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), tcp_sk(sk)->nonagle,
693 * One tasklest per cpu tries to send more skbs.
694 * We run in tasklet context but need to disable irqs when
695 * transfering tsq->head because tcp_wfree() might
696 * interrupt us (non NAPI drivers)
698 static void tcp_tasklet_func(unsigned long data)
700 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
703 struct list_head *q, *n;
707 local_irq_save(flags);
708 list_splice_init(&tsq->head, &list);
709 local_irq_restore(flags);
711 list_for_each_safe(q, n, &list) {
712 tp = list_entry(q, struct tcp_sock, tsq_node);
713 list_del(&tp->tsq_node);
715 sk = (struct sock *)tp;
718 if (!sock_owned_by_user(sk)) {
721 /* defer the work to tcp_release_cb() */
722 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
726 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
731 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
732 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
733 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
734 (1UL << TCP_MTU_REDUCED_DEFERRED))
736 * tcp_release_cb - tcp release_sock() callback
739 * called from release_sock() to perform protocol dependent
740 * actions before socket release.
742 void tcp_release_cb(struct sock *sk)
744 struct tcp_sock *tp = tcp_sk(sk);
745 unsigned long flags, nflags;
747 /* perform an atomic operation only if at least one flag is set */
749 flags = tp->tsq_flags;
750 if (!(flags & TCP_DEFERRED_ALL))
752 nflags = flags & ~TCP_DEFERRED_ALL;
753 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
755 if (flags & (1UL << TCP_TSQ_DEFERRED))
758 /* Here begins the tricky part :
759 * We are called from release_sock() with :
761 * 2) sk_lock.slock spinlock held
762 * 3) socket owned by us (sk->sk_lock.owned == 1)
764 * But following code is meant to be called from BH handlers,
765 * so we should keep BH disabled, but early release socket ownership
767 sock_release_ownership(sk);
769 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
770 tcp_write_timer_handler(sk);
773 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
774 tcp_delack_timer_handler(sk);
777 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
778 sk->sk_prot->mtu_reduced(sk);
782 EXPORT_SYMBOL(tcp_release_cb);
784 void __init tcp_tasklet_init(void)
788 for_each_possible_cpu(i) {
789 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
791 INIT_LIST_HEAD(&tsq->head);
792 tasklet_init(&tsq->tasklet,
799 * Write buffer destructor automatically called from kfree_skb.
800 * We cant xmit new skbs from this context, as we might already
803 void tcp_wfree(struct sk_buff *skb)
805 struct sock *sk = skb->sk;
806 struct tcp_sock *tp = tcp_sk(sk);
808 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
809 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
811 struct tsq_tasklet *tsq;
813 /* Keep a ref on socket.
814 * This last ref will be released in tcp_tasklet_func()
816 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
818 /* queue this socket to tasklet queue */
819 local_irq_save(flags);
820 tsq = &__get_cpu_var(tsq_tasklet);
821 list_add(&tp->tsq_node, &tsq->head);
822 tasklet_schedule(&tsq->tasklet);
823 local_irq_restore(flags);
829 /* This routine actually transmits TCP packets queued in by
830 * tcp_do_sendmsg(). This is used by both the initial
831 * transmission and possible later retransmissions.
832 * All SKB's seen here are completely headerless. It is our
833 * job to build the TCP header, and pass the packet down to
834 * IP so it can do the same plus pass the packet off to the
837 * We are working here with either a clone of the original
838 * SKB, or a fresh unique copy made by the retransmit engine.
840 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
843 const struct inet_connection_sock *icsk = inet_csk(sk);
844 struct inet_sock *inet;
846 struct tcp_skb_cb *tcb;
847 struct tcp_out_options opts;
848 unsigned int tcp_options_size, tcp_header_size;
849 struct tcp_md5sig_key *md5;
853 BUG_ON(!skb || !tcp_skb_pcount(skb));
855 /* If congestion control is doing timestamping, we must
856 * take such a timestamp before we potentially clone/copy.
858 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
859 __net_timestamp(skb);
861 if (likely(clone_it)) {
862 const struct sk_buff *fclone = skb + 1;
864 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
865 fclone->fclone == SKB_FCLONE_CLONE))
866 NET_INC_STATS_BH(sock_net(sk),
867 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
869 if (unlikely(skb_cloned(skb)))
870 skb = pskb_copy(skb, gfp_mask);
872 skb = skb_clone(skb, gfp_mask);
879 tcb = TCP_SKB_CB(skb);
880 memset(&opts, 0, sizeof(opts));
882 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
883 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
885 tcp_options_size = tcp_established_options(sk, skb, &opts,
887 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
889 if (tcp_packets_in_flight(tp) == 0)
890 tcp_ca_event(sk, CA_EVENT_TX_START);
892 /* if no packet is in qdisc/device queue, then allow XPS to select
895 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
897 skb_push(skb, tcp_header_size);
898 skb_reset_transport_header(skb);
902 skb->destructor = tcp_wfree;
903 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
905 /* Build TCP header and checksum it. */
907 th->source = inet->inet_sport;
908 th->dest = inet->inet_dport;
909 th->seq = htonl(tcb->seq);
910 th->ack_seq = htonl(tp->rcv_nxt);
911 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
914 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
915 /* RFC1323: The window in SYN & SYN/ACK segments
918 th->window = htons(min(tp->rcv_wnd, 65535U));
920 th->window = htons(tcp_select_window(sk));
925 /* The urg_mode check is necessary during a below snd_una win probe */
926 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
927 if (before(tp->snd_up, tcb->seq + 0x10000)) {
928 th->urg_ptr = htons(tp->snd_up - tcb->seq);
930 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
931 th->urg_ptr = htons(0xFFFF);
936 tcp_options_write((__be32 *)(th + 1), tp, &opts);
937 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
938 TCP_ECN_send(sk, skb, tcp_header_size);
940 #ifdef CONFIG_TCP_MD5SIG
941 /* Calculate the MD5 hash, as we have all we need now */
943 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
944 tp->af_specific->calc_md5_hash(opts.hash_location,
949 icsk->icsk_af_ops->send_check(sk, skb);
951 if (likely(tcb->tcp_flags & TCPHDR_ACK))
952 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
954 if (skb->len != tcp_header_size)
955 tcp_event_data_sent(tp, sk);
957 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
958 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
959 tcp_skb_pcount(skb));
961 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
962 if (likely(err <= 0))
965 tcp_enter_cwr(sk, 1);
967 return net_xmit_eval(err);
970 /* This routine just queues the buffer for sending.
972 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
973 * otherwise socket can stall.
975 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
977 struct tcp_sock *tp = tcp_sk(sk);
979 /* Advance write_seq and place onto the write_queue. */
980 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
981 skb_header_release(skb);
982 tcp_add_write_queue_tail(sk, skb);
983 sk->sk_wmem_queued += skb->truesize;
984 sk_mem_charge(sk, skb->truesize);
987 /* Initialize TSO segments for a packet. */
988 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
989 unsigned int mss_now)
991 /* Make sure we own this skb before messing gso_size/gso_segs */
992 WARN_ON_ONCE(skb_cloned(skb));
994 if (skb->len <= mss_now || !sk_can_gso(sk) ||
995 skb->ip_summed == CHECKSUM_NONE) {
996 /* Avoid the costly divide in the normal
999 skb_shinfo(skb)->gso_segs = 1;
1000 skb_shinfo(skb)->gso_size = 0;
1001 skb_shinfo(skb)->gso_type = 0;
1003 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1004 skb_shinfo(skb)->gso_size = mss_now;
1005 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1009 /* When a modification to fackets out becomes necessary, we need to check
1010 * skb is counted to fackets_out or not.
1012 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1015 struct tcp_sock *tp = tcp_sk(sk);
1017 if (!tp->sacked_out || tcp_is_reno(tp))
1020 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1021 tp->fackets_out -= decr;
1024 /* Pcount in the middle of the write queue got changed, we need to do various
1025 * tweaks to fix counters
1027 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1029 struct tcp_sock *tp = tcp_sk(sk);
1031 tp->packets_out -= decr;
1033 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1034 tp->sacked_out -= decr;
1035 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1036 tp->retrans_out -= decr;
1037 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1038 tp->lost_out -= decr;
1040 /* Reno case is special. Sigh... */
1041 if (tcp_is_reno(tp) && decr > 0)
1042 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1044 tcp_adjust_fackets_out(sk, skb, decr);
1046 if (tp->lost_skb_hint &&
1047 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1048 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1049 tp->lost_cnt_hint -= decr;
1051 tcp_verify_left_out(tp);
1054 /* Function to create two new TCP segments. Shrinks the given segment
1055 * to the specified size and appends a new segment with the rest of the
1056 * packet to the list. This won't be called frequently, I hope.
1057 * Remember, these are still headerless SKBs at this point.
1059 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1060 unsigned int mss_now)
1062 struct tcp_sock *tp = tcp_sk(sk);
1063 struct sk_buff *buff;
1064 int nsize, old_factor;
1068 if (WARN_ON(len > skb->len))
1071 nsize = skb_headlen(skb) - len;
1075 if (skb_unclone(skb, GFP_ATOMIC))
1078 /* Get a new skb... force flag on. */
1079 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1081 return -ENOMEM; /* We'll just try again later. */
1083 sk->sk_wmem_queued += buff->truesize;
1084 sk_mem_charge(sk, buff->truesize);
1085 nlen = skb->len - len - nsize;
1086 buff->truesize += nlen;
1087 skb->truesize -= nlen;
1089 /* Correct the sequence numbers. */
1090 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1091 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1092 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1094 /* PSH and FIN should only be set in the second packet. */
1095 flags = TCP_SKB_CB(skb)->tcp_flags;
1096 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1097 TCP_SKB_CB(buff)->tcp_flags = flags;
1098 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1100 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1101 /* Copy and checksum data tail into the new buffer. */
1102 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1103 skb_put(buff, nsize),
1108 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1110 skb->ip_summed = CHECKSUM_PARTIAL;
1111 skb_split(skb, buff, len);
1114 buff->ip_summed = skb->ip_summed;
1116 /* Looks stupid, but our code really uses when of
1117 * skbs, which it never sent before. --ANK
1119 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1120 buff->tstamp = skb->tstamp;
1122 old_factor = tcp_skb_pcount(skb);
1124 /* Fix up tso_factor for both original and new SKB. */
1125 tcp_set_skb_tso_segs(sk, skb, mss_now);
1126 tcp_set_skb_tso_segs(sk, buff, mss_now);
1128 /* If this packet has been sent out already, we must
1129 * adjust the various packet counters.
1131 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1132 int diff = old_factor - tcp_skb_pcount(skb) -
1133 tcp_skb_pcount(buff);
1136 tcp_adjust_pcount(sk, skb, diff);
1139 /* Link BUFF into the send queue. */
1140 skb_header_release(buff);
1141 tcp_insert_write_queue_after(skb, buff, sk);
1146 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1147 * eventually). The difference is that pulled data not copied, but
1148 * immediately discarded.
1150 static void __pskb_trim_head(struct sk_buff *skb, int len)
1154 eat = min_t(int, len, skb_headlen(skb));
1156 __skb_pull(skb, eat);
1163 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1164 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1167 skb_frag_unref(skb, i);
1170 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1172 skb_shinfo(skb)->frags[k].page_offset += eat;
1173 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1179 skb_shinfo(skb)->nr_frags = k;
1181 skb_reset_tail_pointer(skb);
1182 skb->data_len -= len;
1183 skb->len = skb->data_len;
1186 /* Remove acked data from a packet in the transmit queue. */
1187 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1189 if (skb_unclone(skb, GFP_ATOMIC))
1192 __pskb_trim_head(skb, len);
1194 TCP_SKB_CB(skb)->seq += len;
1195 skb->ip_summed = CHECKSUM_PARTIAL;
1197 skb->truesize -= len;
1198 sk->sk_wmem_queued -= len;
1199 sk_mem_uncharge(sk, len);
1200 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1202 /* Any change of skb->len requires recalculation of tso factor. */
1203 if (tcp_skb_pcount(skb) > 1)
1204 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1209 /* Calculate MSS not accounting any TCP options. */
1210 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1212 const struct tcp_sock *tp = tcp_sk(sk);
1213 const struct inet_connection_sock *icsk = inet_csk(sk);
1216 /* Calculate base mss without TCP options:
1217 It is MMS_S - sizeof(tcphdr) of rfc1122
1219 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1221 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1222 if (icsk->icsk_af_ops->net_frag_header_len) {
1223 const struct dst_entry *dst = __sk_dst_get(sk);
1225 if (dst && dst_allfrag(dst))
1226 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1229 /* Clamp it (mss_clamp does not include tcp options) */
1230 if (mss_now > tp->rx_opt.mss_clamp)
1231 mss_now = tp->rx_opt.mss_clamp;
1233 /* Now subtract optional transport overhead */
1234 mss_now -= icsk->icsk_ext_hdr_len;
1236 /* Then reserve room for full set of TCP options and 8 bytes of data */
1242 /* Calculate MSS. Not accounting for SACKs here. */
1243 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1245 /* Subtract TCP options size, not including SACKs */
1246 return __tcp_mtu_to_mss(sk, pmtu) -
1247 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1250 /* Inverse of above */
1251 int tcp_mss_to_mtu(struct sock *sk, int mss)
1253 const struct tcp_sock *tp = tcp_sk(sk);
1254 const struct inet_connection_sock *icsk = inet_csk(sk);
1258 tp->tcp_header_len +
1259 icsk->icsk_ext_hdr_len +
1260 icsk->icsk_af_ops->net_header_len;
1262 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1263 if (icsk->icsk_af_ops->net_frag_header_len) {
1264 const struct dst_entry *dst = __sk_dst_get(sk);
1266 if (dst && dst_allfrag(dst))
1267 mtu += icsk->icsk_af_ops->net_frag_header_len;
1272 /* MTU probing init per socket */
1273 void tcp_mtup_init(struct sock *sk)
1275 struct tcp_sock *tp = tcp_sk(sk);
1276 struct inet_connection_sock *icsk = inet_csk(sk);
1278 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1279 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1280 icsk->icsk_af_ops->net_header_len;
1281 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1282 icsk->icsk_mtup.probe_size = 0;
1284 EXPORT_SYMBOL(tcp_mtup_init);
1286 /* This function synchronize snd mss to current pmtu/exthdr set.
1288 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1289 for TCP options, but includes only bare TCP header.
1291 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1292 It is minimum of user_mss and mss received with SYN.
1293 It also does not include TCP options.
1295 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1297 tp->mss_cache is current effective sending mss, including
1298 all tcp options except for SACKs. It is evaluated,
1299 taking into account current pmtu, but never exceeds
1300 tp->rx_opt.mss_clamp.
1302 NOTE1. rfc1122 clearly states that advertised MSS
1303 DOES NOT include either tcp or ip options.
1305 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1306 are READ ONLY outside this function. --ANK (980731)
1308 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1310 struct tcp_sock *tp = tcp_sk(sk);
1311 struct inet_connection_sock *icsk = inet_csk(sk);
1314 if (icsk->icsk_mtup.search_high > pmtu)
1315 icsk->icsk_mtup.search_high = pmtu;
1317 mss_now = tcp_mtu_to_mss(sk, pmtu);
1318 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1320 /* And store cached results */
1321 icsk->icsk_pmtu_cookie = pmtu;
1322 if (icsk->icsk_mtup.enabled)
1323 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1324 tp->mss_cache = mss_now;
1328 EXPORT_SYMBOL(tcp_sync_mss);
1330 /* Compute the current effective MSS, taking SACKs and IP options,
1331 * and even PMTU discovery events into account.
1333 unsigned int tcp_current_mss(struct sock *sk)
1335 const struct tcp_sock *tp = tcp_sk(sk);
1336 const struct dst_entry *dst = __sk_dst_get(sk);
1338 unsigned int header_len;
1339 struct tcp_out_options opts;
1340 struct tcp_md5sig_key *md5;
1342 mss_now = tp->mss_cache;
1345 u32 mtu = dst_mtu(dst);
1346 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1347 mss_now = tcp_sync_mss(sk, mtu);
1350 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1351 sizeof(struct tcphdr);
1352 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1353 * some common options. If this is an odd packet (because we have SACK
1354 * blocks etc) then our calculated header_len will be different, and
1355 * we have to adjust mss_now correspondingly */
1356 if (header_len != tp->tcp_header_len) {
1357 int delta = (int) header_len - tp->tcp_header_len;
1364 /* Congestion window validation. (RFC2861) */
1365 static void tcp_cwnd_validate(struct sock *sk)
1367 struct tcp_sock *tp = tcp_sk(sk);
1369 if (tp->packets_out >= tp->snd_cwnd) {
1370 /* Network is feed fully. */
1371 tp->snd_cwnd_used = 0;
1372 tp->snd_cwnd_stamp = tcp_time_stamp;
1374 /* Network starves. */
1375 if (tp->packets_out > tp->snd_cwnd_used)
1376 tp->snd_cwnd_used = tp->packets_out;
1378 if (sysctl_tcp_slow_start_after_idle &&
1379 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1380 tcp_cwnd_application_limited(sk);
1384 /* Returns the portion of skb which can be sent right away without
1385 * introducing MSS oddities to segment boundaries. In rare cases where
1386 * mss_now != mss_cache, we will request caller to create a small skb
1387 * per input skb which could be mostly avoided here (if desired).
1389 * We explicitly want to create a request for splitting write queue tail
1390 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1391 * thus all the complexity (cwnd_len is always MSS multiple which we
1392 * return whenever allowed by the other factors). Basically we need the
1393 * modulo only when the receiver window alone is the limiting factor or
1394 * when we would be allowed to send the split-due-to-Nagle skb fully.
1396 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1397 unsigned int mss_now, unsigned int max_segs)
1399 const struct tcp_sock *tp = tcp_sk(sk);
1400 u32 needed, window, max_len;
1402 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1403 max_len = mss_now * max_segs;
1405 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1408 needed = min(skb->len, window);
1410 if (max_len <= needed)
1413 return needed - needed % mss_now;
1416 /* Can at least one segment of SKB be sent right now, according to the
1417 * congestion window rules? If so, return how many segments are allowed.
1419 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1420 const struct sk_buff *skb)
1422 u32 in_flight, cwnd;
1424 /* Don't be strict about the congestion window for the final FIN. */
1425 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1426 tcp_skb_pcount(skb) == 1)
1429 in_flight = tcp_packets_in_flight(tp);
1430 cwnd = tp->snd_cwnd;
1431 if (in_flight < cwnd)
1432 return (cwnd - in_flight);
1437 /* Initialize TSO state of a skb.
1438 * This must be invoked the first time we consider transmitting
1439 * SKB onto the wire.
1441 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1442 unsigned int mss_now)
1444 int tso_segs = tcp_skb_pcount(skb);
1446 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1447 tcp_set_skb_tso_segs(sk, skb, mss_now);
1448 tso_segs = tcp_skb_pcount(skb);
1453 /* Minshall's variant of the Nagle send check. */
1454 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1456 return after(tp->snd_sml, tp->snd_una) &&
1457 !after(tp->snd_sml, tp->snd_nxt);
1460 /* Return false, if packet can be sent now without violation Nagle's rules:
1461 * 1. It is full sized.
1462 * 2. Or it contains FIN. (already checked by caller)
1463 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1464 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1465 * With Minshall's modification: all sent small packets are ACKed.
1467 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1468 const struct sk_buff *skb,
1469 unsigned int mss_now, int nonagle)
1471 return skb->len < mss_now &&
1472 ((nonagle & TCP_NAGLE_CORK) ||
1473 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1476 /* Return true if the Nagle test allows this packet to be
1479 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1480 unsigned int cur_mss, int nonagle)
1482 /* Nagle rule does not apply to frames, which sit in the middle of the
1483 * write_queue (they have no chances to get new data).
1485 * This is implemented in the callers, where they modify the 'nonagle'
1486 * argument based upon the location of SKB in the send queue.
1488 if (nonagle & TCP_NAGLE_PUSH)
1491 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1492 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1495 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1501 /* Does at least the first segment of SKB fit into the send window? */
1502 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1503 const struct sk_buff *skb,
1504 unsigned int cur_mss)
1506 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1508 if (skb->len > cur_mss)
1509 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1511 return !after(end_seq, tcp_wnd_end(tp));
1514 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1515 * should be put on the wire right now. If so, it returns the number of
1516 * packets allowed by the congestion window.
1518 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1519 unsigned int cur_mss, int nonagle)
1521 const struct tcp_sock *tp = tcp_sk(sk);
1522 unsigned int cwnd_quota;
1524 tcp_init_tso_segs(sk, skb, cur_mss);
1526 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1529 cwnd_quota = tcp_cwnd_test(tp, skb);
1530 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1536 /* Test if sending is allowed right now. */
1537 bool tcp_may_send_now(struct sock *sk)
1539 const struct tcp_sock *tp = tcp_sk(sk);
1540 struct sk_buff *skb = tcp_send_head(sk);
1543 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1544 (tcp_skb_is_last(sk, skb) ?
1545 tp->nonagle : TCP_NAGLE_PUSH));
1548 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1549 * which is put after SKB on the list. It is very much like
1550 * tcp_fragment() except that it may make several kinds of assumptions
1551 * in order to speed up the splitting operation. In particular, we
1552 * know that all the data is in scatter-gather pages, and that the
1553 * packet has never been sent out before (and thus is not cloned).
1555 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1556 unsigned int mss_now, gfp_t gfp)
1558 struct sk_buff *buff;
1559 int nlen = skb->len - len;
1562 /* All of a TSO frame must be composed of paged data. */
1563 if (skb->len != skb->data_len)
1564 return tcp_fragment(sk, skb, len, mss_now);
1566 buff = sk_stream_alloc_skb(sk, 0, gfp);
1567 if (unlikely(buff == NULL))
1570 sk->sk_wmem_queued += buff->truesize;
1571 sk_mem_charge(sk, buff->truesize);
1572 buff->truesize += nlen;
1573 skb->truesize -= nlen;
1575 /* Correct the sequence numbers. */
1576 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1577 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1578 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1580 /* PSH and FIN should only be set in the second packet. */
1581 flags = TCP_SKB_CB(skb)->tcp_flags;
1582 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1583 TCP_SKB_CB(buff)->tcp_flags = flags;
1585 /* This packet was never sent out yet, so no SACK bits. */
1586 TCP_SKB_CB(buff)->sacked = 0;
1588 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1589 skb_split(skb, buff, len);
1591 /* Fix up tso_factor for both original and new SKB. */
1592 tcp_set_skb_tso_segs(sk, skb, mss_now);
1593 tcp_set_skb_tso_segs(sk, buff, mss_now);
1595 /* Link BUFF into the send queue. */
1596 skb_header_release(buff);
1597 tcp_insert_write_queue_after(skb, buff, sk);
1602 /* Try to defer sending, if possible, in order to minimize the amount
1603 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1605 * This algorithm is from John Heffner.
1607 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1609 struct tcp_sock *tp = tcp_sk(sk);
1610 const struct inet_connection_sock *icsk = inet_csk(sk);
1611 u32 send_win, cong_win, limit, in_flight;
1614 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1617 if (icsk->icsk_ca_state != TCP_CA_Open)
1620 /* Defer for less than two clock ticks. */
1621 if (tp->tso_deferred &&
1622 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1625 in_flight = tcp_packets_in_flight(tp);
1627 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1629 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1631 /* From in_flight test above, we know that cwnd > in_flight. */
1632 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1634 limit = min(send_win, cong_win);
1636 /* If a full-sized TSO skb can be sent, do it. */
1637 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1638 tp->xmit_size_goal_segs * tp->mss_cache))
1641 /* Middle in queue won't get any more data, full sendable already? */
1642 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1645 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1647 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1649 /* If at least some fraction of a window is available,
1652 chunk /= win_divisor;
1656 /* Different approach, try not to defer past a single
1657 * ACK. Receiver should ACK every other full sized
1658 * frame, so if we have space for more than 3 frames
1661 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1665 /* Ok, it looks like it is advisable to defer.
1666 * Do not rearm the timer if already set to not break TCP ACK clocking.
1668 if (!tp->tso_deferred)
1669 tp->tso_deferred = 1 | (jiffies << 1);
1674 tp->tso_deferred = 0;
1678 /* Create a new MTU probe if we are ready.
1679 * MTU probe is regularly attempting to increase the path MTU by
1680 * deliberately sending larger packets. This discovers routing
1681 * changes resulting in larger path MTUs.
1683 * Returns 0 if we should wait to probe (no cwnd available),
1684 * 1 if a probe was sent,
1687 static int tcp_mtu_probe(struct sock *sk)
1689 struct tcp_sock *tp = tcp_sk(sk);
1690 struct inet_connection_sock *icsk = inet_csk(sk);
1691 struct sk_buff *skb, *nskb, *next;
1698 /* Not currently probing/verifying,
1700 * have enough cwnd, and
1701 * not SACKing (the variable headers throw things off) */
1702 if (!icsk->icsk_mtup.enabled ||
1703 icsk->icsk_mtup.probe_size ||
1704 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1705 tp->snd_cwnd < 11 ||
1706 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1709 /* Very simple search strategy: just double the MSS. */
1710 mss_now = tcp_current_mss(sk);
1711 probe_size = 2 * tp->mss_cache;
1712 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1713 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1714 /* TODO: set timer for probe_converge_event */
1718 /* Have enough data in the send queue to probe? */
1719 if (tp->write_seq - tp->snd_nxt < size_needed)
1722 if (tp->snd_wnd < size_needed)
1724 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1727 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1728 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1729 if (!tcp_packets_in_flight(tp))
1735 /* We're allowed to probe. Build it now. */
1736 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1738 sk->sk_wmem_queued += nskb->truesize;
1739 sk_mem_charge(sk, nskb->truesize);
1741 skb = tcp_send_head(sk);
1743 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1744 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1745 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1746 TCP_SKB_CB(nskb)->sacked = 0;
1748 nskb->ip_summed = skb->ip_summed;
1750 tcp_insert_write_queue_before(nskb, skb, sk);
1753 tcp_for_write_queue_from_safe(skb, next, sk) {
1754 copy = min_t(int, skb->len, probe_size - len);
1755 if (nskb->ip_summed)
1756 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1758 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1759 skb_put(nskb, copy),
1762 if (skb->len <= copy) {
1763 /* We've eaten all the data from this skb.
1765 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1766 tcp_unlink_write_queue(skb, sk);
1767 sk_wmem_free_skb(sk, skb);
1769 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1770 ~(TCPHDR_FIN|TCPHDR_PSH);
1771 if (!skb_shinfo(skb)->nr_frags) {
1772 skb_pull(skb, copy);
1773 if (skb->ip_summed != CHECKSUM_PARTIAL)
1774 skb->csum = csum_partial(skb->data,
1777 __pskb_trim_head(skb, copy);
1778 tcp_set_skb_tso_segs(sk, skb, mss_now);
1780 TCP_SKB_CB(skb)->seq += copy;
1785 if (len >= probe_size)
1788 tcp_init_tso_segs(sk, nskb, nskb->len);
1790 /* We're ready to send. If this fails, the probe will
1791 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1792 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1793 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1794 /* Decrement cwnd here because we are sending
1795 * effectively two packets. */
1797 tcp_event_new_data_sent(sk, nskb);
1799 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1800 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1801 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1809 /* This routine writes packets to the network. It advances the
1810 * send_head. This happens as incoming acks open up the remote
1813 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1814 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1815 * account rare use of URG, this is not a big flaw.
1817 * Send at most one packet when push_one > 0. Temporarily ignore
1818 * cwnd limit to force at most one packet out when push_one == 2.
1820 * Returns true, if no segments are in flight and we have queued segments,
1821 * but cannot send anything now because of SWS or another problem.
1823 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1824 int push_one, gfp_t gfp)
1826 struct tcp_sock *tp = tcp_sk(sk);
1827 struct sk_buff *skb;
1828 unsigned int tso_segs, sent_pkts;
1835 /* Do MTU probing. */
1836 result = tcp_mtu_probe(sk);
1839 } else if (result > 0) {
1844 while ((skb = tcp_send_head(sk))) {
1847 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1850 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1851 goto repair; /* Skip network transmission */
1853 cwnd_quota = tcp_cwnd_test(tp, skb);
1856 /* Force out a loss probe pkt. */
1862 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1865 if (tso_segs == 1) {
1866 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1867 (tcp_skb_is_last(sk, skb) ?
1868 nonagle : TCP_NAGLE_PUSH))))
1871 if (!push_one && tcp_tso_should_defer(sk, skb))
1875 /* TCP Small Queues :
1876 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1878 * - better RTT estimation and ACK scheduling
1881 * Alas, some drivers / subsystems require a fair amount
1882 * of queued bytes to ensure line rate.
1883 * One example is wifi aggregation (802.11 AMPDU)
1885 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1886 sk->sk_pacing_rate >> 10);
1888 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1889 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1890 /* It is possible TX completion already happened
1891 * before we set TSQ_THROTTLED, so we must
1892 * test again the condition.
1893 * We abuse smp_mb__after_clear_bit() because
1894 * there is no smp_mb__after_set_bit() yet
1896 smp_mb__after_clear_bit();
1897 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1902 if (tso_segs > 1 && !tcp_urg_mode(tp))
1903 limit = tcp_mss_split_point(sk, skb, mss_now,
1906 sk->sk_gso_max_segs));
1908 if (skb->len > limit &&
1909 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1912 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1914 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1918 /* Advance the send_head. This one is sent out.
1919 * This call will increment packets_out.
1921 tcp_event_new_data_sent(sk, skb);
1923 tcp_minshall_update(tp, mss_now, skb);
1924 sent_pkts += tcp_skb_pcount(skb);
1930 if (likely(sent_pkts)) {
1931 if (tcp_in_cwnd_reduction(sk))
1932 tp->prr_out += sent_pkts;
1934 /* Send one loss probe per tail loss episode. */
1936 tcp_schedule_loss_probe(sk);
1937 tcp_cwnd_validate(sk);
1940 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1943 bool tcp_schedule_loss_probe(struct sock *sk)
1945 struct inet_connection_sock *icsk = inet_csk(sk);
1946 struct tcp_sock *tp = tcp_sk(sk);
1947 u32 timeout, tlp_time_stamp, rto_time_stamp;
1948 u32 rtt = tp->srtt >> 3;
1950 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1952 /* No consecutive loss probes. */
1953 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1957 /* Don't do any loss probe on a Fast Open connection before 3WHS
1960 if (sk->sk_state == TCP_SYN_RECV)
1963 /* TLP is only scheduled when next timer event is RTO. */
1964 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1967 /* Schedule a loss probe in 2*RTT for SACK capable connections
1968 * in Open state, that are either limited by cwnd or application.
1970 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1971 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1974 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1978 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1979 * for delayed ack when there's one outstanding packet.
1982 if (tp->packets_out == 1)
1983 timeout = max_t(u32, timeout,
1984 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1985 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1987 /* If RTO is shorter, just schedule TLP in its place. */
1988 tlp_time_stamp = tcp_time_stamp + timeout;
1989 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1990 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1991 s32 delta = rto_time_stamp - tcp_time_stamp;
1996 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2001 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2002 * retransmit the last segment.
2004 void tcp_send_loss_probe(struct sock *sk)
2006 struct tcp_sock *tp = tcp_sk(sk);
2007 struct sk_buff *skb;
2009 int mss = tcp_current_mss(sk);
2012 if (tcp_send_head(sk) != NULL) {
2013 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2017 /* At most one outstanding TLP retransmission. */
2018 if (tp->tlp_high_seq)
2021 /* Retransmit last segment. */
2022 skb = tcp_write_queue_tail(sk);
2026 pcount = tcp_skb_pcount(skb);
2027 if (WARN_ON(!pcount))
2030 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2031 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2033 skb = tcp_write_queue_tail(sk);
2036 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2039 /* Probe with zero data doesn't trigger fast recovery. */
2041 err = __tcp_retransmit_skb(sk, skb);
2043 /* Record snd_nxt for loss detection. */
2045 tp->tlp_high_seq = tp->snd_nxt;
2048 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2049 inet_csk(sk)->icsk_rto,
2053 NET_INC_STATS_BH(sock_net(sk),
2054 LINUX_MIB_TCPLOSSPROBES);
2058 /* Push out any pending frames which were held back due to
2059 * TCP_CORK or attempt at coalescing tiny packets.
2060 * The socket must be locked by the caller.
2062 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2065 /* If we are closed, the bytes will have to remain here.
2066 * In time closedown will finish, we empty the write queue and
2067 * all will be happy.
2069 if (unlikely(sk->sk_state == TCP_CLOSE))
2072 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2073 sk_gfp_atomic(sk, GFP_ATOMIC)))
2074 tcp_check_probe_timer(sk);
2077 /* Send _single_ skb sitting at the send head. This function requires
2078 * true push pending frames to setup probe timer etc.
2080 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2082 struct sk_buff *skb = tcp_send_head(sk);
2084 BUG_ON(!skb || skb->len < mss_now);
2086 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2089 /* This function returns the amount that we can raise the
2090 * usable window based on the following constraints
2092 * 1. The window can never be shrunk once it is offered (RFC 793)
2093 * 2. We limit memory per socket
2096 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2097 * RECV.NEXT + RCV.WIN fixed until:
2098 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2100 * i.e. don't raise the right edge of the window until you can raise
2101 * it at least MSS bytes.
2103 * Unfortunately, the recommended algorithm breaks header prediction,
2104 * since header prediction assumes th->window stays fixed.
2106 * Strictly speaking, keeping th->window fixed violates the receiver
2107 * side SWS prevention criteria. The problem is that under this rule
2108 * a stream of single byte packets will cause the right side of the
2109 * window to always advance by a single byte.
2111 * Of course, if the sender implements sender side SWS prevention
2112 * then this will not be a problem.
2114 * BSD seems to make the following compromise:
2116 * If the free space is less than the 1/4 of the maximum
2117 * space available and the free space is less than 1/2 mss,
2118 * then set the window to 0.
2119 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2120 * Otherwise, just prevent the window from shrinking
2121 * and from being larger than the largest representable value.
2123 * This prevents incremental opening of the window in the regime
2124 * where TCP is limited by the speed of the reader side taking
2125 * data out of the TCP receive queue. It does nothing about
2126 * those cases where the window is constrained on the sender side
2127 * because the pipeline is full.
2129 * BSD also seems to "accidentally" limit itself to windows that are a
2130 * multiple of MSS, at least until the free space gets quite small.
2131 * This would appear to be a side effect of the mbuf implementation.
2132 * Combining these two algorithms results in the observed behavior
2133 * of having a fixed window size at almost all times.
2135 * Below we obtain similar behavior by forcing the offered window to
2136 * a multiple of the mss when it is feasible to do so.
2138 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2139 * Regular options like TIMESTAMP are taken into account.
2141 u32 __tcp_select_window(struct sock *sk)
2143 struct inet_connection_sock *icsk = inet_csk(sk);
2144 struct tcp_sock *tp = tcp_sk(sk);
2145 /* MSS for the peer's data. Previous versions used mss_clamp
2146 * here. I don't know if the value based on our guesses
2147 * of peer's MSS is better for the performance. It's more correct
2148 * but may be worse for the performance because of rcv_mss
2149 * fluctuations. --SAW 1998/11/1
2151 int mss = icsk->icsk_ack.rcv_mss;
2152 int free_space = tcp_space(sk);
2153 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2156 if (mss > full_space)
2159 if (free_space < (full_space >> 1)) {
2160 icsk->icsk_ack.quick = 0;
2162 if (sk_under_memory_pressure(sk))
2163 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2166 if (free_space < mss)
2170 if (free_space > tp->rcv_ssthresh)
2171 free_space = tp->rcv_ssthresh;
2173 /* Don't do rounding if we are using window scaling, since the
2174 * scaled window will not line up with the MSS boundary anyway.
2176 window = tp->rcv_wnd;
2177 if (tp->rx_opt.rcv_wscale) {
2178 window = free_space;
2180 /* Advertise enough space so that it won't get scaled away.
2181 * Import case: prevent zero window announcement if
2182 * 1<<rcv_wscale > mss.
2184 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2185 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2186 << tp->rx_opt.rcv_wscale);
2188 /* Get the largest window that is a nice multiple of mss.
2189 * Window clamp already applied above.
2190 * If our current window offering is within 1 mss of the
2191 * free space we just keep it. This prevents the divide
2192 * and multiply from happening most of the time.
2193 * We also don't do any window rounding when the free space
2196 if (window <= free_space - mss || window > free_space)
2197 window = (free_space / mss) * mss;
2198 else if (mss == full_space &&
2199 free_space > window + (full_space >> 1))
2200 window = free_space;
2206 /* Collapses two adjacent SKB's during retransmission. */
2207 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2209 struct tcp_sock *tp = tcp_sk(sk);
2210 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2211 int skb_size, next_skb_size;
2213 skb_size = skb->len;
2214 next_skb_size = next_skb->len;
2216 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2218 tcp_highest_sack_combine(sk, next_skb, skb);
2220 tcp_unlink_write_queue(next_skb, sk);
2222 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2225 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2226 skb->ip_summed = CHECKSUM_PARTIAL;
2228 if (skb->ip_summed != CHECKSUM_PARTIAL)
2229 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2231 /* Update sequence range on original skb. */
2232 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2234 /* Merge over control information. This moves PSH/FIN etc. over */
2235 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2237 /* All done, get rid of second SKB and account for it so
2238 * packet counting does not break.
2240 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2242 /* changed transmit queue under us so clear hints */
2243 tcp_clear_retrans_hints_partial(tp);
2244 if (next_skb == tp->retransmit_skb_hint)
2245 tp->retransmit_skb_hint = skb;
2247 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2249 sk_wmem_free_skb(sk, next_skb);
2252 /* Check if coalescing SKBs is legal. */
2253 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2255 if (tcp_skb_pcount(skb) > 1)
2257 /* TODO: SACK collapsing could be used to remove this condition */
2258 if (skb_shinfo(skb)->nr_frags != 0)
2260 if (skb_cloned(skb))
2262 if (skb == tcp_send_head(sk))
2264 /* Some heurestics for collapsing over SACK'd could be invented */
2265 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2271 /* Collapse packets in the retransmit queue to make to create
2272 * less packets on the wire. This is only done on retransmission.
2274 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2277 struct tcp_sock *tp = tcp_sk(sk);
2278 struct sk_buff *skb = to, *tmp;
2281 if (!sysctl_tcp_retrans_collapse)
2283 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2286 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2287 if (!tcp_can_collapse(sk, skb))
2299 /* Punt if not enough space exists in the first SKB for
2300 * the data in the second
2302 if (skb->len > skb_availroom(to))
2305 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2308 tcp_collapse_retrans(sk, to);
2312 /* This retransmits one SKB. Policy decisions and retransmit queue
2313 * state updates are done by the caller. Returns non-zero if an
2314 * error occurred which prevented the send.
2316 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2318 struct tcp_sock *tp = tcp_sk(sk);
2319 struct inet_connection_sock *icsk = inet_csk(sk);
2320 unsigned int cur_mss;
2322 /* Inconslusive MTU probe */
2323 if (icsk->icsk_mtup.probe_size) {
2324 icsk->icsk_mtup.probe_size = 0;
2327 /* Do not sent more than we queued. 1/4 is reserved for possible
2328 * copying overhead: fragmentation, tunneling, mangling etc.
2330 if (atomic_read(&sk->sk_wmem_alloc) >
2331 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2334 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2335 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2337 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2341 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2342 return -EHOSTUNREACH; /* Routing failure or similar. */
2344 cur_mss = tcp_current_mss(sk);
2346 /* If receiver has shrunk his window, and skb is out of
2347 * new window, do not retransmit it. The exception is the
2348 * case, when window is shrunk to zero. In this case
2349 * our retransmit serves as a zero window probe.
2351 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2352 TCP_SKB_CB(skb)->seq != tp->snd_una)
2355 if (skb->len > cur_mss) {
2356 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2357 return -ENOMEM; /* We'll try again later. */
2359 int oldpcount = tcp_skb_pcount(skb);
2361 if (unlikely(oldpcount > 1)) {
2362 if (skb_unclone(skb, GFP_ATOMIC))
2364 tcp_init_tso_segs(sk, skb, cur_mss);
2365 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2369 tcp_retrans_try_collapse(sk, skb, cur_mss);
2371 /* Some Solaris stacks overoptimize and ignore the FIN on a
2372 * retransmit when old data is attached. So strip it off
2373 * since it is cheap to do so and saves bytes on the network.
2376 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2377 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2378 if (!pskb_trim(skb, 0)) {
2379 /* Reuse, even though it does some unnecessary work */
2380 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2381 TCP_SKB_CB(skb)->tcp_flags);
2382 skb->ip_summed = CHECKSUM_NONE;
2386 /* Make a copy, if the first transmission SKB clone we made
2387 * is still in somebody's hands, else make a clone.
2389 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2391 /* make sure skb->data is aligned on arches that require it
2392 * and check if ack-trimming & collapsing extended the headroom
2393 * beyond what csum_start can cover.
2395 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2396 skb_headroom(skb) >= 0xFFFF)) {
2397 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2399 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2402 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2406 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2408 struct tcp_sock *tp = tcp_sk(sk);
2409 int err = __tcp_retransmit_skb(sk, skb);
2412 /* Update global TCP statistics. */
2413 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2415 tp->total_retrans++;
2417 #if FASTRETRANS_DEBUG > 0
2418 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2419 net_dbg_ratelimited("retrans_out leaked\n");
2422 if (!tp->retrans_out)
2423 tp->lost_retrans_low = tp->snd_nxt;
2424 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2425 tp->retrans_out += tcp_skb_pcount(skb);
2427 /* Save stamp of the first retransmit. */
2428 if (!tp->retrans_stamp)
2429 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2431 tp->undo_retrans += tcp_skb_pcount(skb);
2433 /* snd_nxt is stored to detect loss of retransmitted segment,
2434 * see tcp_input.c tcp_sacktag_write_queue().
2436 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2441 /* Check if we forward retransmits are possible in the current
2442 * window/congestion state.
2444 static bool tcp_can_forward_retransmit(struct sock *sk)
2446 const struct inet_connection_sock *icsk = inet_csk(sk);
2447 const struct tcp_sock *tp = tcp_sk(sk);
2449 /* Forward retransmissions are possible only during Recovery. */
2450 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2453 /* No forward retransmissions in Reno are possible. */
2454 if (tcp_is_reno(tp))
2457 /* Yeah, we have to make difficult choice between forward transmission
2458 * and retransmission... Both ways have their merits...
2460 * For now we do not retransmit anything, while we have some new
2461 * segments to send. In the other cases, follow rule 3 for
2462 * NextSeg() specified in RFC3517.
2465 if (tcp_may_send_now(sk))
2471 /* This gets called after a retransmit timeout, and the initially
2472 * retransmitted data is acknowledged. It tries to continue
2473 * resending the rest of the retransmit queue, until either
2474 * we've sent it all or the congestion window limit is reached.
2475 * If doing SACK, the first ACK which comes back for a timeout
2476 * based retransmit packet might feed us FACK information again.
2477 * If so, we use it to avoid unnecessarily retransmissions.
2479 void tcp_xmit_retransmit_queue(struct sock *sk)
2481 const struct inet_connection_sock *icsk = inet_csk(sk);
2482 struct tcp_sock *tp = tcp_sk(sk);
2483 struct sk_buff *skb;
2484 struct sk_buff *hole = NULL;
2487 int fwd_rexmitting = 0;
2489 if (!tp->packets_out)
2493 tp->retransmit_high = tp->snd_una;
2495 if (tp->retransmit_skb_hint) {
2496 skb = tp->retransmit_skb_hint;
2497 last_lost = TCP_SKB_CB(skb)->end_seq;
2498 if (after(last_lost, tp->retransmit_high))
2499 last_lost = tp->retransmit_high;
2501 skb = tcp_write_queue_head(sk);
2502 last_lost = tp->snd_una;
2505 tcp_for_write_queue_from(skb, sk) {
2506 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2508 if (skb == tcp_send_head(sk))
2510 /* we could do better than to assign each time */
2512 tp->retransmit_skb_hint = skb;
2514 /* Assume this retransmit will generate
2515 * only one packet for congestion window
2516 * calculation purposes. This works because
2517 * tcp_retransmit_skb() will chop up the
2518 * packet to be MSS sized and all the
2519 * packet counting works out.
2521 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2524 if (fwd_rexmitting) {
2526 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2528 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2530 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2531 tp->retransmit_high = last_lost;
2532 if (!tcp_can_forward_retransmit(sk))
2534 /* Backtrack if necessary to non-L'ed skb */
2542 } else if (!(sacked & TCPCB_LOST)) {
2543 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2548 last_lost = TCP_SKB_CB(skb)->end_seq;
2549 if (icsk->icsk_ca_state != TCP_CA_Loss)
2550 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2552 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2555 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2558 if (tcp_retransmit_skb(sk, skb)) {
2559 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2562 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2564 if (tcp_in_cwnd_reduction(sk))
2565 tp->prr_out += tcp_skb_pcount(skb);
2567 if (skb == tcp_write_queue_head(sk))
2568 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2569 inet_csk(sk)->icsk_rto,
2574 /* Send a fin. The caller locks the socket for us. This cannot be
2575 * allowed to fail queueing a FIN frame under any circumstances.
2577 void tcp_send_fin(struct sock *sk)
2579 struct tcp_sock *tp = tcp_sk(sk);
2580 struct sk_buff *skb = tcp_write_queue_tail(sk);
2583 /* Optimization, tack on the FIN if we have a queue of
2584 * unsent frames. But be careful about outgoing SACKS
2587 mss_now = tcp_current_mss(sk);
2589 if (tcp_send_head(sk) != NULL) {
2590 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2591 TCP_SKB_CB(skb)->end_seq++;
2594 /* Socket is locked, keep trying until memory is available. */
2596 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2603 /* Reserve space for headers and prepare control bits. */
2604 skb_reserve(skb, MAX_TCP_HEADER);
2605 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2606 tcp_init_nondata_skb(skb, tp->write_seq,
2607 TCPHDR_ACK | TCPHDR_FIN);
2608 tcp_queue_skb(sk, skb);
2610 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2613 /* We get here when a process closes a file descriptor (either due to
2614 * an explicit close() or as a byproduct of exit()'ing) and there
2615 * was unread data in the receive queue. This behavior is recommended
2616 * by RFC 2525, section 2.17. -DaveM
2618 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2620 struct sk_buff *skb;
2622 /* NOTE: No TCP options attached and we never retransmit this. */
2623 skb = alloc_skb(MAX_TCP_HEADER, priority);
2625 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2629 /* Reserve space for headers and prepare control bits. */
2630 skb_reserve(skb, MAX_TCP_HEADER);
2631 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2632 TCPHDR_ACK | TCPHDR_RST);
2634 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2635 if (tcp_transmit_skb(sk, skb, 0, priority))
2636 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2638 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2641 /* Send a crossed SYN-ACK during socket establishment.
2642 * WARNING: This routine must only be called when we have already sent
2643 * a SYN packet that crossed the incoming SYN that caused this routine
2644 * to get called. If this assumption fails then the initial rcv_wnd
2645 * and rcv_wscale values will not be correct.
2647 int tcp_send_synack(struct sock *sk)
2649 struct sk_buff *skb;
2651 skb = tcp_write_queue_head(sk);
2652 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2653 pr_debug("%s: wrong queue state\n", __func__);
2656 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2657 if (skb_cloned(skb)) {
2658 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2661 tcp_unlink_write_queue(skb, sk);
2662 skb_header_release(nskb);
2663 __tcp_add_write_queue_head(sk, nskb);
2664 sk_wmem_free_skb(sk, skb);
2665 sk->sk_wmem_queued += nskb->truesize;
2666 sk_mem_charge(sk, nskb->truesize);
2670 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2671 TCP_ECN_send_synack(tcp_sk(sk), skb);
2673 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2674 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2678 * tcp_make_synack - Prepare a SYN-ACK.
2679 * sk: listener socket
2680 * dst: dst entry attached to the SYNACK
2681 * req: request_sock pointer
2683 * Allocate one skb and build a SYNACK packet.
2684 * @dst is consumed : Caller should not use it again.
2686 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2687 struct request_sock *req,
2688 struct tcp_fastopen_cookie *foc)
2690 struct tcp_out_options opts;
2691 struct inet_request_sock *ireq = inet_rsk(req);
2692 struct tcp_sock *tp = tcp_sk(sk);
2694 struct sk_buff *skb;
2695 struct tcp_md5sig_key *md5;
2696 int tcp_header_size;
2699 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2700 if (unlikely(!skb)) {
2704 /* Reserve space for headers. */
2705 skb_reserve(skb, MAX_TCP_HEADER);
2707 skb_dst_set(skb, dst);
2708 security_skb_owned_by(skb, sk);
2710 mss = dst_metric_advmss(dst);
2711 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2712 mss = tp->rx_opt.user_mss;
2714 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2716 /* Set this up on the first call only */
2717 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2719 /* limit the window selection if the user enforce a smaller rx buffer */
2720 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2721 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2722 req->window_clamp = tcp_full_space(sk);
2724 /* tcp_full_space because it is guaranteed to be the first packet */
2725 tcp_select_initial_window(tcp_full_space(sk),
2726 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2731 dst_metric(dst, RTAX_INITRWND));
2732 ireq->rcv_wscale = rcv_wscale;
2735 memset(&opts, 0, sizeof(opts));
2736 #ifdef CONFIG_SYN_COOKIES
2737 if (unlikely(req->cookie_ts))
2738 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2741 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2742 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2745 skb_push(skb, tcp_header_size);
2746 skb_reset_transport_header(skb);
2749 memset(th, 0, sizeof(struct tcphdr));
2752 TCP_ECN_make_synack(req, th);
2753 th->source = ireq->loc_port;
2754 th->dest = ireq->rmt_port;
2755 /* Setting of flags are superfluous here for callers (and ECE is
2756 * not even correctly set)
2758 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2759 TCPHDR_SYN | TCPHDR_ACK);
2761 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2762 /* XXX data is queued and acked as is. No buffer/window check */
2763 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2765 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2766 th->window = htons(min(req->rcv_wnd, 65535U));
2767 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2768 th->doff = (tcp_header_size >> 2);
2769 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2771 #ifdef CONFIG_TCP_MD5SIG
2772 /* Okay, we have all we need - do the md5 hash if needed */
2774 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2775 md5, NULL, req, skb);
2781 EXPORT_SYMBOL(tcp_make_synack);
2783 /* Do all connect socket setups that can be done AF independent. */
2784 void tcp_connect_init(struct sock *sk)
2786 const struct dst_entry *dst = __sk_dst_get(sk);
2787 struct tcp_sock *tp = tcp_sk(sk);
2790 /* We'll fix this up when we get a response from the other end.
2791 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2793 tp->tcp_header_len = sizeof(struct tcphdr) +
2794 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2796 #ifdef CONFIG_TCP_MD5SIG
2797 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2798 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2801 /* If user gave his TCP_MAXSEG, record it to clamp */
2802 if (tp->rx_opt.user_mss)
2803 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2806 tcp_sync_mss(sk, dst_mtu(dst));
2808 if (!tp->window_clamp)
2809 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2810 tp->advmss = dst_metric_advmss(dst);
2811 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2812 tp->advmss = tp->rx_opt.user_mss;
2814 tcp_initialize_rcv_mss(sk);
2816 /* limit the window selection if the user enforce a smaller rx buffer */
2817 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2818 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2819 tp->window_clamp = tcp_full_space(sk);
2821 tcp_select_initial_window(tcp_full_space(sk),
2822 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2825 sysctl_tcp_window_scaling,
2827 dst_metric(dst, RTAX_INITRWND));
2829 tp->rx_opt.rcv_wscale = rcv_wscale;
2830 tp->rcv_ssthresh = tp->rcv_wnd;
2833 sock_reset_flag(sk, SOCK_DONE);
2836 tp->snd_una = tp->write_seq;
2837 tp->snd_sml = tp->write_seq;
2838 tp->snd_up = tp->write_seq;
2839 tp->snd_nxt = tp->write_seq;
2841 if (likely(!tp->repair))
2844 tp->rcv_tstamp = tcp_time_stamp;
2845 tp->rcv_wup = tp->rcv_nxt;
2846 tp->copied_seq = tp->rcv_nxt;
2848 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2849 inet_csk(sk)->icsk_retransmits = 0;
2850 tcp_clear_retrans(tp);
2853 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2855 struct tcp_sock *tp = tcp_sk(sk);
2856 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2858 tcb->end_seq += skb->len;
2859 skb_header_release(skb);
2860 __tcp_add_write_queue_tail(sk, skb);
2861 sk->sk_wmem_queued += skb->truesize;
2862 sk_mem_charge(sk, skb->truesize);
2863 tp->write_seq = tcb->end_seq;
2864 tp->packets_out += tcp_skb_pcount(skb);
2867 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2868 * queue a data-only packet after the regular SYN, such that regular SYNs
2869 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2870 * only the SYN sequence, the data are retransmitted in the first ACK.
2871 * If cookie is not cached or other error occurs, falls back to send a
2872 * regular SYN with Fast Open cookie request option.
2874 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2876 struct tcp_sock *tp = tcp_sk(sk);
2877 struct tcp_fastopen_request *fo = tp->fastopen_req;
2878 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2879 struct sk_buff *syn_data = NULL, *data;
2880 unsigned long last_syn_loss = 0;
2882 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2883 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2884 &syn_loss, &last_syn_loss);
2885 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2887 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2888 fo->cookie.len = -1;
2892 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2893 fo->cookie.len = -1;
2894 else if (fo->cookie.len <= 0)
2897 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2898 * user-MSS. Reserve maximum option space for middleboxes that add
2899 * private TCP options. The cost is reduced data space in SYN :(
2901 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2902 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2903 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2904 MAX_TCP_OPTION_SPACE;
2906 space = min_t(size_t, space, fo->size);
2908 /* limit to order-0 allocations */
2909 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
2911 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
2913 if (syn_data == NULL)
2916 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2917 struct iovec *iov = &fo->data->msg_iov[i];
2918 unsigned char __user *from = iov->iov_base;
2919 int len = iov->iov_len;
2921 if (syn_data->len + len > space)
2922 len = space - syn_data->len;
2923 else if (i + 1 == iovlen)
2924 /* No more data pending in inet_wait_for_connect() */
2927 if (skb_add_data(syn_data, from, len))
2931 /* Queue a data-only packet after the regular SYN for retransmission */
2932 data = pskb_copy(syn_data, sk->sk_allocation);
2935 TCP_SKB_CB(data)->seq++;
2936 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2937 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2938 tcp_connect_queue_skb(sk, data);
2939 fo->copied = data->len;
2941 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2942 tp->syn_data = (fo->copied > 0);
2943 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2949 /* Send a regular SYN with Fast Open cookie request option */
2950 if (fo->cookie.len > 0)
2952 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2954 tp->syn_fastopen = 0;
2955 kfree_skb(syn_data);
2957 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2961 /* Build a SYN and send it off. */
2962 int tcp_connect(struct sock *sk)
2964 struct tcp_sock *tp = tcp_sk(sk);
2965 struct sk_buff *buff;
2968 tcp_connect_init(sk);
2970 if (unlikely(tp->repair)) {
2971 tcp_finish_connect(sk, NULL);
2975 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2976 if (unlikely(buff == NULL))
2979 /* Reserve space for headers. */
2980 skb_reserve(buff, MAX_TCP_HEADER);
2982 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2983 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2984 tcp_connect_queue_skb(sk, buff);
2985 TCP_ECN_send_syn(sk, buff);
2987 /* Send off SYN; include data in Fast Open. */
2988 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2989 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2990 if (err == -ECONNREFUSED)
2993 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2994 * in order to make this packet get counted in tcpOutSegs.
2996 tp->snd_nxt = tp->write_seq;
2997 tp->pushed_seq = tp->write_seq;
2998 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3000 /* Timer for repeating the SYN until an answer. */
3001 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3002 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3005 EXPORT_SYMBOL(tcp_connect);
3007 /* Send out a delayed ack, the caller does the policy checking
3008 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3011 void tcp_send_delayed_ack(struct sock *sk)
3013 struct inet_connection_sock *icsk = inet_csk(sk);
3014 int ato = icsk->icsk_ack.ato;
3015 unsigned long timeout;
3017 if (ato > TCP_DELACK_MIN) {
3018 const struct tcp_sock *tp = tcp_sk(sk);
3019 int max_ato = HZ / 2;
3021 if (icsk->icsk_ack.pingpong ||
3022 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3023 max_ato = TCP_DELACK_MAX;
3025 /* Slow path, intersegment interval is "high". */
3027 /* If some rtt estimate is known, use it to bound delayed ack.
3028 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3032 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3038 ato = min(ato, max_ato);
3041 /* Stay within the limit we were given */
3042 timeout = jiffies + ato;
3044 /* Use new timeout only if there wasn't a older one earlier. */
3045 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3046 /* If delack timer was blocked or is about to expire,
3049 if (icsk->icsk_ack.blocked ||
3050 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3055 if (!time_before(timeout, icsk->icsk_ack.timeout))
3056 timeout = icsk->icsk_ack.timeout;
3058 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3059 icsk->icsk_ack.timeout = timeout;
3060 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3063 /* This routine sends an ack and also updates the window. */
3064 void tcp_send_ack(struct sock *sk)
3066 struct sk_buff *buff;
3068 /* If we have been reset, we may not send again. */
3069 if (sk->sk_state == TCP_CLOSE)
3072 /* We are not putting this on the write queue, so
3073 * tcp_transmit_skb() will set the ownership to this
3076 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3078 inet_csk_schedule_ack(sk);
3079 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3080 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3081 TCP_DELACK_MAX, TCP_RTO_MAX);
3085 /* Reserve space for headers and prepare control bits. */
3086 skb_reserve(buff, MAX_TCP_HEADER);
3087 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3089 /* Send it off, this clears delayed acks for us. */
3090 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3091 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3094 /* This routine sends a packet with an out of date sequence
3095 * number. It assumes the other end will try to ack it.
3097 * Question: what should we make while urgent mode?
3098 * 4.4BSD forces sending single byte of data. We cannot send
3099 * out of window data, because we have SND.NXT==SND.MAX...
3101 * Current solution: to send TWO zero-length segments in urgent mode:
3102 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3103 * out-of-date with SND.UNA-1 to probe window.
3105 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3107 struct tcp_sock *tp = tcp_sk(sk);
3108 struct sk_buff *skb;
3110 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3111 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3115 /* Reserve space for headers and set control bits. */
3116 skb_reserve(skb, MAX_TCP_HEADER);
3117 /* Use a previous sequence. This should cause the other
3118 * end to send an ack. Don't queue or clone SKB, just
3121 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3122 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3123 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3126 void tcp_send_window_probe(struct sock *sk)
3128 if (sk->sk_state == TCP_ESTABLISHED) {
3129 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3130 tcp_xmit_probe_skb(sk, 0);
3134 /* Initiate keepalive or window probe from timer. */
3135 int tcp_write_wakeup(struct sock *sk)
3137 struct tcp_sock *tp = tcp_sk(sk);
3138 struct sk_buff *skb;
3140 if (sk->sk_state == TCP_CLOSE)
3143 if ((skb = tcp_send_head(sk)) != NULL &&
3144 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3146 unsigned int mss = tcp_current_mss(sk);
3147 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3149 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3150 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3152 /* We are probing the opening of a window
3153 * but the window size is != 0
3154 * must have been a result SWS avoidance ( sender )
3156 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3158 seg_size = min(seg_size, mss);
3159 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3160 if (tcp_fragment(sk, skb, seg_size, mss))
3162 } else if (!tcp_skb_pcount(skb))
3163 tcp_set_skb_tso_segs(sk, skb, mss);
3165 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3166 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3167 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3169 tcp_event_new_data_sent(sk, skb);
3172 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3173 tcp_xmit_probe_skb(sk, 1);
3174 return tcp_xmit_probe_skb(sk, 0);
3178 /* A window probe timeout has occurred. If window is not closed send
3179 * a partial packet else a zero probe.
3181 void tcp_send_probe0(struct sock *sk)
3183 struct inet_connection_sock *icsk = inet_csk(sk);
3184 struct tcp_sock *tp = tcp_sk(sk);
3187 err = tcp_write_wakeup(sk);
3189 if (tp->packets_out || !tcp_send_head(sk)) {
3190 /* Cancel probe timer, if it is not required. */
3191 icsk->icsk_probes_out = 0;
3192 icsk->icsk_backoff = 0;
3197 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3198 icsk->icsk_backoff++;
3199 icsk->icsk_probes_out++;
3200 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3201 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3204 /* If packet was not sent due to local congestion,
3205 * do not backoff and do not remember icsk_probes_out.
3206 * Let local senders to fight for local resources.
3208 * Use accumulated backoff yet.
3210 if (!icsk->icsk_probes_out)
3211 icsk->icsk_probes_out = 1;
3212 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3213 min(icsk->icsk_rto << icsk->icsk_backoff,
3214 TCP_RESOURCE_PROBE_INTERVAL),