2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 /* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
101 return tcp_wnd_end(tp);
104 /* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
118 static __u16 tcp_advertise_mss(struct sock *sk)
120 struct tcp_sock *tp = tcp_sk(sk);
121 const struct dst_entry *dst = __sk_dst_get(sk);
122 int mss = tp->advmss;
125 unsigned int metric = dst_metric_advmss(dst);
136 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
139 void tcp_cwnd_restart(struct sock *sk, s32 delta)
141 struct tcp_sock *tp = tcp_sk(sk);
142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143 u32 cwnd = tp->snd_cwnd;
145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 restart_cwnd = min(restart_cwnd, cwnd);
150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
157 /* Congestion state accounting after a packet has been sent. */
158 static void tcp_event_data_sent(struct tcp_sock *tp,
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
176 /* Account for an ACK we sent. */
177 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
184 u32 tcp_default_init_rwnd(u32 mss)
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
191 u32 init_rwnd = TCP_INIT_CWND * 2;
194 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
198 /* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
205 void tcp_select_initial_window(int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
210 unsigned int space = (__space < 0 ? 0 : __space);
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (65535 << 14);
215 space = min(*window_clamp, space);
217 /* Quantize space offering to a multiple of mss if possible. */
219 space = (space / mss) * mss;
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
229 if (sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
239 space = max_t(u32, space, sysctl_tcp_rmem[2]);
240 space = max_t(u32, space, sysctl_rmem_max);
241 space = min_t(u32, space, *window_clamp);
242 while (space > 65535 && (*rcv_wscale) < 14) {
248 if (mss > (1 << *rcv_wscale)) {
249 if (!init_rcv_wnd) /* Use default unless specified otherwise */
250 init_rcv_wnd = tcp_default_init_rwnd(mss);
251 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
257 EXPORT_SYMBOL(tcp_select_initial_window);
259 /* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
264 static u16 tcp_select_window(struct sock *sk)
266 struct tcp_sock *tp = tcp_sk(sk);
267 u32 old_win = tp->rcv_wnd;
268 u32 cur_win = tcp_receive_window(tp);
269 u32 new_win = __tcp_select_window(sk);
271 /* Never shrink the offered window */
272 if (new_win < cur_win) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
278 * Relax Will Robinson.
281 NET_INC_STATS(sock_net(sk),
282 LINUX_MIB_TCPWANTZEROWINDOWADV);
283 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
285 tp->rcv_wnd = new_win;
286 tp->rcv_wup = tp->rcv_nxt;
288 /* Make sure we do not exceed the maximum possible
291 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
292 new_win = min(new_win, MAX_TCP_WINDOW);
294 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
296 /* RFC1323 scaling applied */
297 new_win >>= tp->rx_opt.rcv_wscale;
299 /* If we advertise zero window, disable fast path. */
303 NET_INC_STATS(sock_net(sk),
304 LINUX_MIB_TCPTOZEROWINDOWADV);
305 } else if (old_win == 0) {
306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
312 /* Packet ECN state for a SYN-ACK */
313 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
315 const struct tcp_sock *tp = tcp_sk(sk);
317 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
318 if (!(tp->ecn_flags & TCP_ECN_OK))
319 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
320 else if (tcp_ca_needs_ecn(sk))
324 /* Packet ECN state for a SYN. */
325 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
327 struct tcp_sock *tp = tcp_sk(sk);
328 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
329 tcp_ca_needs_ecn(sk);
332 const struct dst_entry *dst = __sk_dst_get(sk);
334 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
341 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
342 tp->ecn_flags = TCP_ECN_OK;
343 if (tcp_ca_needs_ecn(sk))
348 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
350 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
351 /* tp->ecn_flags are cleared at a later point in time when
352 * SYN ACK is ultimatively being received.
354 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
358 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
360 if (inet_rsk(req)->ecn_ok)
364 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
367 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
368 struct tcphdr *th, int tcp_header_len)
370 struct tcp_sock *tp = tcp_sk(sk);
372 if (tp->ecn_flags & TCP_ECN_OK) {
373 /* Not-retransmitted data segment: set ECT and inject CWR. */
374 if (skb->len != tcp_header_len &&
375 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
377 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
378 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
380 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
382 } else if (!tcp_ca_needs_ecn(sk)) {
383 /* ACK or retransmitted segment: clear ECT|CE */
384 INET_ECN_dontxmit(sk);
386 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
391 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
392 * auto increment end seqno.
394 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
396 skb->ip_summed = CHECKSUM_PARTIAL;
399 TCP_SKB_CB(skb)->tcp_flags = flags;
400 TCP_SKB_CB(skb)->sacked = 0;
402 tcp_skb_pcount_set(skb, 1);
404 TCP_SKB_CB(skb)->seq = seq;
405 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
407 TCP_SKB_CB(skb)->end_seq = seq;
410 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
412 return tp->snd_una != tp->snd_up;
415 #define OPTION_SACK_ADVERTISE (1 << 0)
416 #define OPTION_TS (1 << 1)
417 #define OPTION_MD5 (1 << 2)
418 #define OPTION_WSCALE (1 << 3)
419 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
421 struct tcp_out_options {
422 u16 options; /* bit field of OPTION_* */
423 u16 mss; /* 0 to disable */
424 u8 ws; /* window scale, 0 to disable */
425 u8 num_sack_blocks; /* number of SACK blocks to include */
426 u8 hash_size; /* bytes in hash_location */
427 __u8 *hash_location; /* temporary pointer, overloaded */
428 __u32 tsval, tsecr; /* need to include OPTION_TS */
429 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
432 /* Write previously computed TCP options to the packet.
434 * Beware: Something in the Internet is very sensitive to the ordering of
435 * TCP options, we learned this through the hard way, so be careful here.
436 * Luckily we can at least blame others for their non-compliance but from
437 * inter-operability perspective it seems that we're somewhat stuck with
438 * the ordering which we have been using if we want to keep working with
439 * those broken things (not that it currently hurts anybody as there isn't
440 * particular reason why the ordering would need to be changed).
442 * At least SACK_PERM as the first option is known to lead to a disaster
443 * (but it may well be that other scenarios fail similarly).
445 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
446 struct tcp_out_options *opts)
448 u16 options = opts->options; /* mungable copy */
450 if (unlikely(OPTION_MD5 & options)) {
451 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
452 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
453 /* overload cookie hash location */
454 opts->hash_location = (__u8 *)ptr;
458 if (unlikely(opts->mss)) {
459 *ptr++ = htonl((TCPOPT_MSS << 24) |
460 (TCPOLEN_MSS << 16) |
464 if (likely(OPTION_TS & options)) {
465 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467 (TCPOLEN_SACK_PERM << 16) |
468 (TCPOPT_TIMESTAMP << 8) |
470 options &= ~OPTION_SACK_ADVERTISE;
472 *ptr++ = htonl((TCPOPT_NOP << 24) |
474 (TCPOPT_TIMESTAMP << 8) |
477 *ptr++ = htonl(opts->tsval);
478 *ptr++ = htonl(opts->tsecr);
481 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
484 (TCPOPT_SACK_PERM << 8) |
488 if (unlikely(OPTION_WSCALE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_WINDOW << 16) |
491 (TCPOLEN_WINDOW << 8) |
495 if (unlikely(opts->num_sack_blocks)) {
496 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
497 tp->duplicate_sack : tp->selective_acks;
500 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
504 TCPOLEN_SACK_PERBLOCK)));
506 for (this_sack = 0; this_sack < opts->num_sack_blocks;
508 *ptr++ = htonl(sp[this_sack].start_seq);
509 *ptr++ = htonl(sp[this_sack].end_seq);
512 tp->rx_opt.dsack = 0;
515 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
516 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
518 u32 len; /* Fast Open option length */
521 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
522 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
523 TCPOPT_FASTOPEN_MAGIC);
524 p += TCPOLEN_EXP_FASTOPEN_BASE;
526 len = TCPOLEN_FASTOPEN_BASE + foc->len;
527 *p++ = TCPOPT_FASTOPEN;
531 memcpy(p, foc->val, foc->len);
532 if ((len & 3) == 2) {
533 p[foc->len] = TCPOPT_NOP;
534 p[foc->len + 1] = TCPOPT_NOP;
536 ptr += (len + 3) >> 2;
540 /* Compute TCP options for SYN packets. This is not the final
541 * network wire format yet.
543 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
544 struct tcp_out_options *opts,
545 struct tcp_md5sig_key **md5)
547 struct tcp_sock *tp = tcp_sk(sk);
548 unsigned int remaining = MAX_TCP_OPTION_SPACE;
549 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
551 #ifdef CONFIG_TCP_MD5SIG
552 *md5 = tp->af_specific->md5_lookup(sk, sk);
554 opts->options |= OPTION_MD5;
555 remaining -= TCPOLEN_MD5SIG_ALIGNED;
561 /* We always get an MSS option. The option bytes which will be seen in
562 * normal data packets should timestamps be used, must be in the MSS
563 * advertised. But we subtract them from tp->mss_cache so that
564 * calculations in tcp_sendmsg are simpler etc. So account for this
565 * fact here if necessary. If we don't do this correctly, as a
566 * receiver we won't recognize data packets as being full sized when we
567 * should, and thus we won't abide by the delayed ACK rules correctly.
568 * SACKs don't matter, we never delay an ACK when we have any of those
570 opts->mss = tcp_advertise_mss(sk);
571 remaining -= TCPOLEN_MSS_ALIGNED;
573 if (likely(sysctl_tcp_timestamps && !*md5)) {
574 opts->options |= OPTION_TS;
575 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
576 opts->tsecr = tp->rx_opt.ts_recent;
577 remaining -= TCPOLEN_TSTAMP_ALIGNED;
579 if (likely(sysctl_tcp_window_scaling)) {
580 opts->ws = tp->rx_opt.rcv_wscale;
581 opts->options |= OPTION_WSCALE;
582 remaining -= TCPOLEN_WSCALE_ALIGNED;
584 if (likely(sysctl_tcp_sack)) {
585 opts->options |= OPTION_SACK_ADVERTISE;
586 if (unlikely(!(OPTION_TS & opts->options)))
587 remaining -= TCPOLEN_SACKPERM_ALIGNED;
590 if (fastopen && fastopen->cookie.len >= 0) {
591 u32 need = fastopen->cookie.len;
593 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
594 TCPOLEN_FASTOPEN_BASE;
595 need = (need + 3) & ~3U; /* Align to 32 bits */
596 if (remaining >= need) {
597 opts->options |= OPTION_FAST_OPEN_COOKIE;
598 opts->fastopen_cookie = &fastopen->cookie;
600 tp->syn_fastopen = 1;
601 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
605 return MAX_TCP_OPTION_SPACE - remaining;
608 /* Set up TCP options for SYN-ACKs. */
609 static unsigned int tcp_synack_options(struct request_sock *req,
610 unsigned int mss, struct sk_buff *skb,
611 struct tcp_out_options *opts,
612 const struct tcp_md5sig_key *md5,
613 struct tcp_fastopen_cookie *foc)
615 struct inet_request_sock *ireq = inet_rsk(req);
616 unsigned int remaining = MAX_TCP_OPTION_SPACE;
618 #ifdef CONFIG_TCP_MD5SIG
620 opts->options |= OPTION_MD5;
621 remaining -= TCPOLEN_MD5SIG_ALIGNED;
623 /* We can't fit any SACK blocks in a packet with MD5 + TS
624 * options. There was discussion about disabling SACK
625 * rather than TS in order to fit in better with old,
626 * buggy kernels, but that was deemed to be unnecessary.
628 ireq->tstamp_ok &= !ireq->sack_ok;
632 /* We always send an MSS option. */
634 remaining -= TCPOLEN_MSS_ALIGNED;
636 if (likely(ireq->wscale_ok)) {
637 opts->ws = ireq->rcv_wscale;
638 opts->options |= OPTION_WSCALE;
639 remaining -= TCPOLEN_WSCALE_ALIGNED;
641 if (likely(ireq->tstamp_ok)) {
642 opts->options |= OPTION_TS;
643 opts->tsval = tcp_skb_timestamp(skb);
644 opts->tsecr = req->ts_recent;
645 remaining -= TCPOLEN_TSTAMP_ALIGNED;
647 if (likely(ireq->sack_ok)) {
648 opts->options |= OPTION_SACK_ADVERTISE;
649 if (unlikely(!ireq->tstamp_ok))
650 remaining -= TCPOLEN_SACKPERM_ALIGNED;
652 if (foc != NULL && foc->len >= 0) {
655 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
656 TCPOLEN_FASTOPEN_BASE;
657 need = (need + 3) & ~3U; /* Align to 32 bits */
658 if (remaining >= need) {
659 opts->options |= OPTION_FAST_OPEN_COOKIE;
660 opts->fastopen_cookie = foc;
665 return MAX_TCP_OPTION_SPACE - remaining;
668 /* Compute TCP options for ESTABLISHED sockets. This is not the
669 * final wire format yet.
671 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
672 struct tcp_out_options *opts,
673 struct tcp_md5sig_key **md5)
675 struct tcp_sock *tp = tcp_sk(sk);
676 unsigned int size = 0;
677 unsigned int eff_sacks;
681 #ifdef CONFIG_TCP_MD5SIG
682 *md5 = tp->af_specific->md5_lookup(sk, sk);
683 if (unlikely(*md5)) {
684 opts->options |= OPTION_MD5;
685 size += TCPOLEN_MD5SIG_ALIGNED;
691 if (likely(tp->rx_opt.tstamp_ok)) {
692 opts->options |= OPTION_TS;
693 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
694 opts->tsecr = tp->rx_opt.ts_recent;
695 size += TCPOLEN_TSTAMP_ALIGNED;
698 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
699 if (unlikely(eff_sacks)) {
700 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
701 opts->num_sack_blocks =
702 min_t(unsigned int, eff_sacks,
703 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
704 TCPOLEN_SACK_PERBLOCK);
705 size += TCPOLEN_SACK_BASE_ALIGNED +
706 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
713 /* TCP SMALL QUEUES (TSQ)
715 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
716 * to reduce RTT and bufferbloat.
717 * We do this using a special skb destructor (tcp_wfree).
719 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
720 * needs to be reallocated in a driver.
721 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
723 * Since transmit from skb destructor is forbidden, we use a tasklet
724 * to process all sockets that eventually need to send more skbs.
725 * We use one tasklet per cpu, with its own queue of sockets.
728 struct tasklet_struct tasklet;
729 struct list_head head; /* queue of tcp sockets */
731 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
733 static void tcp_tsq_handler(struct sock *sk)
735 if ((1 << sk->sk_state) &
736 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
737 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
738 struct tcp_sock *tp = tcp_sk(sk);
740 if (tp->lost_out > tp->retrans_out &&
741 tp->snd_cwnd > tcp_packets_in_flight(tp))
742 tcp_xmit_retransmit_queue(sk);
744 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
749 * One tasklet per cpu tries to send more skbs.
750 * We run in tasklet context but need to disable irqs when
751 * transferring tsq->head because tcp_wfree() might
752 * interrupt us (non NAPI drivers)
754 static void tcp_tasklet_func(unsigned long data)
756 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
759 struct list_head *q, *n;
763 local_irq_save(flags);
764 list_splice_init(&tsq->head, &list);
765 local_irq_restore(flags);
767 list_for_each_safe(q, n, &list) {
768 tp = list_entry(q, struct tcp_sock, tsq_node);
769 list_del(&tp->tsq_node);
771 sk = (struct sock *)tp;
774 if (!sock_owned_by_user(sk)) {
777 /* defer the work to tcp_release_cb() */
778 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
782 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
787 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
788 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
789 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
790 (1UL << TCP_MTU_REDUCED_DEFERRED))
792 * tcp_release_cb - tcp release_sock() callback
795 * called from release_sock() to perform protocol dependent
796 * actions before socket release.
798 void tcp_release_cb(struct sock *sk)
800 struct tcp_sock *tp = tcp_sk(sk);
801 unsigned long flags, nflags;
803 /* perform an atomic operation only if at least one flag is set */
805 flags = tp->tsq_flags;
806 if (!(flags & TCP_DEFERRED_ALL))
808 nflags = flags & ~TCP_DEFERRED_ALL;
809 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
811 if (flags & (1UL << TCP_TSQ_DEFERRED))
814 /* Here begins the tricky part :
815 * We are called from release_sock() with :
817 * 2) sk_lock.slock spinlock held
818 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 * But following code is meant to be called from BH handlers,
821 * so we should keep BH disabled, but early release socket ownership
823 sock_release_ownership(sk);
825 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
826 tcp_write_timer_handler(sk);
829 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
830 tcp_delack_timer_handler(sk);
833 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
834 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
838 EXPORT_SYMBOL(tcp_release_cb);
840 void __init tcp_tasklet_init(void)
844 for_each_possible_cpu(i) {
845 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
847 INIT_LIST_HEAD(&tsq->head);
848 tasklet_init(&tsq->tasklet,
855 * Write buffer destructor automatically called from kfree_skb.
856 * We can't xmit new skbs from this context, as we might already
859 void tcp_wfree(struct sk_buff *skb)
861 struct sock *sk = skb->sk;
862 struct tcp_sock *tp = tcp_sk(sk);
865 /* Keep one reference on sk_wmem_alloc.
866 * Will be released by sk_free() from here or tcp_tasklet_func()
868 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
870 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
871 * Wait until our queues (qdisc + devices) are drained.
873 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
874 * - chance for incoming ACK (processed by another cpu maybe)
875 * to migrate this flow (skb->ooo_okay will be eventually set)
877 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
880 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
881 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
883 struct tsq_tasklet *tsq;
885 /* queue this socket to tasklet queue */
886 local_irq_save(flags);
887 tsq = this_cpu_ptr(&tsq_tasklet);
888 list_add(&tp->tsq_node, &tsq->head);
889 tasklet_schedule(&tsq->tasklet);
890 local_irq_restore(flags);
897 /* This routine actually transmits TCP packets queued in by
898 * tcp_do_sendmsg(). This is used by both the initial
899 * transmission and possible later retransmissions.
900 * All SKB's seen here are completely headerless. It is our
901 * job to build the TCP header, and pass the packet down to
902 * IP so it can do the same plus pass the packet off to the
905 * We are working here with either a clone of the original
906 * SKB, or a fresh unique copy made by the retransmit engine.
908 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
911 const struct inet_connection_sock *icsk = inet_csk(sk);
912 struct inet_sock *inet;
914 struct tcp_skb_cb *tcb;
915 struct tcp_out_options opts;
916 unsigned int tcp_options_size, tcp_header_size;
917 struct tcp_md5sig_key *md5;
921 BUG_ON(!skb || !tcp_skb_pcount(skb));
925 skb_mstamp_get(&skb->skb_mstamp);
926 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
928 tcp_rate_skb_sent(sk, skb);
930 if (unlikely(skb_cloned(skb)))
931 skb = pskb_copy(skb, gfp_mask);
933 skb = skb_clone(skb, gfp_mask);
939 tcb = TCP_SKB_CB(skb);
940 memset(&opts, 0, sizeof(opts));
942 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
943 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
945 tcp_options_size = tcp_established_options(sk, skb, &opts,
947 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
949 /* if no packet is in qdisc/device queue, then allow XPS to select
950 * another queue. We can be called from tcp_tsq_handler()
951 * which holds one reference to sk_wmem_alloc.
953 * TODO: Ideally, in-flight pure ACK packets should not matter here.
954 * One way to get this would be to set skb->truesize = 2 on them.
956 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
958 skb_push(skb, tcp_header_size);
959 skb_reset_transport_header(skb);
963 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
964 skb_set_hash_from_sk(skb, sk);
965 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
967 /* Build TCP header and checksum it. */
968 th = (struct tcphdr *)skb->data;
969 th->source = inet->inet_sport;
970 th->dest = inet->inet_dport;
971 th->seq = htonl(tcb->seq);
972 th->ack_seq = htonl(tp->rcv_nxt);
973 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
979 /* The urg_mode check is necessary during a below snd_una win probe */
980 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
981 if (before(tp->snd_up, tcb->seq + 0x10000)) {
982 th->urg_ptr = htons(tp->snd_up - tcb->seq);
984 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
985 th->urg_ptr = htons(0xFFFF);
990 tcp_options_write((__be32 *)(th + 1), tp, &opts);
991 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
992 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
993 th->window = htons(tcp_select_window(sk));
994 tcp_ecn_send(sk, skb, th, tcp_header_size);
996 /* RFC1323: The window in SYN & SYN/ACK segments
999 th->window = htons(min(tp->rcv_wnd, 65535U));
1001 #ifdef CONFIG_TCP_MD5SIG
1002 /* Calculate the MD5 hash, as we have all we need now */
1004 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1005 tp->af_specific->calc_md5_hash(opts.hash_location,
1010 icsk->icsk_af_ops->send_check(sk, skb);
1012 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1013 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1015 if (skb->len != tcp_header_size) {
1016 tcp_event_data_sent(tp, sk);
1017 tp->data_segs_out += tcp_skb_pcount(skb);
1020 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1021 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1022 tcp_skb_pcount(skb));
1024 tp->segs_out += tcp_skb_pcount(skb);
1025 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1026 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1027 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1029 /* Our usage of tstamp should remain private */
1030 skb->tstamp.tv64 = 0;
1032 /* Cleanup our debris for IP stacks */
1033 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1034 sizeof(struct inet6_skb_parm)));
1036 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1038 if (likely(err <= 0))
1043 return net_xmit_eval(err);
1046 /* This routine just queues the buffer for sending.
1048 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1049 * otherwise socket can stall.
1051 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1053 struct tcp_sock *tp = tcp_sk(sk);
1055 /* Advance write_seq and place onto the write_queue. */
1056 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1057 __skb_header_release(skb);
1058 tcp_add_write_queue_tail(sk, skb);
1059 sk->sk_wmem_queued += skb->truesize;
1060 sk_mem_charge(sk, skb->truesize);
1063 /* Initialize TSO segments for a packet. */
1064 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1066 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1067 /* Avoid the costly divide in the normal
1070 tcp_skb_pcount_set(skb, 1);
1071 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1073 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1074 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1078 /* When a modification to fackets out becomes necessary, we need to check
1079 * skb is counted to fackets_out or not.
1081 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1084 struct tcp_sock *tp = tcp_sk(sk);
1086 if (!tp->sacked_out || tcp_is_reno(tp))
1089 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1090 tp->fackets_out -= decr;
1093 /* Pcount in the middle of the write queue got changed, we need to do various
1094 * tweaks to fix counters
1096 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1098 struct tcp_sock *tp = tcp_sk(sk);
1100 tp->packets_out -= decr;
1102 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1103 tp->sacked_out -= decr;
1104 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1105 tp->retrans_out -= decr;
1106 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1107 tp->lost_out -= decr;
1109 /* Reno case is special. Sigh... */
1110 if (tcp_is_reno(tp) && decr > 0)
1111 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1113 tcp_adjust_fackets_out(sk, skb, decr);
1115 if (tp->lost_skb_hint &&
1116 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1117 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1118 tp->lost_cnt_hint -= decr;
1120 tcp_verify_left_out(tp);
1123 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1125 return TCP_SKB_CB(skb)->txstamp_ack ||
1126 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1129 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1131 struct skb_shared_info *shinfo = skb_shinfo(skb);
1133 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1134 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1135 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1136 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1138 shinfo->tx_flags &= ~tsflags;
1139 shinfo2->tx_flags |= tsflags;
1140 swap(shinfo->tskey, shinfo2->tskey);
1141 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1142 TCP_SKB_CB(skb)->txstamp_ack = 0;
1146 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1148 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1149 TCP_SKB_CB(skb)->eor = 0;
1152 /* Function to create two new TCP segments. Shrinks the given segment
1153 * to the specified size and appends a new segment with the rest of the
1154 * packet to the list. This won't be called frequently, I hope.
1155 * Remember, these are still headerless SKBs at this point.
1157 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1158 unsigned int mss_now, gfp_t gfp)
1160 struct tcp_sock *tp = tcp_sk(sk);
1161 struct sk_buff *buff;
1162 int nsize, old_factor;
1166 if (WARN_ON(len > skb->len))
1169 nsize = skb_headlen(skb) - len;
1173 if (skb_unclone(skb, gfp))
1176 /* Get a new skb... force flag on. */
1177 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1179 return -ENOMEM; /* We'll just try again later. */
1181 sk->sk_wmem_queued += buff->truesize;
1182 sk_mem_charge(sk, buff->truesize);
1183 nlen = skb->len - len - nsize;
1184 buff->truesize += nlen;
1185 skb->truesize -= nlen;
1187 /* Correct the sequence numbers. */
1188 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1189 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1190 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1192 /* PSH and FIN should only be set in the second packet. */
1193 flags = TCP_SKB_CB(skb)->tcp_flags;
1194 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1195 TCP_SKB_CB(buff)->tcp_flags = flags;
1196 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1197 tcp_skb_fragment_eor(skb, buff);
1199 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1200 /* Copy and checksum data tail into the new buffer. */
1201 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1202 skb_put(buff, nsize),
1207 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1209 skb->ip_summed = CHECKSUM_PARTIAL;
1210 skb_split(skb, buff, len);
1213 buff->ip_summed = skb->ip_summed;
1215 buff->tstamp = skb->tstamp;
1216 tcp_fragment_tstamp(skb, buff);
1218 old_factor = tcp_skb_pcount(skb);
1220 /* Fix up tso_factor for both original and new SKB. */
1221 tcp_set_skb_tso_segs(skb, mss_now);
1222 tcp_set_skb_tso_segs(buff, mss_now);
1224 /* Update delivered info for the new segment */
1225 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1227 /* If this packet has been sent out already, we must
1228 * adjust the various packet counters.
1230 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1231 int diff = old_factor - tcp_skb_pcount(skb) -
1232 tcp_skb_pcount(buff);
1235 tcp_adjust_pcount(sk, skb, diff);
1238 /* Link BUFF into the send queue. */
1239 __skb_header_release(buff);
1240 tcp_insert_write_queue_after(skb, buff, sk);
1245 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1246 * eventually). The difference is that pulled data not copied, but
1247 * immediately discarded.
1249 static void __pskb_trim_head(struct sk_buff *skb, int len)
1251 struct skb_shared_info *shinfo;
1254 eat = min_t(int, len, skb_headlen(skb));
1256 __skb_pull(skb, eat);
1263 shinfo = skb_shinfo(skb);
1264 for (i = 0; i < shinfo->nr_frags; i++) {
1265 int size = skb_frag_size(&shinfo->frags[i]);
1268 skb_frag_unref(skb, i);
1271 shinfo->frags[k] = shinfo->frags[i];
1273 shinfo->frags[k].page_offset += eat;
1274 skb_frag_size_sub(&shinfo->frags[k], eat);
1280 shinfo->nr_frags = k;
1282 skb_reset_tail_pointer(skb);
1283 skb->data_len -= len;
1284 skb->len = skb->data_len;
1287 /* Remove acked data from a packet in the transmit queue. */
1288 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1290 if (skb_unclone(skb, GFP_ATOMIC))
1293 __pskb_trim_head(skb, len);
1295 TCP_SKB_CB(skb)->seq += len;
1296 skb->ip_summed = CHECKSUM_PARTIAL;
1298 skb->truesize -= len;
1299 sk->sk_wmem_queued -= len;
1300 sk_mem_uncharge(sk, len);
1301 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1303 /* Any change of skb->len requires recalculation of tso factor. */
1304 if (tcp_skb_pcount(skb) > 1)
1305 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1310 /* Calculate MSS not accounting any TCP options. */
1311 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1313 const struct tcp_sock *tp = tcp_sk(sk);
1314 const struct inet_connection_sock *icsk = inet_csk(sk);
1317 /* Calculate base mss without TCP options:
1318 It is MMS_S - sizeof(tcphdr) of rfc1122
1320 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1322 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1323 if (icsk->icsk_af_ops->net_frag_header_len) {
1324 const struct dst_entry *dst = __sk_dst_get(sk);
1326 if (dst && dst_allfrag(dst))
1327 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1330 /* Clamp it (mss_clamp does not include tcp options) */
1331 if (mss_now > tp->rx_opt.mss_clamp)
1332 mss_now = tp->rx_opt.mss_clamp;
1334 /* Now subtract optional transport overhead */
1335 mss_now -= icsk->icsk_ext_hdr_len;
1337 /* Then reserve room for full set of TCP options and 8 bytes of data */
1343 /* Calculate MSS. Not accounting for SACKs here. */
1344 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1346 /* Subtract TCP options size, not including SACKs */
1347 return __tcp_mtu_to_mss(sk, pmtu) -
1348 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1351 /* Inverse of above */
1352 int tcp_mss_to_mtu(struct sock *sk, int mss)
1354 const struct tcp_sock *tp = tcp_sk(sk);
1355 const struct inet_connection_sock *icsk = inet_csk(sk);
1359 tp->tcp_header_len +
1360 icsk->icsk_ext_hdr_len +
1361 icsk->icsk_af_ops->net_header_len;
1363 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1364 if (icsk->icsk_af_ops->net_frag_header_len) {
1365 const struct dst_entry *dst = __sk_dst_get(sk);
1367 if (dst && dst_allfrag(dst))
1368 mtu += icsk->icsk_af_ops->net_frag_header_len;
1372 EXPORT_SYMBOL(tcp_mss_to_mtu);
1374 /* MTU probing init per socket */
1375 void tcp_mtup_init(struct sock *sk)
1377 struct tcp_sock *tp = tcp_sk(sk);
1378 struct inet_connection_sock *icsk = inet_csk(sk);
1379 struct net *net = sock_net(sk);
1381 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1382 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1383 icsk->icsk_af_ops->net_header_len;
1384 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1385 icsk->icsk_mtup.probe_size = 0;
1386 if (icsk->icsk_mtup.enabled)
1387 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1389 EXPORT_SYMBOL(tcp_mtup_init);
1391 /* This function synchronize snd mss to current pmtu/exthdr set.
1393 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1394 for TCP options, but includes only bare TCP header.
1396 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1397 It is minimum of user_mss and mss received with SYN.
1398 It also does not include TCP options.
1400 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1402 tp->mss_cache is current effective sending mss, including
1403 all tcp options except for SACKs. It is evaluated,
1404 taking into account current pmtu, but never exceeds
1405 tp->rx_opt.mss_clamp.
1407 NOTE1. rfc1122 clearly states that advertised MSS
1408 DOES NOT include either tcp or ip options.
1410 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1411 are READ ONLY outside this function. --ANK (980731)
1413 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1415 struct tcp_sock *tp = tcp_sk(sk);
1416 struct inet_connection_sock *icsk = inet_csk(sk);
1419 if (icsk->icsk_mtup.search_high > pmtu)
1420 icsk->icsk_mtup.search_high = pmtu;
1422 mss_now = tcp_mtu_to_mss(sk, pmtu);
1423 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1425 /* And store cached results */
1426 icsk->icsk_pmtu_cookie = pmtu;
1427 if (icsk->icsk_mtup.enabled)
1428 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1429 tp->mss_cache = mss_now;
1433 EXPORT_SYMBOL(tcp_sync_mss);
1435 /* Compute the current effective MSS, taking SACKs and IP options,
1436 * and even PMTU discovery events into account.
1438 unsigned int tcp_current_mss(struct sock *sk)
1440 const struct tcp_sock *tp = tcp_sk(sk);
1441 const struct dst_entry *dst = __sk_dst_get(sk);
1443 unsigned int header_len;
1444 struct tcp_out_options opts;
1445 struct tcp_md5sig_key *md5;
1447 mss_now = tp->mss_cache;
1450 u32 mtu = dst_mtu(dst);
1451 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1452 mss_now = tcp_sync_mss(sk, mtu);
1455 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1456 sizeof(struct tcphdr);
1457 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1458 * some common options. If this is an odd packet (because we have SACK
1459 * blocks etc) then our calculated header_len will be different, and
1460 * we have to adjust mss_now correspondingly */
1461 if (header_len != tp->tcp_header_len) {
1462 int delta = (int) header_len - tp->tcp_header_len;
1469 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1470 * As additional protections, we do not touch cwnd in retransmission phases,
1471 * and if application hit its sndbuf limit recently.
1473 static void tcp_cwnd_application_limited(struct sock *sk)
1475 struct tcp_sock *tp = tcp_sk(sk);
1477 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1478 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1479 /* Limited by application or receiver window. */
1480 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1481 u32 win_used = max(tp->snd_cwnd_used, init_win);
1482 if (win_used < tp->snd_cwnd) {
1483 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1484 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1486 tp->snd_cwnd_used = 0;
1488 tp->snd_cwnd_stamp = tcp_time_stamp;
1491 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1493 struct tcp_sock *tp = tcp_sk(sk);
1495 /* Track the maximum number of outstanding packets in each
1496 * window, and remember whether we were cwnd-limited then.
1498 if (!before(tp->snd_una, tp->max_packets_seq) ||
1499 tp->packets_out > tp->max_packets_out) {
1500 tp->max_packets_out = tp->packets_out;
1501 tp->max_packets_seq = tp->snd_nxt;
1502 tp->is_cwnd_limited = is_cwnd_limited;
1505 if (tcp_is_cwnd_limited(sk)) {
1506 /* Network is feed fully. */
1507 tp->snd_cwnd_used = 0;
1508 tp->snd_cwnd_stamp = tcp_time_stamp;
1510 /* Network starves. */
1511 if (tp->packets_out > tp->snd_cwnd_used)
1512 tp->snd_cwnd_used = tp->packets_out;
1514 if (sysctl_tcp_slow_start_after_idle &&
1515 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1516 tcp_cwnd_application_limited(sk);
1520 /* Minshall's variant of the Nagle send check. */
1521 static bool tcp_minshall_check(const struct tcp_sock *tp)
1523 return after(tp->snd_sml, tp->snd_una) &&
1524 !after(tp->snd_sml, tp->snd_nxt);
1527 /* Update snd_sml if this skb is under mss
1528 * Note that a TSO packet might end with a sub-mss segment
1529 * The test is really :
1530 * if ((skb->len % mss) != 0)
1531 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1532 * But we can avoid doing the divide again given we already have
1533 * skb_pcount = skb->len / mss_now
1535 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1536 const struct sk_buff *skb)
1538 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1539 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1542 /* Return false, if packet can be sent now without violation Nagle's rules:
1543 * 1. It is full sized. (provided by caller in %partial bool)
1544 * 2. Or it contains FIN. (already checked by caller)
1545 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1546 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1547 * With Minshall's modification: all sent small packets are ACKed.
1549 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1553 ((nonagle & TCP_NAGLE_CORK) ||
1554 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1557 /* Return how many segs we'd like on a TSO packet,
1558 * to send one TSO packet per ms
1560 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1565 bytes = min(sk->sk_pacing_rate >> 10,
1566 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1568 /* Goal is to send at least one packet per ms,
1569 * not one big TSO packet every 100 ms.
1570 * This preserves ACK clocking and is consistent
1571 * with tcp_tso_should_defer() heuristic.
1573 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1575 return min_t(u32, segs, sk->sk_gso_max_segs);
1577 EXPORT_SYMBOL(tcp_tso_autosize);
1579 /* Return the number of segments we want in the skb we are transmitting.
1580 * See if congestion control module wants to decide; otherwise, autosize.
1582 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1584 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1585 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1588 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1591 /* Returns the portion of skb which can be sent right away */
1592 static unsigned int tcp_mss_split_point(const struct sock *sk,
1593 const struct sk_buff *skb,
1594 unsigned int mss_now,
1595 unsigned int max_segs,
1598 const struct tcp_sock *tp = tcp_sk(sk);
1599 u32 partial, needed, window, max_len;
1601 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1602 max_len = mss_now * max_segs;
1604 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1607 needed = min(skb->len, window);
1609 if (max_len <= needed)
1612 partial = needed % mss_now;
1613 /* If last segment is not a full MSS, check if Nagle rules allow us
1614 * to include this last segment in this skb.
1615 * Otherwise, we'll split the skb at last MSS boundary
1617 if (tcp_nagle_check(partial != 0, tp, nonagle))
1618 return needed - partial;
1623 /* Can at least one segment of SKB be sent right now, according to the
1624 * congestion window rules? If so, return how many segments are allowed.
1626 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1627 const struct sk_buff *skb)
1629 u32 in_flight, cwnd, halfcwnd;
1631 /* Don't be strict about the congestion window for the final FIN. */
1632 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1633 tcp_skb_pcount(skb) == 1)
1636 in_flight = tcp_packets_in_flight(tp);
1637 cwnd = tp->snd_cwnd;
1638 if (in_flight >= cwnd)
1641 /* For better scheduling, ensure we have at least
1642 * 2 GSO packets in flight.
1644 halfcwnd = max(cwnd >> 1, 1U);
1645 return min(halfcwnd, cwnd - in_flight);
1648 /* Initialize TSO state of a skb.
1649 * This must be invoked the first time we consider transmitting
1650 * SKB onto the wire.
1652 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1654 int tso_segs = tcp_skb_pcount(skb);
1656 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1657 tcp_set_skb_tso_segs(skb, mss_now);
1658 tso_segs = tcp_skb_pcount(skb);
1664 /* Return true if the Nagle test allows this packet to be
1667 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1668 unsigned int cur_mss, int nonagle)
1670 /* Nagle rule does not apply to frames, which sit in the middle of the
1671 * write_queue (they have no chances to get new data).
1673 * This is implemented in the callers, where they modify the 'nonagle'
1674 * argument based upon the location of SKB in the send queue.
1676 if (nonagle & TCP_NAGLE_PUSH)
1679 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1680 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1683 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1689 /* Does at least the first segment of SKB fit into the send window? */
1690 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1691 const struct sk_buff *skb,
1692 unsigned int cur_mss)
1694 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1696 if (skb->len > cur_mss)
1697 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1699 return !after(end_seq, tcp_wnd_end(tp));
1702 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1703 * should be put on the wire right now. If so, it returns the number of
1704 * packets allowed by the congestion window.
1706 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1707 unsigned int cur_mss, int nonagle)
1709 const struct tcp_sock *tp = tcp_sk(sk);
1710 unsigned int cwnd_quota;
1712 tcp_init_tso_segs(skb, cur_mss);
1714 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1717 cwnd_quota = tcp_cwnd_test(tp, skb);
1718 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1724 /* Test if sending is allowed right now. */
1725 bool tcp_may_send_now(struct sock *sk)
1727 const struct tcp_sock *tp = tcp_sk(sk);
1728 struct sk_buff *skb = tcp_send_head(sk);
1731 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1732 (tcp_skb_is_last(sk, skb) ?
1733 tp->nonagle : TCP_NAGLE_PUSH));
1736 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1737 * which is put after SKB on the list. It is very much like
1738 * tcp_fragment() except that it may make several kinds of assumptions
1739 * in order to speed up the splitting operation. In particular, we
1740 * know that all the data is in scatter-gather pages, and that the
1741 * packet has never been sent out before (and thus is not cloned).
1743 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1744 unsigned int mss_now, gfp_t gfp)
1746 struct sk_buff *buff;
1747 int nlen = skb->len - len;
1750 /* All of a TSO frame must be composed of paged data. */
1751 if (skb->len != skb->data_len)
1752 return tcp_fragment(sk, skb, len, mss_now, gfp);
1754 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1755 if (unlikely(!buff))
1758 sk->sk_wmem_queued += buff->truesize;
1759 sk_mem_charge(sk, buff->truesize);
1760 buff->truesize += nlen;
1761 skb->truesize -= nlen;
1763 /* Correct the sequence numbers. */
1764 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1765 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1766 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1768 /* PSH and FIN should only be set in the second packet. */
1769 flags = TCP_SKB_CB(skb)->tcp_flags;
1770 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1771 TCP_SKB_CB(buff)->tcp_flags = flags;
1773 /* This packet was never sent out yet, so no SACK bits. */
1774 TCP_SKB_CB(buff)->sacked = 0;
1776 tcp_skb_fragment_eor(skb, buff);
1778 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1779 skb_split(skb, buff, len);
1780 tcp_fragment_tstamp(skb, buff);
1782 /* Fix up tso_factor for both original and new SKB. */
1783 tcp_set_skb_tso_segs(skb, mss_now);
1784 tcp_set_skb_tso_segs(buff, mss_now);
1786 /* Link BUFF into the send queue. */
1787 __skb_header_release(buff);
1788 tcp_insert_write_queue_after(skb, buff, sk);
1793 /* Try to defer sending, if possible, in order to minimize the amount
1794 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1796 * This algorithm is from John Heffner.
1798 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1799 bool *is_cwnd_limited, u32 max_segs)
1801 const struct inet_connection_sock *icsk = inet_csk(sk);
1802 u32 age, send_win, cong_win, limit, in_flight;
1803 struct tcp_sock *tp = tcp_sk(sk);
1804 struct skb_mstamp now;
1805 struct sk_buff *head;
1808 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1811 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1814 /* Avoid bursty behavior by allowing defer
1815 * only if the last write was recent.
1817 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1820 in_flight = tcp_packets_in_flight(tp);
1822 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1824 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1826 /* From in_flight test above, we know that cwnd > in_flight. */
1827 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1829 limit = min(send_win, cong_win);
1831 /* If a full-sized TSO skb can be sent, do it. */
1832 if (limit >= max_segs * tp->mss_cache)
1835 /* Middle in queue won't get any more data, full sendable already? */
1836 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1839 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1841 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1843 /* If at least some fraction of a window is available,
1846 chunk /= win_divisor;
1850 /* Different approach, try not to defer past a single
1851 * ACK. Receiver should ACK every other full sized
1852 * frame, so if we have space for more than 3 frames
1855 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1859 head = tcp_write_queue_head(sk);
1860 skb_mstamp_get(&now);
1861 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1862 /* If next ACK is likely to come too late (half srtt), do not defer */
1863 if (age < (tp->srtt_us >> 4))
1866 /* Ok, it looks like it is advisable to defer. */
1868 if (cong_win < send_win && cong_win <= skb->len)
1869 *is_cwnd_limited = true;
1877 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1879 struct inet_connection_sock *icsk = inet_csk(sk);
1880 struct tcp_sock *tp = tcp_sk(sk);
1881 struct net *net = sock_net(sk);
1885 interval = net->ipv4.sysctl_tcp_probe_interval;
1886 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1887 if (unlikely(delta >= interval * HZ)) {
1888 int mss = tcp_current_mss(sk);
1890 /* Update current search range */
1891 icsk->icsk_mtup.probe_size = 0;
1892 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1893 sizeof(struct tcphdr) +
1894 icsk->icsk_af_ops->net_header_len;
1895 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1897 /* Update probe time stamp */
1898 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1902 /* Create a new MTU probe if we are ready.
1903 * MTU probe is regularly attempting to increase the path MTU by
1904 * deliberately sending larger packets. This discovers routing
1905 * changes resulting in larger path MTUs.
1907 * Returns 0 if we should wait to probe (no cwnd available),
1908 * 1 if a probe was sent,
1911 static int tcp_mtu_probe(struct sock *sk)
1913 struct tcp_sock *tp = tcp_sk(sk);
1914 struct inet_connection_sock *icsk = inet_csk(sk);
1915 struct sk_buff *skb, *nskb, *next;
1916 struct net *net = sock_net(sk);
1924 /* Not currently probing/verifying,
1926 * have enough cwnd, and
1927 * not SACKing (the variable headers throw things off) */
1928 if (!icsk->icsk_mtup.enabled ||
1929 icsk->icsk_mtup.probe_size ||
1930 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1931 tp->snd_cwnd < 11 ||
1932 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1935 /* Use binary search for probe_size between tcp_mss_base,
1936 * and current mss_clamp. if (search_high - search_low)
1937 * smaller than a threshold, backoff from probing.
1939 mss_now = tcp_current_mss(sk);
1940 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1941 icsk->icsk_mtup.search_low) >> 1);
1942 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1943 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1944 /* When misfortune happens, we are reprobing actively,
1945 * and then reprobe timer has expired. We stick with current
1946 * probing process by not resetting search range to its orignal.
1948 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1949 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1950 /* Check whether enough time has elaplased for
1951 * another round of probing.
1953 tcp_mtu_check_reprobe(sk);
1957 /* Have enough data in the send queue to probe? */
1958 if (tp->write_seq - tp->snd_nxt < size_needed)
1961 if (tp->snd_wnd < size_needed)
1963 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1966 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1967 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1968 if (!tcp_packets_in_flight(tp))
1974 /* We're allowed to probe. Build it now. */
1975 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1978 sk->sk_wmem_queued += nskb->truesize;
1979 sk_mem_charge(sk, nskb->truesize);
1981 skb = tcp_send_head(sk);
1983 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1984 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1985 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1986 TCP_SKB_CB(nskb)->sacked = 0;
1988 nskb->ip_summed = skb->ip_summed;
1990 tcp_insert_write_queue_before(nskb, skb, sk);
1993 tcp_for_write_queue_from_safe(skb, next, sk) {
1994 copy = min_t(int, skb->len, probe_size - len);
1995 if (nskb->ip_summed)
1996 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1998 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1999 skb_put(nskb, copy),
2002 if (skb->len <= copy) {
2003 /* We've eaten all the data from this skb.
2005 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2006 tcp_unlink_write_queue(skb, sk);
2007 sk_wmem_free_skb(sk, skb);
2009 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2010 ~(TCPHDR_FIN|TCPHDR_PSH);
2011 if (!skb_shinfo(skb)->nr_frags) {
2012 skb_pull(skb, copy);
2013 if (skb->ip_summed != CHECKSUM_PARTIAL)
2014 skb->csum = csum_partial(skb->data,
2017 __pskb_trim_head(skb, copy);
2018 tcp_set_skb_tso_segs(skb, mss_now);
2020 TCP_SKB_CB(skb)->seq += copy;
2025 if (len >= probe_size)
2028 tcp_init_tso_segs(nskb, nskb->len);
2030 /* We're ready to send. If this fails, the probe will
2031 * be resegmented into mss-sized pieces by tcp_write_xmit().
2033 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2034 /* Decrement cwnd here because we are sending
2035 * effectively two packets. */
2037 tcp_event_new_data_sent(sk, nskb);
2039 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2040 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2041 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2049 /* TCP Small Queues :
2050 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2051 * (These limits are doubled for retransmits)
2053 * - better RTT estimation and ACK scheduling
2056 * Alas, some drivers / subsystems require a fair amount
2057 * of queued bytes to ensure line rate.
2058 * One example is wifi aggregation (802.11 AMPDU)
2060 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2061 unsigned int factor)
2065 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2066 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2069 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2070 set_bit(TSQ_THROTTLED, &tcp_sk(sk)->tsq_flags);
2071 /* It is possible TX completion already happened
2072 * before we set TSQ_THROTTLED, so we must
2073 * test again the condition.
2075 smp_mb__after_atomic();
2076 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2082 /* This routine writes packets to the network. It advances the
2083 * send_head. This happens as incoming acks open up the remote
2086 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2087 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2088 * account rare use of URG, this is not a big flaw.
2090 * Send at most one packet when push_one > 0. Temporarily ignore
2091 * cwnd limit to force at most one packet out when push_one == 2.
2093 * Returns true, if no segments are in flight and we have queued segments,
2094 * but cannot send anything now because of SWS or another problem.
2096 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2097 int push_one, gfp_t gfp)
2099 struct tcp_sock *tp = tcp_sk(sk);
2100 struct sk_buff *skb;
2101 unsigned int tso_segs, sent_pkts;
2104 bool is_cwnd_limited = false;
2110 /* Do MTU probing. */
2111 result = tcp_mtu_probe(sk);
2114 } else if (result > 0) {
2119 max_segs = tcp_tso_segs(sk, mss_now);
2120 while ((skb = tcp_send_head(sk))) {
2123 tso_segs = tcp_init_tso_segs(skb, mss_now);
2126 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2127 /* "skb_mstamp" is used as a start point for the retransmit timer */
2128 skb_mstamp_get(&skb->skb_mstamp);
2129 goto repair; /* Skip network transmission */
2132 cwnd_quota = tcp_cwnd_test(tp, skb);
2135 /* Force out a loss probe pkt. */
2141 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2144 if (tso_segs == 1) {
2145 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2146 (tcp_skb_is_last(sk, skb) ?
2147 nonagle : TCP_NAGLE_PUSH))))
2151 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2157 if (tso_segs > 1 && !tcp_urg_mode(tp))
2158 limit = tcp_mss_split_point(sk, skb, mss_now,
2164 if (skb->len > limit &&
2165 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2168 if (tcp_small_queue_check(sk, skb, 0))
2171 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2175 /* Advance the send_head. This one is sent out.
2176 * This call will increment packets_out.
2178 tcp_event_new_data_sent(sk, skb);
2180 tcp_minshall_update(tp, mss_now, skb);
2181 sent_pkts += tcp_skb_pcount(skb);
2187 if (likely(sent_pkts)) {
2188 if (tcp_in_cwnd_reduction(sk))
2189 tp->prr_out += sent_pkts;
2191 /* Send one loss probe per tail loss episode. */
2193 tcp_schedule_loss_probe(sk);
2194 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2195 tcp_cwnd_validate(sk, is_cwnd_limited);
2198 return !tp->packets_out && tcp_send_head(sk);
2201 bool tcp_schedule_loss_probe(struct sock *sk)
2203 struct inet_connection_sock *icsk = inet_csk(sk);
2204 struct tcp_sock *tp = tcp_sk(sk);
2205 u32 timeout, tlp_time_stamp, rto_time_stamp;
2206 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2208 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2210 /* No consecutive loss probes. */
2211 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2215 /* Don't do any loss probe on a Fast Open connection before 3WHS
2218 if (tp->fastopen_rsk)
2221 /* TLP is only scheduled when next timer event is RTO. */
2222 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2225 /* Schedule a loss probe in 2*RTT for SACK capable connections
2226 * in Open state, that are either limited by cwnd or application.
2228 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2229 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2232 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2236 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2237 * for delayed ack when there's one outstanding packet. If no RTT
2238 * sample is available then probe after TCP_TIMEOUT_INIT.
2240 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2241 if (tp->packets_out == 1)
2242 timeout = max_t(u32, timeout,
2243 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2244 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2246 /* If RTO is shorter, just schedule TLP in its place. */
2247 tlp_time_stamp = tcp_time_stamp + timeout;
2248 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2249 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2250 s32 delta = rto_time_stamp - tcp_time_stamp;
2255 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2260 /* Thanks to skb fast clones, we can detect if a prior transmit of
2261 * a packet is still in a qdisc or driver queue.
2262 * In this case, there is very little point doing a retransmit !
2264 static bool skb_still_in_host_queue(const struct sock *sk,
2265 const struct sk_buff *skb)
2267 if (unlikely(skb_fclone_busy(sk, skb))) {
2268 NET_INC_STATS(sock_net(sk),
2269 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2275 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2276 * retransmit the last segment.
2278 void tcp_send_loss_probe(struct sock *sk)
2280 struct tcp_sock *tp = tcp_sk(sk);
2281 struct sk_buff *skb;
2283 int mss = tcp_current_mss(sk);
2285 skb = tcp_send_head(sk);
2287 if (tcp_snd_wnd_test(tp, skb, mss)) {
2288 pcount = tp->packets_out;
2289 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2290 if (tp->packets_out > pcount)
2294 skb = tcp_write_queue_prev(sk, skb);
2296 skb = tcp_write_queue_tail(sk);
2299 /* At most one outstanding TLP retransmission. */
2300 if (tp->tlp_high_seq)
2303 /* Retransmit last segment. */
2307 if (skb_still_in_host_queue(sk, skb))
2310 pcount = tcp_skb_pcount(skb);
2311 if (WARN_ON(!pcount))
2314 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2315 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2318 skb = tcp_write_queue_next(sk, skb);
2321 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2324 if (__tcp_retransmit_skb(sk, skb, 1))
2327 /* Record snd_nxt for loss detection. */
2328 tp->tlp_high_seq = tp->snd_nxt;
2331 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2332 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2333 inet_csk(sk)->icsk_pending = 0;
2338 /* Push out any pending frames which were held back due to
2339 * TCP_CORK or attempt at coalescing tiny packets.
2340 * The socket must be locked by the caller.
2342 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2345 /* If we are closed, the bytes will have to remain here.
2346 * In time closedown will finish, we empty the write queue and
2347 * all will be happy.
2349 if (unlikely(sk->sk_state == TCP_CLOSE))
2352 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2353 sk_gfp_mask(sk, GFP_ATOMIC)))
2354 tcp_check_probe_timer(sk);
2357 /* Send _single_ skb sitting at the send head. This function requires
2358 * true push pending frames to setup probe timer etc.
2360 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2362 struct sk_buff *skb = tcp_send_head(sk);
2364 BUG_ON(!skb || skb->len < mss_now);
2366 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2369 /* This function returns the amount that we can raise the
2370 * usable window based on the following constraints
2372 * 1. The window can never be shrunk once it is offered (RFC 793)
2373 * 2. We limit memory per socket
2376 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2377 * RECV.NEXT + RCV.WIN fixed until:
2378 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2380 * i.e. don't raise the right edge of the window until you can raise
2381 * it at least MSS bytes.
2383 * Unfortunately, the recommended algorithm breaks header prediction,
2384 * since header prediction assumes th->window stays fixed.
2386 * Strictly speaking, keeping th->window fixed violates the receiver
2387 * side SWS prevention criteria. The problem is that under this rule
2388 * a stream of single byte packets will cause the right side of the
2389 * window to always advance by a single byte.
2391 * Of course, if the sender implements sender side SWS prevention
2392 * then this will not be a problem.
2394 * BSD seems to make the following compromise:
2396 * If the free space is less than the 1/4 of the maximum
2397 * space available and the free space is less than 1/2 mss,
2398 * then set the window to 0.
2399 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2400 * Otherwise, just prevent the window from shrinking
2401 * and from being larger than the largest representable value.
2403 * This prevents incremental opening of the window in the regime
2404 * where TCP is limited by the speed of the reader side taking
2405 * data out of the TCP receive queue. It does nothing about
2406 * those cases where the window is constrained on the sender side
2407 * because the pipeline is full.
2409 * BSD also seems to "accidentally" limit itself to windows that are a
2410 * multiple of MSS, at least until the free space gets quite small.
2411 * This would appear to be a side effect of the mbuf implementation.
2412 * Combining these two algorithms results in the observed behavior
2413 * of having a fixed window size at almost all times.
2415 * Below we obtain similar behavior by forcing the offered window to
2416 * a multiple of the mss when it is feasible to do so.
2418 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2419 * Regular options like TIMESTAMP are taken into account.
2421 u32 __tcp_select_window(struct sock *sk)
2423 struct inet_connection_sock *icsk = inet_csk(sk);
2424 struct tcp_sock *tp = tcp_sk(sk);
2425 /* MSS for the peer's data. Previous versions used mss_clamp
2426 * here. I don't know if the value based on our guesses
2427 * of peer's MSS is better for the performance. It's more correct
2428 * but may be worse for the performance because of rcv_mss
2429 * fluctuations. --SAW 1998/11/1
2431 int mss = icsk->icsk_ack.rcv_mss;
2432 int free_space = tcp_space(sk);
2433 int allowed_space = tcp_full_space(sk);
2434 int full_space = min_t(int, tp->window_clamp, allowed_space);
2437 if (mss > full_space)
2440 if (free_space < (full_space >> 1)) {
2441 icsk->icsk_ack.quick = 0;
2443 if (tcp_under_memory_pressure(sk))
2444 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2447 /* free_space might become our new window, make sure we don't
2448 * increase it due to wscale.
2450 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2452 /* if free space is less than mss estimate, or is below 1/16th
2453 * of the maximum allowed, try to move to zero-window, else
2454 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2455 * new incoming data is dropped due to memory limits.
2456 * With large window, mss test triggers way too late in order
2457 * to announce zero window in time before rmem limit kicks in.
2459 if (free_space < (allowed_space >> 4) || free_space < mss)
2463 if (free_space > tp->rcv_ssthresh)
2464 free_space = tp->rcv_ssthresh;
2466 /* Don't do rounding if we are using window scaling, since the
2467 * scaled window will not line up with the MSS boundary anyway.
2469 window = tp->rcv_wnd;
2470 if (tp->rx_opt.rcv_wscale) {
2471 window = free_space;
2473 /* Advertise enough space so that it won't get scaled away.
2474 * Import case: prevent zero window announcement if
2475 * 1<<rcv_wscale > mss.
2477 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2478 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2479 << tp->rx_opt.rcv_wscale);
2481 /* Get the largest window that is a nice multiple of mss.
2482 * Window clamp already applied above.
2483 * If our current window offering is within 1 mss of the
2484 * free space we just keep it. This prevents the divide
2485 * and multiply from happening most of the time.
2486 * We also don't do any window rounding when the free space
2489 if (window <= free_space - mss || window > free_space)
2490 window = (free_space / mss) * mss;
2491 else if (mss == full_space &&
2492 free_space > window + (full_space >> 1))
2493 window = free_space;
2499 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2500 const struct sk_buff *next_skb)
2502 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2503 const struct skb_shared_info *next_shinfo =
2504 skb_shinfo(next_skb);
2505 struct skb_shared_info *shinfo = skb_shinfo(skb);
2507 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2508 shinfo->tskey = next_shinfo->tskey;
2509 TCP_SKB_CB(skb)->txstamp_ack |=
2510 TCP_SKB_CB(next_skb)->txstamp_ack;
2514 /* Collapses two adjacent SKB's during retransmission. */
2515 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2517 struct tcp_sock *tp = tcp_sk(sk);
2518 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2519 int skb_size, next_skb_size;
2521 skb_size = skb->len;
2522 next_skb_size = next_skb->len;
2524 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2526 tcp_highest_sack_combine(sk, next_skb, skb);
2528 tcp_unlink_write_queue(next_skb, sk);
2530 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2533 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2534 skb->ip_summed = CHECKSUM_PARTIAL;
2536 if (skb->ip_summed != CHECKSUM_PARTIAL)
2537 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2539 /* Update sequence range on original skb. */
2540 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2542 /* Merge over control information. This moves PSH/FIN etc. over */
2543 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2545 /* All done, get rid of second SKB and account for it so
2546 * packet counting does not break.
2548 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2549 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2551 /* changed transmit queue under us so clear hints */
2552 tcp_clear_retrans_hints_partial(tp);
2553 if (next_skb == tp->retransmit_skb_hint)
2554 tp->retransmit_skb_hint = skb;
2556 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2558 tcp_skb_collapse_tstamp(skb, next_skb);
2560 sk_wmem_free_skb(sk, next_skb);
2563 /* Check if coalescing SKBs is legal. */
2564 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2566 if (tcp_skb_pcount(skb) > 1)
2568 /* TODO: SACK collapsing could be used to remove this condition */
2569 if (skb_shinfo(skb)->nr_frags != 0)
2571 if (skb_cloned(skb))
2573 if (skb == tcp_send_head(sk))
2575 /* Some heurestics for collapsing over SACK'd could be invented */
2576 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2582 /* Collapse packets in the retransmit queue to make to create
2583 * less packets on the wire. This is only done on retransmission.
2585 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2588 struct tcp_sock *tp = tcp_sk(sk);
2589 struct sk_buff *skb = to, *tmp;
2592 if (!sysctl_tcp_retrans_collapse)
2594 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2597 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2598 if (!tcp_can_collapse(sk, skb))
2601 if (!tcp_skb_can_collapse_to(to))
2613 /* Punt if not enough space exists in the first SKB for
2614 * the data in the second
2616 if (skb->len > skb_availroom(to))
2619 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2622 tcp_collapse_retrans(sk, to);
2626 /* This retransmits one SKB. Policy decisions and retransmit queue
2627 * state updates are done by the caller. Returns non-zero if an
2628 * error occurred which prevented the send.
2630 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2632 struct inet_connection_sock *icsk = inet_csk(sk);
2633 struct tcp_sock *tp = tcp_sk(sk);
2634 unsigned int cur_mss;
2638 /* Inconclusive MTU probe */
2639 if (icsk->icsk_mtup.probe_size)
2640 icsk->icsk_mtup.probe_size = 0;
2642 /* Do not sent more than we queued. 1/4 is reserved for possible
2643 * copying overhead: fragmentation, tunneling, mangling etc.
2645 if (atomic_read(&sk->sk_wmem_alloc) >
2646 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2650 if (skb_still_in_host_queue(sk, skb))
2653 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2654 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2656 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2660 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2661 return -EHOSTUNREACH; /* Routing failure or similar. */
2663 cur_mss = tcp_current_mss(sk);
2665 /* If receiver has shrunk his window, and skb is out of
2666 * new window, do not retransmit it. The exception is the
2667 * case, when window is shrunk to zero. In this case
2668 * our retransmit serves as a zero window probe.
2670 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2671 TCP_SKB_CB(skb)->seq != tp->snd_una)
2674 len = cur_mss * segs;
2675 if (skb->len > len) {
2676 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2677 return -ENOMEM; /* We'll try again later. */
2679 if (skb_unclone(skb, GFP_ATOMIC))
2682 diff = tcp_skb_pcount(skb);
2683 tcp_set_skb_tso_segs(skb, cur_mss);
2684 diff -= tcp_skb_pcount(skb);
2686 tcp_adjust_pcount(sk, skb, diff);
2687 if (skb->len < cur_mss)
2688 tcp_retrans_try_collapse(sk, skb, cur_mss);
2691 /* RFC3168, section 6.1.1.1. ECN fallback */
2692 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2693 tcp_ecn_clear_syn(sk, skb);
2695 /* make sure skb->data is aligned on arches that require it
2696 * and check if ack-trimming & collapsing extended the headroom
2697 * beyond what csum_start can cover.
2699 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2700 skb_headroom(skb) >= 0xFFFF)) {
2701 struct sk_buff *nskb;
2703 skb_mstamp_get(&skb->skb_mstamp);
2704 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2705 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2708 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2712 segs = tcp_skb_pcount(skb);
2714 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2715 /* Update global TCP statistics. */
2716 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2717 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2718 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2719 tp->total_retrans += segs;
2724 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2726 struct tcp_sock *tp = tcp_sk(sk);
2727 int err = __tcp_retransmit_skb(sk, skb, segs);
2730 #if FASTRETRANS_DEBUG > 0
2731 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2732 net_dbg_ratelimited("retrans_out leaked\n");
2735 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2736 tp->retrans_out += tcp_skb_pcount(skb);
2738 /* Save stamp of the first retransmit. */
2739 if (!tp->retrans_stamp)
2740 tp->retrans_stamp = tcp_skb_timestamp(skb);
2742 } else if (err != -EBUSY) {
2743 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2746 if (tp->undo_retrans < 0)
2747 tp->undo_retrans = 0;
2748 tp->undo_retrans += tcp_skb_pcount(skb);
2752 /* Check if we forward retransmits are possible in the current
2753 * window/congestion state.
2755 static bool tcp_can_forward_retransmit(struct sock *sk)
2757 const struct inet_connection_sock *icsk = inet_csk(sk);
2758 const struct tcp_sock *tp = tcp_sk(sk);
2760 /* Forward retransmissions are possible only during Recovery. */
2761 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2764 /* No forward retransmissions in Reno are possible. */
2765 if (tcp_is_reno(tp))
2768 /* Yeah, we have to make difficult choice between forward transmission
2769 * and retransmission... Both ways have their merits...
2771 * For now we do not retransmit anything, while we have some new
2772 * segments to send. In the other cases, follow rule 3 for
2773 * NextSeg() specified in RFC3517.
2776 if (tcp_may_send_now(sk))
2782 /* This gets called after a retransmit timeout, and the initially
2783 * retransmitted data is acknowledged. It tries to continue
2784 * resending the rest of the retransmit queue, until either
2785 * we've sent it all or the congestion window limit is reached.
2786 * If doing SACK, the first ACK which comes back for a timeout
2787 * based retransmit packet might feed us FACK information again.
2788 * If so, we use it to avoid unnecessarily retransmissions.
2790 void tcp_xmit_retransmit_queue(struct sock *sk)
2792 const struct inet_connection_sock *icsk = inet_csk(sk);
2793 struct tcp_sock *tp = tcp_sk(sk);
2794 struct sk_buff *skb;
2795 struct sk_buff *hole = NULL;
2796 u32 max_segs, last_lost;
2798 int fwd_rexmitting = 0;
2800 if (!tp->packets_out)
2804 tp->retransmit_high = tp->snd_una;
2806 if (tp->retransmit_skb_hint) {
2807 skb = tp->retransmit_skb_hint;
2808 last_lost = TCP_SKB_CB(skb)->end_seq;
2809 if (after(last_lost, tp->retransmit_high))
2810 last_lost = tp->retransmit_high;
2812 skb = tcp_write_queue_head(sk);
2813 last_lost = tp->snd_una;
2816 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2817 tcp_for_write_queue_from(skb, sk) {
2821 if (skb == tcp_send_head(sk))
2823 /* we could do better than to assign each time */
2825 tp->retransmit_skb_hint = skb;
2827 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2830 sacked = TCP_SKB_CB(skb)->sacked;
2831 /* In case tcp_shift_skb_data() have aggregated large skbs,
2832 * we need to make sure not sending too bigs TSO packets
2834 segs = min_t(int, segs, max_segs);
2836 if (fwd_rexmitting) {
2838 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2840 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2842 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2843 tp->retransmit_high = last_lost;
2844 if (!tcp_can_forward_retransmit(sk))
2846 /* Backtrack if necessary to non-L'ed skb */
2854 } else if (!(sacked & TCPCB_LOST)) {
2855 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2860 last_lost = TCP_SKB_CB(skb)->end_seq;
2861 if (icsk->icsk_ca_state != TCP_CA_Loss)
2862 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2864 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2867 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2870 if (tcp_small_queue_check(sk, skb, 1))
2873 if (tcp_retransmit_skb(sk, skb, segs))
2876 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2878 if (tcp_in_cwnd_reduction(sk))
2879 tp->prr_out += tcp_skb_pcount(skb);
2881 if (skb == tcp_write_queue_head(sk))
2882 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2883 inet_csk(sk)->icsk_rto,
2888 /* We allow to exceed memory limits for FIN packets to expedite
2889 * connection tear down and (memory) recovery.
2890 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2891 * or even be forced to close flow without any FIN.
2892 * In general, we want to allow one skb per socket to avoid hangs
2893 * with edge trigger epoll()
2895 void sk_forced_mem_schedule(struct sock *sk, int size)
2899 if (size <= sk->sk_forward_alloc)
2901 amt = sk_mem_pages(size);
2902 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2903 sk_memory_allocated_add(sk, amt);
2905 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2906 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2909 /* Send a FIN. The caller locks the socket for us.
2910 * We should try to send a FIN packet really hard, but eventually give up.
2912 void tcp_send_fin(struct sock *sk)
2914 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2915 struct tcp_sock *tp = tcp_sk(sk);
2917 /* Optimization, tack on the FIN if we have one skb in write queue and
2918 * this skb was not yet sent, or we are under memory pressure.
2919 * Note: in the latter case, FIN packet will be sent after a timeout,
2920 * as TCP stack thinks it has already been transmitted.
2922 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2924 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2925 TCP_SKB_CB(tskb)->end_seq++;
2927 if (!tcp_send_head(sk)) {
2928 /* This means tskb was already sent.
2929 * Pretend we included the FIN on previous transmit.
2930 * We need to set tp->snd_nxt to the value it would have
2931 * if FIN had been sent. This is because retransmit path
2932 * does not change tp->snd_nxt.
2938 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2939 if (unlikely(!skb)) {
2944 skb_reserve(skb, MAX_TCP_HEADER);
2945 sk_forced_mem_schedule(sk, skb->truesize);
2946 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2947 tcp_init_nondata_skb(skb, tp->write_seq,
2948 TCPHDR_ACK | TCPHDR_FIN);
2949 tcp_queue_skb(sk, skb);
2951 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2954 /* We get here when a process closes a file descriptor (either due to
2955 * an explicit close() or as a byproduct of exit()'ing) and there
2956 * was unread data in the receive queue. This behavior is recommended
2957 * by RFC 2525, section 2.17. -DaveM
2959 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2961 struct sk_buff *skb;
2963 /* NOTE: No TCP options attached and we never retransmit this. */
2964 skb = alloc_skb(MAX_TCP_HEADER, priority);
2966 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2970 /* Reserve space for headers and prepare control bits. */
2971 skb_reserve(skb, MAX_TCP_HEADER);
2972 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2973 TCPHDR_ACK | TCPHDR_RST);
2974 skb_mstamp_get(&skb->skb_mstamp);
2976 if (tcp_transmit_skb(sk, skb, 0, priority))
2977 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2979 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2982 /* Send a crossed SYN-ACK during socket establishment.
2983 * WARNING: This routine must only be called when we have already sent
2984 * a SYN packet that crossed the incoming SYN that caused this routine
2985 * to get called. If this assumption fails then the initial rcv_wnd
2986 * and rcv_wscale values will not be correct.
2988 int tcp_send_synack(struct sock *sk)
2990 struct sk_buff *skb;
2992 skb = tcp_write_queue_head(sk);
2993 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2994 pr_debug("%s: wrong queue state\n", __func__);
2997 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2998 if (skb_cloned(skb)) {
2999 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3002 tcp_unlink_write_queue(skb, sk);
3003 __skb_header_release(nskb);
3004 __tcp_add_write_queue_head(sk, nskb);
3005 sk_wmem_free_skb(sk, skb);
3006 sk->sk_wmem_queued += nskb->truesize;
3007 sk_mem_charge(sk, nskb->truesize);
3011 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3012 tcp_ecn_send_synack(sk, skb);
3014 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3018 * tcp_make_synack - Prepare a SYN-ACK.
3019 * sk: listener socket
3020 * dst: dst entry attached to the SYNACK
3021 * req: request_sock pointer
3023 * Allocate one skb and build a SYNACK packet.
3024 * @dst is consumed : Caller should not use it again.
3026 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3027 struct request_sock *req,
3028 struct tcp_fastopen_cookie *foc,
3029 enum tcp_synack_type synack_type)
3031 struct inet_request_sock *ireq = inet_rsk(req);
3032 const struct tcp_sock *tp = tcp_sk(sk);
3033 struct tcp_md5sig_key *md5 = NULL;
3034 struct tcp_out_options opts;
3035 struct sk_buff *skb;
3036 int tcp_header_size;
3041 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3042 if (unlikely(!skb)) {
3046 /* Reserve space for headers. */
3047 skb_reserve(skb, MAX_TCP_HEADER);
3049 switch (synack_type) {
3050 case TCP_SYNACK_NORMAL:
3051 skb_set_owner_w(skb, req_to_sk(req));
3053 case TCP_SYNACK_COOKIE:
3054 /* Under synflood, we do not attach skb to a socket,
3055 * to avoid false sharing.
3058 case TCP_SYNACK_FASTOPEN:
3059 /* sk is a const pointer, because we want to express multiple
3060 * cpu might call us concurrently.
3061 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3063 skb_set_owner_w(skb, (struct sock *)sk);
3066 skb_dst_set(skb, dst);
3068 mss = dst_metric_advmss(dst);
3069 user_mss = READ_ONCE(tp->rx_opt.user_mss);
3070 if (user_mss && user_mss < mss)
3073 memset(&opts, 0, sizeof(opts));
3074 #ifdef CONFIG_SYN_COOKIES
3075 if (unlikely(req->cookie_ts))
3076 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3079 skb_mstamp_get(&skb->skb_mstamp);
3081 #ifdef CONFIG_TCP_MD5SIG
3083 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3085 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3086 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3089 skb_push(skb, tcp_header_size);
3090 skb_reset_transport_header(skb);
3092 th = (struct tcphdr *)skb->data;
3093 memset(th, 0, sizeof(struct tcphdr));
3096 tcp_ecn_make_synack(req, th);
3097 th->source = htons(ireq->ir_num);
3098 th->dest = ireq->ir_rmt_port;
3099 /* Setting of flags are superfluous here for callers (and ECE is
3100 * not even correctly set)
3102 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3103 TCPHDR_SYN | TCPHDR_ACK);
3105 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3106 /* XXX data is queued and acked as is. No buffer/window check */
3107 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3109 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3110 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3111 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3112 th->doff = (tcp_header_size >> 2);
3113 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3115 #ifdef CONFIG_TCP_MD5SIG
3116 /* Okay, we have all we need - do the md5 hash if needed */
3118 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3119 md5, req_to_sk(req), skb);
3123 /* Do not fool tcpdump (if any), clean our debris */
3124 skb->tstamp.tv64 = 0;
3127 EXPORT_SYMBOL(tcp_make_synack);
3129 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3131 struct inet_connection_sock *icsk = inet_csk(sk);
3132 const struct tcp_congestion_ops *ca;
3133 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3135 if (ca_key == TCP_CA_UNSPEC)
3139 ca = tcp_ca_find_key(ca_key);
3140 if (likely(ca && try_module_get(ca->owner))) {
3141 module_put(icsk->icsk_ca_ops->owner);
3142 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3143 icsk->icsk_ca_ops = ca;
3148 /* Do all connect socket setups that can be done AF independent. */
3149 static void tcp_connect_init(struct sock *sk)
3151 const struct dst_entry *dst = __sk_dst_get(sk);
3152 struct tcp_sock *tp = tcp_sk(sk);
3155 /* We'll fix this up when we get a response from the other end.
3156 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3158 tp->tcp_header_len = sizeof(struct tcphdr) +
3159 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3161 #ifdef CONFIG_TCP_MD5SIG
3162 if (tp->af_specific->md5_lookup(sk, sk))
3163 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3166 /* If user gave his TCP_MAXSEG, record it to clamp */
3167 if (tp->rx_opt.user_mss)
3168 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3171 tcp_sync_mss(sk, dst_mtu(dst));
3173 tcp_ca_dst_init(sk, dst);
3175 if (!tp->window_clamp)
3176 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3177 tp->advmss = dst_metric_advmss(dst);
3178 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3179 tp->advmss = tp->rx_opt.user_mss;
3181 tcp_initialize_rcv_mss(sk);
3183 /* limit the window selection if the user enforce a smaller rx buffer */
3184 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3185 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3186 tp->window_clamp = tcp_full_space(sk);
3188 tcp_select_initial_window(tcp_full_space(sk),
3189 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3192 sysctl_tcp_window_scaling,
3194 dst_metric(dst, RTAX_INITRWND));
3196 tp->rx_opt.rcv_wscale = rcv_wscale;
3197 tp->rcv_ssthresh = tp->rcv_wnd;
3200 sock_reset_flag(sk, SOCK_DONE);
3203 tp->snd_una = tp->write_seq;
3204 tp->snd_sml = tp->write_seq;
3205 tp->snd_up = tp->write_seq;
3206 tp->snd_nxt = tp->write_seq;
3208 if (likely(!tp->repair))
3211 tp->rcv_tstamp = tcp_time_stamp;
3212 tp->rcv_wup = tp->rcv_nxt;
3213 tp->copied_seq = tp->rcv_nxt;
3215 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3216 inet_csk(sk)->icsk_retransmits = 0;
3217 tcp_clear_retrans(tp);
3220 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3222 struct tcp_sock *tp = tcp_sk(sk);
3223 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3225 tcb->end_seq += skb->len;
3226 __skb_header_release(skb);
3227 __tcp_add_write_queue_tail(sk, skb);
3228 sk->sk_wmem_queued += skb->truesize;
3229 sk_mem_charge(sk, skb->truesize);
3230 tp->write_seq = tcb->end_seq;
3231 tp->packets_out += tcp_skb_pcount(skb);
3234 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3235 * queue a data-only packet after the regular SYN, such that regular SYNs
3236 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3237 * only the SYN sequence, the data are retransmitted in the first ACK.
3238 * If cookie is not cached or other error occurs, falls back to send a
3239 * regular SYN with Fast Open cookie request option.
3241 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3243 struct tcp_sock *tp = tcp_sk(sk);
3244 struct tcp_fastopen_request *fo = tp->fastopen_req;
3245 int syn_loss = 0, space, err = 0;
3246 unsigned long last_syn_loss = 0;
3247 struct sk_buff *syn_data;
3249 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3250 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3251 &syn_loss, &last_syn_loss);
3252 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3254 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3255 fo->cookie.len = -1;
3259 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3260 fo->cookie.len = -1;
3261 else if (fo->cookie.len <= 0)
3264 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3265 * user-MSS. Reserve maximum option space for middleboxes that add
3266 * private TCP options. The cost is reduced data space in SYN :(
3268 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3269 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3270 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3271 MAX_TCP_OPTION_SPACE;
3273 space = min_t(size_t, space, fo->size);
3275 /* limit to order-0 allocations */
3276 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3278 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3281 syn_data->ip_summed = CHECKSUM_PARTIAL;
3282 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3284 int copied = copy_from_iter(skb_put(syn_data, space), space,
3285 &fo->data->msg_iter);
3286 if (unlikely(!copied)) {
3287 kfree_skb(syn_data);
3290 if (copied != space) {
3291 skb_trim(syn_data, copied);
3295 /* No more data pending in inet_wait_for_connect() */
3296 if (space == fo->size)
3300 tcp_connect_queue_skb(sk, syn_data);
3302 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3304 syn->skb_mstamp = syn_data->skb_mstamp;
3306 /* Now full SYN+DATA was cloned and sent (or not),
3307 * remove the SYN from the original skb (syn_data)
3308 * we keep in write queue in case of a retransmit, as we
3309 * also have the SYN packet (with no data) in the same queue.
3311 TCP_SKB_CB(syn_data)->seq++;
3312 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3314 tp->syn_data = (fo->copied > 0);
3315 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3320 /* Send a regular SYN with Fast Open cookie request option */
3321 if (fo->cookie.len > 0)
3323 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3325 tp->syn_fastopen = 0;
3327 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3331 /* Build a SYN and send it off. */
3332 int tcp_connect(struct sock *sk)
3334 struct tcp_sock *tp = tcp_sk(sk);
3335 struct sk_buff *buff;
3338 tcp_connect_init(sk);
3340 if (unlikely(tp->repair)) {
3341 tcp_finish_connect(sk, NULL);
3345 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3346 if (unlikely(!buff))
3349 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3350 tp->retrans_stamp = tcp_time_stamp;
3351 tcp_connect_queue_skb(sk, buff);
3352 tcp_ecn_send_syn(sk, buff);
3354 /* Send off SYN; include data in Fast Open. */
3355 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3356 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3357 if (err == -ECONNREFUSED)
3360 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3361 * in order to make this packet get counted in tcpOutSegs.
3363 tp->snd_nxt = tp->write_seq;
3364 tp->pushed_seq = tp->write_seq;
3365 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3367 /* Timer for repeating the SYN until an answer. */
3368 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3369 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3372 EXPORT_SYMBOL(tcp_connect);
3374 /* Send out a delayed ack, the caller does the policy checking
3375 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3378 void tcp_send_delayed_ack(struct sock *sk)
3380 struct inet_connection_sock *icsk = inet_csk(sk);
3381 int ato = icsk->icsk_ack.ato;
3382 unsigned long timeout;
3384 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3386 if (ato > TCP_DELACK_MIN) {
3387 const struct tcp_sock *tp = tcp_sk(sk);
3388 int max_ato = HZ / 2;
3390 if (icsk->icsk_ack.pingpong ||
3391 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3392 max_ato = TCP_DELACK_MAX;
3394 /* Slow path, intersegment interval is "high". */
3396 /* If some rtt estimate is known, use it to bound delayed ack.
3397 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3401 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3408 ato = min(ato, max_ato);
3411 /* Stay within the limit we were given */
3412 timeout = jiffies + ato;
3414 /* Use new timeout only if there wasn't a older one earlier. */
3415 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3416 /* If delack timer was blocked or is about to expire,
3419 if (icsk->icsk_ack.blocked ||
3420 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3425 if (!time_before(timeout, icsk->icsk_ack.timeout))
3426 timeout = icsk->icsk_ack.timeout;
3428 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3429 icsk->icsk_ack.timeout = timeout;
3430 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3433 /* This routine sends an ack and also updates the window. */
3434 void tcp_send_ack(struct sock *sk)
3436 struct sk_buff *buff;
3438 /* If we have been reset, we may not send again. */
3439 if (sk->sk_state == TCP_CLOSE)
3442 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3444 /* We are not putting this on the write queue, so
3445 * tcp_transmit_skb() will set the ownership to this
3448 buff = alloc_skb(MAX_TCP_HEADER,
3449 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3450 if (unlikely(!buff)) {
3451 inet_csk_schedule_ack(sk);
3452 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3453 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3454 TCP_DELACK_MAX, TCP_RTO_MAX);
3458 /* Reserve space for headers and prepare control bits. */
3459 skb_reserve(buff, MAX_TCP_HEADER);
3460 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3462 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3464 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3465 * We also avoid tcp_wfree() overhead (cache line miss accessing
3466 * tp->tsq_flags) by using regular sock_wfree()
3468 skb_set_tcp_pure_ack(buff);
3470 /* Send it off, this clears delayed acks for us. */
3471 skb_mstamp_get(&buff->skb_mstamp);
3472 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3474 EXPORT_SYMBOL_GPL(tcp_send_ack);
3476 /* This routine sends a packet with an out of date sequence
3477 * number. It assumes the other end will try to ack it.
3479 * Question: what should we make while urgent mode?
3480 * 4.4BSD forces sending single byte of data. We cannot send
3481 * out of window data, because we have SND.NXT==SND.MAX...
3483 * Current solution: to send TWO zero-length segments in urgent mode:
3484 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3485 * out-of-date with SND.UNA-1 to probe window.
3487 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3489 struct tcp_sock *tp = tcp_sk(sk);
3490 struct sk_buff *skb;
3492 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3493 skb = alloc_skb(MAX_TCP_HEADER,
3494 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3498 /* Reserve space for headers and set control bits. */
3499 skb_reserve(skb, MAX_TCP_HEADER);
3500 /* Use a previous sequence. This should cause the other
3501 * end to send an ack. Don't queue or clone SKB, just
3504 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3505 skb_mstamp_get(&skb->skb_mstamp);
3506 NET_INC_STATS(sock_net(sk), mib);
3507 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3510 void tcp_send_window_probe(struct sock *sk)
3512 if (sk->sk_state == TCP_ESTABLISHED) {
3513 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3514 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3518 /* Initiate keepalive or window probe from timer. */
3519 int tcp_write_wakeup(struct sock *sk, int mib)
3521 struct tcp_sock *tp = tcp_sk(sk);
3522 struct sk_buff *skb;
3524 if (sk->sk_state == TCP_CLOSE)
3527 skb = tcp_send_head(sk);
3528 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3530 unsigned int mss = tcp_current_mss(sk);
3531 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3533 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3534 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3536 /* We are probing the opening of a window
3537 * but the window size is != 0
3538 * must have been a result SWS avoidance ( sender )
3540 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3542 seg_size = min(seg_size, mss);
3543 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3544 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3546 } else if (!tcp_skb_pcount(skb))
3547 tcp_set_skb_tso_segs(skb, mss);
3549 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3550 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3552 tcp_event_new_data_sent(sk, skb);
3555 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3556 tcp_xmit_probe_skb(sk, 1, mib);
3557 return tcp_xmit_probe_skb(sk, 0, mib);
3561 /* A window probe timeout has occurred. If window is not closed send
3562 * a partial packet else a zero probe.
3564 void tcp_send_probe0(struct sock *sk)
3566 struct inet_connection_sock *icsk = inet_csk(sk);
3567 struct tcp_sock *tp = tcp_sk(sk);
3568 struct net *net = sock_net(sk);
3569 unsigned long probe_max;
3572 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3574 if (tp->packets_out || !tcp_send_head(sk)) {
3575 /* Cancel probe timer, if it is not required. */
3576 icsk->icsk_probes_out = 0;
3577 icsk->icsk_backoff = 0;
3582 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3583 icsk->icsk_backoff++;
3584 icsk->icsk_probes_out++;
3585 probe_max = TCP_RTO_MAX;
3587 /* If packet was not sent due to local congestion,
3588 * do not backoff and do not remember icsk_probes_out.
3589 * Let local senders to fight for local resources.
3591 * Use accumulated backoff yet.
3593 if (!icsk->icsk_probes_out)
3594 icsk->icsk_probes_out = 1;
3595 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3597 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3598 tcp_probe0_when(sk, probe_max),
3602 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3604 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3608 tcp_rsk(req)->txhash = net_tx_rndhash();
3609 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3611 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3612 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3613 if (unlikely(tcp_passive_fastopen(sk)))
3614 tcp_sk(sk)->total_retrans++;
3618 EXPORT_SYMBOL(tcp_rtx_synack);