2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of two TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
66 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
68 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
69 int push_one, gfp_t gfp);
71 /* Account for new data that has been sent to the network. */
72 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
74 struct inet_connection_sock *icsk = inet_csk(sk);
75 struct tcp_sock *tp = tcp_sk(sk);
76 unsigned int prior_packets = tp->packets_out;
78 tcp_advance_send_head(sk, skb);
79 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
81 tp->packets_out += tcp_skb_pcount(skb);
82 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
83 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
87 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
99 const struct tcp_sock *tp = tcp_sk(sk);
101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
143 struct tcp_sock *tp = tcp_sk(sk);
144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
166 const struct dst_entry *dst = __sk_dst_get(sk);
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 icsk->icsk_ack.pingpong = 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 u32 tcp_default_init_rwnd(u32 mss)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd = TCP_INIT_CWND * 2;
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
213 int wscale_ok, __u8 *rcv_wscale,
216 unsigned int space = (__space < 0 ? 0 : __space);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (65535 << 14);
221 space = min(*window_clamp, space);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space = (space / mss) * mss;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 while (space > 65535 && (*rcv_wscale) < 14) {
253 if (mss > (1 << *rcv_wscale)) {
254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
262 EXPORT_SYMBOL(tcp_select_initial_window);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16 tcp_select_window(struct sock *sk)
271 struct tcp_sock *tp = tcp_sk(sk);
272 u32 old_win = tp->rcv_wnd;
273 u32 cur_win = tcp_receive_window(tp);
274 u32 new_win = __tcp_select_window(sk);
276 /* Never shrink the offered window */
277 if (new_win < cur_win) {
278 /* Danger Will Robinson!
279 * Don't update rcv_wup/rcv_wnd here or else
280 * we will not be able to advertise a zero
281 * window in time. --DaveM
283 * Relax Will Robinson.
286 NET_INC_STATS(sock_net(sk),
287 LINUX_MIB_TCPWANTZEROWINDOWADV);
288 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
290 tp->rcv_wnd = new_win;
291 tp->rcv_wup = tp->rcv_nxt;
293 /* Make sure we do not exceed the maximum possible
296 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
297 new_win = min(new_win, MAX_TCP_WINDOW);
299 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
301 /* RFC1323 scaling applied */
302 new_win >>= tp->rx_opt.rcv_wscale;
304 /* If we advertise zero window, disable fast path. */
308 NET_INC_STATS(sock_net(sk),
309 LINUX_MIB_TCPTOZEROWINDOWADV);
310 } else if (old_win == 0) {
311 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
317 /* Packet ECN state for a SYN-ACK */
318 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
320 const struct tcp_sock *tp = tcp_sk(sk);
322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
323 if (!(tp->ecn_flags & TCP_ECN_OK))
324 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
325 else if (tcp_ca_needs_ecn(sk))
329 /* Packet ECN state for a SYN. */
330 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
332 struct tcp_sock *tp = tcp_sk(sk);
333 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
334 tcp_ca_needs_ecn(sk);
337 const struct dst_entry *dst = __sk_dst_get(sk);
339 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
346 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
347 tp->ecn_flags = TCP_ECN_OK;
348 if (tcp_ca_needs_ecn(sk))
354 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
357 if (inet_rsk(req)->ecn_ok) {
359 if (tcp_ca_needs_ecn(sk))
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,
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;
379 tcp_hdr(skb)->cwr = 1;
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)
387 tcp_hdr(skb)->ece = 1;
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 struct skb_shared_info *shinfo = skb_shinfo(skb);
398 skb->ip_summed = CHECKSUM_PARTIAL;
401 TCP_SKB_CB(skb)->tcp_flags = flags;
402 TCP_SKB_CB(skb)->sacked = 0;
404 tcp_skb_pcount_set(skb, 1);
405 shinfo->gso_size = 0;
406 shinfo->gso_type = 0;
408 TCP_SKB_CB(skb)->seq = seq;
409 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
411 TCP_SKB_CB(skb)->end_seq = seq;
414 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
416 return tp->snd_una != tp->snd_up;
419 #define OPTION_SACK_ADVERTISE (1 << 0)
420 #define OPTION_TS (1 << 1)
421 #define OPTION_MD5 (1 << 2)
422 #define OPTION_WSCALE (1 << 3)
423 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
425 struct tcp_out_options {
426 u16 options; /* bit field of OPTION_* */
427 u16 mss; /* 0 to disable */
428 u8 ws; /* window scale, 0 to disable */
429 u8 num_sack_blocks; /* number of SACK blocks to include */
430 u8 hash_size; /* bytes in hash_location */
431 __u8 *hash_location; /* temporary pointer, overloaded */
432 __u32 tsval, tsecr; /* need to include OPTION_TS */
433 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
436 /* Write previously computed TCP options to the packet.
438 * Beware: Something in the Internet is very sensitive to the ordering of
439 * TCP options, we learned this through the hard way, so be careful here.
440 * Luckily we can at least blame others for their non-compliance but from
441 * inter-operability perspective it seems that we're somewhat stuck with
442 * the ordering which we have been using if we want to keep working with
443 * those broken things (not that it currently hurts anybody as there isn't
444 * particular reason why the ordering would need to be changed).
446 * At least SACK_PERM as the first option is known to lead to a disaster
447 * (but it may well be that other scenarios fail similarly).
449 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
450 struct tcp_out_options *opts)
452 u16 options = opts->options; /* mungable copy */
454 if (unlikely(OPTION_MD5 & options)) {
455 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
456 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
457 /* overload cookie hash location */
458 opts->hash_location = (__u8 *)ptr;
462 if (unlikely(opts->mss)) {
463 *ptr++ = htonl((TCPOPT_MSS << 24) |
464 (TCPOLEN_MSS << 16) |
468 if (likely(OPTION_TS & options)) {
469 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
470 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
471 (TCPOLEN_SACK_PERM << 16) |
472 (TCPOPT_TIMESTAMP << 8) |
474 options &= ~OPTION_SACK_ADVERTISE;
476 *ptr++ = htonl((TCPOPT_NOP << 24) |
478 (TCPOPT_TIMESTAMP << 8) |
481 *ptr++ = htonl(opts->tsval);
482 *ptr++ = htonl(opts->tsecr);
485 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
486 *ptr++ = htonl((TCPOPT_NOP << 24) |
488 (TCPOPT_SACK_PERM << 8) |
492 if (unlikely(OPTION_WSCALE & options)) {
493 *ptr++ = htonl((TCPOPT_NOP << 24) |
494 (TCPOPT_WINDOW << 16) |
495 (TCPOLEN_WINDOW << 8) |
499 if (unlikely(opts->num_sack_blocks)) {
500 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
501 tp->duplicate_sack : tp->selective_acks;
504 *ptr++ = htonl((TCPOPT_NOP << 24) |
507 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
508 TCPOLEN_SACK_PERBLOCK)));
510 for (this_sack = 0; this_sack < opts->num_sack_blocks;
512 *ptr++ = htonl(sp[this_sack].start_seq);
513 *ptr++ = htonl(sp[this_sack].end_seq);
516 tp->rx_opt.dsack = 0;
519 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
520 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
522 *ptr++ = htonl((TCPOPT_EXP << 24) |
523 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
524 TCPOPT_FASTOPEN_MAGIC);
526 memcpy(ptr, foc->val, foc->len);
527 if ((foc->len & 3) == 2) {
528 u8 *align = ((u8 *)ptr) + foc->len;
529 align[0] = align[1] = TCPOPT_NOP;
531 ptr += (foc->len + 3) >> 2;
535 /* Compute TCP options for SYN packets. This is not the final
536 * network wire format yet.
538 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
539 struct tcp_out_options *opts,
540 struct tcp_md5sig_key **md5)
542 struct tcp_sock *tp = tcp_sk(sk);
543 unsigned int remaining = MAX_TCP_OPTION_SPACE;
544 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
546 #ifdef CONFIG_TCP_MD5SIG
547 *md5 = tp->af_specific->md5_lookup(sk, sk);
549 opts->options |= OPTION_MD5;
550 remaining -= TCPOLEN_MD5SIG_ALIGNED;
556 /* We always get an MSS option. The option bytes which will be seen in
557 * normal data packets should timestamps be used, must be in the MSS
558 * advertised. But we subtract them from tp->mss_cache so that
559 * calculations in tcp_sendmsg are simpler etc. So account for this
560 * fact here if necessary. If we don't do this correctly, as a
561 * receiver we won't recognize data packets as being full sized when we
562 * should, and thus we won't abide by the delayed ACK rules correctly.
563 * SACKs don't matter, we never delay an ACK when we have any of those
565 opts->mss = tcp_advertise_mss(sk);
566 remaining -= TCPOLEN_MSS_ALIGNED;
568 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
569 opts->options |= OPTION_TS;
570 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
571 opts->tsecr = tp->rx_opt.ts_recent;
572 remaining -= TCPOLEN_TSTAMP_ALIGNED;
574 if (likely(sysctl_tcp_window_scaling)) {
575 opts->ws = tp->rx_opt.rcv_wscale;
576 opts->options |= OPTION_WSCALE;
577 remaining -= TCPOLEN_WSCALE_ALIGNED;
579 if (likely(sysctl_tcp_sack)) {
580 opts->options |= OPTION_SACK_ADVERTISE;
581 if (unlikely(!(OPTION_TS & opts->options)))
582 remaining -= TCPOLEN_SACKPERM_ALIGNED;
585 if (fastopen && fastopen->cookie.len >= 0) {
586 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
587 need = (need + 3) & ~3U; /* Align to 32 bits */
588 if (remaining >= need) {
589 opts->options |= OPTION_FAST_OPEN_COOKIE;
590 opts->fastopen_cookie = &fastopen->cookie;
592 tp->syn_fastopen = 1;
596 return MAX_TCP_OPTION_SPACE - remaining;
599 /* Set up TCP options for SYN-ACKs. */
600 static unsigned int tcp_synack_options(struct sock *sk,
601 struct request_sock *req,
602 unsigned int mss, struct sk_buff *skb,
603 struct tcp_out_options *opts,
604 struct tcp_md5sig_key **md5,
605 struct tcp_fastopen_cookie *foc)
607 struct inet_request_sock *ireq = inet_rsk(req);
608 unsigned int remaining = MAX_TCP_OPTION_SPACE;
610 #ifdef CONFIG_TCP_MD5SIG
611 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
613 opts->options |= OPTION_MD5;
614 remaining -= TCPOLEN_MD5SIG_ALIGNED;
616 /* We can't fit any SACK blocks in a packet with MD5 + TS
617 * options. There was discussion about disabling SACK
618 * rather than TS in order to fit in better with old,
619 * buggy kernels, but that was deemed to be unnecessary.
621 ireq->tstamp_ok &= !ireq->sack_ok;
627 /* We always send an MSS option. */
629 remaining -= TCPOLEN_MSS_ALIGNED;
631 if (likely(ireq->wscale_ok)) {
632 opts->ws = ireq->rcv_wscale;
633 opts->options |= OPTION_WSCALE;
634 remaining -= TCPOLEN_WSCALE_ALIGNED;
636 if (likely(ireq->tstamp_ok)) {
637 opts->options |= OPTION_TS;
638 opts->tsval = tcp_skb_timestamp(skb);
639 opts->tsecr = req->ts_recent;
640 remaining -= TCPOLEN_TSTAMP_ALIGNED;
642 if (likely(ireq->sack_ok)) {
643 opts->options |= OPTION_SACK_ADVERTISE;
644 if (unlikely(!ireq->tstamp_ok))
645 remaining -= TCPOLEN_SACKPERM_ALIGNED;
647 if (foc != NULL && foc->len >= 0) {
648 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
649 need = (need + 3) & ~3U; /* Align to 32 bits */
650 if (remaining >= need) {
651 opts->options |= OPTION_FAST_OPEN_COOKIE;
652 opts->fastopen_cookie = foc;
657 return MAX_TCP_OPTION_SPACE - remaining;
660 /* Compute TCP options for ESTABLISHED sockets. This is not the
661 * final wire format yet.
663 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
664 struct tcp_out_options *opts,
665 struct tcp_md5sig_key **md5)
667 struct tcp_sock *tp = tcp_sk(sk);
668 unsigned int size = 0;
669 unsigned int eff_sacks;
673 #ifdef CONFIG_TCP_MD5SIG
674 *md5 = tp->af_specific->md5_lookup(sk, sk);
675 if (unlikely(*md5)) {
676 opts->options |= OPTION_MD5;
677 size += TCPOLEN_MD5SIG_ALIGNED;
683 if (likely(tp->rx_opt.tstamp_ok)) {
684 opts->options |= OPTION_TS;
685 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
686 opts->tsecr = tp->rx_opt.ts_recent;
687 size += TCPOLEN_TSTAMP_ALIGNED;
690 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
691 if (unlikely(eff_sacks)) {
692 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
693 opts->num_sack_blocks =
694 min_t(unsigned int, eff_sacks,
695 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
696 TCPOLEN_SACK_PERBLOCK);
697 size += TCPOLEN_SACK_BASE_ALIGNED +
698 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
705 /* TCP SMALL QUEUES (TSQ)
707 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
708 * to reduce RTT and bufferbloat.
709 * We do this using a special skb destructor (tcp_wfree).
711 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
712 * needs to be reallocated in a driver.
713 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
715 * Since transmit from skb destructor is forbidden, we use a tasklet
716 * to process all sockets that eventually need to send more skbs.
717 * We use one tasklet per cpu, with its own queue of sockets.
720 struct tasklet_struct tasklet;
721 struct list_head head; /* queue of tcp sockets */
723 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
725 static void tcp_tsq_handler(struct sock *sk)
727 if ((1 << sk->sk_state) &
728 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
729 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
730 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
734 * One tasklet per cpu tries to send more skbs.
735 * We run in tasklet context but need to disable irqs when
736 * transferring tsq->head because tcp_wfree() might
737 * interrupt us (non NAPI drivers)
739 static void tcp_tasklet_func(unsigned long data)
741 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
744 struct list_head *q, *n;
748 local_irq_save(flags);
749 list_splice_init(&tsq->head, &list);
750 local_irq_restore(flags);
752 list_for_each_safe(q, n, &list) {
753 tp = list_entry(q, struct tcp_sock, tsq_node);
754 list_del(&tp->tsq_node);
756 sk = (struct sock *)tp;
759 if (!sock_owned_by_user(sk)) {
762 /* defer the work to tcp_release_cb() */
763 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
767 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
772 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
773 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
774 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
775 (1UL << TCP_MTU_REDUCED_DEFERRED))
777 * tcp_release_cb - tcp release_sock() callback
780 * called from release_sock() to perform protocol dependent
781 * actions before socket release.
783 void tcp_release_cb(struct sock *sk)
785 struct tcp_sock *tp = tcp_sk(sk);
786 unsigned long flags, nflags;
788 /* perform an atomic operation only if at least one flag is set */
790 flags = tp->tsq_flags;
791 if (!(flags & TCP_DEFERRED_ALL))
793 nflags = flags & ~TCP_DEFERRED_ALL;
794 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
796 if (flags & (1UL << TCP_TSQ_DEFERRED))
799 /* Here begins the tricky part :
800 * We are called from release_sock() with :
802 * 2) sk_lock.slock spinlock held
803 * 3) socket owned by us (sk->sk_lock.owned == 1)
805 * But following code is meant to be called from BH handlers,
806 * so we should keep BH disabled, but early release socket ownership
808 sock_release_ownership(sk);
810 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
811 tcp_write_timer_handler(sk);
814 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
815 tcp_delack_timer_handler(sk);
818 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
819 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
823 EXPORT_SYMBOL(tcp_release_cb);
825 void __init tcp_tasklet_init(void)
829 for_each_possible_cpu(i) {
830 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
832 INIT_LIST_HEAD(&tsq->head);
833 tasklet_init(&tsq->tasklet,
840 * Write buffer destructor automatically called from kfree_skb.
841 * We can't xmit new skbs from this context, as we might already
844 void tcp_wfree(struct sk_buff *skb)
846 struct sock *sk = skb->sk;
847 struct tcp_sock *tp = tcp_sk(sk);
850 /* Keep one reference on sk_wmem_alloc.
851 * Will be released by sk_free() from here or tcp_tasklet_func()
853 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
855 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
856 * Wait until our queues (qdisc + devices) are drained.
858 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
859 * - chance for incoming ACK (processed by another cpu maybe)
860 * to migrate this flow (skb->ooo_okay will be eventually set)
862 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
865 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
866 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
868 struct tsq_tasklet *tsq;
870 /* queue this socket to tasklet queue */
871 local_irq_save(flags);
872 tsq = this_cpu_ptr(&tsq_tasklet);
873 list_add(&tp->tsq_node, &tsq->head);
874 tasklet_schedule(&tsq->tasklet);
875 local_irq_restore(flags);
882 /* This routine actually transmits TCP packets queued in by
883 * tcp_do_sendmsg(). This is used by both the initial
884 * transmission and possible later retransmissions.
885 * All SKB's seen here are completely headerless. It is our
886 * job to build the TCP header, and pass the packet down to
887 * IP so it can do the same plus pass the packet off to the
890 * We are working here with either a clone of the original
891 * SKB, or a fresh unique copy made by the retransmit engine.
893 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
896 const struct inet_connection_sock *icsk = inet_csk(sk);
897 struct inet_sock *inet;
899 struct tcp_skb_cb *tcb;
900 struct tcp_out_options opts;
901 unsigned int tcp_options_size, tcp_header_size;
902 struct tcp_md5sig_key *md5;
906 BUG_ON(!skb || !tcp_skb_pcount(skb));
909 skb_mstamp_get(&skb->skb_mstamp);
911 if (unlikely(skb_cloned(skb)))
912 skb = pskb_copy(skb, gfp_mask);
914 skb = skb_clone(skb, gfp_mask);
921 tcb = TCP_SKB_CB(skb);
922 memset(&opts, 0, sizeof(opts));
924 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
925 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
927 tcp_options_size = tcp_established_options(sk, skb, &opts,
929 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
931 if (tcp_packets_in_flight(tp) == 0)
932 tcp_ca_event(sk, CA_EVENT_TX_START);
934 /* if no packet is in qdisc/device queue, then allow XPS to select
935 * another queue. We can be called from tcp_tsq_handler()
936 * which holds one reference to sk_wmem_alloc.
938 * TODO: Ideally, in-flight pure ACK packets should not matter here.
939 * One way to get this would be to set skb->truesize = 2 on them.
941 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
943 skb_push(skb, tcp_header_size);
944 skb_reset_transport_header(skb);
948 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
949 skb_set_hash_from_sk(skb, sk);
950 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
952 /* Build TCP header and checksum it. */
954 th->source = inet->inet_sport;
955 th->dest = inet->inet_dport;
956 th->seq = htonl(tcb->seq);
957 th->ack_seq = htonl(tp->rcv_nxt);
958 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
961 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
962 /* RFC1323: The window in SYN & SYN/ACK segments
965 th->window = htons(min(tp->rcv_wnd, 65535U));
967 th->window = htons(tcp_select_window(sk));
972 /* The urg_mode check is necessary during a below snd_una win probe */
973 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
974 if (before(tp->snd_up, tcb->seq + 0x10000)) {
975 th->urg_ptr = htons(tp->snd_up - tcb->seq);
977 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
978 th->urg_ptr = htons(0xFFFF);
983 tcp_options_write((__be32 *)(th + 1), tp, &opts);
984 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
985 tcp_ecn_send(sk, skb, tcp_header_size);
987 #ifdef CONFIG_TCP_MD5SIG
988 /* Calculate the MD5 hash, as we have all we need now */
990 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
991 tp->af_specific->calc_md5_hash(opts.hash_location,
996 icsk->icsk_af_ops->send_check(sk, skb);
998 if (likely(tcb->tcp_flags & TCPHDR_ACK))
999 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1001 if (skb->len != tcp_header_size)
1002 tcp_event_data_sent(tp, sk);
1004 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1005 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1006 tcp_skb_pcount(skb));
1008 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
1009 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1011 /* Our usage of tstamp should remain private */
1012 skb->tstamp.tv64 = 0;
1014 /* Cleanup our debris for IP stacks */
1015 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1016 sizeof(struct inet6_skb_parm)));
1018 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1020 if (likely(err <= 0))
1025 return net_xmit_eval(err);
1028 /* This routine just queues the buffer for sending.
1030 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1031 * otherwise socket can stall.
1033 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1035 struct tcp_sock *tp = tcp_sk(sk);
1037 /* Advance write_seq and place onto the write_queue. */
1038 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1039 __skb_header_release(skb);
1040 tcp_add_write_queue_tail(sk, skb);
1041 sk->sk_wmem_queued += skb->truesize;
1042 sk_mem_charge(sk, skb->truesize);
1045 /* Initialize TSO segments for a packet. */
1046 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1047 unsigned int mss_now)
1049 struct skb_shared_info *shinfo = skb_shinfo(skb);
1051 /* Make sure we own this skb before messing gso_size/gso_segs */
1052 WARN_ON_ONCE(skb_cloned(skb));
1054 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1055 /* Avoid the costly divide in the normal
1058 tcp_skb_pcount_set(skb, 1);
1059 shinfo->gso_size = 0;
1060 shinfo->gso_type = 0;
1062 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1063 shinfo->gso_size = mss_now;
1064 shinfo->gso_type = sk->sk_gso_type;
1068 /* When a modification to fackets out becomes necessary, we need to check
1069 * skb is counted to fackets_out or not.
1071 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1074 struct tcp_sock *tp = tcp_sk(sk);
1076 if (!tp->sacked_out || tcp_is_reno(tp))
1079 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1080 tp->fackets_out -= decr;
1083 /* Pcount in the middle of the write queue got changed, we need to do various
1084 * tweaks to fix counters
1086 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1088 struct tcp_sock *tp = tcp_sk(sk);
1090 tp->packets_out -= decr;
1092 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1093 tp->sacked_out -= decr;
1094 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1095 tp->retrans_out -= decr;
1096 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1097 tp->lost_out -= decr;
1099 /* Reno case is special. Sigh... */
1100 if (tcp_is_reno(tp) && decr > 0)
1101 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1103 tcp_adjust_fackets_out(sk, skb, decr);
1105 if (tp->lost_skb_hint &&
1106 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1107 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1108 tp->lost_cnt_hint -= decr;
1110 tcp_verify_left_out(tp);
1113 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1115 struct skb_shared_info *shinfo = skb_shinfo(skb);
1117 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1118 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1119 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1120 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1122 shinfo->tx_flags &= ~tsflags;
1123 shinfo2->tx_flags |= tsflags;
1124 swap(shinfo->tskey, shinfo2->tskey);
1128 /* Function to create two new TCP segments. Shrinks the given segment
1129 * to the specified size and appends a new segment with the rest of the
1130 * packet to the list. This won't be called frequently, I hope.
1131 * Remember, these are still headerless SKBs at this point.
1133 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1134 unsigned int mss_now, gfp_t gfp)
1136 struct tcp_sock *tp = tcp_sk(sk);
1137 struct sk_buff *buff;
1138 int nsize, old_factor;
1142 if (WARN_ON(len > skb->len))
1145 nsize = skb_headlen(skb) - len;
1149 if (skb_unclone(skb, gfp))
1152 /* Get a new skb... force flag on. */
1153 buff = sk_stream_alloc_skb(sk, nsize, gfp);
1155 return -ENOMEM; /* We'll just try again later. */
1157 sk->sk_wmem_queued += buff->truesize;
1158 sk_mem_charge(sk, buff->truesize);
1159 nlen = skb->len - len - nsize;
1160 buff->truesize += nlen;
1161 skb->truesize -= nlen;
1163 /* Correct the sequence numbers. */
1164 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1165 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1166 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1168 /* PSH and FIN should only be set in the second packet. */
1169 flags = TCP_SKB_CB(skb)->tcp_flags;
1170 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1171 TCP_SKB_CB(buff)->tcp_flags = flags;
1172 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1174 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1175 /* Copy and checksum data tail into the new buffer. */
1176 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1177 skb_put(buff, nsize),
1182 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1184 skb->ip_summed = CHECKSUM_PARTIAL;
1185 skb_split(skb, buff, len);
1188 buff->ip_summed = skb->ip_summed;
1190 buff->tstamp = skb->tstamp;
1191 tcp_fragment_tstamp(skb, buff);
1193 old_factor = tcp_skb_pcount(skb);
1195 /* Fix up tso_factor for both original and new SKB. */
1196 tcp_set_skb_tso_segs(sk, skb, mss_now);
1197 tcp_set_skb_tso_segs(sk, buff, mss_now);
1199 /* If this packet has been sent out already, we must
1200 * adjust the various packet counters.
1202 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1203 int diff = old_factor - tcp_skb_pcount(skb) -
1204 tcp_skb_pcount(buff);
1207 tcp_adjust_pcount(sk, skb, diff);
1210 /* Link BUFF into the send queue. */
1211 __skb_header_release(buff);
1212 tcp_insert_write_queue_after(skb, buff, sk);
1217 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1218 * eventually). The difference is that pulled data not copied, but
1219 * immediately discarded.
1221 static void __pskb_trim_head(struct sk_buff *skb, int len)
1223 struct skb_shared_info *shinfo;
1226 eat = min_t(int, len, skb_headlen(skb));
1228 __skb_pull(skb, eat);
1235 shinfo = skb_shinfo(skb);
1236 for (i = 0; i < shinfo->nr_frags; i++) {
1237 int size = skb_frag_size(&shinfo->frags[i]);
1240 skb_frag_unref(skb, i);
1243 shinfo->frags[k] = shinfo->frags[i];
1245 shinfo->frags[k].page_offset += eat;
1246 skb_frag_size_sub(&shinfo->frags[k], eat);
1252 shinfo->nr_frags = k;
1254 skb_reset_tail_pointer(skb);
1255 skb->data_len -= len;
1256 skb->len = skb->data_len;
1259 /* Remove acked data from a packet in the transmit queue. */
1260 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1262 if (skb_unclone(skb, GFP_ATOMIC))
1265 __pskb_trim_head(skb, len);
1267 TCP_SKB_CB(skb)->seq += len;
1268 skb->ip_summed = CHECKSUM_PARTIAL;
1270 skb->truesize -= len;
1271 sk->sk_wmem_queued -= len;
1272 sk_mem_uncharge(sk, len);
1273 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1275 /* Any change of skb->len requires recalculation of tso factor. */
1276 if (tcp_skb_pcount(skb) > 1)
1277 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1282 /* Calculate MSS not accounting any TCP options. */
1283 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1285 const struct tcp_sock *tp = tcp_sk(sk);
1286 const struct inet_connection_sock *icsk = inet_csk(sk);
1289 /* Calculate base mss without TCP options:
1290 It is MMS_S - sizeof(tcphdr) of rfc1122
1292 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1294 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1295 if (icsk->icsk_af_ops->net_frag_header_len) {
1296 const struct dst_entry *dst = __sk_dst_get(sk);
1298 if (dst && dst_allfrag(dst))
1299 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1302 /* Clamp it (mss_clamp does not include tcp options) */
1303 if (mss_now > tp->rx_opt.mss_clamp)
1304 mss_now = tp->rx_opt.mss_clamp;
1306 /* Now subtract optional transport overhead */
1307 mss_now -= icsk->icsk_ext_hdr_len;
1309 /* Then reserve room for full set of TCP options and 8 bytes of data */
1315 /* Calculate MSS. Not accounting for SACKs here. */
1316 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1318 /* Subtract TCP options size, not including SACKs */
1319 return __tcp_mtu_to_mss(sk, pmtu) -
1320 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1323 /* Inverse of above */
1324 int tcp_mss_to_mtu(struct sock *sk, int mss)
1326 const struct tcp_sock *tp = tcp_sk(sk);
1327 const struct inet_connection_sock *icsk = inet_csk(sk);
1331 tp->tcp_header_len +
1332 icsk->icsk_ext_hdr_len +
1333 icsk->icsk_af_ops->net_header_len;
1335 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1336 if (icsk->icsk_af_ops->net_frag_header_len) {
1337 const struct dst_entry *dst = __sk_dst_get(sk);
1339 if (dst && dst_allfrag(dst))
1340 mtu += icsk->icsk_af_ops->net_frag_header_len;
1345 /* MTU probing init per socket */
1346 void tcp_mtup_init(struct sock *sk)
1348 struct tcp_sock *tp = tcp_sk(sk);
1349 struct inet_connection_sock *icsk = inet_csk(sk);
1350 struct net *net = sock_net(sk);
1352 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1353 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1354 icsk->icsk_af_ops->net_header_len;
1355 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1356 icsk->icsk_mtup.probe_size = 0;
1357 if (icsk->icsk_mtup.enabled)
1358 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1360 EXPORT_SYMBOL(tcp_mtup_init);
1362 /* This function synchronize snd mss to current pmtu/exthdr set.
1364 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1365 for TCP options, but includes only bare TCP header.
1367 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1368 It is minimum of user_mss and mss received with SYN.
1369 It also does not include TCP options.
1371 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1373 tp->mss_cache is current effective sending mss, including
1374 all tcp options except for SACKs. It is evaluated,
1375 taking into account current pmtu, but never exceeds
1376 tp->rx_opt.mss_clamp.
1378 NOTE1. rfc1122 clearly states that advertised MSS
1379 DOES NOT include either tcp or ip options.
1381 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1382 are READ ONLY outside this function. --ANK (980731)
1384 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386 struct tcp_sock *tp = tcp_sk(sk);
1387 struct inet_connection_sock *icsk = inet_csk(sk);
1390 if (icsk->icsk_mtup.search_high > pmtu)
1391 icsk->icsk_mtup.search_high = pmtu;
1393 mss_now = tcp_mtu_to_mss(sk, pmtu);
1394 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1396 /* And store cached results */
1397 icsk->icsk_pmtu_cookie = pmtu;
1398 if (icsk->icsk_mtup.enabled)
1399 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1400 tp->mss_cache = mss_now;
1404 EXPORT_SYMBOL(tcp_sync_mss);
1406 /* Compute the current effective MSS, taking SACKs and IP options,
1407 * and even PMTU discovery events into account.
1409 unsigned int tcp_current_mss(struct sock *sk)
1411 const struct tcp_sock *tp = tcp_sk(sk);
1412 const struct dst_entry *dst = __sk_dst_get(sk);
1414 unsigned int header_len;
1415 struct tcp_out_options opts;
1416 struct tcp_md5sig_key *md5;
1418 mss_now = tp->mss_cache;
1421 u32 mtu = dst_mtu(dst);
1422 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1423 mss_now = tcp_sync_mss(sk, mtu);
1426 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1427 sizeof(struct tcphdr);
1428 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1429 * some common options. If this is an odd packet (because we have SACK
1430 * blocks etc) then our calculated header_len will be different, and
1431 * we have to adjust mss_now correspondingly */
1432 if (header_len != tp->tcp_header_len) {
1433 int delta = (int) header_len - tp->tcp_header_len;
1440 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1441 * As additional protections, we do not touch cwnd in retransmission phases,
1442 * and if application hit its sndbuf limit recently.
1444 static void tcp_cwnd_application_limited(struct sock *sk)
1446 struct tcp_sock *tp = tcp_sk(sk);
1448 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1449 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1450 /* Limited by application or receiver window. */
1451 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1452 u32 win_used = max(tp->snd_cwnd_used, init_win);
1453 if (win_used < tp->snd_cwnd) {
1454 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1455 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457 tp->snd_cwnd_used = 0;
1459 tp->snd_cwnd_stamp = tcp_time_stamp;
1462 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1464 struct tcp_sock *tp = tcp_sk(sk);
1466 /* Track the maximum number of outstanding packets in each
1467 * window, and remember whether we were cwnd-limited then.
1469 if (!before(tp->snd_una, tp->max_packets_seq) ||
1470 tp->packets_out > tp->max_packets_out) {
1471 tp->max_packets_out = tp->packets_out;
1472 tp->max_packets_seq = tp->snd_nxt;
1473 tp->is_cwnd_limited = is_cwnd_limited;
1476 if (tcp_is_cwnd_limited(sk)) {
1477 /* Network is feed fully. */
1478 tp->snd_cwnd_used = 0;
1479 tp->snd_cwnd_stamp = tcp_time_stamp;
1481 /* Network starves. */
1482 if (tp->packets_out > tp->snd_cwnd_used)
1483 tp->snd_cwnd_used = tp->packets_out;
1485 if (sysctl_tcp_slow_start_after_idle &&
1486 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1487 tcp_cwnd_application_limited(sk);
1491 /* Minshall's variant of the Nagle send check. */
1492 static bool tcp_minshall_check(const struct tcp_sock *tp)
1494 return after(tp->snd_sml, tp->snd_una) &&
1495 !after(tp->snd_sml, tp->snd_nxt);
1498 /* Update snd_sml if this skb is under mss
1499 * Note that a TSO packet might end with a sub-mss segment
1500 * The test is really :
1501 * if ((skb->len % mss) != 0)
1502 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1503 * But we can avoid doing the divide again given we already have
1504 * skb_pcount = skb->len / mss_now
1506 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1507 const struct sk_buff *skb)
1509 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1510 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1513 /* Return false, if packet can be sent now without violation Nagle's rules:
1514 * 1. It is full sized. (provided by caller in %partial bool)
1515 * 2. Or it contains FIN. (already checked by caller)
1516 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1517 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1518 * With Minshall's modification: all sent small packets are ACKed.
1520 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1524 ((nonagle & TCP_NAGLE_CORK) ||
1525 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1528 /* Return how many segs we'd like on a TSO packet,
1529 * to send one TSO packet per ms
1531 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1535 bytes = min(sk->sk_pacing_rate >> 10,
1536 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1538 /* Goal is to send at least one packet per ms,
1539 * not one big TSO packet every 100 ms.
1540 * This preserves ACK clocking and is consistent
1541 * with tcp_tso_should_defer() heuristic.
1543 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1545 return min_t(u32, segs, sk->sk_gso_max_segs);
1548 /* Returns the portion of skb which can be sent right away */
1549 static unsigned int tcp_mss_split_point(const struct sock *sk,
1550 const struct sk_buff *skb,
1551 unsigned int mss_now,
1552 unsigned int max_segs,
1555 const struct tcp_sock *tp = tcp_sk(sk);
1556 u32 partial, needed, window, max_len;
1558 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1559 max_len = mss_now * max_segs;
1561 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1564 needed = min(skb->len, window);
1566 if (max_len <= needed)
1569 partial = needed % mss_now;
1570 /* If last segment is not a full MSS, check if Nagle rules allow us
1571 * to include this last segment in this skb.
1572 * Otherwise, we'll split the skb at last MSS boundary
1574 if (tcp_nagle_check(partial != 0, tp, nonagle))
1575 return needed - partial;
1580 /* Can at least one segment of SKB be sent right now, according to the
1581 * congestion window rules? If so, return how many segments are allowed.
1583 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1584 const struct sk_buff *skb)
1586 u32 in_flight, cwnd, halfcwnd;
1588 /* Don't be strict about the congestion window for the final FIN. */
1589 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1590 tcp_skb_pcount(skb) == 1)
1593 in_flight = tcp_packets_in_flight(tp);
1594 cwnd = tp->snd_cwnd;
1595 if (in_flight >= cwnd)
1598 /* For better scheduling, ensure we have at least
1599 * 2 GSO packets in flight.
1601 halfcwnd = max(cwnd >> 1, 1U);
1602 return min(halfcwnd, cwnd - in_flight);
1605 /* Initialize TSO state of a skb.
1606 * This must be invoked the first time we consider transmitting
1607 * SKB onto the wire.
1609 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1610 unsigned int mss_now)
1612 int tso_segs = tcp_skb_pcount(skb);
1614 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1615 tcp_set_skb_tso_segs(sk, skb, mss_now);
1616 tso_segs = tcp_skb_pcount(skb);
1622 /* Return true if the Nagle test allows this packet to be
1625 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1626 unsigned int cur_mss, int nonagle)
1628 /* Nagle rule does not apply to frames, which sit in the middle of the
1629 * write_queue (they have no chances to get new data).
1631 * This is implemented in the callers, where they modify the 'nonagle'
1632 * argument based upon the location of SKB in the send queue.
1634 if (nonagle & TCP_NAGLE_PUSH)
1637 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1638 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1641 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1647 /* Does at least the first segment of SKB fit into the send window? */
1648 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1649 const struct sk_buff *skb,
1650 unsigned int cur_mss)
1652 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1654 if (skb->len > cur_mss)
1655 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1657 return !after(end_seq, tcp_wnd_end(tp));
1660 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1661 * should be put on the wire right now. If so, it returns the number of
1662 * packets allowed by the congestion window.
1664 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1665 unsigned int cur_mss, int nonagle)
1667 const struct tcp_sock *tp = tcp_sk(sk);
1668 unsigned int cwnd_quota;
1670 tcp_init_tso_segs(sk, skb, cur_mss);
1672 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1675 cwnd_quota = tcp_cwnd_test(tp, skb);
1676 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1682 /* Test if sending is allowed right now. */
1683 bool tcp_may_send_now(struct sock *sk)
1685 const struct tcp_sock *tp = tcp_sk(sk);
1686 struct sk_buff *skb = tcp_send_head(sk);
1689 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1690 (tcp_skb_is_last(sk, skb) ?
1691 tp->nonagle : TCP_NAGLE_PUSH));
1694 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1695 * which is put after SKB on the list. It is very much like
1696 * tcp_fragment() except that it may make several kinds of assumptions
1697 * in order to speed up the splitting operation. In particular, we
1698 * know that all the data is in scatter-gather pages, and that the
1699 * packet has never been sent out before (and thus is not cloned).
1701 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1702 unsigned int mss_now, gfp_t gfp)
1704 struct sk_buff *buff;
1705 int nlen = skb->len - len;
1708 /* All of a TSO frame must be composed of paged data. */
1709 if (skb->len != skb->data_len)
1710 return tcp_fragment(sk, skb, len, mss_now, gfp);
1712 buff = sk_stream_alloc_skb(sk, 0, gfp);
1713 if (unlikely(buff == NULL))
1716 sk->sk_wmem_queued += buff->truesize;
1717 sk_mem_charge(sk, buff->truesize);
1718 buff->truesize += nlen;
1719 skb->truesize -= nlen;
1721 /* Correct the sequence numbers. */
1722 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1723 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1724 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1726 /* PSH and FIN should only be set in the second packet. */
1727 flags = TCP_SKB_CB(skb)->tcp_flags;
1728 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1729 TCP_SKB_CB(buff)->tcp_flags = flags;
1731 /* This packet was never sent out yet, so no SACK bits. */
1732 TCP_SKB_CB(buff)->sacked = 0;
1734 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1735 skb_split(skb, buff, len);
1736 tcp_fragment_tstamp(skb, buff);
1738 /* Fix up tso_factor for both original and new SKB. */
1739 tcp_set_skb_tso_segs(sk, skb, mss_now);
1740 tcp_set_skb_tso_segs(sk, buff, mss_now);
1742 /* Link BUFF into the send queue. */
1743 __skb_header_release(buff);
1744 tcp_insert_write_queue_after(skb, buff, sk);
1749 /* Try to defer sending, if possible, in order to minimize the amount
1750 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1752 * This algorithm is from John Heffner.
1754 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1755 bool *is_cwnd_limited, u32 max_segs)
1757 const struct inet_connection_sock *icsk = inet_csk(sk);
1758 u32 age, send_win, cong_win, limit, in_flight;
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct skb_mstamp now;
1761 struct sk_buff *head;
1764 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1767 if (!((1 << icsk->icsk_ca_state) & (TCPF_CA_Open | TCPF_CA_CWR)))
1770 /* Avoid bursty behavior by allowing defer
1771 * only if the last write was recent.
1773 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1776 in_flight = tcp_packets_in_flight(tp);
1778 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1780 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1782 /* From in_flight test above, we know that cwnd > in_flight. */
1783 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1785 limit = min(send_win, cong_win);
1787 /* If a full-sized TSO skb can be sent, do it. */
1788 if (limit >= max_segs * tp->mss_cache)
1791 /* Middle in queue won't get any more data, full sendable already? */
1792 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1795 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1797 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1799 /* If at least some fraction of a window is available,
1802 chunk /= win_divisor;
1806 /* Different approach, try not to defer past a single
1807 * ACK. Receiver should ACK every other full sized
1808 * frame, so if we have space for more than 3 frames
1811 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1815 head = tcp_write_queue_head(sk);
1816 skb_mstamp_get(&now);
1817 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1818 /* If next ACK is likely to come too late (half srtt), do not defer */
1819 if (age < (tp->srtt_us >> 4))
1822 /* Ok, it looks like it is advisable to defer. */
1824 if (cong_win < send_win && cong_win < skb->len)
1825 *is_cwnd_limited = true;
1833 static inline void tcp_mtu_check_reprobe(struct sock *sk)
1835 struct inet_connection_sock *icsk = inet_csk(sk);
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct net *net = sock_net(sk);
1841 interval = net->ipv4.sysctl_tcp_probe_interval;
1842 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1843 if (unlikely(delta >= interval * HZ)) {
1844 int mss = tcp_current_mss(sk);
1846 /* Update current search range */
1847 icsk->icsk_mtup.probe_size = 0;
1848 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1849 sizeof(struct tcphdr) +
1850 icsk->icsk_af_ops->net_header_len;
1851 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1853 /* Update probe time stamp */
1854 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1858 /* Create a new MTU probe if we are ready.
1859 * MTU probe is regularly attempting to increase the path MTU by
1860 * deliberately sending larger packets. This discovers routing
1861 * changes resulting in larger path MTUs.
1863 * Returns 0 if we should wait to probe (no cwnd available),
1864 * 1 if a probe was sent,
1867 static int tcp_mtu_probe(struct sock *sk)
1869 struct tcp_sock *tp = tcp_sk(sk);
1870 struct inet_connection_sock *icsk = inet_csk(sk);
1871 struct sk_buff *skb, *nskb, *next;
1872 struct net *net = sock_net(sk);
1880 /* Not currently probing/verifying,
1882 * have enough cwnd, and
1883 * not SACKing (the variable headers throw things off) */
1884 if (!icsk->icsk_mtup.enabled ||
1885 icsk->icsk_mtup.probe_size ||
1886 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1887 tp->snd_cwnd < 11 ||
1888 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1891 /* Use binary search for probe_size between tcp_mss_base,
1892 * and current mss_clamp. if (search_high - search_low)
1893 * smaller than a threshold, backoff from probing.
1895 mss_now = tcp_current_mss(sk);
1896 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1897 icsk->icsk_mtup.search_low) >> 1);
1898 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1899 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1900 /* When misfortune happens, we are reprobing actively,
1901 * and then reprobe timer has expired. We stick with current
1902 * probing process by not resetting search range to its orignal.
1904 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1905 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1906 /* Check whether enough time has elaplased for
1907 * another round of probing.
1909 tcp_mtu_check_reprobe(sk);
1913 /* Have enough data in the send queue to probe? */
1914 if (tp->write_seq - tp->snd_nxt < size_needed)
1917 if (tp->snd_wnd < size_needed)
1919 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1922 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1923 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1924 if (!tcp_packets_in_flight(tp))
1930 /* We're allowed to probe. Build it now. */
1931 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1933 sk->sk_wmem_queued += nskb->truesize;
1934 sk_mem_charge(sk, nskb->truesize);
1936 skb = tcp_send_head(sk);
1938 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1939 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1940 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1941 TCP_SKB_CB(nskb)->sacked = 0;
1943 nskb->ip_summed = skb->ip_summed;
1945 tcp_insert_write_queue_before(nskb, skb, sk);
1948 tcp_for_write_queue_from_safe(skb, next, sk) {
1949 copy = min_t(int, skb->len, probe_size - len);
1950 if (nskb->ip_summed)
1951 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1953 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1954 skb_put(nskb, copy),
1957 if (skb->len <= copy) {
1958 /* We've eaten all the data from this skb.
1960 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1961 tcp_unlink_write_queue(skb, sk);
1962 sk_wmem_free_skb(sk, skb);
1964 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1965 ~(TCPHDR_FIN|TCPHDR_PSH);
1966 if (!skb_shinfo(skb)->nr_frags) {
1967 skb_pull(skb, copy);
1968 if (skb->ip_summed != CHECKSUM_PARTIAL)
1969 skb->csum = csum_partial(skb->data,
1972 __pskb_trim_head(skb, copy);
1973 tcp_set_skb_tso_segs(sk, skb, mss_now);
1975 TCP_SKB_CB(skb)->seq += copy;
1980 if (len >= probe_size)
1983 tcp_init_tso_segs(sk, nskb, nskb->len);
1985 /* We're ready to send. If this fails, the probe will
1986 * be resegmented into mss-sized pieces by tcp_write_xmit().
1988 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1989 /* Decrement cwnd here because we are sending
1990 * effectively two packets. */
1992 tcp_event_new_data_sent(sk, nskb);
1994 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1995 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1996 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2004 /* This routine writes packets to the network. It advances the
2005 * send_head. This happens as incoming acks open up the remote
2008 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2009 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2010 * account rare use of URG, this is not a big flaw.
2012 * Send at most one packet when push_one > 0. Temporarily ignore
2013 * cwnd limit to force at most one packet out when push_one == 2.
2015 * Returns true, if no segments are in flight and we have queued segments,
2016 * but cannot send anything now because of SWS or another problem.
2018 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2019 int push_one, gfp_t gfp)
2021 struct tcp_sock *tp = tcp_sk(sk);
2022 struct sk_buff *skb;
2023 unsigned int tso_segs, sent_pkts;
2026 bool is_cwnd_limited = false;
2032 /* Do MTU probing. */
2033 result = tcp_mtu_probe(sk);
2036 } else if (result > 0) {
2041 max_segs = tcp_tso_autosize(sk, mss_now);
2042 while ((skb = tcp_send_head(sk))) {
2045 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
2048 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2049 /* "skb_mstamp" is used as a start point for the retransmit timer */
2050 skb_mstamp_get(&skb->skb_mstamp);
2051 goto repair; /* Skip network transmission */
2054 cwnd_quota = tcp_cwnd_test(tp, skb);
2056 is_cwnd_limited = true;
2058 /* Force out a loss probe pkt. */
2064 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2067 if (tso_segs == 1 || !max_segs) {
2068 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2069 (tcp_skb_is_last(sk, skb) ?
2070 nonagle : TCP_NAGLE_PUSH))))
2074 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2080 if (tso_segs > 1 && max_segs && !tcp_urg_mode(tp))
2081 limit = tcp_mss_split_point(sk, skb, mss_now,
2087 if (skb->len > limit &&
2088 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2091 /* TCP Small Queues :
2092 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2094 * - better RTT estimation and ACK scheduling
2097 * Alas, some drivers / subsystems require a fair amount
2098 * of queued bytes to ensure line rate.
2099 * One example is wifi aggregation (802.11 AMPDU)
2101 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2102 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2104 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2105 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2106 /* It is possible TX completion already happened
2107 * before we set TSQ_THROTTLED, so we must
2108 * test again the condition.
2110 smp_mb__after_atomic();
2111 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2115 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2119 /* Advance the send_head. This one is sent out.
2120 * This call will increment packets_out.
2122 tcp_event_new_data_sent(sk, skb);
2124 tcp_minshall_update(tp, mss_now, skb);
2125 sent_pkts += tcp_skb_pcount(skb);
2131 if (likely(sent_pkts)) {
2132 if (tcp_in_cwnd_reduction(sk))
2133 tp->prr_out += sent_pkts;
2135 /* Send one loss probe per tail loss episode. */
2137 tcp_schedule_loss_probe(sk);
2138 tcp_cwnd_validate(sk, is_cwnd_limited);
2141 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2144 bool tcp_schedule_loss_probe(struct sock *sk)
2146 struct inet_connection_sock *icsk = inet_csk(sk);
2147 struct tcp_sock *tp = tcp_sk(sk);
2148 u32 timeout, tlp_time_stamp, rto_time_stamp;
2149 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2151 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2153 /* No consecutive loss probes. */
2154 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2158 /* Don't do any loss probe on a Fast Open connection before 3WHS
2161 if (sk->sk_state == TCP_SYN_RECV)
2164 /* TLP is only scheduled when next timer event is RTO. */
2165 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2168 /* Schedule a loss probe in 2*RTT for SACK capable connections
2169 * in Open state, that are either limited by cwnd or application.
2171 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2172 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2175 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2179 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2180 * for delayed ack when there's one outstanding packet.
2183 if (tp->packets_out == 1)
2184 timeout = max_t(u32, timeout,
2185 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2186 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2188 /* If RTO is shorter, just schedule TLP in its place. */
2189 tlp_time_stamp = tcp_time_stamp + timeout;
2190 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2191 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2192 s32 delta = rto_time_stamp - tcp_time_stamp;
2197 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2202 /* Thanks to skb fast clones, we can detect if a prior transmit of
2203 * a packet is still in a qdisc or driver queue.
2204 * In this case, there is very little point doing a retransmit !
2205 * Note: This is called from BH context only.
2207 static bool skb_still_in_host_queue(const struct sock *sk,
2208 const struct sk_buff *skb)
2210 if (unlikely(skb_fclone_busy(sk, skb))) {
2211 NET_INC_STATS_BH(sock_net(sk),
2212 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2218 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2219 * retransmit the last segment.
2221 void tcp_send_loss_probe(struct sock *sk)
2223 struct tcp_sock *tp = tcp_sk(sk);
2224 struct sk_buff *skb;
2226 int mss = tcp_current_mss(sk);
2229 if (tcp_send_head(sk) != NULL) {
2230 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2234 /* At most one outstanding TLP retransmission. */
2235 if (tp->tlp_high_seq)
2238 /* Retransmit last segment. */
2239 skb = tcp_write_queue_tail(sk);
2243 if (skb_still_in_host_queue(sk, skb))
2246 pcount = tcp_skb_pcount(skb);
2247 if (WARN_ON(!pcount))
2250 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2251 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2254 skb = tcp_write_queue_tail(sk);
2257 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2260 err = __tcp_retransmit_skb(sk, skb);
2262 /* Record snd_nxt for loss detection. */
2264 tp->tlp_high_seq = tp->snd_nxt;
2267 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2268 inet_csk(sk)->icsk_rto,
2272 NET_INC_STATS_BH(sock_net(sk),
2273 LINUX_MIB_TCPLOSSPROBES);
2276 /* Push out any pending frames which were held back due to
2277 * TCP_CORK or attempt at coalescing tiny packets.
2278 * The socket must be locked by the caller.
2280 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2283 /* If we are closed, the bytes will have to remain here.
2284 * In time closedown will finish, we empty the write queue and
2285 * all will be happy.
2287 if (unlikely(sk->sk_state == TCP_CLOSE))
2290 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2291 sk_gfp_atomic(sk, GFP_ATOMIC)))
2292 tcp_check_probe_timer(sk);
2295 /* Send _single_ skb sitting at the send head. This function requires
2296 * true push pending frames to setup probe timer etc.
2298 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2300 struct sk_buff *skb = tcp_send_head(sk);
2302 BUG_ON(!skb || skb->len < mss_now);
2304 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2307 /* This function returns the amount that we can raise the
2308 * usable window based on the following constraints
2310 * 1. The window can never be shrunk once it is offered (RFC 793)
2311 * 2. We limit memory per socket
2314 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2315 * RECV.NEXT + RCV.WIN fixed until:
2316 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2318 * i.e. don't raise the right edge of the window until you can raise
2319 * it at least MSS bytes.
2321 * Unfortunately, the recommended algorithm breaks header prediction,
2322 * since header prediction assumes th->window stays fixed.
2324 * Strictly speaking, keeping th->window fixed violates the receiver
2325 * side SWS prevention criteria. The problem is that under this rule
2326 * a stream of single byte packets will cause the right side of the
2327 * window to always advance by a single byte.
2329 * Of course, if the sender implements sender side SWS prevention
2330 * then this will not be a problem.
2332 * BSD seems to make the following compromise:
2334 * If the free space is less than the 1/4 of the maximum
2335 * space available and the free space is less than 1/2 mss,
2336 * then set the window to 0.
2337 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2338 * Otherwise, just prevent the window from shrinking
2339 * and from being larger than the largest representable value.
2341 * This prevents incremental opening of the window in the regime
2342 * where TCP is limited by the speed of the reader side taking
2343 * data out of the TCP receive queue. It does nothing about
2344 * those cases where the window is constrained on the sender side
2345 * because the pipeline is full.
2347 * BSD also seems to "accidentally" limit itself to windows that are a
2348 * multiple of MSS, at least until the free space gets quite small.
2349 * This would appear to be a side effect of the mbuf implementation.
2350 * Combining these two algorithms results in the observed behavior
2351 * of having a fixed window size at almost all times.
2353 * Below we obtain similar behavior by forcing the offered window to
2354 * a multiple of the mss when it is feasible to do so.
2356 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2357 * Regular options like TIMESTAMP are taken into account.
2359 u32 __tcp_select_window(struct sock *sk)
2361 struct inet_connection_sock *icsk = inet_csk(sk);
2362 struct tcp_sock *tp = tcp_sk(sk);
2363 /* MSS for the peer's data. Previous versions used mss_clamp
2364 * here. I don't know if the value based on our guesses
2365 * of peer's MSS is better for the performance. It's more correct
2366 * but may be worse for the performance because of rcv_mss
2367 * fluctuations. --SAW 1998/11/1
2369 int mss = icsk->icsk_ack.rcv_mss;
2370 int free_space = tcp_space(sk);
2371 int allowed_space = tcp_full_space(sk);
2372 int full_space = min_t(int, tp->window_clamp, allowed_space);
2375 if (mss > full_space)
2378 if (free_space < (full_space >> 1)) {
2379 icsk->icsk_ack.quick = 0;
2381 if (sk_under_memory_pressure(sk))
2382 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2385 /* free_space might become our new window, make sure we don't
2386 * increase it due to wscale.
2388 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2390 /* if free space is less than mss estimate, or is below 1/16th
2391 * of the maximum allowed, try to move to zero-window, else
2392 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2393 * new incoming data is dropped due to memory limits.
2394 * With large window, mss test triggers way too late in order
2395 * to announce zero window in time before rmem limit kicks in.
2397 if (free_space < (allowed_space >> 4) || free_space < mss)
2401 if (free_space > tp->rcv_ssthresh)
2402 free_space = tp->rcv_ssthresh;
2404 /* Don't do rounding if we are using window scaling, since the
2405 * scaled window will not line up with the MSS boundary anyway.
2407 window = tp->rcv_wnd;
2408 if (tp->rx_opt.rcv_wscale) {
2409 window = free_space;
2411 /* Advertise enough space so that it won't get scaled away.
2412 * Import case: prevent zero window announcement if
2413 * 1<<rcv_wscale > mss.
2415 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2416 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2417 << tp->rx_opt.rcv_wscale);
2419 /* Get the largest window that is a nice multiple of mss.
2420 * Window clamp already applied above.
2421 * If our current window offering is within 1 mss of the
2422 * free space we just keep it. This prevents the divide
2423 * and multiply from happening most of the time.
2424 * We also don't do any window rounding when the free space
2427 if (window <= free_space - mss || window > free_space)
2428 window = (free_space / mss) * mss;
2429 else if (mss == full_space &&
2430 free_space > window + (full_space >> 1))
2431 window = free_space;
2437 /* Collapses two adjacent SKB's during retransmission. */
2438 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2440 struct tcp_sock *tp = tcp_sk(sk);
2441 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2442 int skb_size, next_skb_size;
2444 skb_size = skb->len;
2445 next_skb_size = next_skb->len;
2447 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2449 tcp_highest_sack_combine(sk, next_skb, skb);
2451 tcp_unlink_write_queue(next_skb, sk);
2453 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2456 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2457 skb->ip_summed = CHECKSUM_PARTIAL;
2459 if (skb->ip_summed != CHECKSUM_PARTIAL)
2460 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2462 /* Update sequence range on original skb. */
2463 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2465 /* Merge over control information. This moves PSH/FIN etc. over */
2466 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2468 /* All done, get rid of second SKB and account for it so
2469 * packet counting does not break.
2471 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2473 /* changed transmit queue under us so clear hints */
2474 tcp_clear_retrans_hints_partial(tp);
2475 if (next_skb == tp->retransmit_skb_hint)
2476 tp->retransmit_skb_hint = skb;
2478 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2480 sk_wmem_free_skb(sk, next_skb);
2483 /* Check if coalescing SKBs is legal. */
2484 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2486 if (tcp_skb_pcount(skb) > 1)
2488 /* TODO: SACK collapsing could be used to remove this condition */
2489 if (skb_shinfo(skb)->nr_frags != 0)
2491 if (skb_cloned(skb))
2493 if (skb == tcp_send_head(sk))
2495 /* Some heurestics for collapsing over SACK'd could be invented */
2496 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2502 /* Collapse packets in the retransmit queue to make to create
2503 * less packets on the wire. This is only done on retransmission.
2505 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2508 struct tcp_sock *tp = tcp_sk(sk);
2509 struct sk_buff *skb = to, *tmp;
2512 if (!sysctl_tcp_retrans_collapse)
2514 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2517 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2518 if (!tcp_can_collapse(sk, skb))
2530 /* Punt if not enough space exists in the first SKB for
2531 * the data in the second
2533 if (skb->len > skb_availroom(to))
2536 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2539 tcp_collapse_retrans(sk, to);
2543 /* This retransmits one SKB. Policy decisions and retransmit queue
2544 * state updates are done by the caller. Returns non-zero if an
2545 * error occurred which prevented the send.
2547 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2549 struct tcp_sock *tp = tcp_sk(sk);
2550 struct inet_connection_sock *icsk = inet_csk(sk);
2551 unsigned int cur_mss;
2554 /* Inconslusive MTU probe */
2555 if (icsk->icsk_mtup.probe_size) {
2556 icsk->icsk_mtup.probe_size = 0;
2559 /* Do not sent more than we queued. 1/4 is reserved for possible
2560 * copying overhead: fragmentation, tunneling, mangling etc.
2562 if (atomic_read(&sk->sk_wmem_alloc) >
2563 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2566 if (skb_still_in_host_queue(sk, skb))
2569 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2570 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2572 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2576 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2577 return -EHOSTUNREACH; /* Routing failure or similar. */
2579 cur_mss = tcp_current_mss(sk);
2581 /* If receiver has shrunk his window, and skb is out of
2582 * new window, do not retransmit it. The exception is the
2583 * case, when window is shrunk to zero. In this case
2584 * our retransmit serves as a zero window probe.
2586 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2587 TCP_SKB_CB(skb)->seq != tp->snd_una)
2590 if (skb->len > cur_mss) {
2591 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2592 return -ENOMEM; /* We'll try again later. */
2594 int oldpcount = tcp_skb_pcount(skb);
2596 if (unlikely(oldpcount > 1)) {
2597 if (skb_unclone(skb, GFP_ATOMIC))
2599 tcp_init_tso_segs(sk, skb, cur_mss);
2600 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2604 tcp_retrans_try_collapse(sk, skb, cur_mss);
2606 /* Make a copy, if the first transmission SKB clone we made
2607 * is still in somebody's hands, else make a clone.
2610 /* make sure skb->data is aligned on arches that require it
2611 * and check if ack-trimming & collapsing extended the headroom
2612 * beyond what csum_start can cover.
2614 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2615 skb_headroom(skb) >= 0xFFFF)) {
2616 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2618 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2621 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2625 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2626 /* Update global TCP statistics. */
2627 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2628 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2629 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2630 tp->total_retrans++;
2635 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2637 struct tcp_sock *tp = tcp_sk(sk);
2638 int err = __tcp_retransmit_skb(sk, skb);
2641 #if FASTRETRANS_DEBUG > 0
2642 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2643 net_dbg_ratelimited("retrans_out leaked\n");
2646 if (!tp->retrans_out)
2647 tp->lost_retrans_low = tp->snd_nxt;
2648 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2649 tp->retrans_out += tcp_skb_pcount(skb);
2651 /* Save stamp of the first retransmit. */
2652 if (!tp->retrans_stamp)
2653 tp->retrans_stamp = tcp_skb_timestamp(skb);
2655 /* snd_nxt is stored to detect loss of retransmitted segment,
2656 * see tcp_input.c tcp_sacktag_write_queue().
2658 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2659 } else if (err != -EBUSY) {
2660 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2663 if (tp->undo_retrans < 0)
2664 tp->undo_retrans = 0;
2665 tp->undo_retrans += tcp_skb_pcount(skb);
2669 /* Check if we forward retransmits are possible in the current
2670 * window/congestion state.
2672 static bool tcp_can_forward_retransmit(struct sock *sk)
2674 const struct inet_connection_sock *icsk = inet_csk(sk);
2675 const struct tcp_sock *tp = tcp_sk(sk);
2677 /* Forward retransmissions are possible only during Recovery. */
2678 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2681 /* No forward retransmissions in Reno are possible. */
2682 if (tcp_is_reno(tp))
2685 /* Yeah, we have to make difficult choice between forward transmission
2686 * and retransmission... Both ways have their merits...
2688 * For now we do not retransmit anything, while we have some new
2689 * segments to send. In the other cases, follow rule 3 for
2690 * NextSeg() specified in RFC3517.
2693 if (tcp_may_send_now(sk))
2699 /* This gets called after a retransmit timeout, and the initially
2700 * retransmitted data is acknowledged. It tries to continue
2701 * resending the rest of the retransmit queue, until either
2702 * we've sent it all or the congestion window limit is reached.
2703 * If doing SACK, the first ACK which comes back for a timeout
2704 * based retransmit packet might feed us FACK information again.
2705 * If so, we use it to avoid unnecessarily retransmissions.
2707 void tcp_xmit_retransmit_queue(struct sock *sk)
2709 const struct inet_connection_sock *icsk = inet_csk(sk);
2710 struct tcp_sock *tp = tcp_sk(sk);
2711 struct sk_buff *skb;
2712 struct sk_buff *hole = NULL;
2715 int fwd_rexmitting = 0;
2717 if (!tp->packets_out)
2721 tp->retransmit_high = tp->snd_una;
2723 if (tp->retransmit_skb_hint) {
2724 skb = tp->retransmit_skb_hint;
2725 last_lost = TCP_SKB_CB(skb)->end_seq;
2726 if (after(last_lost, tp->retransmit_high))
2727 last_lost = tp->retransmit_high;
2729 skb = tcp_write_queue_head(sk);
2730 last_lost = tp->snd_una;
2733 tcp_for_write_queue_from(skb, sk) {
2734 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2736 if (skb == tcp_send_head(sk))
2738 /* we could do better than to assign each time */
2740 tp->retransmit_skb_hint = skb;
2742 /* Assume this retransmit will generate
2743 * only one packet for congestion window
2744 * calculation purposes. This works because
2745 * tcp_retransmit_skb() will chop up the
2746 * packet to be MSS sized and all the
2747 * packet counting works out.
2749 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2752 if (fwd_rexmitting) {
2754 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2756 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2758 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2759 tp->retransmit_high = last_lost;
2760 if (!tcp_can_forward_retransmit(sk))
2762 /* Backtrack if necessary to non-L'ed skb */
2770 } else if (!(sacked & TCPCB_LOST)) {
2771 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2776 last_lost = TCP_SKB_CB(skb)->end_seq;
2777 if (icsk->icsk_ca_state != TCP_CA_Loss)
2778 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2780 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2783 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2786 if (tcp_retransmit_skb(sk, skb))
2789 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2791 if (tcp_in_cwnd_reduction(sk))
2792 tp->prr_out += tcp_skb_pcount(skb);
2794 if (skb == tcp_write_queue_head(sk))
2795 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2796 inet_csk(sk)->icsk_rto,
2801 /* Send a fin. The caller locks the socket for us. This cannot be
2802 * allowed to fail queueing a FIN frame under any circumstances.
2804 void tcp_send_fin(struct sock *sk)
2806 struct tcp_sock *tp = tcp_sk(sk);
2807 struct sk_buff *skb = tcp_write_queue_tail(sk);
2810 /* Optimization, tack on the FIN if we have a queue of
2811 * unsent frames. But be careful about outgoing SACKS
2814 mss_now = tcp_current_mss(sk);
2816 if (tcp_send_head(sk) != NULL) {
2817 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2818 TCP_SKB_CB(skb)->end_seq++;
2821 /* Socket is locked, keep trying until memory is available. */
2823 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2830 /* Reserve space for headers and prepare control bits. */
2831 skb_reserve(skb, MAX_TCP_HEADER);
2832 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2833 tcp_init_nondata_skb(skb, tp->write_seq,
2834 TCPHDR_ACK | TCPHDR_FIN);
2835 tcp_queue_skb(sk, skb);
2837 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2840 /* We get here when a process closes a file descriptor (either due to
2841 * an explicit close() or as a byproduct of exit()'ing) and there
2842 * was unread data in the receive queue. This behavior is recommended
2843 * by RFC 2525, section 2.17. -DaveM
2845 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2847 struct sk_buff *skb;
2849 /* NOTE: No TCP options attached and we never retransmit this. */
2850 skb = alloc_skb(MAX_TCP_HEADER, priority);
2852 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2856 /* Reserve space for headers and prepare control bits. */
2857 skb_reserve(skb, MAX_TCP_HEADER);
2858 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2859 TCPHDR_ACK | TCPHDR_RST);
2861 if (tcp_transmit_skb(sk, skb, 0, priority))
2862 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2864 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2867 /* Send a crossed SYN-ACK during socket establishment.
2868 * WARNING: This routine must only be called when we have already sent
2869 * a SYN packet that crossed the incoming SYN that caused this routine
2870 * to get called. If this assumption fails then the initial rcv_wnd
2871 * and rcv_wscale values will not be correct.
2873 int tcp_send_synack(struct sock *sk)
2875 struct sk_buff *skb;
2877 skb = tcp_write_queue_head(sk);
2878 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2879 pr_debug("%s: wrong queue state\n", __func__);
2882 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2883 if (skb_cloned(skb)) {
2884 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2887 tcp_unlink_write_queue(skb, sk);
2888 __skb_header_release(nskb);
2889 __tcp_add_write_queue_head(sk, nskb);
2890 sk_wmem_free_skb(sk, skb);
2891 sk->sk_wmem_queued += nskb->truesize;
2892 sk_mem_charge(sk, nskb->truesize);
2896 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2897 tcp_ecn_send_synack(sk, skb);
2899 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2903 * tcp_make_synack - Prepare a SYN-ACK.
2904 * sk: listener socket
2905 * dst: dst entry attached to the SYNACK
2906 * req: request_sock pointer
2908 * Allocate one skb and build a SYNACK packet.
2909 * @dst is consumed : Caller should not use it again.
2911 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2912 struct request_sock *req,
2913 struct tcp_fastopen_cookie *foc)
2915 struct tcp_out_options opts;
2916 struct inet_request_sock *ireq = inet_rsk(req);
2917 struct tcp_sock *tp = tcp_sk(sk);
2919 struct sk_buff *skb;
2920 struct tcp_md5sig_key *md5;
2921 int tcp_header_size;
2924 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2925 if (unlikely(!skb)) {
2929 /* Reserve space for headers. */
2930 skb_reserve(skb, MAX_TCP_HEADER);
2932 skb_dst_set(skb, dst);
2933 security_skb_owned_by(skb, sk);
2935 mss = dst_metric_advmss(dst);
2936 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2937 mss = tp->rx_opt.user_mss;
2939 memset(&opts, 0, sizeof(opts));
2940 #ifdef CONFIG_SYN_COOKIES
2941 if (unlikely(req->cookie_ts))
2942 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2945 skb_mstamp_get(&skb->skb_mstamp);
2946 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2949 skb_push(skb, tcp_header_size);
2950 skb_reset_transport_header(skb);
2953 memset(th, 0, sizeof(struct tcphdr));
2956 tcp_ecn_make_synack(req, th, sk);
2957 th->source = htons(ireq->ir_num);
2958 th->dest = ireq->ir_rmt_port;
2959 /* Setting of flags are superfluous here for callers (and ECE is
2960 * not even correctly set)
2962 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2963 TCPHDR_SYN | TCPHDR_ACK);
2965 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2966 /* XXX data is queued and acked as is. No buffer/window check */
2967 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2969 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2970 th->window = htons(min(req->rcv_wnd, 65535U));
2971 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2972 th->doff = (tcp_header_size >> 2);
2973 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2975 #ifdef CONFIG_TCP_MD5SIG
2976 /* Okay, we have all we need - do the md5 hash if needed */
2978 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2979 md5, NULL, req, skb);
2985 EXPORT_SYMBOL(tcp_make_synack);
2987 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
2989 struct inet_connection_sock *icsk = inet_csk(sk);
2990 const struct tcp_congestion_ops *ca;
2991 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
2993 if (ca_key == TCP_CA_UNSPEC)
2997 ca = tcp_ca_find_key(ca_key);
2998 if (likely(ca && try_module_get(ca->owner))) {
2999 module_put(icsk->icsk_ca_ops->owner);
3000 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3001 icsk->icsk_ca_ops = ca;
3006 /* Do all connect socket setups that can be done AF independent. */
3007 static void tcp_connect_init(struct sock *sk)
3009 const struct dst_entry *dst = __sk_dst_get(sk);
3010 struct tcp_sock *tp = tcp_sk(sk);
3013 /* We'll fix this up when we get a response from the other end.
3014 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3016 tp->tcp_header_len = sizeof(struct tcphdr) +
3017 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3019 #ifdef CONFIG_TCP_MD5SIG
3020 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
3021 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3024 /* If user gave his TCP_MAXSEG, record it to clamp */
3025 if (tp->rx_opt.user_mss)
3026 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3029 tcp_sync_mss(sk, dst_mtu(dst));
3031 tcp_ca_dst_init(sk, dst);
3033 if (!tp->window_clamp)
3034 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3035 tp->advmss = dst_metric_advmss(dst);
3036 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3037 tp->advmss = tp->rx_opt.user_mss;
3039 tcp_initialize_rcv_mss(sk);
3041 /* limit the window selection if the user enforce a smaller rx buffer */
3042 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3043 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3044 tp->window_clamp = tcp_full_space(sk);
3046 tcp_select_initial_window(tcp_full_space(sk),
3047 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3050 sysctl_tcp_window_scaling,
3052 dst_metric(dst, RTAX_INITRWND));
3054 tp->rx_opt.rcv_wscale = rcv_wscale;
3055 tp->rcv_ssthresh = tp->rcv_wnd;
3058 sock_reset_flag(sk, SOCK_DONE);
3061 tp->snd_una = tp->write_seq;
3062 tp->snd_sml = tp->write_seq;
3063 tp->snd_up = tp->write_seq;
3064 tp->snd_nxt = tp->write_seq;
3066 if (likely(!tp->repair))
3069 tp->rcv_tstamp = tcp_time_stamp;
3070 tp->rcv_wup = tp->rcv_nxt;
3071 tp->copied_seq = tp->rcv_nxt;
3073 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3074 inet_csk(sk)->icsk_retransmits = 0;
3075 tcp_clear_retrans(tp);
3078 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3080 struct tcp_sock *tp = tcp_sk(sk);
3081 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3083 tcb->end_seq += skb->len;
3084 __skb_header_release(skb);
3085 __tcp_add_write_queue_tail(sk, skb);
3086 sk->sk_wmem_queued += skb->truesize;
3087 sk_mem_charge(sk, skb->truesize);
3088 tp->write_seq = tcb->end_seq;
3089 tp->packets_out += tcp_skb_pcount(skb);
3092 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3093 * queue a data-only packet after the regular SYN, such that regular SYNs
3094 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3095 * only the SYN sequence, the data are retransmitted in the first ACK.
3096 * If cookie is not cached or other error occurs, falls back to send a
3097 * regular SYN with Fast Open cookie request option.
3099 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3101 struct tcp_sock *tp = tcp_sk(sk);
3102 struct tcp_fastopen_request *fo = tp->fastopen_req;
3103 int syn_loss = 0, space, err = 0, copied;
3104 unsigned long last_syn_loss = 0;
3105 struct sk_buff *syn_data;
3107 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3108 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3109 &syn_loss, &last_syn_loss);
3110 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3112 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3113 fo->cookie.len = -1;
3117 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3118 fo->cookie.len = -1;
3119 else if (fo->cookie.len <= 0)
3122 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3123 * user-MSS. Reserve maximum option space for middleboxes that add
3124 * private TCP options. The cost is reduced data space in SYN :(
3126 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3127 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3128 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3129 MAX_TCP_OPTION_SPACE;
3131 space = min_t(size_t, space, fo->size);
3133 /* limit to order-0 allocations */
3134 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3136 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation);
3139 syn_data->ip_summed = CHECKSUM_PARTIAL;
3140 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3141 copied = copy_from_iter(skb_put(syn_data, space), space,
3142 &fo->data->msg_iter);
3143 if (unlikely(!copied)) {
3144 kfree_skb(syn_data);
3147 if (copied != space) {
3148 skb_trim(syn_data, copied);
3152 /* No more data pending in inet_wait_for_connect() */
3153 if (space == fo->size)
3157 tcp_connect_queue_skb(sk, syn_data);
3159 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3161 syn->skb_mstamp = syn_data->skb_mstamp;
3163 /* Now full SYN+DATA was cloned and sent (or not),
3164 * remove the SYN from the original skb (syn_data)
3165 * we keep in write queue in case of a retransmit, as we
3166 * also have the SYN packet (with no data) in the same queue.
3168 TCP_SKB_CB(syn_data)->seq++;
3169 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3171 tp->syn_data = (fo->copied > 0);
3172 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3177 /* Send a regular SYN with Fast Open cookie request option */
3178 if (fo->cookie.len > 0)
3180 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3182 tp->syn_fastopen = 0;
3184 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3188 /* Build a SYN and send it off. */
3189 int tcp_connect(struct sock *sk)
3191 struct tcp_sock *tp = tcp_sk(sk);
3192 struct sk_buff *buff;
3195 tcp_connect_init(sk);
3197 if (unlikely(tp->repair)) {
3198 tcp_finish_connect(sk, NULL);
3202 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation);
3203 if (unlikely(!buff))
3206 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3207 tp->retrans_stamp = tcp_time_stamp;
3208 tcp_connect_queue_skb(sk, buff);
3209 tcp_ecn_send_syn(sk, buff);
3211 /* Send off SYN; include data in Fast Open. */
3212 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3213 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3214 if (err == -ECONNREFUSED)
3217 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3218 * in order to make this packet get counted in tcpOutSegs.
3220 tp->snd_nxt = tp->write_seq;
3221 tp->pushed_seq = tp->write_seq;
3222 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3224 /* Timer for repeating the SYN until an answer. */
3225 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3226 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3229 EXPORT_SYMBOL(tcp_connect);
3231 /* Send out a delayed ack, the caller does the policy checking
3232 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3235 void tcp_send_delayed_ack(struct sock *sk)
3237 struct inet_connection_sock *icsk = inet_csk(sk);
3238 int ato = icsk->icsk_ack.ato;
3239 unsigned long timeout;
3241 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3243 if (ato > TCP_DELACK_MIN) {
3244 const struct tcp_sock *tp = tcp_sk(sk);
3245 int max_ato = HZ / 2;
3247 if (icsk->icsk_ack.pingpong ||
3248 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3249 max_ato = TCP_DELACK_MAX;
3251 /* Slow path, intersegment interval is "high". */
3253 /* If some rtt estimate is known, use it to bound delayed ack.
3254 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3258 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3265 ato = min(ato, max_ato);
3268 /* Stay within the limit we were given */
3269 timeout = jiffies + ato;
3271 /* Use new timeout only if there wasn't a older one earlier. */
3272 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3273 /* If delack timer was blocked or is about to expire,
3276 if (icsk->icsk_ack.blocked ||
3277 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3282 if (!time_before(timeout, icsk->icsk_ack.timeout))
3283 timeout = icsk->icsk_ack.timeout;
3285 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3286 icsk->icsk_ack.timeout = timeout;
3287 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3290 /* This routine sends an ack and also updates the window. */
3291 void tcp_send_ack(struct sock *sk)
3293 struct sk_buff *buff;
3295 /* If we have been reset, we may not send again. */
3296 if (sk->sk_state == TCP_CLOSE)
3299 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3301 /* We are not putting this on the write queue, so
3302 * tcp_transmit_skb() will set the ownership to this
3305 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3307 inet_csk_schedule_ack(sk);
3308 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3309 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3310 TCP_DELACK_MAX, TCP_RTO_MAX);
3314 /* Reserve space for headers and prepare control bits. */
3315 skb_reserve(buff, MAX_TCP_HEADER);
3316 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3318 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3320 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3321 * We also avoid tcp_wfree() overhead (cache line miss accessing
3322 * tp->tsq_flags) by using regular sock_wfree()
3324 skb_set_tcp_pure_ack(buff);
3326 /* Send it off, this clears delayed acks for us. */
3327 skb_mstamp_get(&buff->skb_mstamp);
3328 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3330 EXPORT_SYMBOL_GPL(tcp_send_ack);
3332 /* This routine sends a packet with an out of date sequence
3333 * number. It assumes the other end will try to ack it.
3335 * Question: what should we make while urgent mode?
3336 * 4.4BSD forces sending single byte of data. We cannot send
3337 * out of window data, because we have SND.NXT==SND.MAX...
3339 * Current solution: to send TWO zero-length segments in urgent mode:
3340 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3341 * out-of-date with SND.UNA-1 to probe window.
3343 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3345 struct tcp_sock *tp = tcp_sk(sk);
3346 struct sk_buff *skb;
3348 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3349 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3353 /* Reserve space for headers and set control bits. */
3354 skb_reserve(skb, MAX_TCP_HEADER);
3355 /* Use a previous sequence. This should cause the other
3356 * end to send an ack. Don't queue or clone SKB, just
3359 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3360 skb_mstamp_get(&skb->skb_mstamp);
3361 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3364 void tcp_send_window_probe(struct sock *sk)
3366 if (sk->sk_state == TCP_ESTABLISHED) {
3367 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3368 tcp_xmit_probe_skb(sk, 0);
3372 /* Initiate keepalive or window probe from timer. */
3373 int tcp_write_wakeup(struct sock *sk)
3375 struct tcp_sock *tp = tcp_sk(sk);
3376 struct sk_buff *skb;
3378 if (sk->sk_state == TCP_CLOSE)
3381 if ((skb = tcp_send_head(sk)) != NULL &&
3382 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3384 unsigned int mss = tcp_current_mss(sk);
3385 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3387 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3388 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3390 /* We are probing the opening of a window
3391 * but the window size is != 0
3392 * must have been a result SWS avoidance ( sender )
3394 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3396 seg_size = min(seg_size, mss);
3397 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3398 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3400 } else if (!tcp_skb_pcount(skb))
3401 tcp_set_skb_tso_segs(sk, skb, mss);
3403 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3404 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3406 tcp_event_new_data_sent(sk, skb);
3409 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3410 tcp_xmit_probe_skb(sk, 1);
3411 return tcp_xmit_probe_skb(sk, 0);
3415 /* A window probe timeout has occurred. If window is not closed send
3416 * a partial packet else a zero probe.
3418 void tcp_send_probe0(struct sock *sk)
3420 struct inet_connection_sock *icsk = inet_csk(sk);
3421 struct tcp_sock *tp = tcp_sk(sk);
3422 unsigned long probe_max;
3425 err = tcp_write_wakeup(sk);
3427 if (tp->packets_out || !tcp_send_head(sk)) {
3428 /* Cancel probe timer, if it is not required. */
3429 icsk->icsk_probes_out = 0;
3430 icsk->icsk_backoff = 0;
3435 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3436 icsk->icsk_backoff++;
3437 icsk->icsk_probes_out++;
3438 probe_max = TCP_RTO_MAX;
3440 /* If packet was not sent due to local congestion,
3441 * do not backoff and do not remember icsk_probes_out.
3442 * Let local senders to fight for local resources.
3444 * Use accumulated backoff yet.
3446 if (!icsk->icsk_probes_out)
3447 icsk->icsk_probes_out = 1;
3448 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3450 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3451 inet_csk_rto_backoff(icsk, probe_max),
3455 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3457 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3461 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3463 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3464 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3468 EXPORT_SYMBOL(tcp_rtx_synack);