2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
89 /* SND.NXT, if window was not shrunk.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
95 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
97 const struct tcp_sock *tp = tcp_sk(sk);
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
102 return tcp_wnd_end(tp);
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16 tcp_advertise_mss(struct sock *sk)
121 struct tcp_sock *tp = tcp_sk(sk);
122 const struct dst_entry *dst = __sk_dst_get(sk);
123 int mss = tp->advmss;
126 unsigned int metric = dst_metric_advmss(dst);
137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
140 void tcp_cwnd_restart(struct sock *sk, s32 delta)
142 struct tcp_sock *tp = tcp_sk(sk);
143 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
144 u32 cwnd = tp->snd_cwnd;
146 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
148 tp->snd_ssthresh = tcp_current_ssthresh(sk);
149 restart_cwnd = min(restart_cwnd, cwnd);
151 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
153 tp->snd_cwnd = max(cwnd, restart_cwnd);
154 tp->snd_cwnd_stamp = tcp_time_stamp;
155 tp->snd_cwnd_used = 0;
158 /* Congestion state accounting after a packet has been sent. */
159 static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct inet_connection_sock *icsk = inet_csk(sk);
163 const u32 now = tcp_time_stamp;
165 if (tcp_packets_in_flight(tp) == 0)
166 tcp_ca_event(sk, CA_EVENT_TX_START);
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 icsk->icsk_ack.pingpong = 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 struct tcp_sock *tp = tcp_sk(sk);
183 if (unlikely(rcv_nxt != tp->rcv_nxt))
184 return; /* Special ACK sent by DCTCP to reflect ECN */
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, space, sysctl_tcp_rmem[2]);
246 space = max_t(u32, space, sysctl_rmem_max);
247 space = min_t(u32, space, *window_clamp);
248 while (space > 65535 && (*rcv_wscale) < 14) {
254 if (mss > (1 << *rcv_wscale)) {
255 if (!init_rcv_wnd) /* Use default unless specified otherwise */
256 init_rcv_wnd = tcp_default_init_rwnd(mss);
257 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
260 /* Set the clamp no higher than max representable value */
261 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
263 EXPORT_SYMBOL(tcp_select_initial_window);
265 /* Chose a new window to advertise, update state in tcp_sock for the
266 * socket, and return result with RFC1323 scaling applied. The return
267 * value can be stuffed directly into th->window for an outgoing
270 static u16 tcp_select_window(struct sock *sk)
272 struct tcp_sock *tp = tcp_sk(sk);
273 u32 old_win = tp->rcv_wnd;
274 u32 cur_win = tcp_receive_window(tp);
275 u32 new_win = __tcp_select_window(sk);
277 /* Never shrink the offered window */
278 if (new_win < cur_win) {
279 /* Danger Will Robinson!
280 * Don't update rcv_wup/rcv_wnd here or else
281 * we will not be able to advertise a zero
282 * window in time. --DaveM
284 * Relax Will Robinson.
287 NET_INC_STATS(sock_net(sk),
288 LINUX_MIB_TCPWANTZEROWINDOWADV);
289 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
291 tp->rcv_wnd = new_win;
292 tp->rcv_wup = tp->rcv_nxt;
294 /* Make sure we do not exceed the maximum possible
297 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
298 new_win = min(new_win, MAX_TCP_WINDOW);
300 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
302 /* RFC1323 scaling applied */
303 new_win >>= tp->rx_opt.rcv_wscale;
305 /* If we advertise zero window, disable fast path. */
309 NET_INC_STATS(sock_net(sk),
310 LINUX_MIB_TCPTOZEROWINDOWADV);
311 } else if (old_win == 0) {
312 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
318 /* Packet ECN state for a SYN-ACK */
319 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
321 const struct tcp_sock *tp = tcp_sk(sk);
323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
324 if (!(tp->ecn_flags & TCP_ECN_OK))
325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
326 else if (tcp_ca_needs_ecn(sk))
330 /* Packet ECN state for a SYN. */
331 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
333 struct tcp_sock *tp = tcp_sk(sk);
334 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
335 tcp_ca_needs_ecn(sk);
338 const struct dst_entry *dst = __sk_dst_get(sk);
340 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
347 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
348 tp->ecn_flags = TCP_ECN_OK;
349 if (tcp_ca_needs_ecn(sk))
354 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
356 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
357 /* tp->ecn_flags are cleared at a later point in time when
358 * SYN ACK is ultimatively being received.
360 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
364 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
366 if (inet_rsk(req)->ecn_ok)
370 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
373 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
374 struct tcphdr *th, int tcp_header_len)
376 struct tcp_sock *tp = tcp_sk(sk);
378 if (tp->ecn_flags & TCP_ECN_OK) {
379 /* Not-retransmitted data segment: set ECT and inject CWR. */
380 if (skb->len != tcp_header_len &&
381 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
383 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
384 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
386 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
388 } else if (!tcp_ca_needs_ecn(sk)) {
389 /* ACK or retransmitted segment: clear ECT|CE */
390 INET_ECN_dontxmit(sk);
392 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
397 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
398 * auto increment end seqno.
400 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
402 skb->ip_summed = CHECKSUM_PARTIAL;
405 TCP_SKB_CB(skb)->tcp_flags = flags;
406 TCP_SKB_CB(skb)->sacked = 0;
408 tcp_skb_pcount_set(skb, 1);
410 TCP_SKB_CB(skb)->seq = seq;
411 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
413 TCP_SKB_CB(skb)->end_seq = seq;
416 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
418 return tp->snd_una != tp->snd_up;
421 #define OPTION_SACK_ADVERTISE (1 << 0)
422 #define OPTION_TS (1 << 1)
423 #define OPTION_MD5 (1 << 2)
424 #define OPTION_WSCALE (1 << 3)
425 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
427 struct tcp_out_options {
428 u16 options; /* bit field of OPTION_* */
429 u16 mss; /* 0 to disable */
430 u8 ws; /* window scale, 0 to disable */
431 u8 num_sack_blocks; /* number of SACK blocks to include */
432 u8 hash_size; /* bytes in hash_location */
433 __u8 *hash_location; /* temporary pointer, overloaded */
434 __u32 tsval, tsecr; /* need to include OPTION_TS */
435 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
438 /* Write previously computed TCP options to the packet.
440 * Beware: Something in the Internet is very sensitive to the ordering of
441 * TCP options, we learned this through the hard way, so be careful here.
442 * Luckily we can at least blame others for their non-compliance but from
443 * inter-operability perspective it seems that we're somewhat stuck with
444 * the ordering which we have been using if we want to keep working with
445 * those broken things (not that it currently hurts anybody as there isn't
446 * particular reason why the ordering would need to be changed).
448 * At least SACK_PERM as the first option is known to lead to a disaster
449 * (but it may well be that other scenarios fail similarly).
451 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
452 struct tcp_out_options *opts)
454 u16 options = opts->options; /* mungable copy */
456 if (unlikely(OPTION_MD5 & options)) {
457 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
458 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
459 /* overload cookie hash location */
460 opts->hash_location = (__u8 *)ptr;
464 if (unlikely(opts->mss)) {
465 *ptr++ = htonl((TCPOPT_MSS << 24) |
466 (TCPOLEN_MSS << 16) |
470 if (likely(OPTION_TS & options)) {
471 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
472 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
473 (TCPOLEN_SACK_PERM << 16) |
474 (TCPOPT_TIMESTAMP << 8) |
476 options &= ~OPTION_SACK_ADVERTISE;
478 *ptr++ = htonl((TCPOPT_NOP << 24) |
480 (TCPOPT_TIMESTAMP << 8) |
483 *ptr++ = htonl(opts->tsval);
484 *ptr++ = htonl(opts->tsecr);
487 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_SACK_PERM << 8) |
494 if (unlikely(OPTION_WSCALE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
496 (TCPOPT_WINDOW << 16) |
497 (TCPOLEN_WINDOW << 8) |
501 if (unlikely(opts->num_sack_blocks)) {
502 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
503 tp->duplicate_sack : tp->selective_acks;
506 *ptr++ = htonl((TCPOPT_NOP << 24) |
509 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
510 TCPOLEN_SACK_PERBLOCK)));
512 for (this_sack = 0; this_sack < opts->num_sack_blocks;
514 *ptr++ = htonl(sp[this_sack].start_seq);
515 *ptr++ = htonl(sp[this_sack].end_seq);
518 tp->rx_opt.dsack = 0;
521 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
522 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
524 u32 len; /* Fast Open option length */
527 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
528 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
529 TCPOPT_FASTOPEN_MAGIC);
530 p += TCPOLEN_EXP_FASTOPEN_BASE;
532 len = TCPOLEN_FASTOPEN_BASE + foc->len;
533 *p++ = TCPOPT_FASTOPEN;
537 memcpy(p, foc->val, foc->len);
538 if ((len & 3) == 2) {
539 p[foc->len] = TCPOPT_NOP;
540 p[foc->len + 1] = TCPOPT_NOP;
542 ptr += (len + 3) >> 2;
546 /* Compute TCP options for SYN packets. This is not the final
547 * network wire format yet.
549 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
550 struct tcp_out_options *opts,
551 struct tcp_md5sig_key **md5)
553 struct tcp_sock *tp = tcp_sk(sk);
554 unsigned int remaining = MAX_TCP_OPTION_SPACE;
555 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
557 #ifdef CONFIG_TCP_MD5SIG
558 *md5 = tp->af_specific->md5_lookup(sk, sk);
560 opts->options |= OPTION_MD5;
561 remaining -= TCPOLEN_MD5SIG_ALIGNED;
567 /* We always get an MSS option. The option bytes which will be seen in
568 * normal data packets should timestamps be used, must be in the MSS
569 * advertised. But we subtract them from tp->mss_cache so that
570 * calculations in tcp_sendmsg are simpler etc. So account for this
571 * fact here if necessary. If we don't do this correctly, as a
572 * receiver we won't recognize data packets as being full sized when we
573 * should, and thus we won't abide by the delayed ACK rules correctly.
574 * SACKs don't matter, we never delay an ACK when we have any of those
576 opts->mss = tcp_advertise_mss(sk);
577 remaining -= TCPOLEN_MSS_ALIGNED;
579 if (likely(sysctl_tcp_timestamps && !*md5)) {
580 opts->options |= OPTION_TS;
581 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
582 opts->tsecr = tp->rx_opt.ts_recent;
583 remaining -= TCPOLEN_TSTAMP_ALIGNED;
585 if (likely(sysctl_tcp_window_scaling)) {
586 opts->ws = tp->rx_opt.rcv_wscale;
587 opts->options |= OPTION_WSCALE;
588 remaining -= TCPOLEN_WSCALE_ALIGNED;
590 if (likely(sysctl_tcp_sack)) {
591 opts->options |= OPTION_SACK_ADVERTISE;
592 if (unlikely(!(OPTION_TS & opts->options)))
593 remaining -= TCPOLEN_SACKPERM_ALIGNED;
596 if (fastopen && fastopen->cookie.len >= 0) {
597 u32 need = fastopen->cookie.len;
599 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
600 TCPOLEN_FASTOPEN_BASE;
601 need = (need + 3) & ~3U; /* Align to 32 bits */
602 if (remaining >= need) {
603 opts->options |= OPTION_FAST_OPEN_COOKIE;
604 opts->fastopen_cookie = &fastopen->cookie;
606 tp->syn_fastopen = 1;
607 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
611 return MAX_TCP_OPTION_SPACE - remaining;
614 /* Set up TCP options for SYN-ACKs. */
615 static unsigned int tcp_synack_options(struct request_sock *req,
616 unsigned int mss, struct sk_buff *skb,
617 struct tcp_out_options *opts,
618 const struct tcp_md5sig_key *md5,
619 struct tcp_fastopen_cookie *foc,
620 enum tcp_synack_type synack_type)
622 struct inet_request_sock *ireq = inet_rsk(req);
623 unsigned int remaining = MAX_TCP_OPTION_SPACE;
625 #ifdef CONFIG_TCP_MD5SIG
627 opts->options |= OPTION_MD5;
628 remaining -= TCPOLEN_MD5SIG_ALIGNED;
630 /* We can't fit any SACK blocks in a packet with MD5 + TS
631 * options. There was discussion about disabling SACK
632 * rather than TS in order to fit in better with old,
633 * buggy kernels, but that was deemed to be unnecessary.
635 if (synack_type != TCP_SYNACK_COOKIE)
636 ireq->tstamp_ok &= !ireq->sack_ok;
640 /* We always send an MSS option. */
642 remaining -= TCPOLEN_MSS_ALIGNED;
644 if (likely(ireq->wscale_ok)) {
645 opts->ws = ireq->rcv_wscale;
646 opts->options |= OPTION_WSCALE;
647 remaining -= TCPOLEN_WSCALE_ALIGNED;
649 if (likely(ireq->tstamp_ok)) {
650 opts->options |= OPTION_TS;
651 opts->tsval = tcp_skb_timestamp(skb);
652 opts->tsecr = req->ts_recent;
653 remaining -= TCPOLEN_TSTAMP_ALIGNED;
655 if (likely(ireq->sack_ok)) {
656 opts->options |= OPTION_SACK_ADVERTISE;
657 if (unlikely(!ireq->tstamp_ok))
658 remaining -= TCPOLEN_SACKPERM_ALIGNED;
660 if (foc != NULL && foc->len >= 0) {
663 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
664 TCPOLEN_FASTOPEN_BASE;
665 need = (need + 3) & ~3U; /* Align to 32 bits */
666 if (remaining >= need) {
667 opts->options |= OPTION_FAST_OPEN_COOKIE;
668 opts->fastopen_cookie = foc;
673 return MAX_TCP_OPTION_SPACE - remaining;
676 /* Compute TCP options for ESTABLISHED sockets. This is not the
677 * final wire format yet.
679 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
680 struct tcp_out_options *opts,
681 struct tcp_md5sig_key **md5)
683 struct tcp_sock *tp = tcp_sk(sk);
684 unsigned int size = 0;
685 unsigned int eff_sacks;
689 #ifdef CONFIG_TCP_MD5SIG
690 *md5 = tp->af_specific->md5_lookup(sk, sk);
691 if (unlikely(*md5)) {
692 opts->options |= OPTION_MD5;
693 size += TCPOLEN_MD5SIG_ALIGNED;
699 if (likely(tp->rx_opt.tstamp_ok)) {
700 opts->options |= OPTION_TS;
701 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
702 opts->tsecr = tp->rx_opt.ts_recent;
703 size += TCPOLEN_TSTAMP_ALIGNED;
706 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
707 if (unlikely(eff_sacks)) {
708 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
709 opts->num_sack_blocks =
710 min_t(unsigned int, eff_sacks,
711 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
712 TCPOLEN_SACK_PERBLOCK);
713 if (likely(opts->num_sack_blocks))
714 size += TCPOLEN_SACK_BASE_ALIGNED +
715 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
722 /* TCP SMALL QUEUES (TSQ)
724 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
725 * to reduce RTT and bufferbloat.
726 * We do this using a special skb destructor (tcp_wfree).
728 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
729 * needs to be reallocated in a driver.
730 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
732 * Since transmit from skb destructor is forbidden, we use a tasklet
733 * to process all sockets that eventually need to send more skbs.
734 * We use one tasklet per cpu, with its own queue of sockets.
737 struct tasklet_struct tasklet;
738 struct list_head head; /* queue of tcp sockets */
740 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
742 static void tcp_tsq_handler(struct sock *sk)
744 if ((1 << sk->sk_state) &
745 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
746 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
747 struct tcp_sock *tp = tcp_sk(sk);
749 if (tp->lost_out > tp->retrans_out &&
750 tp->snd_cwnd > tcp_packets_in_flight(tp))
751 tcp_xmit_retransmit_queue(sk);
753 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
758 * One tasklet per cpu tries to send more skbs.
759 * We run in tasklet context but need to disable irqs when
760 * transferring tsq->head because tcp_wfree() might
761 * interrupt us (non NAPI drivers)
763 static void tcp_tasklet_func(unsigned long data)
765 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
768 struct list_head *q, *n;
772 local_irq_save(flags);
773 list_splice_init(&tsq->head, &list);
774 local_irq_restore(flags);
776 list_for_each_safe(q, n, &list) {
777 tp = list_entry(q, struct tcp_sock, tsq_node);
778 list_del(&tp->tsq_node);
780 sk = (struct sock *)tp;
783 if (!sock_owned_by_user(sk)) {
786 /* defer the work to tcp_release_cb() */
787 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
791 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
796 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
797 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
798 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
799 (1UL << TCP_MTU_REDUCED_DEFERRED))
801 * tcp_release_cb - tcp release_sock() callback
804 * called from release_sock() to perform protocol dependent
805 * actions before socket release.
807 void tcp_release_cb(struct sock *sk)
809 struct tcp_sock *tp = tcp_sk(sk);
810 unsigned long flags, nflags;
812 /* perform an atomic operation only if at least one flag is set */
814 flags = tp->tsq_flags;
815 if (!(flags & TCP_DEFERRED_ALL))
817 nflags = flags & ~TCP_DEFERRED_ALL;
818 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
820 if (flags & (1UL << TCP_TSQ_DEFERRED))
823 /* Here begins the tricky part :
824 * We are called from release_sock() with :
826 * 2) sk_lock.slock spinlock held
827 * 3) socket owned by us (sk->sk_lock.owned == 1)
829 * But following code is meant to be called from BH handlers,
830 * so we should keep BH disabled, but early release socket ownership
832 sock_release_ownership(sk);
834 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
835 tcp_write_timer_handler(sk);
838 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
839 tcp_delack_timer_handler(sk);
842 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
843 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
847 EXPORT_SYMBOL(tcp_release_cb);
849 void __init tcp_tasklet_init(void)
853 for_each_possible_cpu(i) {
854 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
856 INIT_LIST_HEAD(&tsq->head);
857 tasklet_init(&tsq->tasklet,
864 * Write buffer destructor automatically called from kfree_skb.
865 * We can't xmit new skbs from this context, as we might already
868 void tcp_wfree(struct sk_buff *skb)
870 struct sock *sk = skb->sk;
871 struct tcp_sock *tp = tcp_sk(sk);
874 /* Keep one reference on sk_wmem_alloc.
875 * Will be released by sk_free() from here or tcp_tasklet_func()
877 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
879 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
880 * Wait until our queues (qdisc + devices) are drained.
882 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
883 * - chance for incoming ACK (processed by another cpu maybe)
884 * to migrate this flow (skb->ooo_okay will be eventually set)
886 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
889 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
890 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
892 struct tsq_tasklet *tsq;
894 /* queue this socket to tasklet queue */
895 local_irq_save(flags);
896 tsq = this_cpu_ptr(&tsq_tasklet);
897 list_add(&tp->tsq_node, &tsq->head);
898 tasklet_schedule(&tsq->tasklet);
899 local_irq_restore(flags);
906 /* This routine actually transmits TCP packets queued in by
907 * tcp_do_sendmsg(). This is used by both the initial
908 * transmission and possible later retransmissions.
909 * All SKB's seen here are completely headerless. It is our
910 * job to build the TCP header, and pass the packet down to
911 * IP so it can do the same plus pass the packet off to the
914 * We are working here with either a clone of the original
915 * SKB, or a fresh unique copy made by the retransmit engine.
917 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
918 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
920 const struct inet_connection_sock *icsk = inet_csk(sk);
921 struct inet_sock *inet;
923 struct tcp_skb_cb *tcb;
924 struct tcp_out_options opts;
925 unsigned int tcp_options_size, tcp_header_size;
926 struct sk_buff *oskb = NULL;
927 struct tcp_md5sig_key *md5;
931 BUG_ON(!skb || !tcp_skb_pcount(skb));
935 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
938 if (unlikely(skb_cloned(skb)))
939 skb = pskb_copy(skb, gfp_mask);
941 skb = skb_clone(skb, gfp_mask);
945 skb_mstamp_get(&skb->skb_mstamp);
948 tcb = TCP_SKB_CB(skb);
949 memset(&opts, 0, sizeof(opts));
951 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
952 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
954 tcp_options_size = tcp_established_options(sk, skb, &opts,
956 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
958 /* if no packet is in qdisc/device queue, then allow XPS to select
959 * another queue. We can be called from tcp_tsq_handler()
960 * which holds one reference to sk_wmem_alloc.
962 * TODO: Ideally, in-flight pure ACK packets should not matter here.
963 * One way to get this would be to set skb->truesize = 2 on them.
965 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
967 skb_push(skb, tcp_header_size);
968 skb_reset_transport_header(skb);
972 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
973 skb_set_hash_from_sk(skb, sk);
974 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
976 /* Build TCP header and checksum it. */
977 th = (struct tcphdr *)skb->data;
978 th->source = inet->inet_sport;
979 th->dest = inet->inet_dport;
980 th->seq = htonl(tcb->seq);
981 th->ack_seq = htonl(rcv_nxt);
982 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
988 /* The urg_mode check is necessary during a below snd_una win probe */
989 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
990 if (before(tp->snd_up, tcb->seq + 0x10000)) {
991 th->urg_ptr = htons(tp->snd_up - tcb->seq);
993 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
994 th->urg_ptr = htons(0xFFFF);
999 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1000 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1001 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1002 th->window = htons(tcp_select_window(sk));
1003 tcp_ecn_send(sk, skb, th, tcp_header_size);
1005 /* RFC1323: The window in SYN & SYN/ACK segments
1008 th->window = htons(min(tp->rcv_wnd, 65535U));
1010 #ifdef CONFIG_TCP_MD5SIG
1011 /* Calculate the MD5 hash, as we have all we need now */
1013 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1014 tp->af_specific->calc_md5_hash(opts.hash_location,
1019 icsk->icsk_af_ops->send_check(sk, skb);
1021 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1022 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1024 if (skb->len != tcp_header_size) {
1025 tcp_event_data_sent(tp, sk);
1026 tp->data_segs_out += tcp_skb_pcount(skb);
1029 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1030 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1031 tcp_skb_pcount(skb));
1033 tp->segs_out += tcp_skb_pcount(skb);
1034 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1035 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1036 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1038 /* Our usage of tstamp should remain private */
1039 skb->tstamp.tv64 = 0;
1041 /* Cleanup our debris for IP stacks */
1042 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1043 sizeof(struct inet6_skb_parm)));
1045 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1047 if (unlikely(err > 0)) {
1049 err = net_xmit_eval(err);
1052 skb_mstamp_get(&oskb->skb_mstamp);
1053 tcp_rate_skb_sent(sk, oskb);
1058 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1061 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1062 tcp_sk(sk)->rcv_nxt);
1065 /* This routine just queues the buffer for sending.
1067 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1068 * otherwise socket can stall.
1070 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1072 struct tcp_sock *tp = tcp_sk(sk);
1074 /* Advance write_seq and place onto the write_queue. */
1075 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1076 __skb_header_release(skb);
1077 tcp_add_write_queue_tail(sk, skb);
1078 sk->sk_wmem_queued += skb->truesize;
1079 sk_mem_charge(sk, skb->truesize);
1082 /* Initialize TSO segments for a packet. */
1083 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1085 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1086 /* Avoid the costly divide in the normal
1089 tcp_skb_pcount_set(skb, 1);
1090 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1092 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1093 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1097 /* When a modification to fackets out becomes necessary, we need to check
1098 * skb is counted to fackets_out or not.
1100 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1103 struct tcp_sock *tp = tcp_sk(sk);
1105 if (!tp->sacked_out || tcp_is_reno(tp))
1108 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1109 tp->fackets_out -= decr;
1112 /* Pcount in the middle of the write queue got changed, we need to do various
1113 * tweaks to fix counters
1115 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1117 struct tcp_sock *tp = tcp_sk(sk);
1119 tp->packets_out -= decr;
1121 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1122 tp->sacked_out -= decr;
1123 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1124 tp->retrans_out -= decr;
1125 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1126 tp->lost_out -= decr;
1128 /* Reno case is special. Sigh... */
1129 if (tcp_is_reno(tp) && decr > 0)
1130 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1132 tcp_adjust_fackets_out(sk, skb, decr);
1134 if (tp->lost_skb_hint &&
1135 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1136 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1137 tp->lost_cnt_hint -= decr;
1139 tcp_verify_left_out(tp);
1142 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1144 return TCP_SKB_CB(skb)->txstamp_ack ||
1145 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1148 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1150 struct skb_shared_info *shinfo = skb_shinfo(skb);
1152 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1153 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1154 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1155 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1157 shinfo->tx_flags &= ~tsflags;
1158 shinfo2->tx_flags |= tsflags;
1159 swap(shinfo->tskey, shinfo2->tskey);
1160 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1161 TCP_SKB_CB(skb)->txstamp_ack = 0;
1165 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1167 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1168 TCP_SKB_CB(skb)->eor = 0;
1171 /* Function to create two new TCP segments. Shrinks the given segment
1172 * to the specified size and appends a new segment with the rest of the
1173 * packet to the list. This won't be called frequently, I hope.
1174 * Remember, these are still headerless SKBs at this point.
1176 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1177 unsigned int mss_now, gfp_t gfp)
1179 struct tcp_sock *tp = tcp_sk(sk);
1180 struct sk_buff *buff;
1181 int nsize, old_factor;
1186 if (WARN_ON(len > skb->len))
1189 nsize = skb_headlen(skb) - len;
1193 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1194 * We need some allowance to not penalize applications setting small
1196 * Also allow first and last skb in retransmit queue to be split.
1198 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1199 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1200 skb != tcp_rtx_queue_head(sk) &&
1201 skb != tcp_rtx_queue_tail(sk))) {
1202 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1206 if (skb_unclone(skb, gfp))
1209 /* Get a new skb... force flag on. */
1210 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1212 return -ENOMEM; /* We'll just try again later. */
1214 sk->sk_wmem_queued += buff->truesize;
1215 sk_mem_charge(sk, buff->truesize);
1216 nlen = skb->len - len - nsize;
1217 buff->truesize += nlen;
1218 skb->truesize -= nlen;
1220 /* Correct the sequence numbers. */
1221 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1222 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1223 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1225 /* PSH and FIN should only be set in the second packet. */
1226 flags = TCP_SKB_CB(skb)->tcp_flags;
1227 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1228 TCP_SKB_CB(buff)->tcp_flags = flags;
1229 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1230 tcp_skb_fragment_eor(skb, buff);
1232 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1233 /* Copy and checksum data tail into the new buffer. */
1234 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1235 skb_put(buff, nsize),
1240 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1242 skb->ip_summed = CHECKSUM_PARTIAL;
1243 skb_split(skb, buff, len);
1246 buff->ip_summed = skb->ip_summed;
1248 buff->tstamp = skb->tstamp;
1249 tcp_fragment_tstamp(skb, buff);
1251 old_factor = tcp_skb_pcount(skb);
1253 /* Fix up tso_factor for both original and new SKB. */
1254 tcp_set_skb_tso_segs(skb, mss_now);
1255 tcp_set_skb_tso_segs(buff, mss_now);
1257 /* Update delivered info for the new segment */
1258 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1260 /* If this packet has been sent out already, we must
1261 * adjust the various packet counters.
1263 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1264 int diff = old_factor - tcp_skb_pcount(skb) -
1265 tcp_skb_pcount(buff);
1268 tcp_adjust_pcount(sk, skb, diff);
1271 /* Link BUFF into the send queue. */
1272 __skb_header_release(buff);
1273 tcp_insert_write_queue_after(skb, buff, sk);
1278 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1279 * eventually). The difference is that pulled data not copied, but
1280 * immediately discarded.
1282 static int __pskb_trim_head(struct sk_buff *skb, int len)
1284 struct skb_shared_info *shinfo;
1287 eat = min_t(int, len, skb_headlen(skb));
1289 __skb_pull(skb, eat);
1296 shinfo = skb_shinfo(skb);
1297 for (i = 0; i < shinfo->nr_frags; i++) {
1298 int size = skb_frag_size(&shinfo->frags[i]);
1301 skb_frag_unref(skb, i);
1304 shinfo->frags[k] = shinfo->frags[i];
1306 shinfo->frags[k].page_offset += eat;
1307 skb_frag_size_sub(&shinfo->frags[k], eat);
1313 shinfo->nr_frags = k;
1315 skb_reset_tail_pointer(skb);
1316 skb->data_len -= len;
1317 skb->len = skb->data_len;
1321 /* Remove acked data from a packet in the transmit queue. */
1322 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1326 if (skb_unclone(skb, GFP_ATOMIC))
1329 delta_truesize = __pskb_trim_head(skb, len);
1331 TCP_SKB_CB(skb)->seq += len;
1332 skb->ip_summed = CHECKSUM_PARTIAL;
1334 if (delta_truesize) {
1335 skb->truesize -= delta_truesize;
1336 sk->sk_wmem_queued -= delta_truesize;
1337 sk_mem_uncharge(sk, delta_truesize);
1338 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1341 /* Any change of skb->len requires recalculation of tso factor. */
1342 if (tcp_skb_pcount(skb) > 1)
1343 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1348 /* Calculate MSS not accounting any TCP options. */
1349 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1351 const struct tcp_sock *tp = tcp_sk(sk);
1352 const struct inet_connection_sock *icsk = inet_csk(sk);
1355 /* Calculate base mss without TCP options:
1356 It is MMS_S - sizeof(tcphdr) of rfc1122
1358 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1360 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1361 if (icsk->icsk_af_ops->net_frag_header_len) {
1362 const struct dst_entry *dst = __sk_dst_get(sk);
1364 if (dst && dst_allfrag(dst))
1365 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1368 /* Clamp it (mss_clamp does not include tcp options) */
1369 if (mss_now > tp->rx_opt.mss_clamp)
1370 mss_now = tp->rx_opt.mss_clamp;
1372 /* Now subtract optional transport overhead */
1373 mss_now -= icsk->icsk_ext_hdr_len;
1375 /* Then reserve room for full set of TCP options and 8 bytes of data */
1376 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1380 /* Calculate MSS. Not accounting for SACKs here. */
1381 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1383 /* Subtract TCP options size, not including SACKs */
1384 return __tcp_mtu_to_mss(sk, pmtu) -
1385 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1387 EXPORT_SYMBOL(tcp_mtu_to_mss);
1389 /* Inverse of above */
1390 int tcp_mss_to_mtu(struct sock *sk, int mss)
1392 const struct tcp_sock *tp = tcp_sk(sk);
1393 const struct inet_connection_sock *icsk = inet_csk(sk);
1397 tp->tcp_header_len +
1398 icsk->icsk_ext_hdr_len +
1399 icsk->icsk_af_ops->net_header_len;
1401 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1402 if (icsk->icsk_af_ops->net_frag_header_len) {
1403 const struct dst_entry *dst = __sk_dst_get(sk);
1405 if (dst && dst_allfrag(dst))
1406 mtu += icsk->icsk_af_ops->net_frag_header_len;
1410 EXPORT_SYMBOL(tcp_mss_to_mtu);
1412 /* MTU probing init per socket */
1413 void tcp_mtup_init(struct sock *sk)
1415 struct tcp_sock *tp = tcp_sk(sk);
1416 struct inet_connection_sock *icsk = inet_csk(sk);
1417 struct net *net = sock_net(sk);
1419 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1420 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1421 icsk->icsk_af_ops->net_header_len;
1422 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1423 icsk->icsk_mtup.probe_size = 0;
1424 if (icsk->icsk_mtup.enabled)
1425 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1427 EXPORT_SYMBOL(tcp_mtup_init);
1429 /* This function synchronize snd mss to current pmtu/exthdr set.
1431 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1432 for TCP options, but includes only bare TCP header.
1434 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1435 It is minimum of user_mss and mss received with SYN.
1436 It also does not include TCP options.
1438 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1440 tp->mss_cache is current effective sending mss, including
1441 all tcp options except for SACKs. It is evaluated,
1442 taking into account current pmtu, but never exceeds
1443 tp->rx_opt.mss_clamp.
1445 NOTE1. rfc1122 clearly states that advertised MSS
1446 DOES NOT include either tcp or ip options.
1448 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1449 are READ ONLY outside this function. --ANK (980731)
1451 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1453 struct tcp_sock *tp = tcp_sk(sk);
1454 struct inet_connection_sock *icsk = inet_csk(sk);
1457 if (icsk->icsk_mtup.search_high > pmtu)
1458 icsk->icsk_mtup.search_high = pmtu;
1460 mss_now = tcp_mtu_to_mss(sk, pmtu);
1461 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1463 /* And store cached results */
1464 icsk->icsk_pmtu_cookie = pmtu;
1465 if (icsk->icsk_mtup.enabled)
1466 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1467 tp->mss_cache = mss_now;
1471 EXPORT_SYMBOL(tcp_sync_mss);
1473 /* Compute the current effective MSS, taking SACKs and IP options,
1474 * and even PMTU discovery events into account.
1476 unsigned int tcp_current_mss(struct sock *sk)
1478 const struct tcp_sock *tp = tcp_sk(sk);
1479 const struct dst_entry *dst = __sk_dst_get(sk);
1481 unsigned int header_len;
1482 struct tcp_out_options opts;
1483 struct tcp_md5sig_key *md5;
1485 mss_now = tp->mss_cache;
1488 u32 mtu = dst_mtu(dst);
1489 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1490 mss_now = tcp_sync_mss(sk, mtu);
1493 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1494 sizeof(struct tcphdr);
1495 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1496 * some common options. If this is an odd packet (because we have SACK
1497 * blocks etc) then our calculated header_len will be different, and
1498 * we have to adjust mss_now correspondingly */
1499 if (header_len != tp->tcp_header_len) {
1500 int delta = (int) header_len - tp->tcp_header_len;
1507 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1508 * As additional protections, we do not touch cwnd in retransmission phases,
1509 * and if application hit its sndbuf limit recently.
1511 static void tcp_cwnd_application_limited(struct sock *sk)
1513 struct tcp_sock *tp = tcp_sk(sk);
1515 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1516 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1517 /* Limited by application or receiver window. */
1518 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1519 u32 win_used = max(tp->snd_cwnd_used, init_win);
1520 if (win_used < tp->snd_cwnd) {
1521 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1522 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1524 tp->snd_cwnd_used = 0;
1526 tp->snd_cwnd_stamp = tcp_time_stamp;
1529 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1531 struct tcp_sock *tp = tcp_sk(sk);
1533 /* Track the strongest available signal of the degree to which the cwnd
1534 * is fully utilized. If cwnd-limited then remember that fact for the
1535 * current window. If not cwnd-limited then track the maximum number of
1536 * outstanding packets in the current window. (If cwnd-limited then we
1537 * chose to not update tp->max_packets_out to avoid an extra else
1538 * clause with no functional impact.)
1540 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1542 (!tp->is_cwnd_limited &&
1543 tp->packets_out > tp->max_packets_out)) {
1544 tp->is_cwnd_limited = is_cwnd_limited;
1545 tp->max_packets_out = tp->packets_out;
1546 tp->cwnd_usage_seq = tp->snd_nxt;
1549 if (tcp_is_cwnd_limited(sk)) {
1550 /* Network is feed fully. */
1551 tp->snd_cwnd_used = 0;
1552 tp->snd_cwnd_stamp = tcp_time_stamp;
1554 /* Network starves. */
1555 if (tp->packets_out > tp->snd_cwnd_used)
1556 tp->snd_cwnd_used = tp->packets_out;
1558 if (sysctl_tcp_slow_start_after_idle &&
1559 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1560 tcp_cwnd_application_limited(sk);
1564 /* Minshall's variant of the Nagle send check. */
1565 static bool tcp_minshall_check(const struct tcp_sock *tp)
1567 return after(tp->snd_sml, tp->snd_una) &&
1568 !after(tp->snd_sml, tp->snd_nxt);
1571 /* Update snd_sml if this skb is under mss
1572 * Note that a TSO packet might end with a sub-mss segment
1573 * The test is really :
1574 * if ((skb->len % mss) != 0)
1575 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1576 * But we can avoid doing the divide again given we already have
1577 * skb_pcount = skb->len / mss_now
1579 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1580 const struct sk_buff *skb)
1582 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1583 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1586 /* Return false, if packet can be sent now without violation Nagle's rules:
1587 * 1. It is full sized. (provided by caller in %partial bool)
1588 * 2. Or it contains FIN. (already checked by caller)
1589 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1590 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1591 * With Minshall's modification: all sent small packets are ACKed.
1593 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1597 ((nonagle & TCP_NAGLE_CORK) ||
1598 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1601 /* Return how many segs we'd like on a TSO packet,
1602 * to send one TSO packet per ms
1604 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1609 bytes = min(sk->sk_pacing_rate >> 10,
1610 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1612 /* Goal is to send at least one packet per ms,
1613 * not one big TSO packet every 100 ms.
1614 * This preserves ACK clocking and is consistent
1615 * with tcp_tso_should_defer() heuristic.
1617 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1621 EXPORT_SYMBOL(tcp_tso_autosize);
1623 /* Return the number of segments we want in the skb we are transmitting.
1624 * See if congestion control module wants to decide; otherwise, autosize.
1626 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1628 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1629 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1632 tso_segs = tcp_tso_autosize(sk, mss_now,
1633 sysctl_tcp_min_tso_segs);
1634 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1637 /* Returns the portion of skb which can be sent right away */
1638 static unsigned int tcp_mss_split_point(const struct sock *sk,
1639 const struct sk_buff *skb,
1640 unsigned int mss_now,
1641 unsigned int max_segs,
1644 const struct tcp_sock *tp = tcp_sk(sk);
1645 u32 partial, needed, window, max_len;
1647 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1648 max_len = mss_now * max_segs;
1650 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1653 needed = min(skb->len, window);
1655 if (max_len <= needed)
1658 partial = needed % mss_now;
1659 /* If last segment is not a full MSS, check if Nagle rules allow us
1660 * to include this last segment in this skb.
1661 * Otherwise, we'll split the skb at last MSS boundary
1663 if (tcp_nagle_check(partial != 0, tp, nonagle))
1664 return needed - partial;
1669 /* Can at least one segment of SKB be sent right now, according to the
1670 * congestion window rules? If so, return how many segments are allowed.
1672 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1673 const struct sk_buff *skb)
1675 u32 in_flight, cwnd, halfcwnd;
1677 /* Don't be strict about the congestion window for the final FIN. */
1678 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1679 tcp_skb_pcount(skb) == 1)
1682 in_flight = tcp_packets_in_flight(tp);
1683 cwnd = tp->snd_cwnd;
1684 if (in_flight >= cwnd)
1687 /* For better scheduling, ensure we have at least
1688 * 2 GSO packets in flight.
1690 halfcwnd = max(cwnd >> 1, 1U);
1691 return min(halfcwnd, cwnd - in_flight);
1694 /* Initialize TSO state of a skb.
1695 * This must be invoked the first time we consider transmitting
1696 * SKB onto the wire.
1698 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1700 int tso_segs = tcp_skb_pcount(skb);
1702 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1703 tcp_set_skb_tso_segs(skb, mss_now);
1704 tso_segs = tcp_skb_pcount(skb);
1710 /* Return true if the Nagle test allows this packet to be
1713 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1714 unsigned int cur_mss, int nonagle)
1716 /* Nagle rule does not apply to frames, which sit in the middle of the
1717 * write_queue (they have no chances to get new data).
1719 * This is implemented in the callers, where they modify the 'nonagle'
1720 * argument based upon the location of SKB in the send queue.
1722 if (nonagle & TCP_NAGLE_PUSH)
1725 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1726 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1729 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1735 /* Does at least the first segment of SKB fit into the send window? */
1736 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1737 const struct sk_buff *skb,
1738 unsigned int cur_mss)
1740 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1742 if (skb->len > cur_mss)
1743 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1745 return !after(end_seq, tcp_wnd_end(tp));
1748 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1749 * should be put on the wire right now. If so, it returns the number of
1750 * packets allowed by the congestion window.
1752 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1753 unsigned int cur_mss, int nonagle)
1755 const struct tcp_sock *tp = tcp_sk(sk);
1756 unsigned int cwnd_quota;
1758 tcp_init_tso_segs(skb, cur_mss);
1760 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1763 cwnd_quota = tcp_cwnd_test(tp, skb);
1764 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1770 /* Test if sending is allowed right now. */
1771 bool tcp_may_send_now(struct sock *sk)
1773 const struct tcp_sock *tp = tcp_sk(sk);
1774 struct sk_buff *skb = tcp_send_head(sk);
1777 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1778 (tcp_skb_is_last(sk, skb) ?
1779 tp->nonagle : TCP_NAGLE_PUSH));
1782 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1783 * which is put after SKB on the list. It is very much like
1784 * tcp_fragment() except that it may make several kinds of assumptions
1785 * in order to speed up the splitting operation. In particular, we
1786 * know that all the data is in scatter-gather pages, and that the
1787 * packet has never been sent out before (and thus is not cloned).
1789 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1790 unsigned int mss_now, gfp_t gfp)
1792 struct sk_buff *buff;
1793 int nlen = skb->len - len;
1796 /* All of a TSO frame must be composed of paged data. */
1797 if (skb->len != skb->data_len)
1798 return tcp_fragment(sk, skb, len, mss_now, gfp);
1800 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1801 if (unlikely(!buff))
1804 sk->sk_wmem_queued += buff->truesize;
1805 sk_mem_charge(sk, buff->truesize);
1806 buff->truesize += nlen;
1807 skb->truesize -= nlen;
1809 /* Correct the sequence numbers. */
1810 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1811 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1812 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1814 /* PSH and FIN should only be set in the second packet. */
1815 flags = TCP_SKB_CB(skb)->tcp_flags;
1816 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1817 TCP_SKB_CB(buff)->tcp_flags = flags;
1819 /* This packet was never sent out yet, so no SACK bits. */
1820 TCP_SKB_CB(buff)->sacked = 0;
1822 tcp_skb_fragment_eor(skb, buff);
1824 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1825 skb_split(skb, buff, len);
1826 tcp_fragment_tstamp(skb, buff);
1828 /* Fix up tso_factor for both original and new SKB. */
1829 tcp_set_skb_tso_segs(skb, mss_now);
1830 tcp_set_skb_tso_segs(buff, mss_now);
1832 /* Link BUFF into the send queue. */
1833 __skb_header_release(buff);
1834 tcp_insert_write_queue_after(skb, buff, sk);
1839 /* Try to defer sending, if possible, in order to minimize the amount
1840 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1842 * This algorithm is from John Heffner.
1844 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1845 bool *is_cwnd_limited, u32 max_segs)
1847 const struct inet_connection_sock *icsk = inet_csk(sk);
1848 u32 age, send_win, cong_win, limit, in_flight;
1849 struct tcp_sock *tp = tcp_sk(sk);
1850 struct skb_mstamp now;
1851 struct sk_buff *head;
1854 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1857 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1860 /* Avoid bursty behavior by allowing defer
1861 * only if the last write was recent.
1863 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1866 in_flight = tcp_packets_in_flight(tp);
1868 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1870 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1872 /* From in_flight test above, we know that cwnd > in_flight. */
1873 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1875 limit = min(send_win, cong_win);
1877 /* If a full-sized TSO skb can be sent, do it. */
1878 if (limit >= max_segs * tp->mss_cache)
1881 /* Middle in queue won't get any more data, full sendable already? */
1882 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1885 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1887 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1889 /* If at least some fraction of a window is available,
1892 chunk /= win_divisor;
1896 /* Different approach, try not to defer past a single
1897 * ACK. Receiver should ACK every other full sized
1898 * frame, so if we have space for more than 3 frames
1901 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1905 head = tcp_write_queue_head(sk);
1906 skb_mstamp_get(&now);
1907 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1908 /* If next ACK is likely to come too late (half srtt), do not defer */
1909 if (age < (tp->srtt_us >> 4))
1912 /* Ok, it looks like it is advisable to defer. */
1914 if (cong_win < send_win && cong_win <= skb->len)
1915 *is_cwnd_limited = true;
1923 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1925 struct inet_connection_sock *icsk = inet_csk(sk);
1926 struct tcp_sock *tp = tcp_sk(sk);
1927 struct net *net = sock_net(sk);
1931 interval = net->ipv4.sysctl_tcp_probe_interval;
1932 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1933 if (unlikely(delta >= interval * HZ)) {
1934 int mss = tcp_current_mss(sk);
1936 /* Update current search range */
1937 icsk->icsk_mtup.probe_size = 0;
1938 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1939 sizeof(struct tcphdr) +
1940 icsk->icsk_af_ops->net_header_len;
1941 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1943 /* Update probe time stamp */
1944 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1948 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
1950 struct sk_buff *skb, *next;
1952 skb = tcp_send_head(sk);
1953 tcp_for_write_queue_from_safe(skb, next, sk) {
1954 if (len <= skb->len)
1957 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
1966 /* Create a new MTU probe if we are ready.
1967 * MTU probe is regularly attempting to increase the path MTU by
1968 * deliberately sending larger packets. This discovers routing
1969 * changes resulting in larger path MTUs.
1971 * Returns 0 if we should wait to probe (no cwnd available),
1972 * 1 if a probe was sent,
1975 static int tcp_mtu_probe(struct sock *sk)
1977 struct tcp_sock *tp = tcp_sk(sk);
1978 struct inet_connection_sock *icsk = inet_csk(sk);
1979 struct sk_buff *skb, *nskb, *next;
1980 struct net *net = sock_net(sk);
1988 /* Not currently probing/verifying,
1990 * have enough cwnd, and
1991 * not SACKing (the variable headers throw things off) */
1992 if (!icsk->icsk_mtup.enabled ||
1993 icsk->icsk_mtup.probe_size ||
1994 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1995 tp->snd_cwnd < 11 ||
1996 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1999 /* Use binary search for probe_size between tcp_mss_base,
2000 * and current mss_clamp. if (search_high - search_low)
2001 * smaller than a threshold, backoff from probing.
2003 mss_now = tcp_current_mss(sk);
2004 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2005 icsk->icsk_mtup.search_low) >> 1);
2006 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2007 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2008 /* When misfortune happens, we are reprobing actively,
2009 * and then reprobe timer has expired. We stick with current
2010 * probing process by not resetting search range to its orignal.
2012 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2013 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2014 /* Check whether enough time has elaplased for
2015 * another round of probing.
2017 tcp_mtu_check_reprobe(sk);
2021 /* Have enough data in the send queue to probe? */
2022 if (tp->write_seq - tp->snd_nxt < size_needed)
2025 if (tp->snd_wnd < size_needed)
2027 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2030 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2031 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2032 if (!tcp_packets_in_flight(tp))
2038 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2041 /* We're allowed to probe. Build it now. */
2042 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2045 sk->sk_wmem_queued += nskb->truesize;
2046 sk_mem_charge(sk, nskb->truesize);
2048 skb = tcp_send_head(sk);
2050 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2051 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2052 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2053 TCP_SKB_CB(nskb)->sacked = 0;
2055 nskb->ip_summed = skb->ip_summed;
2057 tcp_insert_write_queue_before(nskb, skb, sk);
2058 tcp_highest_sack_replace(sk, skb, nskb);
2061 tcp_for_write_queue_from_safe(skb, next, sk) {
2062 copy = min_t(int, skb->len, probe_size - len);
2063 if (nskb->ip_summed) {
2064 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2066 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2067 skb_put(nskb, copy),
2069 nskb->csum = csum_block_add(nskb->csum, csum, len);
2072 if (skb->len <= copy) {
2073 /* We've eaten all the data from this skb.
2075 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2076 /* If this is the last SKB we copy and eor is set
2077 * we need to propagate it to the new skb.
2079 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2080 tcp_skb_collapse_tstamp(nskb, skb);
2081 tcp_unlink_write_queue(skb, sk);
2082 sk_wmem_free_skb(sk, skb);
2084 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2085 ~(TCPHDR_FIN|TCPHDR_PSH);
2086 if (!skb_shinfo(skb)->nr_frags) {
2087 skb_pull(skb, copy);
2088 if (skb->ip_summed != CHECKSUM_PARTIAL)
2089 skb->csum = csum_partial(skb->data,
2092 __pskb_trim_head(skb, copy);
2093 tcp_set_skb_tso_segs(skb, mss_now);
2095 TCP_SKB_CB(skb)->seq += copy;
2100 if (len >= probe_size)
2103 tcp_init_tso_segs(nskb, nskb->len);
2105 /* We're ready to send. If this fails, the probe will
2106 * be resegmented into mss-sized pieces by tcp_write_xmit().
2108 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2109 /* Decrement cwnd here because we are sending
2110 * effectively two packets. */
2112 tcp_event_new_data_sent(sk, nskb);
2114 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2115 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2116 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2124 /* TCP Small Queues :
2125 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2126 * (These limits are doubled for retransmits)
2128 * - better RTT estimation and ACK scheduling
2131 * Alas, some drivers / subsystems require a fair amount
2132 * of queued bytes to ensure line rate.
2133 * One example is wifi aggregation (802.11 AMPDU)
2135 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2136 unsigned int factor)
2140 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2141 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2144 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2145 set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
2146 /* It is possible TX completion already happened
2147 * before we set TSQ_THROTTLED, so we must
2148 * test again the condition.
2150 smp_mb__after_atomic();
2151 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2157 /* This routine writes packets to the network. It advances the
2158 * send_head. This happens as incoming acks open up the remote
2161 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2162 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2163 * account rare use of URG, this is not a big flaw.
2165 * Send at most one packet when push_one > 0. Temporarily ignore
2166 * cwnd limit to force at most one packet out when push_one == 2.
2168 * Returns true, if no segments are in flight and we have queued segments,
2169 * but cannot send anything now because of SWS or another problem.
2171 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2172 int push_one, gfp_t gfp)
2174 struct tcp_sock *tp = tcp_sk(sk);
2175 struct sk_buff *skb;
2176 unsigned int tso_segs, sent_pkts;
2179 bool is_cwnd_limited = false;
2185 /* Do MTU probing. */
2186 result = tcp_mtu_probe(sk);
2189 } else if (result > 0) {
2194 max_segs = tcp_tso_segs(sk, mss_now);
2195 while ((skb = tcp_send_head(sk))) {
2198 tso_segs = tcp_init_tso_segs(skb, mss_now);
2201 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2202 /* "skb_mstamp" is used as a start point for the retransmit timer */
2203 skb_mstamp_get(&skb->skb_mstamp);
2204 goto repair; /* Skip network transmission */
2207 cwnd_quota = tcp_cwnd_test(tp, skb);
2210 /* Force out a loss probe pkt. */
2216 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2219 if (tso_segs == 1) {
2220 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2221 (tcp_skb_is_last(sk, skb) ?
2222 nonagle : TCP_NAGLE_PUSH))))
2226 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2232 if (tso_segs > 1 && !tcp_urg_mode(tp))
2233 limit = tcp_mss_split_point(sk, skb, mss_now,
2239 if (skb->len > limit &&
2240 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2243 if (tcp_small_queue_check(sk, skb, 0))
2246 /* Argh, we hit an empty skb(), presumably a thread
2247 * is sleeping in sendmsg()/sk_stream_wait_memory().
2248 * We do not want to send a pure-ack packet and have
2249 * a strange looking rtx queue with empty packet(s).
2251 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2254 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2258 /* Advance the send_head. This one is sent out.
2259 * This call will increment packets_out.
2261 tcp_event_new_data_sent(sk, skb);
2263 tcp_minshall_update(tp, mss_now, skb);
2264 sent_pkts += tcp_skb_pcount(skb);
2270 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2271 if (likely(sent_pkts || is_cwnd_limited))
2272 tcp_cwnd_validate(sk, is_cwnd_limited);
2274 if (likely(sent_pkts)) {
2275 if (tcp_in_cwnd_reduction(sk))
2276 tp->prr_out += sent_pkts;
2278 /* Send one loss probe per tail loss episode. */
2280 tcp_schedule_loss_probe(sk);
2283 return !tp->packets_out && tcp_send_head(sk);
2286 bool tcp_schedule_loss_probe(struct sock *sk)
2288 struct inet_connection_sock *icsk = inet_csk(sk);
2289 struct tcp_sock *tp = tcp_sk(sk);
2290 u32 timeout, tlp_time_stamp, rto_time_stamp;
2291 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2293 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2295 /* No consecutive loss probes. */
2296 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2300 /* Don't do any loss probe on a Fast Open connection before 3WHS
2303 if (tp->fastopen_rsk)
2306 /* TLP is only scheduled when next timer event is RTO. */
2307 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2310 /* Schedule a loss probe in 2*RTT for SACK capable connections
2311 * in Open state, that are either limited by cwnd or application.
2313 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2314 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2317 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2321 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2322 * for delayed ack when there's one outstanding packet. If no RTT
2323 * sample is available then probe after TCP_TIMEOUT_INIT.
2325 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2326 if (tp->packets_out == 1)
2327 timeout = max_t(u32, timeout,
2328 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2329 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2331 /* If RTO is shorter, just schedule TLP in its place. */
2332 tlp_time_stamp = tcp_time_stamp + timeout;
2333 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2334 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2335 s32 delta = rto_time_stamp - tcp_time_stamp;
2340 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2345 /* Thanks to skb fast clones, we can detect if a prior transmit of
2346 * a packet is still in a qdisc or driver queue.
2347 * In this case, there is very little point doing a retransmit !
2349 static bool skb_still_in_host_queue(const struct sock *sk,
2350 const struct sk_buff *skb)
2352 if (unlikely(skb_fclone_busy(sk, skb))) {
2353 NET_INC_STATS(sock_net(sk),
2354 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2360 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2361 * retransmit the last segment.
2363 void tcp_send_loss_probe(struct sock *sk)
2365 struct tcp_sock *tp = tcp_sk(sk);
2366 struct sk_buff *skb;
2368 int mss = tcp_current_mss(sk);
2370 /* At most one outstanding TLP */
2371 if (tp->tlp_high_seq)
2374 tp->tlp_retrans = 0;
2375 skb = tcp_send_head(sk);
2377 if (tcp_snd_wnd_test(tp, skb, mss)) {
2378 pcount = tp->packets_out;
2379 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2380 if (tp->packets_out > pcount)
2384 skb = tcp_write_queue_prev(sk, skb);
2386 skb = tcp_write_queue_tail(sk);
2389 if (unlikely(!skb)) {
2390 WARN_ONCE(tp->packets_out,
2391 "invalid inflight: %u state %u cwnd %u mss %d\n",
2392 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2393 inet_csk(sk)->icsk_pending = 0;
2397 if (skb_still_in_host_queue(sk, skb))
2400 pcount = tcp_skb_pcount(skb);
2401 if (WARN_ON(!pcount))
2404 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2405 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2408 skb = tcp_write_queue_next(sk, skb);
2411 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2414 if (__tcp_retransmit_skb(sk, skb, 1))
2417 tp->tlp_retrans = 1;
2420 /* Record snd_nxt for loss detection. */
2421 tp->tlp_high_seq = tp->snd_nxt;
2423 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2424 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2425 inet_csk(sk)->icsk_pending = 0;
2430 /* Push out any pending frames which were held back due to
2431 * TCP_CORK or attempt at coalescing tiny packets.
2432 * The socket must be locked by the caller.
2434 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2437 /* If we are closed, the bytes will have to remain here.
2438 * In time closedown will finish, we empty the write queue and
2439 * all will be happy.
2441 if (unlikely(sk->sk_state == TCP_CLOSE))
2444 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2445 sk_gfp_mask(sk, GFP_ATOMIC)))
2446 tcp_check_probe_timer(sk);
2449 /* Send _single_ skb sitting at the send head. This function requires
2450 * true push pending frames to setup probe timer etc.
2452 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2454 struct sk_buff *skb = tcp_send_head(sk);
2456 BUG_ON(!skb || skb->len < mss_now);
2458 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2461 /* This function returns the amount that we can raise the
2462 * usable window based on the following constraints
2464 * 1. The window can never be shrunk once it is offered (RFC 793)
2465 * 2. We limit memory per socket
2468 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2469 * RECV.NEXT + RCV.WIN fixed until:
2470 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2472 * i.e. don't raise the right edge of the window until you can raise
2473 * it at least MSS bytes.
2475 * Unfortunately, the recommended algorithm breaks header prediction,
2476 * since header prediction assumes th->window stays fixed.
2478 * Strictly speaking, keeping th->window fixed violates the receiver
2479 * side SWS prevention criteria. The problem is that under this rule
2480 * a stream of single byte packets will cause the right side of the
2481 * window to always advance by a single byte.
2483 * Of course, if the sender implements sender side SWS prevention
2484 * then this will not be a problem.
2486 * BSD seems to make the following compromise:
2488 * If the free space is less than the 1/4 of the maximum
2489 * space available and the free space is less than 1/2 mss,
2490 * then set the window to 0.
2491 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2492 * Otherwise, just prevent the window from shrinking
2493 * and from being larger than the largest representable value.
2495 * This prevents incremental opening of the window in the regime
2496 * where TCP is limited by the speed of the reader side taking
2497 * data out of the TCP receive queue. It does nothing about
2498 * those cases where the window is constrained on the sender side
2499 * because the pipeline is full.
2501 * BSD also seems to "accidentally" limit itself to windows that are a
2502 * multiple of MSS, at least until the free space gets quite small.
2503 * This would appear to be a side effect of the mbuf implementation.
2504 * Combining these two algorithms results in the observed behavior
2505 * of having a fixed window size at almost all times.
2507 * Below we obtain similar behavior by forcing the offered window to
2508 * a multiple of the mss when it is feasible to do so.
2510 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2511 * Regular options like TIMESTAMP are taken into account.
2513 u32 __tcp_select_window(struct sock *sk)
2515 struct inet_connection_sock *icsk = inet_csk(sk);
2516 struct tcp_sock *tp = tcp_sk(sk);
2517 /* MSS for the peer's data. Previous versions used mss_clamp
2518 * here. I don't know if the value based on our guesses
2519 * of peer's MSS is better for the performance. It's more correct
2520 * but may be worse for the performance because of rcv_mss
2521 * fluctuations. --SAW 1998/11/1
2523 int mss = icsk->icsk_ack.rcv_mss;
2524 int free_space = tcp_space(sk);
2525 int allowed_space = tcp_full_space(sk);
2526 int full_space = min_t(int, tp->window_clamp, allowed_space);
2529 if (unlikely(mss > full_space)) {
2534 if (free_space < (full_space >> 1)) {
2535 icsk->icsk_ack.quick = 0;
2537 if (tcp_under_memory_pressure(sk))
2538 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2541 /* free_space might become our new window, make sure we don't
2542 * increase it due to wscale.
2544 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2546 /* if free space is less than mss estimate, or is below 1/16th
2547 * of the maximum allowed, try to move to zero-window, else
2548 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2549 * new incoming data is dropped due to memory limits.
2550 * With large window, mss test triggers way too late in order
2551 * to announce zero window in time before rmem limit kicks in.
2553 if (free_space < (allowed_space >> 4) || free_space < mss)
2557 if (free_space > tp->rcv_ssthresh)
2558 free_space = tp->rcv_ssthresh;
2560 /* Don't do rounding if we are using window scaling, since the
2561 * scaled window will not line up with the MSS boundary anyway.
2563 window = tp->rcv_wnd;
2564 if (tp->rx_opt.rcv_wscale) {
2565 window = free_space;
2567 /* Advertise enough space so that it won't get scaled away.
2568 * Import case: prevent zero window announcement if
2569 * 1<<rcv_wscale > mss.
2571 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2572 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2573 << tp->rx_opt.rcv_wscale);
2575 /* Get the largest window that is a nice multiple of mss.
2576 * Window clamp already applied above.
2577 * If our current window offering is within 1 mss of the
2578 * free space we just keep it. This prevents the divide
2579 * and multiply from happening most of the time.
2580 * We also don't do any window rounding when the free space
2583 if (window <= free_space - mss || window > free_space)
2584 window = (free_space / mss) * mss;
2585 else if (mss == full_space &&
2586 free_space > window + (full_space >> 1))
2587 window = free_space;
2593 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2594 const struct sk_buff *next_skb)
2596 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2597 const struct skb_shared_info *next_shinfo =
2598 skb_shinfo(next_skb);
2599 struct skb_shared_info *shinfo = skb_shinfo(skb);
2601 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2602 shinfo->tskey = next_shinfo->tskey;
2603 TCP_SKB_CB(skb)->txstamp_ack |=
2604 TCP_SKB_CB(next_skb)->txstamp_ack;
2608 /* Collapses two adjacent SKB's during retransmission. */
2609 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2611 struct tcp_sock *tp = tcp_sk(sk);
2612 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2613 int skb_size, next_skb_size;
2615 skb_size = skb->len;
2616 next_skb_size = next_skb->len;
2618 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2620 tcp_highest_sack_replace(sk, next_skb, skb);
2622 tcp_unlink_write_queue(next_skb, sk);
2624 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2627 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2628 skb->ip_summed = CHECKSUM_PARTIAL;
2630 if (skb->ip_summed != CHECKSUM_PARTIAL)
2631 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2633 /* Update sequence range on original skb. */
2634 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2636 /* Merge over control information. This moves PSH/FIN etc. over */
2637 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2639 /* All done, get rid of second SKB and account for it so
2640 * packet counting does not break.
2642 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2643 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2645 /* changed transmit queue under us so clear hints */
2646 tcp_clear_retrans_hints_partial(tp);
2647 if (next_skb == tp->retransmit_skb_hint)
2648 tp->retransmit_skb_hint = skb;
2650 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2652 tcp_skb_collapse_tstamp(skb, next_skb);
2654 sk_wmem_free_skb(sk, next_skb);
2657 /* Check if coalescing SKBs is legal. */
2658 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2660 if (tcp_skb_pcount(skb) > 1)
2662 /* TODO: SACK collapsing could be used to remove this condition */
2663 if (skb_shinfo(skb)->nr_frags != 0)
2665 if (skb_cloned(skb))
2667 if (skb == tcp_send_head(sk))
2669 /* Some heurestics for collapsing over SACK'd could be invented */
2670 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2676 /* Collapse packets in the retransmit queue to make to create
2677 * less packets on the wire. This is only done on retransmission.
2679 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2682 struct tcp_sock *tp = tcp_sk(sk);
2683 struct sk_buff *skb = to, *tmp;
2686 if (!sysctl_tcp_retrans_collapse)
2688 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2691 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2692 if (!tcp_can_collapse(sk, skb))
2695 if (!tcp_skb_can_collapse_to(to))
2707 /* Punt if not enough space exists in the first SKB for
2708 * the data in the second
2710 if (skb->len > skb_availroom(to))
2713 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2716 tcp_collapse_retrans(sk, to);
2720 /* This retransmits one SKB. Policy decisions and retransmit queue
2721 * state updates are done by the caller. Returns non-zero if an
2722 * error occurred which prevented the send.
2724 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2726 struct inet_connection_sock *icsk = inet_csk(sk);
2727 struct tcp_sock *tp = tcp_sk(sk);
2728 unsigned int cur_mss;
2732 /* Inconclusive MTU probe */
2733 if (icsk->icsk_mtup.probe_size)
2734 icsk->icsk_mtup.probe_size = 0;
2736 /* Do not sent more than we queued. 1/4 is reserved for possible
2737 * copying overhead: fragmentation, tunneling, mangling etc.
2739 if (atomic_read(&sk->sk_wmem_alloc) >
2740 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2744 if (skb_still_in_host_queue(sk, skb))
2747 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2748 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2752 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2756 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2757 return -EHOSTUNREACH; /* Routing failure or similar. */
2759 cur_mss = tcp_current_mss(sk);
2761 /* If receiver has shrunk his window, and skb is out of
2762 * new window, do not retransmit it. The exception is the
2763 * case, when window is shrunk to zero. In this case
2764 * our retransmit serves as a zero window probe.
2766 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2767 TCP_SKB_CB(skb)->seq != tp->snd_una)
2770 len = cur_mss * segs;
2771 if (skb->len > len) {
2772 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2773 return -ENOMEM; /* We'll try again later. */
2775 if (skb_unclone(skb, GFP_ATOMIC))
2778 diff = tcp_skb_pcount(skb);
2779 tcp_set_skb_tso_segs(skb, cur_mss);
2780 diff -= tcp_skb_pcount(skb);
2782 tcp_adjust_pcount(sk, skb, diff);
2783 if (skb->len < cur_mss)
2784 tcp_retrans_try_collapse(sk, skb, cur_mss);
2787 /* RFC3168, section 6.1.1.1. ECN fallback */
2788 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2789 tcp_ecn_clear_syn(sk, skb);
2791 /* make sure skb->data is aligned on arches that require it
2792 * and check if ack-trimming & collapsing extended the headroom
2793 * beyond what csum_start can cover.
2795 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2796 skb_headroom(skb) >= 0xFFFF)) {
2797 struct sk_buff *nskb;
2799 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2800 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2803 skb_mstamp_get(&skb->skb_mstamp);
2805 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2809 segs = tcp_skb_pcount(skb);
2811 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2812 /* Update global TCP statistics. */
2813 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2814 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2815 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2816 tp->total_retrans += segs;
2821 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2823 struct tcp_sock *tp = tcp_sk(sk);
2824 int err = __tcp_retransmit_skb(sk, skb, segs);
2827 #if FASTRETRANS_DEBUG > 0
2828 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2829 net_dbg_ratelimited("retrans_out leaked\n");
2832 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2833 tp->retrans_out += tcp_skb_pcount(skb);
2835 /* Save stamp of the first retransmit. */
2836 if (!tp->retrans_stamp)
2837 tp->retrans_stamp = tcp_skb_timestamp(skb);
2839 } else if (err != -EBUSY) {
2840 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2843 if (tp->undo_retrans < 0)
2844 tp->undo_retrans = 0;
2845 tp->undo_retrans += tcp_skb_pcount(skb);
2849 /* Check if we forward retransmits are possible in the current
2850 * window/congestion state.
2852 static bool tcp_can_forward_retransmit(struct sock *sk)
2854 const struct inet_connection_sock *icsk = inet_csk(sk);
2855 const struct tcp_sock *tp = tcp_sk(sk);
2857 /* Forward retransmissions are possible only during Recovery. */
2858 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2861 /* No forward retransmissions in Reno are possible. */
2862 if (tcp_is_reno(tp))
2865 /* Yeah, we have to make difficult choice between forward transmission
2866 * and retransmission... Both ways have their merits...
2868 * For now we do not retransmit anything, while we have some new
2869 * segments to send. In the other cases, follow rule 3 for
2870 * NextSeg() specified in RFC3517.
2873 if (tcp_may_send_now(sk))
2879 /* This gets called after a retransmit timeout, and the initially
2880 * retransmitted data is acknowledged. It tries to continue
2881 * resending the rest of the retransmit queue, until either
2882 * we've sent it all or the congestion window limit is reached.
2883 * If doing SACK, the first ACK which comes back for a timeout
2884 * based retransmit packet might feed us FACK information again.
2885 * If so, we use it to avoid unnecessarily retransmissions.
2887 void tcp_xmit_retransmit_queue(struct sock *sk)
2889 const struct inet_connection_sock *icsk = inet_csk(sk);
2890 struct tcp_sock *tp = tcp_sk(sk);
2891 struct sk_buff *skb;
2892 struct sk_buff *hole = NULL;
2893 u32 max_segs, last_lost;
2895 int fwd_rexmitting = 0;
2897 if (!tp->packets_out)
2901 tp->retransmit_high = tp->snd_una;
2903 if (tp->retransmit_skb_hint) {
2904 skb = tp->retransmit_skb_hint;
2905 last_lost = TCP_SKB_CB(skb)->end_seq;
2906 if (after(last_lost, tp->retransmit_high))
2907 last_lost = tp->retransmit_high;
2909 skb = tcp_write_queue_head(sk);
2910 last_lost = tp->snd_una;
2913 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2914 tcp_for_write_queue_from(skb, sk) {
2918 if (skb == tcp_send_head(sk))
2920 /* we could do better than to assign each time */
2922 tp->retransmit_skb_hint = skb;
2924 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2927 sacked = TCP_SKB_CB(skb)->sacked;
2928 /* In case tcp_shift_skb_data() have aggregated large skbs,
2929 * we need to make sure not sending too bigs TSO packets
2931 segs = min_t(int, segs, max_segs);
2933 if (fwd_rexmitting) {
2935 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2937 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2939 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2940 tp->retransmit_high = last_lost;
2941 if (!tcp_can_forward_retransmit(sk))
2943 /* Backtrack if necessary to non-L'ed skb */
2951 } else if (!(sacked & TCPCB_LOST)) {
2952 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2957 last_lost = TCP_SKB_CB(skb)->end_seq;
2958 if (icsk->icsk_ca_state != TCP_CA_Loss)
2959 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2961 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2964 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2967 if (tcp_small_queue_check(sk, skb, 1))
2970 if (tcp_retransmit_skb(sk, skb, segs))
2973 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2975 if (tcp_in_cwnd_reduction(sk))
2976 tp->prr_out += tcp_skb_pcount(skb);
2978 if (skb == tcp_write_queue_head(sk))
2979 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2980 inet_csk(sk)->icsk_rto,
2985 /* We allow to exceed memory limits for FIN packets to expedite
2986 * connection tear down and (memory) recovery.
2987 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2988 * or even be forced to close flow without any FIN.
2989 * In general, we want to allow one skb per socket to avoid hangs
2990 * with edge trigger epoll()
2992 void sk_forced_mem_schedule(struct sock *sk, int size)
2996 delta = size - sk->sk_forward_alloc;
2999 amt = sk_mem_pages(delta);
3000 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3001 sk_memory_allocated_add(sk, amt);
3003 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3004 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3007 /* Send a FIN. The caller locks the socket for us.
3008 * We should try to send a FIN packet really hard, but eventually give up.
3010 void tcp_send_fin(struct sock *sk)
3012 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3013 struct tcp_sock *tp = tcp_sk(sk);
3015 /* Optimization, tack on the FIN if we have one skb in write queue and
3016 * this skb was not yet sent, or we are under memory pressure.
3017 * Note: in the latter case, FIN packet will be sent after a timeout,
3018 * as TCP stack thinks it has already been transmitted.
3020 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3022 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3023 TCP_SKB_CB(tskb)->end_seq++;
3025 if (!tcp_send_head(sk)) {
3026 /* This means tskb was already sent.
3027 * Pretend we included the FIN on previous transmit.
3028 * We need to set tp->snd_nxt to the value it would have
3029 * if FIN had been sent. This is because retransmit path
3030 * does not change tp->snd_nxt.
3036 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3037 if (unlikely(!skb)) {
3042 skb_reserve(skb, MAX_TCP_HEADER);
3043 sk_forced_mem_schedule(sk, skb->truesize);
3044 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3045 tcp_init_nondata_skb(skb, tp->write_seq,
3046 TCPHDR_ACK | TCPHDR_FIN);
3047 tcp_queue_skb(sk, skb);
3049 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3052 /* We get here when a process closes a file descriptor (either due to
3053 * an explicit close() or as a byproduct of exit()'ing) and there
3054 * was unread data in the receive queue. This behavior is recommended
3055 * by RFC 2525, section 2.17. -DaveM
3057 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3059 struct sk_buff *skb;
3061 /* NOTE: No TCP options attached and we never retransmit this. */
3062 skb = alloc_skb(MAX_TCP_HEADER, priority);
3064 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3068 /* Reserve space for headers and prepare control bits. */
3069 skb_reserve(skb, MAX_TCP_HEADER);
3070 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3071 TCPHDR_ACK | TCPHDR_RST);
3072 skb_mstamp_get(&skb->skb_mstamp);
3074 if (tcp_transmit_skb(sk, skb, 0, priority))
3075 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3077 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3080 /* Send a crossed SYN-ACK during socket establishment.
3081 * WARNING: This routine must only be called when we have already sent
3082 * a SYN packet that crossed the incoming SYN that caused this routine
3083 * to get called. If this assumption fails then the initial rcv_wnd
3084 * and rcv_wscale values will not be correct.
3086 int tcp_send_synack(struct sock *sk)
3088 struct sk_buff *skb;
3090 skb = tcp_write_queue_head(sk);
3091 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3092 pr_debug("%s: wrong queue state\n", __func__);
3095 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3096 if (skb_cloned(skb)) {
3097 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3100 tcp_unlink_write_queue(skb, sk);
3101 __skb_header_release(nskb);
3102 __tcp_add_write_queue_head(sk, nskb);
3103 sk_wmem_free_skb(sk, skb);
3104 sk->sk_wmem_queued += nskb->truesize;
3105 sk_mem_charge(sk, nskb->truesize);
3109 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3110 tcp_ecn_send_synack(sk, skb);
3112 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3116 * tcp_make_synack - Prepare a SYN-ACK.
3117 * sk: listener socket
3118 * dst: dst entry attached to the SYNACK
3119 * req: request_sock pointer
3121 * Allocate one skb and build a SYNACK packet.
3122 * @dst is consumed : Caller should not use it again.
3124 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3125 struct request_sock *req,
3126 struct tcp_fastopen_cookie *foc,
3127 enum tcp_synack_type synack_type)
3129 struct inet_request_sock *ireq = inet_rsk(req);
3130 const struct tcp_sock *tp = tcp_sk(sk);
3131 struct tcp_md5sig_key *md5 = NULL;
3132 struct tcp_out_options opts;
3133 struct sk_buff *skb;
3134 int tcp_header_size;
3139 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3140 if (unlikely(!skb)) {
3144 /* Reserve space for headers. */
3145 skb_reserve(skb, MAX_TCP_HEADER);
3147 switch (synack_type) {
3148 case TCP_SYNACK_NORMAL:
3149 skb_set_owner_w(skb, req_to_sk(req));
3151 case TCP_SYNACK_COOKIE:
3152 /* Under synflood, we do not attach skb to a socket,
3153 * to avoid false sharing.
3156 case TCP_SYNACK_FASTOPEN:
3157 /* sk is a const pointer, because we want to express multiple
3158 * cpu might call us concurrently.
3159 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3161 skb_set_owner_w(skb, (struct sock *)sk);
3164 skb_dst_set(skb, dst);
3166 mss = dst_metric_advmss(dst);
3167 user_mss = READ_ONCE(tp->rx_opt.user_mss);
3168 if (user_mss && user_mss < mss)
3171 memset(&opts, 0, sizeof(opts));
3172 #ifdef CONFIG_SYN_COOKIES
3173 if (unlikely(req->cookie_ts))
3174 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3177 skb_mstamp_get(&skb->skb_mstamp);
3179 #ifdef CONFIG_TCP_MD5SIG
3181 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3183 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3184 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5,
3185 foc, synack_type) + sizeof(*th);
3187 skb_push(skb, tcp_header_size);
3188 skb_reset_transport_header(skb);
3190 th = (struct tcphdr *)skb->data;
3191 memset(th, 0, sizeof(struct tcphdr));
3194 tcp_ecn_make_synack(req, th);
3195 th->source = htons(ireq->ir_num);
3196 th->dest = ireq->ir_rmt_port;
3197 skb->ip_summed = CHECKSUM_PARTIAL;
3198 th->seq = htonl(tcp_rsk(req)->snt_isn);
3199 /* XXX data is queued and acked as is. No buffer/window check */
3200 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3202 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3203 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3204 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3205 th->doff = (tcp_header_size >> 2);
3206 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3208 #ifdef CONFIG_TCP_MD5SIG
3209 /* Okay, we have all we need - do the md5 hash if needed */
3211 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3212 md5, req_to_sk(req), skb);
3216 /* Do not fool tcpdump (if any), clean our debris */
3217 skb->tstamp.tv64 = 0;
3220 EXPORT_SYMBOL(tcp_make_synack);
3222 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3224 struct inet_connection_sock *icsk = inet_csk(sk);
3225 const struct tcp_congestion_ops *ca;
3226 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3228 if (ca_key == TCP_CA_UNSPEC)
3232 ca = tcp_ca_find_key(ca_key);
3233 if (likely(ca && try_module_get(ca->owner))) {
3234 module_put(icsk->icsk_ca_ops->owner);
3235 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3236 icsk->icsk_ca_ops = ca;
3241 /* Do all connect socket setups that can be done AF independent. */
3242 static void tcp_connect_init(struct sock *sk)
3244 const struct dst_entry *dst = __sk_dst_get(sk);
3245 struct tcp_sock *tp = tcp_sk(sk);
3248 /* We'll fix this up when we get a response from the other end.
3249 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3251 tp->tcp_header_len = sizeof(struct tcphdr) +
3252 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3254 #ifdef CONFIG_TCP_MD5SIG
3255 if (tp->af_specific->md5_lookup(sk, sk))
3256 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3259 /* If user gave his TCP_MAXSEG, record it to clamp */
3260 if (tp->rx_opt.user_mss)
3261 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3264 tcp_sync_mss(sk, dst_mtu(dst));
3266 tcp_ca_dst_init(sk, dst);
3268 if (!tp->window_clamp)
3269 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3270 tp->advmss = dst_metric_advmss(dst);
3271 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3272 tp->advmss = tp->rx_opt.user_mss;
3274 tcp_initialize_rcv_mss(sk);
3276 /* limit the window selection if the user enforce a smaller rx buffer */
3277 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3278 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3279 tp->window_clamp = tcp_full_space(sk);
3281 tcp_select_initial_window(tcp_full_space(sk),
3282 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3285 sysctl_tcp_window_scaling,
3287 dst_metric(dst, RTAX_INITRWND));
3289 tp->rx_opt.rcv_wscale = rcv_wscale;
3290 tp->rcv_ssthresh = tp->rcv_wnd;
3293 sock_reset_flag(sk, SOCK_DONE);
3296 tcp_write_queue_purge(sk);
3297 tp->snd_una = tp->write_seq;
3298 tp->snd_sml = tp->write_seq;
3299 tp->snd_up = tp->write_seq;
3300 tp->snd_nxt = tp->write_seq;
3302 if (likely(!tp->repair))
3305 tp->rcv_tstamp = tcp_time_stamp;
3306 tp->rcv_wup = tp->rcv_nxt;
3307 tp->copied_seq = tp->rcv_nxt;
3309 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3310 inet_csk(sk)->icsk_retransmits = 0;
3311 tcp_clear_retrans(tp);
3314 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3316 struct tcp_sock *tp = tcp_sk(sk);
3317 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3319 tcb->end_seq += skb->len;
3320 __skb_header_release(skb);
3321 __tcp_add_write_queue_tail(sk, skb);
3322 sk->sk_wmem_queued += skb->truesize;
3323 sk_mem_charge(sk, skb->truesize);
3324 tp->write_seq = tcb->end_seq;
3325 tp->packets_out += tcp_skb_pcount(skb);
3328 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3329 * queue a data-only packet after the regular SYN, such that regular SYNs
3330 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3331 * only the SYN sequence, the data are retransmitted in the first ACK.
3332 * If cookie is not cached or other error occurs, falls back to send a
3333 * regular SYN with Fast Open cookie request option.
3335 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3337 struct tcp_sock *tp = tcp_sk(sk);
3338 struct tcp_fastopen_request *fo = tp->fastopen_req;
3339 int syn_loss = 0, space, err = 0;
3340 unsigned long last_syn_loss = 0;
3341 struct sk_buff *syn_data;
3343 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3344 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3345 &syn_loss, &last_syn_loss);
3346 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3348 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3349 fo->cookie.len = -1;
3353 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3354 fo->cookie.len = -1;
3355 else if (fo->cookie.len <= 0)
3358 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3359 * user-MSS. Reserve maximum option space for middleboxes that add
3360 * private TCP options. The cost is reduced data space in SYN :(
3362 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3363 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3364 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3365 MAX_TCP_OPTION_SPACE;
3367 space = min_t(size_t, space, fo->size);
3369 /* limit to order-0 allocations */
3370 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3372 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3375 syn_data->ip_summed = CHECKSUM_PARTIAL;
3376 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3378 int copied = copy_from_iter(skb_put(syn_data, space), space,
3379 &fo->data->msg_iter);
3380 if (unlikely(!copied)) {
3381 kfree_skb(syn_data);
3384 if (copied != space) {
3385 skb_trim(syn_data, copied);
3389 /* No more data pending in inet_wait_for_connect() */
3390 if (space == fo->size)
3394 tcp_connect_queue_skb(sk, syn_data);
3396 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3398 syn->skb_mstamp = syn_data->skb_mstamp;
3400 /* Now full SYN+DATA was cloned and sent (or not),
3401 * remove the SYN from the original skb (syn_data)
3402 * we keep in write queue in case of a retransmit, as we
3403 * also have the SYN packet (with no data) in the same queue.
3405 TCP_SKB_CB(syn_data)->seq++;
3406 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3408 tp->syn_data = (fo->copied > 0);
3409 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3413 /* data was not sent, this is our new send_head */
3414 sk->sk_send_head = syn_data;
3415 tp->packets_out -= tcp_skb_pcount(syn_data);
3418 /* Send a regular SYN with Fast Open cookie request option */
3419 if (fo->cookie.len > 0)
3421 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3423 tp->syn_fastopen = 0;
3425 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3429 /* Build a SYN and send it off. */
3430 int tcp_connect(struct sock *sk)
3432 struct tcp_sock *tp = tcp_sk(sk);
3433 struct sk_buff *buff;
3436 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3437 return -EHOSTUNREACH; /* Routing failure or similar. */
3439 tcp_connect_init(sk);
3441 if (unlikely(tp->repair)) {
3442 tcp_finish_connect(sk, NULL);
3446 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3447 if (unlikely(!buff))
3450 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3451 tp->retrans_stamp = tcp_time_stamp;
3452 tcp_connect_queue_skb(sk, buff);
3453 tcp_ecn_send_syn(sk, buff);
3455 /* Send off SYN; include data in Fast Open. */
3456 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3457 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3458 if (err == -ECONNREFUSED)
3461 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3462 * in order to make this packet get counted in tcpOutSegs.
3464 tp->snd_nxt = tp->write_seq;
3465 tp->pushed_seq = tp->write_seq;
3466 buff = tcp_send_head(sk);
3467 if (unlikely(buff)) {
3468 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3469 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3471 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3473 /* Timer for repeating the SYN until an answer. */
3474 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3475 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3478 EXPORT_SYMBOL(tcp_connect);
3480 /* Send out a delayed ack, the caller does the policy checking
3481 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3484 void tcp_send_delayed_ack(struct sock *sk)
3486 struct inet_connection_sock *icsk = inet_csk(sk);
3487 int ato = icsk->icsk_ack.ato;
3488 unsigned long timeout;
3490 if (ato > TCP_DELACK_MIN) {
3491 const struct tcp_sock *tp = tcp_sk(sk);
3492 int max_ato = HZ / 2;
3494 if (icsk->icsk_ack.pingpong ||
3495 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3496 max_ato = TCP_DELACK_MAX;
3498 /* Slow path, intersegment interval is "high". */
3500 /* If some rtt estimate is known, use it to bound delayed ack.
3501 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3505 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3512 ato = min(ato, max_ato);
3515 /* Stay within the limit we were given */
3516 timeout = jiffies + ato;
3518 /* Use new timeout only if there wasn't a older one earlier. */
3519 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3520 /* If delack timer was blocked or is about to expire,
3523 if (icsk->icsk_ack.blocked ||
3524 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3529 if (!time_before(timeout, icsk->icsk_ack.timeout))
3530 timeout = icsk->icsk_ack.timeout;
3532 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3533 icsk->icsk_ack.timeout = timeout;
3534 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3537 /* This routine sends an ack and also updates the window. */
3538 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3540 struct sk_buff *buff;
3542 /* If we have been reset, we may not send again. */
3543 if (sk->sk_state == TCP_CLOSE)
3546 /* We are not putting this on the write queue, so
3547 * tcp_transmit_skb() will set the ownership to this
3550 buff = alloc_skb(MAX_TCP_HEADER,
3551 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3552 if (unlikely(!buff)) {
3553 inet_csk_schedule_ack(sk);
3554 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3555 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3556 TCP_DELACK_MAX, TCP_RTO_MAX);
3560 /* Reserve space for headers and prepare control bits. */
3561 skb_reserve(buff, MAX_TCP_HEADER);
3562 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3564 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3566 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3567 * We also avoid tcp_wfree() overhead (cache line miss accessing
3568 * tp->tsq_flags) by using regular sock_wfree()
3570 skb_set_tcp_pure_ack(buff);
3572 /* Send it off, this clears delayed acks for us. */
3573 skb_mstamp_get(&buff->skb_mstamp);
3574 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3576 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3578 void tcp_send_ack(struct sock *sk)
3580 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3583 /* This routine sends a packet with an out of date sequence
3584 * number. It assumes the other end will try to ack it.
3586 * Question: what should we make while urgent mode?
3587 * 4.4BSD forces sending single byte of data. We cannot send
3588 * out of window data, because we have SND.NXT==SND.MAX...
3590 * Current solution: to send TWO zero-length segments in urgent mode:
3591 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3592 * out-of-date with SND.UNA-1 to probe window.
3594 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3596 struct tcp_sock *tp = tcp_sk(sk);
3597 struct sk_buff *skb;
3599 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3600 skb = alloc_skb(MAX_TCP_HEADER,
3601 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3605 /* Reserve space for headers and set control bits. */
3606 skb_reserve(skb, MAX_TCP_HEADER);
3607 /* Use a previous sequence. This should cause the other
3608 * end to send an ack. Don't queue or clone SKB, just
3611 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3612 skb_mstamp_get(&skb->skb_mstamp);
3613 NET_INC_STATS(sock_net(sk), mib);
3614 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3617 void tcp_send_window_probe(struct sock *sk)
3619 if (sk->sk_state == TCP_ESTABLISHED) {
3620 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3621 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3625 /* Initiate keepalive or window probe from timer. */
3626 int tcp_write_wakeup(struct sock *sk, int mib)
3628 struct tcp_sock *tp = tcp_sk(sk);
3629 struct sk_buff *skb;
3631 if (sk->sk_state == TCP_CLOSE)
3634 skb = tcp_send_head(sk);
3635 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3637 unsigned int mss = tcp_current_mss(sk);
3638 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3640 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3641 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3643 /* We are probing the opening of a window
3644 * but the window size is != 0
3645 * must have been a result SWS avoidance ( sender )
3647 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3649 seg_size = min(seg_size, mss);
3650 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3651 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3653 } else if (!tcp_skb_pcount(skb))
3654 tcp_set_skb_tso_segs(skb, mss);
3656 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3657 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3659 tcp_event_new_data_sent(sk, skb);
3662 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3663 tcp_xmit_probe_skb(sk, 1, mib);
3664 return tcp_xmit_probe_skb(sk, 0, mib);
3668 /* A window probe timeout has occurred. If window is not closed send
3669 * a partial packet else a zero probe.
3671 void tcp_send_probe0(struct sock *sk)
3673 struct inet_connection_sock *icsk = inet_csk(sk);
3674 struct tcp_sock *tp = tcp_sk(sk);
3675 struct net *net = sock_net(sk);
3676 unsigned long probe_max;
3679 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3681 if (tp->packets_out || !tcp_send_head(sk)) {
3682 /* Cancel probe timer, if it is not required. */
3683 icsk->icsk_probes_out = 0;
3684 icsk->icsk_backoff = 0;
3689 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3690 icsk->icsk_backoff++;
3691 icsk->icsk_probes_out++;
3692 probe_max = TCP_RTO_MAX;
3694 /* If packet was not sent due to local congestion,
3695 * do not backoff and do not remember icsk_probes_out.
3696 * Let local senders to fight for local resources.
3698 * Use accumulated backoff yet.
3700 if (!icsk->icsk_probes_out)
3701 icsk->icsk_probes_out = 1;
3702 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3704 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3705 tcp_probe0_when(sk, probe_max),
3709 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3711 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3715 tcp_rsk(req)->txhash = net_tx_rndhash();
3716 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3718 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3719 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3720 if (unlikely(tcp_passive_fastopen(sk)))
3721 tcp_sk(sk)->total_retrans++;
3725 EXPORT_SYMBOL(tcp_rtx_synack);