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 int sysctl_tcp_mtu_probing __read_mostly = 0;
63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
65 /* By default, RFC2861 behavior. */
66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
68 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
69 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
71 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
72 int push_one, gfp_t gfp);
74 /* Account for new data that has been sent to the network. */
75 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
77 struct inet_connection_sock *icsk = inet_csk(sk);
78 struct tcp_sock *tp = tcp_sk(sk);
79 unsigned int prior_packets = tp->packets_out;
81 tcp_advance_send_head(sk, skb);
82 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
84 tp->packets_out += tcp_skb_pcount(skb);
85 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
86 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
90 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
94 /* SND.NXT, if window was not shrunk.
95 * If window has been shrunk, what should we make? It is not clear at all.
96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
98 * invalid. OK, let's make this for now:
100 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
102 const struct tcp_sock *tp = tcp_sk(sk);
104 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
107 return tcp_wnd_end(tp);
110 /* Calculate mss to advertise in SYN segment.
111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
113 * 1. It is independent of path mtu.
114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
116 * attached devices, because some buggy hosts are confused by
118 * 4. We do not make 3, we advertise MSS, calculated from first
119 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
120 * This may be overridden via information stored in routing table.
121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
122 * probably even Jumbo".
124 static __u16 tcp_advertise_mss(struct sock *sk)
126 struct tcp_sock *tp = tcp_sk(sk);
127 const struct dst_entry *dst = __sk_dst_get(sk);
128 int mss = tp->advmss;
131 unsigned int metric = dst_metric_advmss(dst);
142 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
143 * This is the first part of cwnd validation mechanism. */
144 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
146 struct tcp_sock *tp = tcp_sk(sk);
147 s32 delta = tcp_time_stamp - tp->lsndtime;
148 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
149 u32 cwnd = tp->snd_cwnd;
151 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
153 tp->snd_ssthresh = tcp_current_ssthresh(sk);
154 restart_cwnd = min(restart_cwnd, cwnd);
156 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
158 tp->snd_cwnd = max(cwnd, restart_cwnd);
159 tp->snd_cwnd_stamp = tcp_time_stamp;
160 tp->snd_cwnd_used = 0;
163 /* Congestion state accounting after a packet has been sent. */
164 static void tcp_event_data_sent(struct tcp_sock *tp,
167 struct inet_connection_sock *icsk = inet_csk(sk);
168 const u32 now = tcp_time_stamp;
169 const struct dst_entry *dst = __sk_dst_get(sk);
171 if (sysctl_tcp_slow_start_after_idle &&
172 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
173 tcp_cwnd_restart(sk, __sk_dst_get(sk));
177 /* If it is a reply for ato after last received
178 * packet, enter pingpong mode.
180 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
181 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
182 icsk->icsk_ack.pingpong = 1;
185 /* Account for an ACK we sent. */
186 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
188 tcp_dec_quickack_mode(sk, pkts);
189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
193 u32 tcp_default_init_rwnd(u32 mss)
195 /* Initial receive window should be twice of TCP_INIT_CWND to
196 * enable proper sending of new unsent data during fast recovery
197 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
198 * limit when mss is larger than 1460.
200 u32 init_rwnd = TCP_INIT_CWND * 2;
203 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
207 /* Determine a window scaling and initial window to offer.
208 * Based on the assumption that the given amount of space
209 * will be offered. Store the results in the tp structure.
210 * NOTE: for smooth operation initial space offering should
211 * be a multiple of mss if possible. We assume here that mss >= 1.
212 * This MUST be enforced by all callers.
214 void tcp_select_initial_window(int __space, __u32 mss,
215 __u32 *rcv_wnd, __u32 *window_clamp,
216 int wscale_ok, __u8 *rcv_wscale,
219 unsigned int space = (__space < 0 ? 0 : __space);
221 /* If no clamp set the clamp to the max possible scaled window */
222 if (*window_clamp == 0)
223 (*window_clamp) = (65535 << 14);
224 space = min(*window_clamp, space);
226 /* Quantize space offering to a multiple of mss if possible. */
228 space = (space / mss) * mss;
230 /* NOTE: offering an initial window larger than 32767
231 * will break some buggy TCP stacks. If the admin tells us
232 * it is likely we could be speaking with such a buggy stack
233 * we will truncate our initial window offering to 32K-1
234 * unless the remote has sent us a window scaling option,
235 * which we interpret as a sign the remote TCP is not
236 * misinterpreting the window field as a signed quantity.
238 if (sysctl_tcp_workaround_signed_windows)
239 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
245 /* Set window scaling on max possible window
246 * See RFC1323 for an explanation of the limit to 14
248 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
249 space = min_t(u32, space, *window_clamp);
250 while (space > 65535 && (*rcv_wscale) < 14) {
256 if (mss > (1 << *rcv_wscale)) {
257 if (!init_rcv_wnd) /* Use default unless specified otherwise */
258 init_rcv_wnd = tcp_default_init_rwnd(mss);
259 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
262 /* Set the clamp no higher than max representable value */
263 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
265 EXPORT_SYMBOL(tcp_select_initial_window);
267 /* Chose a new window to advertise, update state in tcp_sock for the
268 * socket, and return result with RFC1323 scaling applied. The return
269 * value can be stuffed directly into th->window for an outgoing
272 static u16 tcp_select_window(struct sock *sk)
274 struct tcp_sock *tp = tcp_sk(sk);
275 u32 old_win = tp->rcv_wnd;
276 u32 cur_win = tcp_receive_window(tp);
277 u32 new_win = __tcp_select_window(sk);
279 /* Never shrink the offered window */
280 if (new_win < cur_win) {
281 /* Danger Will Robinson!
282 * Don't update rcv_wup/rcv_wnd here or else
283 * we will not be able to advertise a zero
284 * window in time. --DaveM
286 * Relax Will Robinson.
289 NET_INC_STATS(sock_net(sk),
290 LINUX_MIB_TCPWANTZEROWINDOWADV);
291 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
293 tp->rcv_wnd = new_win;
294 tp->rcv_wup = tp->rcv_nxt;
296 /* Make sure we do not exceed the maximum possible
299 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
300 new_win = min(new_win, MAX_TCP_WINDOW);
302 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
304 /* RFC1323 scaling applied */
305 new_win >>= tp->rx_opt.rcv_wscale;
307 /* If we advertise zero window, disable fast path. */
311 NET_INC_STATS(sock_net(sk),
312 LINUX_MIB_TCPTOZEROWINDOWADV);
313 } else if (old_win == 0) {
314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
320 /* Packet ECN state for a SYN-ACK */
321 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
323 const struct tcp_sock *tp = tcp_sk(sk);
325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
326 if (!(tp->ecn_flags & TCP_ECN_OK))
327 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
328 else if (tcp_ca_needs_ecn(sk))
332 /* Packet ECN state for a SYN. */
333 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
335 struct tcp_sock *tp = tcp_sk(sk);
338 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
339 tcp_ca_needs_ecn(sk)) {
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk))
348 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
351 if (inet_rsk(req)->ecn_ok) {
353 if (tcp_ca_needs_ecn(sk))
358 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
364 struct tcp_sock *tp = tcp_sk(sk);
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
373 tcp_hdr(skb)->cwr = 1;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
376 } else if (!tcp_ca_needs_ecn(sk)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
381 tcp_hdr(skb)->ece = 1;
385 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
388 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
390 struct skb_shared_info *shinfo = skb_shinfo(skb);
392 skb->ip_summed = CHECKSUM_PARTIAL;
395 TCP_SKB_CB(skb)->tcp_flags = flags;
396 TCP_SKB_CB(skb)->sacked = 0;
398 tcp_skb_pcount_set(skb, 1);
399 shinfo->gso_size = 0;
400 shinfo->gso_type = 0;
402 TCP_SKB_CB(skb)->seq = seq;
403 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
405 TCP_SKB_CB(skb)->end_seq = seq;
408 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
410 return tp->snd_una != tp->snd_up;
413 #define OPTION_SACK_ADVERTISE (1 << 0)
414 #define OPTION_TS (1 << 1)
415 #define OPTION_MD5 (1 << 2)
416 #define OPTION_WSCALE (1 << 3)
417 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
419 struct tcp_out_options {
420 u16 options; /* bit field of OPTION_* */
421 u16 mss; /* 0 to disable */
422 u8 ws; /* window scale, 0 to disable */
423 u8 num_sack_blocks; /* number of SACK blocks to include */
424 u8 hash_size; /* bytes in hash_location */
425 __u8 *hash_location; /* temporary pointer, overloaded */
426 __u32 tsval, tsecr; /* need to include OPTION_TS */
427 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
430 /* Write previously computed TCP options to the packet.
432 * Beware: Something in the Internet is very sensitive to the ordering of
433 * TCP options, we learned this through the hard way, so be careful here.
434 * Luckily we can at least blame others for their non-compliance but from
435 * inter-operability perspective it seems that we're somewhat stuck with
436 * the ordering which we have been using if we want to keep working with
437 * those broken things (not that it currently hurts anybody as there isn't
438 * particular reason why the ordering would need to be changed).
440 * At least SACK_PERM as the first option is known to lead to a disaster
441 * (but it may well be that other scenarios fail similarly).
443 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
444 struct tcp_out_options *opts)
446 u16 options = opts->options; /* mungable copy */
448 if (unlikely(OPTION_MD5 & options)) {
449 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
450 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
451 /* overload cookie hash location */
452 opts->hash_location = (__u8 *)ptr;
456 if (unlikely(opts->mss)) {
457 *ptr++ = htonl((TCPOPT_MSS << 24) |
458 (TCPOLEN_MSS << 16) |
462 if (likely(OPTION_TS & options)) {
463 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
464 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
465 (TCPOLEN_SACK_PERM << 16) |
466 (TCPOPT_TIMESTAMP << 8) |
468 options &= ~OPTION_SACK_ADVERTISE;
470 *ptr++ = htonl((TCPOPT_NOP << 24) |
472 (TCPOPT_TIMESTAMP << 8) |
475 *ptr++ = htonl(opts->tsval);
476 *ptr++ = htonl(opts->tsecr);
479 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
480 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_SACK_PERM << 8) |
486 if (unlikely(OPTION_WSCALE & options)) {
487 *ptr++ = htonl((TCPOPT_NOP << 24) |
488 (TCPOPT_WINDOW << 16) |
489 (TCPOLEN_WINDOW << 8) |
493 if (unlikely(opts->num_sack_blocks)) {
494 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
495 tp->duplicate_sack : tp->selective_acks;
498 *ptr++ = htonl((TCPOPT_NOP << 24) |
501 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
502 TCPOLEN_SACK_PERBLOCK)));
504 for (this_sack = 0; this_sack < opts->num_sack_blocks;
506 *ptr++ = htonl(sp[this_sack].start_seq);
507 *ptr++ = htonl(sp[this_sack].end_seq);
510 tp->rx_opt.dsack = 0;
513 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
514 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
516 *ptr++ = htonl((TCPOPT_EXP << 24) |
517 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
518 TCPOPT_FASTOPEN_MAGIC);
520 memcpy(ptr, foc->val, foc->len);
521 if ((foc->len & 3) == 2) {
522 u8 *align = ((u8 *)ptr) + foc->len;
523 align[0] = align[1] = TCPOPT_NOP;
525 ptr += (foc->len + 3) >> 2;
529 /* Compute TCP options for SYN packets. This is not the final
530 * network wire format yet.
532 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
533 struct tcp_out_options *opts,
534 struct tcp_md5sig_key **md5)
536 struct tcp_sock *tp = tcp_sk(sk);
537 unsigned int remaining = MAX_TCP_OPTION_SPACE;
538 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
540 #ifdef CONFIG_TCP_MD5SIG
541 *md5 = tp->af_specific->md5_lookup(sk, sk);
543 opts->options |= OPTION_MD5;
544 remaining -= TCPOLEN_MD5SIG_ALIGNED;
550 /* We always get an MSS option. The option bytes which will be seen in
551 * normal data packets should timestamps be used, must be in the MSS
552 * advertised. But we subtract them from tp->mss_cache so that
553 * calculations in tcp_sendmsg are simpler etc. So account for this
554 * fact here if necessary. If we don't do this correctly, as a
555 * receiver we won't recognize data packets as being full sized when we
556 * should, and thus we won't abide by the delayed ACK rules correctly.
557 * SACKs don't matter, we never delay an ACK when we have any of those
559 opts->mss = tcp_advertise_mss(sk);
560 remaining -= TCPOLEN_MSS_ALIGNED;
562 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
563 opts->options |= OPTION_TS;
564 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
565 opts->tsecr = tp->rx_opt.ts_recent;
566 remaining -= TCPOLEN_TSTAMP_ALIGNED;
568 if (likely(sysctl_tcp_window_scaling)) {
569 opts->ws = tp->rx_opt.rcv_wscale;
570 opts->options |= OPTION_WSCALE;
571 remaining -= TCPOLEN_WSCALE_ALIGNED;
573 if (likely(sysctl_tcp_sack)) {
574 opts->options |= OPTION_SACK_ADVERTISE;
575 if (unlikely(!(OPTION_TS & opts->options)))
576 remaining -= TCPOLEN_SACKPERM_ALIGNED;
579 if (fastopen && fastopen->cookie.len >= 0) {
580 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
581 need = (need + 3) & ~3U; /* Align to 32 bits */
582 if (remaining >= need) {
583 opts->options |= OPTION_FAST_OPEN_COOKIE;
584 opts->fastopen_cookie = &fastopen->cookie;
586 tp->syn_fastopen = 1;
590 return MAX_TCP_OPTION_SPACE - remaining;
593 /* Set up TCP options for SYN-ACKs. */
594 static unsigned int tcp_synack_options(struct sock *sk,
595 struct request_sock *req,
596 unsigned int mss, struct sk_buff *skb,
597 struct tcp_out_options *opts,
598 struct tcp_md5sig_key **md5,
599 struct tcp_fastopen_cookie *foc)
601 struct inet_request_sock *ireq = inet_rsk(req);
602 unsigned int remaining = MAX_TCP_OPTION_SPACE;
604 #ifdef CONFIG_TCP_MD5SIG
605 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
607 opts->options |= OPTION_MD5;
608 remaining -= TCPOLEN_MD5SIG_ALIGNED;
610 /* We can't fit any SACK blocks in a packet with MD5 + TS
611 * options. There was discussion about disabling SACK
612 * rather than TS in order to fit in better with old,
613 * buggy kernels, but that was deemed to be unnecessary.
615 ireq->tstamp_ok &= !ireq->sack_ok;
621 /* We always send an MSS option. */
623 remaining -= TCPOLEN_MSS_ALIGNED;
625 if (likely(ireq->wscale_ok)) {
626 opts->ws = ireq->rcv_wscale;
627 opts->options |= OPTION_WSCALE;
628 remaining -= TCPOLEN_WSCALE_ALIGNED;
630 if (likely(ireq->tstamp_ok)) {
631 opts->options |= OPTION_TS;
632 opts->tsval = tcp_skb_timestamp(skb);
633 opts->tsecr = req->ts_recent;
634 remaining -= TCPOLEN_TSTAMP_ALIGNED;
636 if (likely(ireq->sack_ok)) {
637 opts->options |= OPTION_SACK_ADVERTISE;
638 if (unlikely(!ireq->tstamp_ok))
639 remaining -= TCPOLEN_SACKPERM_ALIGNED;
641 if (foc != NULL && foc->len >= 0) {
642 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
643 need = (need + 3) & ~3U; /* Align to 32 bits */
644 if (remaining >= need) {
645 opts->options |= OPTION_FAST_OPEN_COOKIE;
646 opts->fastopen_cookie = foc;
651 return MAX_TCP_OPTION_SPACE - remaining;
654 /* Compute TCP options for ESTABLISHED sockets. This is not the
655 * final wire format yet.
657 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
658 struct tcp_out_options *opts,
659 struct tcp_md5sig_key **md5)
661 struct tcp_sock *tp = tcp_sk(sk);
662 unsigned int size = 0;
663 unsigned int eff_sacks;
667 #ifdef CONFIG_TCP_MD5SIG
668 *md5 = tp->af_specific->md5_lookup(sk, sk);
669 if (unlikely(*md5)) {
670 opts->options |= OPTION_MD5;
671 size += TCPOLEN_MD5SIG_ALIGNED;
677 if (likely(tp->rx_opt.tstamp_ok)) {
678 opts->options |= OPTION_TS;
679 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
680 opts->tsecr = tp->rx_opt.ts_recent;
681 size += TCPOLEN_TSTAMP_ALIGNED;
684 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
685 if (unlikely(eff_sacks)) {
686 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
687 opts->num_sack_blocks =
688 min_t(unsigned int, eff_sacks,
689 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
690 TCPOLEN_SACK_PERBLOCK);
691 size += TCPOLEN_SACK_BASE_ALIGNED +
692 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
699 /* TCP SMALL QUEUES (TSQ)
701 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
702 * to reduce RTT and bufferbloat.
703 * We do this using a special skb destructor (tcp_wfree).
705 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
706 * needs to be reallocated in a driver.
707 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
709 * Since transmit from skb destructor is forbidden, we use a tasklet
710 * to process all sockets that eventually need to send more skbs.
711 * We use one tasklet per cpu, with its own queue of sockets.
714 struct tasklet_struct tasklet;
715 struct list_head head; /* queue of tcp sockets */
717 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
719 static void tcp_tsq_handler(struct sock *sk)
721 if ((1 << sk->sk_state) &
722 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
723 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
724 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
728 * One tasklet per cpu tries to send more skbs.
729 * We run in tasklet context but need to disable irqs when
730 * transferring tsq->head because tcp_wfree() might
731 * interrupt us (non NAPI drivers)
733 static void tcp_tasklet_func(unsigned long data)
735 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
738 struct list_head *q, *n;
742 local_irq_save(flags);
743 list_splice_init(&tsq->head, &list);
744 local_irq_restore(flags);
746 list_for_each_safe(q, n, &list) {
747 tp = list_entry(q, struct tcp_sock, tsq_node);
748 list_del(&tp->tsq_node);
750 sk = (struct sock *)tp;
753 if (!sock_owned_by_user(sk)) {
756 /* defer the work to tcp_release_cb() */
757 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
761 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
766 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
767 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
768 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
769 (1UL << TCP_MTU_REDUCED_DEFERRED))
771 * tcp_release_cb - tcp release_sock() callback
774 * called from release_sock() to perform protocol dependent
775 * actions before socket release.
777 void tcp_release_cb(struct sock *sk)
779 struct tcp_sock *tp = tcp_sk(sk);
780 unsigned long flags, nflags;
782 /* perform an atomic operation only if at least one flag is set */
784 flags = tp->tsq_flags;
785 if (!(flags & TCP_DEFERRED_ALL))
787 nflags = flags & ~TCP_DEFERRED_ALL;
788 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
790 if (flags & (1UL << TCP_TSQ_DEFERRED))
793 /* Here begins the tricky part :
794 * We are called from release_sock() with :
796 * 2) sk_lock.slock spinlock held
797 * 3) socket owned by us (sk->sk_lock.owned == 1)
799 * But following code is meant to be called from BH handlers,
800 * so we should keep BH disabled, but early release socket ownership
802 sock_release_ownership(sk);
804 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
805 tcp_write_timer_handler(sk);
808 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
809 tcp_delack_timer_handler(sk);
812 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
813 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
817 EXPORT_SYMBOL(tcp_release_cb);
819 void __init tcp_tasklet_init(void)
823 for_each_possible_cpu(i) {
824 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
826 INIT_LIST_HEAD(&tsq->head);
827 tasklet_init(&tsq->tasklet,
834 * Write buffer destructor automatically called from kfree_skb.
835 * We can't xmit new skbs from this context, as we might already
838 void tcp_wfree(struct sk_buff *skb)
840 struct sock *sk = skb->sk;
841 struct tcp_sock *tp = tcp_sk(sk);
843 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
844 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
846 struct tsq_tasklet *tsq;
848 /* Keep a ref on socket.
849 * This last ref will be released in tcp_tasklet_func()
851 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
853 /* queue this socket to tasklet queue */
854 local_irq_save(flags);
855 tsq = &__get_cpu_var(tsq_tasklet);
856 list_add(&tp->tsq_node, &tsq->head);
857 tasklet_schedule(&tsq->tasklet);
858 local_irq_restore(flags);
864 /* This routine actually transmits TCP packets queued in by
865 * tcp_do_sendmsg(). This is used by both the initial
866 * transmission and possible later retransmissions.
867 * All SKB's seen here are completely headerless. It is our
868 * job to build the TCP header, and pass the packet down to
869 * IP so it can do the same plus pass the packet off to the
872 * We are working here with either a clone of the original
873 * SKB, or a fresh unique copy made by the retransmit engine.
875 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
878 const struct inet_connection_sock *icsk = inet_csk(sk);
879 struct inet_sock *inet;
881 struct tcp_skb_cb *tcb;
882 struct tcp_out_options opts;
883 unsigned int tcp_options_size, tcp_header_size;
884 struct tcp_md5sig_key *md5;
888 BUG_ON(!skb || !tcp_skb_pcount(skb));
891 skb_mstamp_get(&skb->skb_mstamp);
893 if (unlikely(skb_cloned(skb)))
894 skb = pskb_copy(skb, gfp_mask);
896 skb = skb_clone(skb, gfp_mask);
903 tcb = TCP_SKB_CB(skb);
904 memset(&opts, 0, sizeof(opts));
906 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
907 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
909 tcp_options_size = tcp_established_options(sk, skb, &opts,
911 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
913 if (tcp_packets_in_flight(tp) == 0)
914 tcp_ca_event(sk, CA_EVENT_TX_START);
916 /* if no packet is in qdisc/device queue, then allow XPS to select
919 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
921 skb_push(skb, tcp_header_size);
922 skb_reset_transport_header(skb);
926 skb->destructor = tcp_wfree;
927 skb_set_hash_from_sk(skb, sk);
928 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
930 /* Build TCP header and checksum it. */
932 th->source = inet->inet_sport;
933 th->dest = inet->inet_dport;
934 th->seq = htonl(tcb->seq);
935 th->ack_seq = htonl(tp->rcv_nxt);
936 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
939 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
940 /* RFC1323: The window in SYN & SYN/ACK segments
943 th->window = htons(min(tp->rcv_wnd, 65535U));
945 th->window = htons(tcp_select_window(sk));
950 /* The urg_mode check is necessary during a below snd_una win probe */
951 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
952 if (before(tp->snd_up, tcb->seq + 0x10000)) {
953 th->urg_ptr = htons(tp->snd_up - tcb->seq);
955 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
956 th->urg_ptr = htons(0xFFFF);
961 tcp_options_write((__be32 *)(th + 1), tp, &opts);
962 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
963 tcp_ecn_send(sk, skb, tcp_header_size);
965 #ifdef CONFIG_TCP_MD5SIG
966 /* Calculate the MD5 hash, as we have all we need now */
968 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
969 tp->af_specific->calc_md5_hash(opts.hash_location,
974 icsk->icsk_af_ops->send_check(sk, skb);
976 if (likely(tcb->tcp_flags & TCPHDR_ACK))
977 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
979 if (skb->len != tcp_header_size)
980 tcp_event_data_sent(tp, sk);
982 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
983 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
984 tcp_skb_pcount(skb));
986 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
987 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
989 /* Our usage of tstamp should remain private */
990 skb->tstamp.tv64 = 0;
992 /* Cleanup our debris for IP stacks */
993 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
994 sizeof(struct inet6_skb_parm)));
996 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
998 if (likely(err <= 0))
1003 return net_xmit_eval(err);
1006 /* This routine just queues the buffer for sending.
1008 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1009 * otherwise socket can stall.
1011 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1013 struct tcp_sock *tp = tcp_sk(sk);
1015 /* Advance write_seq and place onto the write_queue. */
1016 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1017 __skb_header_release(skb);
1018 tcp_add_write_queue_tail(sk, skb);
1019 sk->sk_wmem_queued += skb->truesize;
1020 sk_mem_charge(sk, skb->truesize);
1023 /* Initialize TSO segments for a packet. */
1024 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1025 unsigned int mss_now)
1027 struct skb_shared_info *shinfo = skb_shinfo(skb);
1029 /* Make sure we own this skb before messing gso_size/gso_segs */
1030 WARN_ON_ONCE(skb_cloned(skb));
1032 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1033 /* Avoid the costly divide in the normal
1036 tcp_skb_pcount_set(skb, 1);
1037 shinfo->gso_size = 0;
1038 shinfo->gso_type = 0;
1040 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1041 shinfo->gso_size = mss_now;
1042 shinfo->gso_type = sk->sk_gso_type;
1046 /* When a modification to fackets out becomes necessary, we need to check
1047 * skb is counted to fackets_out or not.
1049 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1052 struct tcp_sock *tp = tcp_sk(sk);
1054 if (!tp->sacked_out || tcp_is_reno(tp))
1057 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1058 tp->fackets_out -= decr;
1061 /* Pcount in the middle of the write queue got changed, we need to do various
1062 * tweaks to fix counters
1064 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1066 struct tcp_sock *tp = tcp_sk(sk);
1068 tp->packets_out -= decr;
1070 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1071 tp->sacked_out -= decr;
1072 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1073 tp->retrans_out -= decr;
1074 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1075 tp->lost_out -= decr;
1077 /* Reno case is special. Sigh... */
1078 if (tcp_is_reno(tp) && decr > 0)
1079 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1081 tcp_adjust_fackets_out(sk, skb, decr);
1083 if (tp->lost_skb_hint &&
1084 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1085 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1086 tp->lost_cnt_hint -= decr;
1088 tcp_verify_left_out(tp);
1091 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1093 struct skb_shared_info *shinfo = skb_shinfo(skb);
1095 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1096 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1097 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1098 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1100 shinfo->tx_flags &= ~tsflags;
1101 shinfo2->tx_flags |= tsflags;
1102 swap(shinfo->tskey, shinfo2->tskey);
1106 /* Function to create two new TCP segments. Shrinks the given segment
1107 * to the specified size and appends a new segment with the rest of the
1108 * packet to the list. This won't be called frequently, I hope.
1109 * Remember, these are still headerless SKBs at this point.
1111 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1112 unsigned int mss_now, gfp_t gfp)
1114 struct tcp_sock *tp = tcp_sk(sk);
1115 struct sk_buff *buff;
1116 int nsize, old_factor;
1120 if (WARN_ON(len > skb->len))
1123 nsize = skb_headlen(skb) - len;
1127 if (skb_unclone(skb, gfp))
1130 /* Get a new skb... force flag on. */
1131 buff = sk_stream_alloc_skb(sk, nsize, gfp);
1133 return -ENOMEM; /* We'll just try again later. */
1135 sk->sk_wmem_queued += buff->truesize;
1136 sk_mem_charge(sk, buff->truesize);
1137 nlen = skb->len - len - nsize;
1138 buff->truesize += nlen;
1139 skb->truesize -= nlen;
1141 /* Correct the sequence numbers. */
1142 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1143 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1144 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1146 /* PSH and FIN should only be set in the second packet. */
1147 flags = TCP_SKB_CB(skb)->tcp_flags;
1148 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1149 TCP_SKB_CB(buff)->tcp_flags = flags;
1150 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1152 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1153 /* Copy and checksum data tail into the new buffer. */
1154 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1155 skb_put(buff, nsize),
1160 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1162 skb->ip_summed = CHECKSUM_PARTIAL;
1163 skb_split(skb, buff, len);
1166 buff->ip_summed = skb->ip_summed;
1168 buff->tstamp = skb->tstamp;
1169 tcp_fragment_tstamp(skb, buff);
1171 old_factor = tcp_skb_pcount(skb);
1173 /* Fix up tso_factor for both original and new SKB. */
1174 tcp_set_skb_tso_segs(sk, skb, mss_now);
1175 tcp_set_skb_tso_segs(sk, buff, mss_now);
1177 /* If this packet has been sent out already, we must
1178 * adjust the various packet counters.
1180 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1181 int diff = old_factor - tcp_skb_pcount(skb) -
1182 tcp_skb_pcount(buff);
1185 tcp_adjust_pcount(sk, skb, diff);
1188 /* Link BUFF into the send queue. */
1189 __skb_header_release(buff);
1190 tcp_insert_write_queue_after(skb, buff, sk);
1195 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1196 * eventually). The difference is that pulled data not copied, but
1197 * immediately discarded.
1199 static void __pskb_trim_head(struct sk_buff *skb, int len)
1201 struct skb_shared_info *shinfo;
1204 eat = min_t(int, len, skb_headlen(skb));
1206 __skb_pull(skb, eat);
1213 shinfo = skb_shinfo(skb);
1214 for (i = 0; i < shinfo->nr_frags; i++) {
1215 int size = skb_frag_size(&shinfo->frags[i]);
1218 skb_frag_unref(skb, i);
1221 shinfo->frags[k] = shinfo->frags[i];
1223 shinfo->frags[k].page_offset += eat;
1224 skb_frag_size_sub(&shinfo->frags[k], eat);
1230 shinfo->nr_frags = k;
1232 skb_reset_tail_pointer(skb);
1233 skb->data_len -= len;
1234 skb->len = skb->data_len;
1237 /* Remove acked data from a packet in the transmit queue. */
1238 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1240 if (skb_unclone(skb, GFP_ATOMIC))
1243 __pskb_trim_head(skb, len);
1245 TCP_SKB_CB(skb)->seq += len;
1246 skb->ip_summed = CHECKSUM_PARTIAL;
1248 skb->truesize -= len;
1249 sk->sk_wmem_queued -= len;
1250 sk_mem_uncharge(sk, len);
1251 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1253 /* Any change of skb->len requires recalculation of tso factor. */
1254 if (tcp_skb_pcount(skb) > 1)
1255 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1260 /* Calculate MSS not accounting any TCP options. */
1261 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1263 const struct tcp_sock *tp = tcp_sk(sk);
1264 const struct inet_connection_sock *icsk = inet_csk(sk);
1267 /* Calculate base mss without TCP options:
1268 It is MMS_S - sizeof(tcphdr) of rfc1122
1270 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1272 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1273 if (icsk->icsk_af_ops->net_frag_header_len) {
1274 const struct dst_entry *dst = __sk_dst_get(sk);
1276 if (dst && dst_allfrag(dst))
1277 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1280 /* Clamp it (mss_clamp does not include tcp options) */
1281 if (mss_now > tp->rx_opt.mss_clamp)
1282 mss_now = tp->rx_opt.mss_clamp;
1284 /* Now subtract optional transport overhead */
1285 mss_now -= icsk->icsk_ext_hdr_len;
1287 /* Then reserve room for full set of TCP options and 8 bytes of data */
1293 /* Calculate MSS. Not accounting for SACKs here. */
1294 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1296 /* Subtract TCP options size, not including SACKs */
1297 return __tcp_mtu_to_mss(sk, pmtu) -
1298 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1301 /* Inverse of above */
1302 int tcp_mss_to_mtu(struct sock *sk, int mss)
1304 const struct tcp_sock *tp = tcp_sk(sk);
1305 const struct inet_connection_sock *icsk = inet_csk(sk);
1309 tp->tcp_header_len +
1310 icsk->icsk_ext_hdr_len +
1311 icsk->icsk_af_ops->net_header_len;
1313 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1314 if (icsk->icsk_af_ops->net_frag_header_len) {
1315 const struct dst_entry *dst = __sk_dst_get(sk);
1317 if (dst && dst_allfrag(dst))
1318 mtu += icsk->icsk_af_ops->net_frag_header_len;
1323 /* MTU probing init per socket */
1324 void tcp_mtup_init(struct sock *sk)
1326 struct tcp_sock *tp = tcp_sk(sk);
1327 struct inet_connection_sock *icsk = inet_csk(sk);
1329 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1330 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1331 icsk->icsk_af_ops->net_header_len;
1332 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1333 icsk->icsk_mtup.probe_size = 0;
1335 EXPORT_SYMBOL(tcp_mtup_init);
1337 /* This function synchronize snd mss to current pmtu/exthdr set.
1339 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1340 for TCP options, but includes only bare TCP header.
1342 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1343 It is minimum of user_mss and mss received with SYN.
1344 It also does not include TCP options.
1346 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1348 tp->mss_cache is current effective sending mss, including
1349 all tcp options except for SACKs. It is evaluated,
1350 taking into account current pmtu, but never exceeds
1351 tp->rx_opt.mss_clamp.
1353 NOTE1. rfc1122 clearly states that advertised MSS
1354 DOES NOT include either tcp or ip options.
1356 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1357 are READ ONLY outside this function. --ANK (980731)
1359 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1361 struct tcp_sock *tp = tcp_sk(sk);
1362 struct inet_connection_sock *icsk = inet_csk(sk);
1365 if (icsk->icsk_mtup.search_high > pmtu)
1366 icsk->icsk_mtup.search_high = pmtu;
1368 mss_now = tcp_mtu_to_mss(sk, pmtu);
1369 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1371 /* And store cached results */
1372 icsk->icsk_pmtu_cookie = pmtu;
1373 if (icsk->icsk_mtup.enabled)
1374 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1375 tp->mss_cache = mss_now;
1379 EXPORT_SYMBOL(tcp_sync_mss);
1381 /* Compute the current effective MSS, taking SACKs and IP options,
1382 * and even PMTU discovery events into account.
1384 unsigned int tcp_current_mss(struct sock *sk)
1386 const struct tcp_sock *tp = tcp_sk(sk);
1387 const struct dst_entry *dst = __sk_dst_get(sk);
1389 unsigned int header_len;
1390 struct tcp_out_options opts;
1391 struct tcp_md5sig_key *md5;
1393 mss_now = tp->mss_cache;
1396 u32 mtu = dst_mtu(dst);
1397 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1398 mss_now = tcp_sync_mss(sk, mtu);
1401 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1402 sizeof(struct tcphdr);
1403 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1404 * some common options. If this is an odd packet (because we have SACK
1405 * blocks etc) then our calculated header_len will be different, and
1406 * we have to adjust mss_now correspondingly */
1407 if (header_len != tp->tcp_header_len) {
1408 int delta = (int) header_len - tp->tcp_header_len;
1415 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1416 * As additional protections, we do not touch cwnd in retransmission phases,
1417 * and if application hit its sndbuf limit recently.
1419 static void tcp_cwnd_application_limited(struct sock *sk)
1421 struct tcp_sock *tp = tcp_sk(sk);
1423 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1424 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1425 /* Limited by application or receiver window. */
1426 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1427 u32 win_used = max(tp->snd_cwnd_used, init_win);
1428 if (win_used < tp->snd_cwnd) {
1429 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1430 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1432 tp->snd_cwnd_used = 0;
1434 tp->snd_cwnd_stamp = tcp_time_stamp;
1437 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1439 struct tcp_sock *tp = tcp_sk(sk);
1441 /* Track the maximum number of outstanding packets in each
1442 * window, and remember whether we were cwnd-limited then.
1444 if (!before(tp->snd_una, tp->max_packets_seq) ||
1445 tp->packets_out > tp->max_packets_out) {
1446 tp->max_packets_out = tp->packets_out;
1447 tp->max_packets_seq = tp->snd_nxt;
1448 tp->is_cwnd_limited = is_cwnd_limited;
1451 if (tcp_is_cwnd_limited(sk)) {
1452 /* Network is feed fully. */
1453 tp->snd_cwnd_used = 0;
1454 tp->snd_cwnd_stamp = tcp_time_stamp;
1456 /* Network starves. */
1457 if (tp->packets_out > tp->snd_cwnd_used)
1458 tp->snd_cwnd_used = tp->packets_out;
1460 if (sysctl_tcp_slow_start_after_idle &&
1461 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1462 tcp_cwnd_application_limited(sk);
1466 /* Minshall's variant of the Nagle send check. */
1467 static bool tcp_minshall_check(const struct tcp_sock *tp)
1469 return after(tp->snd_sml, tp->snd_una) &&
1470 !after(tp->snd_sml, tp->snd_nxt);
1473 /* Update snd_sml if this skb is under mss
1474 * Note that a TSO packet might end with a sub-mss segment
1475 * The test is really :
1476 * if ((skb->len % mss) != 0)
1477 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1478 * But we can avoid doing the divide again given we already have
1479 * skb_pcount = skb->len / mss_now
1481 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1482 const struct sk_buff *skb)
1484 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1485 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1488 /* Return false, if packet can be sent now without violation Nagle's rules:
1489 * 1. It is full sized. (provided by caller in %partial bool)
1490 * 2. Or it contains FIN. (already checked by caller)
1491 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1492 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1493 * With Minshall's modification: all sent small packets are ACKed.
1495 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1499 ((nonagle & TCP_NAGLE_CORK) ||
1500 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1502 /* Returns the portion of skb which can be sent right away */
1503 static unsigned int tcp_mss_split_point(const struct sock *sk,
1504 const struct sk_buff *skb,
1505 unsigned int mss_now,
1506 unsigned int max_segs,
1509 const struct tcp_sock *tp = tcp_sk(sk);
1510 u32 partial, needed, window, max_len;
1512 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1513 max_len = mss_now * max_segs;
1515 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1518 needed = min(skb->len, window);
1520 if (max_len <= needed)
1523 partial = needed % mss_now;
1524 /* If last segment is not a full MSS, check if Nagle rules allow us
1525 * to include this last segment in this skb.
1526 * Otherwise, we'll split the skb at last MSS boundary
1528 if (tcp_nagle_check(partial != 0, tp, nonagle))
1529 return needed - partial;
1534 /* Can at least one segment of SKB be sent right now, according to the
1535 * congestion window rules? If so, return how many segments are allowed.
1537 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1538 const struct sk_buff *skb)
1540 u32 in_flight, cwnd;
1542 /* Don't be strict about the congestion window for the final FIN. */
1543 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1544 tcp_skb_pcount(skb) == 1)
1547 in_flight = tcp_packets_in_flight(tp);
1548 cwnd = tp->snd_cwnd;
1549 if (in_flight < cwnd)
1550 return (cwnd - in_flight);
1555 /* Initialize TSO state of a skb.
1556 * This must be invoked the first time we consider transmitting
1557 * SKB onto the wire.
1559 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1560 unsigned int mss_now)
1562 int tso_segs = tcp_skb_pcount(skb);
1564 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1565 tcp_set_skb_tso_segs(sk, skb, mss_now);
1566 tso_segs = tcp_skb_pcount(skb);
1572 /* Return true if the Nagle test allows this packet to be
1575 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1576 unsigned int cur_mss, int nonagle)
1578 /* Nagle rule does not apply to frames, which sit in the middle of the
1579 * write_queue (they have no chances to get new data).
1581 * This is implemented in the callers, where they modify the 'nonagle'
1582 * argument based upon the location of SKB in the send queue.
1584 if (nonagle & TCP_NAGLE_PUSH)
1587 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1588 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1591 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1597 /* Does at least the first segment of SKB fit into the send window? */
1598 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1599 const struct sk_buff *skb,
1600 unsigned int cur_mss)
1602 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1604 if (skb->len > cur_mss)
1605 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1607 return !after(end_seq, tcp_wnd_end(tp));
1610 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1611 * should be put on the wire right now. If so, it returns the number of
1612 * packets allowed by the congestion window.
1614 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1615 unsigned int cur_mss, int nonagle)
1617 const struct tcp_sock *tp = tcp_sk(sk);
1618 unsigned int cwnd_quota;
1620 tcp_init_tso_segs(sk, skb, cur_mss);
1622 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1625 cwnd_quota = tcp_cwnd_test(tp, skb);
1626 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1632 /* Test if sending is allowed right now. */
1633 bool tcp_may_send_now(struct sock *sk)
1635 const struct tcp_sock *tp = tcp_sk(sk);
1636 struct sk_buff *skb = tcp_send_head(sk);
1639 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1640 (tcp_skb_is_last(sk, skb) ?
1641 tp->nonagle : TCP_NAGLE_PUSH));
1644 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1645 * which is put after SKB on the list. It is very much like
1646 * tcp_fragment() except that it may make several kinds of assumptions
1647 * in order to speed up the splitting operation. In particular, we
1648 * know that all the data is in scatter-gather pages, and that the
1649 * packet has never been sent out before (and thus is not cloned).
1651 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1652 unsigned int mss_now, gfp_t gfp)
1654 struct sk_buff *buff;
1655 int nlen = skb->len - len;
1658 /* All of a TSO frame must be composed of paged data. */
1659 if (skb->len != skb->data_len)
1660 return tcp_fragment(sk, skb, len, mss_now, gfp);
1662 buff = sk_stream_alloc_skb(sk, 0, gfp);
1663 if (unlikely(buff == NULL))
1666 sk->sk_wmem_queued += buff->truesize;
1667 sk_mem_charge(sk, buff->truesize);
1668 buff->truesize += nlen;
1669 skb->truesize -= nlen;
1671 /* Correct the sequence numbers. */
1672 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1673 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1674 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1676 /* PSH and FIN should only be set in the second packet. */
1677 flags = TCP_SKB_CB(skb)->tcp_flags;
1678 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1679 TCP_SKB_CB(buff)->tcp_flags = flags;
1681 /* This packet was never sent out yet, so no SACK bits. */
1682 TCP_SKB_CB(buff)->sacked = 0;
1684 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1685 skb_split(skb, buff, len);
1686 tcp_fragment_tstamp(skb, buff);
1688 /* Fix up tso_factor for both original and new SKB. */
1689 tcp_set_skb_tso_segs(sk, skb, mss_now);
1690 tcp_set_skb_tso_segs(sk, buff, mss_now);
1692 /* Link BUFF into the send queue. */
1693 __skb_header_release(buff);
1694 tcp_insert_write_queue_after(skb, buff, sk);
1699 /* Try to defer sending, if possible, in order to minimize the amount
1700 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1702 * This algorithm is from John Heffner.
1704 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1705 bool *is_cwnd_limited)
1707 struct tcp_sock *tp = tcp_sk(sk);
1708 const struct inet_connection_sock *icsk = inet_csk(sk);
1709 u32 send_win, cong_win, limit, in_flight;
1712 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1715 if (icsk->icsk_ca_state != TCP_CA_Open)
1718 /* Defer for less than two clock ticks. */
1719 if (tp->tso_deferred &&
1720 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1723 in_flight = tcp_packets_in_flight(tp);
1725 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1727 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1729 /* From in_flight test above, we know that cwnd > in_flight. */
1730 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1732 limit = min(send_win, cong_win);
1734 /* If a full-sized TSO skb can be sent, do it. */
1735 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1736 tp->xmit_size_goal_segs * tp->mss_cache))
1739 /* Middle in queue won't get any more data, full sendable already? */
1740 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1743 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1745 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1747 /* If at least some fraction of a window is available,
1750 chunk /= win_divisor;
1754 /* Different approach, try not to defer past a single
1755 * ACK. Receiver should ACK every other full sized
1756 * frame, so if we have space for more than 3 frames
1759 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1763 /* Ok, it looks like it is advisable to defer.
1764 * Do not rearm the timer if already set to not break TCP ACK clocking.
1766 if (!tp->tso_deferred)
1767 tp->tso_deferred = 1 | (jiffies << 1);
1769 if (cong_win < send_win && cong_win < skb->len)
1770 *is_cwnd_limited = true;
1775 tp->tso_deferred = 0;
1779 /* Create a new MTU probe if we are ready.
1780 * MTU probe is regularly attempting to increase the path MTU by
1781 * deliberately sending larger packets. This discovers routing
1782 * changes resulting in larger path MTUs.
1784 * Returns 0 if we should wait to probe (no cwnd available),
1785 * 1 if a probe was sent,
1788 static int tcp_mtu_probe(struct sock *sk)
1790 struct tcp_sock *tp = tcp_sk(sk);
1791 struct inet_connection_sock *icsk = inet_csk(sk);
1792 struct sk_buff *skb, *nskb, *next;
1799 /* Not currently probing/verifying,
1801 * have enough cwnd, and
1802 * not SACKing (the variable headers throw things off) */
1803 if (!icsk->icsk_mtup.enabled ||
1804 icsk->icsk_mtup.probe_size ||
1805 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1806 tp->snd_cwnd < 11 ||
1807 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1810 /* Very simple search strategy: just double the MSS. */
1811 mss_now = tcp_current_mss(sk);
1812 probe_size = 2 * tp->mss_cache;
1813 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1814 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1815 /* TODO: set timer for probe_converge_event */
1819 /* Have enough data in the send queue to probe? */
1820 if (tp->write_seq - tp->snd_nxt < size_needed)
1823 if (tp->snd_wnd < size_needed)
1825 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1828 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1829 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1830 if (!tcp_packets_in_flight(tp))
1836 /* We're allowed to probe. Build it now. */
1837 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1839 sk->sk_wmem_queued += nskb->truesize;
1840 sk_mem_charge(sk, nskb->truesize);
1842 skb = tcp_send_head(sk);
1844 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1845 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1846 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1847 TCP_SKB_CB(nskb)->sacked = 0;
1849 nskb->ip_summed = skb->ip_summed;
1851 tcp_insert_write_queue_before(nskb, skb, sk);
1854 tcp_for_write_queue_from_safe(skb, next, sk) {
1855 copy = min_t(int, skb->len, probe_size - len);
1856 if (nskb->ip_summed)
1857 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1859 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1860 skb_put(nskb, copy),
1863 if (skb->len <= copy) {
1864 /* We've eaten all the data from this skb.
1866 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1867 tcp_unlink_write_queue(skb, sk);
1868 sk_wmem_free_skb(sk, skb);
1870 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1871 ~(TCPHDR_FIN|TCPHDR_PSH);
1872 if (!skb_shinfo(skb)->nr_frags) {
1873 skb_pull(skb, copy);
1874 if (skb->ip_summed != CHECKSUM_PARTIAL)
1875 skb->csum = csum_partial(skb->data,
1878 __pskb_trim_head(skb, copy);
1879 tcp_set_skb_tso_segs(sk, skb, mss_now);
1881 TCP_SKB_CB(skb)->seq += copy;
1886 if (len >= probe_size)
1889 tcp_init_tso_segs(sk, nskb, nskb->len);
1891 /* We're ready to send. If this fails, the probe will
1892 * be resegmented into mss-sized pieces by tcp_write_xmit().
1894 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1895 /* Decrement cwnd here because we are sending
1896 * effectively two packets. */
1898 tcp_event_new_data_sent(sk, nskb);
1900 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1901 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1902 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1910 /* This routine writes packets to the network. It advances the
1911 * send_head. This happens as incoming acks open up the remote
1914 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1915 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1916 * account rare use of URG, this is not a big flaw.
1918 * Send at most one packet when push_one > 0. Temporarily ignore
1919 * cwnd limit to force at most one packet out when push_one == 2.
1921 * Returns true, if no segments are in flight and we have queued segments,
1922 * but cannot send anything now because of SWS or another problem.
1924 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1925 int push_one, gfp_t gfp)
1927 struct tcp_sock *tp = tcp_sk(sk);
1928 struct sk_buff *skb;
1929 unsigned int tso_segs, sent_pkts;
1932 bool is_cwnd_limited = false;
1937 /* Do MTU probing. */
1938 result = tcp_mtu_probe(sk);
1941 } else if (result > 0) {
1946 while ((skb = tcp_send_head(sk))) {
1949 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1952 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
1953 /* "skb_mstamp" is used as a start point for the retransmit timer */
1954 skb_mstamp_get(&skb->skb_mstamp);
1955 goto repair; /* Skip network transmission */
1958 cwnd_quota = tcp_cwnd_test(tp, skb);
1960 is_cwnd_limited = true;
1962 /* Force out a loss probe pkt. */
1968 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1971 if (tso_segs == 1) {
1972 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1973 (tcp_skb_is_last(sk, skb) ?
1974 nonagle : TCP_NAGLE_PUSH))))
1978 tcp_tso_should_defer(sk, skb, &is_cwnd_limited))
1982 /* TCP Small Queues :
1983 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1985 * - better RTT estimation and ACK scheduling
1988 * Alas, some drivers / subsystems require a fair amount
1989 * of queued bytes to ensure line rate.
1990 * One example is wifi aggregation (802.11 AMPDU)
1992 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1993 sk->sk_pacing_rate >> 10);
1995 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1996 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1997 /* It is possible TX completion already happened
1998 * before we set TSQ_THROTTLED, so we must
1999 * test again the condition.
2001 smp_mb__after_atomic();
2002 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2007 if (tso_segs > 1 && !tcp_urg_mode(tp))
2008 limit = tcp_mss_split_point(sk, skb, mss_now,
2011 sk->sk_gso_max_segs),
2014 if (skb->len > limit &&
2015 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2018 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2022 /* Advance the send_head. This one is sent out.
2023 * This call will increment packets_out.
2025 tcp_event_new_data_sent(sk, skb);
2027 tcp_minshall_update(tp, mss_now, skb);
2028 sent_pkts += tcp_skb_pcount(skb);
2034 if (likely(sent_pkts)) {
2035 if (tcp_in_cwnd_reduction(sk))
2036 tp->prr_out += sent_pkts;
2038 /* Send one loss probe per tail loss episode. */
2040 tcp_schedule_loss_probe(sk);
2041 tcp_cwnd_validate(sk, is_cwnd_limited);
2044 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2047 bool tcp_schedule_loss_probe(struct sock *sk)
2049 struct inet_connection_sock *icsk = inet_csk(sk);
2050 struct tcp_sock *tp = tcp_sk(sk);
2051 u32 timeout, tlp_time_stamp, rto_time_stamp;
2052 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2054 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2056 /* No consecutive loss probes. */
2057 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2061 /* Don't do any loss probe on a Fast Open connection before 3WHS
2064 if (sk->sk_state == TCP_SYN_RECV)
2067 /* TLP is only scheduled when next timer event is RTO. */
2068 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2071 /* Schedule a loss probe in 2*RTT for SACK capable connections
2072 * in Open state, that are either limited by cwnd or application.
2074 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2075 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2078 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2082 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2083 * for delayed ack when there's one outstanding packet.
2086 if (tp->packets_out == 1)
2087 timeout = max_t(u32, timeout,
2088 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2089 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2091 /* If RTO is shorter, just schedule TLP in its place. */
2092 tlp_time_stamp = tcp_time_stamp + timeout;
2093 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2094 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2095 s32 delta = rto_time_stamp - tcp_time_stamp;
2100 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2105 /* Thanks to skb fast clones, we can detect if a prior transmit of
2106 * a packet is still in a qdisc or driver queue.
2107 * In this case, there is very little point doing a retransmit !
2108 * Note: This is called from BH context only.
2110 static bool skb_still_in_host_queue(const struct sock *sk,
2111 const struct sk_buff *skb)
2113 if (unlikely(skb_fclone_busy(skb))) {
2114 NET_INC_STATS_BH(sock_net(sk),
2115 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2121 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2122 * retransmit the last segment.
2124 void tcp_send_loss_probe(struct sock *sk)
2126 struct tcp_sock *tp = tcp_sk(sk);
2127 struct sk_buff *skb;
2129 int mss = tcp_current_mss(sk);
2132 if (tcp_send_head(sk) != NULL) {
2133 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2137 /* At most one outstanding TLP retransmission. */
2138 if (tp->tlp_high_seq)
2141 /* Retransmit last segment. */
2142 skb = tcp_write_queue_tail(sk);
2146 if (skb_still_in_host_queue(sk, skb))
2149 pcount = tcp_skb_pcount(skb);
2150 if (WARN_ON(!pcount))
2153 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2154 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2157 skb = tcp_write_queue_tail(sk);
2160 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2163 err = __tcp_retransmit_skb(sk, skb);
2165 /* Record snd_nxt for loss detection. */
2167 tp->tlp_high_seq = tp->snd_nxt;
2170 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2171 inet_csk(sk)->icsk_rto,
2175 NET_INC_STATS_BH(sock_net(sk),
2176 LINUX_MIB_TCPLOSSPROBES);
2179 /* Push out any pending frames which were held back due to
2180 * TCP_CORK or attempt at coalescing tiny packets.
2181 * The socket must be locked by the caller.
2183 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2186 /* If we are closed, the bytes will have to remain here.
2187 * In time closedown will finish, we empty the write queue and
2188 * all will be happy.
2190 if (unlikely(sk->sk_state == TCP_CLOSE))
2193 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2194 sk_gfp_atomic(sk, GFP_ATOMIC)))
2195 tcp_check_probe_timer(sk);
2198 /* Send _single_ skb sitting at the send head. This function requires
2199 * true push pending frames to setup probe timer etc.
2201 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2203 struct sk_buff *skb = tcp_send_head(sk);
2205 BUG_ON(!skb || skb->len < mss_now);
2207 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2210 /* This function returns the amount that we can raise the
2211 * usable window based on the following constraints
2213 * 1. The window can never be shrunk once it is offered (RFC 793)
2214 * 2. We limit memory per socket
2217 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2218 * RECV.NEXT + RCV.WIN fixed until:
2219 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2221 * i.e. don't raise the right edge of the window until you can raise
2222 * it at least MSS bytes.
2224 * Unfortunately, the recommended algorithm breaks header prediction,
2225 * since header prediction assumes th->window stays fixed.
2227 * Strictly speaking, keeping th->window fixed violates the receiver
2228 * side SWS prevention criteria. The problem is that under this rule
2229 * a stream of single byte packets will cause the right side of the
2230 * window to always advance by a single byte.
2232 * Of course, if the sender implements sender side SWS prevention
2233 * then this will not be a problem.
2235 * BSD seems to make the following compromise:
2237 * If the free space is less than the 1/4 of the maximum
2238 * space available and the free space is less than 1/2 mss,
2239 * then set the window to 0.
2240 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2241 * Otherwise, just prevent the window from shrinking
2242 * and from being larger than the largest representable value.
2244 * This prevents incremental opening of the window in the regime
2245 * where TCP is limited by the speed of the reader side taking
2246 * data out of the TCP receive queue. It does nothing about
2247 * those cases where the window is constrained on the sender side
2248 * because the pipeline is full.
2250 * BSD also seems to "accidentally" limit itself to windows that are a
2251 * multiple of MSS, at least until the free space gets quite small.
2252 * This would appear to be a side effect of the mbuf implementation.
2253 * Combining these two algorithms results in the observed behavior
2254 * of having a fixed window size at almost all times.
2256 * Below we obtain similar behavior by forcing the offered window to
2257 * a multiple of the mss when it is feasible to do so.
2259 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2260 * Regular options like TIMESTAMP are taken into account.
2262 u32 __tcp_select_window(struct sock *sk)
2264 struct inet_connection_sock *icsk = inet_csk(sk);
2265 struct tcp_sock *tp = tcp_sk(sk);
2266 /* MSS for the peer's data. Previous versions used mss_clamp
2267 * here. I don't know if the value based on our guesses
2268 * of peer's MSS is better for the performance. It's more correct
2269 * but may be worse for the performance because of rcv_mss
2270 * fluctuations. --SAW 1998/11/1
2272 int mss = icsk->icsk_ack.rcv_mss;
2273 int free_space = tcp_space(sk);
2274 int allowed_space = tcp_full_space(sk);
2275 int full_space = min_t(int, tp->window_clamp, allowed_space);
2278 if (mss > full_space)
2281 if (free_space < (full_space >> 1)) {
2282 icsk->icsk_ack.quick = 0;
2284 if (sk_under_memory_pressure(sk))
2285 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2288 /* free_space might become our new window, make sure we don't
2289 * increase it due to wscale.
2291 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2293 /* if free space is less than mss estimate, or is below 1/16th
2294 * of the maximum allowed, try to move to zero-window, else
2295 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2296 * new incoming data is dropped due to memory limits.
2297 * With large window, mss test triggers way too late in order
2298 * to announce zero window in time before rmem limit kicks in.
2300 if (free_space < (allowed_space >> 4) || free_space < mss)
2304 if (free_space > tp->rcv_ssthresh)
2305 free_space = tp->rcv_ssthresh;
2307 /* Don't do rounding if we are using window scaling, since the
2308 * scaled window will not line up with the MSS boundary anyway.
2310 window = tp->rcv_wnd;
2311 if (tp->rx_opt.rcv_wscale) {
2312 window = free_space;
2314 /* Advertise enough space so that it won't get scaled away.
2315 * Import case: prevent zero window announcement if
2316 * 1<<rcv_wscale > mss.
2318 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2319 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2320 << tp->rx_opt.rcv_wscale);
2322 /* Get the largest window that is a nice multiple of mss.
2323 * Window clamp already applied above.
2324 * If our current window offering is within 1 mss of the
2325 * free space we just keep it. This prevents the divide
2326 * and multiply from happening most of the time.
2327 * We also don't do any window rounding when the free space
2330 if (window <= free_space - mss || window > free_space)
2331 window = (free_space / mss) * mss;
2332 else if (mss == full_space &&
2333 free_space > window + (full_space >> 1))
2334 window = free_space;
2340 /* Collapses two adjacent SKB's during retransmission. */
2341 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2343 struct tcp_sock *tp = tcp_sk(sk);
2344 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2345 int skb_size, next_skb_size;
2347 skb_size = skb->len;
2348 next_skb_size = next_skb->len;
2350 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2352 tcp_highest_sack_combine(sk, next_skb, skb);
2354 tcp_unlink_write_queue(next_skb, sk);
2356 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2359 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2360 skb->ip_summed = CHECKSUM_PARTIAL;
2362 if (skb->ip_summed != CHECKSUM_PARTIAL)
2363 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2365 /* Update sequence range on original skb. */
2366 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2368 /* Merge over control information. This moves PSH/FIN etc. over */
2369 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2371 /* All done, get rid of second SKB and account for it so
2372 * packet counting does not break.
2374 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2376 /* changed transmit queue under us so clear hints */
2377 tcp_clear_retrans_hints_partial(tp);
2378 if (next_skb == tp->retransmit_skb_hint)
2379 tp->retransmit_skb_hint = skb;
2381 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2383 sk_wmem_free_skb(sk, next_skb);
2386 /* Check if coalescing SKBs is legal. */
2387 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2389 if (tcp_skb_pcount(skb) > 1)
2391 /* TODO: SACK collapsing could be used to remove this condition */
2392 if (skb_shinfo(skb)->nr_frags != 0)
2394 if (skb_cloned(skb))
2396 if (skb == tcp_send_head(sk))
2398 /* Some heurestics for collapsing over SACK'd could be invented */
2399 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2405 /* Collapse packets in the retransmit queue to make to create
2406 * less packets on the wire. This is only done on retransmission.
2408 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2411 struct tcp_sock *tp = tcp_sk(sk);
2412 struct sk_buff *skb = to, *tmp;
2415 if (!sysctl_tcp_retrans_collapse)
2417 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2420 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2421 if (!tcp_can_collapse(sk, skb))
2433 /* Punt if not enough space exists in the first SKB for
2434 * the data in the second
2436 if (skb->len > skb_availroom(to))
2439 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2442 tcp_collapse_retrans(sk, to);
2446 /* This retransmits one SKB. Policy decisions and retransmit queue
2447 * state updates are done by the caller. Returns non-zero if an
2448 * error occurred which prevented the send.
2450 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2452 struct tcp_sock *tp = tcp_sk(sk);
2453 struct inet_connection_sock *icsk = inet_csk(sk);
2454 unsigned int cur_mss;
2457 /* Inconslusive MTU probe */
2458 if (icsk->icsk_mtup.probe_size) {
2459 icsk->icsk_mtup.probe_size = 0;
2462 /* Do not sent more than we queued. 1/4 is reserved for possible
2463 * copying overhead: fragmentation, tunneling, mangling etc.
2465 if (atomic_read(&sk->sk_wmem_alloc) >
2466 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2469 if (skb_still_in_host_queue(sk, skb))
2472 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2473 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2475 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2479 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2480 return -EHOSTUNREACH; /* Routing failure or similar. */
2482 cur_mss = tcp_current_mss(sk);
2484 /* If receiver has shrunk his window, and skb is out of
2485 * new window, do not retransmit it. The exception is the
2486 * case, when window is shrunk to zero. In this case
2487 * our retransmit serves as a zero window probe.
2489 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2490 TCP_SKB_CB(skb)->seq != tp->snd_una)
2493 if (skb->len > cur_mss) {
2494 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2495 return -ENOMEM; /* We'll try again later. */
2497 int oldpcount = tcp_skb_pcount(skb);
2499 if (unlikely(oldpcount > 1)) {
2500 if (skb_unclone(skb, GFP_ATOMIC))
2502 tcp_init_tso_segs(sk, skb, cur_mss);
2503 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2507 tcp_retrans_try_collapse(sk, skb, cur_mss);
2509 /* Make a copy, if the first transmission SKB clone we made
2510 * is still in somebody's hands, else make a clone.
2513 /* make sure skb->data is aligned on arches that require it
2514 * and check if ack-trimming & collapsing extended the headroom
2515 * beyond what csum_start can cover.
2517 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2518 skb_headroom(skb) >= 0xFFFF)) {
2519 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2521 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2524 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2528 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2529 /* Update global TCP statistics. */
2530 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2531 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2532 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2533 tp->total_retrans++;
2538 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2540 struct tcp_sock *tp = tcp_sk(sk);
2541 int err = __tcp_retransmit_skb(sk, skb);
2544 #if FASTRETRANS_DEBUG > 0
2545 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2546 net_dbg_ratelimited("retrans_out leaked\n");
2549 if (!tp->retrans_out)
2550 tp->lost_retrans_low = tp->snd_nxt;
2551 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2552 tp->retrans_out += tcp_skb_pcount(skb);
2554 /* Save stamp of the first retransmit. */
2555 if (!tp->retrans_stamp)
2556 tp->retrans_stamp = tcp_skb_timestamp(skb);
2558 /* snd_nxt is stored to detect loss of retransmitted segment,
2559 * see tcp_input.c tcp_sacktag_write_queue().
2561 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2562 } else if (err != -EBUSY) {
2563 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2566 if (tp->undo_retrans < 0)
2567 tp->undo_retrans = 0;
2568 tp->undo_retrans += tcp_skb_pcount(skb);
2572 /* Check if we forward retransmits are possible in the current
2573 * window/congestion state.
2575 static bool tcp_can_forward_retransmit(struct sock *sk)
2577 const struct inet_connection_sock *icsk = inet_csk(sk);
2578 const struct tcp_sock *tp = tcp_sk(sk);
2580 /* Forward retransmissions are possible only during Recovery. */
2581 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2584 /* No forward retransmissions in Reno are possible. */
2585 if (tcp_is_reno(tp))
2588 /* Yeah, we have to make difficult choice between forward transmission
2589 * and retransmission... Both ways have their merits...
2591 * For now we do not retransmit anything, while we have some new
2592 * segments to send. In the other cases, follow rule 3 for
2593 * NextSeg() specified in RFC3517.
2596 if (tcp_may_send_now(sk))
2602 /* This gets called after a retransmit timeout, and the initially
2603 * retransmitted data is acknowledged. It tries to continue
2604 * resending the rest of the retransmit queue, until either
2605 * we've sent it all or the congestion window limit is reached.
2606 * If doing SACK, the first ACK which comes back for a timeout
2607 * based retransmit packet might feed us FACK information again.
2608 * If so, we use it to avoid unnecessarily retransmissions.
2610 void tcp_xmit_retransmit_queue(struct sock *sk)
2612 const struct inet_connection_sock *icsk = inet_csk(sk);
2613 struct tcp_sock *tp = tcp_sk(sk);
2614 struct sk_buff *skb;
2615 struct sk_buff *hole = NULL;
2618 int fwd_rexmitting = 0;
2620 if (!tp->packets_out)
2624 tp->retransmit_high = tp->snd_una;
2626 if (tp->retransmit_skb_hint) {
2627 skb = tp->retransmit_skb_hint;
2628 last_lost = TCP_SKB_CB(skb)->end_seq;
2629 if (after(last_lost, tp->retransmit_high))
2630 last_lost = tp->retransmit_high;
2632 skb = tcp_write_queue_head(sk);
2633 last_lost = tp->snd_una;
2636 tcp_for_write_queue_from(skb, sk) {
2637 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2639 if (skb == tcp_send_head(sk))
2641 /* we could do better than to assign each time */
2643 tp->retransmit_skb_hint = skb;
2645 /* Assume this retransmit will generate
2646 * only one packet for congestion window
2647 * calculation purposes. This works because
2648 * tcp_retransmit_skb() will chop up the
2649 * packet to be MSS sized and all the
2650 * packet counting works out.
2652 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2655 if (fwd_rexmitting) {
2657 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2659 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2661 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2662 tp->retransmit_high = last_lost;
2663 if (!tcp_can_forward_retransmit(sk))
2665 /* Backtrack if necessary to non-L'ed skb */
2673 } else if (!(sacked & TCPCB_LOST)) {
2674 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2679 last_lost = TCP_SKB_CB(skb)->end_seq;
2680 if (icsk->icsk_ca_state != TCP_CA_Loss)
2681 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2683 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2686 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2689 if (tcp_retransmit_skb(sk, skb))
2692 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2694 if (tcp_in_cwnd_reduction(sk))
2695 tp->prr_out += tcp_skb_pcount(skb);
2697 if (skb == tcp_write_queue_head(sk))
2698 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2699 inet_csk(sk)->icsk_rto,
2704 /* Send a fin. The caller locks the socket for us. This cannot be
2705 * allowed to fail queueing a FIN frame under any circumstances.
2707 void tcp_send_fin(struct sock *sk)
2709 struct tcp_sock *tp = tcp_sk(sk);
2710 struct sk_buff *skb = tcp_write_queue_tail(sk);
2713 /* Optimization, tack on the FIN if we have a queue of
2714 * unsent frames. But be careful about outgoing SACKS
2717 mss_now = tcp_current_mss(sk);
2719 if (tcp_send_head(sk) != NULL) {
2720 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2721 TCP_SKB_CB(skb)->end_seq++;
2724 /* Socket is locked, keep trying until memory is available. */
2726 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2733 /* Reserve space for headers and prepare control bits. */
2734 skb_reserve(skb, MAX_TCP_HEADER);
2735 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2736 tcp_init_nondata_skb(skb, tp->write_seq,
2737 TCPHDR_ACK | TCPHDR_FIN);
2738 tcp_queue_skb(sk, skb);
2740 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2743 /* We get here when a process closes a file descriptor (either due to
2744 * an explicit close() or as a byproduct of exit()'ing) and there
2745 * was unread data in the receive queue. This behavior is recommended
2746 * by RFC 2525, section 2.17. -DaveM
2748 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2750 struct sk_buff *skb;
2752 /* NOTE: No TCP options attached and we never retransmit this. */
2753 skb = alloc_skb(MAX_TCP_HEADER, priority);
2755 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2759 /* Reserve space for headers and prepare control bits. */
2760 skb_reserve(skb, MAX_TCP_HEADER);
2761 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2762 TCPHDR_ACK | TCPHDR_RST);
2764 if (tcp_transmit_skb(sk, skb, 0, priority))
2765 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2767 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2770 /* Send a crossed SYN-ACK during socket establishment.
2771 * WARNING: This routine must only be called when we have already sent
2772 * a SYN packet that crossed the incoming SYN that caused this routine
2773 * to get called. If this assumption fails then the initial rcv_wnd
2774 * and rcv_wscale values will not be correct.
2776 int tcp_send_synack(struct sock *sk)
2778 struct sk_buff *skb;
2780 skb = tcp_write_queue_head(sk);
2781 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2782 pr_debug("%s: wrong queue state\n", __func__);
2785 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2786 if (skb_cloned(skb)) {
2787 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2790 tcp_unlink_write_queue(skb, sk);
2791 __skb_header_release(nskb);
2792 __tcp_add_write_queue_head(sk, nskb);
2793 sk_wmem_free_skb(sk, skb);
2794 sk->sk_wmem_queued += nskb->truesize;
2795 sk_mem_charge(sk, nskb->truesize);
2799 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2800 tcp_ecn_send_synack(sk, skb);
2802 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2806 * tcp_make_synack - Prepare a SYN-ACK.
2807 * sk: listener socket
2808 * dst: dst entry attached to the SYNACK
2809 * req: request_sock pointer
2811 * Allocate one skb and build a SYNACK packet.
2812 * @dst is consumed : Caller should not use it again.
2814 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2815 struct request_sock *req,
2816 struct tcp_fastopen_cookie *foc)
2818 struct tcp_out_options opts;
2819 struct inet_request_sock *ireq = inet_rsk(req);
2820 struct tcp_sock *tp = tcp_sk(sk);
2822 struct sk_buff *skb;
2823 struct tcp_md5sig_key *md5;
2824 int tcp_header_size;
2827 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2828 if (unlikely(!skb)) {
2832 /* Reserve space for headers. */
2833 skb_reserve(skb, MAX_TCP_HEADER);
2835 skb_dst_set(skb, dst);
2836 security_skb_owned_by(skb, sk);
2838 mss = dst_metric_advmss(dst);
2839 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2840 mss = tp->rx_opt.user_mss;
2842 memset(&opts, 0, sizeof(opts));
2843 #ifdef CONFIG_SYN_COOKIES
2844 if (unlikely(req->cookie_ts))
2845 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2848 skb_mstamp_get(&skb->skb_mstamp);
2849 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2852 skb_push(skb, tcp_header_size);
2853 skb_reset_transport_header(skb);
2856 memset(th, 0, sizeof(struct tcphdr));
2859 tcp_ecn_make_synack(req, th, sk);
2860 th->source = htons(ireq->ir_num);
2861 th->dest = ireq->ir_rmt_port;
2862 /* Setting of flags are superfluous here for callers (and ECE is
2863 * not even correctly set)
2865 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2866 TCPHDR_SYN | TCPHDR_ACK);
2868 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2869 /* XXX data is queued and acked as is. No buffer/window check */
2870 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2872 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2873 th->window = htons(min(req->rcv_wnd, 65535U));
2874 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2875 th->doff = (tcp_header_size >> 2);
2876 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2878 #ifdef CONFIG_TCP_MD5SIG
2879 /* Okay, we have all we need - do the md5 hash if needed */
2881 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2882 md5, NULL, req, skb);
2888 EXPORT_SYMBOL(tcp_make_synack);
2890 /* Do all connect socket setups that can be done AF independent. */
2891 static void tcp_connect_init(struct sock *sk)
2893 const struct dst_entry *dst = __sk_dst_get(sk);
2894 struct tcp_sock *tp = tcp_sk(sk);
2897 /* We'll fix this up when we get a response from the other end.
2898 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2900 tp->tcp_header_len = sizeof(struct tcphdr) +
2901 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2903 #ifdef CONFIG_TCP_MD5SIG
2904 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2905 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2908 /* If user gave his TCP_MAXSEG, record it to clamp */
2909 if (tp->rx_opt.user_mss)
2910 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2913 tcp_sync_mss(sk, dst_mtu(dst));
2915 if (!tp->window_clamp)
2916 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2917 tp->advmss = dst_metric_advmss(dst);
2918 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2919 tp->advmss = tp->rx_opt.user_mss;
2921 tcp_initialize_rcv_mss(sk);
2923 /* limit the window selection if the user enforce a smaller rx buffer */
2924 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2925 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2926 tp->window_clamp = tcp_full_space(sk);
2928 tcp_select_initial_window(tcp_full_space(sk),
2929 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2932 sysctl_tcp_window_scaling,
2934 dst_metric(dst, RTAX_INITRWND));
2936 tp->rx_opt.rcv_wscale = rcv_wscale;
2937 tp->rcv_ssthresh = tp->rcv_wnd;
2940 sock_reset_flag(sk, SOCK_DONE);
2943 tp->snd_una = tp->write_seq;
2944 tp->snd_sml = tp->write_seq;
2945 tp->snd_up = tp->write_seq;
2946 tp->snd_nxt = tp->write_seq;
2948 if (likely(!tp->repair))
2951 tp->rcv_tstamp = tcp_time_stamp;
2952 tp->rcv_wup = tp->rcv_nxt;
2953 tp->copied_seq = tp->rcv_nxt;
2955 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2956 inet_csk(sk)->icsk_retransmits = 0;
2957 tcp_clear_retrans(tp);
2960 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2962 struct tcp_sock *tp = tcp_sk(sk);
2963 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2965 tcb->end_seq += skb->len;
2966 __skb_header_release(skb);
2967 __tcp_add_write_queue_tail(sk, skb);
2968 sk->sk_wmem_queued += skb->truesize;
2969 sk_mem_charge(sk, skb->truesize);
2970 tp->write_seq = tcb->end_seq;
2971 tp->packets_out += tcp_skb_pcount(skb);
2974 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2975 * queue a data-only packet after the regular SYN, such that regular SYNs
2976 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2977 * only the SYN sequence, the data are retransmitted in the first ACK.
2978 * If cookie is not cached or other error occurs, falls back to send a
2979 * regular SYN with Fast Open cookie request option.
2981 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2983 struct tcp_sock *tp = tcp_sk(sk);
2984 struct tcp_fastopen_request *fo = tp->fastopen_req;
2985 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2986 struct sk_buff *syn_data = NULL, *data;
2987 unsigned long last_syn_loss = 0;
2989 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2990 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2991 &syn_loss, &last_syn_loss);
2992 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2994 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2995 fo->cookie.len = -1;
2999 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3000 fo->cookie.len = -1;
3001 else if (fo->cookie.len <= 0)
3004 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3005 * user-MSS. Reserve maximum option space for middleboxes that add
3006 * private TCP options. The cost is reduced data space in SYN :(
3008 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3009 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3010 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3011 MAX_TCP_OPTION_SPACE;
3013 space = min_t(size_t, space, fo->size);
3015 /* limit to order-0 allocations */
3016 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3018 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
3020 if (syn_data == NULL)
3023 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3024 struct iovec *iov = &fo->data->msg_iov[i];
3025 unsigned char __user *from = iov->iov_base;
3026 int len = iov->iov_len;
3028 if (syn_data->len + len > space)
3029 len = space - syn_data->len;
3030 else if (i + 1 == iovlen)
3031 /* No more data pending in inet_wait_for_connect() */
3034 if (skb_add_data(syn_data, from, len))
3038 /* Queue a data-only packet after the regular SYN for retransmission */
3039 data = pskb_copy(syn_data, sk->sk_allocation);
3042 TCP_SKB_CB(data)->seq++;
3043 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3044 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3045 tcp_connect_queue_skb(sk, data);
3046 fo->copied = data->len;
3048 /* syn_data is about to be sent, we need to take current time stamps
3049 * for the packets that are in write queue : SYN packet and DATA
3051 skb_mstamp_get(&syn->skb_mstamp);
3052 data->skb_mstamp = syn->skb_mstamp;
3054 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
3055 tp->syn_data = (fo->copied > 0);
3056 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3062 /* Send a regular SYN with Fast Open cookie request option */
3063 if (fo->cookie.len > 0)
3065 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3067 tp->syn_fastopen = 0;
3068 kfree_skb(syn_data);
3070 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3074 /* Build a SYN and send it off. */
3075 int tcp_connect(struct sock *sk)
3077 struct tcp_sock *tp = tcp_sk(sk);
3078 struct sk_buff *buff;
3081 tcp_connect_init(sk);
3083 if (unlikely(tp->repair)) {
3084 tcp_finish_connect(sk, NULL);
3088 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3089 if (unlikely(buff == NULL))
3092 /* Reserve space for headers. */
3093 skb_reserve(buff, MAX_TCP_HEADER);
3095 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3096 tp->retrans_stamp = tcp_time_stamp;
3097 tcp_connect_queue_skb(sk, buff);
3098 tcp_ecn_send_syn(sk, buff);
3100 /* Send off SYN; include data in Fast Open. */
3101 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3102 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3103 if (err == -ECONNREFUSED)
3106 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3107 * in order to make this packet get counted in tcpOutSegs.
3109 tp->snd_nxt = tp->write_seq;
3110 tp->pushed_seq = tp->write_seq;
3111 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3113 /* Timer for repeating the SYN until an answer. */
3114 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3115 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3118 EXPORT_SYMBOL(tcp_connect);
3120 /* Send out a delayed ack, the caller does the policy checking
3121 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3124 void tcp_send_delayed_ack(struct sock *sk)
3126 struct inet_connection_sock *icsk = inet_csk(sk);
3127 int ato = icsk->icsk_ack.ato;
3128 unsigned long timeout;
3130 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3132 if (ato > TCP_DELACK_MIN) {
3133 const struct tcp_sock *tp = tcp_sk(sk);
3134 int max_ato = HZ / 2;
3136 if (icsk->icsk_ack.pingpong ||
3137 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3138 max_ato = TCP_DELACK_MAX;
3140 /* Slow path, intersegment interval is "high". */
3142 /* If some rtt estimate is known, use it to bound delayed ack.
3143 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3147 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3154 ato = min(ato, max_ato);
3157 /* Stay within the limit we were given */
3158 timeout = jiffies + ato;
3160 /* Use new timeout only if there wasn't a older one earlier. */
3161 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3162 /* If delack timer was blocked or is about to expire,
3165 if (icsk->icsk_ack.blocked ||
3166 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3171 if (!time_before(timeout, icsk->icsk_ack.timeout))
3172 timeout = icsk->icsk_ack.timeout;
3174 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3175 icsk->icsk_ack.timeout = timeout;
3176 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3179 /* This routine sends an ack and also updates the window. */
3180 void tcp_send_ack(struct sock *sk)
3182 struct sk_buff *buff;
3184 /* If we have been reset, we may not send again. */
3185 if (sk->sk_state == TCP_CLOSE)
3188 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3190 /* We are not putting this on the write queue, so
3191 * tcp_transmit_skb() will set the ownership to this
3194 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3196 inet_csk_schedule_ack(sk);
3197 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3198 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3199 TCP_DELACK_MAX, TCP_RTO_MAX);
3203 /* Reserve space for headers and prepare control bits. */
3204 skb_reserve(buff, MAX_TCP_HEADER);
3205 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3207 /* Send it off, this clears delayed acks for us. */
3208 skb_mstamp_get(&buff->skb_mstamp);
3209 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3211 EXPORT_SYMBOL_GPL(tcp_send_ack);
3213 /* This routine sends a packet with an out of date sequence
3214 * number. It assumes the other end will try to ack it.
3216 * Question: what should we make while urgent mode?
3217 * 4.4BSD forces sending single byte of data. We cannot send
3218 * out of window data, because we have SND.NXT==SND.MAX...
3220 * Current solution: to send TWO zero-length segments in urgent mode:
3221 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3222 * out-of-date with SND.UNA-1 to probe window.
3224 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3226 struct tcp_sock *tp = tcp_sk(sk);
3227 struct sk_buff *skb;
3229 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3230 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3234 /* Reserve space for headers and set control bits. */
3235 skb_reserve(skb, MAX_TCP_HEADER);
3236 /* Use a previous sequence. This should cause the other
3237 * end to send an ack. Don't queue or clone SKB, just
3240 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3241 skb_mstamp_get(&skb->skb_mstamp);
3242 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3245 void tcp_send_window_probe(struct sock *sk)
3247 if (sk->sk_state == TCP_ESTABLISHED) {
3248 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3249 tcp_xmit_probe_skb(sk, 0);
3253 /* Initiate keepalive or window probe from timer. */
3254 int tcp_write_wakeup(struct sock *sk)
3256 struct tcp_sock *tp = tcp_sk(sk);
3257 struct sk_buff *skb;
3259 if (sk->sk_state == TCP_CLOSE)
3262 if ((skb = tcp_send_head(sk)) != NULL &&
3263 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3265 unsigned int mss = tcp_current_mss(sk);
3266 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3268 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3269 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3271 /* We are probing the opening of a window
3272 * but the window size is != 0
3273 * must have been a result SWS avoidance ( sender )
3275 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3277 seg_size = min(seg_size, mss);
3278 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3279 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3281 } else if (!tcp_skb_pcount(skb))
3282 tcp_set_skb_tso_segs(sk, skb, mss);
3284 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3285 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3287 tcp_event_new_data_sent(sk, skb);
3290 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3291 tcp_xmit_probe_skb(sk, 1);
3292 return tcp_xmit_probe_skb(sk, 0);
3296 /* A window probe timeout has occurred. If window is not closed send
3297 * a partial packet else a zero probe.
3299 void tcp_send_probe0(struct sock *sk)
3301 struct inet_connection_sock *icsk = inet_csk(sk);
3302 struct tcp_sock *tp = tcp_sk(sk);
3303 unsigned long probe_max;
3306 err = tcp_write_wakeup(sk);
3308 if (tp->packets_out || !tcp_send_head(sk)) {
3309 /* Cancel probe timer, if it is not required. */
3310 icsk->icsk_probes_out = 0;
3311 icsk->icsk_backoff = 0;
3316 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3317 icsk->icsk_backoff++;
3318 icsk->icsk_probes_out++;
3319 probe_max = TCP_RTO_MAX;
3321 /* If packet was not sent due to local congestion,
3322 * do not backoff and do not remember icsk_probes_out.
3323 * Let local senders to fight for local resources.
3325 * Use accumulated backoff yet.
3327 if (!icsk->icsk_probes_out)
3328 icsk->icsk_probes_out = 1;
3329 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3331 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3332 inet_csk_rto_backoff(icsk, probe_max),
3336 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3338 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3342 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3344 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3345 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3349 EXPORT_SYMBOL(tcp_rtx_synack);