set the window to 0.
2076 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2077 * Otherwise, just prevent the window from shrinking
2078 * and from being larger than the largest representable value.
2080 * This prevents incremental opening of the window in the regime
2081 * where TCP is limited by the speed of the reader side taking
2082 * data out of the TCP receive queue. It does nothing about
2083 * those cases where the window is constrained on the sender side
2084 * because the pipeline is full.
2086 * BSD also seems to "accidentally" limit itself to windows that are a
2087 * multiple of MSS, at least until the free space gets quite small.
2088 * This would appear to be a side effect of the mbuf implementation.
2089 * Combining these two algorithms results in the observed behavior
2090 * of having a fixed window size at almost all times.
2092 * Below we obtain similar behavior by forcing the offered window to
2093 * a multiple of the mss when it is feasible to do so.
2095 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2096 * Regular options like TIMESTAMP are taken into account.
2098 u32 __tcp_select_window(struct sock *sk)
2100 struct inet_connection_sock *icsk = inet_csk(sk);
2101 struct tcp_sock *tp = tcp_sk(sk);
2102 /* MSS for the peer's data. Previous versions used mss_clamp
2103 * here. I don't know if the value based on our guesses
2104 * of peer's MSS is better for the performance. It's more correct
2105 * but may be worse for the performance because of rcv_mss
2106 * fluctuations. --SAW 1998/11/1
2108 int mss = icsk->icsk_ack.rcv_mss;
2109 int free_space = tcp_space(sk);
2110 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2113 if (mss > full_space)
2116 if (free_space < (full_space >> 1)) {
2117 icsk->icsk_ack.quick = 0;
2119 if (sk_under_memory_pressure(sk))
2120 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2123 if (free_space < mss)
2127 if (free_space > tp->rcv_ssthresh)
2128 free_space = tp->rcv_ssthresh;
2130 /* Don't do rounding if we are using window scaling, since the
2131 * scaled window will not line up with the MSS boundary anyway.
2133 window = tp->rcv_wnd;
2134 if (tp->rx_opt.rcv_wscale) {
2135 window = free_space;
2137 /* Advertise enough space so that it won't get scaled away.
2138 * Import case: prevent zero window announcement if
2139 * 1<<rcv_wscale > mss.
2141 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2142 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2143 << tp->rx_opt.rcv_wscale);
2145 /* Get the largest window that is a nice multiple of mss.
2146 * Window clamp already applied above.
2147 * If our current window offering is within 1 mss of the
2148 * free space we just keep it. This prevents the divide
2149 * and multiply from happening most of the time.
2150 * We also don't do any window rounding when the free space
2153 if (window <= free_space - mss || window > free_space)
2154 window = (free_space / mss) * mss;
2155 else if (mss == full_space &&
2156 free_space > window + (full_space >> 1))
2157 window = free_space;
2163 /* Collapses two adjacent SKB's during retransmission. */
2164 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2166 struct tcp_sock *tp = tcp_sk(sk);
2167 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2168 int skb_size, next_skb_size;
2170 skb_size = skb->len;
2171 next_skb_size = next_skb->len;
2173 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2175 tcp_highest_sack_combine(sk, next_skb, skb);
2177 tcp_unlink_write_queue(next_skb, sk);
2179 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2182 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2183 skb->ip_summed = CHECKSUM_PARTIAL;
2185 if (skb->ip_summed != CHECKSUM_PARTIAL)
2186 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2188 /* Update sequence range on original skb. */
2189 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2191 /* Merge over control information. This moves PSH/FIN etc. over */
2192 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2194 /* All done, get rid of second SKB and account for it so
2195 * packet counting does not break.
2197 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2199 /* changed transmit queue under us so clear hints */
2200 tcp_clear_retrans_hints_partial(tp);
2201 if (next_skb == tp->retransmit_skb_hint)
2202 tp->retransmit_skb_hint = skb;
2204 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2206 sk_wmem_free_skb(sk, next_skb);
2209 /* Check if coalescing SKBs is legal. */
2210 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2212 if (tcp_skb_pcount(skb) > 1)
2214 /* TODO: SACK collapsing could be used to remove this condition */
2215 if (skb_shinfo(skb)->nr_frags != 0)
2217 if (skb_cloned(skb))
2219 if (skb == tcp_send_head(sk))
2221 /* Some heurestics for collapsing over SACK'd could be invented */
2222 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2228 /* Collapse packets in the retransmit queue to make to create
2229 * less packets on the wire. This is only done on retransmission.
2231 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2234 struct tcp_sock *tp = tcp_sk(sk);
2235 struct sk_buff *skb = to, *tmp;
2238 if (!sysctl_tcp_retrans_collapse)
2240 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2243 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2244 if (!tcp_can_collapse(sk, skb))
2256 /* Punt if not enough space exists in the first SKB for
2257 * the data in the second
2259 if (skb->len > skb_availroom(to))
2262 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2265 tcp_collapse_retrans(sk, to);
2269 /* This retransmits one SKB. Policy decisions and retransmit queue
2270 * state updates are done by the caller. Returns non-zero if an
2271 * error occurred which prevented the send.
2273 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2275 struct tcp_sock *tp = tcp_sk(sk);
2276 struct inet_connection_sock *icsk = inet_csk(sk);
2277 unsigned int cur_mss;
2279 /* Inconslusive MTU probe */
2280 if (icsk->icsk_mtup.probe_size) {
2281 icsk->icsk_mtup.probe_size = 0;
2284 /* Do not sent more than we queued. 1/4 is reserved for possible
2285 * copying overhead: fragmentation, tunneling, mangling etc.
2287 if (atomic_read(&sk->sk_wmem_alloc) >
2288 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2291 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2292 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2294 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2298 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2299 return -EHOSTUNREACH; /* Routing failure or similar. */
2301 cur_mss = tcp_current_mss(sk);
2303 /* If receiver has shrunk his window, and skb is out of
2304 * new window, do not retransmit it. The exception is the
2305 * case, when window is shrunk to zero. In this case
2306 * our retransmit serves as a zero window probe.
2308 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2309 TCP_SKB_CB(skb)->seq != tp->snd_una)
2312 if (skb->len > cur_mss) {
2313 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2314 return -ENOMEM; /* We'll try again later. */
2316 int oldpcount = tcp_skb_pcount(skb);
2318 if (unlikely(oldpcount > 1)) {
2319 tcp_init_tso_segs(sk, skb, cur_mss);
2320 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2324 tcp_retrans_try_collapse(sk, skb, cur_mss);
2326 /* Some Solaris stacks overoptimize and ignore the FIN on a
2327 * retransmit when old data is attached. So strip it off
2328 * since it is cheap to do so and saves bytes on the network.
2331 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2332 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2333 if (!pskb_trim(skb, 0)) {
2334 /* Reuse, even though it does some unnecessary work */
2335 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2336 TCP_SKB_CB(skb)->tcp_flags);
2337 skb->ip_summed = CHECKSUM_NONE;
2341 /* Make a copy, if the first transmission SKB clone we made
2342 * is still in somebody's hands, else make a clone.
2344 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2346 /* make sure skb->data is aligned on arches that require it */
2347 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2348 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2350 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2353 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2357 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2359 struct tcp_sock *tp = tcp_sk(sk);
2360 int err = __tcp_retransmit_skb(sk, skb);
2363 /* Update global TCP statistics. */
2364 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2366 tp->total_retrans++;
2368 #if FASTRETRANS_DEBUG > 0
2369 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2370 net_dbg_ratelimited("retrans_out leaked\n");
2373 if (!tp->retrans_out)
2374 tp->lost_retrans_low = tp->snd_nxt;
2375 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2376 tp->retrans_out += tcp_skb_pcount(skb);
2378 /* Save stamp of the first retransmit. */
2379 if (!tp->retrans_stamp)
2380 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2382 tp->undo_retrans += tcp_skb_pcount(skb);
2384 /* snd_nxt is stored to detect loss of retransmitted segment,
2385 * see tcp_input.c tcp_sacktag_write_queue().
2387 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2392 /* Check if we forward retransmits are possible in the current
2393 * window/congestion state.
2395 static bool tcp_can_forward_retransmit(struct sock *sk)
2397 const struct inet_connection_sock *icsk = inet_csk(sk);
2398 const struct tcp_sock *tp = tcp_sk(sk);
2400 /* Forward retransmissions are possible only during Recovery. */
2401 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2404 /* No forward retransmissions in Reno are possible. */
2405 if (tcp_is_reno(tp))
2408 /* Yeah, we have to make difficult choice between forward transmission
2409 * and retransmission... Both ways have their merits...
2411 * For now we do not retransmit anything, while we have some new
2412 * segments to send. In the other cases, follow rule 3 for
2413 * NextSeg() specified in RFC3517.
2416 if (tcp_may_send_now(sk))
2422 /* This gets called after a retransmit timeout, and the initially
2423 * retransmitted data is acknowledged. It tries to continue
2424 * resending the rest of the retransmit queue, until either
2425 * we've sent it all or the congestion window limit is reached.
2426 * If doing SACK, the first ACK which comes back for a timeout
2427 * based retransmit packet might feed us FACK information again.
2428 * If so, we use it to avoid unnecessarily retransmissions.
2430 void tcp_xmit_retransmit_queue(struct sock *sk)
2432 const struct inet_connection_sock *icsk = inet_csk(sk);
2433 struct tcp_sock *tp = tcp_sk(sk);
2434 struct sk_buff *skb;
2435 struct sk_buff *hole = NULL;
2438 int fwd_rexmitting = 0;
2440 if (!tp->packets_out)
2444 tp->retransmit_high = tp->snd_una;
2446 if (tp->retransmit_skb_hint) {
2447 skb = tp->retransmit_skb_hint;
2448 last_lost = TCP_SKB_CB(skb)->end_seq;
2449 if (after(last_lost, tp->retransmit_high))
2450 last_lost = tp->retransmit_high;
2452 skb = tcp_write_queue_head(sk);
2453 last_lost = tp->snd_una;
2456 tcp_for_write_queue_from(skb, sk) {
2457 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2459 if (skb == tcp_send_head(sk))
2461 /* we could do better than to assign each time */
2463 tp->retransmit_skb_hint = skb;
2465 /* Assume this retransmit will generate
2466 * only one packet for congestion window
2467 * calculation purposes. This works because
2468 * tcp_retransmit_skb() will chop up the
2469 * packet to be MSS sized and all the
2470 * packet counting works out.
2472 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2475 if (fwd_rexmitting) {
2477 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2479 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2481 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2482 tp->retransmit_high = last_lost;
2483 if (!tcp_can_forward_retransmit(sk))
2485 /* Backtrack if necessary to non-L'ed skb */
2493 } else if (!(sacked & TCPCB_LOST)) {
2494 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2499 last_lost = TCP_SKB_CB(skb)->end_seq;
2500 if (icsk->icsk_ca_state != TCP_CA_Loss)
2501 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2503 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2506 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2509 if (tcp_retransmit_skb(sk, skb)) {
2510 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2513 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2515 if (tcp_in_cwnd_reduction(sk))
2516 tp->prr_out += tcp_skb_pcount(skb);
2518 if (skb == tcp_write_queue_head(sk))
2519 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2520 inet_csk(sk)->icsk_rto,
2525 /* Send a fin. The caller locks the socket for us. This cannot be
2526 * allowed to fail queueing a FIN frame under any circumstances.
2528 void tcp_send_fin(struct sock *sk)
2530 struct tcp_sock *tp = tcp_sk(sk);
2531 struct sk_buff *skb = tcp_write_queue_tail(sk);
2534 /* Optimization, tack on the FIN if we have a queue of
2535 * unsent frames. But be careful about outgoing SACKS
2538 mss_now = tcp_current_mss(sk);
2540 if (tcp_send_head(sk) != NULL) {
2541 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2542 TCP_SKB_CB(skb)->end_seq++;
2545 /* Socket is locked, keep trying until memory is available. */
2547 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2554 /* Reserve space for headers and prepare control bits. */
2555 skb_reserve(skb, MAX_TCP_HEADER);
2556 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2557 tcp_init_nondata_skb(skb, tp->write_seq,
2558 TCPHDR_ACK | TCPHDR_FIN);
2559 tcp_queue_skb(sk, skb);
2561 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2564 /* We get here when a process closes a file descriptor (either due to
2565 * an explicit close() or as a byproduct of exit()'ing) and there
2566 * was unread data in the receive queue. This behavior is recommended
2567 * by RFC 2525, section 2.17. -DaveM
2569 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2571 struct sk_buff *skb;
2573 /* NOTE: No TCP options attached and we never retransmit this. */
2574 skb = alloc_skb(MAX_TCP_HEADER, priority);
2576 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2580 /* Reserve space for headers and prepare control bits. */
2581 skb_reserve(skb, MAX_TCP_HEADER);
2582 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2583 TCPHDR_ACK | TCPHDR_RST);
2585 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2586 if (tcp_transmit_skb(sk, skb, 0, priority))
2587 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2589 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2592 /* Send a crossed SYN-ACK during socket establishment.
2593 * WARNING: This routine must only be called when we have already sent
2594 * a SYN packet that crossed the incoming SYN that caused this routine
2595 * to get called. If this assumption fails then the initial rcv_wnd
2596 * and rcv_wscale values will not be correct.
2598 int tcp_send_synack(struct sock *sk)
2600 struct sk_buff *skb;
2602 skb = tcp_write_queue_head(sk);
2603 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2604 pr_debug("%s: wrong queue state\n", __func__);
2607 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2608 if (skb_cloned(skb)) {
2609 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2612 tcp_unlink_write_queue(skb, sk);
2613 skb_header_release(nskb);
2614 __tcp_add_write_queue_head(sk, nskb);
2615 sk_wmem_free_skb(sk, skb);
2616 sk->sk_wmem_queued += nskb->truesize;
2617 sk_mem_charge(sk, nskb->truesize);
2621 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2622 TCP_ECN_send_synack(tcp_sk(sk), skb);
2624 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2625 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2629 * tcp_make_synack - Prepare a SYN-ACK.
2630 * sk: listener socket
2631 * dst: dst entry attached to the SYNACK
2632 * req: request_sock pointer
2634 * Allocate one skb and build a SYNACK packet.
2635 * @dst is consumed : Caller should not use it again.
2637 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2638 struct request_sock *req,
2639 struct tcp_fastopen_cookie *foc)
2641 struct tcp_out_options opts;
2642 struct inet_request_sock *ireq = inet_rsk(req);
2643 struct tcp_sock *tp = tcp_sk(sk);
2645 struct sk_buff *skb;
2646 struct tcp_md5sig_key *md5;
2647 int tcp_header_size;
2650 skb = alloc_skb(MAX_TCP_HEADER + 15, sk_gfp_atomic(sk, GFP_ATOMIC));
2651 if (unlikely(!skb)) {
2655 /* Reserve space for headers. */
2656 skb_reserve(skb, MAX_TCP_HEADER);
2658 skb_dst_set(skb, dst);
2660 mss = dst_metric_advmss(dst);
2661 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2662 mss = tp->rx_opt.user_mss;
2664 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2666 /* Set this up on the first call only */
2667 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2669 /* limit the window selection if the user enforce a smaller rx buffer */
2670 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2671 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2672 req->window_clamp = tcp_full_space(sk);
2674 /* tcp_full_space because it is guaranteed to be the first packet */
2675 tcp_select_initial_window(tcp_full_space(sk),
2676 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2681 dst_metric(dst, RTAX_INITRWND));
2682 ireq->rcv_wscale = rcv_wscale;
2685 memset(&opts, 0, sizeof(opts));
2686 #ifdef CONFIG_SYN_COOKIES
2687 if (unlikely(req->cookie_ts))
2688 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2691 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2692 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2695 skb_push(skb, tcp_header_size);
2696 skb_reset_transport_header(skb);
2699 memset(th, 0, sizeof(struct tcphdr));
2702 TCP_ECN_make_synack(req, th);
2703 th->source = ireq->loc_port;
2704 th->dest = ireq->rmt_port;
2705 /* Setting of flags are superfluous here for callers (and ECE is
2706 * not even correctly set)
2708 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2709 TCPHDR_SYN | TCPHDR_ACK);
2711 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2712 /* XXX data is queued and acked as is. No buffer/window check */
2713 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2715 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2716 th->window = htons(min(req->rcv_wnd, 65535U));
2717 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2718 th->doff = (tcp_header_size >> 2);
2719 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2721 #ifdef CONFIG_TCP_MD5SIG
2722 /* Okay, we have all we need - do the md5 hash if needed */
2724 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2725 md5, NULL, req, skb);
2731 EXPORT_SYMBOL(tcp_make_synack);
2733 /* Do all connect socket setups that can be done AF independent. */
2734 void tcp_connect_init(struct sock *sk)
2736 const struct dst_entry *dst = __sk_dst_get(sk);
2737 struct tcp_sock *tp = tcp_sk(sk);
2740 /* We'll fix this up when we get a response from the other end.
2741 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2743 tp->tcp_header_len = sizeof(struct tcphdr) +
2744 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2746 #ifdef CONFIG_TCP_MD5SIG
2747 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2748 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2751 /* If user gave his TCP_MAXSEG, record it to clamp */
2752 if (tp->rx_opt.user_mss)
2753 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2756 tcp_sync_mss(sk, dst_mtu(dst));
2758 if (!tp->window_clamp)
2759 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2760 tp->advmss = dst_metric_advmss(dst);
2761 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2762 tp->advmss = tp->rx_opt.user_mss;
2764 tcp_initialize_rcv_mss(sk);
2766 /* limit the window selection if the user enforce a smaller rx buffer */
2767 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2768 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2769 tp->window_clamp = tcp_full_space(sk);
2771 tcp_select_initial_window(tcp_full_space(sk),
2772 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2775 sysctl_tcp_window_scaling,
2777 dst_metric(dst, RTAX_INITRWND));
2779 tp->rx_opt.rcv_wscale = rcv_wscale;
2780 tp->rcv_ssthresh = tp->rcv_wnd;
2783 sock_reset_flag(sk, SOCK_DONE);
2786 tp->snd_una = tp->write_seq;
2787 tp->snd_sml = tp->write_seq;
2788 tp->snd_up = tp->write_seq;
2789 tp->snd_nxt = tp->write_seq;
2791 if (likely(!tp->repair))
2793 tp->rcv_wup = tp->rcv_nxt;
2794 tp->copied_seq = tp->rcv_nxt;
2796 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2797 inet_csk(sk)->icsk_retransmits = 0;
2798 tcp_clear_retrans(tp);
2801 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2803 struct tcp_sock *tp = tcp_sk(sk);
2804 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2806 tcb->end_seq += skb->len;
2807 skb_header_release(skb);
2808 __tcp_add_write_queue_tail(sk, skb);
2809 sk->sk_wmem_queued += skb->truesize;
2810 sk_mem_charge(sk, skb->truesize);
2811 tp->write_seq = tcb->end_seq;
2812 tp->packets_out += tcp_skb_pcount(skb);
2815 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2816 * queue a data-only packet after the regular SYN, such that regular SYNs
2817 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2818 * only the SYN sequence, the data are retransmitted in the first ACK.
2819 * If cookie is not cached or other error occurs, falls back to send a
2820 * regular SYN with Fast Open cookie request option.
2822 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2824 struct tcp_sock *tp = tcp_sk(sk);
2825 struct tcp_fastopen_request *fo = tp->fastopen_req;
2826 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2827 struct sk_buff *syn_data = NULL, *data;
2828 unsigned long last_syn_loss = 0;
2830 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2831 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2832 &syn_loss, &last_syn_loss);
2833 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2835 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2836 fo->cookie.len = -1;
2840 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2841 fo->cookie.len = -1;
2842 else if (fo->cookie.len <= 0)
2845 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2846 * user-MSS. Reserve maximum option space for middleboxes that add
2847 * private TCP options. The cost is reduced data space in SYN :(
2849 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2850 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2851 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2852 MAX_TCP_OPTION_SPACE;
2854 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2856 if (syn_data == NULL)
2859 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2860 struct iovec *iov = &fo->data->msg_iov[i];
2861 unsigned char __user *from = iov->iov_base;
2862 int len = iov->iov_len;
2864 if (syn_data->len + len > space)
2865 len = space - syn_data->len;
2866 else if (i + 1 == iovlen)
2867 /* No more data pending in inet_wait_for_connect() */
2870 if (skb_add_data(syn_data, from, len))
2874 /* Queue a data-only packet after the regular SYN for retransmission */
2875 data = pskb_copy(syn_data, sk->sk_allocation);
2878 TCP_SKB_CB(data)->seq++;
2879 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2880 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2881 tcp_connect_queue_skb(sk, data);
2882 fo->copied = data->len;
2884 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2885 tp->syn_data = (fo->copied > 0);
2886 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2892 /* Send a regular SYN with Fast Open cookie request option */
2893 if (fo->cookie.len > 0)
2895 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2897 tp->syn_fastopen = 0;
2898 kfree_skb(syn_data);
2900 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2904 /* Build a SYN and send it off. */
2905 int tcp_connect(struct sock *sk)
2907 struct tcp_sock *tp = tcp_sk(sk);
2908 struct sk_buff *buff;
2911 tcp_connect_init(sk);
2913 if (unlikely(tp->repair)) {
2914 tcp_finish_connect(sk, NULL);
2918 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2919 if (unlikely(buff == NULL))
2922 /* Reserve space for headers. */
2923 skb_reserve(buff, MAX_TCP_HEADER);
2925 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2926 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2927 tcp_connect_queue_skb(sk, buff);
2928 TCP_ECN_send_syn(sk, buff);
2930 /* Send off SYN; include data in Fast Open. */
2931 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2932 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2933 if (err == -ECONNREFUSED)
2936 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2937 * in order to make this packet get counted in tcpOutSegs.
2939 tp->snd_nxt = tp->write_seq;
2940 tp->pushed_seq = tp->write_seq;
2941 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2943 /* Timer for repeating the SYN until an answer. */
2944 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2945 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2948 EXPORT_SYMBOL(tcp_connect);
2950 /* Send out a delayed ack, the caller does the policy checking
2951 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2954 void tcp_send_delayed_ack(struct sock *sk)
2956 struct inet_connection_sock *icsk = inet_csk(sk);
2957 int ato = icsk->icsk_ack.ato;
2958 unsigned long timeout;
2960 if (ato > TCP_DELACK_MIN) {
2961 const struct tcp_sock *tp = tcp_sk(sk);
2962 int max_ato = HZ / 2;
2964 if (icsk->icsk_ack.pingpong ||
2965 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2966 max_ato = TCP_DELACK_MAX;
2968 /* Slow path, intersegment interval is "high". */
2970 /* If some rtt estimate is known, use it to bound delayed ack.
2971 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2975 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2981 ato = min(ato, max_ato);
2984 /* Stay within the limit we were given */
2985 timeout = jiffies + ato;
2987 /* Use new timeout only if there wasn't a older one earlier. */
2988 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2989 /* If delack timer was blocked or is about to expire,
2992 if (icsk->icsk_ack.blocked ||
2993 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2998 if (!time_before(timeout, icsk->icsk_ack.timeout))
2999 timeout = icsk->icsk_ack.timeout;
3001 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3002 icsk->icsk_ack.timeout = timeout;
3003 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3006 /* This routine sends an ack and also updates the window. */
3007 void tcp_send_ack(struct sock *sk)
3009 struct sk_buff *buff;
3011 /* If we have been reset, we may not send again. */
3012 if (sk->sk_state == TCP_CLOSE)
3015 /* We are not putting this on the write queue, so
3016 * tcp_transmit_skb() will set the ownership to this
3019 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3021 inet_csk_schedule_ack(sk);
3022 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3023 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3024 TCP_DELACK_MAX, TCP_RTO_MAX);
3028 /* Reserve space for headers and prepare control bits. */
3029 skb_reserve(buff, MAX_TCP_HEADER);
3030 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3032 /* Send it off, this clears delayed acks for us. */
3033 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3034 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3037 /* This routine sends a packet with an out of date sequence
3038 * number. It assumes the other end will try to ack it.
3040 * Question: what should we make while urgent mode?
3041 * 4.4BSD forces sending single byte of data. We cannot send
3042 * out of window data, because we have SND.NXT==SND.MAX...
3044 * Current solution: to send TWO zero-length segments in urgent mode:
3045 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3046 * out-of-date with SND.UNA-1 to probe window.
3048 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3050 struct tcp_sock *tp = tcp_sk(sk);
3051 struct sk_buff *skb;
3053 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3054 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3058 /* Reserve space for headers and set control bits. */
3059 skb_reserve(skb, MAX_TCP_HEADER);
3060 /* Use a previous sequence. This should cause the other
3061 * end to send an ack. Don't queue or clone SKB, just
3064 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3065 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3066 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3069 void tcp_send_window_probe(struct sock *sk)
3071 if (sk->sk_state == TCP_ESTABLISHED) {
3072 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3073 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3074 tcp_xmit_probe_skb(sk, 0);
3078 /* Initiate keepalive or window probe from timer. */
3079 int tcp_write_wakeup(struct sock *sk)
3081 struct tcp_sock *tp = tcp_sk(sk);
3082 struct sk_buff *skb;
3084 if (sk->sk_state == TCP_CLOSE)
3087 if ((skb = tcp_send_head(sk)) != NULL &&
3088 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3090 unsigned int mss = tcp_current_mss(sk);
3091 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3093 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3094 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3096 /* We are probing the opening of a window
3097 * but the window size is != 0
3098 * must have been a result SWS avoidance ( sender )
3100 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3102 seg_size = min(seg_size, mss);
3103 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3104 if (tcp_fragment(sk, skb, seg_size, mss))
3106 } else if (!tcp_skb_pcount(skb))
3107 tcp_set_skb_tso_segs(sk, skb, mss);
3109 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3110 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3111 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3113 tcp_event_new_data_sent(sk, skb);
3116 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3117 tcp_xmit_probe_skb(sk, 1);
3118 return tcp_xmit_probe_skb(sk, 0);
3122 /* A window probe timeout has occurred. If window is not closed send
3123 * a partial packet else a zero probe.
3125 void tcp_send_probe0(struct sock *sk)
3127 struct inet_connection_sock *icsk = inet_csk(sk);
3128 struct tcp_sock *tp = tcp_sk(sk);
3131 err = tcp_write_wakeup(sk);
3133 if (tp->packets_out || !tcp_send_head(sk)) {
3134 /* Cancel probe timer, if it is not required. */
3135 icsk->icsk_probes_out = 0;
3136 icsk->icsk_backoff = 0;
3141 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3142 icsk->icsk_backoff++;
3143 icsk->icsk_probes_out++;
3144 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3145 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3148 /* If packet was not sent due to local congestion,
3149 * do not backoff and do not remember icsk_probes_out.
3150 * Let local senders to fight for local resources.
3152 * Use accumulated backoff yet.
3154 if (!icsk->icsk_probes_out)
3155 icsk->icsk_probes_out = 1;
3156 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3157 min(icsk->icsk_rto << icsk->icsk_backoff,
3158 TCP_RESOURCE_PROBE_INTERVAL),