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);
336 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
337 tcp_ca_needs_ecn(sk);
340 const struct dst_entry *dst = __sk_dst_get(sk);
342 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
349 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
350 tp->ecn_flags = TCP_ECN_OK;
351 if (tcp_ca_needs_ecn(sk))
357 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th,
360 if (inet_rsk(req)->ecn_ok) {
362 if (tcp_ca_needs_ecn(sk))
367 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
370 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
373 struct tcp_sock *tp = tcp_sk(sk);
375 if (tp->ecn_flags & TCP_ECN_OK) {
376 /* Not-retransmitted data segment: set ECT and inject CWR. */
377 if (skb->len != tcp_header_len &&
378 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
380 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
381 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
382 tcp_hdr(skb)->cwr = 1;
383 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
385 } else if (!tcp_ca_needs_ecn(sk)) {
386 /* ACK or retransmitted segment: clear ECT|CE */
387 INET_ECN_dontxmit(sk);
389 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
390 tcp_hdr(skb)->ece = 1;
394 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
395 * auto increment end seqno.
397 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
399 struct skb_shared_info *shinfo = skb_shinfo(skb);
401 skb->ip_summed = CHECKSUM_PARTIAL;
404 TCP_SKB_CB(skb)->tcp_flags = flags;
405 TCP_SKB_CB(skb)->sacked = 0;
407 tcp_skb_pcount_set(skb, 1);
408 shinfo->gso_size = 0;
409 shinfo->gso_type = 0;
411 TCP_SKB_CB(skb)->seq = seq;
412 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
414 TCP_SKB_CB(skb)->end_seq = seq;
417 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
419 return tp->snd_una != tp->snd_up;
422 #define OPTION_SACK_ADVERTISE (1 << 0)
423 #define OPTION_TS (1 << 1)
424 #define OPTION_MD5 (1 << 2)
425 #define OPTION_WSCALE (1 << 3)
426 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
428 struct tcp_out_options {
429 u16 options; /* bit field of OPTION_* */
430 u16 mss; /* 0 to disable */
431 u8 ws; /* window scale, 0 to disable */
432 u8 num_sack_blocks; /* number of SACK blocks to include */
433 u8 hash_size; /* bytes in hash_location */
434 __u8 *hash_location; /* temporary pointer, overloaded */
435 __u32 tsval, tsecr; /* need to include OPTION_TS */
436 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
439 /* Write previously computed TCP options to the packet.
441 * Beware: Something in the Internet is very sensitive to the ordering of
442 * TCP options, we learned this through the hard way, so be careful here.
443 * Luckily we can at least blame others for their non-compliance but from
444 * inter-operability perspective it seems that we're somewhat stuck with
445 * the ordering which we have been using if we want to keep working with
446 * those broken things (not that it currently hurts anybody as there isn't
447 * particular reason why the ordering would need to be changed).
449 * At least SACK_PERM as the first option is known to lead to a disaster
450 * (but it may well be that other scenarios fail similarly).
452 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
453 struct tcp_out_options *opts)
455 u16 options = opts->options; /* mungable copy */
457 if (unlikely(OPTION_MD5 & options)) {
458 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
459 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
460 /* overload cookie hash location */
461 opts->hash_location = (__u8 *)ptr;
465 if (unlikely(opts->mss)) {
466 *ptr++ = htonl((TCPOPT_MSS << 24) |
467 (TCPOLEN_MSS << 16) |
471 if (likely(OPTION_TS & options)) {
472 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
473 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
474 (TCPOLEN_SACK_PERM << 16) |
475 (TCPOPT_TIMESTAMP << 8) |
477 options &= ~OPTION_SACK_ADVERTISE;
479 *ptr++ = htonl((TCPOPT_NOP << 24) |
481 (TCPOPT_TIMESTAMP << 8) |
484 *ptr++ = htonl(opts->tsval);
485 *ptr++ = htonl(opts->tsecr);
488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
491 (TCPOPT_SACK_PERM << 8) |
495 if (unlikely(OPTION_WSCALE & options)) {
496 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_WINDOW << 16) |
498 (TCPOLEN_WINDOW << 8) |
502 if (unlikely(opts->num_sack_blocks)) {
503 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
504 tp->duplicate_sack : tp->selective_acks;
507 *ptr++ = htonl((TCPOPT_NOP << 24) |
510 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
511 TCPOLEN_SACK_PERBLOCK)));
513 for (this_sack = 0; this_sack < opts->num_sack_blocks;
515 *ptr++ = htonl(sp[this_sack].start_seq);
516 *ptr++ = htonl(sp[this_sack].end_seq);
519 tp->rx_opt.dsack = 0;
522 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
523 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
525 *ptr++ = htonl((TCPOPT_EXP << 24) |
526 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
527 TCPOPT_FASTOPEN_MAGIC);
529 memcpy(ptr, foc->val, foc->len);
530 if ((foc->len & 3) == 2) {
531 u8 *align = ((u8 *)ptr) + foc->len;
532 align[0] = align[1] = TCPOPT_NOP;
534 ptr += (foc->len + 3) >> 2;
538 /* Compute TCP options for SYN packets. This is not the final
539 * network wire format yet.
541 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
542 struct tcp_out_options *opts,
543 struct tcp_md5sig_key **md5)
545 struct tcp_sock *tp = tcp_sk(sk);
546 unsigned int remaining = MAX_TCP_OPTION_SPACE;
547 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
549 #ifdef CONFIG_TCP_MD5SIG
550 *md5 = tp->af_specific->md5_lookup(sk, sk);
552 opts->options |= OPTION_MD5;
553 remaining -= TCPOLEN_MD5SIG_ALIGNED;
559 /* We always get an MSS option. The option bytes which will be seen in
560 * normal data packets should timestamps be used, must be in the MSS
561 * advertised. But we subtract them from tp->mss_cache so that
562 * calculations in tcp_sendmsg are simpler etc. So account for this
563 * fact here if necessary. If we don't do this correctly, as a
564 * receiver we won't recognize data packets as being full sized when we
565 * should, and thus we won't abide by the delayed ACK rules correctly.
566 * SACKs don't matter, we never delay an ACK when we have any of those
568 opts->mss = tcp_advertise_mss(sk);
569 remaining -= TCPOLEN_MSS_ALIGNED;
571 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
572 opts->options |= OPTION_TS;
573 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
574 opts->tsecr = tp->rx_opt.ts_recent;
575 remaining -= TCPOLEN_TSTAMP_ALIGNED;
577 if (likely(sysctl_tcp_window_scaling)) {
578 opts->ws = tp->rx_opt.rcv_wscale;
579 opts->options |= OPTION_WSCALE;
580 remaining -= TCPOLEN_WSCALE_ALIGNED;
582 if (likely(sysctl_tcp_sack)) {
583 opts->options |= OPTION_SACK_ADVERTISE;
584 if (unlikely(!(OPTION_TS & opts->options)))
585 remaining -= TCPOLEN_SACKPERM_ALIGNED;
588 if (fastopen && fastopen->cookie.len >= 0) {
589 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
590 need = (need + 3) & ~3U; /* Align to 32 bits */
591 if (remaining >= need) {
592 opts->options |= OPTION_FAST_OPEN_COOKIE;
593 opts->fastopen_cookie = &fastopen->cookie;
595 tp->syn_fastopen = 1;
599 return MAX_TCP_OPTION_SPACE - remaining;
602 /* Set up TCP options for SYN-ACKs. */
603 static unsigned int tcp_synack_options(struct sock *sk,
604 struct request_sock *req,
605 unsigned int mss, struct sk_buff *skb,
606 struct tcp_out_options *opts,
607 struct tcp_md5sig_key **md5,
608 struct tcp_fastopen_cookie *foc)
610 struct inet_request_sock *ireq = inet_rsk(req);
611 unsigned int remaining = MAX_TCP_OPTION_SPACE;
613 #ifdef CONFIG_TCP_MD5SIG
614 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
616 opts->options |= OPTION_MD5;
617 remaining -= TCPOLEN_MD5SIG_ALIGNED;
619 /* We can't fit any SACK blocks in a packet with MD5 + TS
620 * options. There was discussion about disabling SACK
621 * rather than TS in order to fit in better with old,
622 * buggy kernels, but that was deemed to be unnecessary.
624 ireq->tstamp_ok &= !ireq->sack_ok;
630 /* We always send an MSS option. */
632 remaining -= TCPOLEN_MSS_ALIGNED;
634 if (likely(ireq->wscale_ok)) {
635 opts->ws = ireq->rcv_wscale;
636 opts->options |= OPTION_WSCALE;
637 remaining -= TCPOLEN_WSCALE_ALIGNED;
639 if (likely(ireq->tstamp_ok)) {
640 opts->options |= OPTION_TS;
641 opts->tsval = tcp_skb_timestamp(skb);
642 opts->tsecr = req->ts_recent;
643 remaining -= TCPOLEN_TSTAMP_ALIGNED;
645 if (likely(ireq->sack_ok)) {
646 opts->options |= OPTION_SACK_ADVERTISE;
647 if (unlikely(!ireq->tstamp_ok))
648 remaining -= TCPOLEN_SACKPERM_ALIGNED;
650 if (foc != NULL && foc->len >= 0) {
651 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
652 need = (need + 3) & ~3U; /* Align to 32 bits */
653 if (remaining >= need) {
654 opts->options |= OPTION_FAST_OPEN_COOKIE;
655 opts->fastopen_cookie = foc;
660 return MAX_TCP_OPTION_SPACE - remaining;
663 /* Compute TCP options for ESTABLISHED sockets. This is not the
664 * final wire format yet.
666 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
667 struct tcp_out_options *opts,
668 struct tcp_md5sig_key **md5)
670 struct tcp_sock *tp = tcp_sk(sk);
671 unsigned int size = 0;
672 unsigned int eff_sacks;
676 #ifdef CONFIG_TCP_MD5SIG
677 *md5 = tp->af_specific->md5_lookup(sk, sk);
678 if (unlikely(*md5)) {
679 opts->options |= OPTION_MD5;
680 size += TCPOLEN_MD5SIG_ALIGNED;
686 if (likely(tp->rx_opt.tstamp_ok)) {
687 opts->options |= OPTION_TS;
688 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
689 opts->tsecr = tp->rx_opt.ts_recent;
690 size += TCPOLEN_TSTAMP_ALIGNED;
693 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
694 if (unlikely(eff_sacks)) {
695 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
696 opts->num_sack_blocks =
697 min_t(unsigned int, eff_sacks,
698 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
699 TCPOLEN_SACK_PERBLOCK);
700 size += TCPOLEN_SACK_BASE_ALIGNED +
701 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
708 /* TCP SMALL QUEUES (TSQ)
710 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
711 * to reduce RTT and bufferbloat.
712 * We do this using a special skb destructor (tcp_wfree).
714 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
715 * needs to be reallocated in a driver.
716 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
718 * Since transmit from skb destructor is forbidden, we use a tasklet
719 * to process all sockets that eventually need to send more skbs.
720 * We use one tasklet per cpu, with its own queue of sockets.
723 struct tasklet_struct tasklet;
724 struct list_head head; /* queue of tcp sockets */
726 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
728 static void tcp_tsq_handler(struct sock *sk)
730 if ((1 << sk->sk_state) &
731 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
732 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
733 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
737 * One tasklet per cpu tries to send more skbs.
738 * We run in tasklet context but need to disable irqs when
739 * transferring tsq->head because tcp_wfree() might
740 * interrupt us (non NAPI drivers)
742 static void tcp_tasklet_func(unsigned long data)
744 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
747 struct list_head *q, *n;
751 local_irq_save(flags);
752 list_splice_init(&tsq->head, &list);
753 local_irq_restore(flags);
755 list_for_each_safe(q, n, &list) {
756 tp = list_entry(q, struct tcp_sock, tsq_node);
757 list_del(&tp->tsq_node);
759 sk = (struct sock *)tp;
762 if (!sock_owned_by_user(sk)) {
765 /* defer the work to tcp_release_cb() */
766 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
770 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
775 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
776 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
777 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
778 (1UL << TCP_MTU_REDUCED_DEFERRED))
780 * tcp_release_cb - tcp release_sock() callback
783 * called from release_sock() to perform protocol dependent
784 * actions before socket release.
786 void tcp_release_cb(struct sock *sk)
788 struct tcp_sock *tp = tcp_sk(sk);
789 unsigned long flags, nflags;
791 /* perform an atomic operation only if at least one flag is set */
793 flags = tp->tsq_flags;
794 if (!(flags & TCP_DEFERRED_ALL))
796 nflags = flags & ~TCP_DEFERRED_ALL;
797 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
799 if (flags & (1UL << TCP_TSQ_DEFERRED))
802 /* Here begins the tricky part :
803 * We are called from release_sock() with :
805 * 2) sk_lock.slock spinlock held
806 * 3) socket owned by us (sk->sk_lock.owned == 1)
808 * But following code is meant to be called from BH handlers,
809 * so we should keep BH disabled, but early release socket ownership
811 sock_release_ownership(sk);
813 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
814 tcp_write_timer_handler(sk);
817 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
818 tcp_delack_timer_handler(sk);
821 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
822 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
826 EXPORT_SYMBOL(tcp_release_cb);
828 void __init tcp_tasklet_init(void)
832 for_each_possible_cpu(i) {
833 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
835 INIT_LIST_HEAD(&tsq->head);
836 tasklet_init(&tsq->tasklet,
843 * Write buffer destructor automatically called from kfree_skb.
844 * We can't xmit new skbs from this context, as we might already
847 void tcp_wfree(struct sk_buff *skb)
849 struct sock *sk = skb->sk;
850 struct tcp_sock *tp = tcp_sk(sk);
853 /* Keep one reference on sk_wmem_alloc.
854 * Will be released by sk_free() from here or tcp_tasklet_func()
856 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
858 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
859 * Wait until our queues (qdisc + devices) are drained.
861 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
862 * - chance for incoming ACK (processed by another cpu maybe)
863 * to migrate this flow (skb->ooo_okay will be eventually set)
865 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
868 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
869 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
871 struct tsq_tasklet *tsq;
873 /* queue this socket to tasklet queue */
874 local_irq_save(flags);
875 tsq = this_cpu_ptr(&tsq_tasklet);
876 list_add(&tp->tsq_node, &tsq->head);
877 tasklet_schedule(&tsq->tasklet);
878 local_irq_restore(flags);
885 /* This routine actually transmits TCP packets queued in by
886 * tcp_do_sendmsg(). This is used by both the initial
887 * transmission and possible later retransmissions.
888 * All SKB's seen here are completely headerless. It is our
889 * job to build the TCP header, and pass the packet down to
890 * IP so it can do the same plus pass the packet off to the
893 * We are working here with either a clone of the original
894 * SKB, or a fresh unique copy made by the retransmit engine.
896 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
899 const struct inet_connection_sock *icsk = inet_csk(sk);
900 struct inet_sock *inet;
902 struct tcp_skb_cb *tcb;
903 struct tcp_out_options opts;
904 unsigned int tcp_options_size, tcp_header_size;
905 struct tcp_md5sig_key *md5;
909 BUG_ON(!skb || !tcp_skb_pcount(skb));
912 skb_mstamp_get(&skb->skb_mstamp);
914 if (unlikely(skb_cloned(skb)))
915 skb = pskb_copy(skb, gfp_mask);
917 skb = skb_clone(skb, gfp_mask);
924 tcb = TCP_SKB_CB(skb);
925 memset(&opts, 0, sizeof(opts));
927 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
928 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
930 tcp_options_size = tcp_established_options(sk, skb, &opts,
932 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
934 if (tcp_packets_in_flight(tp) == 0)
935 tcp_ca_event(sk, CA_EVENT_TX_START);
937 /* if no packet is in qdisc/device queue, then allow XPS to select
938 * another queue. We can be called from tcp_tsq_handler()
939 * which holds one reference to sk_wmem_alloc.
941 * TODO: Ideally, in-flight pure ACK packets should not matter here.
942 * One way to get this would be to set skb->truesize = 2 on them.
944 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
946 skb_push(skb, tcp_header_size);
947 skb_reset_transport_header(skb);
951 skb->destructor = tcp_wfree;
952 skb_set_hash_from_sk(skb, sk);
953 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
955 /* Build TCP header and checksum it. */
957 th->source = inet->inet_sport;
958 th->dest = inet->inet_dport;
959 th->seq = htonl(tcb->seq);
960 th->ack_seq = htonl(tp->rcv_nxt);
961 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
964 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
965 /* RFC1323: The window in SYN & SYN/ACK segments
968 th->window = htons(min(tp->rcv_wnd, 65535U));
970 th->window = htons(tcp_select_window(sk));
975 /* The urg_mode check is necessary during a below snd_una win probe */
976 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
977 if (before(tp->snd_up, tcb->seq + 0x10000)) {
978 th->urg_ptr = htons(tp->snd_up - tcb->seq);
980 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
981 th->urg_ptr = htons(0xFFFF);
986 tcp_options_write((__be32 *)(th + 1), tp, &opts);
987 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
988 tcp_ecn_send(sk, skb, tcp_header_size);
990 #ifdef CONFIG_TCP_MD5SIG
991 /* Calculate the MD5 hash, as we have all we need now */
993 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
994 tp->af_specific->calc_md5_hash(opts.hash_location,
999 icsk->icsk_af_ops->send_check(sk, skb);
1001 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1002 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1004 if (skb->len != tcp_header_size)
1005 tcp_event_data_sent(tp, sk);
1007 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1008 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1009 tcp_skb_pcount(skb));
1011 /* OK, its time to fill skb_shinfo(skb)->gso_segs */
1012 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1014 /* Our usage of tstamp should remain private */
1015 skb->tstamp.tv64 = 0;
1017 /* Cleanup our debris for IP stacks */
1018 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1019 sizeof(struct inet6_skb_parm)));
1021 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1023 if (likely(err <= 0))
1028 return net_xmit_eval(err);
1031 /* This routine just queues the buffer for sending.
1033 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1034 * otherwise socket can stall.
1036 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1038 struct tcp_sock *tp = tcp_sk(sk);
1040 /* Advance write_seq and place onto the write_queue. */
1041 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1042 __skb_header_release(skb);
1043 tcp_add_write_queue_tail(sk, skb);
1044 sk->sk_wmem_queued += skb->truesize;
1045 sk_mem_charge(sk, skb->truesize);
1048 /* Initialize TSO segments for a packet. */
1049 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1050 unsigned int mss_now)
1052 struct skb_shared_info *shinfo = skb_shinfo(skb);
1054 /* Make sure we own this skb before messing gso_size/gso_segs */
1055 WARN_ON_ONCE(skb_cloned(skb));
1057 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1058 /* Avoid the costly divide in the normal
1061 tcp_skb_pcount_set(skb, 1);
1062 shinfo->gso_size = 0;
1063 shinfo->gso_type = 0;
1065 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1066 shinfo->gso_size = mss_now;
1067 shinfo->gso_type = sk->sk_gso_type;
1071 /* When a modification to fackets out becomes necessary, we need to check
1072 * skb is counted to fackets_out or not.
1074 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1077 struct tcp_sock *tp = tcp_sk(sk);
1079 if (!tp->sacked_out || tcp_is_reno(tp))
1082 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1083 tp->fackets_out -= decr;
1086 /* Pcount in the middle of the write queue got changed, we need to do various
1087 * tweaks to fix counters
1089 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1091 struct tcp_sock *tp = tcp_sk(sk);
1093 tp->packets_out -= decr;
1095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1096 tp->sacked_out -= decr;
1097 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1098 tp->retrans_out -= decr;
1099 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1100 tp->lost_out -= decr;
1102 /* Reno case is special. Sigh... */
1103 if (tcp_is_reno(tp) && decr > 0)
1104 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1106 tcp_adjust_fackets_out(sk, skb, decr);
1108 if (tp->lost_skb_hint &&
1109 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1110 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1111 tp->lost_cnt_hint -= decr;
1113 tcp_verify_left_out(tp);
1116 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1118 struct skb_shared_info *shinfo = skb_shinfo(skb);
1120 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1121 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1122 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1123 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1125 shinfo->tx_flags &= ~tsflags;
1126 shinfo2->tx_flags |= tsflags;
1127 swap(shinfo->tskey, shinfo2->tskey);
1131 /* Function to create two new TCP segments. Shrinks the given segment
1132 * to the specified size and appends a new segment with the rest of the
1133 * packet to the list. This won't be called frequently, I hope.
1134 * Remember, these are still headerless SKBs at this point.
1136 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1137 unsigned int mss_now, gfp_t gfp)
1139 struct tcp_sock *tp = tcp_sk(sk);
1140 struct sk_buff *buff;
1141 int nsize, old_factor;
1145 if (WARN_ON(len > skb->len))
1148 nsize = skb_headlen(skb) - len;
1152 if (skb_unclone(skb, gfp))
1155 /* Get a new skb... force flag on. */
1156 buff = sk_stream_alloc_skb(sk, nsize, gfp);
1158 return -ENOMEM; /* We'll just try again later. */
1160 sk->sk_wmem_queued += buff->truesize;
1161 sk_mem_charge(sk, buff->truesize);
1162 nlen = skb->len - len - nsize;
1163 buff->truesize += nlen;
1164 skb->truesize -= nlen;
1166 /* Correct the sequence numbers. */
1167 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1168 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1169 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1171 /* PSH and FIN should only be set in the second packet. */
1172 flags = TCP_SKB_CB(skb)->tcp_flags;
1173 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1174 TCP_SKB_CB(buff)->tcp_flags = flags;
1175 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1177 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1178 /* Copy and checksum data tail into the new buffer. */
1179 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1180 skb_put(buff, nsize),
1185 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1187 skb->ip_summed = CHECKSUM_PARTIAL;
1188 skb_split(skb, buff, len);
1191 buff->ip_summed = skb->ip_summed;
1193 buff->tstamp = skb->tstamp;
1194 tcp_fragment_tstamp(skb, buff);
1196 old_factor = tcp_skb_pcount(skb);
1198 /* Fix up tso_factor for both original and new SKB. */
1199 tcp_set_skb_tso_segs(sk, skb, mss_now);
1200 tcp_set_skb_tso_segs(sk, buff, mss_now);
1202 /* If this packet has been sent out already, we must
1203 * adjust the various packet counters.
1205 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1206 int diff = old_factor - tcp_skb_pcount(skb) -
1207 tcp_skb_pcount(buff);
1210 tcp_adjust_pcount(sk, skb, diff);
1213 /* Link BUFF into the send queue. */
1214 __skb_header_release(buff);
1215 tcp_insert_write_queue_after(skb, buff, sk);
1220 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1221 * eventually). The difference is that pulled data not copied, but
1222 * immediately discarded.
1224 static void __pskb_trim_head(struct sk_buff *skb, int len)
1226 struct skb_shared_info *shinfo;
1229 eat = min_t(int, len, skb_headlen(skb));
1231 __skb_pull(skb, eat);
1238 shinfo = skb_shinfo(skb);
1239 for (i = 0; i < shinfo->nr_frags; i++) {
1240 int size = skb_frag_size(&shinfo->frags[i]);
1243 skb_frag_unref(skb, i);
1246 shinfo->frags[k] = shinfo->frags[i];
1248 shinfo->frags[k].page_offset += eat;
1249 skb_frag_size_sub(&shinfo->frags[k], eat);
1255 shinfo->nr_frags = k;
1257 skb_reset_tail_pointer(skb);
1258 skb->data_len -= len;
1259 skb->len = skb->data_len;
1262 /* Remove acked data from a packet in the transmit queue. */
1263 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1265 if (skb_unclone(skb, GFP_ATOMIC))
1268 __pskb_trim_head(skb, len);
1270 TCP_SKB_CB(skb)->seq += len;
1271 skb->ip_summed = CHECKSUM_PARTIAL;
1273 skb->truesize -= len;
1274 sk->sk_wmem_queued -= len;
1275 sk_mem_uncharge(sk, len);
1276 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1278 /* Any change of skb->len requires recalculation of tso factor. */
1279 if (tcp_skb_pcount(skb) > 1)
1280 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1285 /* Calculate MSS not accounting any TCP options. */
1286 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1288 const struct tcp_sock *tp = tcp_sk(sk);
1289 const struct inet_connection_sock *icsk = inet_csk(sk);
1292 /* Calculate base mss without TCP options:
1293 It is MMS_S - sizeof(tcphdr) of rfc1122
1295 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1297 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1298 if (icsk->icsk_af_ops->net_frag_header_len) {
1299 const struct dst_entry *dst = __sk_dst_get(sk);
1301 if (dst && dst_allfrag(dst))
1302 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1305 /* Clamp it (mss_clamp does not include tcp options) */
1306 if (mss_now > tp->rx_opt.mss_clamp)
1307 mss_now = tp->rx_opt.mss_clamp;
1309 /* Now subtract optional transport overhead */
1310 mss_now -= icsk->icsk_ext_hdr_len;
1312 /* Then reserve room for full set of TCP options and 8 bytes of data */
1318 /* Calculate MSS. Not accounting for SACKs here. */
1319 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1321 /* Subtract TCP options size, not including SACKs */
1322 return __tcp_mtu_to_mss(sk, pmtu) -
1323 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1326 /* Inverse of above */
1327 int tcp_mss_to_mtu(struct sock *sk, int mss)
1329 const struct tcp_sock *tp = tcp_sk(sk);
1330 const struct inet_connection_sock *icsk = inet_csk(sk);
1334 tp->tcp_header_len +
1335 icsk->icsk_ext_hdr_len +
1336 icsk->icsk_af_ops->net_header_len;
1338 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1339 if (icsk->icsk_af_ops->net_frag_header_len) {
1340 const struct dst_entry *dst = __sk_dst_get(sk);
1342 if (dst && dst_allfrag(dst))
1343 mtu += icsk->icsk_af_ops->net_frag_header_len;
1348 /* MTU probing init per socket */
1349 void tcp_mtup_init(struct sock *sk)
1351 struct tcp_sock *tp = tcp_sk(sk);
1352 struct inet_connection_sock *icsk = inet_csk(sk);
1354 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1355 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1356 icsk->icsk_af_ops->net_header_len;
1357 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1358 icsk->icsk_mtup.probe_size = 0;
1360 EXPORT_SYMBOL(tcp_mtup_init);
1362 /* This function synchronize snd mss to current pmtu/exthdr set.
1364 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1365 for TCP options, but includes only bare TCP header.
1367 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1368 It is minimum of user_mss and mss received with SYN.
1369 It also does not include TCP options.
1371 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1373 tp->mss_cache is current effective sending mss, including
1374 all tcp options except for SACKs. It is evaluated,
1375 taking into account current pmtu, but never exceeds
1376 tp->rx_opt.mss_clamp.
1378 NOTE1. rfc1122 clearly states that advertised MSS
1379 DOES NOT include either tcp or ip options.
1381 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1382 are READ ONLY outside this function. --ANK (980731)
1384 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386 struct tcp_sock *tp = tcp_sk(sk);
1387 struct inet_connection_sock *icsk = inet_csk(sk);
1390 if (icsk->icsk_mtup.search_high > pmtu)
1391 icsk->icsk_mtup.search_high = pmtu;
1393 mss_now = tcp_mtu_to_mss(sk, pmtu);
1394 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1396 /* And store cached results */
1397 icsk->icsk_pmtu_cookie = pmtu;
1398 if (icsk->icsk_mtup.enabled)
1399 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1400 tp->mss_cache = mss_now;
1404 EXPORT_SYMBOL(tcp_sync_mss);
1406 /* Compute the current effective MSS, taking SACKs and IP options,
1407 * and even PMTU discovery events into account.
1409 unsigned int tcp_current_mss(struct sock *sk)
1411 const struct tcp_sock *tp = tcp_sk(sk);
1412 const struct dst_entry *dst = __sk_dst_get(sk);
1414 unsigned int header_len;
1415 struct tcp_out_options opts;
1416 struct tcp_md5sig_key *md5;
1418 mss_now = tp->mss_cache;
1421 u32 mtu = dst_mtu(dst);
1422 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1423 mss_now = tcp_sync_mss(sk, mtu);
1426 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1427 sizeof(struct tcphdr);
1428 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1429 * some common options. If this is an odd packet (because we have SACK
1430 * blocks etc) then our calculated header_len will be different, and
1431 * we have to adjust mss_now correspondingly */
1432 if (header_len != tp->tcp_header_len) {
1433 int delta = (int) header_len - tp->tcp_header_len;
1440 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1441 * As additional protections, we do not touch cwnd in retransmission phases,
1442 * and if application hit its sndbuf limit recently.
1444 static void tcp_cwnd_application_limited(struct sock *sk)
1446 struct tcp_sock *tp = tcp_sk(sk);
1448 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1449 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1450 /* Limited by application or receiver window. */
1451 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1452 u32 win_used = max(tp->snd_cwnd_used, init_win);
1453 if (win_used < tp->snd_cwnd) {
1454 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1455 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457 tp->snd_cwnd_used = 0;
1459 tp->snd_cwnd_stamp = tcp_time_stamp;
1462 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1464 struct tcp_sock *tp = tcp_sk(sk);
1466 /* Track the maximum number of outstanding packets in each
1467 * window, and remember whether we were cwnd-limited then.
1469 if (!before(tp->snd_una, tp->max_packets_seq) ||
1470 tp->packets_out > tp->max_packets_out) {
1471 tp->max_packets_out = tp->packets_out;
1472 tp->max_packets_seq = tp->snd_nxt;
1473 tp->is_cwnd_limited = is_cwnd_limited;
1476 if (tcp_is_cwnd_limited(sk)) {
1477 /* Network is feed fully. */
1478 tp->snd_cwnd_used = 0;
1479 tp->snd_cwnd_stamp = tcp_time_stamp;
1481 /* Network starves. */
1482 if (tp->packets_out > tp->snd_cwnd_used)
1483 tp->snd_cwnd_used = tp->packets_out;
1485 if (sysctl_tcp_slow_start_after_idle &&
1486 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1487 tcp_cwnd_application_limited(sk);
1491 /* Minshall's variant of the Nagle send check. */
1492 static bool tcp_minshall_check(const struct tcp_sock *tp)
1494 return after(tp->snd_sml, tp->snd_una) &&
1495 !after(tp->snd_sml, tp->snd_nxt);
1498 /* Update snd_sml if this skb is under mss
1499 * Note that a TSO packet might end with a sub-mss segment
1500 * The test is really :
1501 * if ((skb->len % mss) != 0)
1502 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1503 * But we can avoid doing the divide again given we already have
1504 * skb_pcount = skb->len / mss_now
1506 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1507 const struct sk_buff *skb)
1509 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1510 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1513 /* Return false, if packet can be sent now without violation Nagle's rules:
1514 * 1. It is full sized. (provided by caller in %partial bool)
1515 * 2. Or it contains FIN. (already checked by caller)
1516 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1517 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1518 * With Minshall's modification: all sent small packets are ACKed.
1520 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1524 ((nonagle & TCP_NAGLE_CORK) ||
1525 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1527 /* Returns the portion of skb which can be sent right away */
1528 static unsigned int tcp_mss_split_point(const struct sock *sk,
1529 const struct sk_buff *skb,
1530 unsigned int mss_now,
1531 unsigned int max_segs,
1534 const struct tcp_sock *tp = tcp_sk(sk);
1535 u32 partial, needed, window, max_len;
1537 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1538 max_len = mss_now * max_segs;
1540 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1543 needed = min(skb->len, window);
1545 if (max_len <= needed)
1548 partial = needed % mss_now;
1549 /* If last segment is not a full MSS, check if Nagle rules allow us
1550 * to include this last segment in this skb.
1551 * Otherwise, we'll split the skb at last MSS boundary
1553 if (tcp_nagle_check(partial != 0, tp, nonagle))
1554 return needed - partial;
1559 /* Can at least one segment of SKB be sent right now, according to the
1560 * congestion window rules? If so, return how many segments are allowed.
1562 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1563 const struct sk_buff *skb)
1565 u32 in_flight, cwnd, halfcwnd;
1567 /* Don't be strict about the congestion window for the final FIN. */
1568 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1569 tcp_skb_pcount(skb) == 1)
1572 in_flight = tcp_packets_in_flight(tp);
1573 cwnd = tp->snd_cwnd;
1574 if (in_flight >= cwnd)
1577 /* For better scheduling, ensure we have at least
1578 * 2 GSO packets in flight.
1580 halfcwnd = max(cwnd >> 1, 1U);
1581 return min(halfcwnd, cwnd - in_flight);
1584 /* Initialize TSO state of a skb.
1585 * This must be invoked the first time we consider transmitting
1586 * SKB onto the wire.
1588 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1589 unsigned int mss_now)
1591 int tso_segs = tcp_skb_pcount(skb);
1593 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1594 tcp_set_skb_tso_segs(sk, skb, mss_now);
1595 tso_segs = tcp_skb_pcount(skb);
1601 /* Return true if the Nagle test allows this packet to be
1604 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1605 unsigned int cur_mss, int nonagle)
1607 /* Nagle rule does not apply to frames, which sit in the middle of the
1608 * write_queue (they have no chances to get new data).
1610 * This is implemented in the callers, where they modify the 'nonagle'
1611 * argument based upon the location of SKB in the send queue.
1613 if (nonagle & TCP_NAGLE_PUSH)
1616 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1617 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1620 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1626 /* Does at least the first segment of SKB fit into the send window? */
1627 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1628 const struct sk_buff *skb,
1629 unsigned int cur_mss)
1631 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1633 if (skb->len > cur_mss)
1634 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1636 return !after(end_seq, tcp_wnd_end(tp));
1639 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1640 * should be put on the wire right now. If so, it returns the number of
1641 * packets allowed by the congestion window.
1643 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1644 unsigned int cur_mss, int nonagle)
1646 const struct tcp_sock *tp = tcp_sk(sk);
1647 unsigned int cwnd_quota;
1649 tcp_init_tso_segs(sk, skb, cur_mss);
1651 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1654 cwnd_quota = tcp_cwnd_test(tp, skb);
1655 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1661 /* Test if sending is allowed right now. */
1662 bool tcp_may_send_now(struct sock *sk)
1664 const struct tcp_sock *tp = tcp_sk(sk);
1665 struct sk_buff *skb = tcp_send_head(sk);
1668 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1669 (tcp_skb_is_last(sk, skb) ?
1670 tp->nonagle : TCP_NAGLE_PUSH));
1673 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1674 * which is put after SKB on the list. It is very much like
1675 * tcp_fragment() except that it may make several kinds of assumptions
1676 * in order to speed up the splitting operation. In particular, we
1677 * know that all the data is in scatter-gather pages, and that the
1678 * packet has never been sent out before (and thus is not cloned).
1680 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1681 unsigned int mss_now, gfp_t gfp)
1683 struct sk_buff *buff;
1684 int nlen = skb->len - len;
1687 /* All of a TSO frame must be composed of paged data. */
1688 if (skb->len != skb->data_len)
1689 return tcp_fragment(sk, skb, len, mss_now, gfp);
1691 buff = sk_stream_alloc_skb(sk, 0, gfp);
1692 if (unlikely(buff == NULL))
1695 sk->sk_wmem_queued += buff->truesize;
1696 sk_mem_charge(sk, buff->truesize);
1697 buff->truesize += nlen;
1698 skb->truesize -= nlen;
1700 /* Correct the sequence numbers. */
1701 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1702 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1703 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1705 /* PSH and FIN should only be set in the second packet. */
1706 flags = TCP_SKB_CB(skb)->tcp_flags;
1707 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1708 TCP_SKB_CB(buff)->tcp_flags = flags;
1710 /* This packet was never sent out yet, so no SACK bits. */
1711 TCP_SKB_CB(buff)->sacked = 0;
1713 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1714 skb_split(skb, buff, len);
1715 tcp_fragment_tstamp(skb, buff);
1717 /* Fix up tso_factor for both original and new SKB. */
1718 tcp_set_skb_tso_segs(sk, skb, mss_now);
1719 tcp_set_skb_tso_segs(sk, buff, mss_now);
1721 /* Link BUFF into the send queue. */
1722 __skb_header_release(buff);
1723 tcp_insert_write_queue_after(skb, buff, sk);
1728 /* Try to defer sending, if possible, in order to minimize the amount
1729 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1731 * This algorithm is from John Heffner.
1733 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1734 bool *is_cwnd_limited)
1736 struct tcp_sock *tp = tcp_sk(sk);
1737 const struct inet_connection_sock *icsk = inet_csk(sk);
1738 u32 send_win, cong_win, limit, in_flight;
1741 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1744 if (icsk->icsk_ca_state != TCP_CA_Open)
1747 /* Defer for less than two clock ticks. */
1748 if (tp->tso_deferred &&
1749 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1752 in_flight = tcp_packets_in_flight(tp);
1754 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1756 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1758 /* From in_flight test above, we know that cwnd > in_flight. */
1759 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1761 limit = min(send_win, cong_win);
1763 /* If a full-sized TSO skb can be sent, do it. */
1764 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1765 tp->xmit_size_goal_segs * tp->mss_cache))
1768 /* Middle in queue won't get any more data, full sendable already? */
1769 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1772 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1774 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1776 /* If at least some fraction of a window is available,
1779 chunk /= win_divisor;
1783 /* Different approach, try not to defer past a single
1784 * ACK. Receiver should ACK every other full sized
1785 * frame, so if we have space for more than 3 frames
1788 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1792 /* Ok, it looks like it is advisable to defer.
1793 * Do not rearm the timer if already set to not break TCP ACK clocking.
1795 if (!tp->tso_deferred)
1796 tp->tso_deferred = 1 | (jiffies << 1);
1798 if (cong_win < send_win && cong_win < skb->len)
1799 *is_cwnd_limited = true;
1804 tp->tso_deferred = 0;
1808 /* Create a new MTU probe if we are ready.
1809 * MTU probe is regularly attempting to increase the path MTU by
1810 * deliberately sending larger packets. This discovers routing
1811 * changes resulting in larger path MTUs.
1813 * Returns 0 if we should wait to probe (no cwnd available),
1814 * 1 if a probe was sent,
1817 static int tcp_mtu_probe(struct sock *sk)
1819 struct tcp_sock *tp = tcp_sk(sk);
1820 struct inet_connection_sock *icsk = inet_csk(sk);
1821 struct sk_buff *skb, *nskb, *next;
1828 /* Not currently probing/verifying,
1830 * have enough cwnd, and
1831 * not SACKing (the variable headers throw things off) */
1832 if (!icsk->icsk_mtup.enabled ||
1833 icsk->icsk_mtup.probe_size ||
1834 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1835 tp->snd_cwnd < 11 ||
1836 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1839 /* Very simple search strategy: just double the MSS. */
1840 mss_now = tcp_current_mss(sk);
1841 probe_size = 2 * tp->mss_cache;
1842 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1843 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1844 /* TODO: set timer for probe_converge_event */
1848 /* Have enough data in the send queue to probe? */
1849 if (tp->write_seq - tp->snd_nxt < size_needed)
1852 if (tp->snd_wnd < size_needed)
1854 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1857 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1858 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1859 if (!tcp_packets_in_flight(tp))
1865 /* We're allowed to probe. Build it now. */
1866 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1868 sk->sk_wmem_queued += nskb->truesize;
1869 sk_mem_charge(sk, nskb->truesize);
1871 skb = tcp_send_head(sk);
1873 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1874 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1875 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1876 TCP_SKB_CB(nskb)->sacked = 0;
1878 nskb->ip_summed = skb->ip_summed;
1880 tcp_insert_write_queue_before(nskb, skb, sk);
1883 tcp_for_write_queue_from_safe(skb, next, sk) {
1884 copy = min_t(int, skb->len, probe_size - len);
1885 if (nskb->ip_summed)
1886 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1888 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1889 skb_put(nskb, copy),
1892 if (skb->len <= copy) {
1893 /* We've eaten all the data from this skb.
1895 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1896 tcp_unlink_write_queue(skb, sk);
1897 sk_wmem_free_skb(sk, skb);
1899 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1900 ~(TCPHDR_FIN|TCPHDR_PSH);
1901 if (!skb_shinfo(skb)->nr_frags) {
1902 skb_pull(skb, copy);
1903 if (skb->ip_summed != CHECKSUM_PARTIAL)
1904 skb->csum = csum_partial(skb->data,
1907 __pskb_trim_head(skb, copy);
1908 tcp_set_skb_tso_segs(sk, skb, mss_now);
1910 TCP_SKB_CB(skb)->seq += copy;
1915 if (len >= probe_size)
1918 tcp_init_tso_segs(sk, nskb, nskb->len);
1920 /* We're ready to send. If this fails, the probe will
1921 * be resegmented into mss-sized pieces by tcp_write_xmit().
1923 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1924 /* Decrement cwnd here because we are sending
1925 * effectively two packets. */
1927 tcp_event_new_data_sent(sk, nskb);
1929 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1930 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1931 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1939 /* This routine writes packets to the network. It advances the
1940 * send_head. This happens as incoming acks open up the remote
1943 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1944 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1945 * account rare use of URG, this is not a big flaw.
1947 * Send at most one packet when push_one > 0. Temporarily ignore
1948 * cwnd limit to force at most one packet out when push_one == 2.
1950 * Returns true, if no segments are in flight and we have queued segments,
1951 * but cannot send anything now because of SWS or another problem.
1953 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1954 int push_one, gfp_t gfp)
1956 struct tcp_sock *tp = tcp_sk(sk);
1957 struct sk_buff *skb;
1958 unsigned int tso_segs, sent_pkts;
1961 bool is_cwnd_limited = false;
1966 /* Do MTU probing. */
1967 result = tcp_mtu_probe(sk);
1970 } else if (result > 0) {
1975 while ((skb = tcp_send_head(sk))) {
1978 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1981 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
1982 /* "skb_mstamp" is used as a start point for the retransmit timer */
1983 skb_mstamp_get(&skb->skb_mstamp);
1984 goto repair; /* Skip network transmission */
1987 cwnd_quota = tcp_cwnd_test(tp, skb);
1989 is_cwnd_limited = true;
1991 /* Force out a loss probe pkt. */
1997 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2000 if (tso_segs == 1) {
2001 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2002 (tcp_skb_is_last(sk, skb) ?
2003 nonagle : TCP_NAGLE_PUSH))))
2007 tcp_tso_should_defer(sk, skb, &is_cwnd_limited))
2011 /* TCP Small Queues :
2012 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2014 * - better RTT estimation and ACK scheduling
2017 * Alas, some drivers / subsystems require a fair amount
2018 * of queued bytes to ensure line rate.
2019 * One example is wifi aggregation (802.11 AMPDU)
2021 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
2022 sk->sk_pacing_rate >> 10);
2024 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2025 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2026 /* It is possible TX completion already happened
2027 * before we set TSQ_THROTTLED, so we must
2028 * test again the condition.
2030 smp_mb__after_atomic();
2031 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2036 if (tso_segs > 1 && !tcp_urg_mode(tp))
2037 limit = tcp_mss_split_point(sk, skb, mss_now,
2040 sk->sk_gso_max_segs),
2043 if (skb->len > limit &&
2044 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2047 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2051 /* Advance the send_head. This one is sent out.
2052 * This call will increment packets_out.
2054 tcp_event_new_data_sent(sk, skb);
2056 tcp_minshall_update(tp, mss_now, skb);
2057 sent_pkts += tcp_skb_pcount(skb);
2063 if (likely(sent_pkts)) {
2064 if (tcp_in_cwnd_reduction(sk))
2065 tp->prr_out += sent_pkts;
2067 /* Send one loss probe per tail loss episode. */
2069 tcp_schedule_loss_probe(sk);
2070 tcp_cwnd_validate(sk, is_cwnd_limited);
2073 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2076 bool tcp_schedule_loss_probe(struct sock *sk)
2078 struct inet_connection_sock *icsk = inet_csk(sk);
2079 struct tcp_sock *tp = tcp_sk(sk);
2080 u32 timeout, tlp_time_stamp, rto_time_stamp;
2081 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2083 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2085 /* No consecutive loss probes. */
2086 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2090 /* Don't do any loss probe on a Fast Open connection before 3WHS
2093 if (sk->sk_state == TCP_SYN_RECV)
2096 /* TLP is only scheduled when next timer event is RTO. */
2097 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2100 /* Schedule a loss probe in 2*RTT for SACK capable connections
2101 * in Open state, that are either limited by cwnd or application.
2103 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2104 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2107 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2111 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2112 * for delayed ack when there's one outstanding packet.
2115 if (tp->packets_out == 1)
2116 timeout = max_t(u32, timeout,
2117 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2118 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2120 /* If RTO is shorter, just schedule TLP in its place. */
2121 tlp_time_stamp = tcp_time_stamp + timeout;
2122 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2123 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2124 s32 delta = rto_time_stamp - tcp_time_stamp;
2129 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2134 /* Thanks to skb fast clones, we can detect if a prior transmit of
2135 * a packet is still in a qdisc or driver queue.
2136 * In this case, there is very little point doing a retransmit !
2137 * Note: This is called from BH context only.
2139 static bool skb_still_in_host_queue(const struct sock *sk,
2140 const struct sk_buff *skb)
2142 if (unlikely(skb_fclone_busy(sk, skb))) {
2143 NET_INC_STATS_BH(sock_net(sk),
2144 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2150 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2151 * retransmit the last segment.
2153 void tcp_send_loss_probe(struct sock *sk)
2155 struct tcp_sock *tp = tcp_sk(sk);
2156 struct sk_buff *skb;
2158 int mss = tcp_current_mss(sk);
2161 if (tcp_send_head(sk) != NULL) {
2162 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2166 /* At most one outstanding TLP retransmission. */
2167 if (tp->tlp_high_seq)
2170 /* Retransmit last segment. */
2171 skb = tcp_write_queue_tail(sk);
2175 if (skb_still_in_host_queue(sk, skb))
2178 pcount = tcp_skb_pcount(skb);
2179 if (WARN_ON(!pcount))
2182 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2183 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2186 skb = tcp_write_queue_tail(sk);
2189 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2192 err = __tcp_retransmit_skb(sk, skb);
2194 /* Record snd_nxt for loss detection. */
2196 tp->tlp_high_seq = tp->snd_nxt;
2199 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2200 inet_csk(sk)->icsk_rto,
2204 NET_INC_STATS_BH(sock_net(sk),
2205 LINUX_MIB_TCPLOSSPROBES);
2208 /* Push out any pending frames which were held back due to
2209 * TCP_CORK or attempt at coalescing tiny packets.
2210 * The socket must be locked by the caller.
2212 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2215 /* If we are closed, the bytes will have to remain here.
2216 * In time closedown will finish, we empty the write queue and
2217 * all will be happy.
2219 if (unlikely(sk->sk_state == TCP_CLOSE))
2222 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2223 sk_gfp_atomic(sk, GFP_ATOMIC)))
2224 tcp_check_probe_timer(sk);
2227 /* Send _single_ skb sitting at the send head. This function requires
2228 * true push pending frames to setup probe timer etc.
2230 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2232 struct sk_buff *skb = tcp_send_head(sk);
2234 BUG_ON(!skb || skb->len < mss_now);
2236 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2239 /* This function returns the amount that we can raise the
2240 * usable window based on the following constraints
2242 * 1. The window can never be shrunk once it is offered (RFC 793)
2243 * 2. We limit memory per socket
2246 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2247 * RECV.NEXT + RCV.WIN fixed until:
2248 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2250 * i.e. don't raise the right edge of the window until you can raise
2251 * it at least MSS bytes.
2253 * Unfortunately, the recommended algorithm breaks header prediction,
2254 * since header prediction assumes th->window stays fixed.
2256 * Strictly speaking, keeping th->window fixed violates the receiver
2257 * side SWS prevention criteria. The problem is that under this rule
2258 * a stream of single byte packets will cause the right side of the
2259 * window to always advance by a single byte.
2261 * Of course, if the sender implements sender side SWS prevention
2262 * then this will not be a problem.
2264 * BSD seems to make the following compromise:
2266 * If the free space is less than the 1/4 of the maximum
2267 * space available and the free space is less than 1/2 mss,
2268 * then set the window to 0.
2269 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2270 * Otherwise, just prevent the window from shrinking
2271 * and from being larger than the largest representable value.
2273 * This prevents incremental opening of the window in the regime
2274 * where TCP is limited by the speed of the reader side taking
2275 * data out of the TCP receive queue. It does nothing about
2276 * those cases where the window is constrained on the sender side
2277 * because the pipeline is full.
2279 * BSD also seems to "accidentally" limit itself to windows that are a
2280 * multiple of MSS, at least until the free space gets quite small.
2281 * This would appear to be a side effect of the mbuf implementation.
2282 * Combining these two algorithms results in the observed behavior
2283 * of having a fixed window size at almost all times.
2285 * Below we obtain similar behavior by forcing the offered window to
2286 * a multiple of the mss when it is feasible to do so.
2288 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2289 * Regular options like TIMESTAMP are taken into account.
2291 u32 __tcp_select_window(struct sock *sk)
2293 struct inet_connection_sock *icsk = inet_csk(sk);
2294 struct tcp_sock *tp = tcp_sk(sk);
2295 /* MSS for the peer's data. Previous versions used mss_clamp
2296 * here. I don't know if the value based on our guesses
2297 * of peer's MSS is better for the performance. It's more correct
2298 * but may be worse for the performance because of rcv_mss
2299 * fluctuations. --SAW 1998/11/1
2301 int mss = icsk->icsk_ack.rcv_mss;
2302 int free_space = tcp_space(sk);
2303 int allowed_space = tcp_full_space(sk);
2304 int full_space = min_t(int, tp->window_clamp, allowed_space);
2307 if (mss > full_space)
2310 if (free_space < (full_space >> 1)) {
2311 icsk->icsk_ack.quick = 0;
2313 if (sk_under_memory_pressure(sk))
2314 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2317 /* free_space might become our new window, make sure we don't
2318 * increase it due to wscale.
2320 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2322 /* if free space is less than mss estimate, or is below 1/16th
2323 * of the maximum allowed, try to move to zero-window, else
2324 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2325 * new incoming data is dropped due to memory limits.
2326 * With large window, mss test triggers way too late in order
2327 * to announce zero window in time before rmem limit kicks in.
2329 if (free_space < (allowed_space >> 4) || free_space < mss)
2333 if (free_space > tp->rcv_ssthresh)
2334 free_space = tp->rcv_ssthresh;
2336 /* Don't do rounding if we are using window scaling, since the
2337 * scaled window will not line up with the MSS boundary anyway.
2339 window = tp->rcv_wnd;
2340 if (tp->rx_opt.rcv_wscale) {
2341 window = free_space;
2343 /* Advertise enough space so that it won't get scaled away.
2344 * Import case: prevent zero window announcement if
2345 * 1<<rcv_wscale > mss.
2347 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2348 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2349 << tp->rx_opt.rcv_wscale);
2351 /* Get the largest window that is a nice multiple of mss.
2352 * Window clamp already applied above.
2353 * If our current window offering is within 1 mss of the
2354 * free space we just keep it. This prevents the divide
2355 * and multiply from happening most of the time.
2356 * We also don't do any window rounding when the free space
2359 if (window <= free_space - mss || window > free_space)
2360 window = (free_space / mss) * mss;
2361 else if (mss == full_space &&
2362 free_space > window + (full_space >> 1))
2363 window = free_space;
2369 /* Collapses two adjacent SKB's during retransmission. */
2370 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2372 struct tcp_sock *tp = tcp_sk(sk);
2373 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2374 int skb_size, next_skb_size;
2376 skb_size = skb->len;
2377 next_skb_size = next_skb->len;
2379 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2381 tcp_highest_sack_combine(sk, next_skb, skb);
2383 tcp_unlink_write_queue(next_skb, sk);
2385 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2388 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2389 skb->ip_summed = CHECKSUM_PARTIAL;
2391 if (skb->ip_summed != CHECKSUM_PARTIAL)
2392 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2394 /* Update sequence range on original skb. */
2395 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2397 /* Merge over control information. This moves PSH/FIN etc. over */
2398 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2400 /* All done, get rid of second SKB and account for it so
2401 * packet counting does not break.
2403 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2405 /* changed transmit queue under us so clear hints */
2406 tcp_clear_retrans_hints_partial(tp);
2407 if (next_skb == tp->retransmit_skb_hint)
2408 tp->retransmit_skb_hint = skb;
2410 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2412 sk_wmem_free_skb(sk, next_skb);
2415 /* Check if coalescing SKBs is legal. */
2416 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2418 if (tcp_skb_pcount(skb) > 1)
2420 /* TODO: SACK collapsing could be used to remove this condition */
2421 if (skb_shinfo(skb)->nr_frags != 0)
2423 if (skb_cloned(skb))
2425 if (skb == tcp_send_head(sk))
2427 /* Some heurestics for collapsing over SACK'd could be invented */
2428 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2434 /* Collapse packets in the retransmit queue to make to create
2435 * less packets on the wire. This is only done on retransmission.
2437 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2440 struct tcp_sock *tp = tcp_sk(sk);
2441 struct sk_buff *skb = to, *tmp;
2444 if (!sysctl_tcp_retrans_collapse)
2446 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2449 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2450 if (!tcp_can_collapse(sk, skb))
2462 /* Punt if not enough space exists in the first SKB for
2463 * the data in the second
2465 if (skb->len > skb_availroom(to))
2468 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2471 tcp_collapse_retrans(sk, to);
2475 /* This retransmits one SKB. Policy decisions and retransmit queue
2476 * state updates are done by the caller. Returns non-zero if an
2477 * error occurred which prevented the send.
2479 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2481 struct tcp_sock *tp = tcp_sk(sk);
2482 struct inet_connection_sock *icsk = inet_csk(sk);
2483 unsigned int cur_mss;
2486 /* Inconslusive MTU probe */
2487 if (icsk->icsk_mtup.probe_size) {
2488 icsk->icsk_mtup.probe_size = 0;
2491 /* Do not sent more than we queued. 1/4 is reserved for possible
2492 * copying overhead: fragmentation, tunneling, mangling etc.
2494 if (atomic_read(&sk->sk_wmem_alloc) >
2495 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2498 if (skb_still_in_host_queue(sk, skb))
2501 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2502 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2504 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2508 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2509 return -EHOSTUNREACH; /* Routing failure or similar. */
2511 cur_mss = tcp_current_mss(sk);
2513 /* If receiver has shrunk his window, and skb is out of
2514 * new window, do not retransmit it. The exception is the
2515 * case, when window is shrunk to zero. In this case
2516 * our retransmit serves as a zero window probe.
2518 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2519 TCP_SKB_CB(skb)->seq != tp->snd_una)
2522 if (skb->len > cur_mss) {
2523 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2524 return -ENOMEM; /* We'll try again later. */
2526 int oldpcount = tcp_skb_pcount(skb);
2528 if (unlikely(oldpcount > 1)) {
2529 if (skb_unclone(skb, GFP_ATOMIC))
2531 tcp_init_tso_segs(sk, skb, cur_mss);
2532 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2536 tcp_retrans_try_collapse(sk, skb, cur_mss);
2538 /* Make a copy, if the first transmission SKB clone we made
2539 * is still in somebody's hands, else make a clone.
2542 /* make sure skb->data is aligned on arches that require it
2543 * and check if ack-trimming & collapsing extended the headroom
2544 * beyond what csum_start can cover.
2546 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2547 skb_headroom(skb) >= 0xFFFF)) {
2548 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2550 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2553 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2557 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2558 /* Update global TCP statistics. */
2559 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2560 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2561 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2562 tp->total_retrans++;
2567 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2569 struct tcp_sock *tp = tcp_sk(sk);
2570 int err = __tcp_retransmit_skb(sk, skb);
2573 #if FASTRETRANS_DEBUG > 0
2574 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2575 net_dbg_ratelimited("retrans_out leaked\n");
2578 if (!tp->retrans_out)
2579 tp->lost_retrans_low = tp->snd_nxt;
2580 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2581 tp->retrans_out += tcp_skb_pcount(skb);
2583 /* Save stamp of the first retransmit. */
2584 if (!tp->retrans_stamp)
2585 tp->retrans_stamp = tcp_skb_timestamp(skb);
2587 /* snd_nxt is stored to detect loss of retransmitted segment,
2588 * see tcp_input.c tcp_sacktag_write_queue().
2590 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2591 } else if (err != -EBUSY) {
2592 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2595 if (tp->undo_retrans < 0)
2596 tp->undo_retrans = 0;
2597 tp->undo_retrans += tcp_skb_pcount(skb);
2601 /* Check if we forward retransmits are possible in the current
2602 * window/congestion state.
2604 static bool tcp_can_forward_retransmit(struct sock *sk)
2606 const struct inet_connection_sock *icsk = inet_csk(sk);
2607 const struct tcp_sock *tp = tcp_sk(sk);
2609 /* Forward retransmissions are possible only during Recovery. */
2610 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2613 /* No forward retransmissions in Reno are possible. */
2614 if (tcp_is_reno(tp))
2617 /* Yeah, we have to make difficult choice between forward transmission
2618 * and retransmission... Both ways have their merits...
2620 * For now we do not retransmit anything, while we have some new
2621 * segments to send. In the other cases, follow rule 3 for
2622 * NextSeg() specified in RFC3517.
2625 if (tcp_may_send_now(sk))
2631 /* This gets called after a retransmit timeout, and the initially
2632 * retransmitted data is acknowledged. It tries to continue
2633 * resending the rest of the retransmit queue, until either
2634 * we've sent it all or the congestion window limit is reached.
2635 * If doing SACK, the first ACK which comes back for a timeout
2636 * based retransmit packet might feed us FACK information again.
2637 * If so, we use it to avoid unnecessarily retransmissions.
2639 void tcp_xmit_retransmit_queue(struct sock *sk)
2641 const struct inet_connection_sock *icsk = inet_csk(sk);
2642 struct tcp_sock *tp = tcp_sk(sk);
2643 struct sk_buff *skb;
2644 struct sk_buff *hole = NULL;
2647 int fwd_rexmitting = 0;
2649 if (!tp->packets_out)
2653 tp->retransmit_high = tp->snd_una;
2655 if (tp->retransmit_skb_hint) {
2656 skb = tp->retransmit_skb_hint;
2657 last_lost = TCP_SKB_CB(skb)->end_seq;
2658 if (after(last_lost, tp->retransmit_high))
2659 last_lost = tp->retransmit_high;
2661 skb = tcp_write_queue_head(sk);
2662 last_lost = tp->snd_una;
2665 tcp_for_write_queue_from(skb, sk) {
2666 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2668 if (skb == tcp_send_head(sk))
2670 /* we could do better than to assign each time */
2672 tp->retransmit_skb_hint = skb;
2674 /* Assume this retransmit will generate
2675 * only one packet for congestion window
2676 * calculation purposes. This works because
2677 * tcp_retransmit_skb() will chop up the
2678 * packet to be MSS sized and all the
2679 * packet counting works out.
2681 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2684 if (fwd_rexmitting) {
2686 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2688 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2690 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2691 tp->retransmit_high = last_lost;
2692 if (!tcp_can_forward_retransmit(sk))
2694 /* Backtrack if necessary to non-L'ed skb */
2702 } else if (!(sacked & TCPCB_LOST)) {
2703 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2708 last_lost = TCP_SKB_CB(skb)->end_seq;
2709 if (icsk->icsk_ca_state != TCP_CA_Loss)
2710 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2712 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2715 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2718 if (tcp_retransmit_skb(sk, skb))
2721 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2723 if (tcp_in_cwnd_reduction(sk))
2724 tp->prr_out += tcp_skb_pcount(skb);
2726 if (skb == tcp_write_queue_head(sk))
2727 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2728 inet_csk(sk)->icsk_rto,
2733 /* Send a fin. The caller locks the socket for us. This cannot be
2734 * allowed to fail queueing a FIN frame under any circumstances.
2736 void tcp_send_fin(struct sock *sk)
2738 struct tcp_sock *tp = tcp_sk(sk);
2739 struct sk_buff *skb = tcp_write_queue_tail(sk);
2742 /* Optimization, tack on the FIN if we have a queue of
2743 * unsent frames. But be careful about outgoing SACKS
2746 mss_now = tcp_current_mss(sk);
2748 if (tcp_send_head(sk) != NULL) {
2749 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2750 TCP_SKB_CB(skb)->end_seq++;
2753 /* Socket is locked, keep trying until memory is available. */
2755 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2762 /* Reserve space for headers and prepare control bits. */
2763 skb_reserve(skb, MAX_TCP_HEADER);
2764 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2765 tcp_init_nondata_skb(skb, tp->write_seq,
2766 TCPHDR_ACK | TCPHDR_FIN);
2767 tcp_queue_skb(sk, skb);
2769 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2772 /* We get here when a process closes a file descriptor (either due to
2773 * an explicit close() or as a byproduct of exit()'ing) and there
2774 * was unread data in the receive queue. This behavior is recommended
2775 * by RFC 2525, section 2.17. -DaveM
2777 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2779 struct sk_buff *skb;
2781 /* NOTE: No TCP options attached and we never retransmit this. */
2782 skb = alloc_skb(MAX_TCP_HEADER, priority);
2784 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2788 /* Reserve space for headers and prepare control bits. */
2789 skb_reserve(skb, MAX_TCP_HEADER);
2790 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2791 TCPHDR_ACK | TCPHDR_RST);
2793 if (tcp_transmit_skb(sk, skb, 0, priority))
2794 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2796 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2799 /* Send a crossed SYN-ACK during socket establishment.
2800 * WARNING: This routine must only be called when we have already sent
2801 * a SYN packet that crossed the incoming SYN that caused this routine
2802 * to get called. If this assumption fails then the initial rcv_wnd
2803 * and rcv_wscale values will not be correct.
2805 int tcp_send_synack(struct sock *sk)
2807 struct sk_buff *skb;
2809 skb = tcp_write_queue_head(sk);
2810 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2811 pr_debug("%s: wrong queue state\n", __func__);
2814 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2815 if (skb_cloned(skb)) {
2816 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2819 tcp_unlink_write_queue(skb, sk);
2820 __skb_header_release(nskb);
2821 __tcp_add_write_queue_head(sk, nskb);
2822 sk_wmem_free_skb(sk, skb);
2823 sk->sk_wmem_queued += nskb->truesize;
2824 sk_mem_charge(sk, nskb->truesize);
2828 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2829 tcp_ecn_send_synack(sk, skb);
2831 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2835 * tcp_make_synack - Prepare a SYN-ACK.
2836 * sk: listener socket
2837 * dst: dst entry attached to the SYNACK
2838 * req: request_sock pointer
2840 * Allocate one skb and build a SYNACK packet.
2841 * @dst is consumed : Caller should not use it again.
2843 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2844 struct request_sock *req,
2845 struct tcp_fastopen_cookie *foc)
2847 struct tcp_out_options opts;
2848 struct inet_request_sock *ireq = inet_rsk(req);
2849 struct tcp_sock *tp = tcp_sk(sk);
2851 struct sk_buff *skb;
2852 struct tcp_md5sig_key *md5;
2853 int tcp_header_size;
2856 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2857 if (unlikely(!skb)) {
2861 /* Reserve space for headers. */
2862 skb_reserve(skb, MAX_TCP_HEADER);
2864 skb_dst_set(skb, dst);
2865 security_skb_owned_by(skb, sk);
2867 mss = dst_metric_advmss(dst);
2868 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2869 mss = tp->rx_opt.user_mss;
2871 memset(&opts, 0, sizeof(opts));
2872 #ifdef CONFIG_SYN_COOKIES
2873 if (unlikely(req->cookie_ts))
2874 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2877 skb_mstamp_get(&skb->skb_mstamp);
2878 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2881 skb_push(skb, tcp_header_size);
2882 skb_reset_transport_header(skb);
2885 memset(th, 0, sizeof(struct tcphdr));
2888 tcp_ecn_make_synack(req, th, sk);
2889 th->source = htons(ireq->ir_num);
2890 th->dest = ireq->ir_rmt_port;
2891 /* Setting of flags are superfluous here for callers (and ECE is
2892 * not even correctly set)
2894 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2895 TCPHDR_SYN | TCPHDR_ACK);
2897 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2898 /* XXX data is queued and acked as is. No buffer/window check */
2899 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2901 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2902 th->window = htons(min(req->rcv_wnd, 65535U));
2903 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2904 th->doff = (tcp_header_size >> 2);
2905 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2907 #ifdef CONFIG_TCP_MD5SIG
2908 /* Okay, we have all we need - do the md5 hash if needed */
2910 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2911 md5, NULL, req, skb);
2917 EXPORT_SYMBOL(tcp_make_synack);
2919 /* Do all connect socket setups that can be done AF independent. */
2920 static void tcp_connect_init(struct sock *sk)
2922 const struct dst_entry *dst = __sk_dst_get(sk);
2923 struct tcp_sock *tp = tcp_sk(sk);
2926 /* We'll fix this up when we get a response from the other end.
2927 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2929 tp->tcp_header_len = sizeof(struct tcphdr) +
2930 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2932 #ifdef CONFIG_TCP_MD5SIG
2933 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2934 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2937 /* If user gave his TCP_MAXSEG, record it to clamp */
2938 if (tp->rx_opt.user_mss)
2939 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2942 tcp_sync_mss(sk, dst_mtu(dst));
2944 if (!tp->window_clamp)
2945 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2946 tp->advmss = dst_metric_advmss(dst);
2947 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2948 tp->advmss = tp->rx_opt.user_mss;
2950 tcp_initialize_rcv_mss(sk);
2952 /* limit the window selection if the user enforce a smaller rx buffer */
2953 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2954 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2955 tp->window_clamp = tcp_full_space(sk);
2957 tcp_select_initial_window(tcp_full_space(sk),
2958 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2961 sysctl_tcp_window_scaling,
2963 dst_metric(dst, RTAX_INITRWND));
2965 tp->rx_opt.rcv_wscale = rcv_wscale;
2966 tp->rcv_ssthresh = tp->rcv_wnd;
2969 sock_reset_flag(sk, SOCK_DONE);
2972 tp->snd_una = tp->write_seq;
2973 tp->snd_sml = tp->write_seq;
2974 tp->snd_up = tp->write_seq;
2975 tp->snd_nxt = tp->write_seq;
2977 if (likely(!tp->repair))
2980 tp->rcv_tstamp = tcp_time_stamp;
2981 tp->rcv_wup = tp->rcv_nxt;
2982 tp->copied_seq = tp->rcv_nxt;
2984 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2985 inet_csk(sk)->icsk_retransmits = 0;
2986 tcp_clear_retrans(tp);
2989 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2991 struct tcp_sock *tp = tcp_sk(sk);
2992 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2994 tcb->end_seq += skb->len;
2995 __skb_header_release(skb);
2996 __tcp_add_write_queue_tail(sk, skb);
2997 sk->sk_wmem_queued += skb->truesize;
2998 sk_mem_charge(sk, skb->truesize);
2999 tp->write_seq = tcb->end_seq;
3000 tp->packets_out += tcp_skb_pcount(skb);
3003 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3004 * queue a data-only packet after the regular SYN, such that regular SYNs
3005 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3006 * only the SYN sequence, the data are retransmitted in the first ACK.
3007 * If cookie is not cached or other error occurs, falls back to send a
3008 * regular SYN with Fast Open cookie request option.
3010 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3012 struct tcp_sock *tp = tcp_sk(sk);
3013 struct tcp_fastopen_request *fo = tp->fastopen_req;
3014 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
3015 struct sk_buff *syn_data = NULL, *data;
3016 unsigned long last_syn_loss = 0;
3018 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3019 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3020 &syn_loss, &last_syn_loss);
3021 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3023 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3024 fo->cookie.len = -1;
3028 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3029 fo->cookie.len = -1;
3030 else if (fo->cookie.len <= 0)
3033 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3034 * user-MSS. Reserve maximum option space for middleboxes that add
3035 * private TCP options. The cost is reduced data space in SYN :(
3037 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3038 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3039 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3040 MAX_TCP_OPTION_SPACE;
3042 space = min_t(size_t, space, fo->size);
3044 /* limit to order-0 allocations */
3045 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3047 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
3049 if (syn_data == NULL)
3052 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3053 struct iovec *iov = &fo->data->msg_iov[i];
3054 unsigned char __user *from = iov->iov_base;
3055 int len = iov->iov_len;
3057 if (syn_data->len + len > space)
3058 len = space - syn_data->len;
3059 else if (i + 1 == iovlen)
3060 /* No more data pending in inet_wait_for_connect() */
3063 if (skb_add_data(syn_data, from, len))
3067 /* Queue a data-only packet after the regular SYN for retransmission */
3068 data = pskb_copy(syn_data, sk->sk_allocation);
3071 TCP_SKB_CB(data)->seq++;
3072 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3073 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3074 tcp_connect_queue_skb(sk, data);
3075 fo->copied = data->len;
3077 /* syn_data is about to be sent, we need to take current time stamps
3078 * for the packets that are in write queue : SYN packet and DATA
3080 skb_mstamp_get(&syn->skb_mstamp);
3081 data->skb_mstamp = syn->skb_mstamp;
3083 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
3084 tp->syn_data = (fo->copied > 0);
3085 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3091 /* Send a regular SYN with Fast Open cookie request option */
3092 if (fo->cookie.len > 0)
3094 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3096 tp->syn_fastopen = 0;
3097 kfree_skb(syn_data);
3099 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3103 /* Build a SYN and send it off. */
3104 int tcp_connect(struct sock *sk)
3106 struct tcp_sock *tp = tcp_sk(sk);
3107 struct sk_buff *buff;
3110 tcp_connect_init(sk);
3112 if (unlikely(tp->repair)) {
3113 tcp_finish_connect(sk, NULL);
3117 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3118 if (unlikely(buff == NULL))
3121 /* Reserve space for headers. */
3122 skb_reserve(buff, MAX_TCP_HEADER);
3124 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3125 tp->retrans_stamp = tcp_time_stamp;
3126 tcp_connect_queue_skb(sk, buff);
3127 tcp_ecn_send_syn(sk, buff);
3129 /* Send off SYN; include data in Fast Open. */
3130 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3131 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3132 if (err == -ECONNREFUSED)
3135 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3136 * in order to make this packet get counted in tcpOutSegs.
3138 tp->snd_nxt = tp->write_seq;
3139 tp->pushed_seq = tp->write_seq;
3140 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3142 /* Timer for repeating the SYN until an answer. */
3143 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3144 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3147 EXPORT_SYMBOL(tcp_connect);
3149 /* Send out a delayed ack, the caller does the policy checking
3150 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3153 void tcp_send_delayed_ack(struct sock *sk)
3155 struct inet_connection_sock *icsk = inet_csk(sk);
3156 int ato = icsk->icsk_ack.ato;
3157 unsigned long timeout;
3159 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3161 if (ato > TCP_DELACK_MIN) {
3162 const struct tcp_sock *tp = tcp_sk(sk);
3163 int max_ato = HZ / 2;
3165 if (icsk->icsk_ack.pingpong ||
3166 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3167 max_ato = TCP_DELACK_MAX;
3169 /* Slow path, intersegment interval is "high". */
3171 /* If some rtt estimate is known, use it to bound delayed ack.
3172 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3176 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3183 ato = min(ato, max_ato);
3186 /* Stay within the limit we were given */
3187 timeout = jiffies + ato;
3189 /* Use new timeout only if there wasn't a older one earlier. */
3190 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3191 /* If delack timer was blocked or is about to expire,
3194 if (icsk->icsk_ack.blocked ||
3195 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3200 if (!time_before(timeout, icsk->icsk_ack.timeout))
3201 timeout = icsk->icsk_ack.timeout;
3203 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3204 icsk->icsk_ack.timeout = timeout;
3205 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3208 /* This routine sends an ack and also updates the window. */
3209 void tcp_send_ack(struct sock *sk)
3211 struct sk_buff *buff;
3213 /* If we have been reset, we may not send again. */
3214 if (sk->sk_state == TCP_CLOSE)
3217 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3219 /* We are not putting this on the write queue, so
3220 * tcp_transmit_skb() will set the ownership to this
3223 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3225 inet_csk_schedule_ack(sk);
3226 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3227 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3228 TCP_DELACK_MAX, TCP_RTO_MAX);
3232 /* Reserve space for headers and prepare control bits. */
3233 skb_reserve(buff, MAX_TCP_HEADER);
3234 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3236 /* Send it off, this clears delayed acks for us. */
3237 skb_mstamp_get(&buff->skb_mstamp);
3238 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3240 EXPORT_SYMBOL_GPL(tcp_send_ack);
3242 /* This routine sends a packet with an out of date sequence
3243 * number. It assumes the other end will try to ack it.
3245 * Question: what should we make while urgent mode?
3246 * 4.4BSD forces sending single byte of data. We cannot send
3247 * out of window data, because we have SND.NXT==SND.MAX...
3249 * Current solution: to send TWO zero-length segments in urgent mode:
3250 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3251 * out-of-date with SND.UNA-1 to probe window.
3253 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3255 struct tcp_sock *tp = tcp_sk(sk);
3256 struct sk_buff *skb;
3258 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3259 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3263 /* Reserve space for headers and set control bits. */
3264 skb_reserve(skb, MAX_TCP_HEADER);
3265 /* Use a previous sequence. This should cause the other
3266 * end to send an ack. Don't queue or clone SKB, just
3269 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3270 skb_mstamp_get(&skb->skb_mstamp);
3271 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3274 void tcp_send_window_probe(struct sock *sk)
3276 if (sk->sk_state == TCP_ESTABLISHED) {
3277 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3278 tcp_xmit_probe_skb(sk, 0);
3282 /* Initiate keepalive or window probe from timer. */
3283 int tcp_write_wakeup(struct sock *sk)
3285 struct tcp_sock *tp = tcp_sk(sk);
3286 struct sk_buff *skb;
3288 if (sk->sk_state == TCP_CLOSE)
3291 if ((skb = tcp_send_head(sk)) != NULL &&
3292 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3294 unsigned int mss = tcp_current_mss(sk);
3295 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3297 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3298 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3300 /* We are probing the opening of a window
3301 * but the window size is != 0
3302 * must have been a result SWS avoidance ( sender )
3304 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3306 seg_size = min(seg_size, mss);
3307 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3308 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3310 } else if (!tcp_skb_pcount(skb))
3311 tcp_set_skb_tso_segs(sk, skb, mss);
3313 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3314 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3316 tcp_event_new_data_sent(sk, skb);
3319 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3320 tcp_xmit_probe_skb(sk, 1);
3321 return tcp_xmit_probe_skb(sk, 0);
3325 /* A window probe timeout has occurred. If window is not closed send
3326 * a partial packet else a zero probe.
3328 void tcp_send_probe0(struct sock *sk)
3330 struct inet_connection_sock *icsk = inet_csk(sk);
3331 struct tcp_sock *tp = tcp_sk(sk);
3332 unsigned long probe_max;
3335 err = tcp_write_wakeup(sk);
3337 if (tp->packets_out || !tcp_send_head(sk)) {
3338 /* Cancel probe timer, if it is not required. */
3339 icsk->icsk_probes_out = 0;
3340 icsk->icsk_backoff = 0;
3345 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3346 icsk->icsk_backoff++;
3347 icsk->icsk_probes_out++;
3348 probe_max = TCP_RTO_MAX;
3350 /* If packet was not sent due to local congestion,
3351 * do not backoff and do not remember icsk_probes_out.
3352 * Let local senders to fight for local resources.
3354 * Use accumulated backoff yet.
3356 if (!icsk->icsk_probes_out)
3357 icsk->icsk_probes_out = 1;
3358 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3360 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3361 inet_csk_rto_backoff(icsk, probe_max),
3365 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3367 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3371 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3373 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3374 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3378 EXPORT_SYMBOL(tcp_rtx_synack);