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 four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
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 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
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.
139 void tcp_cwnd_restart(struct sock *sk, s32 delta)
141 struct tcp_sock *tp = tcp_sk(sk);
142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
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 (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
180 struct tcp_sock *tp = tcp_sk(sk);
182 if (unlikely(rcv_nxt != tp->rcv_nxt))
183 return; /* Special ACK sent by DCTCP to reflect ECN */
184 tcp_dec_quickack_mode(sk, pkts);
185 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
189 u32 tcp_default_init_rwnd(u32 mss)
191 /* Initial receive window should be twice of TCP_INIT_CWND to
192 * enable proper sending of new unsent data during fast recovery
193 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
194 * limit when mss is larger than 1460.
196 u32 init_rwnd = TCP_INIT_CWND * 2;
199 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
203 /* Determine a window scaling and initial window to offer.
204 * Based on the assumption that the given amount of space
205 * will be offered. Store the results in the tp structure.
206 * NOTE: for smooth operation initial space offering should
207 * be a multiple of mss if possible. We assume here that mss >= 1.
208 * This MUST be enforced by all callers.
210 void tcp_select_initial_window(int __space, __u32 mss,
211 __u32 *rcv_wnd, __u32 *window_clamp,
212 int wscale_ok, __u8 *rcv_wscale,
215 unsigned int space = (__space < 0 ? 0 : __space);
217 /* If no clamp set the clamp to the max possible scaled window */
218 if (*window_clamp == 0)
219 (*window_clamp) = (65535 << 14);
220 space = min(*window_clamp, space);
222 /* Quantize space offering to a multiple of mss if possible. */
224 space = (space / mss) * mss;
226 /* NOTE: offering an initial window larger than 32767
227 * will break some buggy TCP stacks. If the admin tells us
228 * it is likely we could be speaking with such a buggy stack
229 * we will truncate our initial window offering to 32K-1
230 * unless the remote has sent us a window scaling option,
231 * which we interpret as a sign the remote TCP is not
232 * misinterpreting the window field as a signed quantity.
234 if (sysctl_tcp_workaround_signed_windows)
235 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
241 /* Set window scaling on max possible window
242 * See RFC1323 for an explanation of the limit to 14
244 space = max_t(u32, space, sysctl_tcp_rmem[2]);
245 space = max_t(u32, space, sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 while (space > 65535 && (*rcv_wscale) < 14) {
253 if (mss > (1 << *rcv_wscale)) {
254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
262 EXPORT_SYMBOL(tcp_select_initial_window);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16 tcp_select_window(struct sock *sk)
271 struct tcp_sock *tp = tcp_sk(sk);
272 u32 old_win = tp->rcv_wnd;
273 u32 cur_win = tcp_receive_window(tp);
274 u32 new_win = __tcp_select_window(sk);
276 /* Never shrink the offered window */
277 if (new_win < cur_win) {
278 /* Danger Will Robinson!
279 * Don't update rcv_wup/rcv_wnd here or else
280 * we will not be able to advertise a zero
281 * window in time. --DaveM
283 * Relax Will Robinson.
286 NET_INC_STATS(sock_net(sk),
287 LINUX_MIB_TCPWANTZEROWINDOWADV);
288 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
290 tp->rcv_wnd = new_win;
291 tp->rcv_wup = tp->rcv_nxt;
293 /* Make sure we do not exceed the maximum possible
296 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
297 new_win = min(new_win, MAX_TCP_WINDOW);
299 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
301 /* RFC1323 scaling applied */
302 new_win >>= tp->rx_opt.rcv_wscale;
304 /* If we advertise zero window, disable fast path. */
308 NET_INC_STATS(sock_net(sk),
309 LINUX_MIB_TCPTOZEROWINDOWADV);
310 } else if (old_win == 0) {
311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
317 /* Packet ECN state for a SYN-ACK */
318 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
320 const struct tcp_sock *tp = tcp_sk(sk);
322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
323 if (!(tp->ecn_flags & TCP_ECN_OK))
324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
325 else if (tcp_ca_needs_ecn(sk))
329 /* Packet ECN state for a SYN. */
330 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
332 struct tcp_sock *tp = tcp_sk(sk);
333 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
334 tcp_ca_needs_ecn(sk);
337 const struct dst_entry *dst = __sk_dst_get(sk);
339 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
346 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
347 tp->ecn_flags = TCP_ECN_OK;
348 if (tcp_ca_needs_ecn(sk))
353 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
355 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
356 /* tp->ecn_flags are cleared at a later point in time when
357 * SYN ACK is ultimatively being received.
359 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
363 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
365 if (inet_rsk(req)->ecn_ok)
369 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
372 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
373 struct tcphdr *th, int tcp_header_len)
375 struct tcp_sock *tp = tcp_sk(sk);
377 if (tp->ecn_flags & TCP_ECN_OK) {
378 /* Not-retransmitted data segment: set ECT and inject CWR. */
379 if (skb->len != tcp_header_len &&
380 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
382 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
383 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
385 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
387 } else if (!tcp_ca_needs_ecn(sk)) {
388 /* ACK or retransmitted segment: clear ECT|CE */
389 INET_ECN_dontxmit(sk);
391 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
396 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
397 * auto increment end seqno.
399 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
401 skb->ip_summed = CHECKSUM_PARTIAL;
404 TCP_SKB_CB(skb)->tcp_flags = flags;
405 TCP_SKB_CB(skb)->sacked = 0;
407 tcp_skb_pcount_set(skb, 1);
409 TCP_SKB_CB(skb)->seq = seq;
410 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
412 TCP_SKB_CB(skb)->end_seq = seq;
415 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
417 return tp->snd_una != tp->snd_up;
420 #define OPTION_SACK_ADVERTISE (1 << 0)
421 #define OPTION_TS (1 << 1)
422 #define OPTION_MD5 (1 << 2)
423 #define OPTION_WSCALE (1 << 3)
424 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
426 struct tcp_out_options {
427 u16 options; /* bit field of OPTION_* */
428 u16 mss; /* 0 to disable */
429 u8 ws; /* window scale, 0 to disable */
430 u8 num_sack_blocks; /* number of SACK blocks to include */
431 u8 hash_size; /* bytes in hash_location */
432 __u8 *hash_location; /* temporary pointer, overloaded */
433 __u32 tsval, tsecr; /* need to include OPTION_TS */
434 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
437 /* Write previously computed TCP options to the packet.
439 * Beware: Something in the Internet is very sensitive to the ordering of
440 * TCP options, we learned this through the hard way, so be careful here.
441 * Luckily we can at least blame others for their non-compliance but from
442 * inter-operability perspective it seems that we're somewhat stuck with
443 * the ordering which we have been using if we want to keep working with
444 * those broken things (not that it currently hurts anybody as there isn't
445 * particular reason why the ordering would need to be changed).
447 * At least SACK_PERM as the first option is known to lead to a disaster
448 * (but it may well be that other scenarios fail similarly).
450 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
451 struct tcp_out_options *opts)
453 u16 options = opts->options; /* mungable copy */
455 if (unlikely(OPTION_MD5 & options)) {
456 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
457 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
458 /* overload cookie hash location */
459 opts->hash_location = (__u8 *)ptr;
463 if (unlikely(opts->mss)) {
464 *ptr++ = htonl((TCPOPT_MSS << 24) |
465 (TCPOLEN_MSS << 16) |
469 if (likely(OPTION_TS & options)) {
470 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
471 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
472 (TCPOLEN_SACK_PERM << 16) |
473 (TCPOPT_TIMESTAMP << 8) |
475 options &= ~OPTION_SACK_ADVERTISE;
477 *ptr++ = htonl((TCPOPT_NOP << 24) |
479 (TCPOPT_TIMESTAMP << 8) |
482 *ptr++ = htonl(opts->tsval);
483 *ptr++ = htonl(opts->tsecr);
486 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
487 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_SACK_PERM << 8) |
493 if (unlikely(OPTION_WSCALE & options)) {
494 *ptr++ = htonl((TCPOPT_NOP << 24) |
495 (TCPOPT_WINDOW << 16) |
496 (TCPOLEN_WINDOW << 8) |
500 if (unlikely(opts->num_sack_blocks)) {
501 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
502 tp->duplicate_sack : tp->selective_acks;
505 *ptr++ = htonl((TCPOPT_NOP << 24) |
508 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
509 TCPOLEN_SACK_PERBLOCK)));
511 for (this_sack = 0; this_sack < opts->num_sack_blocks;
513 *ptr++ = htonl(sp[this_sack].start_seq);
514 *ptr++ = htonl(sp[this_sack].end_seq);
517 tp->rx_opt.dsack = 0;
520 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
521 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
523 u32 len; /* Fast Open option length */
526 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
527 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
528 TCPOPT_FASTOPEN_MAGIC);
529 p += TCPOLEN_EXP_FASTOPEN_BASE;
531 len = TCPOLEN_FASTOPEN_BASE + foc->len;
532 *p++ = TCPOPT_FASTOPEN;
536 memcpy(p, foc->val, foc->len);
537 if ((len & 3) == 2) {
538 p[foc->len] = TCPOPT_NOP;
539 p[foc->len + 1] = TCPOPT_NOP;
541 ptr += (len + 3) >> 2;
545 /* Compute TCP options for SYN packets. This is not the final
546 * network wire format yet.
548 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
549 struct tcp_out_options *opts,
550 struct tcp_md5sig_key **md5)
552 struct tcp_sock *tp = tcp_sk(sk);
553 unsigned int remaining = MAX_TCP_OPTION_SPACE;
554 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
556 #ifdef CONFIG_TCP_MD5SIG
557 *md5 = tp->af_specific->md5_lookup(sk, sk);
559 opts->options |= OPTION_MD5;
560 remaining -= TCPOLEN_MD5SIG_ALIGNED;
566 /* We always get an MSS option. The option bytes which will be seen in
567 * normal data packets should timestamps be used, must be in the MSS
568 * advertised. But we subtract them from tp->mss_cache so that
569 * calculations in tcp_sendmsg are simpler etc. So account for this
570 * fact here if necessary. If we don't do this correctly, as a
571 * receiver we won't recognize data packets as being full sized when we
572 * should, and thus we won't abide by the delayed ACK rules correctly.
573 * SACKs don't matter, we never delay an ACK when we have any of those
575 opts->mss = tcp_advertise_mss(sk);
576 remaining -= TCPOLEN_MSS_ALIGNED;
578 if (likely(sysctl_tcp_timestamps && !*md5)) {
579 opts->options |= OPTION_TS;
580 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
581 opts->tsecr = tp->rx_opt.ts_recent;
582 remaining -= TCPOLEN_TSTAMP_ALIGNED;
584 if (likely(sysctl_tcp_window_scaling)) {
585 opts->ws = tp->rx_opt.rcv_wscale;
586 opts->options |= OPTION_WSCALE;
587 remaining -= TCPOLEN_WSCALE_ALIGNED;
589 if (likely(sysctl_tcp_sack)) {
590 opts->options |= OPTION_SACK_ADVERTISE;
591 if (unlikely(!(OPTION_TS & opts->options)))
592 remaining -= TCPOLEN_SACKPERM_ALIGNED;
595 if (fastopen && fastopen->cookie.len >= 0) {
596 u32 need = fastopen->cookie.len;
598 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
599 TCPOLEN_FASTOPEN_BASE;
600 need = (need + 3) & ~3U; /* Align to 32 bits */
601 if (remaining >= need) {
602 opts->options |= OPTION_FAST_OPEN_COOKIE;
603 opts->fastopen_cookie = &fastopen->cookie;
605 tp->syn_fastopen = 1;
606 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
610 return MAX_TCP_OPTION_SPACE - remaining;
613 /* Set up TCP options for SYN-ACKs. */
614 static unsigned int tcp_synack_options(struct request_sock *req,
615 unsigned int mss, struct sk_buff *skb,
616 struct tcp_out_options *opts,
617 const struct tcp_md5sig_key *md5,
618 struct tcp_fastopen_cookie *foc,
619 enum tcp_synack_type synack_type)
621 struct inet_request_sock *ireq = inet_rsk(req);
622 unsigned int remaining = MAX_TCP_OPTION_SPACE;
624 #ifdef CONFIG_TCP_MD5SIG
626 opts->options |= OPTION_MD5;
627 remaining -= TCPOLEN_MD5SIG_ALIGNED;
629 /* We can't fit any SACK blocks in a packet with MD5 + TS
630 * options. There was discussion about disabling SACK
631 * rather than TS in order to fit in better with old,
632 * buggy kernels, but that was deemed to be unnecessary.
634 if (synack_type != TCP_SYNACK_COOKIE)
635 ireq->tstamp_ok &= !ireq->sack_ok;
639 /* We always send an MSS option. */
641 remaining -= TCPOLEN_MSS_ALIGNED;
643 if (likely(ireq->wscale_ok)) {
644 opts->ws = ireq->rcv_wscale;
645 opts->options |= OPTION_WSCALE;
646 remaining -= TCPOLEN_WSCALE_ALIGNED;
648 if (likely(ireq->tstamp_ok)) {
649 opts->options |= OPTION_TS;
650 opts->tsval = tcp_skb_timestamp(skb);
651 opts->tsecr = req->ts_recent;
652 remaining -= TCPOLEN_TSTAMP_ALIGNED;
654 if (likely(ireq->sack_ok)) {
655 opts->options |= OPTION_SACK_ADVERTISE;
656 if (unlikely(!ireq->tstamp_ok))
657 remaining -= TCPOLEN_SACKPERM_ALIGNED;
659 if (foc != NULL && foc->len >= 0) {
662 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
663 TCPOLEN_FASTOPEN_BASE;
664 need = (need + 3) & ~3U; /* Align to 32 bits */
665 if (remaining >= need) {
666 opts->options |= OPTION_FAST_OPEN_COOKIE;
667 opts->fastopen_cookie = foc;
672 return MAX_TCP_OPTION_SPACE - remaining;
675 /* Compute TCP options for ESTABLISHED sockets. This is not the
676 * final wire format yet.
678 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
679 struct tcp_out_options *opts,
680 struct tcp_md5sig_key **md5)
682 struct tcp_sock *tp = tcp_sk(sk);
683 unsigned int size = 0;
684 unsigned int eff_sacks;
688 #ifdef CONFIG_TCP_MD5SIG
689 *md5 = tp->af_specific->md5_lookup(sk, sk);
690 if (unlikely(*md5)) {
691 opts->options |= OPTION_MD5;
692 size += TCPOLEN_MD5SIG_ALIGNED;
698 if (likely(tp->rx_opt.tstamp_ok)) {
699 opts->options |= OPTION_TS;
700 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
701 opts->tsecr = tp->rx_opt.ts_recent;
702 size += TCPOLEN_TSTAMP_ALIGNED;
705 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
706 if (unlikely(eff_sacks)) {
707 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
708 opts->num_sack_blocks =
709 min_t(unsigned int, eff_sacks,
710 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
711 TCPOLEN_SACK_PERBLOCK);
712 if (likely(opts->num_sack_blocks))
713 size += TCPOLEN_SACK_BASE_ALIGNED +
714 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
721 /* TCP SMALL QUEUES (TSQ)
723 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
724 * to reduce RTT and bufferbloat.
725 * We do this using a special skb destructor (tcp_wfree).
727 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
728 * needs to be reallocated in a driver.
729 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
731 * Since transmit from skb destructor is forbidden, we use a tasklet
732 * to process all sockets that eventually need to send more skbs.
733 * We use one tasklet per cpu, with its own queue of sockets.
736 struct tasklet_struct tasklet;
737 struct list_head head; /* queue of tcp sockets */
739 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
741 static void tcp_tsq_handler(struct sock *sk)
743 if ((1 << sk->sk_state) &
744 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
745 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
746 struct tcp_sock *tp = tcp_sk(sk);
748 if (tp->lost_out > tp->retrans_out &&
749 tp->snd_cwnd > tcp_packets_in_flight(tp))
750 tcp_xmit_retransmit_queue(sk);
752 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
757 * One tasklet per cpu tries to send more skbs.
758 * We run in tasklet context but need to disable irqs when
759 * transferring tsq->head because tcp_wfree() might
760 * interrupt us (non NAPI drivers)
762 static void tcp_tasklet_func(unsigned long data)
764 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
767 struct list_head *q, *n;
771 local_irq_save(flags);
772 list_splice_init(&tsq->head, &list);
773 local_irq_restore(flags);
775 list_for_each_safe(q, n, &list) {
776 tp = list_entry(q, struct tcp_sock, tsq_node);
777 list_del(&tp->tsq_node);
779 sk = (struct sock *)tp;
782 if (!sock_owned_by_user(sk)) {
785 /* defer the work to tcp_release_cb() */
786 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
790 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
795 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
796 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
797 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
798 (1UL << TCP_MTU_REDUCED_DEFERRED))
800 * tcp_release_cb - tcp release_sock() callback
803 * called from release_sock() to perform protocol dependent
804 * actions before socket release.
806 void tcp_release_cb(struct sock *sk)
808 struct tcp_sock *tp = tcp_sk(sk);
809 unsigned long flags, nflags;
811 /* perform an atomic operation only if at least one flag is set */
813 flags = tp->tsq_flags;
814 if (!(flags & TCP_DEFERRED_ALL))
816 nflags = flags & ~TCP_DEFERRED_ALL;
817 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
819 if (flags & (1UL << TCP_TSQ_DEFERRED))
822 /* Here begins the tricky part :
823 * We are called from release_sock() with :
825 * 2) sk_lock.slock spinlock held
826 * 3) socket owned by us (sk->sk_lock.owned == 1)
828 * But following code is meant to be called from BH handlers,
829 * so we should keep BH disabled, but early release socket ownership
831 sock_release_ownership(sk);
833 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
834 tcp_write_timer_handler(sk);
837 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
838 tcp_delack_timer_handler(sk);
841 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
842 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
846 EXPORT_SYMBOL(tcp_release_cb);
848 void __init tcp_tasklet_init(void)
852 for_each_possible_cpu(i) {
853 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
855 INIT_LIST_HEAD(&tsq->head);
856 tasklet_init(&tsq->tasklet,
863 * Write buffer destructor automatically called from kfree_skb.
864 * We can't xmit new skbs from this context, as we might already
867 void tcp_wfree(struct sk_buff *skb)
869 struct sock *sk = skb->sk;
870 struct tcp_sock *tp = tcp_sk(sk);
873 /* Keep one reference on sk_wmem_alloc.
874 * Will be released by sk_free() from here or tcp_tasklet_func()
876 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
878 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
879 * Wait until our queues (qdisc + devices) are drained.
881 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
882 * - chance for incoming ACK (processed by another cpu maybe)
883 * to migrate this flow (skb->ooo_okay will be eventually set)
885 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
888 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
889 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
891 struct tsq_tasklet *tsq;
893 /* queue this socket to tasklet queue */
894 local_irq_save(flags);
895 tsq = this_cpu_ptr(&tsq_tasklet);
896 list_add(&tp->tsq_node, &tsq->head);
897 tasklet_schedule(&tsq->tasklet);
898 local_irq_restore(flags);
905 /* This routine actually transmits TCP packets queued in by
906 * tcp_do_sendmsg(). This is used by both the initial
907 * transmission and possible later retransmissions.
908 * All SKB's seen here are completely headerless. It is our
909 * job to build the TCP header, and pass the packet down to
910 * IP so it can do the same plus pass the packet off to the
913 * We are working here with either a clone of the original
914 * SKB, or a fresh unique copy made by the retransmit engine.
916 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
917 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
919 const struct inet_connection_sock *icsk = inet_csk(sk);
920 struct inet_sock *inet;
922 struct tcp_skb_cb *tcb;
923 struct tcp_out_options opts;
924 unsigned int tcp_options_size, tcp_header_size;
925 struct sk_buff *oskb = NULL;
926 struct tcp_md5sig_key *md5;
930 BUG_ON(!skb || !tcp_skb_pcount(skb));
934 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
937 if (unlikely(skb_cloned(skb)))
938 skb = pskb_copy(skb, gfp_mask);
940 skb = skb_clone(skb, gfp_mask);
944 skb_mstamp_get(&skb->skb_mstamp);
947 tcb = TCP_SKB_CB(skb);
948 memset(&opts, 0, sizeof(opts));
950 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
951 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
953 tcp_options_size = tcp_established_options(sk, skb, &opts,
955 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
957 /* if no packet is in qdisc/device queue, then allow XPS to select
958 * another queue. We can be called from tcp_tsq_handler()
959 * which holds one reference to sk_wmem_alloc.
961 * TODO: Ideally, in-flight pure ACK packets should not matter here.
962 * One way to get this would be to set skb->truesize = 2 on them.
964 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
966 skb_push(skb, tcp_header_size);
967 skb_reset_transport_header(skb);
971 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
972 skb_set_hash_from_sk(skb, sk);
973 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
975 /* Build TCP header and checksum it. */
976 th = (struct tcphdr *)skb->data;
977 th->source = inet->inet_sport;
978 th->dest = inet->inet_dport;
979 th->seq = htonl(tcb->seq);
980 th->ack_seq = htonl(rcv_nxt);
981 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
987 /* The urg_mode check is necessary during a below snd_una win probe */
988 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
989 if (before(tp->snd_up, tcb->seq + 0x10000)) {
990 th->urg_ptr = htons(tp->snd_up - tcb->seq);
992 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
993 th->urg_ptr = htons(0xFFFF);
998 tcp_options_write((__be32 *)(th + 1), tp, &opts);
999 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1000 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1001 th->window = htons(tcp_select_window(sk));
1002 tcp_ecn_send(sk, skb, th, tcp_header_size);
1004 /* RFC1323: The window in SYN & SYN/ACK segments
1007 th->window = htons(min(tp->rcv_wnd, 65535U));
1009 #ifdef CONFIG_TCP_MD5SIG
1010 /* Calculate the MD5 hash, as we have all we need now */
1012 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1013 tp->af_specific->calc_md5_hash(opts.hash_location,
1018 icsk->icsk_af_ops->send_check(sk, skb);
1020 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1021 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1023 if (skb->len != tcp_header_size) {
1024 tcp_event_data_sent(tp, sk);
1025 tp->data_segs_out += tcp_skb_pcount(skb);
1028 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1029 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1030 tcp_skb_pcount(skb));
1032 tp->segs_out += tcp_skb_pcount(skb);
1033 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1034 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1035 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1037 /* Our usage of tstamp should remain private */
1038 skb->tstamp.tv64 = 0;
1040 /* Cleanup our debris for IP stacks */
1041 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1042 sizeof(struct inet6_skb_parm)));
1044 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1046 if (unlikely(err > 0)) {
1048 err = net_xmit_eval(err);
1051 skb_mstamp_get(&oskb->skb_mstamp);
1052 tcp_rate_skb_sent(sk, oskb);
1057 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1060 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1061 tcp_sk(sk)->rcv_nxt);
1064 /* This routine just queues the buffer for sending.
1066 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1067 * otherwise socket can stall.
1069 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1071 struct tcp_sock *tp = tcp_sk(sk);
1073 /* Advance write_seq and place onto the write_queue. */
1074 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1075 __skb_header_release(skb);
1076 tcp_add_write_queue_tail(sk, skb);
1077 sk->sk_wmem_queued += skb->truesize;
1078 sk_mem_charge(sk, skb->truesize);
1081 /* Initialize TSO segments for a packet. */
1082 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1084 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1085 /* Avoid the costly divide in the normal
1088 tcp_skb_pcount_set(skb, 1);
1089 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1091 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1092 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1096 /* When a modification to fackets out becomes necessary, we need to check
1097 * skb is counted to fackets_out or not.
1099 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1102 struct tcp_sock *tp = tcp_sk(sk);
1104 if (!tp->sacked_out || tcp_is_reno(tp))
1107 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1108 tp->fackets_out -= decr;
1111 /* Pcount in the middle of the write queue got changed, we need to do various
1112 * tweaks to fix counters
1114 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1116 struct tcp_sock *tp = tcp_sk(sk);
1118 tp->packets_out -= decr;
1120 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1121 tp->sacked_out -= decr;
1122 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1123 tp->retrans_out -= decr;
1124 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1125 tp->lost_out -= decr;
1127 /* Reno case is special. Sigh... */
1128 if (tcp_is_reno(tp) && decr > 0)
1129 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1131 tcp_adjust_fackets_out(sk, skb, decr);
1133 if (tp->lost_skb_hint &&
1134 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1135 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1136 tp->lost_cnt_hint -= decr;
1138 tcp_verify_left_out(tp);
1141 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1143 return TCP_SKB_CB(skb)->txstamp_ack ||
1144 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1147 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1149 struct skb_shared_info *shinfo = skb_shinfo(skb);
1151 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1152 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1153 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1154 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1156 shinfo->tx_flags &= ~tsflags;
1157 shinfo2->tx_flags |= tsflags;
1158 swap(shinfo->tskey, shinfo2->tskey);
1159 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1160 TCP_SKB_CB(skb)->txstamp_ack = 0;
1164 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1166 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1167 TCP_SKB_CB(skb)->eor = 0;
1170 /* Function to create two new TCP segments. Shrinks the given segment
1171 * to the specified size and appends a new segment with the rest of the
1172 * packet to the list. This won't be called frequently, I hope.
1173 * Remember, these are still headerless SKBs at this point.
1175 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1176 unsigned int mss_now, gfp_t gfp)
1178 struct tcp_sock *tp = tcp_sk(sk);
1179 struct sk_buff *buff;
1180 int nsize, old_factor;
1185 if (WARN_ON(len > skb->len))
1188 nsize = skb_headlen(skb) - len;
1192 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1193 * We need some allowance to not penalize applications setting small
1195 * Also allow first and last skb in retransmit queue to be split.
1197 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1198 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1199 skb != tcp_rtx_queue_head(sk) &&
1200 skb != tcp_rtx_queue_tail(sk))) {
1201 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1205 if (skb_unclone(skb, gfp))
1208 /* Get a new skb... force flag on. */
1209 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1211 return -ENOMEM; /* We'll just try again later. */
1213 sk->sk_wmem_queued += buff->truesize;
1214 sk_mem_charge(sk, buff->truesize);
1215 nlen = skb->len - len - nsize;
1216 buff->truesize += nlen;
1217 skb->truesize -= nlen;
1219 /* Correct the sequence numbers. */
1220 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1221 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1222 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1224 /* PSH and FIN should only be set in the second packet. */
1225 flags = TCP_SKB_CB(skb)->tcp_flags;
1226 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1227 TCP_SKB_CB(buff)->tcp_flags = flags;
1228 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1229 tcp_skb_fragment_eor(skb, buff);
1231 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1232 /* Copy and checksum data tail into the new buffer. */
1233 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1234 skb_put(buff, nsize),
1239 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1241 skb->ip_summed = CHECKSUM_PARTIAL;
1242 skb_split(skb, buff, len);
1245 buff->ip_summed = skb->ip_summed;
1247 buff->tstamp = skb->tstamp;
1248 tcp_fragment_tstamp(skb, buff);
1250 old_factor = tcp_skb_pcount(skb);
1252 /* Fix up tso_factor for both original and new SKB. */
1253 tcp_set_skb_tso_segs(skb, mss_now);
1254 tcp_set_skb_tso_segs(buff, mss_now);
1256 /* Update delivered info for the new segment */
1257 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1259 /* If this packet has been sent out already, we must
1260 * adjust the various packet counters.
1262 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1263 int diff = old_factor - tcp_skb_pcount(skb) -
1264 tcp_skb_pcount(buff);
1267 tcp_adjust_pcount(sk, skb, diff);
1270 /* Link BUFF into the send queue. */
1271 __skb_header_release(buff);
1272 tcp_insert_write_queue_after(skb, buff, sk);
1277 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1278 * eventually). The difference is that pulled data not copied, but
1279 * immediately discarded.
1281 static int __pskb_trim_head(struct sk_buff *skb, int len)
1283 struct skb_shared_info *shinfo;
1286 eat = min_t(int, len, skb_headlen(skb));
1288 __skb_pull(skb, eat);
1295 shinfo = skb_shinfo(skb);
1296 for (i = 0; i < shinfo->nr_frags; i++) {
1297 int size = skb_frag_size(&shinfo->frags[i]);
1300 skb_frag_unref(skb, i);
1303 shinfo->frags[k] = shinfo->frags[i];
1305 shinfo->frags[k].page_offset += eat;
1306 skb_frag_size_sub(&shinfo->frags[k], eat);
1312 shinfo->nr_frags = k;
1314 skb_reset_tail_pointer(skb);
1315 skb->data_len -= len;
1316 skb->len = skb->data_len;
1320 /* Remove acked data from a packet in the transmit queue. */
1321 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1325 if (skb_unclone(skb, GFP_ATOMIC))
1328 delta_truesize = __pskb_trim_head(skb, len);
1330 TCP_SKB_CB(skb)->seq += len;
1331 skb->ip_summed = CHECKSUM_PARTIAL;
1333 if (delta_truesize) {
1334 skb->truesize -= delta_truesize;
1335 sk->sk_wmem_queued -= delta_truesize;
1336 sk_mem_uncharge(sk, delta_truesize);
1337 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1340 /* Any change of skb->len requires recalculation of tso factor. */
1341 if (tcp_skb_pcount(skb) > 1)
1342 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1347 /* Calculate MSS not accounting any TCP options. */
1348 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1350 const struct tcp_sock *tp = tcp_sk(sk);
1351 const struct inet_connection_sock *icsk = inet_csk(sk);
1354 /* Calculate base mss without TCP options:
1355 It is MMS_S - sizeof(tcphdr) of rfc1122
1357 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1359 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1360 if (icsk->icsk_af_ops->net_frag_header_len) {
1361 const struct dst_entry *dst = __sk_dst_get(sk);
1363 if (dst && dst_allfrag(dst))
1364 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1367 /* Clamp it (mss_clamp does not include tcp options) */
1368 if (mss_now > tp->rx_opt.mss_clamp)
1369 mss_now = tp->rx_opt.mss_clamp;
1371 /* Now subtract optional transport overhead */
1372 mss_now -= icsk->icsk_ext_hdr_len;
1374 /* Then reserve room for full set of TCP options and 8 bytes of data */
1375 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1379 /* Calculate MSS. Not accounting for SACKs here. */
1380 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1382 /* Subtract TCP options size, not including SACKs */
1383 return __tcp_mtu_to_mss(sk, pmtu) -
1384 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1386 EXPORT_SYMBOL(tcp_mtu_to_mss);
1388 /* Inverse of above */
1389 int tcp_mss_to_mtu(struct sock *sk, int mss)
1391 const struct tcp_sock *tp = tcp_sk(sk);
1392 const struct inet_connection_sock *icsk = inet_csk(sk);
1396 tp->tcp_header_len +
1397 icsk->icsk_ext_hdr_len +
1398 icsk->icsk_af_ops->net_header_len;
1400 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1401 if (icsk->icsk_af_ops->net_frag_header_len) {
1402 const struct dst_entry *dst = __sk_dst_get(sk);
1404 if (dst && dst_allfrag(dst))
1405 mtu += icsk->icsk_af_ops->net_frag_header_len;
1409 EXPORT_SYMBOL(tcp_mss_to_mtu);
1411 /* MTU probing init per socket */
1412 void tcp_mtup_init(struct sock *sk)
1414 struct tcp_sock *tp = tcp_sk(sk);
1415 struct inet_connection_sock *icsk = inet_csk(sk);
1416 struct net *net = sock_net(sk);
1418 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1419 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1420 icsk->icsk_af_ops->net_header_len;
1421 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1422 icsk->icsk_mtup.probe_size = 0;
1423 if (icsk->icsk_mtup.enabled)
1424 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1426 EXPORT_SYMBOL(tcp_mtup_init);
1428 /* This function synchronize snd mss to current pmtu/exthdr set.
1430 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1431 for TCP options, but includes only bare TCP header.
1433 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1434 It is minimum of user_mss and mss received with SYN.
1435 It also does not include TCP options.
1437 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1439 tp->mss_cache is current effective sending mss, including
1440 all tcp options except for SACKs. It is evaluated,
1441 taking into account current pmtu, but never exceeds
1442 tp->rx_opt.mss_clamp.
1444 NOTE1. rfc1122 clearly states that advertised MSS
1445 DOES NOT include either tcp or ip options.
1447 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1448 are READ ONLY outside this function. --ANK (980731)
1450 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1452 struct tcp_sock *tp = tcp_sk(sk);
1453 struct inet_connection_sock *icsk = inet_csk(sk);
1456 if (icsk->icsk_mtup.search_high > pmtu)
1457 icsk->icsk_mtup.search_high = pmtu;
1459 mss_now = tcp_mtu_to_mss(sk, pmtu);
1460 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1462 /* And store cached results */
1463 icsk->icsk_pmtu_cookie = pmtu;
1464 if (icsk->icsk_mtup.enabled)
1465 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1466 tp->mss_cache = mss_now;
1470 EXPORT_SYMBOL(tcp_sync_mss);
1472 /* Compute the current effective MSS, taking SACKs and IP options,
1473 * and even PMTU discovery events into account.
1475 unsigned int tcp_current_mss(struct sock *sk)
1477 const struct tcp_sock *tp = tcp_sk(sk);
1478 const struct dst_entry *dst = __sk_dst_get(sk);
1480 unsigned int header_len;
1481 struct tcp_out_options opts;
1482 struct tcp_md5sig_key *md5;
1484 mss_now = tp->mss_cache;
1487 u32 mtu = dst_mtu(dst);
1488 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1489 mss_now = tcp_sync_mss(sk, mtu);
1492 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1493 sizeof(struct tcphdr);
1494 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1495 * some common options. If this is an odd packet (because we have SACK
1496 * blocks etc) then our calculated header_len will be different, and
1497 * we have to adjust mss_now correspondingly */
1498 if (header_len != tp->tcp_header_len) {
1499 int delta = (int) header_len - tp->tcp_header_len;
1506 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1507 * As additional protections, we do not touch cwnd in retransmission phases,
1508 * and if application hit its sndbuf limit recently.
1510 static void tcp_cwnd_application_limited(struct sock *sk)
1512 struct tcp_sock *tp = tcp_sk(sk);
1514 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1515 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1516 /* Limited by application or receiver window. */
1517 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1518 u32 win_used = max(tp->snd_cwnd_used, init_win);
1519 if (win_used < tp->snd_cwnd) {
1520 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1521 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1523 tp->snd_cwnd_used = 0;
1525 tp->snd_cwnd_stamp = tcp_time_stamp;
1528 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1530 struct tcp_sock *tp = tcp_sk(sk);
1532 /* Track the maximum number of outstanding packets in each
1533 * window, and remember whether we were cwnd-limited then.
1535 if (!before(tp->snd_una, tp->max_packets_seq) ||
1536 tp->packets_out > tp->max_packets_out ||
1538 tp->max_packets_out = tp->packets_out;
1539 tp->max_packets_seq = tp->snd_nxt;
1540 tp->is_cwnd_limited = is_cwnd_limited;
1543 if (tcp_is_cwnd_limited(sk)) {
1544 /* Network is feed fully. */
1545 tp->snd_cwnd_used = 0;
1546 tp->snd_cwnd_stamp = tcp_time_stamp;
1548 /* Network starves. */
1549 if (tp->packets_out > tp->snd_cwnd_used)
1550 tp->snd_cwnd_used = tp->packets_out;
1552 if (sysctl_tcp_slow_start_after_idle &&
1553 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1554 tcp_cwnd_application_limited(sk);
1558 /* Minshall's variant of the Nagle send check. */
1559 static bool tcp_minshall_check(const struct tcp_sock *tp)
1561 return after(tp->snd_sml, tp->snd_una) &&
1562 !after(tp->snd_sml, tp->snd_nxt);
1565 /* Update snd_sml if this skb is under mss
1566 * Note that a TSO packet might end with a sub-mss segment
1567 * The test is really :
1568 * if ((skb->len % mss) != 0)
1569 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1570 * But we can avoid doing the divide again given we already have
1571 * skb_pcount = skb->len / mss_now
1573 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1574 const struct sk_buff *skb)
1576 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1577 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1580 /* Return false, if packet can be sent now without violation Nagle's rules:
1581 * 1. It is full sized. (provided by caller in %partial bool)
1582 * 2. Or it contains FIN. (already checked by caller)
1583 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1584 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1585 * With Minshall's modification: all sent small packets are ACKed.
1587 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1591 ((nonagle & TCP_NAGLE_CORK) ||
1592 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1595 /* Return how many segs we'd like on a TSO packet,
1596 * to send one TSO packet per ms
1598 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1603 bytes = min(sk->sk_pacing_rate >> 10,
1604 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1606 /* Goal is to send at least one packet per ms,
1607 * not one big TSO packet every 100 ms.
1608 * This preserves ACK clocking and is consistent
1609 * with tcp_tso_should_defer() heuristic.
1611 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1615 EXPORT_SYMBOL(tcp_tso_autosize);
1617 /* Return the number of segments we want in the skb we are transmitting.
1618 * See if congestion control module wants to decide; otherwise, autosize.
1620 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1622 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1623 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1626 tso_segs = tcp_tso_autosize(sk, mss_now,
1627 sysctl_tcp_min_tso_segs);
1628 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1631 /* Returns the portion of skb which can be sent right away */
1632 static unsigned int tcp_mss_split_point(const struct sock *sk,
1633 const struct sk_buff *skb,
1634 unsigned int mss_now,
1635 unsigned int max_segs,
1638 const struct tcp_sock *tp = tcp_sk(sk);
1639 u32 partial, needed, window, max_len;
1641 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1642 max_len = mss_now * max_segs;
1644 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1647 needed = min(skb->len, window);
1649 if (max_len <= needed)
1652 partial = needed % mss_now;
1653 /* If last segment is not a full MSS, check if Nagle rules allow us
1654 * to include this last segment in this skb.
1655 * Otherwise, we'll split the skb at last MSS boundary
1657 if (tcp_nagle_check(partial != 0, tp, nonagle))
1658 return needed - partial;
1663 /* Can at least one segment of SKB be sent right now, according to the
1664 * congestion window rules? If so, return how many segments are allowed.
1666 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1667 const struct sk_buff *skb)
1669 u32 in_flight, cwnd, halfcwnd;
1671 /* Don't be strict about the congestion window for the final FIN. */
1672 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1673 tcp_skb_pcount(skb) == 1)
1676 in_flight = tcp_packets_in_flight(tp);
1677 cwnd = tp->snd_cwnd;
1678 if (in_flight >= cwnd)
1681 /* For better scheduling, ensure we have at least
1682 * 2 GSO packets in flight.
1684 halfcwnd = max(cwnd >> 1, 1U);
1685 return min(halfcwnd, cwnd - in_flight);
1688 /* Initialize TSO state of a skb.
1689 * This must be invoked the first time we consider transmitting
1690 * SKB onto the wire.
1692 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1694 int tso_segs = tcp_skb_pcount(skb);
1696 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1697 tcp_set_skb_tso_segs(skb, mss_now);
1698 tso_segs = tcp_skb_pcount(skb);
1704 /* Return true if the Nagle test allows this packet to be
1707 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1708 unsigned int cur_mss, int nonagle)
1710 /* Nagle rule does not apply to frames, which sit in the middle of the
1711 * write_queue (they have no chances to get new data).
1713 * This is implemented in the callers, where they modify the 'nonagle'
1714 * argument based upon the location of SKB in the send queue.
1716 if (nonagle & TCP_NAGLE_PUSH)
1719 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1720 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1723 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1729 /* Does at least the first segment of SKB fit into the send window? */
1730 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1731 const struct sk_buff *skb,
1732 unsigned int cur_mss)
1734 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1736 if (skb->len > cur_mss)
1737 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1739 return !after(end_seq, tcp_wnd_end(tp));
1742 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1743 * should be put on the wire right now. If so, it returns the number of
1744 * packets allowed by the congestion window.
1746 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1747 unsigned int cur_mss, int nonagle)
1749 const struct tcp_sock *tp = tcp_sk(sk);
1750 unsigned int cwnd_quota;
1752 tcp_init_tso_segs(skb, cur_mss);
1754 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1757 cwnd_quota = tcp_cwnd_test(tp, skb);
1758 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1764 /* Test if sending is allowed right now. */
1765 bool tcp_may_send_now(struct sock *sk)
1767 const struct tcp_sock *tp = tcp_sk(sk);
1768 struct sk_buff *skb = tcp_send_head(sk);
1771 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1772 (tcp_skb_is_last(sk, skb) ?
1773 tp->nonagle : TCP_NAGLE_PUSH));
1776 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1777 * which is put after SKB on the list. It is very much like
1778 * tcp_fragment() except that it may make several kinds of assumptions
1779 * in order to speed up the splitting operation. In particular, we
1780 * know that all the data is in scatter-gather pages, and that the
1781 * packet has never been sent out before (and thus is not cloned).
1783 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1784 unsigned int mss_now, gfp_t gfp)
1786 struct sk_buff *buff;
1787 int nlen = skb->len - len;
1790 /* All of a TSO frame must be composed of paged data. */
1791 if (skb->len != skb->data_len)
1792 return tcp_fragment(sk, skb, len, mss_now, gfp);
1794 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1795 if (unlikely(!buff))
1798 sk->sk_wmem_queued += buff->truesize;
1799 sk_mem_charge(sk, buff->truesize);
1800 buff->truesize += nlen;
1801 skb->truesize -= nlen;
1803 /* Correct the sequence numbers. */
1804 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1805 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1806 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1808 /* PSH and FIN should only be set in the second packet. */
1809 flags = TCP_SKB_CB(skb)->tcp_flags;
1810 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1811 TCP_SKB_CB(buff)->tcp_flags = flags;
1813 /* This packet was never sent out yet, so no SACK bits. */
1814 TCP_SKB_CB(buff)->sacked = 0;
1816 tcp_skb_fragment_eor(skb, buff);
1818 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1819 skb_split(skb, buff, len);
1820 tcp_fragment_tstamp(skb, buff);
1822 /* Fix up tso_factor for both original and new SKB. */
1823 tcp_set_skb_tso_segs(skb, mss_now);
1824 tcp_set_skb_tso_segs(buff, mss_now);
1826 /* Link BUFF into the send queue. */
1827 __skb_header_release(buff);
1828 tcp_insert_write_queue_after(skb, buff, sk);
1833 /* Try to defer sending, if possible, in order to minimize the amount
1834 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1836 * This algorithm is from John Heffner.
1838 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1839 bool *is_cwnd_limited, u32 max_segs)
1841 const struct inet_connection_sock *icsk = inet_csk(sk);
1842 u32 age, send_win, cong_win, limit, in_flight;
1843 struct tcp_sock *tp = tcp_sk(sk);
1844 struct skb_mstamp now;
1845 struct sk_buff *head;
1848 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1851 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1854 /* Avoid bursty behavior by allowing defer
1855 * only if the last write was recent.
1857 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1860 in_flight = tcp_packets_in_flight(tp);
1862 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1864 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1866 /* From in_flight test above, we know that cwnd > in_flight. */
1867 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1869 limit = min(send_win, cong_win);
1871 /* If a full-sized TSO skb can be sent, do it. */
1872 if (limit >= max_segs * tp->mss_cache)
1875 /* Middle in queue won't get any more data, full sendable already? */
1876 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1879 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1881 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1883 /* If at least some fraction of a window is available,
1886 chunk /= win_divisor;
1890 /* Different approach, try not to defer past a single
1891 * ACK. Receiver should ACK every other full sized
1892 * frame, so if we have space for more than 3 frames
1895 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1899 head = tcp_write_queue_head(sk);
1900 skb_mstamp_get(&now);
1901 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1902 /* If next ACK is likely to come too late (half srtt), do not defer */
1903 if (age < (tp->srtt_us >> 4))
1906 /* Ok, it looks like it is advisable to defer. */
1908 if (cong_win < send_win && cong_win <= skb->len)
1909 *is_cwnd_limited = true;
1917 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1919 struct inet_connection_sock *icsk = inet_csk(sk);
1920 struct tcp_sock *tp = tcp_sk(sk);
1921 struct net *net = sock_net(sk);
1925 interval = net->ipv4.sysctl_tcp_probe_interval;
1926 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1927 if (unlikely(delta >= interval * HZ)) {
1928 int mss = tcp_current_mss(sk);
1930 /* Update current search range */
1931 icsk->icsk_mtup.probe_size = 0;
1932 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1933 sizeof(struct tcphdr) +
1934 icsk->icsk_af_ops->net_header_len;
1935 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1937 /* Update probe time stamp */
1938 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1942 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
1944 struct sk_buff *skb, *next;
1946 skb = tcp_send_head(sk);
1947 tcp_for_write_queue_from_safe(skb, next, sk) {
1948 if (len <= skb->len)
1951 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
1960 /* Create a new MTU probe if we are ready.
1961 * MTU probe is regularly attempting to increase the path MTU by
1962 * deliberately sending larger packets. This discovers routing
1963 * changes resulting in larger path MTUs.
1965 * Returns 0 if we should wait to probe (no cwnd available),
1966 * 1 if a probe was sent,
1969 static int tcp_mtu_probe(struct sock *sk)
1971 struct tcp_sock *tp = tcp_sk(sk);
1972 struct inet_connection_sock *icsk = inet_csk(sk);
1973 struct sk_buff *skb, *nskb, *next;
1974 struct net *net = sock_net(sk);
1982 /* Not currently probing/verifying,
1984 * have enough cwnd, and
1985 * not SACKing (the variable headers throw things off) */
1986 if (!icsk->icsk_mtup.enabled ||
1987 icsk->icsk_mtup.probe_size ||
1988 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1989 tp->snd_cwnd < 11 ||
1990 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1993 /* Use binary search for probe_size between tcp_mss_base,
1994 * and current mss_clamp. if (search_high - search_low)
1995 * smaller than a threshold, backoff from probing.
1997 mss_now = tcp_current_mss(sk);
1998 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1999 icsk->icsk_mtup.search_low) >> 1);
2000 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2001 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2002 /* When misfortune happens, we are reprobing actively,
2003 * and then reprobe timer has expired. We stick with current
2004 * probing process by not resetting search range to its orignal.
2006 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2007 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2008 /* Check whether enough time has elaplased for
2009 * another round of probing.
2011 tcp_mtu_check_reprobe(sk);
2015 /* Have enough data in the send queue to probe? */
2016 if (tp->write_seq - tp->snd_nxt < size_needed)
2019 if (tp->snd_wnd < size_needed)
2021 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2024 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2025 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2026 if (!tcp_packets_in_flight(tp))
2032 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2035 /* We're allowed to probe. Build it now. */
2036 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2039 sk->sk_wmem_queued += nskb->truesize;
2040 sk_mem_charge(sk, nskb->truesize);
2042 skb = tcp_send_head(sk);
2044 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2045 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2046 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2047 TCP_SKB_CB(nskb)->sacked = 0;
2049 nskb->ip_summed = skb->ip_summed;
2051 tcp_insert_write_queue_before(nskb, skb, sk);
2052 tcp_highest_sack_replace(sk, skb, nskb);
2055 tcp_for_write_queue_from_safe(skb, next, sk) {
2056 copy = min_t(int, skb->len, probe_size - len);
2057 if (nskb->ip_summed) {
2058 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2060 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2061 skb_put(nskb, copy),
2063 nskb->csum = csum_block_add(nskb->csum, csum, len);
2066 if (skb->len <= copy) {
2067 /* We've eaten all the data from this skb.
2069 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2070 /* If this is the last SKB we copy and eor is set
2071 * we need to propagate it to the new skb.
2073 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2074 tcp_skb_collapse_tstamp(nskb, skb);
2075 tcp_unlink_write_queue(skb, sk);
2076 sk_wmem_free_skb(sk, skb);
2078 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2079 ~(TCPHDR_FIN|TCPHDR_PSH);
2080 if (!skb_shinfo(skb)->nr_frags) {
2081 skb_pull(skb, copy);
2082 if (skb->ip_summed != CHECKSUM_PARTIAL)
2083 skb->csum = csum_partial(skb->data,
2086 __pskb_trim_head(skb, copy);
2087 tcp_set_skb_tso_segs(skb, mss_now);
2089 TCP_SKB_CB(skb)->seq += copy;
2094 if (len >= probe_size)
2097 tcp_init_tso_segs(nskb, nskb->len);
2099 /* We're ready to send. If this fails, the probe will
2100 * be resegmented into mss-sized pieces by tcp_write_xmit().
2102 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2103 /* Decrement cwnd here because we are sending
2104 * effectively two packets. */
2106 tcp_event_new_data_sent(sk, nskb);
2108 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2109 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2110 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2118 /* TCP Small Queues :
2119 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2120 * (These limits are doubled for retransmits)
2122 * - better RTT estimation and ACK scheduling
2125 * Alas, some drivers / subsystems require a fair amount
2126 * of queued bytes to ensure line rate.
2127 * One example is wifi aggregation (802.11 AMPDU)
2129 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2130 unsigned int factor)
2134 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2135 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2138 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2139 set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
2140 /* It is possible TX completion already happened
2141 * before we set TSQ_THROTTLED, so we must
2142 * test again the condition.
2144 smp_mb__after_atomic();
2145 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2151 /* This routine writes packets to the network. It advances the
2152 * send_head. This happens as incoming acks open up the remote
2155 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2156 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2157 * account rare use of URG, this is not a big flaw.
2159 * Send at most one packet when push_one > 0. Temporarily ignore
2160 * cwnd limit to force at most one packet out when push_one == 2.
2162 * Returns true, if no segments are in flight and we have queued segments,
2163 * but cannot send anything now because of SWS or another problem.
2165 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2166 int push_one, gfp_t gfp)
2168 struct tcp_sock *tp = tcp_sk(sk);
2169 struct sk_buff *skb;
2170 unsigned int tso_segs, sent_pkts;
2173 bool is_cwnd_limited = false;
2179 /* Do MTU probing. */
2180 result = tcp_mtu_probe(sk);
2183 } else if (result > 0) {
2188 max_segs = tcp_tso_segs(sk, mss_now);
2189 while ((skb = tcp_send_head(sk))) {
2192 tso_segs = tcp_init_tso_segs(skb, mss_now);
2195 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2196 /* "skb_mstamp" is used as a start point for the retransmit timer */
2197 skb_mstamp_get(&skb->skb_mstamp);
2198 goto repair; /* Skip network transmission */
2201 cwnd_quota = tcp_cwnd_test(tp, skb);
2204 /* Force out a loss probe pkt. */
2210 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2213 if (tso_segs == 1) {
2214 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2215 (tcp_skb_is_last(sk, skb) ?
2216 nonagle : TCP_NAGLE_PUSH))))
2220 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2226 if (tso_segs > 1 && !tcp_urg_mode(tp))
2227 limit = tcp_mss_split_point(sk, skb, mss_now,
2233 if (skb->len > limit &&
2234 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2237 if (tcp_small_queue_check(sk, skb, 0))
2240 /* Argh, we hit an empty skb(), presumably a thread
2241 * is sleeping in sendmsg()/sk_stream_wait_memory().
2242 * We do not want to send a pure-ack packet and have
2243 * a strange looking rtx queue with empty packet(s).
2245 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2248 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2252 /* Advance the send_head. This one is sent out.
2253 * This call will increment packets_out.
2255 tcp_event_new_data_sent(sk, skb);
2257 tcp_minshall_update(tp, mss_now, skb);
2258 sent_pkts += tcp_skb_pcount(skb);
2264 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2265 if (likely(sent_pkts || is_cwnd_limited))
2266 tcp_cwnd_validate(sk, is_cwnd_limited);
2268 if (likely(sent_pkts)) {
2269 if (tcp_in_cwnd_reduction(sk))
2270 tp->prr_out += sent_pkts;
2272 /* Send one loss probe per tail loss episode. */
2274 tcp_schedule_loss_probe(sk);
2277 return !tp->packets_out && tcp_send_head(sk);
2280 bool tcp_schedule_loss_probe(struct sock *sk)
2282 struct inet_connection_sock *icsk = inet_csk(sk);
2283 struct tcp_sock *tp = tcp_sk(sk);
2284 u32 timeout, tlp_time_stamp, rto_time_stamp;
2285 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2287 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2289 /* No consecutive loss probes. */
2290 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2294 /* Don't do any loss probe on a Fast Open connection before 3WHS
2297 if (tp->fastopen_rsk)
2300 /* TLP is only scheduled when next timer event is RTO. */
2301 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2304 /* Schedule a loss probe in 2*RTT for SACK capable connections
2305 * in Open state, that are either limited by cwnd or application.
2307 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2308 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2311 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2315 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2316 * for delayed ack when there's one outstanding packet. If no RTT
2317 * sample is available then probe after TCP_TIMEOUT_INIT.
2319 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2320 if (tp->packets_out == 1)
2321 timeout = max_t(u32, timeout,
2322 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2323 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2325 /* If RTO is shorter, just schedule TLP in its place. */
2326 tlp_time_stamp = tcp_time_stamp + timeout;
2327 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2328 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2329 s32 delta = rto_time_stamp - tcp_time_stamp;
2334 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2339 /* Thanks to skb fast clones, we can detect if a prior transmit of
2340 * a packet is still in a qdisc or driver queue.
2341 * In this case, there is very little point doing a retransmit !
2343 static bool skb_still_in_host_queue(const struct sock *sk,
2344 const struct sk_buff *skb)
2346 if (unlikely(skb_fclone_busy(sk, skb))) {
2347 NET_INC_STATS(sock_net(sk),
2348 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2354 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2355 * retransmit the last segment.
2357 void tcp_send_loss_probe(struct sock *sk)
2359 struct tcp_sock *tp = tcp_sk(sk);
2360 struct sk_buff *skb;
2362 int mss = tcp_current_mss(sk);
2364 /* At most one outstanding TLP */
2365 if (tp->tlp_high_seq)
2368 tp->tlp_retrans = 0;
2369 skb = tcp_send_head(sk);
2371 if (tcp_snd_wnd_test(tp, skb, mss)) {
2372 pcount = tp->packets_out;
2373 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2374 if (tp->packets_out > pcount)
2378 skb = tcp_write_queue_prev(sk, skb);
2380 skb = tcp_write_queue_tail(sk);
2383 if (unlikely(!skb)) {
2384 WARN_ONCE(tp->packets_out,
2385 "invalid inflight: %u state %u cwnd %u mss %d\n",
2386 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2387 inet_csk(sk)->icsk_pending = 0;
2391 if (skb_still_in_host_queue(sk, skb))
2394 pcount = tcp_skb_pcount(skb);
2395 if (WARN_ON(!pcount))
2398 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2399 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2402 skb = tcp_write_queue_next(sk, skb);
2405 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2408 if (__tcp_retransmit_skb(sk, skb, 1))
2411 tp->tlp_retrans = 1;
2414 /* Record snd_nxt for loss detection. */
2415 tp->tlp_high_seq = tp->snd_nxt;
2417 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2418 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2419 inet_csk(sk)->icsk_pending = 0;
2424 /* Push out any pending frames which were held back due to
2425 * TCP_CORK or attempt at coalescing tiny packets.
2426 * The socket must be locked by the caller.
2428 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2431 /* If we are closed, the bytes will have to remain here.
2432 * In time closedown will finish, we empty the write queue and
2433 * all will be happy.
2435 if (unlikely(sk->sk_state == TCP_CLOSE))
2438 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2439 sk_gfp_mask(sk, GFP_ATOMIC)))
2440 tcp_check_probe_timer(sk);
2443 /* Send _single_ skb sitting at the send head. This function requires
2444 * true push pending frames to setup probe timer etc.
2446 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2448 struct sk_buff *skb = tcp_send_head(sk);
2450 BUG_ON(!skb || skb->len < mss_now);
2452 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2455 /* This function returns the amount that we can raise the
2456 * usable window based on the following constraints
2458 * 1. The window can never be shrunk once it is offered (RFC 793)
2459 * 2. We limit memory per socket
2462 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2463 * RECV.NEXT + RCV.WIN fixed until:
2464 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2466 * i.e. don't raise the right edge of the window until you can raise
2467 * it at least MSS bytes.
2469 * Unfortunately, the recommended algorithm breaks header prediction,
2470 * since header prediction assumes th->window stays fixed.
2472 * Strictly speaking, keeping th->window fixed violates the receiver
2473 * side SWS prevention criteria. The problem is that under this rule
2474 * a stream of single byte packets will cause the right side of the
2475 * window to always advance by a single byte.
2477 * Of course, if the sender implements sender side SWS prevention
2478 * then this will not be a problem.
2480 * BSD seems to make the following compromise:
2482 * If the free space is less than the 1/4 of the maximum
2483 * space available and the free space is less than 1/2 mss,
2484 * then set the window to 0.
2485 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2486 * Otherwise, just prevent the window from shrinking
2487 * and from being larger than the largest representable value.
2489 * This prevents incremental opening of the window in the regime
2490 * where TCP is limited by the speed of the reader side taking
2491 * data out of the TCP receive queue. It does nothing about
2492 * those cases where the window is constrained on the sender side
2493 * because the pipeline is full.
2495 * BSD also seems to "accidentally" limit itself to windows that are a
2496 * multiple of MSS, at least until the free space gets quite small.
2497 * This would appear to be a side effect of the mbuf implementation.
2498 * Combining these two algorithms results in the observed behavior
2499 * of having a fixed window size at almost all times.
2501 * Below we obtain similar behavior by forcing the offered window to
2502 * a multiple of the mss when it is feasible to do so.
2504 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2505 * Regular options like TIMESTAMP are taken into account.
2507 u32 __tcp_select_window(struct sock *sk)
2509 struct inet_connection_sock *icsk = inet_csk(sk);
2510 struct tcp_sock *tp = tcp_sk(sk);
2511 /* MSS for the peer's data. Previous versions used mss_clamp
2512 * here. I don't know if the value based on our guesses
2513 * of peer's MSS is better for the performance. It's more correct
2514 * but may be worse for the performance because of rcv_mss
2515 * fluctuations. --SAW 1998/11/1
2517 int mss = icsk->icsk_ack.rcv_mss;
2518 int free_space = tcp_space(sk);
2519 int allowed_space = tcp_full_space(sk);
2520 int full_space = min_t(int, tp->window_clamp, allowed_space);
2523 if (unlikely(mss > full_space)) {
2528 if (free_space < (full_space >> 1)) {
2529 icsk->icsk_ack.quick = 0;
2531 if (tcp_under_memory_pressure(sk))
2532 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2535 /* free_space might become our new window, make sure we don't
2536 * increase it due to wscale.
2538 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2540 /* if free space is less than mss estimate, or is below 1/16th
2541 * of the maximum allowed, try to move to zero-window, else
2542 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2543 * new incoming data is dropped due to memory limits.
2544 * With large window, mss test triggers way too late in order
2545 * to announce zero window in time before rmem limit kicks in.
2547 if (free_space < (allowed_space >> 4) || free_space < mss)
2551 if (free_space > tp->rcv_ssthresh)
2552 free_space = tp->rcv_ssthresh;
2554 /* Don't do rounding if we are using window scaling, since the
2555 * scaled window will not line up with the MSS boundary anyway.
2557 window = tp->rcv_wnd;
2558 if (tp->rx_opt.rcv_wscale) {
2559 window = free_space;
2561 /* Advertise enough space so that it won't get scaled away.
2562 * Import case: prevent zero window announcement if
2563 * 1<<rcv_wscale > mss.
2565 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2566 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2567 << tp->rx_opt.rcv_wscale);
2569 /* Get the largest window that is a nice multiple of mss.
2570 * Window clamp already applied above.
2571 * If our current window offering is within 1 mss of the
2572 * free space we just keep it. This prevents the divide
2573 * and multiply from happening most of the time.
2574 * We also don't do any window rounding when the free space
2577 if (window <= free_space - mss || window > free_space)
2578 window = (free_space / mss) * mss;
2579 else if (mss == full_space &&
2580 free_space > window + (full_space >> 1))
2581 window = free_space;
2587 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2588 const struct sk_buff *next_skb)
2590 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2591 const struct skb_shared_info *next_shinfo =
2592 skb_shinfo(next_skb);
2593 struct skb_shared_info *shinfo = skb_shinfo(skb);
2595 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2596 shinfo->tskey = next_shinfo->tskey;
2597 TCP_SKB_CB(skb)->txstamp_ack |=
2598 TCP_SKB_CB(next_skb)->txstamp_ack;
2602 /* Collapses two adjacent SKB's during retransmission. */
2603 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2605 struct tcp_sock *tp = tcp_sk(sk);
2606 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2607 int skb_size, next_skb_size;
2609 skb_size = skb->len;
2610 next_skb_size = next_skb->len;
2612 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2614 tcp_highest_sack_replace(sk, next_skb, skb);
2616 tcp_unlink_write_queue(next_skb, sk);
2618 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2621 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2622 skb->ip_summed = CHECKSUM_PARTIAL;
2624 if (skb->ip_summed != CHECKSUM_PARTIAL)
2625 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2627 /* Update sequence range on original skb. */
2628 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2630 /* Merge over control information. This moves PSH/FIN etc. over */
2631 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2633 /* All done, get rid of second SKB and account for it so
2634 * packet counting does not break.
2636 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2637 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2639 /* changed transmit queue under us so clear hints */
2640 tcp_clear_retrans_hints_partial(tp);
2641 if (next_skb == tp->retransmit_skb_hint)
2642 tp->retransmit_skb_hint = skb;
2644 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2646 tcp_skb_collapse_tstamp(skb, next_skb);
2648 sk_wmem_free_skb(sk, next_skb);
2651 /* Check if coalescing SKBs is legal. */
2652 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2654 if (tcp_skb_pcount(skb) > 1)
2656 /* TODO: SACK collapsing could be used to remove this condition */
2657 if (skb_shinfo(skb)->nr_frags != 0)
2659 if (skb_cloned(skb))
2661 if (skb == tcp_send_head(sk))
2663 /* Some heurestics for collapsing over SACK'd could be invented */
2664 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2670 /* Collapse packets in the retransmit queue to make to create
2671 * less packets on the wire. This is only done on retransmission.
2673 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2676 struct tcp_sock *tp = tcp_sk(sk);
2677 struct sk_buff *skb = to, *tmp;
2680 if (!sysctl_tcp_retrans_collapse)
2682 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2685 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2686 if (!tcp_can_collapse(sk, skb))
2689 if (!tcp_skb_can_collapse_to(to))
2701 /* Punt if not enough space exists in the first SKB for
2702 * the data in the second
2704 if (skb->len > skb_availroom(to))
2707 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2710 tcp_collapse_retrans(sk, to);
2714 /* This retransmits one SKB. Policy decisions and retransmit queue
2715 * state updates are done by the caller. Returns non-zero if an
2716 * error occurred which prevented the send.
2718 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2720 struct inet_connection_sock *icsk = inet_csk(sk);
2721 struct tcp_sock *tp = tcp_sk(sk);
2722 unsigned int cur_mss;
2726 /* Inconclusive MTU probe */
2727 if (icsk->icsk_mtup.probe_size)
2728 icsk->icsk_mtup.probe_size = 0;
2730 /* Do not sent more than we queued. 1/4 is reserved for possible
2731 * copying overhead: fragmentation, tunneling, mangling etc.
2733 if (atomic_read(&sk->sk_wmem_alloc) >
2734 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2738 if (skb_still_in_host_queue(sk, skb))
2741 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2742 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2746 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2750 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2751 return -EHOSTUNREACH; /* Routing failure or similar. */
2753 cur_mss = tcp_current_mss(sk);
2755 /* If receiver has shrunk his window, and skb is out of
2756 * new window, do not retransmit it. The exception is the
2757 * case, when window is shrunk to zero. In this case
2758 * our retransmit serves as a zero window probe.
2760 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2761 TCP_SKB_CB(skb)->seq != tp->snd_una)
2764 len = cur_mss * segs;
2765 if (skb->len > len) {
2766 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2767 return -ENOMEM; /* We'll try again later. */
2769 if (skb_unclone(skb, GFP_ATOMIC))
2772 diff = tcp_skb_pcount(skb);
2773 tcp_set_skb_tso_segs(skb, cur_mss);
2774 diff -= tcp_skb_pcount(skb);
2776 tcp_adjust_pcount(sk, skb, diff);
2777 if (skb->len < cur_mss)
2778 tcp_retrans_try_collapse(sk, skb, cur_mss);
2781 /* RFC3168, section 6.1.1.1. ECN fallback */
2782 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2783 tcp_ecn_clear_syn(sk, skb);
2785 /* make sure skb->data is aligned on arches that require it
2786 * and check if ack-trimming & collapsing extended the headroom
2787 * beyond what csum_start can cover.
2789 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2790 skb_headroom(skb) >= 0xFFFF)) {
2791 struct sk_buff *nskb;
2793 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2794 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2797 skb_mstamp_get(&skb->skb_mstamp);
2799 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2803 segs = tcp_skb_pcount(skb);
2805 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2806 /* Update global TCP statistics. */
2807 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2808 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2809 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2810 tp->total_retrans += segs;
2815 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2817 struct tcp_sock *tp = tcp_sk(sk);
2818 int err = __tcp_retransmit_skb(sk, skb, segs);
2821 #if FASTRETRANS_DEBUG > 0
2822 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2823 net_dbg_ratelimited("retrans_out leaked\n");
2826 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2827 tp->retrans_out += tcp_skb_pcount(skb);
2829 /* Save stamp of the first retransmit. */
2830 if (!tp->retrans_stamp)
2831 tp->retrans_stamp = tcp_skb_timestamp(skb);
2833 } else if (err != -EBUSY) {
2834 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2837 if (tp->undo_retrans < 0)
2838 tp->undo_retrans = 0;
2839 tp->undo_retrans += tcp_skb_pcount(skb);
2843 /* Check if we forward retransmits are possible in the current
2844 * window/congestion state.
2846 static bool tcp_can_forward_retransmit(struct sock *sk)
2848 const struct inet_connection_sock *icsk = inet_csk(sk);
2849 const struct tcp_sock *tp = tcp_sk(sk);
2851 /* Forward retransmissions are possible only during Recovery. */
2852 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2855 /* No forward retransmissions in Reno are possible. */
2856 if (tcp_is_reno(tp))
2859 /* Yeah, we have to make difficult choice between forward transmission
2860 * and retransmission... Both ways have their merits...
2862 * For now we do not retransmit anything, while we have some new
2863 * segments to send. In the other cases, follow rule 3 for
2864 * NextSeg() specified in RFC3517.
2867 if (tcp_may_send_now(sk))
2873 /* This gets called after a retransmit timeout, and the initially
2874 * retransmitted data is acknowledged. It tries to continue
2875 * resending the rest of the retransmit queue, until either
2876 * we've sent it all or the congestion window limit is reached.
2877 * If doing SACK, the first ACK which comes back for a timeout
2878 * based retransmit packet might feed us FACK information again.
2879 * If so, we use it to avoid unnecessarily retransmissions.
2881 void tcp_xmit_retransmit_queue(struct sock *sk)
2883 const struct inet_connection_sock *icsk = inet_csk(sk);
2884 struct tcp_sock *tp = tcp_sk(sk);
2885 struct sk_buff *skb;
2886 struct sk_buff *hole = NULL;
2887 u32 max_segs, last_lost;
2889 int fwd_rexmitting = 0;
2891 if (!tp->packets_out)
2895 tp->retransmit_high = tp->snd_una;
2897 if (tp->retransmit_skb_hint) {
2898 skb = tp->retransmit_skb_hint;
2899 last_lost = TCP_SKB_CB(skb)->end_seq;
2900 if (after(last_lost, tp->retransmit_high))
2901 last_lost = tp->retransmit_high;
2903 skb = tcp_write_queue_head(sk);
2904 last_lost = tp->snd_una;
2907 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2908 tcp_for_write_queue_from(skb, sk) {
2912 if (skb == tcp_send_head(sk))
2914 /* we could do better than to assign each time */
2916 tp->retransmit_skb_hint = skb;
2918 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2921 sacked = TCP_SKB_CB(skb)->sacked;
2922 /* In case tcp_shift_skb_data() have aggregated large skbs,
2923 * we need to make sure not sending too bigs TSO packets
2925 segs = min_t(int, segs, max_segs);
2927 if (fwd_rexmitting) {
2929 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2931 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2933 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2934 tp->retransmit_high = last_lost;
2935 if (!tcp_can_forward_retransmit(sk))
2937 /* Backtrack if necessary to non-L'ed skb */
2945 } else if (!(sacked & TCPCB_LOST)) {
2946 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2951 last_lost = TCP_SKB_CB(skb)->end_seq;
2952 if (icsk->icsk_ca_state != TCP_CA_Loss)
2953 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2955 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2958 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2961 if (tcp_small_queue_check(sk, skb, 1))
2964 if (tcp_retransmit_skb(sk, skb, segs))
2967 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2969 if (tcp_in_cwnd_reduction(sk))
2970 tp->prr_out += tcp_skb_pcount(skb);
2972 if (skb == tcp_write_queue_head(sk))
2973 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2974 inet_csk(sk)->icsk_rto,
2979 /* We allow to exceed memory limits for FIN packets to expedite
2980 * connection tear down and (memory) recovery.
2981 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2982 * or even be forced to close flow without any FIN.
2983 * In general, we want to allow one skb per socket to avoid hangs
2984 * with edge trigger epoll()
2986 void sk_forced_mem_schedule(struct sock *sk, int size)
2990 if (size <= sk->sk_forward_alloc)
2992 amt = sk_mem_pages(size);
2993 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2994 sk_memory_allocated_add(sk, amt);
2996 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2997 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3000 /* Send a FIN. The caller locks the socket for us.
3001 * We should try to send a FIN packet really hard, but eventually give up.
3003 void tcp_send_fin(struct sock *sk)
3005 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3006 struct tcp_sock *tp = tcp_sk(sk);
3008 /* Optimization, tack on the FIN if we have one skb in write queue and
3009 * this skb was not yet sent, or we are under memory pressure.
3010 * Note: in the latter case, FIN packet will be sent after a timeout,
3011 * as TCP stack thinks it has already been transmitted.
3013 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3015 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3016 TCP_SKB_CB(tskb)->end_seq++;
3018 if (!tcp_send_head(sk)) {
3019 /* This means tskb was already sent.
3020 * Pretend we included the FIN on previous transmit.
3021 * We need to set tp->snd_nxt to the value it would have
3022 * if FIN had been sent. This is because retransmit path
3023 * does not change tp->snd_nxt.
3029 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3030 if (unlikely(!skb)) {
3035 skb_reserve(skb, MAX_TCP_HEADER);
3036 sk_forced_mem_schedule(sk, skb->truesize);
3037 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3038 tcp_init_nondata_skb(skb, tp->write_seq,
3039 TCPHDR_ACK | TCPHDR_FIN);
3040 tcp_queue_skb(sk, skb);
3042 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3045 /* We get here when a process closes a file descriptor (either due to
3046 * an explicit close() or as a byproduct of exit()'ing) and there
3047 * was unread data in the receive queue. This behavior is recommended
3048 * by RFC 2525, section 2.17. -DaveM
3050 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3052 struct sk_buff *skb;
3054 /* NOTE: No TCP options attached and we never retransmit this. */
3055 skb = alloc_skb(MAX_TCP_HEADER, priority);
3057 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3061 /* Reserve space for headers and prepare control bits. */
3062 skb_reserve(skb, MAX_TCP_HEADER);
3063 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3064 TCPHDR_ACK | TCPHDR_RST);
3065 skb_mstamp_get(&skb->skb_mstamp);
3067 if (tcp_transmit_skb(sk, skb, 0, priority))
3068 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3070 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3073 /* Send a crossed SYN-ACK during socket establishment.
3074 * WARNING: This routine must only be called when we have already sent
3075 * a SYN packet that crossed the incoming SYN that caused this routine
3076 * to get called. If this assumption fails then the initial rcv_wnd
3077 * and rcv_wscale values will not be correct.
3079 int tcp_send_synack(struct sock *sk)
3081 struct sk_buff *skb;
3083 skb = tcp_write_queue_head(sk);
3084 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3085 pr_debug("%s: wrong queue state\n", __func__);
3088 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3089 if (skb_cloned(skb)) {
3090 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3093 tcp_unlink_write_queue(skb, sk);
3094 __skb_header_release(nskb);
3095 __tcp_add_write_queue_head(sk, nskb);
3096 sk_wmem_free_skb(sk, skb);
3097 sk->sk_wmem_queued += nskb->truesize;
3098 sk_mem_charge(sk, nskb->truesize);
3102 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3103 tcp_ecn_send_synack(sk, skb);
3105 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3109 * tcp_make_synack - Prepare a SYN-ACK.
3110 * sk: listener socket
3111 * dst: dst entry attached to the SYNACK
3112 * req: request_sock pointer
3114 * Allocate one skb and build a SYNACK packet.
3115 * @dst is consumed : Caller should not use it again.
3117 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3118 struct request_sock *req,
3119 struct tcp_fastopen_cookie *foc,
3120 enum tcp_synack_type synack_type)
3122 struct inet_request_sock *ireq = inet_rsk(req);
3123 const struct tcp_sock *tp = tcp_sk(sk);
3124 struct tcp_md5sig_key *md5 = NULL;
3125 struct tcp_out_options opts;
3126 struct sk_buff *skb;
3127 int tcp_header_size;
3132 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3133 if (unlikely(!skb)) {
3137 /* Reserve space for headers. */
3138 skb_reserve(skb, MAX_TCP_HEADER);
3140 switch (synack_type) {
3141 case TCP_SYNACK_NORMAL:
3142 skb_set_owner_w(skb, req_to_sk(req));
3144 case TCP_SYNACK_COOKIE:
3145 /* Under synflood, we do not attach skb to a socket,
3146 * to avoid false sharing.
3149 case TCP_SYNACK_FASTOPEN:
3150 /* sk is a const pointer, because we want to express multiple
3151 * cpu might call us concurrently.
3152 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3154 skb_set_owner_w(skb, (struct sock *)sk);
3157 skb_dst_set(skb, dst);
3159 mss = dst_metric_advmss(dst);
3160 user_mss = READ_ONCE(tp->rx_opt.user_mss);
3161 if (user_mss && user_mss < mss)
3164 memset(&opts, 0, sizeof(opts));
3165 #ifdef CONFIG_SYN_COOKIES
3166 if (unlikely(req->cookie_ts))
3167 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3170 skb_mstamp_get(&skb->skb_mstamp);
3172 #ifdef CONFIG_TCP_MD5SIG
3174 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3176 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3177 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5,
3178 foc, synack_type) + sizeof(*th);
3180 skb_push(skb, tcp_header_size);
3181 skb_reset_transport_header(skb);
3183 th = (struct tcphdr *)skb->data;
3184 memset(th, 0, sizeof(struct tcphdr));
3187 tcp_ecn_make_synack(req, th);
3188 th->source = htons(ireq->ir_num);
3189 th->dest = ireq->ir_rmt_port;
3190 skb->ip_summed = CHECKSUM_PARTIAL;
3191 th->seq = htonl(tcp_rsk(req)->snt_isn);
3192 /* XXX data is queued and acked as is. No buffer/window check */
3193 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3195 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3196 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3197 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3198 th->doff = (tcp_header_size >> 2);
3199 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3201 #ifdef CONFIG_TCP_MD5SIG
3202 /* Okay, we have all we need - do the md5 hash if needed */
3204 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3205 md5, req_to_sk(req), skb);
3209 /* Do not fool tcpdump (if any), clean our debris */
3210 skb->tstamp.tv64 = 0;
3213 EXPORT_SYMBOL(tcp_make_synack);
3215 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3217 struct inet_connection_sock *icsk = inet_csk(sk);
3218 const struct tcp_congestion_ops *ca;
3219 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3221 if (ca_key == TCP_CA_UNSPEC)
3225 ca = tcp_ca_find_key(ca_key);
3226 if (likely(ca && try_module_get(ca->owner))) {
3227 module_put(icsk->icsk_ca_ops->owner);
3228 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3229 icsk->icsk_ca_ops = ca;
3234 /* Do all connect socket setups that can be done AF independent. */
3235 static void tcp_connect_init(struct sock *sk)
3237 const struct dst_entry *dst = __sk_dst_get(sk);
3238 struct tcp_sock *tp = tcp_sk(sk);
3241 /* We'll fix this up when we get a response from the other end.
3242 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3244 tp->tcp_header_len = sizeof(struct tcphdr) +
3245 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3247 #ifdef CONFIG_TCP_MD5SIG
3248 if (tp->af_specific->md5_lookup(sk, sk))
3249 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3252 /* If user gave his TCP_MAXSEG, record it to clamp */
3253 if (tp->rx_opt.user_mss)
3254 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3257 tcp_sync_mss(sk, dst_mtu(dst));
3259 tcp_ca_dst_init(sk, dst);
3261 if (!tp->window_clamp)
3262 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3263 tp->advmss = dst_metric_advmss(dst);
3264 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3265 tp->advmss = tp->rx_opt.user_mss;
3267 tcp_initialize_rcv_mss(sk);
3269 /* limit the window selection if the user enforce a smaller rx buffer */
3270 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3271 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3272 tp->window_clamp = tcp_full_space(sk);
3274 tcp_select_initial_window(tcp_full_space(sk),
3275 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3278 sysctl_tcp_window_scaling,
3280 dst_metric(dst, RTAX_INITRWND));
3282 tp->rx_opt.rcv_wscale = rcv_wscale;
3283 tp->rcv_ssthresh = tp->rcv_wnd;
3286 sock_reset_flag(sk, SOCK_DONE);
3289 tcp_write_queue_purge(sk);
3290 tp->snd_una = tp->write_seq;
3291 tp->snd_sml = tp->write_seq;
3292 tp->snd_up = tp->write_seq;
3293 tp->snd_nxt = tp->write_seq;
3295 if (likely(!tp->repair))
3298 tp->rcv_tstamp = tcp_time_stamp;
3299 tp->rcv_wup = tp->rcv_nxt;
3300 tp->copied_seq = tp->rcv_nxt;
3302 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3303 inet_csk(sk)->icsk_retransmits = 0;
3304 tcp_clear_retrans(tp);
3307 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3309 struct tcp_sock *tp = tcp_sk(sk);
3310 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3312 tcb->end_seq += skb->len;
3313 __skb_header_release(skb);
3314 __tcp_add_write_queue_tail(sk, skb);
3315 sk->sk_wmem_queued += skb->truesize;
3316 sk_mem_charge(sk, skb->truesize);
3317 tp->write_seq = tcb->end_seq;
3318 tp->packets_out += tcp_skb_pcount(skb);
3321 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3322 * queue a data-only packet after the regular SYN, such that regular SYNs
3323 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3324 * only the SYN sequence, the data are retransmitted in the first ACK.
3325 * If cookie is not cached or other error occurs, falls back to send a
3326 * regular SYN with Fast Open cookie request option.
3328 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3330 struct tcp_sock *tp = tcp_sk(sk);
3331 struct tcp_fastopen_request *fo = tp->fastopen_req;
3332 int syn_loss = 0, space, err = 0;
3333 unsigned long last_syn_loss = 0;
3334 struct sk_buff *syn_data;
3336 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3337 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3338 &syn_loss, &last_syn_loss);
3339 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3341 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3342 fo->cookie.len = -1;
3346 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3347 fo->cookie.len = -1;
3348 else if (fo->cookie.len <= 0)
3351 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3352 * user-MSS. Reserve maximum option space for middleboxes that add
3353 * private TCP options. The cost is reduced data space in SYN :(
3355 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3356 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3357 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3358 MAX_TCP_OPTION_SPACE;
3360 space = min_t(size_t, space, fo->size);
3362 /* limit to order-0 allocations */
3363 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3365 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3368 syn_data->ip_summed = CHECKSUM_PARTIAL;
3369 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3371 int copied = copy_from_iter(skb_put(syn_data, space), space,
3372 &fo->data->msg_iter);
3373 if (unlikely(!copied)) {
3374 kfree_skb(syn_data);
3377 if (copied != space) {
3378 skb_trim(syn_data, copied);
3382 /* No more data pending in inet_wait_for_connect() */
3383 if (space == fo->size)
3387 tcp_connect_queue_skb(sk, syn_data);
3389 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3391 syn->skb_mstamp = syn_data->skb_mstamp;
3393 /* Now full SYN+DATA was cloned and sent (or not),
3394 * remove the SYN from the original skb (syn_data)
3395 * we keep in write queue in case of a retransmit, as we
3396 * also have the SYN packet (with no data) in the same queue.
3398 TCP_SKB_CB(syn_data)->seq++;
3399 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3401 tp->syn_data = (fo->copied > 0);
3402 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3406 /* data was not sent, this is our new send_head */
3407 sk->sk_send_head = syn_data;
3408 tp->packets_out -= tcp_skb_pcount(syn_data);
3411 /* Send a regular SYN with Fast Open cookie request option */
3412 if (fo->cookie.len > 0)
3414 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3416 tp->syn_fastopen = 0;
3418 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3422 /* Build a SYN and send it off. */
3423 int tcp_connect(struct sock *sk)
3425 struct tcp_sock *tp = tcp_sk(sk);
3426 struct sk_buff *buff;
3429 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3430 return -EHOSTUNREACH; /* Routing failure or similar. */
3432 tcp_connect_init(sk);
3434 if (unlikely(tp->repair)) {
3435 tcp_finish_connect(sk, NULL);
3439 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3440 if (unlikely(!buff))
3443 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3444 tp->retrans_stamp = tcp_time_stamp;
3445 tcp_connect_queue_skb(sk, buff);
3446 tcp_ecn_send_syn(sk, buff);
3448 /* Send off SYN; include data in Fast Open. */
3449 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3450 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3451 if (err == -ECONNREFUSED)
3454 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3455 * in order to make this packet get counted in tcpOutSegs.
3457 tp->snd_nxt = tp->write_seq;
3458 tp->pushed_seq = tp->write_seq;
3459 buff = tcp_send_head(sk);
3460 if (unlikely(buff)) {
3461 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3462 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3464 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3466 /* Timer for repeating the SYN until an answer. */
3467 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3468 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3471 EXPORT_SYMBOL(tcp_connect);
3473 /* Send out a delayed ack, the caller does the policy checking
3474 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3477 void tcp_send_delayed_ack(struct sock *sk)
3479 struct inet_connection_sock *icsk = inet_csk(sk);
3480 int ato = icsk->icsk_ack.ato;
3481 unsigned long timeout;
3483 if (ato > TCP_DELACK_MIN) {
3484 const struct tcp_sock *tp = tcp_sk(sk);
3485 int max_ato = HZ / 2;
3487 if (icsk->icsk_ack.pingpong ||
3488 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3489 max_ato = TCP_DELACK_MAX;
3491 /* Slow path, intersegment interval is "high". */
3493 /* If some rtt estimate is known, use it to bound delayed ack.
3494 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3498 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3505 ato = min(ato, max_ato);
3508 /* Stay within the limit we were given */
3509 timeout = jiffies + ato;
3511 /* Use new timeout only if there wasn't a older one earlier. */
3512 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3513 /* If delack timer was blocked or is about to expire,
3516 if (icsk->icsk_ack.blocked ||
3517 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3522 if (!time_before(timeout, icsk->icsk_ack.timeout))
3523 timeout = icsk->icsk_ack.timeout;
3525 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3526 icsk->icsk_ack.timeout = timeout;
3527 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3530 /* This routine sends an ack and also updates the window. */
3531 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3533 struct sk_buff *buff;
3535 /* If we have been reset, we may not send again. */
3536 if (sk->sk_state == TCP_CLOSE)
3539 /* We are not putting this on the write queue, so
3540 * tcp_transmit_skb() will set the ownership to this
3543 buff = alloc_skb(MAX_TCP_HEADER,
3544 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3545 if (unlikely(!buff)) {
3546 inet_csk_schedule_ack(sk);
3547 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3548 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3549 TCP_DELACK_MAX, TCP_RTO_MAX);
3553 /* Reserve space for headers and prepare control bits. */
3554 skb_reserve(buff, MAX_TCP_HEADER);
3555 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3557 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3559 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3560 * We also avoid tcp_wfree() overhead (cache line miss accessing
3561 * tp->tsq_flags) by using regular sock_wfree()
3563 skb_set_tcp_pure_ack(buff);
3565 /* Send it off, this clears delayed acks for us. */
3566 skb_mstamp_get(&buff->skb_mstamp);
3567 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3569 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3571 void tcp_send_ack(struct sock *sk)
3573 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3576 /* This routine sends a packet with an out of date sequence
3577 * number. It assumes the other end will try to ack it.
3579 * Question: what should we make while urgent mode?
3580 * 4.4BSD forces sending single byte of data. We cannot send
3581 * out of window data, because we have SND.NXT==SND.MAX...
3583 * Current solution: to send TWO zero-length segments in urgent mode:
3584 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3585 * out-of-date with SND.UNA-1 to probe window.
3587 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3589 struct tcp_sock *tp = tcp_sk(sk);
3590 struct sk_buff *skb;
3592 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3593 skb = alloc_skb(MAX_TCP_HEADER,
3594 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3598 /* Reserve space for headers and set control bits. */
3599 skb_reserve(skb, MAX_TCP_HEADER);
3600 /* Use a previous sequence. This should cause the other
3601 * end to send an ack. Don't queue or clone SKB, just
3604 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3605 skb_mstamp_get(&skb->skb_mstamp);
3606 NET_INC_STATS(sock_net(sk), mib);
3607 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3610 void tcp_send_window_probe(struct sock *sk)
3612 if (sk->sk_state == TCP_ESTABLISHED) {
3613 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3614 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3618 /* Initiate keepalive or window probe from timer. */
3619 int tcp_write_wakeup(struct sock *sk, int mib)
3621 struct tcp_sock *tp = tcp_sk(sk);
3622 struct sk_buff *skb;
3624 if (sk->sk_state == TCP_CLOSE)
3627 skb = tcp_send_head(sk);
3628 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3630 unsigned int mss = tcp_current_mss(sk);
3631 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3633 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3634 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3636 /* We are probing the opening of a window
3637 * but the window size is != 0
3638 * must have been a result SWS avoidance ( sender )
3640 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3642 seg_size = min(seg_size, mss);
3643 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3644 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3646 } else if (!tcp_skb_pcount(skb))
3647 tcp_set_skb_tso_segs(skb, mss);
3649 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3650 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3652 tcp_event_new_data_sent(sk, skb);
3655 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3656 tcp_xmit_probe_skb(sk, 1, mib);
3657 return tcp_xmit_probe_skb(sk, 0, mib);
3661 /* A window probe timeout has occurred. If window is not closed send
3662 * a partial packet else a zero probe.
3664 void tcp_send_probe0(struct sock *sk)
3666 struct inet_connection_sock *icsk = inet_csk(sk);
3667 struct tcp_sock *tp = tcp_sk(sk);
3668 struct net *net = sock_net(sk);
3669 unsigned long probe_max;
3672 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3674 if (tp->packets_out || !tcp_send_head(sk)) {
3675 /* Cancel probe timer, if it is not required. */
3676 icsk->icsk_probes_out = 0;
3677 icsk->icsk_backoff = 0;
3682 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3683 icsk->icsk_backoff++;
3684 icsk->icsk_probes_out++;
3685 probe_max = TCP_RTO_MAX;
3687 /* If packet was not sent due to local congestion,
3688 * do not backoff and do not remember icsk_probes_out.
3689 * Let local senders to fight for local resources.
3691 * Use accumulated backoff yet.
3693 if (!icsk->icsk_probes_out)
3694 icsk->icsk_probes_out = 1;
3695 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3697 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3698 tcp_probe0_when(sk, probe_max),
3702 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3704 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3708 tcp_rsk(req)->txhash = net_tx_rndhash();
3709 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3711 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3712 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3713 if (unlikely(tcp_passive_fastopen(sk)))
3714 tcp_sk(sk)->total_retrans++;
3718 EXPORT_SYMBOL(tcp_rtx_synack);