2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
74 struct inet_connection_sock *icsk = inet_csk(sk);
75 struct tcp_sock *tp = tcp_sk(sk);
76 unsigned int prior_packets = tp->packets_out;
78 tcp_advance_send_head(sk, skb);
79 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
81 tp->packets_out += tcp_skb_pcount(skb);
82 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
83 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
87 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
99 const struct tcp_sock *tp = tcp_sk(sk);
101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
143 struct tcp_sock *tp = tcp_sk(sk);
144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
166 const struct dst_entry *dst = __sk_dst_get(sk);
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 icsk->icsk_ack.pingpong = 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 u32 tcp_default_init_rwnd(u32 mss)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd = TCP_INIT_CWND * 2;
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
213 int wscale_ok, __u8 *rcv_wscale,
216 unsigned int space = (__space < 0 ? 0 : __space);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (65535 << 14);
221 space = min(*window_clamp, space);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space = (space / mss) * mss;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space = max_t(u32, sysctl_tcp_rmem[2], 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,
366 if (inet_rsk(req)->ecn_ok) {
368 if (tcp_ca_needs_ecn(sk))
373 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
376 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
379 struct tcp_sock *tp = tcp_sk(sk);
381 if (tp->ecn_flags & TCP_ECN_OK) {
382 /* Not-retransmitted data segment: set ECT and inject CWR. */
383 if (skb->len != tcp_header_len &&
384 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
386 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
387 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
388 tcp_hdr(skb)->cwr = 1;
389 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
391 } else if (!tcp_ca_needs_ecn(sk)) {
392 /* ACK or retransmitted segment: clear ECT|CE */
393 INET_ECN_dontxmit(sk);
395 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
396 tcp_hdr(skb)->ece = 1;
400 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
401 * auto increment end seqno.
403 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
405 struct skb_shared_info *shinfo = skb_shinfo(skb);
407 skb->ip_summed = CHECKSUM_PARTIAL;
410 TCP_SKB_CB(skb)->tcp_flags = flags;
411 TCP_SKB_CB(skb)->sacked = 0;
413 tcp_skb_pcount_set(skb, 1);
414 shinfo->gso_size = 0;
415 shinfo->gso_type = 0;
417 TCP_SKB_CB(skb)->seq = seq;
418 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
420 TCP_SKB_CB(skb)->end_seq = seq;
423 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
425 return tp->snd_una != tp->snd_up;
428 #define OPTION_SACK_ADVERTISE (1 << 0)
429 #define OPTION_TS (1 << 1)
430 #define OPTION_MD5 (1 << 2)
431 #define OPTION_WSCALE (1 << 3)
432 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
434 struct tcp_out_options {
435 u16 options; /* bit field of OPTION_* */
436 u16 mss; /* 0 to disable */
437 u8 ws; /* window scale, 0 to disable */
438 u8 num_sack_blocks; /* number of SACK blocks to include */
439 u8 hash_size; /* bytes in hash_location */
440 __u8 *hash_location; /* temporary pointer, overloaded */
441 __u32 tsval, tsecr; /* need to include OPTION_TS */
442 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
445 /* Write previously computed TCP options to the packet.
447 * Beware: Something in the Internet is very sensitive to the ordering of
448 * TCP options, we learned this through the hard way, so be careful here.
449 * Luckily we can at least blame others for their non-compliance but from
450 * inter-operability perspective it seems that we're somewhat stuck with
451 * the ordering which we have been using if we want to keep working with
452 * those broken things (not that it currently hurts anybody as there isn't
453 * particular reason why the ordering would need to be changed).
455 * At least SACK_PERM as the first option is known to lead to a disaster
456 * (but it may well be that other scenarios fail similarly).
458 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
459 struct tcp_out_options *opts)
461 u16 options = opts->options; /* mungable copy */
463 if (unlikely(OPTION_MD5 & options)) {
464 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
465 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
466 /* overload cookie hash location */
467 opts->hash_location = (__u8 *)ptr;
471 if (unlikely(opts->mss)) {
472 *ptr++ = htonl((TCPOPT_MSS << 24) |
473 (TCPOLEN_MSS << 16) |
477 if (likely(OPTION_TS & options)) {
478 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
479 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
480 (TCPOLEN_SACK_PERM << 16) |
481 (TCPOPT_TIMESTAMP << 8) |
483 options &= ~OPTION_SACK_ADVERTISE;
485 *ptr++ = htonl((TCPOPT_NOP << 24) |
487 (TCPOPT_TIMESTAMP << 8) |
490 *ptr++ = htonl(opts->tsval);
491 *ptr++ = htonl(opts->tsecr);
494 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_SACK_PERM << 8) |
501 if (unlikely(OPTION_WSCALE & options)) {
502 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOPT_WINDOW << 16) |
504 (TCPOLEN_WINDOW << 8) |
508 if (unlikely(opts->num_sack_blocks)) {
509 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
510 tp->duplicate_sack : tp->selective_acks;
513 *ptr++ = htonl((TCPOPT_NOP << 24) |
516 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
517 TCPOLEN_SACK_PERBLOCK)));
519 for (this_sack = 0; this_sack < opts->num_sack_blocks;
521 *ptr++ = htonl(sp[this_sack].start_seq);
522 *ptr++ = htonl(sp[this_sack].end_seq);
525 tp->rx_opt.dsack = 0;
528 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
529 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
531 u32 len; /* Fast Open option length */
534 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
535 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
536 TCPOPT_FASTOPEN_MAGIC);
537 p += TCPOLEN_EXP_FASTOPEN_BASE;
539 len = TCPOLEN_FASTOPEN_BASE + foc->len;
540 *p++ = TCPOPT_FASTOPEN;
544 memcpy(p, foc->val, foc->len);
545 if ((len & 3) == 2) {
546 p[foc->len] = TCPOPT_NOP;
547 p[foc->len + 1] = TCPOPT_NOP;
549 ptr += (len + 3) >> 2;
553 /* Compute TCP options for SYN packets. This is not the final
554 * network wire format yet.
556 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
557 struct tcp_out_options *opts,
558 struct tcp_md5sig_key **md5)
560 struct tcp_sock *tp = tcp_sk(sk);
561 unsigned int remaining = MAX_TCP_OPTION_SPACE;
562 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
564 #ifdef CONFIG_TCP_MD5SIG
565 *md5 = tp->af_specific->md5_lookup(sk, sk);
567 opts->options |= OPTION_MD5;
568 remaining -= TCPOLEN_MD5SIG_ALIGNED;
574 /* We always get an MSS option. The option bytes which will be seen in
575 * normal data packets should timestamps be used, must be in the MSS
576 * advertised. But we subtract them from tp->mss_cache so that
577 * calculations in tcp_sendmsg are simpler etc. So account for this
578 * fact here if necessary. If we don't do this correctly, as a
579 * receiver we won't recognize data packets as being full sized when we
580 * should, and thus we won't abide by the delayed ACK rules correctly.
581 * SACKs don't matter, we never delay an ACK when we have any of those
583 opts->mss = tcp_advertise_mss(sk);
584 remaining -= TCPOLEN_MSS_ALIGNED;
586 if (likely(sysctl_tcp_timestamps && !*md5)) {
587 opts->options |= OPTION_TS;
588 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
589 opts->tsecr = tp->rx_opt.ts_recent;
590 remaining -= TCPOLEN_TSTAMP_ALIGNED;
592 if (likely(sysctl_tcp_window_scaling)) {
593 opts->ws = tp->rx_opt.rcv_wscale;
594 opts->options |= OPTION_WSCALE;
595 remaining -= TCPOLEN_WSCALE_ALIGNED;
597 if (likely(sysctl_tcp_sack)) {
598 opts->options |= OPTION_SACK_ADVERTISE;
599 if (unlikely(!(OPTION_TS & opts->options)))
600 remaining -= TCPOLEN_SACKPERM_ALIGNED;
603 if (fastopen && fastopen->cookie.len >= 0) {
604 u32 need = fastopen->cookie.len;
606 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
607 TCPOLEN_FASTOPEN_BASE;
608 need = (need + 3) & ~3U; /* Align to 32 bits */
609 if (remaining >= need) {
610 opts->options |= OPTION_FAST_OPEN_COOKIE;
611 opts->fastopen_cookie = &fastopen->cookie;
613 tp->syn_fastopen = 1;
614 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
618 return MAX_TCP_OPTION_SPACE - remaining;
621 /* Set up TCP options for SYN-ACKs. */
622 static unsigned int tcp_synack_options(struct sock *sk,
623 struct request_sock *req,
624 unsigned int mss, struct sk_buff *skb,
625 struct tcp_out_options *opts,
626 const struct tcp_md5sig_key *md5,
627 struct tcp_fastopen_cookie *foc)
629 struct inet_request_sock *ireq = inet_rsk(req);
630 unsigned int remaining = MAX_TCP_OPTION_SPACE;
632 #ifdef CONFIG_TCP_MD5SIG
634 opts->options |= OPTION_MD5;
635 remaining -= TCPOLEN_MD5SIG_ALIGNED;
637 /* We can't fit any SACK blocks in a packet with MD5 + TS
638 * options. There was discussion about disabling SACK
639 * rather than TS in order to fit in better with old,
640 * buggy kernels, but that was deemed to be unnecessary.
642 ireq->tstamp_ok &= !ireq->sack_ok;
646 /* We always send an MSS option. */
648 remaining -= TCPOLEN_MSS_ALIGNED;
650 if (likely(ireq->wscale_ok)) {
651 opts->ws = ireq->rcv_wscale;
652 opts->options |= OPTION_WSCALE;
653 remaining -= TCPOLEN_WSCALE_ALIGNED;
655 if (likely(ireq->tstamp_ok)) {
656 opts->options |= OPTION_TS;
657 opts->tsval = tcp_skb_timestamp(skb);
658 opts->tsecr = req->ts_recent;
659 remaining -= TCPOLEN_TSTAMP_ALIGNED;
661 if (likely(ireq->sack_ok)) {
662 opts->options |= OPTION_SACK_ADVERTISE;
663 if (unlikely(!ireq->tstamp_ok))
664 remaining -= TCPOLEN_SACKPERM_ALIGNED;
666 if (foc != NULL && foc->len >= 0) {
669 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
670 TCPOLEN_FASTOPEN_BASE;
671 need = (need + 3) & ~3U; /* Align to 32 bits */
672 if (remaining >= need) {
673 opts->options |= OPTION_FAST_OPEN_COOKIE;
674 opts->fastopen_cookie = foc;
679 return MAX_TCP_OPTION_SPACE - remaining;
682 /* Compute TCP options for ESTABLISHED sockets. This is not the
683 * final wire format yet.
685 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
686 struct tcp_out_options *opts,
687 struct tcp_md5sig_key **md5)
689 struct tcp_sock *tp = tcp_sk(sk);
690 unsigned int size = 0;
691 unsigned int eff_sacks;
695 #ifdef CONFIG_TCP_MD5SIG
696 *md5 = tp->af_specific->md5_lookup(sk, sk);
697 if (unlikely(*md5)) {
698 opts->options |= OPTION_MD5;
699 size += TCPOLEN_MD5SIG_ALIGNED;
705 if (likely(tp->rx_opt.tstamp_ok)) {
706 opts->options |= OPTION_TS;
707 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
708 opts->tsecr = tp->rx_opt.ts_recent;
709 size += TCPOLEN_TSTAMP_ALIGNED;
712 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
713 if (unlikely(eff_sacks)) {
714 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
715 opts->num_sack_blocks =
716 min_t(unsigned int, eff_sacks,
717 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
718 TCPOLEN_SACK_PERBLOCK);
719 size += TCPOLEN_SACK_BASE_ALIGNED +
720 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
727 /* TCP SMALL QUEUES (TSQ)
729 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
730 * to reduce RTT and bufferbloat.
731 * We do this using a special skb destructor (tcp_wfree).
733 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
734 * needs to be reallocated in a driver.
735 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
737 * Since transmit from skb destructor is forbidden, we use a tasklet
738 * to process all sockets that eventually need to send more skbs.
739 * We use one tasklet per cpu, with its own queue of sockets.
742 struct tasklet_struct tasklet;
743 struct list_head head; /* queue of tcp sockets */
745 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
747 static void tcp_tsq_handler(struct sock *sk)
749 if ((1 << sk->sk_state) &
750 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
751 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
752 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
756 * One tasklet per cpu tries to send more skbs.
757 * We run in tasklet context but need to disable irqs when
758 * transferring tsq->head because tcp_wfree() might
759 * interrupt us (non NAPI drivers)
761 static void tcp_tasklet_func(unsigned long data)
763 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
766 struct list_head *q, *n;
770 local_irq_save(flags);
771 list_splice_init(&tsq->head, &list);
772 local_irq_restore(flags);
774 list_for_each_safe(q, n, &list) {
775 tp = list_entry(q, struct tcp_sock, tsq_node);
776 list_del(&tp->tsq_node);
778 sk = (struct sock *)tp;
781 if (!sock_owned_by_user(sk)) {
784 /* defer the work to tcp_release_cb() */
785 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
789 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
794 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
795 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
796 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
797 (1UL << TCP_MTU_REDUCED_DEFERRED))
799 * tcp_release_cb - tcp release_sock() callback
802 * called from release_sock() to perform protocol dependent
803 * actions before socket release.
805 void tcp_release_cb(struct sock *sk)
807 struct tcp_sock *tp = tcp_sk(sk);
808 unsigned long flags, nflags;
810 /* perform an atomic operation only if at least one flag is set */
812 flags = tp->tsq_flags;
813 if (!(flags & TCP_DEFERRED_ALL))
815 nflags = flags & ~TCP_DEFERRED_ALL;
816 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
818 if (flags & (1UL << TCP_TSQ_DEFERRED))
821 /* Here begins the tricky part :
822 * We are called from release_sock() with :
824 * 2) sk_lock.slock spinlock held
825 * 3) socket owned by us (sk->sk_lock.owned == 1)
827 * But following code is meant to be called from BH handlers,
828 * so we should keep BH disabled, but early release socket ownership
830 sock_release_ownership(sk);
832 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
833 tcp_write_timer_handler(sk);
836 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
837 tcp_delack_timer_handler(sk);
840 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
841 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
845 EXPORT_SYMBOL(tcp_release_cb);
847 void __init tcp_tasklet_init(void)
851 for_each_possible_cpu(i) {
852 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
854 INIT_LIST_HEAD(&tsq->head);
855 tasklet_init(&tsq->tasklet,
862 * Write buffer destructor automatically called from kfree_skb.
863 * We can't xmit new skbs from this context, as we might already
866 void tcp_wfree(struct sk_buff *skb)
868 struct sock *sk = skb->sk;
869 struct tcp_sock *tp = tcp_sk(sk);
872 /* Keep one reference on sk_wmem_alloc.
873 * Will be released by sk_free() from here or tcp_tasklet_func()
875 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
877 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
878 * Wait until our queues (qdisc + devices) are drained.
880 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
881 * - chance for incoming ACK (processed by another cpu maybe)
882 * to migrate this flow (skb->ooo_okay will be eventually set)
884 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
887 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
888 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
890 struct tsq_tasklet *tsq;
892 /* queue this socket to tasklet queue */
893 local_irq_save(flags);
894 tsq = this_cpu_ptr(&tsq_tasklet);
895 list_add(&tp->tsq_node, &tsq->head);
896 tasklet_schedule(&tsq->tasklet);
897 local_irq_restore(flags);
904 /* This routine actually transmits TCP packets queued in by
905 * tcp_do_sendmsg(). This is used by both the initial
906 * transmission and possible later retransmissions.
907 * All SKB's seen here are completely headerless. It is our
908 * job to build the TCP header, and pass the packet down to
909 * IP so it can do the same plus pass the packet off to the
912 * We are working here with either a clone of the original
913 * SKB, or a fresh unique copy made by the retransmit engine.
915 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
918 const struct inet_connection_sock *icsk = inet_csk(sk);
919 struct inet_sock *inet;
921 struct tcp_skb_cb *tcb;
922 struct tcp_out_options opts;
923 unsigned int tcp_options_size, tcp_header_size;
924 struct tcp_md5sig_key *md5;
928 BUG_ON(!skb || !tcp_skb_pcount(skb));
931 skb_mstamp_get(&skb->skb_mstamp);
933 if (unlikely(skb_cloned(skb)))
934 skb = pskb_copy(skb, gfp_mask);
936 skb = skb_clone(skb, gfp_mask);
943 tcb = TCP_SKB_CB(skb);
944 memset(&opts, 0, sizeof(opts));
946 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
947 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
949 tcp_options_size = tcp_established_options(sk, skb, &opts,
951 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
953 if (tcp_packets_in_flight(tp) == 0)
954 tcp_ca_event(sk, CA_EVENT_TX_START);
956 /* if no packet is in qdisc/device queue, then allow XPS to select
957 * another queue. We can be called from tcp_tsq_handler()
958 * which holds one reference to sk_wmem_alloc.
960 * TODO: Ideally, in-flight pure ACK packets should not matter here.
961 * One way to get this would be to set skb->truesize = 2 on them.
963 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
965 skb_push(skb, tcp_header_size);
966 skb_reset_transport_header(skb);
970 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
971 skb_set_hash_from_sk(skb, sk);
972 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
974 /* Build TCP header and checksum it. */
976 th->source = inet->inet_sport;
977 th->dest = inet->inet_dport;
978 th->seq = htonl(tcb->seq);
979 th->ack_seq = htonl(tp->rcv_nxt);
980 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
983 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
984 /* RFC1323: The window in SYN & SYN/ACK segments
987 th->window = htons(min(tp->rcv_wnd, 65535U));
989 th->window = htons(tcp_select_window(sk));
994 /* The urg_mode check is necessary during a below snd_una win probe */
995 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
996 if (before(tp->snd_up, tcb->seq + 0x10000)) {
997 th->urg_ptr = htons(tp->snd_up - tcb->seq);
999 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1000 th->urg_ptr = htons(0xFFFF);
1005 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1006 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1007 tcp_ecn_send(sk, skb, tcp_header_size);
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));
1023 if (skb->len != tcp_header_size)
1024 tcp_event_data_sent(tp, sk);
1026 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1027 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1028 tcp_skb_pcount(skb));
1030 tp->segs_out += tcp_skb_pcount(skb);
1031 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
1032 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1034 /* Our usage of tstamp should remain private */
1035 skb->tstamp.tv64 = 0;
1037 /* Cleanup our debris for IP stacks */
1038 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1039 sizeof(struct inet6_skb_parm)));
1041 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1043 if (likely(err <= 0))
1048 return net_xmit_eval(err);
1051 /* This routine just queues the buffer for sending.
1053 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1054 * otherwise socket can stall.
1056 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1058 struct tcp_sock *tp = tcp_sk(sk);
1060 /* Advance write_seq and place onto the write_queue. */
1061 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1062 __skb_header_release(skb);
1063 tcp_add_write_queue_tail(sk, skb);
1064 sk->sk_wmem_queued += skb->truesize;
1065 sk_mem_charge(sk, skb->truesize);
1068 /* Initialize TSO segments for a packet. */
1069 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1070 unsigned int mss_now)
1072 struct skb_shared_info *shinfo = skb_shinfo(skb);
1074 /* Make sure we own this skb before messing gso_size/gso_segs */
1075 WARN_ON_ONCE(skb_cloned(skb));
1077 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1078 /* Avoid the costly divide in the normal
1081 tcp_skb_pcount_set(skb, 1);
1082 shinfo->gso_size = 0;
1083 shinfo->gso_type = 0;
1085 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1086 shinfo->gso_size = mss_now;
1087 shinfo->gso_type = sk->sk_gso_type;
1091 /* When a modification to fackets out becomes necessary, we need to check
1092 * skb is counted to fackets_out or not.
1094 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1097 struct tcp_sock *tp = tcp_sk(sk);
1099 if (!tp->sacked_out || tcp_is_reno(tp))
1102 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1103 tp->fackets_out -= decr;
1106 /* Pcount in the middle of the write queue got changed, we need to do various
1107 * tweaks to fix counters
1109 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1111 struct tcp_sock *tp = tcp_sk(sk);
1113 tp->packets_out -= decr;
1115 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1116 tp->sacked_out -= decr;
1117 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1118 tp->retrans_out -= decr;
1119 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1120 tp->lost_out -= decr;
1122 /* Reno case is special. Sigh... */
1123 if (tcp_is_reno(tp) && decr > 0)
1124 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1126 tcp_adjust_fackets_out(sk, skb, decr);
1128 if (tp->lost_skb_hint &&
1129 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1130 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1131 tp->lost_cnt_hint -= decr;
1133 tcp_verify_left_out(tp);
1136 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1138 struct skb_shared_info *shinfo = skb_shinfo(skb);
1140 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1141 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1142 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1143 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1145 shinfo->tx_flags &= ~tsflags;
1146 shinfo2->tx_flags |= tsflags;
1147 swap(shinfo->tskey, shinfo2->tskey);
1151 /* Function to create two new TCP segments. Shrinks the given segment
1152 * to the specified size and appends a new segment with the rest of the
1153 * packet to the list. This won't be called frequently, I hope.
1154 * Remember, these are still headerless SKBs at this point.
1156 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1157 unsigned int mss_now, gfp_t gfp)
1159 struct tcp_sock *tp = tcp_sk(sk);
1160 struct sk_buff *buff;
1161 int nsize, old_factor;
1165 if (WARN_ON(len > skb->len))
1168 nsize = skb_headlen(skb) - len;
1172 if (skb_unclone(skb, gfp))
1175 /* Get a new skb... force flag on. */
1176 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1178 return -ENOMEM; /* We'll just try again later. */
1180 sk->sk_wmem_queued += buff->truesize;
1181 sk_mem_charge(sk, buff->truesize);
1182 nlen = skb->len - len - nsize;
1183 buff->truesize += nlen;
1184 skb->truesize -= nlen;
1186 /* Correct the sequence numbers. */
1187 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1188 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1189 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1191 /* PSH and FIN should only be set in the second packet. */
1192 flags = TCP_SKB_CB(skb)->tcp_flags;
1193 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1194 TCP_SKB_CB(buff)->tcp_flags = flags;
1195 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1197 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1198 /* Copy and checksum data tail into the new buffer. */
1199 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1200 skb_put(buff, nsize),
1205 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1207 skb->ip_summed = CHECKSUM_PARTIAL;
1208 skb_split(skb, buff, len);
1211 buff->ip_summed = skb->ip_summed;
1213 buff->tstamp = skb->tstamp;
1214 tcp_fragment_tstamp(skb, buff);
1216 old_factor = tcp_skb_pcount(skb);
1218 /* Fix up tso_factor for both original and new SKB. */
1219 tcp_set_skb_tso_segs(sk, skb, mss_now);
1220 tcp_set_skb_tso_segs(sk, buff, mss_now);
1222 /* If this packet has been sent out already, we must
1223 * adjust the various packet counters.
1225 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1226 int diff = old_factor - tcp_skb_pcount(skb) -
1227 tcp_skb_pcount(buff);
1230 tcp_adjust_pcount(sk, skb, diff);
1233 /* Link BUFF into the send queue. */
1234 __skb_header_release(buff);
1235 tcp_insert_write_queue_after(skb, buff, sk);
1240 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1241 * eventually). The difference is that pulled data not copied, but
1242 * immediately discarded.
1244 static void __pskb_trim_head(struct sk_buff *skb, int len)
1246 struct skb_shared_info *shinfo;
1249 eat = min_t(int, len, skb_headlen(skb));
1251 __skb_pull(skb, eat);
1258 shinfo = skb_shinfo(skb);
1259 for (i = 0; i < shinfo->nr_frags; i++) {
1260 int size = skb_frag_size(&shinfo->frags[i]);
1263 skb_frag_unref(skb, i);
1266 shinfo->frags[k] = shinfo->frags[i];
1268 shinfo->frags[k].page_offset += eat;
1269 skb_frag_size_sub(&shinfo->frags[k], eat);
1275 shinfo->nr_frags = k;
1277 skb_reset_tail_pointer(skb);
1278 skb->data_len -= len;
1279 skb->len = skb->data_len;
1282 /* Remove acked data from a packet in the transmit queue. */
1283 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1285 if (skb_unclone(skb, GFP_ATOMIC))
1288 __pskb_trim_head(skb, len);
1290 TCP_SKB_CB(skb)->seq += len;
1291 skb->ip_summed = CHECKSUM_PARTIAL;
1293 skb->truesize -= len;
1294 sk->sk_wmem_queued -= len;
1295 sk_mem_uncharge(sk, len);
1296 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1298 /* Any change of skb->len requires recalculation of tso factor. */
1299 if (tcp_skb_pcount(skb) > 1)
1300 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1305 /* Calculate MSS not accounting any TCP options. */
1306 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1308 const struct tcp_sock *tp = tcp_sk(sk);
1309 const struct inet_connection_sock *icsk = inet_csk(sk);
1312 /* Calculate base mss without TCP options:
1313 It is MMS_S - sizeof(tcphdr) of rfc1122
1315 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1317 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1318 if (icsk->icsk_af_ops->net_frag_header_len) {
1319 const struct dst_entry *dst = __sk_dst_get(sk);
1321 if (dst && dst_allfrag(dst))
1322 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1325 /* Clamp it (mss_clamp does not include tcp options) */
1326 if (mss_now > tp->rx_opt.mss_clamp)
1327 mss_now = tp->rx_opt.mss_clamp;
1329 /* Now subtract optional transport overhead */
1330 mss_now -= icsk->icsk_ext_hdr_len;
1332 /* Then reserve room for full set of TCP options and 8 bytes of data */
1338 /* Calculate MSS. Not accounting for SACKs here. */
1339 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1341 /* Subtract TCP options size, not including SACKs */
1342 return __tcp_mtu_to_mss(sk, pmtu) -
1343 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1346 /* Inverse of above */
1347 int tcp_mss_to_mtu(struct sock *sk, int mss)
1349 const struct tcp_sock *tp = tcp_sk(sk);
1350 const struct inet_connection_sock *icsk = inet_csk(sk);
1354 tp->tcp_header_len +
1355 icsk->icsk_ext_hdr_len +
1356 icsk->icsk_af_ops->net_header_len;
1358 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1359 if (icsk->icsk_af_ops->net_frag_header_len) {
1360 const struct dst_entry *dst = __sk_dst_get(sk);
1362 if (dst && dst_allfrag(dst))
1363 mtu += icsk->icsk_af_ops->net_frag_header_len;
1368 /* MTU probing init per socket */
1369 void tcp_mtup_init(struct sock *sk)
1371 struct tcp_sock *tp = tcp_sk(sk);
1372 struct inet_connection_sock *icsk = inet_csk(sk);
1373 struct net *net = sock_net(sk);
1375 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1376 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1377 icsk->icsk_af_ops->net_header_len;
1378 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1379 icsk->icsk_mtup.probe_size = 0;
1380 if (icsk->icsk_mtup.enabled)
1381 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1383 EXPORT_SYMBOL(tcp_mtup_init);
1385 /* This function synchronize snd mss to current pmtu/exthdr set.
1387 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1388 for TCP options, but includes only bare TCP header.
1390 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1391 It is minimum of user_mss and mss received with SYN.
1392 It also does not include TCP options.
1394 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1396 tp->mss_cache is current effective sending mss, including
1397 all tcp options except for SACKs. It is evaluated,
1398 taking into account current pmtu, but never exceeds
1399 tp->rx_opt.mss_clamp.
1401 NOTE1. rfc1122 clearly states that advertised MSS
1402 DOES NOT include either tcp or ip options.
1404 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1405 are READ ONLY outside this function. --ANK (980731)
1407 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1409 struct tcp_sock *tp = tcp_sk(sk);
1410 struct inet_connection_sock *icsk = inet_csk(sk);
1413 if (icsk->icsk_mtup.search_high > pmtu)
1414 icsk->icsk_mtup.search_high = pmtu;
1416 mss_now = tcp_mtu_to_mss(sk, pmtu);
1417 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1419 /* And store cached results */
1420 icsk->icsk_pmtu_cookie = pmtu;
1421 if (icsk->icsk_mtup.enabled)
1422 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1423 tp->mss_cache = mss_now;
1427 EXPORT_SYMBOL(tcp_sync_mss);
1429 /* Compute the current effective MSS, taking SACKs and IP options,
1430 * and even PMTU discovery events into account.
1432 unsigned int tcp_current_mss(struct sock *sk)
1434 const struct tcp_sock *tp = tcp_sk(sk);
1435 const struct dst_entry *dst = __sk_dst_get(sk);
1437 unsigned int header_len;
1438 struct tcp_out_options opts;
1439 struct tcp_md5sig_key *md5;
1441 mss_now = tp->mss_cache;
1444 u32 mtu = dst_mtu(dst);
1445 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1446 mss_now = tcp_sync_mss(sk, mtu);
1449 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1450 sizeof(struct tcphdr);
1451 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1452 * some common options. If this is an odd packet (because we have SACK
1453 * blocks etc) then our calculated header_len will be different, and
1454 * we have to adjust mss_now correspondingly */
1455 if (header_len != tp->tcp_header_len) {
1456 int delta = (int) header_len - tp->tcp_header_len;
1463 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1464 * As additional protections, we do not touch cwnd in retransmission phases,
1465 * and if application hit its sndbuf limit recently.
1467 static void tcp_cwnd_application_limited(struct sock *sk)
1469 struct tcp_sock *tp = tcp_sk(sk);
1471 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1472 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1473 /* Limited by application or receiver window. */
1474 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1475 u32 win_used = max(tp->snd_cwnd_used, init_win);
1476 if (win_used < tp->snd_cwnd) {
1477 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1478 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1480 tp->snd_cwnd_used = 0;
1482 tp->snd_cwnd_stamp = tcp_time_stamp;
1485 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1487 struct tcp_sock *tp = tcp_sk(sk);
1489 /* Track the maximum number of outstanding packets in each
1490 * window, and remember whether we were cwnd-limited then.
1492 if (!before(tp->snd_una, tp->max_packets_seq) ||
1493 tp->packets_out > tp->max_packets_out) {
1494 tp->max_packets_out = tp->packets_out;
1495 tp->max_packets_seq = tp->snd_nxt;
1496 tp->is_cwnd_limited = is_cwnd_limited;
1499 if (tcp_is_cwnd_limited(sk)) {
1500 /* Network is feed fully. */
1501 tp->snd_cwnd_used = 0;
1502 tp->snd_cwnd_stamp = tcp_time_stamp;
1504 /* Network starves. */
1505 if (tp->packets_out > tp->snd_cwnd_used)
1506 tp->snd_cwnd_used = tp->packets_out;
1508 if (sysctl_tcp_slow_start_after_idle &&
1509 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1510 tcp_cwnd_application_limited(sk);
1514 /* Minshall's variant of the Nagle send check. */
1515 static bool tcp_minshall_check(const struct tcp_sock *tp)
1517 return after(tp->snd_sml, tp->snd_una) &&
1518 !after(tp->snd_sml, tp->snd_nxt);
1521 /* Update snd_sml if this skb is under mss
1522 * Note that a TSO packet might end with a sub-mss segment
1523 * The test is really :
1524 * if ((skb->len % mss) != 0)
1525 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1526 * But we can avoid doing the divide again given we already have
1527 * skb_pcount = skb->len / mss_now
1529 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1530 const struct sk_buff *skb)
1532 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1533 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1536 /* Return false, if packet can be sent now without violation Nagle's rules:
1537 * 1. It is full sized. (provided by caller in %partial bool)
1538 * 2. Or it contains FIN. (already checked by caller)
1539 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1540 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1541 * With Minshall's modification: all sent small packets are ACKed.
1543 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1547 ((nonagle & TCP_NAGLE_CORK) ||
1548 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1551 /* Return how many segs we'd like on a TSO packet,
1552 * to send one TSO packet per ms
1554 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1558 bytes = min(sk->sk_pacing_rate >> 10,
1559 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1561 /* Goal is to send at least one packet per ms,
1562 * not one big TSO packet every 100 ms.
1563 * This preserves ACK clocking and is consistent
1564 * with tcp_tso_should_defer() heuristic.
1566 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1568 return min_t(u32, segs, sk->sk_gso_max_segs);
1571 /* Returns the portion of skb which can be sent right away */
1572 static unsigned int tcp_mss_split_point(const struct sock *sk,
1573 const struct sk_buff *skb,
1574 unsigned int mss_now,
1575 unsigned int max_segs,
1578 const struct tcp_sock *tp = tcp_sk(sk);
1579 u32 partial, needed, window, max_len;
1581 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1582 max_len = mss_now * max_segs;
1584 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1587 needed = min(skb->len, window);
1589 if (max_len <= needed)
1592 partial = needed % mss_now;
1593 /* If last segment is not a full MSS, check if Nagle rules allow us
1594 * to include this last segment in this skb.
1595 * Otherwise, we'll split the skb at last MSS boundary
1597 if (tcp_nagle_check(partial != 0, tp, nonagle))
1598 return needed - partial;
1603 /* Can at least one segment of SKB be sent right now, according to the
1604 * congestion window rules? If so, return how many segments are allowed.
1606 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1607 const struct sk_buff *skb)
1609 u32 in_flight, cwnd, halfcwnd;
1611 /* Don't be strict about the congestion window for the final FIN. */
1612 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1613 tcp_skb_pcount(skb) == 1)
1616 in_flight = tcp_packets_in_flight(tp);
1617 cwnd = tp->snd_cwnd;
1618 if (in_flight >= cwnd)
1621 /* For better scheduling, ensure we have at least
1622 * 2 GSO packets in flight.
1624 halfcwnd = max(cwnd >> 1, 1U);
1625 return min(halfcwnd, cwnd - in_flight);
1628 /* Initialize TSO state of a skb.
1629 * This must be invoked the first time we consider transmitting
1630 * SKB onto the wire.
1632 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1633 unsigned int mss_now)
1635 int tso_segs = tcp_skb_pcount(skb);
1637 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1638 tcp_set_skb_tso_segs(sk, skb, mss_now);
1639 tso_segs = tcp_skb_pcount(skb);
1645 /* Return true if the Nagle test allows this packet to be
1648 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1649 unsigned int cur_mss, int nonagle)
1651 /* Nagle rule does not apply to frames, which sit in the middle of the
1652 * write_queue (they have no chances to get new data).
1654 * This is implemented in the callers, where they modify the 'nonagle'
1655 * argument based upon the location of SKB in the send queue.
1657 if (nonagle & TCP_NAGLE_PUSH)
1660 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1661 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1664 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1670 /* Does at least the first segment of SKB fit into the send window? */
1671 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1672 const struct sk_buff *skb,
1673 unsigned int cur_mss)
1675 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1677 if (skb->len > cur_mss)
1678 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1680 return !after(end_seq, tcp_wnd_end(tp));
1683 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1684 * should be put on the wire right now. If so, it returns the number of
1685 * packets allowed by the congestion window.
1687 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1688 unsigned int cur_mss, int nonagle)
1690 const struct tcp_sock *tp = tcp_sk(sk);
1691 unsigned int cwnd_quota;
1693 tcp_init_tso_segs(sk, skb, cur_mss);
1695 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1698 cwnd_quota = tcp_cwnd_test(tp, skb);
1699 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1705 /* Test if sending is allowed right now. */
1706 bool tcp_may_send_now(struct sock *sk)
1708 const struct tcp_sock *tp = tcp_sk(sk);
1709 struct sk_buff *skb = tcp_send_head(sk);
1712 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1713 (tcp_skb_is_last(sk, skb) ?
1714 tp->nonagle : TCP_NAGLE_PUSH));
1717 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1718 * which is put after SKB on the list. It is very much like
1719 * tcp_fragment() except that it may make several kinds of assumptions
1720 * in order to speed up the splitting operation. In particular, we
1721 * know that all the data is in scatter-gather pages, and that the
1722 * packet has never been sent out before (and thus is not cloned).
1724 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1725 unsigned int mss_now, gfp_t gfp)
1727 struct sk_buff *buff;
1728 int nlen = skb->len - len;
1731 /* All of a TSO frame must be composed of paged data. */
1732 if (skb->len != skb->data_len)
1733 return tcp_fragment(sk, skb, len, mss_now, gfp);
1735 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1736 if (unlikely(!buff))
1739 sk->sk_wmem_queued += buff->truesize;
1740 sk_mem_charge(sk, buff->truesize);
1741 buff->truesize += nlen;
1742 skb->truesize -= nlen;
1744 /* Correct the sequence numbers. */
1745 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1746 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1747 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1749 /* PSH and FIN should only be set in the second packet. */
1750 flags = TCP_SKB_CB(skb)->tcp_flags;
1751 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1752 TCP_SKB_CB(buff)->tcp_flags = flags;
1754 /* This packet was never sent out yet, so no SACK bits. */
1755 TCP_SKB_CB(buff)->sacked = 0;
1757 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1758 skb_split(skb, buff, len);
1759 tcp_fragment_tstamp(skb, buff);
1761 /* Fix up tso_factor for both original and new SKB. */
1762 tcp_set_skb_tso_segs(sk, skb, mss_now);
1763 tcp_set_skb_tso_segs(sk, buff, mss_now);
1765 /* Link BUFF into the send queue. */
1766 __skb_header_release(buff);
1767 tcp_insert_write_queue_after(skb, buff, sk);
1772 /* Try to defer sending, if possible, in order to minimize the amount
1773 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1775 * This algorithm is from John Heffner.
1777 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1778 bool *is_cwnd_limited, u32 max_segs)
1780 const struct inet_connection_sock *icsk = inet_csk(sk);
1781 u32 age, send_win, cong_win, limit, in_flight;
1782 struct tcp_sock *tp = tcp_sk(sk);
1783 struct skb_mstamp now;
1784 struct sk_buff *head;
1787 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1790 if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR)))
1793 /* Avoid bursty behavior by allowing defer
1794 * only if the last write was recent.
1796 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1799 in_flight = tcp_packets_in_flight(tp);
1801 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1803 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1805 /* From in_flight test above, we know that cwnd > in_flight. */
1806 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1808 limit = min(send_win, cong_win);
1810 /* If a full-sized TSO skb can be sent, do it. */
1811 if (limit >= max_segs * tp->mss_cache)
1814 /* Middle in queue won't get any more data, full sendable already? */
1815 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1818 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1820 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1822 /* If at least some fraction of a window is available,
1825 chunk /= win_divisor;
1829 /* Different approach, try not to defer past a single
1830 * ACK. Receiver should ACK every other full sized
1831 * frame, so if we have space for more than 3 frames
1834 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1838 head = tcp_write_queue_head(sk);
1839 skb_mstamp_get(&now);
1840 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1841 /* If next ACK is likely to come too late (half srtt), do not defer */
1842 if (age < (tp->srtt_us >> 4))
1845 /* Ok, it looks like it is advisable to defer. */
1847 if (cong_win < send_win && cong_win < skb->len)
1848 *is_cwnd_limited = true;
1856 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1858 struct inet_connection_sock *icsk = inet_csk(sk);
1859 struct tcp_sock *tp = tcp_sk(sk);
1860 struct net *net = sock_net(sk);
1864 interval = net->ipv4.sysctl_tcp_probe_interval;
1865 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1866 if (unlikely(delta >= interval * HZ)) {
1867 int mss = tcp_current_mss(sk);
1869 /* Update current search range */
1870 icsk->icsk_mtup.probe_size = 0;
1871 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1872 sizeof(struct tcphdr) +
1873 icsk->icsk_af_ops->net_header_len;
1874 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1876 /* Update probe time stamp */
1877 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1881 /* Create a new MTU probe if we are ready.
1882 * MTU probe is regularly attempting to increase the path MTU by
1883 * deliberately sending larger packets. This discovers routing
1884 * changes resulting in larger path MTUs.
1886 * Returns 0 if we should wait to probe (no cwnd available),
1887 * 1 if a probe was sent,
1890 static int tcp_mtu_probe(struct sock *sk)
1892 struct tcp_sock *tp = tcp_sk(sk);
1893 struct inet_connection_sock *icsk = inet_csk(sk);
1894 struct sk_buff *skb, *nskb, *next;
1895 struct net *net = sock_net(sk);
1903 /* Not currently probing/verifying,
1905 * have enough cwnd, and
1906 * not SACKing (the variable headers throw things off) */
1907 if (!icsk->icsk_mtup.enabled ||
1908 icsk->icsk_mtup.probe_size ||
1909 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1910 tp->snd_cwnd < 11 ||
1911 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1914 /* Use binary search for probe_size between tcp_mss_base,
1915 * and current mss_clamp. if (search_high - search_low)
1916 * smaller than a threshold, backoff from probing.
1918 mss_now = tcp_current_mss(sk);
1919 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1920 icsk->icsk_mtup.search_low) >> 1);
1921 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1922 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1923 /* When misfortune happens, we are reprobing actively,
1924 * and then reprobe timer has expired. We stick with current
1925 * probing process by not resetting search range to its orignal.
1927 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1928 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1929 /* Check whether enough time has elaplased for
1930 * another round of probing.
1932 tcp_mtu_check_reprobe(sk);
1936 /* Have enough data in the send queue to probe? */
1937 if (tp->write_seq - tp->snd_nxt < size_needed)
1940 if (tp->snd_wnd < size_needed)
1942 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1945 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1946 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1947 if (!tcp_packets_in_flight(tp))
1953 /* We're allowed to probe. Build it now. */
1954 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1957 sk->sk_wmem_queued += nskb->truesize;
1958 sk_mem_charge(sk, nskb->truesize);
1960 skb = tcp_send_head(sk);
1962 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1963 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1964 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1965 TCP_SKB_CB(nskb)->sacked = 0;
1967 nskb->ip_summed = skb->ip_summed;
1969 tcp_insert_write_queue_before(nskb, skb, sk);
1972 tcp_for_write_queue_from_safe(skb, next, sk) {
1973 copy = min_t(int, skb->len, probe_size - len);
1974 if (nskb->ip_summed)
1975 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1977 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1978 skb_put(nskb, copy),
1981 if (skb->len <= copy) {
1982 /* We've eaten all the data from this skb.
1984 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1985 tcp_unlink_write_queue(skb, sk);
1986 sk_wmem_free_skb(sk, skb);
1988 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1989 ~(TCPHDR_FIN|TCPHDR_PSH);
1990 if (!skb_shinfo(skb)->nr_frags) {
1991 skb_pull(skb, copy);
1992 if (skb->ip_summed != CHECKSUM_PARTIAL)
1993 skb->csum = csum_partial(skb->data,
1996 __pskb_trim_head(skb, copy);
1997 tcp_set_skb_tso_segs(sk, skb, mss_now);
1999 TCP_SKB_CB(skb)->seq += copy;
2004 if (len >= probe_size)
2007 tcp_init_tso_segs(sk, nskb, nskb->len);
2009 /* We're ready to send. If this fails, the probe will
2010 * be resegmented into mss-sized pieces by tcp_write_xmit().
2012 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2013 /* Decrement cwnd here because we are sending
2014 * effectively two packets. */
2016 tcp_event_new_data_sent(sk, nskb);
2018 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2019 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2020 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2028 /* This routine writes packets to the network. It advances the
2029 * send_head. This happens as incoming acks open up the remote
2032 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2033 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2034 * account rare use of URG, this is not a big flaw.
2036 * Send at most one packet when push_one > 0. Temporarily ignore
2037 * cwnd limit to force at most one packet out when push_one == 2.
2039 * Returns true, if no segments are in flight and we have queued segments,
2040 * but cannot send anything now because of SWS or another problem.
2042 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2043 int push_one, gfp_t gfp)
2045 struct tcp_sock *tp = tcp_sk(sk);
2046 struct sk_buff *skb;
2047 unsigned int tso_segs, sent_pkts;
2050 bool is_cwnd_limited = false;
2056 /* Do MTU probing. */
2057 result = tcp_mtu_probe(sk);
2060 } else if (result > 0) {
2065 max_segs = tcp_tso_autosize(sk, mss_now);
2066 while ((skb = tcp_send_head(sk))) {
2069 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
2072 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2073 /* "skb_mstamp" is used as a start point for the retransmit timer */
2074 skb_mstamp_get(&skb->skb_mstamp);
2075 goto repair; /* Skip network transmission */
2078 cwnd_quota = tcp_cwnd_test(tp, skb);
2080 is_cwnd_limited = true;
2082 /* Force out a loss probe pkt. */
2088 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2091 if (tso_segs == 1 || !max_segs) {
2092 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2093 (tcp_skb_is_last(sk, skb) ?
2094 nonagle : TCP_NAGLE_PUSH))))
2098 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2104 if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp))
2105 limit = tcp_mss_split_point(sk, skb, mss_now,
2111 if (skb->len > limit &&
2112 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2115 /* TCP Small Queues :
2116 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2118 * - better RTT estimation and ACK scheduling
2121 * Alas, some drivers / subsystems require a fair amount
2122 * of queued bytes to ensure line rate.
2123 * One example is wifi aggregation (802.11 AMPDU)
2125 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2126 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2128 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2129 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2130 /* It is possible TX completion already happened
2131 * before we set TSQ_THROTTLED, so we must
2132 * test again the condition.
2134 smp_mb__after_atomic();
2135 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2139 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2143 /* Advance the send_head. This one is sent out.
2144 * This call will increment packets_out.
2146 tcp_event_new_data_sent(sk, skb);
2148 tcp_minshall_update(tp, mss_now, skb);
2149 sent_pkts += tcp_skb_pcount(skb);
2155 if (likely(sent_pkts)) {
2156 if (tcp_in_cwnd_reduction(sk))
2157 tp->prr_out += sent_pkts;
2159 /* Send one loss probe per tail loss episode. */
2161 tcp_schedule_loss_probe(sk);
2162 tcp_cwnd_validate(sk, is_cwnd_limited);
2165 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2168 bool tcp_schedule_loss_probe(struct sock *sk)
2170 struct inet_connection_sock *icsk = inet_csk(sk);
2171 struct tcp_sock *tp = tcp_sk(sk);
2172 u32 timeout, tlp_time_stamp, rto_time_stamp;
2173 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2175 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2177 /* No consecutive loss probes. */
2178 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2182 /* Don't do any loss probe on a Fast Open connection before 3WHS
2185 if (sk->sk_state == TCP_SYN_RECV)
2188 /* TLP is only scheduled when next timer event is RTO. */
2189 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2192 /* Schedule a loss probe in 2*RTT for SACK capable connections
2193 * in Open state, that are either limited by cwnd or application.
2195 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2196 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2199 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2203 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2204 * for delayed ack when there's one outstanding packet.
2207 if (tp->packets_out == 1)
2208 timeout = max_t(u32, timeout,
2209 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2210 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2212 /* If RTO is shorter, just schedule TLP in its place. */
2213 tlp_time_stamp = tcp_time_stamp + timeout;
2214 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2215 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2216 s32 delta = rto_time_stamp - tcp_time_stamp;
2221 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2226 /* Thanks to skb fast clones, we can detect if a prior transmit of
2227 * a packet is still in a qdisc or driver queue.
2228 * In this case, there is very little point doing a retransmit !
2229 * Note: This is called from BH context only.
2231 static bool skb_still_in_host_queue(const struct sock *sk,
2232 const struct sk_buff *skb)
2234 if (unlikely(skb_fclone_busy(sk, skb))) {
2235 NET_INC_STATS_BH(sock_net(sk),
2236 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2242 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2243 * retransmit the last segment.
2245 void tcp_send_loss_probe(struct sock *sk)
2247 struct tcp_sock *tp = tcp_sk(sk);
2248 struct sk_buff *skb;
2250 int mss = tcp_current_mss(sk);
2253 if (tcp_send_head(sk)) {
2254 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2258 /* At most one outstanding TLP retransmission. */
2259 if (tp->tlp_high_seq)
2262 /* Retransmit last segment. */
2263 skb = tcp_write_queue_tail(sk);
2267 if (skb_still_in_host_queue(sk, skb))
2270 pcount = tcp_skb_pcount(skb);
2271 if (WARN_ON(!pcount))
2274 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2275 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2278 skb = tcp_write_queue_tail(sk);
2281 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2284 err = __tcp_retransmit_skb(sk, skb);
2286 /* Record snd_nxt for loss detection. */
2288 tp->tlp_high_seq = tp->snd_nxt;
2291 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2292 inet_csk(sk)->icsk_rto,
2296 NET_INC_STATS_BH(sock_net(sk),
2297 LINUX_MIB_TCPLOSSPROBES);
2300 /* Push out any pending frames which were held back due to
2301 * TCP_CORK or attempt at coalescing tiny packets.
2302 * The socket must be locked by the caller.
2304 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2307 /* If we are closed, the bytes will have to remain here.
2308 * In time closedown will finish, we empty the write queue and
2309 * all will be happy.
2311 if (unlikely(sk->sk_state == TCP_CLOSE))
2314 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2315 sk_gfp_atomic(sk, GFP_ATOMIC)))
2316 tcp_check_probe_timer(sk);
2319 /* Send _single_ skb sitting at the send head. This function requires
2320 * true push pending frames to setup probe timer etc.
2322 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2324 struct sk_buff *skb = tcp_send_head(sk);
2326 BUG_ON(!skb || skb->len < mss_now);
2328 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2331 /* This function returns the amount that we can raise the
2332 * usable window based on the following constraints
2334 * 1. The window can never be shrunk once it is offered (RFC 793)
2335 * 2. We limit memory per socket
2338 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2339 * RECV.NEXT + RCV.WIN fixed until:
2340 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2342 * i.e. don't raise the right edge of the window until you can raise
2343 * it at least MSS bytes.
2345 * Unfortunately, the recommended algorithm breaks header prediction,
2346 * since header prediction assumes th->window stays fixed.
2348 * Strictly speaking, keeping th->window fixed violates the receiver
2349 * side SWS prevention criteria. The problem is that under this rule
2350 * a stream of single byte packets will cause the right side of the
2351 * window to always advance by a single byte.
2353 * Of course, if the sender implements sender side SWS prevention
2354 * then this will not be a problem.
2356 * BSD seems to make the following compromise:
2358 * If the free space is less than the 1/4 of the maximum
2359 * space available and the free space is less than 1/2 mss,
2360 * then set the window to 0.
2361 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2362 * Otherwise, just prevent the window from shrinking
2363 * and from being larger than the largest representable value.
2365 * This prevents incremental opening of the window in the regime
2366 * where TCP is limited by the speed of the reader side taking
2367 * data out of the TCP receive queue. It does nothing about
2368 * those cases where the window is constrained on the sender side
2369 * because the pipeline is full.
2371 * BSD also seems to "accidentally" limit itself to windows that are a
2372 * multiple of MSS, at least until the free space gets quite small.
2373 * This would appear to be a side effect of the mbuf implementation.
2374 * Combining these two algorithms results in the observed behavior
2375 * of having a fixed window size at almost all times.
2377 * Below we obtain similar behavior by forcing the offered window to
2378 * a multiple of the mss when it is feasible to do so.
2380 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2381 * Regular options like TIMESTAMP are taken into account.
2383 u32 __tcp_select_window(struct sock *sk)
2385 struct inet_connection_sock *icsk = inet_csk(sk);
2386 struct tcp_sock *tp = tcp_sk(sk);
2387 /* MSS for the peer's data. Previous versions used mss_clamp
2388 * here. I don't know if the value based on our guesses
2389 * of peer's MSS is better for the performance. It's more correct
2390 * but may be worse for the performance because of rcv_mss
2391 * fluctuations. --SAW 1998/11/1
2393 int mss = icsk->icsk_ack.rcv_mss;
2394 int free_space = tcp_space(sk);
2395 int allowed_space = tcp_full_space(sk);
2396 int full_space = min_t(int, tp->window_clamp, allowed_space);
2399 if (mss > full_space)
2402 if (free_space < (full_space >> 1)) {
2403 icsk->icsk_ack.quick = 0;
2405 if (tcp_under_memory_pressure(sk))
2406 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2409 /* free_space might become our new window, make sure we don't
2410 * increase it due to wscale.
2412 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2414 /* if free space is less than mss estimate, or is below 1/16th
2415 * of the maximum allowed, try to move to zero-window, else
2416 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2417 * new incoming data is dropped due to memory limits.
2418 * With large window, mss test triggers way too late in order
2419 * to announce zero window in time before rmem limit kicks in.
2421 if (free_space < (allowed_space >> 4) || free_space < mss)
2425 if (free_space > tp->rcv_ssthresh)
2426 free_space = tp->rcv_ssthresh;
2428 /* Don't do rounding if we are using window scaling, since the
2429 * scaled window will not line up with the MSS boundary anyway.
2431 window = tp->rcv_wnd;
2432 if (tp->rx_opt.rcv_wscale) {
2433 window = free_space;
2435 /* Advertise enough space so that it won't get scaled away.
2436 * Import case: prevent zero window announcement if
2437 * 1<<rcv_wscale > mss.
2439 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2440 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2441 << tp->rx_opt.rcv_wscale);
2443 /* Get the largest window that is a nice multiple of mss.
2444 * Window clamp already applied above.
2445 * If our current window offering is within 1 mss of the
2446 * free space we just keep it. This prevents the divide
2447 * and multiply from happening most of the time.
2448 * We also don't do any window rounding when the free space
2451 if (window <= free_space - mss || window > free_space)
2452 window = (free_space / mss) * mss;
2453 else if (mss == full_space &&
2454 free_space > window + (full_space >> 1))
2455 window = free_space;
2461 /* Collapses two adjacent SKB's during retransmission. */
2462 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2464 struct tcp_sock *tp = tcp_sk(sk);
2465 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2466 int skb_size, next_skb_size;
2468 skb_size = skb->len;
2469 next_skb_size = next_skb->len;
2471 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2473 tcp_highest_sack_combine(sk, next_skb, skb);
2475 tcp_unlink_write_queue(next_skb, sk);
2477 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2480 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2481 skb->ip_summed = CHECKSUM_PARTIAL;
2483 if (skb->ip_summed != CHECKSUM_PARTIAL)
2484 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2486 /* Update sequence range on original skb. */
2487 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2489 /* Merge over control information. This moves PSH/FIN etc. over */
2490 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2492 /* All done, get rid of second SKB and account for it so
2493 * packet counting does not break.
2495 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2497 /* changed transmit queue under us so clear hints */
2498 tcp_clear_retrans_hints_partial(tp);
2499 if (next_skb == tp->retransmit_skb_hint)
2500 tp->retransmit_skb_hint = skb;
2502 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2504 sk_wmem_free_skb(sk, next_skb);
2507 /* Check if coalescing SKBs is legal. */
2508 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2510 if (tcp_skb_pcount(skb) > 1)
2512 /* TODO: SACK collapsing could be used to remove this condition */
2513 if (skb_shinfo(skb)->nr_frags != 0)
2515 if (skb_cloned(skb))
2517 if (skb == tcp_send_head(sk))
2519 /* Some heurestics for collapsing over SACK'd could be invented */
2520 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2526 /* Collapse packets in the retransmit queue to make to create
2527 * less packets on the wire. This is only done on retransmission.
2529 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2532 struct tcp_sock *tp = tcp_sk(sk);
2533 struct sk_buff *skb = to, *tmp;
2536 if (!sysctl_tcp_retrans_collapse)
2538 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2541 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2542 if (!tcp_can_collapse(sk, skb))
2554 /* Punt if not enough space exists in the first SKB for
2555 * the data in the second
2557 if (skb->len > skb_availroom(to))
2560 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2563 tcp_collapse_retrans(sk, to);
2567 /* This retransmits one SKB. Policy decisions and retransmit queue
2568 * state updates are done by the caller. Returns non-zero if an
2569 * error occurred which prevented the send.
2571 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2573 struct tcp_sock *tp = tcp_sk(sk);
2574 struct inet_connection_sock *icsk = inet_csk(sk);
2575 unsigned int cur_mss;
2578 /* Inconslusive MTU probe */
2579 if (icsk->icsk_mtup.probe_size) {
2580 icsk->icsk_mtup.probe_size = 0;
2583 /* Do not sent more than we queued. 1/4 is reserved for possible
2584 * copying overhead: fragmentation, tunneling, mangling etc.
2586 if (atomic_read(&sk->sk_wmem_alloc) >
2587 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2590 if (skb_still_in_host_queue(sk, skb))
2593 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2594 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2596 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2600 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2601 return -EHOSTUNREACH; /* Routing failure or similar. */
2603 cur_mss = tcp_current_mss(sk);
2605 /* If receiver has shrunk his window, and skb is out of
2606 * new window, do not retransmit it. The exception is the
2607 * case, when window is shrunk to zero. In this case
2608 * our retransmit serves as a zero window probe.
2610 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2611 TCP_SKB_CB(skb)->seq != tp->snd_una)
2614 if (skb->len > cur_mss) {
2615 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2616 return -ENOMEM; /* We'll try again later. */
2618 int oldpcount = tcp_skb_pcount(skb);
2620 if (unlikely(oldpcount > 1)) {
2621 if (skb_unclone(skb, GFP_ATOMIC))
2623 tcp_init_tso_segs(sk, skb, cur_mss);
2624 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2628 /* RFC3168, section 6.1.1.1. ECN fallback */
2629 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2630 tcp_ecn_clear_syn(sk, skb);
2632 tcp_retrans_try_collapse(sk, skb, cur_mss);
2634 /* Make a copy, if the first transmission SKB clone we made
2635 * is still in somebody's hands, else make a clone.
2638 /* make sure skb->data is aligned on arches that require it
2639 * and check if ack-trimming & collapsing extended the headroom
2640 * beyond what csum_start can cover.
2642 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2643 skb_headroom(skb) >= 0xFFFF)) {
2644 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2646 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2649 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2653 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2654 /* Update global TCP statistics. */
2655 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2656 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2657 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2658 tp->total_retrans++;
2663 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2665 struct tcp_sock *tp = tcp_sk(sk);
2666 int err = __tcp_retransmit_skb(sk, skb);
2669 #if FASTRETRANS_DEBUG > 0
2670 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2671 net_dbg_ratelimited("retrans_out leaked\n");
2674 if (!tp->retrans_out)
2675 tp->lost_retrans_low = tp->snd_nxt;
2676 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2677 tp->retrans_out += tcp_skb_pcount(skb);
2679 /* Save stamp of the first retransmit. */
2680 if (!tp->retrans_stamp)
2681 tp->retrans_stamp = tcp_skb_timestamp(skb);
2683 /* snd_nxt is stored to detect loss of retransmitted segment,
2684 * see tcp_input.c tcp_sacktag_write_queue().
2686 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2687 } else if (err != -EBUSY) {
2688 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2691 if (tp->undo_retrans < 0)
2692 tp->undo_retrans = 0;
2693 tp->undo_retrans += tcp_skb_pcount(skb);
2697 /* Check if we forward retransmits are possible in the current
2698 * window/congestion state.
2700 static bool tcp_can_forward_retransmit(struct sock *sk)
2702 const struct inet_connection_sock *icsk = inet_csk(sk);
2703 const struct tcp_sock *tp = tcp_sk(sk);
2705 /* Forward retransmissions are possible only during Recovery. */
2706 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2709 /* No forward retransmissions in Reno are possible. */
2710 if (tcp_is_reno(tp))
2713 /* Yeah, we have to make difficult choice between forward transmission
2714 * and retransmission... Both ways have their merits...
2716 * For now we do not retransmit anything, while we have some new
2717 * segments to send. In the other cases, follow rule 3 for
2718 * NextSeg() specified in RFC3517.
2721 if (tcp_may_send_now(sk))
2727 /* This gets called after a retransmit timeout, and the initially
2728 * retransmitted data is acknowledged. It tries to continue
2729 * resending the rest of the retransmit queue, until either
2730 * we've sent it all or the congestion window limit is reached.
2731 * If doing SACK, the first ACK which comes back for a timeout
2732 * based retransmit packet might feed us FACK information again.
2733 * If so, we use it to avoid unnecessarily retransmissions.
2735 void tcp_xmit_retransmit_queue(struct sock *sk)
2737 const struct inet_connection_sock *icsk = inet_csk(sk);
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 struct sk_buff *skb;
2740 struct sk_buff *hole = NULL;
2743 int fwd_rexmitting = 0;
2745 if (!tp->packets_out)
2749 tp->retransmit_high = tp->snd_una;
2751 if (tp->retransmit_skb_hint) {
2752 skb = tp->retransmit_skb_hint;
2753 last_lost = TCP_SKB_CB(skb)->end_seq;
2754 if (after(last_lost, tp->retransmit_high))
2755 last_lost = tp->retransmit_high;
2757 skb = tcp_write_queue_head(sk);
2758 last_lost = tp->snd_una;
2761 tcp_for_write_queue_from(skb, sk) {
2762 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2764 if (skb == tcp_send_head(sk))
2766 /* we could do better than to assign each time */
2768 tp->retransmit_skb_hint = skb;
2770 /* Assume this retransmit will generate
2771 * only one packet for congestion window
2772 * calculation purposes. This works because
2773 * tcp_retransmit_skb() will chop up the
2774 * packet to be MSS sized and all the
2775 * packet counting works out.
2777 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2780 if (fwd_rexmitting) {
2782 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2784 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2786 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2787 tp->retransmit_high = last_lost;
2788 if (!tcp_can_forward_retransmit(sk))
2790 /* Backtrack if necessary to non-L'ed skb */
2798 } else if (!(sacked & TCPCB_LOST)) {
2799 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2804 last_lost = TCP_SKB_CB(skb)->end_seq;
2805 if (icsk->icsk_ca_state != TCP_CA_Loss)
2806 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2808 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2811 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2814 if (tcp_retransmit_skb(sk, skb))
2817 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2819 if (tcp_in_cwnd_reduction(sk))
2820 tp->prr_out += tcp_skb_pcount(skb);
2822 if (skb == tcp_write_queue_head(sk))
2823 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2824 inet_csk(sk)->icsk_rto,
2829 /* We allow to exceed memory limits for FIN packets to expedite
2830 * connection tear down and (memory) recovery.
2831 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2832 * or even be forced to close flow without any FIN.
2833 * In general, we want to allow one skb per socket to avoid hangs
2834 * with edge trigger epoll()
2836 void sk_forced_mem_schedule(struct sock *sk, int size)
2840 if (size <= sk->sk_forward_alloc)
2842 amt = sk_mem_pages(size);
2843 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2844 sk_memory_allocated_add(sk, amt, &status);
2847 /* Send a FIN. The caller locks the socket for us.
2848 * We should try to send a FIN packet really hard, but eventually give up.
2850 void tcp_send_fin(struct sock *sk)
2852 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2853 struct tcp_sock *tp = tcp_sk(sk);
2855 /* Optimization, tack on the FIN if we have one skb in write queue and
2856 * this skb was not yet sent, or we are under memory pressure.
2857 * Note: in the latter case, FIN packet will be sent after a timeout,
2858 * as TCP stack thinks it has already been transmitted.
2860 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2862 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2863 TCP_SKB_CB(tskb)->end_seq++;
2865 if (!tcp_send_head(sk)) {
2866 /* This means tskb was already sent.
2867 * Pretend we included the FIN on previous transmit.
2868 * We need to set tp->snd_nxt to the value it would have
2869 * if FIN had been sent. This is because retransmit path
2870 * does not change tp->snd_nxt.
2876 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2877 if (unlikely(!skb)) {
2882 skb_reserve(skb, MAX_TCP_HEADER);
2883 sk_forced_mem_schedule(sk, skb->truesize);
2884 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2885 tcp_init_nondata_skb(skb, tp->write_seq,
2886 TCPHDR_ACK | TCPHDR_FIN);
2887 tcp_queue_skb(sk, skb);
2889 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2892 /* We get here when a process closes a file descriptor (either due to
2893 * an explicit close() or as a byproduct of exit()'ing) and there
2894 * was unread data in the receive queue. This behavior is recommended
2895 * by RFC 2525, section 2.17. -DaveM
2897 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2899 struct sk_buff *skb;
2901 /* NOTE: No TCP options attached and we never retransmit this. */
2902 skb = alloc_skb(MAX_TCP_HEADER, priority);
2904 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2908 /* Reserve space for headers and prepare control bits. */
2909 skb_reserve(skb, MAX_TCP_HEADER);
2910 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2911 TCPHDR_ACK | TCPHDR_RST);
2913 if (tcp_transmit_skb(sk, skb, 0, priority))
2914 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2916 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2919 /* Send a crossed SYN-ACK during socket establishment.
2920 * WARNING: This routine must only be called when we have already sent
2921 * a SYN packet that crossed the incoming SYN that caused this routine
2922 * to get called. If this assumption fails then the initial rcv_wnd
2923 * and rcv_wscale values will not be correct.
2925 int tcp_send_synack(struct sock *sk)
2927 struct sk_buff *skb;
2929 skb = tcp_write_queue_head(sk);
2930 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2931 pr_debug("%s: wrong queue state\n", __func__);
2934 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2935 if (skb_cloned(skb)) {
2936 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2939 tcp_unlink_write_queue(skb, sk);
2940 __skb_header_release(nskb);
2941 __tcp_add_write_queue_head(sk, nskb);
2942 sk_wmem_free_skb(sk, skb);
2943 sk->sk_wmem_queued += nskb->truesize;
2944 sk_mem_charge(sk, nskb->truesize);
2948 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2949 tcp_ecn_send_synack(sk, skb);
2951 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2955 * tcp_make_synack - Prepare a SYN-ACK.
2956 * sk: listener socket
2957 * dst: dst entry attached to the SYNACK
2958 * req: request_sock pointer
2960 * Allocate one skb and build a SYNACK packet.
2961 * @dst is consumed : Caller should not use it again.
2963 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2964 struct request_sock *req,
2965 struct tcp_fastopen_cookie *foc)
2967 struct tcp_out_options opts;
2968 struct inet_request_sock *ireq = inet_rsk(req);
2969 struct tcp_sock *tp = tcp_sk(sk);
2971 struct sk_buff *skb;
2972 struct tcp_md5sig_key *md5 = NULL;
2973 int tcp_header_size;
2976 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2977 if (unlikely(!skb)) {
2981 /* Reserve space for headers. */
2982 skb_reserve(skb, MAX_TCP_HEADER);
2984 skb_dst_set(skb, dst);
2986 mss = dst_metric_advmss(dst);
2987 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2988 mss = tp->rx_opt.user_mss;
2990 memset(&opts, 0, sizeof(opts));
2991 #ifdef CONFIG_SYN_COOKIES
2992 if (unlikely(req->cookie_ts))
2993 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2996 skb_mstamp_get(&skb->skb_mstamp);
2998 #ifdef CONFIG_TCP_MD5SIG
3000 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3002 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3005 skb_push(skb, tcp_header_size);
3006 skb_reset_transport_header(skb);
3009 memset(th, 0, sizeof(struct tcphdr));
3012 tcp_ecn_make_synack(req, th, sk);
3013 th->source = htons(ireq->ir_num);
3014 th->dest = ireq->ir_rmt_port;
3015 /* Setting of flags are superfluous here for callers (and ECE is
3016 * not even correctly set)
3018 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3019 TCPHDR_SYN | TCPHDR_ACK);
3021 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3022 /* XXX data is queued and acked as is. No buffer/window check */
3023 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3025 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3026 th->window = htons(min(req->rcv_wnd, 65535U));
3027 tcp_options_write((__be32 *)(th + 1), tp, &opts);
3028 th->doff = (tcp_header_size >> 2);
3029 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3031 #ifdef CONFIG_TCP_MD5SIG
3032 /* Okay, we have all we need - do the md5 hash if needed */
3034 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3035 md5, req_to_sk(req), skb);
3039 /* Do not fool tcpdump (if any), clean our debris */
3040 skb->tstamp.tv64 = 0;
3043 EXPORT_SYMBOL(tcp_make_synack);
3045 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3047 struct inet_connection_sock *icsk = inet_csk(sk);
3048 const struct tcp_congestion_ops *ca;
3049 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3051 if (ca_key == TCP_CA_UNSPEC)
3055 ca = tcp_ca_find_key(ca_key);
3056 if (likely(ca && try_module_get(ca->owner))) {
3057 module_put(icsk->icsk_ca_ops->owner);
3058 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3059 icsk->icsk_ca_ops = ca;
3064 /* Do all connect socket setups that can be done AF independent. */
3065 static void tcp_connect_init(struct sock *sk)
3067 const struct dst_entry *dst = __sk_dst_get(sk);
3068 struct tcp_sock *tp = tcp_sk(sk);
3071 /* We'll fix this up when we get a response from the other end.
3072 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3074 tp->tcp_header_len = sizeof(struct tcphdr) +
3075 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3077 #ifdef CONFIG_TCP_MD5SIG
3078 if (tp->af_specific->md5_lookup(sk, sk))
3079 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3082 /* If user gave his TCP_MAXSEG, record it to clamp */
3083 if (tp->rx_opt.user_mss)
3084 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3087 tcp_sync_mss(sk, dst_mtu(dst));
3089 tcp_ca_dst_init(sk, dst);
3091 if (!tp->window_clamp)
3092 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3093 tp->advmss = dst_metric_advmss(dst);
3094 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3095 tp->advmss = tp->rx_opt.user_mss;
3097 tcp_initialize_rcv_mss(sk);
3099 /* limit the window selection if the user enforce a smaller rx buffer */
3100 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3101 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3102 tp->window_clamp = tcp_full_space(sk);
3104 tcp_select_initial_window(tcp_full_space(sk),
3105 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3108 sysctl_tcp_window_scaling,
3110 dst_metric(dst, RTAX_INITRWND));
3112 tp->rx_opt.rcv_wscale = rcv_wscale;
3113 tp->rcv_ssthresh = tp->rcv_wnd;
3116 sock_reset_flag(sk, SOCK_DONE);
3119 tp->snd_una = tp->write_seq;
3120 tp->snd_sml = tp->write_seq;
3121 tp->snd_up = tp->write_seq;
3122 tp->snd_nxt = tp->write_seq;
3124 if (likely(!tp->repair))
3127 tp->rcv_tstamp = tcp_time_stamp;
3128 tp->rcv_wup = tp->rcv_nxt;
3129 tp->copied_seq = tp->rcv_nxt;
3131 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3132 inet_csk(sk)->icsk_retransmits = 0;
3133 tcp_clear_retrans(tp);
3136 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3138 struct tcp_sock *tp = tcp_sk(sk);
3139 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3141 tcb->end_seq += skb->len;
3142 __skb_header_release(skb);
3143 __tcp_add_write_queue_tail(sk, skb);
3144 sk->sk_wmem_queued += skb->truesize;
3145 sk_mem_charge(sk, skb->truesize);
3146 tp->write_seq = tcb->end_seq;
3147 tp->packets_out += tcp_skb_pcount(skb);
3150 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3151 * queue a data-only packet after the regular SYN, such that regular SYNs
3152 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3153 * only the SYN sequence, the data are retransmitted in the first ACK.
3154 * If cookie is not cached or other error occurs, falls back to send a
3155 * regular SYN with Fast Open cookie request option.
3157 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3159 struct tcp_sock *tp = tcp_sk(sk);
3160 struct tcp_fastopen_request *fo = tp->fastopen_req;
3161 int syn_loss = 0, space, err = 0, copied;
3162 unsigned long last_syn_loss = 0;
3163 struct sk_buff *syn_data;
3165 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3166 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3167 &syn_loss, &last_syn_loss);
3168 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3170 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3171 fo->cookie.len = -1;
3175 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3176 fo->cookie.len = -1;
3177 else if (fo->cookie.len <= 0)
3180 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3181 * user-MSS. Reserve maximum option space for middleboxes that add
3182 * private TCP options. The cost is reduced data space in SYN :(
3184 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3185 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3186 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3187 MAX_TCP_OPTION_SPACE;
3189 space = min_t(size_t, space, fo->size);
3191 /* limit to order-0 allocations */
3192 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3194 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3197 syn_data->ip_summed = CHECKSUM_PARTIAL;
3198 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3199 copied = copy_from_iter(skb_put(syn_data, space), space,
3200 &fo->data->msg_iter);
3201 if (unlikely(!copied)) {
3202 kfree_skb(syn_data);
3205 if (copied != space) {
3206 skb_trim(syn_data, copied);
3210 /* No more data pending in inet_wait_for_connect() */
3211 if (space == fo->size)
3215 tcp_connect_queue_skb(sk, syn_data);
3217 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3219 syn->skb_mstamp = syn_data->skb_mstamp;
3221 /* Now full SYN+DATA was cloned and sent (or not),
3222 * remove the SYN from the original skb (syn_data)
3223 * we keep in write queue in case of a retransmit, as we
3224 * also have the SYN packet (with no data) in the same queue.
3226 TCP_SKB_CB(syn_data)->seq++;
3227 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3229 tp->syn_data = (fo->copied > 0);
3230 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3235 /* Send a regular SYN with Fast Open cookie request option */
3236 if (fo->cookie.len > 0)
3238 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3240 tp->syn_fastopen = 0;
3242 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3246 /* Build a SYN and send it off. */
3247 int tcp_connect(struct sock *sk)
3249 struct tcp_sock *tp = tcp_sk(sk);
3250 struct sk_buff *buff;
3253 tcp_connect_init(sk);
3255 if (unlikely(tp->repair)) {
3256 tcp_finish_connect(sk, NULL);
3260 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3261 if (unlikely(!buff))
3264 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3265 tp->retrans_stamp = tcp_time_stamp;
3266 tcp_connect_queue_skb(sk, buff);
3267 tcp_ecn_send_syn(sk, buff);
3269 /* Send off SYN; include data in Fast Open. */
3270 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3271 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3272 if (err == -ECONNREFUSED)
3275 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3276 * in order to make this packet get counted in tcpOutSegs.
3278 tp->snd_nxt = tp->write_seq;
3279 tp->pushed_seq = tp->write_seq;
3280 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3282 /* Timer for repeating the SYN until an answer. */
3283 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3284 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3287 EXPORT_SYMBOL(tcp_connect);
3289 /* Send out a delayed ack, the caller does the policy checking
3290 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3293 void tcp_send_delayed_ack(struct sock *sk)
3295 struct inet_connection_sock *icsk = inet_csk(sk);
3296 int ato = icsk->icsk_ack.ato;
3297 unsigned long timeout;
3299 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3301 if (ato > TCP_DELACK_MIN) {
3302 const struct tcp_sock *tp = tcp_sk(sk);
3303 int max_ato = HZ / 2;
3305 if (icsk->icsk_ack.pingpong ||
3306 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3307 max_ato = TCP_DELACK_MAX;
3309 /* Slow path, intersegment interval is "high". */
3311 /* If some rtt estimate is known, use it to bound delayed ack.
3312 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3316 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3323 ato = min(ato, max_ato);
3326 /* Stay within the limit we were given */
3327 timeout = jiffies + ato;
3329 /* Use new timeout only if there wasn't a older one earlier. */
3330 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3331 /* If delack timer was blocked or is about to expire,
3334 if (icsk->icsk_ack.blocked ||
3335 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3340 if (!time_before(timeout, icsk->icsk_ack.timeout))
3341 timeout = icsk->icsk_ack.timeout;
3343 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3344 icsk->icsk_ack.timeout = timeout;
3345 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3348 /* This routine sends an ack and also updates the window. */
3349 void tcp_send_ack(struct sock *sk)
3351 struct sk_buff *buff;
3353 /* If we have been reset, we may not send again. */
3354 if (sk->sk_state == TCP_CLOSE)
3357 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3359 /* We are not putting this on the write queue, so
3360 * tcp_transmit_skb() will set the ownership to this
3363 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3365 inet_csk_schedule_ack(sk);
3366 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3367 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3368 TCP_DELACK_MAX, TCP_RTO_MAX);
3372 /* Reserve space for headers and prepare control bits. */
3373 skb_reserve(buff, MAX_TCP_HEADER);
3374 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3376 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3378 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3379 * We also avoid tcp_wfree() overhead (cache line miss accessing
3380 * tp->tsq_flags) by using regular sock_wfree()
3382 skb_set_tcp_pure_ack(buff);
3384 /* Send it off, this clears delayed acks for us. */
3385 skb_mstamp_get(&buff->skb_mstamp);
3386 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3388 EXPORT_SYMBOL_GPL(tcp_send_ack);
3390 /* This routine sends a packet with an out of date sequence
3391 * number. It assumes the other end will try to ack it.
3393 * Question: what should we make while urgent mode?
3394 * 4.4BSD forces sending single byte of data. We cannot send
3395 * out of window data, because we have SND.NXT==SND.MAX...
3397 * Current solution: to send TWO zero-length segments in urgent mode:
3398 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3399 * out-of-date with SND.UNA-1 to probe window.
3401 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3403 struct tcp_sock *tp = tcp_sk(sk);
3404 struct sk_buff *skb;
3406 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3407 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3411 /* Reserve space for headers and set control bits. */
3412 skb_reserve(skb, MAX_TCP_HEADER);
3413 /* Use a previous sequence. This should cause the other
3414 * end to send an ack. Don't queue or clone SKB, just
3417 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3418 skb_mstamp_get(&skb->skb_mstamp);
3419 NET_INC_STATS_BH(sock_net(sk), mib);
3420 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3423 void tcp_send_window_probe(struct sock *sk)
3425 if (sk->sk_state == TCP_ESTABLISHED) {
3426 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3427 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3431 /* Initiate keepalive or window probe from timer. */
3432 int tcp_write_wakeup(struct sock *sk, int mib)
3434 struct tcp_sock *tp = tcp_sk(sk);
3435 struct sk_buff *skb;
3437 if (sk->sk_state == TCP_CLOSE)
3440 skb = tcp_send_head(sk);
3441 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3443 unsigned int mss = tcp_current_mss(sk);
3444 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3446 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3447 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3449 /* We are probing the opening of a window
3450 * but the window size is != 0
3451 * must have been a result SWS avoidance ( sender )
3453 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3455 seg_size = min(seg_size, mss);
3456 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3457 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3459 } else if (!tcp_skb_pcount(skb))
3460 tcp_set_skb_tso_segs(sk, skb, mss);
3462 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3463 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3465 tcp_event_new_data_sent(sk, skb);
3468 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3469 tcp_xmit_probe_skb(sk, 1, mib);
3470 return tcp_xmit_probe_skb(sk, 0, mib);
3474 /* A window probe timeout has occurred. If window is not closed send
3475 * a partial packet else a zero probe.
3477 void tcp_send_probe0(struct sock *sk)
3479 struct inet_connection_sock *icsk = inet_csk(sk);
3480 struct tcp_sock *tp = tcp_sk(sk);
3481 unsigned long probe_max;
3484 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3486 if (tp->packets_out || !tcp_send_head(sk)) {
3487 /* Cancel probe timer, if it is not required. */
3488 icsk->icsk_probes_out = 0;
3489 icsk->icsk_backoff = 0;
3494 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3495 icsk->icsk_backoff++;
3496 icsk->icsk_probes_out++;
3497 probe_max = TCP_RTO_MAX;
3499 /* If packet was not sent due to local congestion,
3500 * do not backoff and do not remember icsk_probes_out.
3501 * Let local senders to fight for local resources.
3503 * Use accumulated backoff yet.
3505 if (!icsk->icsk_probes_out)
3506 icsk->icsk_probes_out = 1;
3507 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3509 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3510 tcp_probe0_when(sk, probe_max),
3514 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3516 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3520 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3522 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3523 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3527 EXPORT_SYMBOL(tcp_rtx_synack);