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 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
759 tcp_write_timer_handler(sk);
762 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
763 tcp_delack_timer_handler(sk);
766 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
767 sk->sk_prot->mtu_reduced(sk);
771 EXPORT_SYMBOL(tcp_release_cb);
773 void __init tcp_tasklet_init(void)
777 for_each_possible_cpu(i) {
778 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
780 INIT_LIST_HEAD(&tsq->head);
781 tasklet_init(&tsq->tasklet,
788 * Write buffer destructor automatically called from kfree_skb.
789 * We cant xmit new skbs from this context, as we might already
792 void tcp_wfree(struct sk_buff *skb)
794 struct sock *sk = skb->sk;
795 struct tcp_sock *tp = tcp_sk(sk);
797 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
798 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
800 struct tsq_tasklet *tsq;
802 /* Keep a ref on socket.
803 * This last ref will be released in tcp_tasklet_func()
805 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
807 /* queue this socket to tasklet queue */
808 local_irq_save(flags);
809 tsq = &__get_cpu_var(tsq_tasklet);
810 list_add(&tp->tsq_node, &tsq->head);
811 tasklet_schedule(&tsq->tasklet);
812 local_irq_restore(flags);
818 /* This routine actually transmits TCP packets queued in by
819 * tcp_do_sendmsg(). This is used by both the initial
820 * transmission and possible later retransmissions.
821 * All SKB's seen here are completely headerless. It is our
822 * job to build the TCP header, and pass the packet down to
823 * IP so it can do the same plus pass the packet off to the
826 * We are working here with either a clone of the original
827 * SKB, or a fresh unique copy made by the retransmit engine.
829 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
832 const struct inet_connection_sock *icsk = inet_csk(sk);
833 struct inet_sock *inet;
835 struct tcp_skb_cb *tcb;
836 struct tcp_out_options opts;
837 unsigned int tcp_options_size, tcp_header_size;
838 struct tcp_md5sig_key *md5;
842 BUG_ON(!skb || !tcp_skb_pcount(skb));
844 /* If congestion control is doing timestamping, we must
845 * take such a timestamp before we potentially clone/copy.
847 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
848 __net_timestamp(skb);
850 if (likely(clone_it)) {
851 const struct sk_buff *fclone = skb + 1;
853 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
854 fclone->fclone == SKB_FCLONE_CLONE))
855 NET_INC_STATS_BH(sock_net(sk),
856 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
858 if (unlikely(skb_cloned(skb)))
859 skb = pskb_copy(skb, gfp_mask);
861 skb = skb_clone(skb, gfp_mask);
868 tcb = TCP_SKB_CB(skb);
869 memset(&opts, 0, sizeof(opts));
871 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
872 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
874 tcp_options_size = tcp_established_options(sk, skb, &opts,
876 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
878 if (tcp_packets_in_flight(tp) == 0)
879 tcp_ca_event(sk, CA_EVENT_TX_START);
881 /* if no packet is in qdisc/device queue, then allow XPS to select
884 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
886 skb_push(skb, tcp_header_size);
887 skb_reset_transport_header(skb);
891 skb->destructor = tcp_wfree;
892 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
894 /* Build TCP header and checksum it. */
896 th->source = inet->inet_sport;
897 th->dest = inet->inet_dport;
898 th->seq = htonl(tcb->seq);
899 th->ack_seq = htonl(tp->rcv_nxt);
900 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
903 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
904 /* RFC1323: The window in SYN & SYN/ACK segments
907 th->window = htons(min(tp->rcv_wnd, 65535U));
909 th->window = htons(tcp_select_window(sk));
914 /* The urg_mode check is necessary during a below snd_una win probe */
915 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
916 if (before(tp->snd_up, tcb->seq + 0x10000)) {
917 th->urg_ptr = htons(tp->snd_up - tcb->seq);
919 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
920 th->urg_ptr = htons(0xFFFF);
925 tcp_options_write((__be32 *)(th + 1), tp, &opts);
926 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
927 TCP_ECN_send(sk, skb, tcp_header_size);
929 #ifdef CONFIG_TCP_MD5SIG
930 /* Calculate the MD5 hash, as we have all we need now */
932 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
933 tp->af_specific->calc_md5_hash(opts.hash_location,
938 icsk->icsk_af_ops->send_check(sk, skb);
940 if (likely(tcb->tcp_flags & TCPHDR_ACK))
941 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
943 if (skb->len != tcp_header_size)
944 tcp_event_data_sent(tp, sk);
946 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
947 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
948 tcp_skb_pcount(skb));
950 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
951 if (likely(err <= 0))
954 tcp_enter_cwr(sk, 1);
956 return net_xmit_eval(err);
959 /* This routine just queues the buffer for sending.
961 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
962 * otherwise socket can stall.
964 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
966 struct tcp_sock *tp = tcp_sk(sk);
968 /* Advance write_seq and place onto the write_queue. */
969 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
970 skb_header_release(skb);
971 tcp_add_write_queue_tail(sk, skb);
972 sk->sk_wmem_queued += skb->truesize;
973 sk_mem_charge(sk, skb->truesize);
976 /* Initialize TSO segments for a packet. */
977 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
978 unsigned int mss_now)
980 /* Make sure we own this skb before messing gso_size/gso_segs */
981 WARN_ON_ONCE(skb_cloned(skb));
983 if (skb->len <= mss_now || !sk_can_gso(sk) ||
984 skb->ip_summed == CHECKSUM_NONE) {
985 /* Avoid the costly divide in the normal
988 skb_shinfo(skb)->gso_segs = 1;
989 skb_shinfo(skb)->gso_size = 0;
990 skb_shinfo(skb)->gso_type = 0;
992 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
993 skb_shinfo(skb)->gso_size = mss_now;
994 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
998 /* When a modification to fackets out becomes necessary, we need to check
999 * skb is counted to fackets_out or not.
1001 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1004 struct tcp_sock *tp = tcp_sk(sk);
1006 if (!tp->sacked_out || tcp_is_reno(tp))
1009 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1010 tp->fackets_out -= decr;
1013 /* Pcount in the middle of the write queue got changed, we need to do various
1014 * tweaks to fix counters
1016 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1018 struct tcp_sock *tp = tcp_sk(sk);
1020 tp->packets_out -= decr;
1022 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1023 tp->sacked_out -= decr;
1024 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1025 tp->retrans_out -= decr;
1026 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1027 tp->lost_out -= decr;
1029 /* Reno case is special. Sigh... */
1030 if (tcp_is_reno(tp) && decr > 0)
1031 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1033 tcp_adjust_fackets_out(sk, skb, decr);
1035 if (tp->lost_skb_hint &&
1036 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1037 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1038 tp->lost_cnt_hint -= decr;
1040 tcp_verify_left_out(tp);
1043 /* Function to create two new TCP segments. Shrinks the given segment
1044 * to the specified size and appends a new segment with the rest of the
1045 * packet to the list. This won't be called frequently, I hope.
1046 * Remember, these are still headerless SKBs at this point.
1048 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1049 unsigned int mss_now)
1051 struct tcp_sock *tp = tcp_sk(sk);
1052 struct sk_buff *buff;
1053 int nsize, old_factor;
1057 if (WARN_ON(len > skb->len))
1060 nsize = skb_headlen(skb) - len;
1064 if (skb_unclone(skb, GFP_ATOMIC))
1067 /* Get a new skb... force flag on. */
1068 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1070 return -ENOMEM; /* We'll just try again later. */
1072 sk->sk_wmem_queued += buff->truesize;
1073 sk_mem_charge(sk, buff->truesize);
1074 nlen = skb->len - len - nsize;
1075 buff->truesize += nlen;
1076 skb->truesize -= nlen;
1078 /* Correct the sequence numbers. */
1079 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1080 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1081 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1083 /* PSH and FIN should only be set in the second packet. */
1084 flags = TCP_SKB_CB(skb)->tcp_flags;
1085 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1086 TCP_SKB_CB(buff)->tcp_flags = flags;
1087 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1089 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1090 /* Copy and checksum data tail into the new buffer. */
1091 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1092 skb_put(buff, nsize),
1097 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1099 skb->ip_summed = CHECKSUM_PARTIAL;
1100 skb_split(skb, buff, len);
1103 buff->ip_summed = skb->ip_summed;
1105 /* Looks stupid, but our code really uses when of
1106 * skbs, which it never sent before. --ANK
1108 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1109 buff->tstamp = skb->tstamp;
1111 old_factor = tcp_skb_pcount(skb);
1113 /* Fix up tso_factor for both original and new SKB. */
1114 tcp_set_skb_tso_segs(sk, skb, mss_now);
1115 tcp_set_skb_tso_segs(sk, buff, mss_now);
1117 /* If this packet has been sent out already, we must
1118 * adjust the various packet counters.
1120 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1121 int diff = old_factor - tcp_skb_pcount(skb) -
1122 tcp_skb_pcount(buff);
1125 tcp_adjust_pcount(sk, skb, diff);
1128 /* Link BUFF into the send queue. */
1129 skb_header_release(buff);
1130 tcp_insert_write_queue_after(skb, buff, sk);
1135 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1136 * eventually). The difference is that pulled data not copied, but
1137 * immediately discarded.
1139 static void __pskb_trim_head(struct sk_buff *skb, int len)
1143 eat = min_t(int, len, skb_headlen(skb));
1145 __skb_pull(skb, eat);
1152 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1153 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1156 skb_frag_unref(skb, i);
1159 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1161 skb_shinfo(skb)->frags[k].page_offset += eat;
1162 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1168 skb_shinfo(skb)->nr_frags = k;
1170 skb_reset_tail_pointer(skb);
1171 skb->data_len -= len;
1172 skb->len = skb->data_len;
1175 /* Remove acked data from a packet in the transmit queue. */
1176 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1178 if (skb_unclone(skb, GFP_ATOMIC))
1181 __pskb_trim_head(skb, len);
1183 TCP_SKB_CB(skb)->seq += len;
1184 skb->ip_summed = CHECKSUM_PARTIAL;
1186 skb->truesize -= len;
1187 sk->sk_wmem_queued -= len;
1188 sk_mem_uncharge(sk, len);
1189 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1191 /* Any change of skb->len requires recalculation of tso factor. */
1192 if (tcp_skb_pcount(skb) > 1)
1193 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1198 /* Calculate MSS not accounting any TCP options. */
1199 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1201 const struct tcp_sock *tp = tcp_sk(sk);
1202 const struct inet_connection_sock *icsk = inet_csk(sk);
1205 /* Calculate base mss without TCP options:
1206 It is MMS_S - sizeof(tcphdr) of rfc1122
1208 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1210 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1211 if (icsk->icsk_af_ops->net_frag_header_len) {
1212 const struct dst_entry *dst = __sk_dst_get(sk);
1214 if (dst && dst_allfrag(dst))
1215 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1218 /* Clamp it (mss_clamp does not include tcp options) */
1219 if (mss_now > tp->rx_opt.mss_clamp)
1220 mss_now = tp->rx_opt.mss_clamp;
1222 /* Now subtract optional transport overhead */
1223 mss_now -= icsk->icsk_ext_hdr_len;
1225 /* Then reserve room for full set of TCP options and 8 bytes of data */
1231 /* Calculate MSS. Not accounting for SACKs here. */
1232 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1234 /* Subtract TCP options size, not including SACKs */
1235 return __tcp_mtu_to_mss(sk, pmtu) -
1236 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1239 /* Inverse of above */
1240 int tcp_mss_to_mtu(struct sock *sk, int mss)
1242 const struct tcp_sock *tp = tcp_sk(sk);
1243 const struct inet_connection_sock *icsk = inet_csk(sk);
1247 tp->tcp_header_len +
1248 icsk->icsk_ext_hdr_len +
1249 icsk->icsk_af_ops->net_header_len;
1251 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1252 if (icsk->icsk_af_ops->net_frag_header_len) {
1253 const struct dst_entry *dst = __sk_dst_get(sk);
1255 if (dst && dst_allfrag(dst))
1256 mtu += icsk->icsk_af_ops->net_frag_header_len;
1261 /* MTU probing init per socket */
1262 void tcp_mtup_init(struct sock *sk)
1264 struct tcp_sock *tp = tcp_sk(sk);
1265 struct inet_connection_sock *icsk = inet_csk(sk);
1267 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1268 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1269 icsk->icsk_af_ops->net_header_len;
1270 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1271 icsk->icsk_mtup.probe_size = 0;
1273 EXPORT_SYMBOL(tcp_mtup_init);
1275 /* This function synchronize snd mss to current pmtu/exthdr set.
1277 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1278 for TCP options, but includes only bare TCP header.
1280 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1281 It is minimum of user_mss and mss received with SYN.
1282 It also does not include TCP options.
1284 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1286 tp->mss_cache is current effective sending mss, including
1287 all tcp options except for SACKs. It is evaluated,
1288 taking into account current pmtu, but never exceeds
1289 tp->rx_opt.mss_clamp.
1291 NOTE1. rfc1122 clearly states that advertised MSS
1292 DOES NOT include either tcp or ip options.
1294 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1295 are READ ONLY outside this function. --ANK (980731)
1297 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1299 struct tcp_sock *tp = tcp_sk(sk);
1300 struct inet_connection_sock *icsk = inet_csk(sk);
1303 if (icsk->icsk_mtup.search_high > pmtu)
1304 icsk->icsk_mtup.search_high = pmtu;
1306 mss_now = tcp_mtu_to_mss(sk, pmtu);
1307 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1309 /* And store cached results */
1310 icsk->icsk_pmtu_cookie = pmtu;
1311 if (icsk->icsk_mtup.enabled)
1312 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1313 tp->mss_cache = mss_now;
1317 EXPORT_SYMBOL(tcp_sync_mss);
1319 /* Compute the current effective MSS, taking SACKs and IP options,
1320 * and even PMTU discovery events into account.
1322 unsigned int tcp_current_mss(struct sock *sk)
1324 const struct tcp_sock *tp = tcp_sk(sk);
1325 const struct dst_entry *dst = __sk_dst_get(sk);
1327 unsigned int header_len;
1328 struct tcp_out_options opts;
1329 struct tcp_md5sig_key *md5;
1331 mss_now = tp->mss_cache;
1334 u32 mtu = dst_mtu(dst);
1335 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1336 mss_now = tcp_sync_mss(sk, mtu);
1339 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1340 sizeof(struct tcphdr);
1341 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1342 * some common options. If this is an odd packet (because we have SACK
1343 * blocks etc) then our calculated header_len will be different, and
1344 * we have to adjust mss_now correspondingly */
1345 if (header_len != tp->tcp_header_len) {
1346 int delta = (int) header_len - tp->tcp_header_len;
1353 /* Congestion window validation. (RFC2861) */
1354 static void tcp_cwnd_validate(struct sock *sk)
1356 struct tcp_sock *tp = tcp_sk(sk);
1358 if (tp->packets_out >= tp->snd_cwnd) {
1359 /* Network is feed fully. */
1360 tp->snd_cwnd_used = 0;
1361 tp->snd_cwnd_stamp = tcp_time_stamp;
1363 /* Network starves. */
1364 if (tp->packets_out > tp->snd_cwnd_used)
1365 tp->snd_cwnd_used = tp->packets_out;
1367 if (sysctl_tcp_slow_start_after_idle &&
1368 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1369 tcp_cwnd_application_limited(sk);
1373 /* Returns the portion of skb which can be sent right away without
1374 * introducing MSS oddities to segment boundaries. In rare cases where
1375 * mss_now != mss_cache, we will request caller to create a small skb
1376 * per input skb which could be mostly avoided here (if desired).
1378 * We explicitly want to create a request for splitting write queue tail
1379 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1380 * thus all the complexity (cwnd_len is always MSS multiple which we
1381 * return whenever allowed by the other factors). Basically we need the
1382 * modulo only when the receiver window alone is the limiting factor or
1383 * when we would be allowed to send the split-due-to-Nagle skb fully.
1385 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1386 unsigned int mss_now, unsigned int max_segs)
1388 const struct tcp_sock *tp = tcp_sk(sk);
1389 u32 needed, window, max_len;
1391 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1392 max_len = mss_now * max_segs;
1394 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1397 needed = min(skb->len, window);
1399 if (max_len <= needed)
1402 return needed - needed % mss_now;
1405 /* Can at least one segment of SKB be sent right now, according to the
1406 * congestion window rules? If so, return how many segments are allowed.
1408 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1409 const struct sk_buff *skb)
1411 u32 in_flight, cwnd;
1413 /* Don't be strict about the congestion window for the final FIN. */
1414 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1415 tcp_skb_pcount(skb) == 1)
1418 in_flight = tcp_packets_in_flight(tp);
1419 cwnd = tp->snd_cwnd;
1420 if (in_flight < cwnd)
1421 return (cwnd - in_flight);
1426 /* Initialize TSO state of a skb.
1427 * This must be invoked the first time we consider transmitting
1428 * SKB onto the wire.
1430 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1431 unsigned int mss_now)
1433 int tso_segs = tcp_skb_pcount(skb);
1435 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1436 tcp_set_skb_tso_segs(sk, skb, mss_now);
1437 tso_segs = tcp_skb_pcount(skb);
1442 /* Minshall's variant of the Nagle send check. */
1443 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1445 return after(tp->snd_sml, tp->snd_una) &&
1446 !after(tp->snd_sml, tp->snd_nxt);
1449 /* Return false, if packet can be sent now without violation Nagle's rules:
1450 * 1. It is full sized.
1451 * 2. Or it contains FIN. (already checked by caller)
1452 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1453 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1454 * With Minshall's modification: all sent small packets are ACKed.
1456 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1457 const struct sk_buff *skb,
1458 unsigned int mss_now, int nonagle)
1460 return skb->len < mss_now &&
1461 ((nonagle & TCP_NAGLE_CORK) ||
1462 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1465 /* Return true if the Nagle test allows this packet to be
1468 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1469 unsigned int cur_mss, int nonagle)
1471 /* Nagle rule does not apply to frames, which sit in the middle of the
1472 * write_queue (they have no chances to get new data).
1474 * This is implemented in the callers, where they modify the 'nonagle'
1475 * argument based upon the location of SKB in the send queue.
1477 if (nonagle & TCP_NAGLE_PUSH)
1480 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1481 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1484 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1490 /* Does at least the first segment of SKB fit into the send window? */
1491 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1492 const struct sk_buff *skb,
1493 unsigned int cur_mss)
1495 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1497 if (skb->len > cur_mss)
1498 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1500 return !after(end_seq, tcp_wnd_end(tp));
1503 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1504 * should be put on the wire right now. If so, it returns the number of
1505 * packets allowed by the congestion window.
1507 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1508 unsigned int cur_mss, int nonagle)
1510 const struct tcp_sock *tp = tcp_sk(sk);
1511 unsigned int cwnd_quota;
1513 tcp_init_tso_segs(sk, skb, cur_mss);
1515 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1518 cwnd_quota = tcp_cwnd_test(tp, skb);
1519 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1525 /* Test if sending is allowed right now. */
1526 bool tcp_may_send_now(struct sock *sk)
1528 const struct tcp_sock *tp = tcp_sk(sk);
1529 struct sk_buff *skb = tcp_send_head(sk);
1532 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1533 (tcp_skb_is_last(sk, skb) ?
1534 tp->nonagle : TCP_NAGLE_PUSH));
1537 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1538 * which is put after SKB on the list. It is very much like
1539 * tcp_fragment() except that it may make several kinds of assumptions
1540 * in order to speed up the splitting operation. In particular, we
1541 * know that all the data is in scatter-gather pages, and that the
1542 * packet has never been sent out before (and thus is not cloned).
1544 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1545 unsigned int mss_now, gfp_t gfp)
1547 struct sk_buff *buff;
1548 int nlen = skb->len - len;
1551 /* All of a TSO frame must be composed of paged data. */
1552 if (skb->len != skb->data_len)
1553 return tcp_fragment(sk, skb, len, mss_now);
1555 buff = sk_stream_alloc_skb(sk, 0, gfp);
1556 if (unlikely(buff == NULL))
1559 sk->sk_wmem_queued += buff->truesize;
1560 sk_mem_charge(sk, buff->truesize);
1561 buff->truesize += nlen;
1562 skb->truesize -= nlen;
1564 /* Correct the sequence numbers. */
1565 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1566 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1567 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1569 /* PSH and FIN should only be set in the second packet. */
1570 flags = TCP_SKB_CB(skb)->tcp_flags;
1571 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1572 TCP_SKB_CB(buff)->tcp_flags = flags;
1574 /* This packet was never sent out yet, so no SACK bits. */
1575 TCP_SKB_CB(buff)->sacked = 0;
1577 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1578 skb_split(skb, buff, len);
1580 /* Fix up tso_factor for both original and new SKB. */
1581 tcp_set_skb_tso_segs(sk, skb, mss_now);
1582 tcp_set_skb_tso_segs(sk, buff, mss_now);
1584 /* Link BUFF into the send queue. */
1585 skb_header_release(buff);
1586 tcp_insert_write_queue_after(skb, buff, sk);
1591 /* Try to defer sending, if possible, in order to minimize the amount
1592 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1594 * This algorithm is from John Heffner.
1596 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1598 struct tcp_sock *tp = tcp_sk(sk);
1599 const struct inet_connection_sock *icsk = inet_csk(sk);
1600 u32 send_win, cong_win, limit, in_flight;
1603 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1606 if (icsk->icsk_ca_state != TCP_CA_Open)
1609 /* Defer for less than two clock ticks. */
1610 if (tp->tso_deferred &&
1611 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1614 in_flight = tcp_packets_in_flight(tp);
1616 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1618 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1620 /* From in_flight test above, we know that cwnd > in_flight. */
1621 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1623 limit = min(send_win, cong_win);
1625 /* If a full-sized TSO skb can be sent, do it. */
1626 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1627 tp->xmit_size_goal_segs * tp->mss_cache))
1630 /* Middle in queue won't get any more data, full sendable already? */
1631 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1634 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1636 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1638 /* If at least some fraction of a window is available,
1641 chunk /= win_divisor;
1645 /* Different approach, try not to defer past a single
1646 * ACK. Receiver should ACK every other full sized
1647 * frame, so if we have space for more than 3 frames
1650 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1654 /* Ok, it looks like it is advisable to defer.
1655 * Do not rearm the timer if already set to not break TCP ACK clocking.
1657 if (!tp->tso_deferred)
1658 tp->tso_deferred = 1 | (jiffies << 1);
1663 tp->tso_deferred = 0;
1667 /* Create a new MTU probe if we are ready.
1668 * MTU probe is regularly attempting to increase the path MTU by
1669 * deliberately sending larger packets. This discovers routing
1670 * changes resulting in larger path MTUs.
1672 * Returns 0 if we should wait to probe (no cwnd available),
1673 * 1 if a probe was sent,
1676 static int tcp_mtu_probe(struct sock *sk)
1678 struct tcp_sock *tp = tcp_sk(sk);
1679 struct inet_connection_sock *icsk = inet_csk(sk);
1680 struct sk_buff *skb, *nskb, *next;
1687 /* Not currently probing/verifying,
1689 * have enough cwnd, and
1690 * not SACKing (the variable headers throw things off) */
1691 if (!icsk->icsk_mtup.enabled ||
1692 icsk->icsk_mtup.probe_size ||
1693 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1694 tp->snd_cwnd < 11 ||
1695 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1698 /* Very simple search strategy: just double the MSS. */
1699 mss_now = tcp_current_mss(sk);
1700 probe_size = 2 * tp->mss_cache;
1701 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1702 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1703 /* TODO: set timer for probe_converge_event */
1707 /* Have enough data in the send queue to probe? */
1708 if (tp->write_seq - tp->snd_nxt < size_needed)
1711 if (tp->snd_wnd < size_needed)
1713 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1716 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1717 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1718 if (!tcp_packets_in_flight(tp))
1724 /* We're allowed to probe. Build it now. */
1725 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1727 sk->sk_wmem_queued += nskb->truesize;
1728 sk_mem_charge(sk, nskb->truesize);
1730 skb = tcp_send_head(sk);
1732 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1733 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1734 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1735 TCP_SKB_CB(nskb)->sacked = 0;
1737 nskb->ip_summed = skb->ip_summed;
1739 tcp_insert_write_queue_before(nskb, skb, sk);
1742 tcp_for_write_queue_from_safe(skb, next, sk) {
1743 copy = min_t(int, skb->len, probe_size - len);
1744 if (nskb->ip_summed)
1745 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1747 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1748 skb_put(nskb, copy),
1751 if (skb->len <= copy) {
1752 /* We've eaten all the data from this skb.
1754 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1755 tcp_unlink_write_queue(skb, sk);
1756 sk_wmem_free_skb(sk, skb);
1758 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1759 ~(TCPHDR_FIN|TCPHDR_PSH);
1760 if (!skb_shinfo(skb)->nr_frags) {
1761 skb_pull(skb, copy);
1762 if (skb->ip_summed != CHECKSUM_PARTIAL)
1763 skb->csum = csum_partial(skb->data,
1766 __pskb_trim_head(skb, copy);
1767 tcp_set_skb_tso_segs(sk, skb, mss_now);
1769 TCP_SKB_CB(skb)->seq += copy;
1774 if (len >= probe_size)
1777 tcp_init_tso_segs(sk, nskb, nskb->len);
1779 /* We're ready to send. If this fails, the probe will
1780 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1781 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1782 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1783 /* Decrement cwnd here because we are sending
1784 * effectively two packets. */
1786 tcp_event_new_data_sent(sk, nskb);
1788 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1789 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1790 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1798 /* This routine writes packets to the network. It advances the
1799 * send_head. This happens as incoming acks open up the remote
1802 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1803 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1804 * account rare use of URG, this is not a big flaw.
1806 * Send at most one packet when push_one > 0. Temporarily ignore
1807 * cwnd limit to force at most one packet out when push_one == 2.
1809 * Returns true, if no segments are in flight and we have queued segments,
1810 * but cannot send anything now because of SWS or another problem.
1812 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1813 int push_one, gfp_t gfp)
1815 struct tcp_sock *tp = tcp_sk(sk);
1816 struct sk_buff *skb;
1817 unsigned int tso_segs, sent_pkts;
1824 /* Do MTU probing. */
1825 result = tcp_mtu_probe(sk);
1828 } else if (result > 0) {
1833 while ((skb = tcp_send_head(sk))) {
1836 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1839 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1840 goto repair; /* Skip network transmission */
1842 cwnd_quota = tcp_cwnd_test(tp, skb);
1845 /* Force out a loss probe pkt. */
1851 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1854 if (tso_segs == 1) {
1855 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1856 (tcp_skb_is_last(sk, skb) ?
1857 nonagle : TCP_NAGLE_PUSH))))
1860 if (!push_one && tcp_tso_should_defer(sk, skb))
1864 /* TCP Small Queues :
1865 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1867 * - better RTT estimation and ACK scheduling
1870 * Alas, some drivers / subsystems require a fair amount
1871 * of queued bytes to ensure line rate.
1872 * One example is wifi aggregation (802.11 AMPDU)
1874 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1875 sk->sk_pacing_rate >> 10);
1877 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1878 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1879 /* It is possible TX completion already happened
1880 * before we set TSQ_THROTTLED, so we must
1881 * test again the condition.
1882 * We abuse smp_mb__after_clear_bit() because
1883 * there is no smp_mb__after_set_bit() yet
1885 smp_mb__after_clear_bit();
1886 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1891 if (tso_segs > 1 && !tcp_urg_mode(tp))
1892 limit = tcp_mss_split_point(sk, skb, mss_now,
1895 sk->sk_gso_max_segs));
1897 if (skb->len > limit &&
1898 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1901 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1903 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1907 /* Advance the send_head. This one is sent out.
1908 * This call will increment packets_out.
1910 tcp_event_new_data_sent(sk, skb);
1912 tcp_minshall_update(tp, mss_now, skb);
1913 sent_pkts += tcp_skb_pcount(skb);
1919 if (likely(sent_pkts)) {
1920 if (tcp_in_cwnd_reduction(sk))
1921 tp->prr_out += sent_pkts;
1923 /* Send one loss probe per tail loss episode. */
1925 tcp_schedule_loss_probe(sk);
1926 tcp_cwnd_validate(sk);
1929 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1932 bool tcp_schedule_loss_probe(struct sock *sk)
1934 struct inet_connection_sock *icsk = inet_csk(sk);
1935 struct tcp_sock *tp = tcp_sk(sk);
1936 u32 timeout, tlp_time_stamp, rto_time_stamp;
1937 u32 rtt = tp->srtt >> 3;
1939 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1941 /* No consecutive loss probes. */
1942 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1946 /* Don't do any loss probe on a Fast Open connection before 3WHS
1949 if (sk->sk_state == TCP_SYN_RECV)
1952 /* TLP is only scheduled when next timer event is RTO. */
1953 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1956 /* Schedule a loss probe in 2*RTT for SACK capable connections
1957 * in Open state, that are either limited by cwnd or application.
1959 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1960 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1963 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1967 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1968 * for delayed ack when there's one outstanding packet.
1971 if (tp->packets_out == 1)
1972 timeout = max_t(u32, timeout,
1973 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1974 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1976 /* If RTO is shorter, just schedule TLP in its place. */
1977 tlp_time_stamp = tcp_time_stamp + timeout;
1978 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1979 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1980 s32 delta = rto_time_stamp - tcp_time_stamp;
1985 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1990 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1991 * retransmit the last segment.
1993 void tcp_send_loss_probe(struct sock *sk)
1995 struct tcp_sock *tp = tcp_sk(sk);
1996 struct sk_buff *skb;
1998 int mss = tcp_current_mss(sk);
2001 if (tcp_send_head(sk) != NULL) {
2002 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2006 /* At most one outstanding TLP retransmission. */
2007 if (tp->tlp_high_seq)
2010 /* Retransmit last segment. */
2011 skb = tcp_write_queue_tail(sk);
2015 pcount = tcp_skb_pcount(skb);
2016 if (WARN_ON(!pcount))
2019 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2020 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2022 skb = tcp_write_queue_tail(sk);
2025 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2028 /* Probe with zero data doesn't trigger fast recovery. */
2030 err = __tcp_retransmit_skb(sk, skb);
2032 /* Record snd_nxt for loss detection. */
2034 tp->tlp_high_seq = tp->snd_nxt;
2037 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2038 inet_csk(sk)->icsk_rto,
2042 NET_INC_STATS_BH(sock_net(sk),
2043 LINUX_MIB_TCPLOSSPROBES);
2047 /* Push out any pending frames which were held back due to
2048 * TCP_CORK or attempt at coalescing tiny packets.
2049 * The socket must be locked by the caller.
2051 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2054 /* If we are closed, the bytes will have to remain here.
2055 * In time closedown will finish, we empty the write queue and
2056 * all will be happy.
2058 if (unlikely(sk->sk_state == TCP_CLOSE))
2061 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2062 sk_gfp_atomic(sk, GFP_ATOMIC)))
2063 tcp_check_probe_timer(sk);
2066 /* Send _single_ skb sitting at the send head. This function requires
2067 * true push pending frames to setup probe timer etc.
2069 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2071 struct sk_buff *skb = tcp_send_head(sk);
2073 BUG_ON(!skb || skb->len < mss_now);
2075 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2078 /* This function returns the amount that we can raise the
2079 * usable window based on the following constraints
2081 * 1. The window can never be shrunk once it is offered (RFC 793)
2082 * 2. We limit memory per socket
2085 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2086 * RECV.NEXT + RCV.WIN fixed until:
2087 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2089 * i.e. don't raise the right edge of the window until you can raise
2090 * it at least MSS bytes.
2092 * Unfortunately, the recommended algorithm breaks header prediction,
2093 * since header prediction assumes th->window stays fixed.
2095 * Strictly speaking, keeping th->window fixed violates the receiver
2096 * side SWS prevention criteria. The problem is that under this rule
2097 * a stream of single byte packets will cause the right side of the
2098 * window to always advance by a single byte.
2100 * Of course, if the sender implements sender side SWS prevention
2101 * then this will not be a problem.
2103 * BSD seems to make the following compromise:
2105 * If the free space is less than the 1/4 of the maximum
2106 * space available and the free space is less than 1/2 mss,
2107 * then set the window to 0.
2108 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2109 * Otherwise, just prevent the window from shrinking
2110 * and from being larger than the largest representable value.
2112 * This prevents incremental opening of the window in the regime
2113 * where TCP is limited by the speed of the reader side taking
2114 * data out of the TCP receive queue. It does nothing about
2115 * those cases where the window is constrained on the sender side
2116 * because the pipeline is full.
2118 * BSD also seems to "accidentally" limit itself to windows that are a
2119 * multiple of MSS, at least until the free space gets quite small.
2120 * This would appear to be a side effect of the mbuf implementation.
2121 * Combining these two algorithms results in the observed behavior
2122 * of having a fixed window size at almost all times.
2124 * Below we obtain similar behavior by forcing the offered window to
2125 * a multiple of the mss when it is feasible to do so.
2127 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2128 * Regular options like TIMESTAMP are taken into account.
2130 u32 __tcp_select_window(struct sock *sk)
2132 struct inet_connection_sock *icsk = inet_csk(sk);
2133 struct tcp_sock *tp = tcp_sk(sk);
2134 /* MSS for the peer's data. Previous versions used mss_clamp
2135 * here. I don't know if the value based on our guesses
2136 * of peer's MSS is better for the performance. It's more correct
2137 * but may be worse for the performance because of rcv_mss
2138 * fluctuations. --SAW 1998/11/1
2140 int mss = icsk->icsk_ack.rcv_mss;
2141 int free_space = tcp_space(sk);
2142 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2145 if (mss > full_space)
2148 if (free_space < (full_space >> 1)) {
2149 icsk->icsk_ack.quick = 0;
2151 if (sk_under_memory_pressure(sk))
2152 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2155 if (free_space < mss)
2159 if (free_space > tp->rcv_ssthresh)
2160 free_space = tp->rcv_ssthresh;
2162 /* Don't do rounding if we are using window scaling, since the
2163 * scaled window will not line up with the MSS boundary anyway.
2165 window = tp->rcv_wnd;
2166 if (tp->rx_opt.rcv_wscale) {
2167 window = free_space;
2169 /* Advertise enough space so that it won't get scaled away.
2170 * Import case: prevent zero window announcement if
2171 * 1<<rcv_wscale > mss.
2173 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2174 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2175 << tp->rx_opt.rcv_wscale);
2177 /* Get the largest window that is a nice multiple of mss.
2178 * Window clamp already applied above.
2179 * If our current window offering is within 1 mss of the
2180 * free space we just keep it. This prevents the divide
2181 * and multiply from happening most of the time.
2182 * We also don't do any window rounding when the free space
2185 if (window <= free_space - mss || window > free_space)
2186 window = (free_space / mss) * mss;
2187 else if (mss == full_space &&
2188 free_space > window + (full_space >> 1))
2189 window = free_space;
2195 /* Collapses two adjacent SKB's during retransmission. */
2196 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2198 struct tcp_sock *tp = tcp_sk(sk);
2199 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2200 int skb_size, next_skb_size;
2202 skb_size = skb->len;
2203 next_skb_size = next_skb->len;
2205 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2207 tcp_highest_sack_combine(sk, next_skb, skb);
2209 tcp_unlink_write_queue(next_skb, sk);
2211 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2214 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2215 skb->ip_summed = CHECKSUM_PARTIAL;
2217 if (skb->ip_summed != CHECKSUM_PARTIAL)
2218 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2220 /* Update sequence range on original skb. */
2221 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2223 /* Merge over control information. This moves PSH/FIN etc. over */
2224 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2226 /* All done, get rid of second SKB and account for it so
2227 * packet counting does not break.
2229 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2231 /* changed transmit queue under us so clear hints */
2232 tcp_clear_retrans_hints_partial(tp);
2233 if (next_skb == tp->retransmit_skb_hint)
2234 tp->retransmit_skb_hint = skb;
2236 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2238 sk_wmem_free_skb(sk, next_skb);
2241 /* Check if coalescing SKBs is legal. */
2242 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2244 if (tcp_skb_pcount(skb) > 1)
2246 /* TODO: SACK collapsing could be used to remove this condition */
2247 if (skb_shinfo(skb)->nr_frags != 0)
2249 if (skb_cloned(skb))
2251 if (skb == tcp_send_head(sk))
2253 /* Some heurestics for collapsing over SACK'd could be invented */
2254 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2260 /* Collapse packets in the retransmit queue to make to create
2261 * less packets on the wire. This is only done on retransmission.
2263 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2266 struct tcp_sock *tp = tcp_sk(sk);
2267 struct sk_buff *skb = to, *tmp;
2270 if (!sysctl_tcp_retrans_collapse)
2272 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2275 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2276 if (!tcp_can_collapse(sk, skb))
2288 /* Punt if not enough space exists in the first SKB for
2289 * the data in the second
2291 if (skb->len > skb_availroom(to))
2294 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2297 tcp_collapse_retrans(sk, to);
2301 /* This retransmits one SKB. Policy decisions and retransmit queue
2302 * state updates are done by the caller. Returns non-zero if an
2303 * error occurred which prevented the send.
2305 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2307 struct tcp_sock *tp = tcp_sk(sk);
2308 struct inet_connection_sock *icsk = inet_csk(sk);
2309 unsigned int cur_mss;
2311 /* Inconslusive MTU probe */
2312 if (icsk->icsk_mtup.probe_size) {
2313 icsk->icsk_mtup.probe_size = 0;
2316 /* Do not sent more than we queued. 1/4 is reserved for possible
2317 * copying overhead: fragmentation, tunneling, mangling etc.
2319 if (atomic_read(&sk->sk_wmem_alloc) >
2320 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2323 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2324 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2326 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2330 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2331 return -EHOSTUNREACH; /* Routing failure or similar. */
2333 cur_mss = tcp_current_mss(sk);
2335 /* If receiver has shrunk his window, and skb is out of
2336 * new window, do not retransmit it. The exception is the
2337 * case, when window is shrunk to zero. In this case
2338 * our retransmit serves as a zero window probe.
2340 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2341 TCP_SKB_CB(skb)->seq != tp->snd_una)
2344 if (skb->len > cur_mss) {
2345 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2346 return -ENOMEM; /* We'll try again later. */
2348 int oldpcount = tcp_skb_pcount(skb);
2350 if (unlikely(oldpcount > 1)) {
2351 if (skb_unclone(skb, GFP_ATOMIC))
2353 tcp_init_tso_segs(sk, skb, cur_mss);
2354 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2358 tcp_retrans_try_collapse(sk, skb, cur_mss);
2360 /* Some Solaris stacks overoptimize and ignore the FIN on a
2361 * retransmit when old data is attached. So strip it off
2362 * since it is cheap to do so and saves bytes on the network.
2365 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2366 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2367 if (!pskb_trim(skb, 0)) {
2368 /* Reuse, even though it does some unnecessary work */
2369 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2370 TCP_SKB_CB(skb)->tcp_flags);
2371 skb->ip_summed = CHECKSUM_NONE;
2375 /* Make a copy, if the first transmission SKB clone we made
2376 * is still in somebody's hands, else make a clone.
2378 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2380 /* make sure skb->data is aligned on arches that require it
2381 * and check if ack-trimming & collapsing extended the headroom
2382 * beyond what csum_start can cover.
2384 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2385 skb_headroom(skb) >= 0xFFFF)) {
2386 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2388 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2391 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2395 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2397 struct tcp_sock *tp = tcp_sk(sk);
2398 int err = __tcp_retransmit_skb(sk, skb);
2401 /* Update global TCP statistics. */
2402 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2404 tp->total_retrans++;
2406 #if FASTRETRANS_DEBUG > 0
2407 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2408 net_dbg_ratelimited("retrans_out leaked\n");
2411 if (!tp->retrans_out)
2412 tp->lost_retrans_low = tp->snd_nxt;
2413 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2414 tp->retrans_out += tcp_skb_pcount(skb);
2416 /* Save stamp of the first retransmit. */
2417 if (!tp->retrans_stamp)
2418 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2420 tp->undo_retrans += tcp_skb_pcount(skb);
2422 /* snd_nxt is stored to detect loss of retransmitted segment,
2423 * see tcp_input.c tcp_sacktag_write_queue().
2425 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2430 /* Check if we forward retransmits are possible in the current
2431 * window/congestion state.
2433 static bool tcp_can_forward_retransmit(struct sock *sk)
2435 const struct inet_connection_sock *icsk = inet_csk(sk);
2436 const struct tcp_sock *tp = tcp_sk(sk);
2438 /* Forward retransmissions are possible only during Recovery. */
2439 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2442 /* No forward retransmissions in Reno are possible. */
2443 if (tcp_is_reno(tp))
2446 /* Yeah, we have to make difficult choice between forward transmission
2447 * and retransmission... Both ways have their merits...
2449 * For now we do not retransmit anything, while we have some new
2450 * segments to send. In the other cases, follow rule 3 for
2451 * NextSeg() specified in RFC3517.
2454 if (tcp_may_send_now(sk))
2460 /* This gets called after a retransmit timeout, and the initially
2461 * retransmitted data is acknowledged. It tries to continue
2462 * resending the rest of the retransmit queue, until either
2463 * we've sent it all or the congestion window limit is reached.
2464 * If doing SACK, the first ACK which comes back for a timeout
2465 * based retransmit packet might feed us FACK information again.
2466 * If so, we use it to avoid unnecessarily retransmissions.
2468 void tcp_xmit_retransmit_queue(struct sock *sk)
2470 const struct inet_connection_sock *icsk = inet_csk(sk);
2471 struct tcp_sock *tp = tcp_sk(sk);
2472 struct sk_buff *skb;
2473 struct sk_buff *hole = NULL;
2476 int fwd_rexmitting = 0;
2478 if (!tp->packets_out)
2482 tp->retransmit_high = tp->snd_una;
2484 if (tp->retransmit_skb_hint) {
2485 skb = tp->retransmit_skb_hint;
2486 last_lost = TCP_SKB_CB(skb)->end_seq;
2487 if (after(last_lost, tp->retransmit_high))
2488 last_lost = tp->retransmit_high;
2490 skb = tcp_write_queue_head(sk);
2491 last_lost = tp->snd_una;
2494 tcp_for_write_queue_from(skb, sk) {
2495 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2497 if (skb == tcp_send_head(sk))
2499 /* we could do better than to assign each time */
2501 tp->retransmit_skb_hint = skb;
2503 /* Assume this retransmit will generate
2504 * only one packet for congestion window
2505 * calculation purposes. This works because
2506 * tcp_retransmit_skb() will chop up the
2507 * packet to be MSS sized and all the
2508 * packet counting works out.
2510 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2513 if (fwd_rexmitting) {
2515 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2517 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2519 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2520 tp->retransmit_high = last_lost;
2521 if (!tcp_can_forward_retransmit(sk))
2523 /* Backtrack if necessary to non-L'ed skb */
2531 } else if (!(sacked & TCPCB_LOST)) {
2532 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2537 last_lost = TCP_SKB_CB(skb)->end_seq;
2538 if (icsk->icsk_ca_state != TCP_CA_Loss)
2539 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2541 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2544 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2547 if (tcp_retransmit_skb(sk, skb)) {
2548 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2551 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2553 if (tcp_in_cwnd_reduction(sk))
2554 tp->prr_out += tcp_skb_pcount(skb);
2556 if (skb == tcp_write_queue_head(sk))
2557 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2558 inet_csk(sk)->icsk_rto,
2563 /* Send a fin. The caller locks the socket for us. This cannot be
2564 * allowed to fail queueing a FIN frame under any circumstances.
2566 void tcp_send_fin(struct sock *sk)
2568 struct tcp_sock *tp = tcp_sk(sk);
2569 struct sk_buff *skb = tcp_write_queue_tail(sk);
2572 /* Optimization, tack on the FIN if we have a queue of
2573 * unsent frames. But be careful about outgoing SACKS
2576 mss_now = tcp_current_mss(sk);
2578 if (tcp_send_head(sk) != NULL) {
2579 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2580 TCP_SKB_CB(skb)->end_seq++;
2583 /* Socket is locked, keep trying until memory is available. */
2585 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2592 /* Reserve space for headers and prepare control bits. */
2593 skb_reserve(skb, MAX_TCP_HEADER);
2594 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2595 tcp_init_nondata_skb(skb, tp->write_seq,
2596 TCPHDR_ACK | TCPHDR_FIN);
2597 tcp_queue_skb(sk, skb);
2599 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2602 /* We get here when a process closes a file descriptor (either due to
2603 * an explicit close() or as a byproduct of exit()'ing) and there
2604 * was unread data in the receive queue. This behavior is recommended
2605 * by RFC 2525, section 2.17. -DaveM
2607 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2609 struct sk_buff *skb;
2611 /* NOTE: No TCP options attached and we never retransmit this. */
2612 skb = alloc_skb(MAX_TCP_HEADER, priority);
2614 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2618 /* Reserve space for headers and prepare control bits. */
2619 skb_reserve(skb, MAX_TCP_HEADER);
2620 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2621 TCPHDR_ACK | TCPHDR_RST);
2623 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2624 if (tcp_transmit_skb(sk, skb, 0, priority))
2625 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2627 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2630 /* Send a crossed SYN-ACK during socket establishment.
2631 * WARNING: This routine must only be called when we have already sent
2632 * a SYN packet that crossed the incoming SYN that caused this routine
2633 * to get called. If this assumption fails then the initial rcv_wnd
2634 * and rcv_wscale values will not be correct.
2636 int tcp_send_synack(struct sock *sk)
2638 struct sk_buff *skb;
2640 skb = tcp_write_queue_head(sk);
2641 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2642 pr_debug("%s: wrong queue state\n", __func__);
2645 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2646 if (skb_cloned(skb)) {
2647 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2650 tcp_unlink_write_queue(skb, sk);
2651 skb_header_release(nskb);
2652 __tcp_add_write_queue_head(sk, nskb);
2653 sk_wmem_free_skb(sk, skb);
2654 sk->sk_wmem_queued += nskb->truesize;
2655 sk_mem_charge(sk, nskb->truesize);
2659 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2660 TCP_ECN_send_synack(tcp_sk(sk), skb);
2662 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2663 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2667 * tcp_make_synack - Prepare a SYN-ACK.
2668 * sk: listener socket
2669 * dst: dst entry attached to the SYNACK
2670 * req: request_sock pointer
2672 * Allocate one skb and build a SYNACK packet.
2673 * @dst is consumed : Caller should not use it again.
2675 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2676 struct request_sock *req,
2677 struct tcp_fastopen_cookie *foc)
2679 struct tcp_out_options opts;
2680 struct inet_request_sock *ireq = inet_rsk(req);
2681 struct tcp_sock *tp = tcp_sk(sk);
2683 struct sk_buff *skb;
2684 struct tcp_md5sig_key *md5;
2685 int tcp_header_size;
2688 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2689 if (unlikely(!skb)) {
2693 /* Reserve space for headers. */
2694 skb_reserve(skb, MAX_TCP_HEADER);
2696 skb_dst_set(skb, dst);
2697 security_skb_owned_by(skb, sk);
2699 mss = dst_metric_advmss(dst);
2700 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2701 mss = tp->rx_opt.user_mss;
2703 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2705 /* Set this up on the first call only */
2706 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2708 /* limit the window selection if the user enforce a smaller rx buffer */
2709 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2710 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2711 req->window_clamp = tcp_full_space(sk);
2713 /* tcp_full_space because it is guaranteed to be the first packet */
2714 tcp_select_initial_window(tcp_full_space(sk),
2715 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2720 dst_metric(dst, RTAX_INITRWND));
2721 ireq->rcv_wscale = rcv_wscale;
2724 memset(&opts, 0, sizeof(opts));
2725 #ifdef CONFIG_SYN_COOKIES
2726 if (unlikely(req->cookie_ts))
2727 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2730 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2731 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2734 skb_push(skb, tcp_header_size);
2735 skb_reset_transport_header(skb);
2738 memset(th, 0, sizeof(struct tcphdr));
2741 TCP_ECN_make_synack(req, th);
2742 th->source = ireq->loc_port;
2743 th->dest = ireq->rmt_port;
2744 /* Setting of flags are superfluous here for callers (and ECE is
2745 * not even correctly set)
2747 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2748 TCPHDR_SYN | TCPHDR_ACK);
2750 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2751 /* XXX data is queued and acked as is. No buffer/window check */
2752 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2754 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2755 th->window = htons(min(req->rcv_wnd, 65535U));
2756 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2757 th->doff = (tcp_header_size >> 2);
2758 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2760 #ifdef CONFIG_TCP_MD5SIG
2761 /* Okay, we have all we need - do the md5 hash if needed */
2763 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2764 md5, NULL, req, skb);
2770 EXPORT_SYMBOL(tcp_make_synack);
2772 /* Do all connect socket setups that can be done AF independent. */
2773 void tcp_connect_init(struct sock *sk)
2775 const struct dst_entry *dst = __sk_dst_get(sk);
2776 struct tcp_sock *tp = tcp_sk(sk);
2779 /* We'll fix this up when we get a response from the other end.
2780 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2782 tp->tcp_header_len = sizeof(struct tcphdr) +
2783 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2785 #ifdef CONFIG_TCP_MD5SIG
2786 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2787 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2790 /* If user gave his TCP_MAXSEG, record it to clamp */
2791 if (tp->rx_opt.user_mss)
2792 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2795 tcp_sync_mss(sk, dst_mtu(dst));
2797 if (!tp->window_clamp)
2798 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2799 tp->advmss = dst_metric_advmss(dst);
2800 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2801 tp->advmss = tp->rx_opt.user_mss;
2803 tcp_initialize_rcv_mss(sk);
2805 /* limit the window selection if the user enforce a smaller rx buffer */
2806 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2807 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2808 tp->window_clamp = tcp_full_space(sk);
2810 tcp_select_initial_window(tcp_full_space(sk),
2811 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2814 sysctl_tcp_window_scaling,
2816 dst_metric(dst, RTAX_INITRWND));
2818 tp->rx_opt.rcv_wscale = rcv_wscale;
2819 tp->rcv_ssthresh = tp->rcv_wnd;
2822 sock_reset_flag(sk, SOCK_DONE);
2825 tp->snd_una = tp->write_seq;
2826 tp->snd_sml = tp->write_seq;
2827 tp->snd_up = tp->write_seq;
2828 tp->snd_nxt = tp->write_seq;
2830 if (likely(!tp->repair))
2833 tp->rcv_tstamp = tcp_time_stamp;
2834 tp->rcv_wup = tp->rcv_nxt;
2835 tp->copied_seq = tp->rcv_nxt;
2837 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2838 inet_csk(sk)->icsk_retransmits = 0;
2839 tcp_clear_retrans(tp);
2842 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2844 struct tcp_sock *tp = tcp_sk(sk);
2845 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2847 tcb->end_seq += skb->len;
2848 skb_header_release(skb);
2849 __tcp_add_write_queue_tail(sk, skb);
2850 sk->sk_wmem_queued += skb->truesize;
2851 sk_mem_charge(sk, skb->truesize);
2852 tp->write_seq = tcb->end_seq;
2853 tp->packets_out += tcp_skb_pcount(skb);
2856 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2857 * queue a data-only packet after the regular SYN, such that regular SYNs
2858 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2859 * only the SYN sequence, the data are retransmitted in the first ACK.
2860 * If cookie is not cached or other error occurs, falls back to send a
2861 * regular SYN with Fast Open cookie request option.
2863 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2865 struct tcp_sock *tp = tcp_sk(sk);
2866 struct tcp_fastopen_request *fo = tp->fastopen_req;
2867 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2868 struct sk_buff *syn_data = NULL, *data;
2869 unsigned long last_syn_loss = 0;
2871 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2872 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2873 &syn_loss, &last_syn_loss);
2874 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2876 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2877 fo->cookie.len = -1;
2881 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2882 fo->cookie.len = -1;
2883 else if (fo->cookie.len <= 0)
2886 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2887 * user-MSS. Reserve maximum option space for middleboxes that add
2888 * private TCP options. The cost is reduced data space in SYN :(
2890 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2891 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2892 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2893 MAX_TCP_OPTION_SPACE;
2895 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2897 if (syn_data == NULL)
2900 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2901 struct iovec *iov = &fo->data->msg_iov[i];
2902 unsigned char __user *from = iov->iov_base;
2903 int len = iov->iov_len;
2905 if (syn_data->len + len > space)
2906 len = space - syn_data->len;
2907 else if (i + 1 == iovlen)
2908 /* No more data pending in inet_wait_for_connect() */
2911 if (skb_add_data(syn_data, from, len))
2915 /* Queue a data-only packet after the regular SYN for retransmission */
2916 data = pskb_copy(syn_data, sk->sk_allocation);
2919 TCP_SKB_CB(data)->seq++;
2920 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2921 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2922 tcp_connect_queue_skb(sk, data);
2923 fo->copied = data->len;
2925 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2926 tp->syn_data = (fo->copied > 0);
2927 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2933 /* Send a regular SYN with Fast Open cookie request option */
2934 if (fo->cookie.len > 0)
2936 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2938 tp->syn_fastopen = 0;
2939 kfree_skb(syn_data);
2941 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2945 /* Build a SYN and send it off. */
2946 int tcp_connect(struct sock *sk)
2948 struct tcp_sock *tp = tcp_sk(sk);
2949 struct sk_buff *buff;
2952 tcp_connect_init(sk);
2954 if (unlikely(tp->repair)) {
2955 tcp_finish_connect(sk, NULL);
2959 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2960 if (unlikely(buff == NULL))
2963 /* Reserve space for headers. */
2964 skb_reserve(buff, MAX_TCP_HEADER);
2966 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2967 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2968 tcp_connect_queue_skb(sk, buff);
2969 TCP_ECN_send_syn(sk, buff);
2971 /* Send off SYN; include data in Fast Open. */
2972 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2973 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2974 if (err == -ECONNREFUSED)
2977 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2978 * in order to make this packet get counted in tcpOutSegs.
2980 tp->snd_nxt = tp->write_seq;
2981 tp->pushed_seq = tp->write_seq;
2982 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2984 /* Timer for repeating the SYN until an answer. */
2985 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2986 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2989 EXPORT_SYMBOL(tcp_connect);
2991 /* Send out a delayed ack, the caller does the policy checking
2992 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2995 void tcp_send_delayed_ack(struct sock *sk)
2997 struct inet_connection_sock *icsk = inet_csk(sk);
2998 int ato = icsk->icsk_ack.ato;
2999 unsigned long timeout;
3001 if (ato > TCP_DELACK_MIN) {
3002 const struct tcp_sock *tp = tcp_sk(sk);
3003 int max_ato = HZ / 2;
3005 if (icsk->icsk_ack.pingpong ||
3006 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3007 max_ato = TCP_DELACK_MAX;
3009 /* Slow path, intersegment interval is "high". */
3011 /* If some rtt estimate is known, use it to bound delayed ack.
3012 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3016 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3022 ato = min(ato, max_ato);
3025 /* Stay within the limit we were given */
3026 timeout = jiffies + ato;
3028 /* Use new timeout only if there wasn't a older one earlier. */
3029 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3030 /* If delack timer was blocked or is about to expire,
3033 if (icsk->icsk_ack.blocked ||
3034 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3039 if (!time_before(timeout, icsk->icsk_ack.timeout))
3040 timeout = icsk->icsk_ack.timeout;
3042 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3043 icsk->icsk_ack.timeout = timeout;
3044 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3047 /* This routine sends an ack and also updates the window. */
3048 void tcp_send_ack(struct sock *sk)
3050 struct sk_buff *buff;
3052 /* If we have been reset, we may not send again. */
3053 if (sk->sk_state == TCP_CLOSE)
3056 /* We are not putting this on the write queue, so
3057 * tcp_transmit_skb() will set the ownership to this
3060 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3062 inet_csk_schedule_ack(sk);
3063 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3064 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3065 TCP_DELACK_MAX, TCP_RTO_MAX);
3069 /* Reserve space for headers and prepare control bits. */
3070 skb_reserve(buff, MAX_TCP_HEADER);
3071 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3073 /* Send it off, this clears delayed acks for us. */
3074 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3075 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3078 /* This routine sends a packet with an out of date sequence
3079 * number. It assumes the other end will try to ack it.
3081 * Question: what should we make while urgent mode?
3082 * 4.4BSD forces sending single byte of data. We cannot send
3083 * out of window data, because we have SND.NXT==SND.MAX...
3085 * Current solution: to send TWO zero-length segments in urgent mode:
3086 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3087 * out-of-date with SND.UNA-1 to probe window.
3089 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3091 struct tcp_sock *tp = tcp_sk(sk);
3092 struct sk_buff *skb;
3094 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3095 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3099 /* Reserve space for headers and set control bits. */
3100 skb_reserve(skb, MAX_TCP_HEADER);
3101 /* Use a previous sequence. This should cause the other
3102 * end to send an ack. Don't queue or clone SKB, just
3105 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3106 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3107 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3110 void tcp_send_window_probe(struct sock *sk)
3112 if (sk->sk_state == TCP_ESTABLISHED) {
3113 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3114 tcp_xmit_probe_skb(sk, 0);
3118 /* Initiate keepalive or window probe from timer. */
3119 int tcp_write_wakeup(struct sock *sk)
3121 struct tcp_sock *tp = tcp_sk(sk);
3122 struct sk_buff *skb;
3124 if (sk->sk_state == TCP_CLOSE)
3127 if ((skb = tcp_send_head(sk)) != NULL &&
3128 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3130 unsigned int mss = tcp_current_mss(sk);
3131 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3133 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3134 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3136 /* We are probing the opening of a window
3137 * but the window size is != 0
3138 * must have been a result SWS avoidance ( sender )
3140 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3142 seg_size = min(seg_size, mss);
3143 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3144 if (tcp_fragment(sk, skb, seg_size, mss))
3146 } else if (!tcp_skb_pcount(skb))
3147 tcp_set_skb_tso_segs(sk, skb, mss);
3149 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3150 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3151 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3153 tcp_event_new_data_sent(sk, skb);
3156 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3157 tcp_xmit_probe_skb(sk, 1);
3158 return tcp_xmit_probe_skb(sk, 0);
3162 /* A window probe timeout has occurred. If window is not closed send
3163 * a partial packet else a zero probe.
3165 void tcp_send_probe0(struct sock *sk)
3167 struct inet_connection_sock *icsk = inet_csk(sk);
3168 struct tcp_sock *tp = tcp_sk(sk);
3171 err = tcp_write_wakeup(sk);
3173 if (tp->packets_out || !tcp_send_head(sk)) {
3174 /* Cancel probe timer, if it is not required. */
3175 icsk->icsk_probes_out = 0;
3176 icsk->icsk_backoff = 0;
3181 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3182 icsk->icsk_backoff++;
3183 icsk->icsk_probes_out++;
3184 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3185 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3188 /* If packet was not sent due to local congestion,
3189 * do not backoff and do not remember icsk_probes_out.
3190 * Let local senders to fight for local resources.
3192 * Use accumulated backoff yet.
3194 if (!icsk->icsk_probes_out)
3195 icsk->icsk_probes_out = 1;
3196 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3197 min(icsk->icsk_rto << icsk->icsk_backoff,
3198 TCP_RESOURCE_PROBE_INTERVAL),