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) {
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
101 return tcp_wnd_end(tp);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16 tcp_advertise_mss(struct sock *sk)
120 struct tcp_sock *tp = tcp_sk(sk);
121 const struct dst_entry *dst = __sk_dst_get(sk);
122 int mss = tp->advmss;
125 unsigned int metric = dst_metric_advmss(dst);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism. */
138 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
140 struct tcp_sock *tp = tcp_sk(sk);
141 s32 delta = tcp_time_stamp - tp->lsndtime;
142 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
143 u32 cwnd = tp->snd_cwnd;
145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 restart_cwnd = min(restart_cwnd, cwnd);
150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock *tp,
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
164 if (sysctl_tcp_slow_start_after_idle &&
165 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
166 tcp_cwnd_restart(sk, __sk_dst_get(sk));
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 icsk->icsk_ack.pingpong = 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
180 tcp_dec_quickack_mode(sk, pkts);
181 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
184 /* Determine a window scaling and initial window to offer.
185 * Based on the assumption that the given amount of space
186 * will be offered. Store the results in the tp structure.
187 * NOTE: for smooth operation initial space offering should
188 * be a multiple of mss if possible. We assume here that mss >= 1.
189 * This MUST be enforced by all callers.
191 void tcp_select_initial_window(int __space, __u32 mss,
192 __u32 *rcv_wnd, __u32 *window_clamp,
193 int wscale_ok, __u8 *rcv_wscale,
196 unsigned int space = (__space < 0 ? 0 : __space);
198 /* If no clamp set the clamp to the max possible scaled window */
199 if (*window_clamp == 0)
200 (*window_clamp) = (65535 << 14);
201 space = min(*window_clamp, space);
203 /* Quantize space offering to a multiple of mss if possible. */
205 space = (space / mss) * mss;
207 /* NOTE: offering an initial window larger than 32767
208 * will break some buggy TCP stacks. If the admin tells us
209 * it is likely we could be speaking with such a buggy stack
210 * we will truncate our initial window offering to 32K-1
211 * unless the remote has sent us a window scaling option,
212 * which we interpret as a sign the remote TCP is not
213 * misinterpreting the window field as a signed quantity.
215 if (sysctl_tcp_workaround_signed_windows)
216 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
222 /* Set window scaling on max possible window
223 * See RFC1323 for an explanation of the limit to 14
225 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
226 space = min_t(u32, space, *window_clamp);
227 while (space > 65535 && (*rcv_wscale) < 14) {
233 /* Set initial window to a value enough for senders starting with
234 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
235 * a limit on the initial window when mss is larger than 1460.
237 if (mss > (1 << *rcv_wscale)) {
238 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
241 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
242 /* when initializing use the value from init_rcv_wnd
243 * rather than the default from above
246 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
248 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
251 /* Set the clamp no higher than max representable value */
252 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
254 EXPORT_SYMBOL(tcp_select_initial_window);
256 /* Chose a new window to advertise, update state in tcp_sock for the
257 * socket, and return result with RFC1323 scaling applied. The return
258 * value can be stuffed directly into th->window for an outgoing
261 static u16 tcp_select_window(struct sock *sk)
263 struct tcp_sock *tp = tcp_sk(sk);
264 u32 cur_win = tcp_receive_window(tp);
265 u32 new_win = __tcp_select_window(sk);
267 /* Never shrink the offered window */
268 if (new_win < cur_win) {
269 /* Danger Will Robinson!
270 * Don't update rcv_wup/rcv_wnd here or else
271 * we will not be able to advertise a zero
272 * window in time. --DaveM
274 * Relax Will Robinson.
276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
278 tp->rcv_wnd = new_win;
279 tp->rcv_wup = tp->rcv_nxt;
281 /* Make sure we do not exceed the maximum possible
284 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
285 new_win = min(new_win, MAX_TCP_WINDOW);
287 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
289 /* RFC1323 scaling applied */
290 new_win >>= tp->rx_opt.rcv_wscale;
292 /* If we advertise zero window, disable fast path. */
299 /* Packet ECN state for a SYN-ACK */
300 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
302 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
303 if (!(tp->ecn_flags & TCP_ECN_OK))
304 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
307 /* Packet ECN state for a SYN. */
308 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
310 struct tcp_sock *tp = tcp_sk(sk);
313 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
314 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
315 tp->ecn_flags = TCP_ECN_OK;
319 static __inline__ void
320 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
322 if (inet_rsk(req)->ecn_ok)
326 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
329 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
332 struct tcp_sock *tp = tcp_sk(sk);
334 if (tp->ecn_flags & TCP_ECN_OK) {
335 /* Not-retransmitted data segment: set ECT and inject CWR. */
336 if (skb->len != tcp_header_len &&
337 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
339 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
340 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
341 tcp_hdr(skb)->cwr = 1;
342 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
345 /* ACK or retransmitted segment: clear ECT|CE */
346 INET_ECN_dontxmit(sk);
348 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
349 tcp_hdr(skb)->ece = 1;
353 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
354 * auto increment end seqno.
356 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
358 skb->ip_summed = CHECKSUM_PARTIAL;
361 TCP_SKB_CB(skb)->tcp_flags = flags;
362 TCP_SKB_CB(skb)->sacked = 0;
364 skb_shinfo(skb)->gso_segs = 1;
365 skb_shinfo(skb)->gso_size = 0;
366 skb_shinfo(skb)->gso_type = 0;
368 TCP_SKB_CB(skb)->seq = seq;
369 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
371 TCP_SKB_CB(skb)->end_seq = seq;
374 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
376 return tp->snd_una != tp->snd_up;
379 #define OPTION_SACK_ADVERTISE (1 << 0)
380 #define OPTION_TS (1 << 1)
381 #define OPTION_MD5 (1 << 2)
382 #define OPTION_WSCALE (1 << 3)
383 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
385 struct tcp_out_options {
386 u16 options; /* bit field of OPTION_* */
387 u16 mss; /* 0 to disable */
388 u8 ws; /* window scale, 0 to disable */
389 u8 num_sack_blocks; /* number of SACK blocks to include */
390 u8 hash_size; /* bytes in hash_location */
391 __u8 *hash_location; /* temporary pointer, overloaded */
392 __u32 tsval, tsecr; /* need to include OPTION_TS */
393 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
396 /* Write previously computed TCP options to the packet.
398 * Beware: Something in the Internet is very sensitive to the ordering of
399 * TCP options, we learned this through the hard way, so be careful here.
400 * Luckily we can at least blame others for their non-compliance but from
401 * inter-operatibility perspective it seems that we're somewhat stuck with
402 * the ordering which we have been using if we want to keep working with
403 * those broken things (not that it currently hurts anybody as there isn't
404 * particular reason why the ordering would need to be changed).
406 * At least SACK_PERM as the first option is known to lead to a disaster
407 * (but it may well be that other scenarios fail similarly).
409 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
410 struct tcp_out_options *opts)
412 u16 options = opts->options; /* mungable copy */
414 if (unlikely(OPTION_MD5 & options)) {
415 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
416 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
417 /* overload cookie hash location */
418 opts->hash_location = (__u8 *)ptr;
422 if (unlikely(opts->mss)) {
423 *ptr++ = htonl((TCPOPT_MSS << 24) |
424 (TCPOLEN_MSS << 16) |
428 if (likely(OPTION_TS & options)) {
429 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
430 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
431 (TCPOLEN_SACK_PERM << 16) |
432 (TCPOPT_TIMESTAMP << 8) |
434 options &= ~OPTION_SACK_ADVERTISE;
436 *ptr++ = htonl((TCPOPT_NOP << 24) |
438 (TCPOPT_TIMESTAMP << 8) |
441 *ptr++ = htonl(opts->tsval);
442 *ptr++ = htonl(opts->tsecr);
445 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
446 *ptr++ = htonl((TCPOPT_NOP << 24) |
448 (TCPOPT_SACK_PERM << 8) |
452 if (unlikely(OPTION_WSCALE & options)) {
453 *ptr++ = htonl((TCPOPT_NOP << 24) |
454 (TCPOPT_WINDOW << 16) |
455 (TCPOLEN_WINDOW << 8) |
459 if (unlikely(opts->num_sack_blocks)) {
460 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
461 tp->duplicate_sack : tp->selective_acks;
464 *ptr++ = htonl((TCPOPT_NOP << 24) |
467 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
468 TCPOLEN_SACK_PERBLOCK)));
470 for (this_sack = 0; this_sack < opts->num_sack_blocks;
472 *ptr++ = htonl(sp[this_sack].start_seq);
473 *ptr++ = htonl(sp[this_sack].end_seq);
476 tp->rx_opt.dsack = 0;
479 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
480 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
482 *ptr++ = htonl((TCPOPT_EXP << 24) |
483 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
484 TCPOPT_FASTOPEN_MAGIC);
486 memcpy(ptr, foc->val, foc->len);
487 if ((foc->len & 3) == 2) {
488 u8 *align = ((u8 *)ptr) + foc->len;
489 align[0] = align[1] = TCPOPT_NOP;
491 ptr += (foc->len + 3) >> 2;
495 /* Compute TCP options for SYN packets. This is not the final
496 * network wire format yet.
498 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
499 struct tcp_out_options *opts,
500 struct tcp_md5sig_key **md5)
502 struct tcp_sock *tp = tcp_sk(sk);
503 unsigned int remaining = MAX_TCP_OPTION_SPACE;
504 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
506 #ifdef CONFIG_TCP_MD5SIG
507 *md5 = tp->af_specific->md5_lookup(sk, sk);
509 opts->options |= OPTION_MD5;
510 remaining -= TCPOLEN_MD5SIG_ALIGNED;
516 /* We always get an MSS option. The option bytes which will be seen in
517 * normal data packets should timestamps be used, must be in the MSS
518 * advertised. But we subtract them from tp->mss_cache so that
519 * calculations in tcp_sendmsg are simpler etc. So account for this
520 * fact here if necessary. If we don't do this correctly, as a
521 * receiver we won't recognize data packets as being full sized when we
522 * should, and thus we won't abide by the delayed ACK rules correctly.
523 * SACKs don't matter, we never delay an ACK when we have any of those
525 opts->mss = tcp_advertise_mss(sk);
526 remaining -= TCPOLEN_MSS_ALIGNED;
528 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
529 opts->options |= OPTION_TS;
530 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
531 opts->tsecr = tp->rx_opt.ts_recent;
532 remaining -= TCPOLEN_TSTAMP_ALIGNED;
534 if (likely(sysctl_tcp_window_scaling)) {
535 opts->ws = tp->rx_opt.rcv_wscale;
536 opts->options |= OPTION_WSCALE;
537 remaining -= TCPOLEN_WSCALE_ALIGNED;
539 if (likely(sysctl_tcp_sack)) {
540 opts->options |= OPTION_SACK_ADVERTISE;
541 if (unlikely(!(OPTION_TS & opts->options)))
542 remaining -= TCPOLEN_SACKPERM_ALIGNED;
545 if (fastopen && fastopen->cookie.len >= 0) {
546 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
547 need = (need + 3) & ~3U; /* Align to 32 bits */
548 if (remaining >= need) {
549 opts->options |= OPTION_FAST_OPEN_COOKIE;
550 opts->fastopen_cookie = &fastopen->cookie;
552 tp->syn_fastopen = 1;
556 return MAX_TCP_OPTION_SPACE - remaining;
559 /* Set up TCP options for SYN-ACKs. */
560 static unsigned int tcp_synack_options(struct sock *sk,
561 struct request_sock *req,
562 unsigned int mss, struct sk_buff *skb,
563 struct tcp_out_options *opts,
564 struct tcp_md5sig_key **md5,
565 struct tcp_fastopen_cookie *foc)
567 struct inet_request_sock *ireq = inet_rsk(req);
568 unsigned int remaining = MAX_TCP_OPTION_SPACE;
570 #ifdef CONFIG_TCP_MD5SIG
571 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
573 opts->options |= OPTION_MD5;
574 remaining -= TCPOLEN_MD5SIG_ALIGNED;
576 /* We can't fit any SACK blocks in a packet with MD5 + TS
577 * options. There was discussion about disabling SACK
578 * rather than TS in order to fit in better with old,
579 * buggy kernels, but that was deemed to be unnecessary.
581 ireq->tstamp_ok &= !ireq->sack_ok;
587 /* We always send an MSS option. */
589 remaining -= TCPOLEN_MSS_ALIGNED;
591 if (likely(ireq->wscale_ok)) {
592 opts->ws = ireq->rcv_wscale;
593 opts->options |= OPTION_WSCALE;
594 remaining -= TCPOLEN_WSCALE_ALIGNED;
596 if (likely(ireq->tstamp_ok)) {
597 opts->options |= OPTION_TS;
598 opts->tsval = TCP_SKB_CB(skb)->when;
599 opts->tsecr = req->ts_recent;
600 remaining -= TCPOLEN_TSTAMP_ALIGNED;
602 if (likely(ireq->sack_ok)) {
603 opts->options |= OPTION_SACK_ADVERTISE;
604 if (unlikely(!ireq->tstamp_ok))
605 remaining -= TCPOLEN_SACKPERM_ALIGNED;
608 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
609 need = (need + 3) & ~3U; /* Align to 32 bits */
610 if (remaining >= need) {
611 opts->options |= OPTION_FAST_OPEN_COOKIE;
612 opts->fastopen_cookie = foc;
617 return MAX_TCP_OPTION_SPACE - remaining;
620 /* Compute TCP options for ESTABLISHED sockets. This is not the
621 * final wire format yet.
623 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
624 struct tcp_out_options *opts,
625 struct tcp_md5sig_key **md5)
627 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
628 struct tcp_sock *tp = tcp_sk(sk);
629 unsigned int size = 0;
630 unsigned int eff_sacks;
632 #ifdef CONFIG_TCP_MD5SIG
633 *md5 = tp->af_specific->md5_lookup(sk, sk);
634 if (unlikely(*md5)) {
635 opts->options |= OPTION_MD5;
636 size += TCPOLEN_MD5SIG_ALIGNED;
642 if (likely(tp->rx_opt.tstamp_ok)) {
643 opts->options |= OPTION_TS;
644 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
645 opts->tsecr = tp->rx_opt.ts_recent;
646 size += TCPOLEN_TSTAMP_ALIGNED;
649 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
650 if (unlikely(eff_sacks)) {
651 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
652 opts->num_sack_blocks =
653 min_t(unsigned int, eff_sacks,
654 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
655 TCPOLEN_SACK_PERBLOCK);
656 size += TCPOLEN_SACK_BASE_ALIGNED +
657 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
664 /* TCP SMALL QUEUES (TSQ)
666 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
667 * to reduce RTT and bufferbloat.
668 * We do this using a special skb destructor (tcp_wfree).
670 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
671 * needs to be reallocated in a driver.
672 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
674 * Since transmit from skb destructor is forbidden, we use a tasklet
675 * to process all sockets that eventually need to send more skbs.
676 * We use one tasklet per cpu, with its own queue of sockets.
679 struct tasklet_struct tasklet;
680 struct list_head head; /* queue of tcp sockets */
682 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
684 static void tcp_tsq_handler(struct sock *sk)
686 if ((1 << sk->sk_state) &
687 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
688 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
689 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
692 * One tasklest per cpu tries to send more skbs.
693 * We run in tasklet context but need to disable irqs when
694 * transfering tsq->head because tcp_wfree() might
695 * interrupt us (non NAPI drivers)
697 static void tcp_tasklet_func(unsigned long data)
699 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
702 struct list_head *q, *n;
706 local_irq_save(flags);
707 list_splice_init(&tsq->head, &list);
708 local_irq_restore(flags);
710 list_for_each_safe(q, n, &list) {
711 tp = list_entry(q, struct tcp_sock, tsq_node);
712 list_del(&tp->tsq_node);
714 sk = (struct sock *)tp;
717 if (!sock_owned_by_user(sk)) {
720 /* defer the work to tcp_release_cb() */
721 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
725 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
730 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
731 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
732 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
733 (1UL << TCP_MTU_REDUCED_DEFERRED))
735 * tcp_release_cb - tcp release_sock() callback
738 * called from release_sock() to perform protocol dependent
739 * actions before socket release.
741 void tcp_release_cb(struct sock *sk)
743 struct tcp_sock *tp = tcp_sk(sk);
744 unsigned long flags, nflags;
746 /* perform an atomic operation only if at least one flag is set */
748 flags = tp->tsq_flags;
749 if (!(flags & TCP_DEFERRED_ALL))
751 nflags = flags & ~TCP_DEFERRED_ALL;
752 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
754 if (flags & (1UL << TCP_TSQ_DEFERRED))
757 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
758 tcp_write_timer_handler(sk);
761 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
762 tcp_delack_timer_handler(sk);
765 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
766 sk->sk_prot->mtu_reduced(sk);
770 EXPORT_SYMBOL(tcp_release_cb);
772 void __init tcp_tasklet_init(void)
776 for_each_possible_cpu(i) {
777 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
779 INIT_LIST_HEAD(&tsq->head);
780 tasklet_init(&tsq->tasklet,
787 * Write buffer destructor automatically called from kfree_skb.
788 * We cant xmit new skbs from this context, as we might already
791 void tcp_wfree(struct sk_buff *skb)
793 struct sock *sk = skb->sk;
794 struct tcp_sock *tp = tcp_sk(sk);
796 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
797 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
799 struct tsq_tasklet *tsq;
801 /* Keep a ref on socket.
802 * This last ref will be released in tcp_tasklet_func()
804 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
806 /* queue this socket to tasklet queue */
807 local_irq_save(flags);
808 tsq = &__get_cpu_var(tsq_tasklet);
809 list_add(&tp->tsq_node, &tsq->head);
810 tasklet_schedule(&tsq->tasklet);
811 local_irq_restore(flags);
817 /* This routine actually transmits TCP packets queued in by
818 * tcp_do_sendmsg(). This is used by both the initial
819 * transmission and possible later retransmissions.
820 * All SKB's seen here are completely headerless. It is our
821 * job to build the TCP header, and pass the packet down to
822 * IP so it can do the same plus pass the packet off to the
825 * We are working here with either a clone of the original
826 * SKB, or a fresh unique copy made by the retransmit engine.
828 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
831 const struct inet_connection_sock *icsk = inet_csk(sk);
832 struct inet_sock *inet;
834 struct tcp_skb_cb *tcb;
835 struct tcp_out_options opts;
836 unsigned int tcp_options_size, tcp_header_size;
837 struct tcp_md5sig_key *md5;
841 BUG_ON(!skb || !tcp_skb_pcount(skb));
843 /* If congestion control is doing timestamping, we must
844 * take such a timestamp before we potentially clone/copy.
846 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
847 __net_timestamp(skb);
849 if (likely(clone_it)) {
850 const struct sk_buff *fclone = skb + 1;
852 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
853 fclone->fclone == SKB_FCLONE_CLONE))
854 NET_INC_STATS_BH(sock_net(sk),
855 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
857 if (unlikely(skb_cloned(skb)))
858 skb = pskb_copy(skb, gfp_mask);
860 skb = skb_clone(skb, gfp_mask);
867 tcb = TCP_SKB_CB(skb);
868 memset(&opts, 0, sizeof(opts));
870 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
871 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
873 tcp_options_size = tcp_established_options(sk, skb, &opts,
875 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
877 if (tcp_packets_in_flight(tp) == 0)
878 tcp_ca_event(sk, CA_EVENT_TX_START);
880 /* if no packet is in qdisc/device queue, then allow XPS to select
883 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
885 skb_push(skb, tcp_header_size);
886 skb_reset_transport_header(skb);
890 skb->destructor = tcp_wfree;
891 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
893 /* Build TCP header and checksum it. */
895 th->source = inet->inet_sport;
896 th->dest = inet->inet_dport;
897 th->seq = htonl(tcb->seq);
898 th->ack_seq = htonl(tp->rcv_nxt);
899 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
902 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
903 /* RFC1323: The window in SYN & SYN/ACK segments
906 th->window = htons(min(tp->rcv_wnd, 65535U));
908 th->window = htons(tcp_select_window(sk));
913 /* The urg_mode check is necessary during a below snd_una win probe */
914 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
915 if (before(tp->snd_up, tcb->seq + 0x10000)) {
916 th->urg_ptr = htons(tp->snd_up - tcb->seq);
918 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
919 th->urg_ptr = htons(0xFFFF);
924 tcp_options_write((__be32 *)(th + 1), tp, &opts);
925 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
926 TCP_ECN_send(sk, skb, tcp_header_size);
928 #ifdef CONFIG_TCP_MD5SIG
929 /* Calculate the MD5 hash, as we have all we need now */
931 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
932 tp->af_specific->calc_md5_hash(opts.hash_location,
937 icsk->icsk_af_ops->send_check(sk, skb);
939 if (likely(tcb->tcp_flags & TCPHDR_ACK))
940 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
942 if (skb->len != tcp_header_size)
943 tcp_event_data_sent(tp, sk);
945 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
946 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
947 tcp_skb_pcount(skb));
949 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
950 if (likely(err <= 0))
953 tcp_enter_cwr(sk, 1);
955 return net_xmit_eval(err);
958 /* This routine just queues the buffer for sending.
960 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
961 * otherwise socket can stall.
963 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
965 struct tcp_sock *tp = tcp_sk(sk);
967 /* Advance write_seq and place onto the write_queue. */
968 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
969 skb_header_release(skb);
970 tcp_add_write_queue_tail(sk, skb);
971 sk->sk_wmem_queued += skb->truesize;
972 sk_mem_charge(sk, skb->truesize);
975 /* Initialize TSO segments for a packet. */
976 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
977 unsigned int mss_now)
979 if (skb->len <= mss_now || !sk_can_gso(sk) ||
980 skb->ip_summed == CHECKSUM_NONE) {
981 /* Avoid the costly divide in the normal
984 skb_shinfo(skb)->gso_segs = 1;
985 skb_shinfo(skb)->gso_size = 0;
986 skb_shinfo(skb)->gso_type = 0;
988 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
989 skb_shinfo(skb)->gso_size = mss_now;
990 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
994 /* When a modification to fackets out becomes necessary, we need to check
995 * skb is counted to fackets_out or not.
997 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1000 struct tcp_sock *tp = tcp_sk(sk);
1002 if (!tp->sacked_out || tcp_is_reno(tp))
1005 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1006 tp->fackets_out -= decr;
1009 /* Pcount in the middle of the write queue got changed, we need to do various
1010 * tweaks to fix counters
1012 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1014 struct tcp_sock *tp = tcp_sk(sk);
1016 tp->packets_out -= decr;
1018 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1019 tp->sacked_out -= decr;
1020 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1021 tp->retrans_out -= decr;
1022 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1023 tp->lost_out -= decr;
1025 /* Reno case is special. Sigh... */
1026 if (tcp_is_reno(tp) && decr > 0)
1027 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1029 tcp_adjust_fackets_out(sk, skb, decr);
1031 if (tp->lost_skb_hint &&
1032 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1033 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1034 tp->lost_cnt_hint -= decr;
1036 tcp_verify_left_out(tp);
1039 /* Function to create two new TCP segments. Shrinks the given segment
1040 * to the specified size and appends a new segment with the rest of the
1041 * packet to the list. This won't be called frequently, I hope.
1042 * Remember, these are still headerless SKBs at this point.
1044 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1045 unsigned int mss_now)
1047 struct tcp_sock *tp = tcp_sk(sk);
1048 struct sk_buff *buff;
1049 int nsize, old_factor;
1053 if (WARN_ON(len > skb->len))
1056 nsize = skb_headlen(skb) - len;
1060 if (skb_cloned(skb) &&
1061 skb_is_nonlinear(skb) &&
1062 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1065 /* Get a new skb... force flag on. */
1066 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1068 return -ENOMEM; /* We'll just try again later. */
1070 sk->sk_wmem_queued += buff->truesize;
1071 sk_mem_charge(sk, buff->truesize);
1072 nlen = skb->len - len - nsize;
1073 buff->truesize += nlen;
1074 skb->truesize -= nlen;
1076 /* Correct the sequence numbers. */
1077 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1078 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1079 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1081 /* PSH and FIN should only be set in the second packet. */
1082 flags = TCP_SKB_CB(skb)->tcp_flags;
1083 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1084 TCP_SKB_CB(buff)->tcp_flags = flags;
1085 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1087 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1088 /* Copy and checksum data tail into the new buffer. */
1089 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1090 skb_put(buff, nsize),
1095 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1097 skb->ip_summed = CHECKSUM_PARTIAL;
1098 skb_split(skb, buff, len);
1101 buff->ip_summed = skb->ip_summed;
1103 /* Looks stupid, but our code really uses when of
1104 * skbs, which it never sent before. --ANK
1106 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1107 buff->tstamp = skb->tstamp;
1109 old_factor = tcp_skb_pcount(skb);
1111 /* Fix up tso_factor for both original and new SKB. */
1112 tcp_set_skb_tso_segs(sk, skb, mss_now);
1113 tcp_set_skb_tso_segs(sk, buff, mss_now);
1115 /* If this packet has been sent out already, we must
1116 * adjust the various packet counters.
1118 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1119 int diff = old_factor - tcp_skb_pcount(skb) -
1120 tcp_skb_pcount(buff);
1123 tcp_adjust_pcount(sk, skb, diff);
1126 /* Link BUFF into the send queue. */
1127 skb_header_release(buff);
1128 tcp_insert_write_queue_after(skb, buff, sk);
1133 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1134 * eventually). The difference is that pulled data not copied, but
1135 * immediately discarded.
1137 static void __pskb_trim_head(struct sk_buff *skb, int len)
1141 eat = min_t(int, len, skb_headlen(skb));
1143 __skb_pull(skb, eat);
1150 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1151 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1154 skb_frag_unref(skb, i);
1157 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1159 skb_shinfo(skb)->frags[k].page_offset += eat;
1160 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1166 skb_shinfo(skb)->nr_frags = k;
1168 skb_reset_tail_pointer(skb);
1169 skb->data_len -= len;
1170 skb->len = skb->data_len;
1173 /* Remove acked data from a packet in the transmit queue. */
1174 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1176 if (skb_unclone(skb, GFP_ATOMIC))
1179 __pskb_trim_head(skb, len);
1181 TCP_SKB_CB(skb)->seq += len;
1182 skb->ip_summed = CHECKSUM_PARTIAL;
1184 skb->truesize -= len;
1185 sk->sk_wmem_queued -= len;
1186 sk_mem_uncharge(sk, len);
1187 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1189 /* Any change of skb->len requires recalculation of tso factor. */
1190 if (tcp_skb_pcount(skb) > 1)
1191 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1196 /* Calculate MSS not accounting any TCP options. */
1197 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1199 const struct tcp_sock *tp = tcp_sk(sk);
1200 const struct inet_connection_sock *icsk = inet_csk(sk);
1203 /* Calculate base mss without TCP options:
1204 It is MMS_S - sizeof(tcphdr) of rfc1122
1206 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1208 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1209 if (icsk->icsk_af_ops->net_frag_header_len) {
1210 const struct dst_entry *dst = __sk_dst_get(sk);
1212 if (dst && dst_allfrag(dst))
1213 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1216 /* Clamp it (mss_clamp does not include tcp options) */
1217 if (mss_now > tp->rx_opt.mss_clamp)
1218 mss_now = tp->rx_opt.mss_clamp;
1220 /* Now subtract optional transport overhead */
1221 mss_now -= icsk->icsk_ext_hdr_len;
1223 /* Then reserve room for full set of TCP options and 8 bytes of data */
1229 /* Calculate MSS. Not accounting for SACKs here. */
1230 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1232 /* Subtract TCP options size, not including SACKs */
1233 return __tcp_mtu_to_mss(sk, pmtu) -
1234 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1237 /* Inverse of above */
1238 int tcp_mss_to_mtu(struct sock *sk, int mss)
1240 const struct tcp_sock *tp = tcp_sk(sk);
1241 const struct inet_connection_sock *icsk = inet_csk(sk);
1245 tp->tcp_header_len +
1246 icsk->icsk_ext_hdr_len +
1247 icsk->icsk_af_ops->net_header_len;
1249 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1250 if (icsk->icsk_af_ops->net_frag_header_len) {
1251 const struct dst_entry *dst = __sk_dst_get(sk);
1253 if (dst && dst_allfrag(dst))
1254 mtu += icsk->icsk_af_ops->net_frag_header_len;
1259 /* MTU probing init per socket */
1260 void tcp_mtup_init(struct sock *sk)
1262 struct tcp_sock *tp = tcp_sk(sk);
1263 struct inet_connection_sock *icsk = inet_csk(sk);
1265 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1266 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1267 icsk->icsk_af_ops->net_header_len;
1268 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1269 icsk->icsk_mtup.probe_size = 0;
1271 EXPORT_SYMBOL(tcp_mtup_init);
1273 /* This function synchronize snd mss to current pmtu/exthdr set.
1275 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1276 for TCP options, but includes only bare TCP header.
1278 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1279 It is minimum of user_mss and mss received with SYN.
1280 It also does not include TCP options.
1282 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1284 tp->mss_cache is current effective sending mss, including
1285 all tcp options except for SACKs. It is evaluated,
1286 taking into account current pmtu, but never exceeds
1287 tp->rx_opt.mss_clamp.
1289 NOTE1. rfc1122 clearly states that advertised MSS
1290 DOES NOT include either tcp or ip options.
1292 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1293 are READ ONLY outside this function. --ANK (980731)
1295 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1297 struct tcp_sock *tp = tcp_sk(sk);
1298 struct inet_connection_sock *icsk = inet_csk(sk);
1301 if (icsk->icsk_mtup.search_high > pmtu)
1302 icsk->icsk_mtup.search_high = pmtu;
1304 mss_now = tcp_mtu_to_mss(sk, pmtu);
1305 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1307 /* And store cached results */
1308 icsk->icsk_pmtu_cookie = pmtu;
1309 if (icsk->icsk_mtup.enabled)
1310 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1311 tp->mss_cache = mss_now;
1315 EXPORT_SYMBOL(tcp_sync_mss);
1317 /* Compute the current effective MSS, taking SACKs and IP options,
1318 * and even PMTU discovery events into account.
1320 unsigned int tcp_current_mss(struct sock *sk)
1322 const struct tcp_sock *tp = tcp_sk(sk);
1323 const struct dst_entry *dst = __sk_dst_get(sk);
1325 unsigned int header_len;
1326 struct tcp_out_options opts;
1327 struct tcp_md5sig_key *md5;
1329 mss_now = tp->mss_cache;
1332 u32 mtu = dst_mtu(dst);
1333 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1334 mss_now = tcp_sync_mss(sk, mtu);
1337 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1338 sizeof(struct tcphdr);
1339 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1340 * some common options. If this is an odd packet (because we have SACK
1341 * blocks etc) then our calculated header_len will be different, and
1342 * we have to adjust mss_now correspondingly */
1343 if (header_len != tp->tcp_header_len) {
1344 int delta = (int) header_len - tp->tcp_header_len;
1351 /* Congestion window validation. (RFC2861) */
1352 static void tcp_cwnd_validate(struct sock *sk)
1354 struct tcp_sock *tp = tcp_sk(sk);
1356 if (tp->packets_out >= tp->snd_cwnd) {
1357 /* Network is feed fully. */
1358 tp->snd_cwnd_used = 0;
1359 tp->snd_cwnd_stamp = tcp_time_stamp;
1361 /* Network starves. */
1362 if (tp->packets_out > tp->snd_cwnd_used)
1363 tp->snd_cwnd_used = tp->packets_out;
1365 if (sysctl_tcp_slow_start_after_idle &&
1366 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1367 tcp_cwnd_application_limited(sk);
1371 /* Returns the portion of skb which can be sent right away without
1372 * introducing MSS oddities to segment boundaries. In rare cases where
1373 * mss_now != mss_cache, we will request caller to create a small skb
1374 * per input skb which could be mostly avoided here (if desired).
1376 * We explicitly want to create a request for splitting write queue tail
1377 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1378 * thus all the complexity (cwnd_len is always MSS multiple which we
1379 * return whenever allowed by the other factors). Basically we need the
1380 * modulo only when the receiver window alone is the limiting factor or
1381 * when we would be allowed to send the split-due-to-Nagle skb fully.
1383 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1384 unsigned int mss_now, unsigned int max_segs)
1386 const struct tcp_sock *tp = tcp_sk(sk);
1387 u32 needed, window, max_len;
1389 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1390 max_len = mss_now * max_segs;
1392 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1395 needed = min(skb->len, window);
1397 if (max_len <= needed)
1400 return needed - needed % mss_now;
1403 /* Can at least one segment of SKB be sent right now, according to the
1404 * congestion window rules? If so, return how many segments are allowed.
1406 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1407 const struct sk_buff *skb)
1409 u32 in_flight, cwnd;
1411 /* Don't be strict about the congestion window for the final FIN. */
1412 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1413 tcp_skb_pcount(skb) == 1)
1416 in_flight = tcp_packets_in_flight(tp);
1417 cwnd = tp->snd_cwnd;
1418 if (in_flight < cwnd)
1419 return (cwnd - in_flight);
1424 /* Initialize TSO state of a skb.
1425 * This must be invoked the first time we consider transmitting
1426 * SKB onto the wire.
1428 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1429 unsigned int mss_now)
1431 int tso_segs = tcp_skb_pcount(skb);
1433 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1434 tcp_set_skb_tso_segs(sk, skb, mss_now);
1435 tso_segs = tcp_skb_pcount(skb);
1440 /* Minshall's variant of the Nagle send check. */
1441 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1443 return after(tp->snd_sml, tp->snd_una) &&
1444 !after(tp->snd_sml, tp->snd_nxt);
1447 /* Return false, if packet can be sent now without violation Nagle's rules:
1448 * 1. It is full sized.
1449 * 2. Or it contains FIN. (already checked by caller)
1450 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1451 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1452 * With Minshall's modification: all sent small packets are ACKed.
1454 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1455 const struct sk_buff *skb,
1456 unsigned int mss_now, int nonagle)
1458 return skb->len < mss_now &&
1459 ((nonagle & TCP_NAGLE_CORK) ||
1460 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1463 /* Return true if the Nagle test allows this packet to be
1466 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1467 unsigned int cur_mss, int nonagle)
1469 /* Nagle rule does not apply to frames, which sit in the middle of the
1470 * write_queue (they have no chances to get new data).
1472 * This is implemented in the callers, where they modify the 'nonagle'
1473 * argument based upon the location of SKB in the send queue.
1475 if (nonagle & TCP_NAGLE_PUSH)
1478 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1479 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1482 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1488 /* Does at least the first segment of SKB fit into the send window? */
1489 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1490 const struct sk_buff *skb,
1491 unsigned int cur_mss)
1493 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1495 if (skb->len > cur_mss)
1496 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1498 return !after(end_seq, tcp_wnd_end(tp));
1501 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1502 * should be put on the wire right now. If so, it returns the number of
1503 * packets allowed by the congestion window.
1505 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1506 unsigned int cur_mss, int nonagle)
1508 const struct tcp_sock *tp = tcp_sk(sk);
1509 unsigned int cwnd_quota;
1511 tcp_init_tso_segs(sk, skb, cur_mss);
1513 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1516 cwnd_quota = tcp_cwnd_test(tp, skb);
1517 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1523 /* Test if sending is allowed right now. */
1524 bool tcp_may_send_now(struct sock *sk)
1526 const struct tcp_sock *tp = tcp_sk(sk);
1527 struct sk_buff *skb = tcp_send_head(sk);
1530 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1531 (tcp_skb_is_last(sk, skb) ?
1532 tp->nonagle : TCP_NAGLE_PUSH));
1535 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1536 * which is put after SKB on the list. It is very much like
1537 * tcp_fragment() except that it may make several kinds of assumptions
1538 * in order to speed up the splitting operation. In particular, we
1539 * know that all the data is in scatter-gather pages, and that the
1540 * packet has never been sent out before (and thus is not cloned).
1542 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1543 unsigned int mss_now, gfp_t gfp)
1545 struct sk_buff *buff;
1546 int nlen = skb->len - len;
1549 /* All of a TSO frame must be composed of paged data. */
1550 if (skb->len != skb->data_len)
1551 return tcp_fragment(sk, skb, len, mss_now);
1553 buff = sk_stream_alloc_skb(sk, 0, gfp);
1554 if (unlikely(buff == NULL))
1557 sk->sk_wmem_queued += buff->truesize;
1558 sk_mem_charge(sk, buff->truesize);
1559 buff->truesize += nlen;
1560 skb->truesize -= nlen;
1562 /* Correct the sequence numbers. */
1563 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1564 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1565 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1567 /* PSH and FIN should only be set in the second packet. */
1568 flags = TCP_SKB_CB(skb)->tcp_flags;
1569 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1570 TCP_SKB_CB(buff)->tcp_flags = flags;
1572 /* This packet was never sent out yet, so no SACK bits. */
1573 TCP_SKB_CB(buff)->sacked = 0;
1575 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1576 skb_split(skb, buff, len);
1578 /* Fix up tso_factor for both original and new SKB. */
1579 tcp_set_skb_tso_segs(sk, skb, mss_now);
1580 tcp_set_skb_tso_segs(sk, buff, mss_now);
1582 /* Link BUFF into the send queue. */
1583 skb_header_release(buff);
1584 tcp_insert_write_queue_after(skb, buff, sk);
1589 /* Try to defer sending, if possible, in order to minimize the amount
1590 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1592 * This algorithm is from John Heffner.
1594 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1596 struct tcp_sock *tp = tcp_sk(sk);
1597 const struct inet_connection_sock *icsk = inet_csk(sk);
1598 u32 send_win, cong_win, limit, in_flight;
1601 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1604 if (icsk->icsk_ca_state != TCP_CA_Open)
1607 /* Defer for less than two clock ticks. */
1608 if (tp->tso_deferred &&
1609 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1612 in_flight = tcp_packets_in_flight(tp);
1614 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1616 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1618 /* From in_flight test above, we know that cwnd > in_flight. */
1619 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1621 limit = min(send_win, cong_win);
1623 /* If a full-sized TSO skb can be sent, do it. */
1624 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1625 tp->xmit_size_goal_segs * tp->mss_cache))
1628 /* Middle in queue won't get any more data, full sendable already? */
1629 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1632 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1634 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1636 /* If at least some fraction of a window is available,
1639 chunk /= win_divisor;
1643 /* Different approach, try not to defer past a single
1644 * ACK. Receiver should ACK every other full sized
1645 * frame, so if we have space for more than 3 frames
1648 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1652 /* Ok, it looks like it is advisable to defer.
1653 * Do not rearm the timer if already set to not break TCP ACK clocking.
1655 if (!tp->tso_deferred)
1656 tp->tso_deferred = 1 | (jiffies << 1);
1661 tp->tso_deferred = 0;
1665 /* Create a new MTU probe if we are ready.
1666 * MTU probe is regularly attempting to increase the path MTU by
1667 * deliberately sending larger packets. This discovers routing
1668 * changes resulting in larger path MTUs.
1670 * Returns 0 if we should wait to probe (no cwnd available),
1671 * 1 if a probe was sent,
1674 static int tcp_mtu_probe(struct sock *sk)
1676 struct tcp_sock *tp = tcp_sk(sk);
1677 struct inet_connection_sock *icsk = inet_csk(sk);
1678 struct sk_buff *skb, *nskb, *next;
1685 /* Not currently probing/verifying,
1687 * have enough cwnd, and
1688 * not SACKing (the variable headers throw things off) */
1689 if (!icsk->icsk_mtup.enabled ||
1690 icsk->icsk_mtup.probe_size ||
1691 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1692 tp->snd_cwnd < 11 ||
1693 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1696 /* Very simple search strategy: just double the MSS. */
1697 mss_now = tcp_current_mss(sk);
1698 probe_size = 2 * tp->mss_cache;
1699 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1700 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1701 /* TODO: set timer for probe_converge_event */
1705 /* Have enough data in the send queue to probe? */
1706 if (tp->write_seq - tp->snd_nxt < size_needed)
1709 if (tp->snd_wnd < size_needed)
1711 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1714 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1715 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1716 if (!tcp_packets_in_flight(tp))
1722 /* We're allowed to probe. Build it now. */
1723 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1725 sk->sk_wmem_queued += nskb->truesize;
1726 sk_mem_charge(sk, nskb->truesize);
1728 skb = tcp_send_head(sk);
1730 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1731 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1732 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1733 TCP_SKB_CB(nskb)->sacked = 0;
1735 nskb->ip_summed = skb->ip_summed;
1737 tcp_insert_write_queue_before(nskb, skb, sk);
1740 tcp_for_write_queue_from_safe(skb, next, sk) {
1741 copy = min_t(int, skb->len, probe_size - len);
1742 if (nskb->ip_summed)
1743 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1745 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1746 skb_put(nskb, copy),
1749 if (skb->len <= copy) {
1750 /* We've eaten all the data from this skb.
1752 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1753 tcp_unlink_write_queue(skb, sk);
1754 sk_wmem_free_skb(sk, skb);
1756 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1757 ~(TCPHDR_FIN|TCPHDR_PSH);
1758 if (!skb_shinfo(skb)->nr_frags) {
1759 skb_pull(skb, copy);
1760 if (skb->ip_summed != CHECKSUM_PARTIAL)
1761 skb->csum = csum_partial(skb->data,
1764 __pskb_trim_head(skb, copy);
1765 tcp_set_skb_tso_segs(sk, skb, mss_now);
1767 TCP_SKB_CB(skb)->seq += copy;
1772 if (len >= probe_size)
1775 tcp_init_tso_segs(sk, nskb, nskb->len);
1777 /* We're ready to send. If this fails, the probe will
1778 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1779 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1780 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1781 /* Decrement cwnd here because we are sending
1782 * effectively two packets. */
1784 tcp_event_new_data_sent(sk, nskb);
1786 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1787 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1788 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1796 /* This routine writes packets to the network. It advances the
1797 * send_head. This happens as incoming acks open up the remote
1800 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1801 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1802 * account rare use of URG, this is not a big flaw.
1804 * Send at most one packet when push_one > 0. Temporarily ignore
1805 * cwnd limit to force at most one packet out when push_one == 2.
1807 * Returns true, if no segments are in flight and we have queued segments,
1808 * but cannot send anything now because of SWS or another problem.
1810 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1811 int push_one, gfp_t gfp)
1813 struct tcp_sock *tp = tcp_sk(sk);
1814 struct sk_buff *skb;
1815 unsigned int tso_segs, sent_pkts;
1822 /* Do MTU probing. */
1823 result = tcp_mtu_probe(sk);
1826 } else if (result > 0) {
1831 while ((skb = tcp_send_head(sk))) {
1834 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1837 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1838 goto repair; /* Skip network transmission */
1840 cwnd_quota = tcp_cwnd_test(tp, skb);
1843 /* Force out a loss probe pkt. */
1849 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1852 if (tso_segs == 1) {
1853 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1854 (tcp_skb_is_last(sk, skb) ?
1855 nonagle : TCP_NAGLE_PUSH))))
1858 if (!push_one && tcp_tso_should_defer(sk, skb))
1862 /* TCP Small Queues :
1863 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1865 * - better RTT estimation and ACK scheduling
1869 limit = max(skb->truesize, sk->sk_pacing_rate >> 10);
1871 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1872 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1877 if (tso_segs > 1 && !tcp_urg_mode(tp))
1878 limit = tcp_mss_split_point(sk, skb, mss_now,
1881 sk->sk_gso_max_segs));
1883 if (skb->len > limit &&
1884 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1887 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1889 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1893 /* Advance the send_head. This one is sent out.
1894 * This call will increment packets_out.
1896 tcp_event_new_data_sent(sk, skb);
1898 tcp_minshall_update(tp, mss_now, skb);
1899 sent_pkts += tcp_skb_pcount(skb);
1905 if (likely(sent_pkts)) {
1906 if (tcp_in_cwnd_reduction(sk))
1907 tp->prr_out += sent_pkts;
1909 /* Send one loss probe per tail loss episode. */
1911 tcp_schedule_loss_probe(sk);
1912 tcp_cwnd_validate(sk);
1915 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1918 bool tcp_schedule_loss_probe(struct sock *sk)
1920 struct inet_connection_sock *icsk = inet_csk(sk);
1921 struct tcp_sock *tp = tcp_sk(sk);
1922 u32 timeout, tlp_time_stamp, rto_time_stamp;
1923 u32 rtt = tp->srtt >> 3;
1925 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1927 /* No consecutive loss probes. */
1928 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1932 /* Don't do any loss probe on a Fast Open connection before 3WHS
1935 if (sk->sk_state == TCP_SYN_RECV)
1938 /* TLP is only scheduled when next timer event is RTO. */
1939 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1942 /* Schedule a loss probe in 2*RTT for SACK capable connections
1943 * in Open state, that are either limited by cwnd or application.
1945 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1946 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1949 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1953 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1954 * for delayed ack when there's one outstanding packet.
1957 if (tp->packets_out == 1)
1958 timeout = max_t(u32, timeout,
1959 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1960 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1962 /* If RTO is shorter, just schedule TLP in its place. */
1963 tlp_time_stamp = tcp_time_stamp + timeout;
1964 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1965 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1966 s32 delta = rto_time_stamp - tcp_time_stamp;
1971 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1976 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1977 * retransmit the last segment.
1979 void tcp_send_loss_probe(struct sock *sk)
1981 struct tcp_sock *tp = tcp_sk(sk);
1982 struct sk_buff *skb;
1984 int mss = tcp_current_mss(sk);
1987 if (tcp_send_head(sk) != NULL) {
1988 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
1992 /* At most one outstanding TLP retransmission. */
1993 if (tp->tlp_high_seq)
1996 /* Retransmit last segment. */
1997 skb = tcp_write_queue_tail(sk);
2001 pcount = tcp_skb_pcount(skb);
2002 if (WARN_ON(!pcount))
2005 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2006 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2008 skb = tcp_write_queue_tail(sk);
2011 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2014 /* Probe with zero data doesn't trigger fast recovery. */
2016 err = __tcp_retransmit_skb(sk, skb);
2018 /* Record snd_nxt for loss detection. */
2020 tp->tlp_high_seq = tp->snd_nxt;
2023 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2024 inet_csk(sk)->icsk_rto,
2028 NET_INC_STATS_BH(sock_net(sk),
2029 LINUX_MIB_TCPLOSSPROBES);
2033 /* Push out any pending frames which were held back due to
2034 * TCP_CORK or attempt at coalescing tiny packets.
2035 * The socket must be locked by the caller.
2037 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2040 /* If we are closed, the bytes will have to remain here.
2041 * In time closedown will finish, we empty the write queue and
2042 * all will be happy.
2044 if (unlikely(sk->sk_state == TCP_CLOSE))
2047 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2048 sk_gfp_atomic(sk, GFP_ATOMIC)))
2049 tcp_check_probe_timer(sk);
2052 /* Send _single_ skb sitting at the send head. This function requires
2053 * true push pending frames to setup probe timer etc.
2055 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2057 struct sk_buff *skb = tcp_send_head(sk);
2059 BUG_ON(!skb || skb->len < mss_now);
2061 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2064 /* This function returns the amount that we can raise the
2065 * usable window based on the following constraints
2067 * 1. The window can never be shrunk once it is offered (RFC 793)
2068 * 2. We limit memory per socket
2071 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2072 * RECV.NEXT + RCV.WIN fixed until:
2073 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2075 * i.e. don't raise the right edge of the window until you can raise
2076 * it at least MSS bytes.
2078 * Unfortunately, the recommended algorithm breaks header prediction,
2079 * since header prediction assumes th->window stays fixed.
2081 * Strictly speaking, keeping th->window fixed violates the receiver
2082 * side SWS prevention criteria. The problem is that under this rule
2083 * a stream of single byte packets will cause the right side of the
2084 * window to always advance by a single byte.
2086 * Of course, if the sender implements sender side SWS prevention
2087 * then this will not be a problem.
2089 * BSD seems to make the following compromise:
2091 * If the free space is less than the 1/4 of the maximum
2092 * space available and the free space is less than 1/2 mss,
2093 * then set the window to 0.
2094 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2095 * Otherwise, just prevent the window from shrinking
2096 * and from being larger than the largest representable value.
2098 * This prevents incremental opening of the window in the regime
2099 * where TCP is limited by the speed of the reader side taking
2100 * data out of the TCP receive queue. It does nothing about
2101 * those cases where the window is constrained on the sender side
2102 * because the pipeline is full.
2104 * BSD also seems to "accidentally" limit itself to windows that are a
2105 * multiple of MSS, at least until the free space gets quite small.
2106 * This would appear to be a side effect of the mbuf implementation.
2107 * Combining these two algorithms results in the observed behavior
2108 * of having a fixed window size at almost all times.
2110 * Below we obtain similar behavior by forcing the offered window to
2111 * a multiple of the mss when it is feasible to do so.
2113 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2114 * Regular options like TIMESTAMP are taken into account.
2116 u32 __tcp_select_window(struct sock *sk)
2118 struct inet_connection_sock *icsk = inet_csk(sk);
2119 struct tcp_sock *tp = tcp_sk(sk);
2120 /* MSS for the peer's data. Previous versions used mss_clamp
2121 * here. I don't know if the value based on our guesses
2122 * of peer's MSS is better for the performance. It's more correct
2123 * but may be worse for the performance because of rcv_mss
2124 * fluctuations. --SAW 1998/11/1
2126 int mss = icsk->icsk_ack.rcv_mss;
2127 int free_space = tcp_space(sk);
2128 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2131 if (mss > full_space)
2134 if (free_space < (full_space >> 1)) {
2135 icsk->icsk_ack.quick = 0;
2137 if (sk_under_memory_pressure(sk))
2138 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2141 if (free_space < mss)
2145 if (free_space > tp->rcv_ssthresh)
2146 free_space = tp->rcv_ssthresh;
2148 /* Don't do rounding if we are using window scaling, since the
2149 * scaled window will not line up with the MSS boundary anyway.
2151 window = tp->rcv_wnd;
2152 if (tp->rx_opt.rcv_wscale) {
2153 window = free_space;
2155 /* Advertise enough space so that it won't get scaled away.
2156 * Import case: prevent zero window announcement if
2157 * 1<<rcv_wscale > mss.
2159 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2160 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2161 << tp->rx_opt.rcv_wscale);
2163 /* Get the largest window that is a nice multiple of mss.
2164 * Window clamp already applied above.
2165 * If our current window offering is within 1 mss of the
2166 * free space we just keep it. This prevents the divide
2167 * and multiply from happening most of the time.
2168 * We also don't do any window rounding when the free space
2171 if (window <= free_space - mss || window > free_space)
2172 window = (free_space / mss) * mss;
2173 else if (mss == full_space &&
2174 free_space > window + (full_space >> 1))
2175 window = free_space;
2181 /* Collapses two adjacent SKB's during retransmission. */
2182 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2184 struct tcp_sock *tp = tcp_sk(sk);
2185 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2186 int skb_size, next_skb_size;
2188 skb_size = skb->len;
2189 next_skb_size = next_skb->len;
2191 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2193 tcp_highest_sack_combine(sk, next_skb, skb);
2195 tcp_unlink_write_queue(next_skb, sk);
2197 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2200 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2201 skb->ip_summed = CHECKSUM_PARTIAL;
2203 if (skb->ip_summed != CHECKSUM_PARTIAL)
2204 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2206 /* Update sequence range on original skb. */
2207 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2209 /* Merge over control information. This moves PSH/FIN etc. over */
2210 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2212 /* All done, get rid of second SKB and account for it so
2213 * packet counting does not break.
2215 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2217 /* changed transmit queue under us so clear hints */
2218 tcp_clear_retrans_hints_partial(tp);
2219 if (next_skb == tp->retransmit_skb_hint)
2220 tp->retransmit_skb_hint = skb;
2222 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2224 sk_wmem_free_skb(sk, next_skb);
2227 /* Check if coalescing SKBs is legal. */
2228 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2230 if (tcp_skb_pcount(skb) > 1)
2232 /* TODO: SACK collapsing could be used to remove this condition */
2233 if (skb_shinfo(skb)->nr_frags != 0)
2235 if (skb_cloned(skb))
2237 if (skb == tcp_send_head(sk))
2239 /* Some heurestics for collapsing over SACK'd could be invented */
2240 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2246 /* Collapse packets in the retransmit queue to make to create
2247 * less packets on the wire. This is only done on retransmission.
2249 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2252 struct tcp_sock *tp = tcp_sk(sk);
2253 struct sk_buff *skb = to, *tmp;
2256 if (!sysctl_tcp_retrans_collapse)
2258 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2261 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2262 if (!tcp_can_collapse(sk, skb))
2274 /* Punt if not enough space exists in the first SKB for
2275 * the data in the second
2277 if (skb->len > skb_availroom(to))
2280 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2283 tcp_collapse_retrans(sk, to);
2287 /* This retransmits one SKB. Policy decisions and retransmit queue
2288 * state updates are done by the caller. Returns non-zero if an
2289 * error occurred which prevented the send.
2291 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2293 struct tcp_sock *tp = tcp_sk(sk);
2294 struct inet_connection_sock *icsk = inet_csk(sk);
2295 unsigned int cur_mss;
2297 /* Inconslusive MTU probe */
2298 if (icsk->icsk_mtup.probe_size) {
2299 icsk->icsk_mtup.probe_size = 0;
2302 /* Do not sent more than we queued. 1/4 is reserved for possible
2303 * copying overhead: fragmentation, tunneling, mangling etc.
2305 if (atomic_read(&sk->sk_wmem_alloc) >
2306 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2309 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2310 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2312 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2316 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2317 return -EHOSTUNREACH; /* Routing failure or similar. */
2319 cur_mss = tcp_current_mss(sk);
2321 /* If receiver has shrunk his window, and skb is out of
2322 * new window, do not retransmit it. The exception is the
2323 * case, when window is shrunk to zero. In this case
2324 * our retransmit serves as a zero window probe.
2326 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2327 TCP_SKB_CB(skb)->seq != tp->snd_una)
2330 if (skb->len > cur_mss) {
2331 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2332 return -ENOMEM; /* We'll try again later. */
2334 int oldpcount = tcp_skb_pcount(skb);
2336 if (unlikely(oldpcount > 1)) {
2337 tcp_init_tso_segs(sk, skb, cur_mss);
2338 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2342 tcp_retrans_try_collapse(sk, skb, cur_mss);
2344 /* Some Solaris stacks overoptimize and ignore the FIN on a
2345 * retransmit when old data is attached. So strip it off
2346 * since it is cheap to do so and saves bytes on the network.
2349 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2350 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2351 if (!pskb_trim(skb, 0)) {
2352 /* Reuse, even though it does some unnecessary work */
2353 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2354 TCP_SKB_CB(skb)->tcp_flags);
2355 skb->ip_summed = CHECKSUM_NONE;
2359 /* Make a copy, if the first transmission SKB clone we made
2360 * is still in somebody's hands, else make a clone.
2362 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2364 /* make sure skb->data is aligned on arches that require it
2365 * and check if ack-trimming & collapsing extended the headroom
2366 * beyond what csum_start can cover.
2368 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2369 skb_headroom(skb) >= 0xFFFF)) {
2370 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2372 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2375 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2379 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2381 struct tcp_sock *tp = tcp_sk(sk);
2382 int err = __tcp_retransmit_skb(sk, skb);
2385 /* Update global TCP statistics. */
2386 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2388 tp->total_retrans++;
2390 #if FASTRETRANS_DEBUG > 0
2391 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2392 net_dbg_ratelimited("retrans_out leaked\n");
2395 if (!tp->retrans_out)
2396 tp->lost_retrans_low = tp->snd_nxt;
2397 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2398 tp->retrans_out += tcp_skb_pcount(skb);
2400 /* Save stamp of the first retransmit. */
2401 if (!tp->retrans_stamp)
2402 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2404 tp->undo_retrans += tcp_skb_pcount(skb);
2406 /* snd_nxt is stored to detect loss of retransmitted segment,
2407 * see tcp_input.c tcp_sacktag_write_queue().
2409 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2414 /* Check if we forward retransmits are possible in the current
2415 * window/congestion state.
2417 static bool tcp_can_forward_retransmit(struct sock *sk)
2419 const struct inet_connection_sock *icsk = inet_csk(sk);
2420 const struct tcp_sock *tp = tcp_sk(sk);
2422 /* Forward retransmissions are possible only during Recovery. */
2423 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2426 /* No forward retransmissions in Reno are possible. */
2427 if (tcp_is_reno(tp))
2430 /* Yeah, we have to make difficult choice between forward transmission
2431 * and retransmission... Both ways have their merits...
2433 * For now we do not retransmit anything, while we have some new
2434 * segments to send. In the other cases, follow rule 3 for
2435 * NextSeg() specified in RFC3517.
2438 if (tcp_may_send_now(sk))
2444 /* This gets called after a retransmit timeout, and the initially
2445 * retransmitted data is acknowledged. It tries to continue
2446 * resending the rest of the retransmit queue, until either
2447 * we've sent it all or the congestion window limit is reached.
2448 * If doing SACK, the first ACK which comes back for a timeout
2449 * based retransmit packet might feed us FACK information again.
2450 * If so, we use it to avoid unnecessarily retransmissions.
2452 void tcp_xmit_retransmit_queue(struct sock *sk)
2454 const struct inet_connection_sock *icsk = inet_csk(sk);
2455 struct tcp_sock *tp = tcp_sk(sk);
2456 struct sk_buff *skb;
2457 struct sk_buff *hole = NULL;
2460 int fwd_rexmitting = 0;
2462 if (!tp->packets_out)
2466 tp->retransmit_high = tp->snd_una;
2468 if (tp->retransmit_skb_hint) {
2469 skb = tp->retransmit_skb_hint;
2470 last_lost = TCP_SKB_CB(skb)->end_seq;
2471 if (after(last_lost, tp->retransmit_high))
2472 last_lost = tp->retransmit_high;
2474 skb = tcp_write_queue_head(sk);
2475 last_lost = tp->snd_una;
2478 tcp_for_write_queue_from(skb, sk) {
2479 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2481 if (skb == tcp_send_head(sk))
2483 /* we could do better than to assign each time */
2485 tp->retransmit_skb_hint = skb;
2487 /* Assume this retransmit will generate
2488 * only one packet for congestion window
2489 * calculation purposes. This works because
2490 * tcp_retransmit_skb() will chop up the
2491 * packet to be MSS sized and all the
2492 * packet counting works out.
2494 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2497 if (fwd_rexmitting) {
2499 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2501 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2503 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2504 tp->retransmit_high = last_lost;
2505 if (!tcp_can_forward_retransmit(sk))
2507 /* Backtrack if necessary to non-L'ed skb */
2515 } else if (!(sacked & TCPCB_LOST)) {
2516 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2521 last_lost = TCP_SKB_CB(skb)->end_seq;
2522 if (icsk->icsk_ca_state != TCP_CA_Loss)
2523 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2525 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2528 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2531 if (tcp_retransmit_skb(sk, skb)) {
2532 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2535 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2537 if (tcp_in_cwnd_reduction(sk))
2538 tp->prr_out += tcp_skb_pcount(skb);
2540 if (skb == tcp_write_queue_head(sk))
2541 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2542 inet_csk(sk)->icsk_rto,
2547 /* Send a fin. The caller locks the socket for us. This cannot be
2548 * allowed to fail queueing a FIN frame under any circumstances.
2550 void tcp_send_fin(struct sock *sk)
2552 struct tcp_sock *tp = tcp_sk(sk);
2553 struct sk_buff *skb = tcp_write_queue_tail(sk);
2556 /* Optimization, tack on the FIN if we have a queue of
2557 * unsent frames. But be careful about outgoing SACKS
2560 mss_now = tcp_current_mss(sk);
2562 if (tcp_send_head(sk) != NULL) {
2563 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2564 TCP_SKB_CB(skb)->end_seq++;
2567 /* Socket is locked, keep trying until memory is available. */
2569 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2576 /* Reserve space for headers and prepare control bits. */
2577 skb_reserve(skb, MAX_TCP_HEADER);
2578 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2579 tcp_init_nondata_skb(skb, tp->write_seq,
2580 TCPHDR_ACK | TCPHDR_FIN);
2581 tcp_queue_skb(sk, skb);
2583 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2586 /* We get here when a process closes a file descriptor (either due to
2587 * an explicit close() or as a byproduct of exit()'ing) and there
2588 * was unread data in the receive queue. This behavior is recommended
2589 * by RFC 2525, section 2.17. -DaveM
2591 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2593 struct sk_buff *skb;
2595 /* NOTE: No TCP options attached and we never retransmit this. */
2596 skb = alloc_skb(MAX_TCP_HEADER, priority);
2598 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2602 /* Reserve space for headers and prepare control bits. */
2603 skb_reserve(skb, MAX_TCP_HEADER);
2604 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2605 TCPHDR_ACK | TCPHDR_RST);
2607 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2608 if (tcp_transmit_skb(sk, skb, 0, priority))
2609 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2611 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2614 /* Send a crossed SYN-ACK during socket establishment.
2615 * WARNING: This routine must only be called when we have already sent
2616 * a SYN packet that crossed the incoming SYN that caused this routine
2617 * to get called. If this assumption fails then the initial rcv_wnd
2618 * and rcv_wscale values will not be correct.
2620 int tcp_send_synack(struct sock *sk)
2622 struct sk_buff *skb;
2624 skb = tcp_write_queue_head(sk);
2625 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2626 pr_debug("%s: wrong queue state\n", __func__);
2629 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2630 if (skb_cloned(skb)) {
2631 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2634 tcp_unlink_write_queue(skb, sk);
2635 skb_header_release(nskb);
2636 __tcp_add_write_queue_head(sk, nskb);
2637 sk_wmem_free_skb(sk, skb);
2638 sk->sk_wmem_queued += nskb->truesize;
2639 sk_mem_charge(sk, nskb->truesize);
2643 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2644 TCP_ECN_send_synack(tcp_sk(sk), skb);
2646 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2647 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2651 * tcp_make_synack - Prepare a SYN-ACK.
2652 * sk: listener socket
2653 * dst: dst entry attached to the SYNACK
2654 * req: request_sock pointer
2656 * Allocate one skb and build a SYNACK packet.
2657 * @dst is consumed : Caller should not use it again.
2659 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2660 struct request_sock *req,
2661 struct tcp_fastopen_cookie *foc)
2663 struct tcp_out_options opts;
2664 struct inet_request_sock *ireq = inet_rsk(req);
2665 struct tcp_sock *tp = tcp_sk(sk);
2667 struct sk_buff *skb;
2668 struct tcp_md5sig_key *md5;
2669 int tcp_header_size;
2672 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2673 if (unlikely(!skb)) {
2677 /* Reserve space for headers. */
2678 skb_reserve(skb, MAX_TCP_HEADER);
2680 skb_dst_set(skb, dst);
2681 security_skb_owned_by(skb, sk);
2683 mss = dst_metric_advmss(dst);
2684 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2685 mss = tp->rx_opt.user_mss;
2687 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2689 /* Set this up on the first call only */
2690 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2692 /* limit the window selection if the user enforce a smaller rx buffer */
2693 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2694 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2695 req->window_clamp = tcp_full_space(sk);
2697 /* tcp_full_space because it is guaranteed to be the first packet */
2698 tcp_select_initial_window(tcp_full_space(sk),
2699 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2704 dst_metric(dst, RTAX_INITRWND));
2705 ireq->rcv_wscale = rcv_wscale;
2708 memset(&opts, 0, sizeof(opts));
2709 #ifdef CONFIG_SYN_COOKIES
2710 if (unlikely(req->cookie_ts))
2711 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2714 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2715 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2718 skb_push(skb, tcp_header_size);
2719 skb_reset_transport_header(skb);
2722 memset(th, 0, sizeof(struct tcphdr));
2725 TCP_ECN_make_synack(req, th);
2726 th->source = ireq->loc_port;
2727 th->dest = ireq->rmt_port;
2728 /* Setting of flags are superfluous here for callers (and ECE is
2729 * not even correctly set)
2731 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2732 TCPHDR_SYN | TCPHDR_ACK);
2734 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2735 /* XXX data is queued and acked as is. No buffer/window check */
2736 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2738 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2739 th->window = htons(min(req->rcv_wnd, 65535U));
2740 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2741 th->doff = (tcp_header_size >> 2);
2742 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2744 #ifdef CONFIG_TCP_MD5SIG
2745 /* Okay, we have all we need - do the md5 hash if needed */
2747 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2748 md5, NULL, req, skb);
2754 EXPORT_SYMBOL(tcp_make_synack);
2756 /* Do all connect socket setups that can be done AF independent. */
2757 void tcp_connect_init(struct sock *sk)
2759 const struct dst_entry *dst = __sk_dst_get(sk);
2760 struct tcp_sock *tp = tcp_sk(sk);
2763 /* We'll fix this up when we get a response from the other end.
2764 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2766 tp->tcp_header_len = sizeof(struct tcphdr) +
2767 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2769 #ifdef CONFIG_TCP_MD5SIG
2770 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2771 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2774 /* If user gave his TCP_MAXSEG, record it to clamp */
2775 if (tp->rx_opt.user_mss)
2776 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2779 tcp_sync_mss(sk, dst_mtu(dst));
2781 if (!tp->window_clamp)
2782 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2783 tp->advmss = dst_metric_advmss(dst);
2784 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2785 tp->advmss = tp->rx_opt.user_mss;
2787 tcp_initialize_rcv_mss(sk);
2789 /* limit the window selection if the user enforce a smaller rx buffer */
2790 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2791 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2792 tp->window_clamp = tcp_full_space(sk);
2794 tcp_select_initial_window(tcp_full_space(sk),
2795 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2798 sysctl_tcp_window_scaling,
2800 dst_metric(dst, RTAX_INITRWND));
2802 tp->rx_opt.rcv_wscale = rcv_wscale;
2803 tp->rcv_ssthresh = tp->rcv_wnd;
2806 sock_reset_flag(sk, SOCK_DONE);
2809 tp->snd_una = tp->write_seq;
2810 tp->snd_sml = tp->write_seq;
2811 tp->snd_up = tp->write_seq;
2812 tp->snd_nxt = tp->write_seq;
2814 if (likely(!tp->repair))
2817 tp->rcv_tstamp = tcp_time_stamp;
2818 tp->rcv_wup = tp->rcv_nxt;
2819 tp->copied_seq = tp->rcv_nxt;
2821 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2822 inet_csk(sk)->icsk_retransmits = 0;
2823 tcp_clear_retrans(tp);
2826 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2828 struct tcp_sock *tp = tcp_sk(sk);
2829 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2831 tcb->end_seq += skb->len;
2832 skb_header_release(skb);
2833 __tcp_add_write_queue_tail(sk, skb);
2834 sk->sk_wmem_queued += skb->truesize;
2835 sk_mem_charge(sk, skb->truesize);
2836 tp->write_seq = tcb->end_seq;
2837 tp->packets_out += tcp_skb_pcount(skb);
2840 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2841 * queue a data-only packet after the regular SYN, such that regular SYNs
2842 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2843 * only the SYN sequence, the data are retransmitted in the first ACK.
2844 * If cookie is not cached or other error occurs, falls back to send a
2845 * regular SYN with Fast Open cookie request option.
2847 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2849 struct tcp_sock *tp = tcp_sk(sk);
2850 struct tcp_fastopen_request *fo = tp->fastopen_req;
2851 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2852 struct sk_buff *syn_data = NULL, *data;
2853 unsigned long last_syn_loss = 0;
2855 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2856 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2857 &syn_loss, &last_syn_loss);
2858 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2860 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2861 fo->cookie.len = -1;
2865 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2866 fo->cookie.len = -1;
2867 else if (fo->cookie.len <= 0)
2870 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2871 * user-MSS. Reserve maximum option space for middleboxes that add
2872 * private TCP options. The cost is reduced data space in SYN :(
2874 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2875 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2876 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2877 MAX_TCP_OPTION_SPACE;
2879 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2881 if (syn_data == NULL)
2884 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2885 struct iovec *iov = &fo->data->msg_iov[i];
2886 unsigned char __user *from = iov->iov_base;
2887 int len = iov->iov_len;
2889 if (syn_data->len + len > space)
2890 len = space - syn_data->len;
2891 else if (i + 1 == iovlen)
2892 /* No more data pending in inet_wait_for_connect() */
2895 if (skb_add_data(syn_data, from, len))
2899 /* Queue a data-only packet after the regular SYN for retransmission */
2900 data = pskb_copy(syn_data, sk->sk_allocation);
2903 TCP_SKB_CB(data)->seq++;
2904 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2905 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2906 tcp_connect_queue_skb(sk, data);
2907 fo->copied = data->len;
2909 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2910 tp->syn_data = (fo->copied > 0);
2911 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2917 /* Send a regular SYN with Fast Open cookie request option */
2918 if (fo->cookie.len > 0)
2920 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2922 tp->syn_fastopen = 0;
2923 kfree_skb(syn_data);
2925 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2929 /* Build a SYN and send it off. */
2930 int tcp_connect(struct sock *sk)
2932 struct tcp_sock *tp = tcp_sk(sk);
2933 struct sk_buff *buff;
2936 tcp_connect_init(sk);
2938 if (unlikely(tp->repair)) {
2939 tcp_finish_connect(sk, NULL);
2943 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2944 if (unlikely(buff == NULL))
2947 /* Reserve space for headers. */
2948 skb_reserve(buff, MAX_TCP_HEADER);
2950 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2951 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2952 tcp_connect_queue_skb(sk, buff);
2953 TCP_ECN_send_syn(sk, buff);
2955 /* Send off SYN; include data in Fast Open. */
2956 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2957 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2958 if (err == -ECONNREFUSED)
2961 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2962 * in order to make this packet get counted in tcpOutSegs.
2964 tp->snd_nxt = tp->write_seq;
2965 tp->pushed_seq = tp->write_seq;
2966 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2968 /* Timer for repeating the SYN until an answer. */
2969 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2970 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2973 EXPORT_SYMBOL(tcp_connect);
2975 /* Send out a delayed ack, the caller does the policy checking
2976 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2979 void tcp_send_delayed_ack(struct sock *sk)
2981 struct inet_connection_sock *icsk = inet_csk(sk);
2982 int ato = icsk->icsk_ack.ato;
2983 unsigned long timeout;
2985 if (ato > TCP_DELACK_MIN) {
2986 const struct tcp_sock *tp = tcp_sk(sk);
2987 int max_ato = HZ / 2;
2989 if (icsk->icsk_ack.pingpong ||
2990 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2991 max_ato = TCP_DELACK_MAX;
2993 /* Slow path, intersegment interval is "high". */
2995 /* If some rtt estimate is known, use it to bound delayed ack.
2996 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3000 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3006 ato = min(ato, max_ato);
3009 /* Stay within the limit we were given */
3010 timeout = jiffies + ato;
3012 /* Use new timeout only if there wasn't a older one earlier. */
3013 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3014 /* If delack timer was blocked or is about to expire,
3017 if (icsk->icsk_ack.blocked ||
3018 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3023 if (!time_before(timeout, icsk->icsk_ack.timeout))
3024 timeout = icsk->icsk_ack.timeout;
3026 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3027 icsk->icsk_ack.timeout = timeout;
3028 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3031 /* This routine sends an ack and also updates the window. */
3032 void tcp_send_ack(struct sock *sk)
3034 struct sk_buff *buff;
3036 /* If we have been reset, we may not send again. */
3037 if (sk->sk_state == TCP_CLOSE)
3040 /* We are not putting this on the write queue, so
3041 * tcp_transmit_skb() will set the ownership to this
3044 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3046 inet_csk_schedule_ack(sk);
3047 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3048 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3049 TCP_DELACK_MAX, TCP_RTO_MAX);
3053 /* Reserve space for headers and prepare control bits. */
3054 skb_reserve(buff, MAX_TCP_HEADER);
3055 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3057 /* Send it off, this clears delayed acks for us. */
3058 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3059 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3062 /* This routine sends a packet with an out of date sequence
3063 * number. It assumes the other end will try to ack it.
3065 * Question: what should we make while urgent mode?
3066 * 4.4BSD forces sending single byte of data. We cannot send
3067 * out of window data, because we have SND.NXT==SND.MAX...
3069 * Current solution: to send TWO zero-length segments in urgent mode:
3070 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3071 * out-of-date with SND.UNA-1 to probe window.
3073 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3075 struct tcp_sock *tp = tcp_sk(sk);
3076 struct sk_buff *skb;
3078 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3079 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3083 /* Reserve space for headers and set control bits. */
3084 skb_reserve(skb, MAX_TCP_HEADER);
3085 /* Use a previous sequence. This should cause the other
3086 * end to send an ack. Don't queue or clone SKB, just
3089 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3090 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3091 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3094 void tcp_send_window_probe(struct sock *sk)
3096 if (sk->sk_state == TCP_ESTABLISHED) {
3097 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3098 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3099 tcp_xmit_probe_skb(sk, 0);
3103 /* Initiate keepalive or window probe from timer. */
3104 int tcp_write_wakeup(struct sock *sk)
3106 struct tcp_sock *tp = tcp_sk(sk);
3107 struct sk_buff *skb;
3109 if (sk->sk_state == TCP_CLOSE)
3112 if ((skb = tcp_send_head(sk)) != NULL &&
3113 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3115 unsigned int mss = tcp_current_mss(sk);
3116 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3118 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3119 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3121 /* We are probing the opening of a window
3122 * but the window size is != 0
3123 * must have been a result SWS avoidance ( sender )
3125 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3127 seg_size = min(seg_size, mss);
3128 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3129 if (tcp_fragment(sk, skb, seg_size, mss))
3131 } else if (!tcp_skb_pcount(skb))
3132 tcp_set_skb_tso_segs(sk, skb, mss);
3134 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3135 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3136 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3138 tcp_event_new_data_sent(sk, skb);
3141 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3142 tcp_xmit_probe_skb(sk, 1);
3143 return tcp_xmit_probe_skb(sk, 0);
3147 /* A window probe timeout has occurred. If window is not closed send
3148 * a partial packet else a zero probe.
3150 void tcp_send_probe0(struct sock *sk)
3152 struct inet_connection_sock *icsk = inet_csk(sk);
3153 struct tcp_sock *tp = tcp_sk(sk);
3156 err = tcp_write_wakeup(sk);
3158 if (tp->packets_out || !tcp_send_head(sk)) {
3159 /* Cancel probe timer, if it is not required. */
3160 icsk->icsk_probes_out = 0;
3161 icsk->icsk_backoff = 0;
3166 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3167 icsk->icsk_backoff++;
3168 icsk->icsk_probes_out++;
3169 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3170 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3173 /* If packet was not sent due to local congestion,
3174 * do not backoff and do not remember icsk_probes_out.
3175 * Let local senders to fight for local resources.
3177 * Use accumulated backoff yet.
3179 if (!icsk->icsk_probes_out)
3180 icsk->icsk_probes_out = 1;
3181 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3182 min(icsk->icsk_rto << icsk->icsk_backoff,
3183 TCP_RESOURCE_PROBE_INTERVAL),