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 inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
324 if (!(tp->ecn_flags & TCP_ECN_OK))
325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
328 /* Packet ECN state for a SYN. */
329 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
331 struct tcp_sock *tp = tcp_sk(sk);
334 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
340 static __inline__ void
341 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
343 if (inet_rsk(req)->ecn_ok)
347 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
350 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
353 struct tcp_sock *tp = tcp_sk(sk);
355 if (tp->ecn_flags & TCP_ECN_OK) {
356 /* Not-retransmitted data segment: set ECT and inject CWR. */
357 if (skb->len != tcp_header_len &&
358 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
360 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
361 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
362 tcp_hdr(skb)->cwr = 1;
363 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
366 /* ACK or retransmitted segment: clear ECT|CE */
367 INET_ECN_dontxmit(sk);
369 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
370 tcp_hdr(skb)->ece = 1;
374 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
375 * auto increment end seqno.
377 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
379 struct skb_shared_info *shinfo = skb_shinfo(skb);
381 skb->ip_summed = CHECKSUM_PARTIAL;
384 TCP_SKB_CB(skb)->tcp_flags = flags;
385 TCP_SKB_CB(skb)->sacked = 0;
387 shinfo->gso_segs = 1;
388 shinfo->gso_size = 0;
389 shinfo->gso_type = 0;
391 TCP_SKB_CB(skb)->seq = seq;
392 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
394 TCP_SKB_CB(skb)->end_seq = seq;
397 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
399 return tp->snd_una != tp->snd_up;
402 #define OPTION_SACK_ADVERTISE (1 << 0)
403 #define OPTION_TS (1 << 1)
404 #define OPTION_MD5 (1 << 2)
405 #define OPTION_WSCALE (1 << 3)
406 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
408 struct tcp_out_options {
409 u16 options; /* bit field of OPTION_* */
410 u16 mss; /* 0 to disable */
411 u8 ws; /* window scale, 0 to disable */
412 u8 num_sack_blocks; /* number of SACK blocks to include */
413 u8 hash_size; /* bytes in hash_location */
414 __u8 *hash_location; /* temporary pointer, overloaded */
415 __u32 tsval, tsecr; /* need to include OPTION_TS */
416 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
419 /* Write previously computed TCP options to the packet.
421 * Beware: Something in the Internet is very sensitive to the ordering of
422 * TCP options, we learned this through the hard way, so be careful here.
423 * Luckily we can at least blame others for their non-compliance but from
424 * inter-operability perspective it seems that we're somewhat stuck with
425 * the ordering which we have been using if we want to keep working with
426 * those broken things (not that it currently hurts anybody as there isn't
427 * particular reason why the ordering would need to be changed).
429 * At least SACK_PERM as the first option is known to lead to a disaster
430 * (but it may well be that other scenarios fail similarly).
432 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
433 struct tcp_out_options *opts)
435 u16 options = opts->options; /* mungable copy */
437 if (unlikely(OPTION_MD5 & options)) {
438 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
439 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
440 /* overload cookie hash location */
441 opts->hash_location = (__u8 *)ptr;
445 if (unlikely(opts->mss)) {
446 *ptr++ = htonl((TCPOPT_MSS << 24) |
447 (TCPOLEN_MSS << 16) |
451 if (likely(OPTION_TS & options)) {
452 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
453 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
454 (TCPOLEN_SACK_PERM << 16) |
455 (TCPOPT_TIMESTAMP << 8) |
457 options &= ~OPTION_SACK_ADVERTISE;
459 *ptr++ = htonl((TCPOPT_NOP << 24) |
461 (TCPOPT_TIMESTAMP << 8) |
464 *ptr++ = htonl(opts->tsval);
465 *ptr++ = htonl(opts->tsecr);
468 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
469 *ptr++ = htonl((TCPOPT_NOP << 24) |
471 (TCPOPT_SACK_PERM << 8) |
475 if (unlikely(OPTION_WSCALE & options)) {
476 *ptr++ = htonl((TCPOPT_NOP << 24) |
477 (TCPOPT_WINDOW << 16) |
478 (TCPOLEN_WINDOW << 8) |
482 if (unlikely(opts->num_sack_blocks)) {
483 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
484 tp->duplicate_sack : tp->selective_acks;
487 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
491 TCPOLEN_SACK_PERBLOCK)));
493 for (this_sack = 0; this_sack < opts->num_sack_blocks;
495 *ptr++ = htonl(sp[this_sack].start_seq);
496 *ptr++ = htonl(sp[this_sack].end_seq);
499 tp->rx_opt.dsack = 0;
502 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
503 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
505 *ptr++ = htonl((TCPOPT_EXP << 24) |
506 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
507 TCPOPT_FASTOPEN_MAGIC);
509 memcpy(ptr, foc->val, foc->len);
510 if ((foc->len & 3) == 2) {
511 u8 *align = ((u8 *)ptr) + foc->len;
512 align[0] = align[1] = TCPOPT_NOP;
514 ptr += (foc->len + 3) >> 2;
518 /* Compute TCP options for SYN packets. This is not the final
519 * network wire format yet.
521 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
522 struct tcp_out_options *opts,
523 struct tcp_md5sig_key **md5)
525 struct tcp_sock *tp = tcp_sk(sk);
526 unsigned int remaining = MAX_TCP_OPTION_SPACE;
527 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
529 #ifdef CONFIG_TCP_MD5SIG
530 *md5 = tp->af_specific->md5_lookup(sk, sk);
532 opts->options |= OPTION_MD5;
533 remaining -= TCPOLEN_MD5SIG_ALIGNED;
539 /* We always get an MSS option. The option bytes which will be seen in
540 * normal data packets should timestamps be used, must be in the MSS
541 * advertised. But we subtract them from tp->mss_cache so that
542 * calculations in tcp_sendmsg are simpler etc. So account for this
543 * fact here if necessary. If we don't do this correctly, as a
544 * receiver we won't recognize data packets as being full sized when we
545 * should, and thus we won't abide by the delayed ACK rules correctly.
546 * SACKs don't matter, we never delay an ACK when we have any of those
548 opts->mss = tcp_advertise_mss(sk);
549 remaining -= TCPOLEN_MSS_ALIGNED;
551 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
552 opts->options |= OPTION_TS;
553 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
554 opts->tsecr = tp->rx_opt.ts_recent;
555 remaining -= TCPOLEN_TSTAMP_ALIGNED;
557 if (likely(sysctl_tcp_window_scaling)) {
558 opts->ws = tp->rx_opt.rcv_wscale;
559 opts->options |= OPTION_WSCALE;
560 remaining -= TCPOLEN_WSCALE_ALIGNED;
562 if (likely(sysctl_tcp_sack)) {
563 opts->options |= OPTION_SACK_ADVERTISE;
564 if (unlikely(!(OPTION_TS & opts->options)))
565 remaining -= TCPOLEN_SACKPERM_ALIGNED;
568 if (fastopen && fastopen->cookie.len >= 0) {
569 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
570 need = (need + 3) & ~3U; /* Align to 32 bits */
571 if (remaining >= need) {
572 opts->options |= OPTION_FAST_OPEN_COOKIE;
573 opts->fastopen_cookie = &fastopen->cookie;
575 tp->syn_fastopen = 1;
579 return MAX_TCP_OPTION_SPACE - remaining;
582 /* Set up TCP options for SYN-ACKs. */
583 static unsigned int tcp_synack_options(struct sock *sk,
584 struct request_sock *req,
585 unsigned int mss, struct sk_buff *skb,
586 struct tcp_out_options *opts,
587 struct tcp_md5sig_key **md5,
588 struct tcp_fastopen_cookie *foc)
590 struct inet_request_sock *ireq = inet_rsk(req);
591 unsigned int remaining = MAX_TCP_OPTION_SPACE;
593 #ifdef CONFIG_TCP_MD5SIG
594 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
596 opts->options |= OPTION_MD5;
597 remaining -= TCPOLEN_MD5SIG_ALIGNED;
599 /* We can't fit any SACK blocks in a packet with MD5 + TS
600 * options. There was discussion about disabling SACK
601 * rather than TS in order to fit in better with old,
602 * buggy kernels, but that was deemed to be unnecessary.
604 ireq->tstamp_ok &= !ireq->sack_ok;
610 /* We always send an MSS option. */
612 remaining -= TCPOLEN_MSS_ALIGNED;
614 if (likely(ireq->wscale_ok)) {
615 opts->ws = ireq->rcv_wscale;
616 opts->options |= OPTION_WSCALE;
617 remaining -= TCPOLEN_WSCALE_ALIGNED;
619 if (likely(ireq->tstamp_ok)) {
620 opts->options |= OPTION_TS;
621 opts->tsval = tcp_skb_timestamp(skb);
622 opts->tsecr = req->ts_recent;
623 remaining -= TCPOLEN_TSTAMP_ALIGNED;
625 if (likely(ireq->sack_ok)) {
626 opts->options |= OPTION_SACK_ADVERTISE;
627 if (unlikely(!ireq->tstamp_ok))
628 remaining -= TCPOLEN_SACKPERM_ALIGNED;
630 if (foc != NULL && foc->len >= 0) {
631 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
632 need = (need + 3) & ~3U; /* Align to 32 bits */
633 if (remaining >= need) {
634 opts->options |= OPTION_FAST_OPEN_COOKIE;
635 opts->fastopen_cookie = foc;
640 return MAX_TCP_OPTION_SPACE - remaining;
643 /* Compute TCP options for ESTABLISHED sockets. This is not the
644 * final wire format yet.
646 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
647 struct tcp_out_options *opts,
648 struct tcp_md5sig_key **md5)
650 struct tcp_sock *tp = tcp_sk(sk);
651 unsigned int size = 0;
652 unsigned int eff_sacks;
656 #ifdef CONFIG_TCP_MD5SIG
657 *md5 = tp->af_specific->md5_lookup(sk, sk);
658 if (unlikely(*md5)) {
659 opts->options |= OPTION_MD5;
660 size += TCPOLEN_MD5SIG_ALIGNED;
666 if (likely(tp->rx_opt.tstamp_ok)) {
667 opts->options |= OPTION_TS;
668 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
669 opts->tsecr = tp->rx_opt.ts_recent;
670 size += TCPOLEN_TSTAMP_ALIGNED;
673 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
674 if (unlikely(eff_sacks)) {
675 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
676 opts->num_sack_blocks =
677 min_t(unsigned int, eff_sacks,
678 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
679 TCPOLEN_SACK_PERBLOCK);
680 size += TCPOLEN_SACK_BASE_ALIGNED +
681 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
688 /* TCP SMALL QUEUES (TSQ)
690 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
691 * to reduce RTT and bufferbloat.
692 * We do this using a special skb destructor (tcp_wfree).
694 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
695 * needs to be reallocated in a driver.
696 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
698 * Since transmit from skb destructor is forbidden, we use a tasklet
699 * to process all sockets that eventually need to send more skbs.
700 * We use one tasklet per cpu, with its own queue of sockets.
703 struct tasklet_struct tasklet;
704 struct list_head head; /* queue of tcp sockets */
706 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
708 static void tcp_tsq_handler(struct sock *sk)
710 if ((1 << sk->sk_state) &
711 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
712 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
713 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
717 * One tasklet per cpu tries to send more skbs.
718 * We run in tasklet context but need to disable irqs when
719 * transferring tsq->head because tcp_wfree() might
720 * interrupt us (non NAPI drivers)
722 static void tcp_tasklet_func(unsigned long data)
724 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
727 struct list_head *q, *n;
731 local_irq_save(flags);
732 list_splice_init(&tsq->head, &list);
733 local_irq_restore(flags);
735 list_for_each_safe(q, n, &list) {
736 tp = list_entry(q, struct tcp_sock, tsq_node);
737 list_del(&tp->tsq_node);
739 sk = (struct sock *)tp;
742 if (!sock_owned_by_user(sk)) {
745 /* defer the work to tcp_release_cb() */
746 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
750 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
755 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
756 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
757 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
758 (1UL << TCP_MTU_REDUCED_DEFERRED))
760 * tcp_release_cb - tcp release_sock() callback
763 * called from release_sock() to perform protocol dependent
764 * actions before socket release.
766 void tcp_release_cb(struct sock *sk)
768 struct tcp_sock *tp = tcp_sk(sk);
769 unsigned long flags, nflags;
771 /* perform an atomic operation only if at least one flag is set */
773 flags = tp->tsq_flags;
774 if (!(flags & TCP_DEFERRED_ALL))
776 nflags = flags & ~TCP_DEFERRED_ALL;
777 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
779 if (flags & (1UL << TCP_TSQ_DEFERRED))
782 /* Here begins the tricky part :
783 * We are called from release_sock() with :
785 * 2) sk_lock.slock spinlock held
786 * 3) socket owned by us (sk->sk_lock.owned == 1)
788 * But following code is meant to be called from BH handlers,
789 * so we should keep BH disabled, but early release socket ownership
791 sock_release_ownership(sk);
793 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
794 tcp_write_timer_handler(sk);
797 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
798 tcp_delack_timer_handler(sk);
801 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
802 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
806 EXPORT_SYMBOL(tcp_release_cb);
808 void __init tcp_tasklet_init(void)
812 for_each_possible_cpu(i) {
813 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
815 INIT_LIST_HEAD(&tsq->head);
816 tasklet_init(&tsq->tasklet,
823 * Write buffer destructor automatically called from kfree_skb.
824 * We can't xmit new skbs from this context, as we might already
827 void tcp_wfree(struct sk_buff *skb)
829 struct sock *sk = skb->sk;
830 struct tcp_sock *tp = tcp_sk(sk);
832 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
833 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
835 struct tsq_tasklet *tsq;
837 /* Keep a ref on socket.
838 * This last ref will be released in tcp_tasklet_func()
840 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
842 /* queue this socket to tasklet queue */
843 local_irq_save(flags);
844 tsq = &__get_cpu_var(tsq_tasklet);
845 list_add(&tp->tsq_node, &tsq->head);
846 tasklet_schedule(&tsq->tasklet);
847 local_irq_restore(flags);
853 /* This routine actually transmits TCP packets queued in by
854 * tcp_do_sendmsg(). This is used by both the initial
855 * transmission and possible later retransmissions.
856 * All SKB's seen here are completely headerless. It is our
857 * job to build the TCP header, and pass the packet down to
858 * IP so it can do the same plus pass the packet off to the
861 * We are working here with either a clone of the original
862 * SKB, or a fresh unique copy made by the retransmit engine.
864 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
867 const struct inet_connection_sock *icsk = inet_csk(sk);
868 struct inet_sock *inet;
870 struct tcp_skb_cb *tcb;
871 struct tcp_out_options opts;
872 unsigned int tcp_options_size, tcp_header_size;
873 struct tcp_md5sig_key *md5;
877 BUG_ON(!skb || !tcp_skb_pcount(skb));
880 skb_mstamp_get(&skb->skb_mstamp);
882 if (unlikely(skb_cloned(skb)))
883 skb = pskb_copy(skb, gfp_mask);
885 skb = skb_clone(skb, gfp_mask);
892 tcb = TCP_SKB_CB(skb);
893 memset(&opts, 0, sizeof(opts));
895 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
896 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
898 tcp_options_size = tcp_established_options(sk, skb, &opts,
900 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
902 if (tcp_packets_in_flight(tp) == 0)
903 tcp_ca_event(sk, CA_EVENT_TX_START);
905 /* if no packet is in qdisc/device queue, then allow XPS to select
908 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
910 skb_push(skb, tcp_header_size);
911 skb_reset_transport_header(skb);
915 skb->destructor = tcp_wfree;
916 skb_set_hash_from_sk(skb, sk);
917 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
919 /* Build TCP header and checksum it. */
921 th->source = inet->inet_sport;
922 th->dest = inet->inet_dport;
923 th->seq = htonl(tcb->seq);
924 th->ack_seq = htonl(tp->rcv_nxt);
925 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
928 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
929 /* RFC1323: The window in SYN & SYN/ACK segments
932 th->window = htons(min(tp->rcv_wnd, 65535U));
934 th->window = htons(tcp_select_window(sk));
939 /* The urg_mode check is necessary during a below snd_una win probe */
940 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
941 if (before(tp->snd_up, tcb->seq + 0x10000)) {
942 th->urg_ptr = htons(tp->snd_up - tcb->seq);
944 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
945 th->urg_ptr = htons(0xFFFF);
950 tcp_options_write((__be32 *)(th + 1), tp, &opts);
951 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
952 TCP_ECN_send(sk, skb, tcp_header_size);
954 #ifdef CONFIG_TCP_MD5SIG
955 /* Calculate the MD5 hash, as we have all we need now */
957 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
958 tp->af_specific->calc_md5_hash(opts.hash_location,
963 icsk->icsk_af_ops->send_check(sk, skb);
965 if (likely(tcb->tcp_flags & TCPHDR_ACK))
966 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
968 if (skb->len != tcp_header_size)
969 tcp_event_data_sent(tp, sk);
971 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
972 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
973 tcp_skb_pcount(skb));
975 /* Our usage of tstamp should remain private */
976 skb->tstamp.tv64 = 0;
978 /* Cleanup our debris for IP stacks */
979 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
980 sizeof(struct inet6_skb_parm)));
982 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
984 if (likely(err <= 0))
989 return net_xmit_eval(err);
992 /* This routine just queues the buffer for sending.
994 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
995 * otherwise socket can stall.
997 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
999 struct tcp_sock *tp = tcp_sk(sk);
1001 /* Advance write_seq and place onto the write_queue. */
1002 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1003 __skb_header_release(skb);
1004 tcp_add_write_queue_tail(sk, skb);
1005 sk->sk_wmem_queued += skb->truesize;
1006 sk_mem_charge(sk, skb->truesize);
1009 /* Initialize TSO segments for a packet. */
1010 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1011 unsigned int mss_now)
1013 struct skb_shared_info *shinfo = skb_shinfo(skb);
1015 /* Make sure we own this skb before messing gso_size/gso_segs */
1016 WARN_ON_ONCE(skb_cloned(skb));
1018 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1019 /* Avoid the costly divide in the normal
1022 shinfo->gso_segs = 1;
1023 shinfo->gso_size = 0;
1024 shinfo->gso_type = 0;
1026 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1027 shinfo->gso_size = mss_now;
1028 shinfo->gso_type = sk->sk_gso_type;
1032 /* When a modification to fackets out becomes necessary, we need to check
1033 * skb is counted to fackets_out or not.
1035 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1038 struct tcp_sock *tp = tcp_sk(sk);
1040 if (!tp->sacked_out || tcp_is_reno(tp))
1043 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1044 tp->fackets_out -= decr;
1047 /* Pcount in the middle of the write queue got changed, we need to do various
1048 * tweaks to fix counters
1050 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1052 struct tcp_sock *tp = tcp_sk(sk);
1054 tp->packets_out -= decr;
1056 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1057 tp->sacked_out -= decr;
1058 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1059 tp->retrans_out -= decr;
1060 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1061 tp->lost_out -= decr;
1063 /* Reno case is special. Sigh... */
1064 if (tcp_is_reno(tp) && decr > 0)
1065 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1067 tcp_adjust_fackets_out(sk, skb, decr);
1069 if (tp->lost_skb_hint &&
1070 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1071 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1072 tp->lost_cnt_hint -= decr;
1074 tcp_verify_left_out(tp);
1077 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1079 struct skb_shared_info *shinfo = skb_shinfo(skb);
1081 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1082 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1083 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1084 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1086 shinfo->tx_flags &= ~tsflags;
1087 shinfo2->tx_flags |= tsflags;
1088 swap(shinfo->tskey, shinfo2->tskey);
1092 /* Function to create two new TCP segments. Shrinks the given segment
1093 * to the specified size and appends a new segment with the rest of the
1094 * packet to the list. This won't be called frequently, I hope.
1095 * Remember, these are still headerless SKBs at this point.
1097 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1098 unsigned int mss_now, gfp_t gfp)
1100 struct tcp_sock *tp = tcp_sk(sk);
1101 struct sk_buff *buff;
1102 int nsize, old_factor;
1106 if (WARN_ON(len > skb->len))
1109 nsize = skb_headlen(skb) - len;
1113 if (skb_unclone(skb, gfp))
1116 /* Get a new skb... force flag on. */
1117 buff = sk_stream_alloc_skb(sk, nsize, gfp);
1119 return -ENOMEM; /* We'll just try again later. */
1121 sk->sk_wmem_queued += buff->truesize;
1122 sk_mem_charge(sk, buff->truesize);
1123 nlen = skb->len - len - nsize;
1124 buff->truesize += nlen;
1125 skb->truesize -= nlen;
1127 /* Correct the sequence numbers. */
1128 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1129 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1130 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1132 /* PSH and FIN should only be set in the second packet. */
1133 flags = TCP_SKB_CB(skb)->tcp_flags;
1134 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1135 TCP_SKB_CB(buff)->tcp_flags = flags;
1136 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1138 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1139 /* Copy and checksum data tail into the new buffer. */
1140 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1141 skb_put(buff, nsize),
1146 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1148 skb->ip_summed = CHECKSUM_PARTIAL;
1149 skb_split(skb, buff, len);
1152 buff->ip_summed = skb->ip_summed;
1154 buff->tstamp = skb->tstamp;
1155 tcp_fragment_tstamp(skb, buff);
1157 old_factor = tcp_skb_pcount(skb);
1159 /* Fix up tso_factor for both original and new SKB. */
1160 tcp_set_skb_tso_segs(sk, skb, mss_now);
1161 tcp_set_skb_tso_segs(sk, buff, mss_now);
1163 /* If this packet has been sent out already, we must
1164 * adjust the various packet counters.
1166 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1167 int diff = old_factor - tcp_skb_pcount(skb) -
1168 tcp_skb_pcount(buff);
1171 tcp_adjust_pcount(sk, skb, diff);
1174 /* Link BUFF into the send queue. */
1175 __skb_header_release(buff);
1176 tcp_insert_write_queue_after(skb, buff, sk);
1181 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1182 * eventually). The difference is that pulled data not copied, but
1183 * immediately discarded.
1185 static void __pskb_trim_head(struct sk_buff *skb, int len)
1187 struct skb_shared_info *shinfo;
1190 eat = min_t(int, len, skb_headlen(skb));
1192 __skb_pull(skb, eat);
1199 shinfo = skb_shinfo(skb);
1200 for (i = 0; i < shinfo->nr_frags; i++) {
1201 int size = skb_frag_size(&shinfo->frags[i]);
1204 skb_frag_unref(skb, i);
1207 shinfo->frags[k] = shinfo->frags[i];
1209 shinfo->frags[k].page_offset += eat;
1210 skb_frag_size_sub(&shinfo->frags[k], eat);
1216 shinfo->nr_frags = k;
1218 skb_reset_tail_pointer(skb);
1219 skb->data_len -= len;
1220 skb->len = skb->data_len;
1223 /* Remove acked data from a packet in the transmit queue. */
1224 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1226 if (skb_unclone(skb, GFP_ATOMIC))
1229 __pskb_trim_head(skb, len);
1231 TCP_SKB_CB(skb)->seq += len;
1232 skb->ip_summed = CHECKSUM_PARTIAL;
1234 skb->truesize -= len;
1235 sk->sk_wmem_queued -= len;
1236 sk_mem_uncharge(sk, len);
1237 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1239 /* Any change of skb->len requires recalculation of tso factor. */
1240 if (tcp_skb_pcount(skb) > 1)
1241 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1246 /* Calculate MSS not accounting any TCP options. */
1247 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1249 const struct tcp_sock *tp = tcp_sk(sk);
1250 const struct inet_connection_sock *icsk = inet_csk(sk);
1253 /* Calculate base mss without TCP options:
1254 It is MMS_S - sizeof(tcphdr) of rfc1122
1256 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1258 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1259 if (icsk->icsk_af_ops->net_frag_header_len) {
1260 const struct dst_entry *dst = __sk_dst_get(sk);
1262 if (dst && dst_allfrag(dst))
1263 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1266 /* Clamp it (mss_clamp does not include tcp options) */
1267 if (mss_now > tp->rx_opt.mss_clamp)
1268 mss_now = tp->rx_opt.mss_clamp;
1270 /* Now subtract optional transport overhead */
1271 mss_now -= icsk->icsk_ext_hdr_len;
1273 /* Then reserve room for full set of TCP options and 8 bytes of data */
1279 /* Calculate MSS. Not accounting for SACKs here. */
1280 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1282 /* Subtract TCP options size, not including SACKs */
1283 return __tcp_mtu_to_mss(sk, pmtu) -
1284 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1287 /* Inverse of above */
1288 int tcp_mss_to_mtu(struct sock *sk, int mss)
1290 const struct tcp_sock *tp = tcp_sk(sk);
1291 const struct inet_connection_sock *icsk = inet_csk(sk);
1295 tp->tcp_header_len +
1296 icsk->icsk_ext_hdr_len +
1297 icsk->icsk_af_ops->net_header_len;
1299 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1300 if (icsk->icsk_af_ops->net_frag_header_len) {
1301 const struct dst_entry *dst = __sk_dst_get(sk);
1303 if (dst && dst_allfrag(dst))
1304 mtu += icsk->icsk_af_ops->net_frag_header_len;
1309 /* MTU probing init per socket */
1310 void tcp_mtup_init(struct sock *sk)
1312 struct tcp_sock *tp = tcp_sk(sk);
1313 struct inet_connection_sock *icsk = inet_csk(sk);
1315 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1316 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1317 icsk->icsk_af_ops->net_header_len;
1318 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1319 icsk->icsk_mtup.probe_size = 0;
1321 EXPORT_SYMBOL(tcp_mtup_init);
1323 /* This function synchronize snd mss to current pmtu/exthdr set.
1325 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1326 for TCP options, but includes only bare TCP header.
1328 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1329 It is minimum of user_mss and mss received with SYN.
1330 It also does not include TCP options.
1332 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1334 tp->mss_cache is current effective sending mss, including
1335 all tcp options except for SACKs. It is evaluated,
1336 taking into account current pmtu, but never exceeds
1337 tp->rx_opt.mss_clamp.
1339 NOTE1. rfc1122 clearly states that advertised MSS
1340 DOES NOT include either tcp or ip options.
1342 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1343 are READ ONLY outside this function. --ANK (980731)
1345 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1347 struct tcp_sock *tp = tcp_sk(sk);
1348 struct inet_connection_sock *icsk = inet_csk(sk);
1351 if (icsk->icsk_mtup.search_high > pmtu)
1352 icsk->icsk_mtup.search_high = pmtu;
1354 mss_now = tcp_mtu_to_mss(sk, pmtu);
1355 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1357 /* And store cached results */
1358 icsk->icsk_pmtu_cookie = pmtu;
1359 if (icsk->icsk_mtup.enabled)
1360 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1361 tp->mss_cache = mss_now;
1365 EXPORT_SYMBOL(tcp_sync_mss);
1367 /* Compute the current effective MSS, taking SACKs and IP options,
1368 * and even PMTU discovery events into account.
1370 unsigned int tcp_current_mss(struct sock *sk)
1372 const struct tcp_sock *tp = tcp_sk(sk);
1373 const struct dst_entry *dst = __sk_dst_get(sk);
1375 unsigned int header_len;
1376 struct tcp_out_options opts;
1377 struct tcp_md5sig_key *md5;
1379 mss_now = tp->mss_cache;
1382 u32 mtu = dst_mtu(dst);
1383 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1384 mss_now = tcp_sync_mss(sk, mtu);
1387 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1388 sizeof(struct tcphdr);
1389 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1390 * some common options. If this is an odd packet (because we have SACK
1391 * blocks etc) then our calculated header_len will be different, and
1392 * we have to adjust mss_now correspondingly */
1393 if (header_len != tp->tcp_header_len) {
1394 int delta = (int) header_len - tp->tcp_header_len;
1401 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1402 * As additional protections, we do not touch cwnd in retransmission phases,
1403 * and if application hit its sndbuf limit recently.
1405 static void tcp_cwnd_application_limited(struct sock *sk)
1407 struct tcp_sock *tp = tcp_sk(sk);
1409 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1410 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1411 /* Limited by application or receiver window. */
1412 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1413 u32 win_used = max(tp->snd_cwnd_used, init_win);
1414 if (win_used < tp->snd_cwnd) {
1415 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1416 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1418 tp->snd_cwnd_used = 0;
1420 tp->snd_cwnd_stamp = tcp_time_stamp;
1423 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1425 struct tcp_sock *tp = tcp_sk(sk);
1427 /* Track the maximum number of outstanding packets in each
1428 * window, and remember whether we were cwnd-limited then.
1430 if (!before(tp->snd_una, tp->max_packets_seq) ||
1431 tp->packets_out > tp->max_packets_out) {
1432 tp->max_packets_out = tp->packets_out;
1433 tp->max_packets_seq = tp->snd_nxt;
1434 tp->is_cwnd_limited = is_cwnd_limited;
1437 if (tcp_is_cwnd_limited(sk)) {
1438 /* Network is feed fully. */
1439 tp->snd_cwnd_used = 0;
1440 tp->snd_cwnd_stamp = tcp_time_stamp;
1442 /* Network starves. */
1443 if (tp->packets_out > tp->snd_cwnd_used)
1444 tp->snd_cwnd_used = tp->packets_out;
1446 if (sysctl_tcp_slow_start_after_idle &&
1447 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1448 tcp_cwnd_application_limited(sk);
1452 /* Minshall's variant of the Nagle send check. */
1453 static bool tcp_minshall_check(const struct tcp_sock *tp)
1455 return after(tp->snd_sml, tp->snd_una) &&
1456 !after(tp->snd_sml, tp->snd_nxt);
1459 /* Update snd_sml if this skb is under mss
1460 * Note that a TSO packet might end with a sub-mss segment
1461 * The test is really :
1462 * if ((skb->len % mss) != 0)
1463 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1464 * But we can avoid doing the divide again given we already have
1465 * skb_pcount = skb->len / mss_now
1467 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1468 const struct sk_buff *skb)
1470 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1471 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1474 /* Return false, if packet can be sent now without violation Nagle's rules:
1475 * 1. It is full sized. (provided by caller in %partial bool)
1476 * 2. Or it contains FIN. (already checked by caller)
1477 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1478 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1479 * With Minshall's modification: all sent small packets are ACKed.
1481 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1485 ((nonagle & TCP_NAGLE_CORK) ||
1486 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1488 /* Returns the portion of skb which can be sent right away */
1489 static unsigned int tcp_mss_split_point(const struct sock *sk,
1490 const struct sk_buff *skb,
1491 unsigned int mss_now,
1492 unsigned int max_segs,
1495 const struct tcp_sock *tp = tcp_sk(sk);
1496 u32 partial, needed, window, max_len;
1498 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1499 max_len = mss_now * max_segs;
1501 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1504 needed = min(skb->len, window);
1506 if (max_len <= needed)
1509 partial = needed % mss_now;
1510 /* If last segment is not a full MSS, check if Nagle rules allow us
1511 * to include this last segment in this skb.
1512 * Otherwise, we'll split the skb at last MSS boundary
1514 if (tcp_nagle_check(partial != 0, tp, nonagle))
1515 return needed - partial;
1520 /* Can at least one segment of SKB be sent right now, according to the
1521 * congestion window rules? If so, return how many segments are allowed.
1523 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1524 const struct sk_buff *skb)
1526 u32 in_flight, cwnd;
1528 /* Don't be strict about the congestion window for the final FIN. */
1529 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1530 tcp_skb_pcount(skb) == 1)
1533 in_flight = tcp_packets_in_flight(tp);
1534 cwnd = tp->snd_cwnd;
1535 if (in_flight < cwnd)
1536 return (cwnd - in_flight);
1541 /* Initialize TSO state of a skb.
1542 * This must be invoked the first time we consider transmitting
1543 * SKB onto the wire.
1545 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1546 unsigned int mss_now)
1548 int tso_segs = tcp_skb_pcount(skb);
1550 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1551 tcp_set_skb_tso_segs(sk, skb, mss_now);
1552 tso_segs = tcp_skb_pcount(skb);
1558 /* Return true if the Nagle test allows this packet to be
1561 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1562 unsigned int cur_mss, int nonagle)
1564 /* Nagle rule does not apply to frames, which sit in the middle of the
1565 * write_queue (they have no chances to get new data).
1567 * This is implemented in the callers, where they modify the 'nonagle'
1568 * argument based upon the location of SKB in the send queue.
1570 if (nonagle & TCP_NAGLE_PUSH)
1573 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1574 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1577 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1583 /* Does at least the first segment of SKB fit into the send window? */
1584 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1585 const struct sk_buff *skb,
1586 unsigned int cur_mss)
1588 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1590 if (skb->len > cur_mss)
1591 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1593 return !after(end_seq, tcp_wnd_end(tp));
1596 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1597 * should be put on the wire right now. If so, it returns the number of
1598 * packets allowed by the congestion window.
1600 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1601 unsigned int cur_mss, int nonagle)
1603 const struct tcp_sock *tp = tcp_sk(sk);
1604 unsigned int cwnd_quota;
1606 tcp_init_tso_segs(sk, skb, cur_mss);
1608 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1611 cwnd_quota = tcp_cwnd_test(tp, skb);
1612 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1618 /* Test if sending is allowed right now. */
1619 bool tcp_may_send_now(struct sock *sk)
1621 const struct tcp_sock *tp = tcp_sk(sk);
1622 struct sk_buff *skb = tcp_send_head(sk);
1625 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1626 (tcp_skb_is_last(sk, skb) ?
1627 tp->nonagle : TCP_NAGLE_PUSH));
1630 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1631 * which is put after SKB on the list. It is very much like
1632 * tcp_fragment() except that it may make several kinds of assumptions
1633 * in order to speed up the splitting operation. In particular, we
1634 * know that all the data is in scatter-gather pages, and that the
1635 * packet has never been sent out before (and thus is not cloned).
1637 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1638 unsigned int mss_now, gfp_t gfp)
1640 struct sk_buff *buff;
1641 int nlen = skb->len - len;
1644 /* All of a TSO frame must be composed of paged data. */
1645 if (skb->len != skb->data_len)
1646 return tcp_fragment(sk, skb, len, mss_now, gfp);
1648 buff = sk_stream_alloc_skb(sk, 0, gfp);
1649 if (unlikely(buff == NULL))
1652 sk->sk_wmem_queued += buff->truesize;
1653 sk_mem_charge(sk, buff->truesize);
1654 buff->truesize += nlen;
1655 skb->truesize -= nlen;
1657 /* Correct the sequence numbers. */
1658 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1659 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1660 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1662 /* PSH and FIN should only be set in the second packet. */
1663 flags = TCP_SKB_CB(skb)->tcp_flags;
1664 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1665 TCP_SKB_CB(buff)->tcp_flags = flags;
1667 /* This packet was never sent out yet, so no SACK bits. */
1668 TCP_SKB_CB(buff)->sacked = 0;
1670 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1671 skb_split(skb, buff, len);
1672 tcp_fragment_tstamp(skb, buff);
1674 /* Fix up tso_factor for both original and new SKB. */
1675 tcp_set_skb_tso_segs(sk, skb, mss_now);
1676 tcp_set_skb_tso_segs(sk, buff, mss_now);
1678 /* Link BUFF into the send queue. */
1679 __skb_header_release(buff);
1680 tcp_insert_write_queue_after(skb, buff, sk);
1685 /* Try to defer sending, if possible, in order to minimize the amount
1686 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1688 * This algorithm is from John Heffner.
1690 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1691 bool *is_cwnd_limited)
1693 struct tcp_sock *tp = tcp_sk(sk);
1694 const struct inet_connection_sock *icsk = inet_csk(sk);
1695 u32 send_win, cong_win, limit, in_flight;
1698 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1701 if (icsk->icsk_ca_state != TCP_CA_Open)
1704 /* Defer for less than two clock ticks. */
1705 if (tp->tso_deferred &&
1706 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1709 in_flight = tcp_packets_in_flight(tp);
1711 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1713 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1715 /* From in_flight test above, we know that cwnd > in_flight. */
1716 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1718 limit = min(send_win, cong_win);
1720 /* If a full-sized TSO skb can be sent, do it. */
1721 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1722 tp->xmit_size_goal_segs * tp->mss_cache))
1725 /* Middle in queue won't get any more data, full sendable already? */
1726 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1729 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1731 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1733 /* If at least some fraction of a window is available,
1736 chunk /= win_divisor;
1740 /* Different approach, try not to defer past a single
1741 * ACK. Receiver should ACK every other full sized
1742 * frame, so if we have space for more than 3 frames
1745 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1749 /* Ok, it looks like it is advisable to defer.
1750 * Do not rearm the timer if already set to not break TCP ACK clocking.
1752 if (!tp->tso_deferred)
1753 tp->tso_deferred = 1 | (jiffies << 1);
1755 if (cong_win < send_win && cong_win < skb->len)
1756 *is_cwnd_limited = true;
1761 tp->tso_deferred = 0;
1765 /* Create a new MTU probe if we are ready.
1766 * MTU probe is regularly attempting to increase the path MTU by
1767 * deliberately sending larger packets. This discovers routing
1768 * changes resulting in larger path MTUs.
1770 * Returns 0 if we should wait to probe (no cwnd available),
1771 * 1 if a probe was sent,
1774 static int tcp_mtu_probe(struct sock *sk)
1776 struct tcp_sock *tp = tcp_sk(sk);
1777 struct inet_connection_sock *icsk = inet_csk(sk);
1778 struct sk_buff *skb, *nskb, *next;
1785 /* Not currently probing/verifying,
1787 * have enough cwnd, and
1788 * not SACKing (the variable headers throw things off) */
1789 if (!icsk->icsk_mtup.enabled ||
1790 icsk->icsk_mtup.probe_size ||
1791 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1792 tp->snd_cwnd < 11 ||
1793 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1796 /* Very simple search strategy: just double the MSS. */
1797 mss_now = tcp_current_mss(sk);
1798 probe_size = 2 * tp->mss_cache;
1799 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1800 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1801 /* TODO: set timer for probe_converge_event */
1805 /* Have enough data in the send queue to probe? */
1806 if (tp->write_seq - tp->snd_nxt < size_needed)
1809 if (tp->snd_wnd < size_needed)
1811 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1814 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1815 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1816 if (!tcp_packets_in_flight(tp))
1822 /* We're allowed to probe. Build it now. */
1823 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1825 sk->sk_wmem_queued += nskb->truesize;
1826 sk_mem_charge(sk, nskb->truesize);
1828 skb = tcp_send_head(sk);
1830 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1831 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1832 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1833 TCP_SKB_CB(nskb)->sacked = 0;
1835 nskb->ip_summed = skb->ip_summed;
1837 tcp_insert_write_queue_before(nskb, skb, sk);
1840 tcp_for_write_queue_from_safe(skb, next, sk) {
1841 copy = min_t(int, skb->len, probe_size - len);
1842 if (nskb->ip_summed)
1843 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1845 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1846 skb_put(nskb, copy),
1849 if (skb->len <= copy) {
1850 /* We've eaten all the data from this skb.
1852 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1853 tcp_unlink_write_queue(skb, sk);
1854 sk_wmem_free_skb(sk, skb);
1856 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1857 ~(TCPHDR_FIN|TCPHDR_PSH);
1858 if (!skb_shinfo(skb)->nr_frags) {
1859 skb_pull(skb, copy);
1860 if (skb->ip_summed != CHECKSUM_PARTIAL)
1861 skb->csum = csum_partial(skb->data,
1864 __pskb_trim_head(skb, copy);
1865 tcp_set_skb_tso_segs(sk, skb, mss_now);
1867 TCP_SKB_CB(skb)->seq += copy;
1872 if (len >= probe_size)
1875 tcp_init_tso_segs(sk, nskb, nskb->len);
1877 /* We're ready to send. If this fails, the probe will
1878 * be resegmented into mss-sized pieces by tcp_write_xmit().
1880 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1881 /* Decrement cwnd here because we are sending
1882 * effectively two packets. */
1884 tcp_event_new_data_sent(sk, nskb);
1886 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1887 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1888 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1896 /* This routine writes packets to the network. It advances the
1897 * send_head. This happens as incoming acks open up the remote
1900 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1901 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1902 * account rare use of URG, this is not a big flaw.
1904 * Send at most one packet when push_one > 0. Temporarily ignore
1905 * cwnd limit to force at most one packet out when push_one == 2.
1907 * Returns true, if no segments are in flight and we have queued segments,
1908 * but cannot send anything now because of SWS or another problem.
1910 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1911 int push_one, gfp_t gfp)
1913 struct tcp_sock *tp = tcp_sk(sk);
1914 struct sk_buff *skb;
1915 unsigned int tso_segs, sent_pkts;
1918 bool is_cwnd_limited = false;
1923 /* Do MTU probing. */
1924 result = tcp_mtu_probe(sk);
1927 } else if (result > 0) {
1932 while ((skb = tcp_send_head(sk))) {
1935 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1938 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
1939 /* "skb_mstamp" is used as a start point for the retransmit timer */
1940 skb_mstamp_get(&skb->skb_mstamp);
1941 goto repair; /* Skip network transmission */
1944 cwnd_quota = tcp_cwnd_test(tp, skb);
1946 is_cwnd_limited = true;
1948 /* Force out a loss probe pkt. */
1954 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1957 if (tso_segs == 1) {
1958 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1959 (tcp_skb_is_last(sk, skb) ?
1960 nonagle : TCP_NAGLE_PUSH))))
1964 tcp_tso_should_defer(sk, skb, &is_cwnd_limited))
1968 /* TCP Small Queues :
1969 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1971 * - better RTT estimation and ACK scheduling
1974 * Alas, some drivers / subsystems require a fair amount
1975 * of queued bytes to ensure line rate.
1976 * One example is wifi aggregation (802.11 AMPDU)
1978 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1979 sk->sk_pacing_rate >> 10);
1981 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1982 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1983 /* It is possible TX completion already happened
1984 * before we set TSQ_THROTTLED, so we must
1985 * test again the condition.
1987 smp_mb__after_atomic();
1988 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1993 if (tso_segs > 1 && !tcp_urg_mode(tp))
1994 limit = tcp_mss_split_point(sk, skb, mss_now,
1997 sk->sk_gso_max_segs),
2000 if (skb->len > limit &&
2001 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2004 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2008 /* Advance the send_head. This one is sent out.
2009 * This call will increment packets_out.
2011 tcp_event_new_data_sent(sk, skb);
2013 tcp_minshall_update(tp, mss_now, skb);
2014 sent_pkts += tcp_skb_pcount(skb);
2020 if (likely(sent_pkts)) {
2021 if (tcp_in_cwnd_reduction(sk))
2022 tp->prr_out += sent_pkts;
2024 /* Send one loss probe per tail loss episode. */
2026 tcp_schedule_loss_probe(sk);
2027 tcp_cwnd_validate(sk, is_cwnd_limited);
2030 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2033 bool tcp_schedule_loss_probe(struct sock *sk)
2035 struct inet_connection_sock *icsk = inet_csk(sk);
2036 struct tcp_sock *tp = tcp_sk(sk);
2037 u32 timeout, tlp_time_stamp, rto_time_stamp;
2038 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2040 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2042 /* No consecutive loss probes. */
2043 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2047 /* Don't do any loss probe on a Fast Open connection before 3WHS
2050 if (sk->sk_state == TCP_SYN_RECV)
2053 /* TLP is only scheduled when next timer event is RTO. */
2054 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2057 /* Schedule a loss probe in 2*RTT for SACK capable connections
2058 * in Open state, that are either limited by cwnd or application.
2060 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2061 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2064 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2068 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2069 * for delayed ack when there's one outstanding packet.
2072 if (tp->packets_out == 1)
2073 timeout = max_t(u32, timeout,
2074 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2075 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2077 /* If RTO is shorter, just schedule TLP in its place. */
2078 tlp_time_stamp = tcp_time_stamp + timeout;
2079 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2080 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2081 s32 delta = rto_time_stamp - tcp_time_stamp;
2086 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2091 /* Thanks to skb fast clones, we can detect if a prior transmit of
2092 * a packet is still in a qdisc or driver queue.
2093 * In this case, there is very little point doing a retransmit !
2094 * Note: This is called from BH context only.
2096 static bool skb_still_in_host_queue(const struct sock *sk,
2097 const struct sk_buff *skb)
2099 const struct sk_buff *fclone = skb + 1;
2101 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
2102 fclone->fclone == SKB_FCLONE_CLONE)) {
2103 NET_INC_STATS_BH(sock_net(sk),
2104 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2110 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2111 * retransmit the last segment.
2113 void tcp_send_loss_probe(struct sock *sk)
2115 struct tcp_sock *tp = tcp_sk(sk);
2116 struct sk_buff *skb;
2118 int mss = tcp_current_mss(sk);
2121 if (tcp_send_head(sk) != NULL) {
2122 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2126 /* At most one outstanding TLP retransmission. */
2127 if (tp->tlp_high_seq)
2130 /* Retransmit last segment. */
2131 skb = tcp_write_queue_tail(sk);
2135 if (skb_still_in_host_queue(sk, skb))
2138 pcount = tcp_skb_pcount(skb);
2139 if (WARN_ON(!pcount))
2142 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2143 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2146 skb = tcp_write_queue_tail(sk);
2149 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2152 err = __tcp_retransmit_skb(sk, skb);
2154 /* Record snd_nxt for loss detection. */
2156 tp->tlp_high_seq = tp->snd_nxt;
2159 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2160 inet_csk(sk)->icsk_rto,
2164 NET_INC_STATS_BH(sock_net(sk),
2165 LINUX_MIB_TCPLOSSPROBES);
2168 /* Push out any pending frames which were held back due to
2169 * TCP_CORK or attempt at coalescing tiny packets.
2170 * The socket must be locked by the caller.
2172 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2175 /* If we are closed, the bytes will have to remain here.
2176 * In time closedown will finish, we empty the write queue and
2177 * all will be happy.
2179 if (unlikely(sk->sk_state == TCP_CLOSE))
2182 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2183 sk_gfp_atomic(sk, GFP_ATOMIC)))
2184 tcp_check_probe_timer(sk);
2187 /* Send _single_ skb sitting at the send head. This function requires
2188 * true push pending frames to setup probe timer etc.
2190 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2192 struct sk_buff *skb = tcp_send_head(sk);
2194 BUG_ON(!skb || skb->len < mss_now);
2196 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2199 /* This function returns the amount that we can raise the
2200 * usable window based on the following constraints
2202 * 1. The window can never be shrunk once it is offered (RFC 793)
2203 * 2. We limit memory per socket
2206 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2207 * RECV.NEXT + RCV.WIN fixed until:
2208 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2210 * i.e. don't raise the right edge of the window until you can raise
2211 * it at least MSS bytes.
2213 * Unfortunately, the recommended algorithm breaks header prediction,
2214 * since header prediction assumes th->window stays fixed.
2216 * Strictly speaking, keeping th->window fixed violates the receiver
2217 * side SWS prevention criteria. The problem is that under this rule
2218 * a stream of single byte packets will cause the right side of the
2219 * window to always advance by a single byte.
2221 * Of course, if the sender implements sender side SWS prevention
2222 * then this will not be a problem.
2224 * BSD seems to make the following compromise:
2226 * If the free space is less than the 1/4 of the maximum
2227 * space available and the free space is less than 1/2 mss,
2228 * then set the window to 0.
2229 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2230 * Otherwise, just prevent the window from shrinking
2231 * and from being larger than the largest representable value.
2233 * This prevents incremental opening of the window in the regime
2234 * where TCP is limited by the speed of the reader side taking
2235 * data out of the TCP receive queue. It does nothing about
2236 * those cases where the window is constrained on the sender side
2237 * because the pipeline is full.
2239 * BSD also seems to "accidentally" limit itself to windows that are a
2240 * multiple of MSS, at least until the free space gets quite small.
2241 * This would appear to be a side effect of the mbuf implementation.
2242 * Combining these two algorithms results in the observed behavior
2243 * of having a fixed window size at almost all times.
2245 * Below we obtain similar behavior by forcing the offered window to
2246 * a multiple of the mss when it is feasible to do so.
2248 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2249 * Regular options like TIMESTAMP are taken into account.
2251 u32 __tcp_select_window(struct sock *sk)
2253 struct inet_connection_sock *icsk = inet_csk(sk);
2254 struct tcp_sock *tp = tcp_sk(sk);
2255 /* MSS for the peer's data. Previous versions used mss_clamp
2256 * here. I don't know if the value based on our guesses
2257 * of peer's MSS is better for the performance. It's more correct
2258 * but may be worse for the performance because of rcv_mss
2259 * fluctuations. --SAW 1998/11/1
2261 int mss = icsk->icsk_ack.rcv_mss;
2262 int free_space = tcp_space(sk);
2263 int allowed_space = tcp_full_space(sk);
2264 int full_space = min_t(int, tp->window_clamp, allowed_space);
2267 if (mss > full_space)
2270 if (free_space < (full_space >> 1)) {
2271 icsk->icsk_ack.quick = 0;
2273 if (sk_under_memory_pressure(sk))
2274 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2277 /* free_space might become our new window, make sure we don't
2278 * increase it due to wscale.
2280 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2282 /* if free space is less than mss estimate, or is below 1/16th
2283 * of the maximum allowed, try to move to zero-window, else
2284 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2285 * new incoming data is dropped due to memory limits.
2286 * With large window, mss test triggers way too late in order
2287 * to announce zero window in time before rmem limit kicks in.
2289 if (free_space < (allowed_space >> 4) || free_space < mss)
2293 if (free_space > tp->rcv_ssthresh)
2294 free_space = tp->rcv_ssthresh;
2296 /* Don't do rounding if we are using window scaling, since the
2297 * scaled window will not line up with the MSS boundary anyway.
2299 window = tp->rcv_wnd;
2300 if (tp->rx_opt.rcv_wscale) {
2301 window = free_space;
2303 /* Advertise enough space so that it won't get scaled away.
2304 * Import case: prevent zero window announcement if
2305 * 1<<rcv_wscale > mss.
2307 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2308 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2309 << tp->rx_opt.rcv_wscale);
2311 /* Get the largest window that is a nice multiple of mss.
2312 * Window clamp already applied above.
2313 * If our current window offering is within 1 mss of the
2314 * free space we just keep it. This prevents the divide
2315 * and multiply from happening most of the time.
2316 * We also don't do any window rounding when the free space
2319 if (window <= free_space - mss || window > free_space)
2320 window = (free_space / mss) * mss;
2321 else if (mss == full_space &&
2322 free_space > window + (full_space >> 1))
2323 window = free_space;
2329 /* Collapses two adjacent SKB's during retransmission. */
2330 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2332 struct tcp_sock *tp = tcp_sk(sk);
2333 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2334 int skb_size, next_skb_size;
2336 skb_size = skb->len;
2337 next_skb_size = next_skb->len;
2339 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2341 tcp_highest_sack_combine(sk, next_skb, skb);
2343 tcp_unlink_write_queue(next_skb, sk);
2345 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2348 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2349 skb->ip_summed = CHECKSUM_PARTIAL;
2351 if (skb->ip_summed != CHECKSUM_PARTIAL)
2352 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2354 /* Update sequence range on original skb. */
2355 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2357 /* Merge over control information. This moves PSH/FIN etc. over */
2358 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2360 /* All done, get rid of second SKB and account for it so
2361 * packet counting does not break.
2363 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2365 /* changed transmit queue under us so clear hints */
2366 tcp_clear_retrans_hints_partial(tp);
2367 if (next_skb == tp->retransmit_skb_hint)
2368 tp->retransmit_skb_hint = skb;
2370 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2372 sk_wmem_free_skb(sk, next_skb);
2375 /* Check if coalescing SKBs is legal. */
2376 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2378 if (tcp_skb_pcount(skb) > 1)
2380 /* TODO: SACK collapsing could be used to remove this condition */
2381 if (skb_shinfo(skb)->nr_frags != 0)
2383 if (skb_cloned(skb))
2385 if (skb == tcp_send_head(sk))
2387 /* Some heurestics for collapsing over SACK'd could be invented */
2388 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2394 /* Collapse packets in the retransmit queue to make to create
2395 * less packets on the wire. This is only done on retransmission.
2397 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2400 struct tcp_sock *tp = tcp_sk(sk);
2401 struct sk_buff *skb = to, *tmp;
2404 if (!sysctl_tcp_retrans_collapse)
2406 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2409 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2410 if (!tcp_can_collapse(sk, skb))
2422 /* Punt if not enough space exists in the first SKB for
2423 * the data in the second
2425 if (skb->len > skb_availroom(to))
2428 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2431 tcp_collapse_retrans(sk, to);
2435 /* This retransmits one SKB. Policy decisions and retransmit queue
2436 * state updates are done by the caller. Returns non-zero if an
2437 * error occurred which prevented the send.
2439 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2441 struct tcp_sock *tp = tcp_sk(sk);
2442 struct inet_connection_sock *icsk = inet_csk(sk);
2443 unsigned int cur_mss;
2446 /* Inconslusive MTU probe */
2447 if (icsk->icsk_mtup.probe_size) {
2448 icsk->icsk_mtup.probe_size = 0;
2451 /* Do not sent more than we queued. 1/4 is reserved for possible
2452 * copying overhead: fragmentation, tunneling, mangling etc.
2454 if (atomic_read(&sk->sk_wmem_alloc) >
2455 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2458 if (skb_still_in_host_queue(sk, skb))
2461 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2462 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2464 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2468 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2469 return -EHOSTUNREACH; /* Routing failure or similar. */
2471 cur_mss = tcp_current_mss(sk);
2473 /* If receiver has shrunk his window, and skb is out of
2474 * new window, do not retransmit it. The exception is the
2475 * case, when window is shrunk to zero. In this case
2476 * our retransmit serves as a zero window probe.
2478 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2479 TCP_SKB_CB(skb)->seq != tp->snd_una)
2482 if (skb->len > cur_mss) {
2483 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2484 return -ENOMEM; /* We'll try again later. */
2486 int oldpcount = tcp_skb_pcount(skb);
2488 if (unlikely(oldpcount > 1)) {
2489 if (skb_unclone(skb, GFP_ATOMIC))
2491 tcp_init_tso_segs(sk, skb, cur_mss);
2492 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2496 tcp_retrans_try_collapse(sk, skb, cur_mss);
2498 /* Make a copy, if the first transmission SKB clone we made
2499 * is still in somebody's hands, else make a clone.
2502 /* make sure skb->data is aligned on arches that require it
2503 * and check if ack-trimming & collapsing extended the headroom
2504 * beyond what csum_start can cover.
2506 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2507 skb_headroom(skb) >= 0xFFFF)) {
2508 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2510 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2513 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2517 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2518 /* Update global TCP statistics. */
2519 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2520 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2521 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2522 tp->total_retrans++;
2527 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2529 struct tcp_sock *tp = tcp_sk(sk);
2530 int err = __tcp_retransmit_skb(sk, skb);
2533 #if FASTRETRANS_DEBUG > 0
2534 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2535 net_dbg_ratelimited("retrans_out leaked\n");
2538 if (!tp->retrans_out)
2539 tp->lost_retrans_low = tp->snd_nxt;
2540 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2541 tp->retrans_out += tcp_skb_pcount(skb);
2543 /* Save stamp of the first retransmit. */
2544 if (!tp->retrans_stamp)
2545 tp->retrans_stamp = tcp_skb_timestamp(skb);
2547 /* snd_nxt is stored to detect loss of retransmitted segment,
2548 * see tcp_input.c tcp_sacktag_write_queue().
2550 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2551 } else if (err != -EBUSY) {
2552 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2555 if (tp->undo_retrans < 0)
2556 tp->undo_retrans = 0;
2557 tp->undo_retrans += tcp_skb_pcount(skb);
2561 /* Check if we forward retransmits are possible in the current
2562 * window/congestion state.
2564 static bool tcp_can_forward_retransmit(struct sock *sk)
2566 const struct inet_connection_sock *icsk = inet_csk(sk);
2567 const struct tcp_sock *tp = tcp_sk(sk);
2569 /* Forward retransmissions are possible only during Recovery. */
2570 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2573 /* No forward retransmissions in Reno are possible. */
2574 if (tcp_is_reno(tp))
2577 /* Yeah, we have to make difficult choice between forward transmission
2578 * and retransmission... Both ways have their merits...
2580 * For now we do not retransmit anything, while we have some new
2581 * segments to send. In the other cases, follow rule 3 for
2582 * NextSeg() specified in RFC3517.
2585 if (tcp_may_send_now(sk))
2591 /* This gets called after a retransmit timeout, and the initially
2592 * retransmitted data is acknowledged. It tries to continue
2593 * resending the rest of the retransmit queue, until either
2594 * we've sent it all or the congestion window limit is reached.
2595 * If doing SACK, the first ACK which comes back for a timeout
2596 * based retransmit packet might feed us FACK information again.
2597 * If so, we use it to avoid unnecessarily retransmissions.
2599 void tcp_xmit_retransmit_queue(struct sock *sk)
2601 const struct inet_connection_sock *icsk = inet_csk(sk);
2602 struct tcp_sock *tp = tcp_sk(sk);
2603 struct sk_buff *skb;
2604 struct sk_buff *hole = NULL;
2607 int fwd_rexmitting = 0;
2609 if (!tp->packets_out)
2613 tp->retransmit_high = tp->snd_una;
2615 if (tp->retransmit_skb_hint) {
2616 skb = tp->retransmit_skb_hint;
2617 last_lost = TCP_SKB_CB(skb)->end_seq;
2618 if (after(last_lost, tp->retransmit_high))
2619 last_lost = tp->retransmit_high;
2621 skb = tcp_write_queue_head(sk);
2622 last_lost = tp->snd_una;
2625 tcp_for_write_queue_from(skb, sk) {
2626 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2628 if (skb == tcp_send_head(sk))
2630 /* we could do better than to assign each time */
2632 tp->retransmit_skb_hint = skb;
2634 /* Assume this retransmit will generate
2635 * only one packet for congestion window
2636 * calculation purposes. This works because
2637 * tcp_retransmit_skb() will chop up the
2638 * packet to be MSS sized and all the
2639 * packet counting works out.
2641 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2644 if (fwd_rexmitting) {
2646 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2648 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2650 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2651 tp->retransmit_high = last_lost;
2652 if (!tcp_can_forward_retransmit(sk))
2654 /* Backtrack if necessary to non-L'ed skb */
2662 } else if (!(sacked & TCPCB_LOST)) {
2663 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2668 last_lost = TCP_SKB_CB(skb)->end_seq;
2669 if (icsk->icsk_ca_state != TCP_CA_Loss)
2670 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2672 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2675 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2678 if (tcp_retransmit_skb(sk, skb))
2681 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2683 if (tcp_in_cwnd_reduction(sk))
2684 tp->prr_out += tcp_skb_pcount(skb);
2686 if (skb == tcp_write_queue_head(sk))
2687 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2688 inet_csk(sk)->icsk_rto,
2693 /* Send a fin. The caller locks the socket for us. This cannot be
2694 * allowed to fail queueing a FIN frame under any circumstances.
2696 void tcp_send_fin(struct sock *sk)
2698 struct tcp_sock *tp = tcp_sk(sk);
2699 struct sk_buff *skb = tcp_write_queue_tail(sk);
2702 /* Optimization, tack on the FIN if we have a queue of
2703 * unsent frames. But be careful about outgoing SACKS
2706 mss_now = tcp_current_mss(sk);
2708 if (tcp_send_head(sk) != NULL) {
2709 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2710 TCP_SKB_CB(skb)->end_seq++;
2713 /* Socket is locked, keep trying until memory is available. */
2715 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2722 /* Reserve space for headers and prepare control bits. */
2723 skb_reserve(skb, MAX_TCP_HEADER);
2724 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2725 tcp_init_nondata_skb(skb, tp->write_seq,
2726 TCPHDR_ACK | TCPHDR_FIN);
2727 tcp_queue_skb(sk, skb);
2729 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2732 /* We get here when a process closes a file descriptor (either due to
2733 * an explicit close() or as a byproduct of exit()'ing) and there
2734 * was unread data in the receive queue. This behavior is recommended
2735 * by RFC 2525, section 2.17. -DaveM
2737 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2739 struct sk_buff *skb;
2741 /* NOTE: No TCP options attached and we never retransmit this. */
2742 skb = alloc_skb(MAX_TCP_HEADER, priority);
2744 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2748 /* Reserve space for headers and prepare control bits. */
2749 skb_reserve(skb, MAX_TCP_HEADER);
2750 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2751 TCPHDR_ACK | TCPHDR_RST);
2753 if (tcp_transmit_skb(sk, skb, 0, priority))
2754 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2756 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2759 /* Send a crossed SYN-ACK during socket establishment.
2760 * WARNING: This routine must only be called when we have already sent
2761 * a SYN packet that crossed the incoming SYN that caused this routine
2762 * to get called. If this assumption fails then the initial rcv_wnd
2763 * and rcv_wscale values will not be correct.
2765 int tcp_send_synack(struct sock *sk)
2767 struct sk_buff *skb;
2769 skb = tcp_write_queue_head(sk);
2770 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2771 pr_debug("%s: wrong queue state\n", __func__);
2774 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2775 if (skb_cloned(skb)) {
2776 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2779 tcp_unlink_write_queue(skb, sk);
2780 __skb_header_release(nskb);
2781 __tcp_add_write_queue_head(sk, nskb);
2782 sk_wmem_free_skb(sk, skb);
2783 sk->sk_wmem_queued += nskb->truesize;
2784 sk_mem_charge(sk, nskb->truesize);
2788 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2789 TCP_ECN_send_synack(tcp_sk(sk), skb);
2791 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2795 * tcp_make_synack - Prepare a SYN-ACK.
2796 * sk: listener socket
2797 * dst: dst entry attached to the SYNACK
2798 * req: request_sock pointer
2800 * Allocate one skb and build a SYNACK packet.
2801 * @dst is consumed : Caller should not use it again.
2803 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2804 struct request_sock *req,
2805 struct tcp_fastopen_cookie *foc)
2807 struct tcp_out_options opts;
2808 struct inet_request_sock *ireq = inet_rsk(req);
2809 struct tcp_sock *tp = tcp_sk(sk);
2811 struct sk_buff *skb;
2812 struct tcp_md5sig_key *md5;
2813 int tcp_header_size;
2816 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2817 if (unlikely(!skb)) {
2821 /* Reserve space for headers. */
2822 skb_reserve(skb, MAX_TCP_HEADER);
2824 skb_dst_set(skb, dst);
2825 security_skb_owned_by(skb, sk);
2827 mss = dst_metric_advmss(dst);
2828 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2829 mss = tp->rx_opt.user_mss;
2831 memset(&opts, 0, sizeof(opts));
2832 #ifdef CONFIG_SYN_COOKIES
2833 if (unlikely(req->cookie_ts))
2834 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2837 skb_mstamp_get(&skb->skb_mstamp);
2838 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2841 skb_push(skb, tcp_header_size);
2842 skb_reset_transport_header(skb);
2845 memset(th, 0, sizeof(struct tcphdr));
2848 TCP_ECN_make_synack(req, th);
2849 th->source = htons(ireq->ir_num);
2850 th->dest = ireq->ir_rmt_port;
2851 /* Setting of flags are superfluous here for callers (and ECE is
2852 * not even correctly set)
2854 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2855 TCPHDR_SYN | TCPHDR_ACK);
2857 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2858 /* XXX data is queued and acked as is. No buffer/window check */
2859 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2861 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2862 th->window = htons(min(req->rcv_wnd, 65535U));
2863 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2864 th->doff = (tcp_header_size >> 2);
2865 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2867 #ifdef CONFIG_TCP_MD5SIG
2868 /* Okay, we have all we need - do the md5 hash if needed */
2870 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2871 md5, NULL, req, skb);
2877 EXPORT_SYMBOL(tcp_make_synack);
2879 /* Do all connect socket setups that can be done AF independent. */
2880 static void tcp_connect_init(struct sock *sk)
2882 const struct dst_entry *dst = __sk_dst_get(sk);
2883 struct tcp_sock *tp = tcp_sk(sk);
2886 /* We'll fix this up when we get a response from the other end.
2887 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2889 tp->tcp_header_len = sizeof(struct tcphdr) +
2890 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2892 #ifdef CONFIG_TCP_MD5SIG
2893 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2894 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2897 /* If user gave his TCP_MAXSEG, record it to clamp */
2898 if (tp->rx_opt.user_mss)
2899 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2902 tcp_sync_mss(sk, dst_mtu(dst));
2904 if (!tp->window_clamp)
2905 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2906 tp->advmss = dst_metric_advmss(dst);
2907 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2908 tp->advmss = tp->rx_opt.user_mss;
2910 tcp_initialize_rcv_mss(sk);
2912 /* limit the window selection if the user enforce a smaller rx buffer */
2913 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2914 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2915 tp->window_clamp = tcp_full_space(sk);
2917 tcp_select_initial_window(tcp_full_space(sk),
2918 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2921 sysctl_tcp_window_scaling,
2923 dst_metric(dst, RTAX_INITRWND));
2925 tp->rx_opt.rcv_wscale = rcv_wscale;
2926 tp->rcv_ssthresh = tp->rcv_wnd;
2929 sock_reset_flag(sk, SOCK_DONE);
2932 tp->snd_una = tp->write_seq;
2933 tp->snd_sml = tp->write_seq;
2934 tp->snd_up = tp->write_seq;
2935 tp->snd_nxt = tp->write_seq;
2937 if (likely(!tp->repair))
2940 tp->rcv_tstamp = tcp_time_stamp;
2941 tp->rcv_wup = tp->rcv_nxt;
2942 tp->copied_seq = tp->rcv_nxt;
2944 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2945 inet_csk(sk)->icsk_retransmits = 0;
2946 tcp_clear_retrans(tp);
2949 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2951 struct tcp_sock *tp = tcp_sk(sk);
2952 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2954 tcb->end_seq += skb->len;
2955 __skb_header_release(skb);
2956 __tcp_add_write_queue_tail(sk, skb);
2957 sk->sk_wmem_queued += skb->truesize;
2958 sk_mem_charge(sk, skb->truesize);
2959 tp->write_seq = tcb->end_seq;
2960 tp->packets_out += tcp_skb_pcount(skb);
2963 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2964 * queue a data-only packet after the regular SYN, such that regular SYNs
2965 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2966 * only the SYN sequence, the data are retransmitted in the first ACK.
2967 * If cookie is not cached or other error occurs, falls back to send a
2968 * regular SYN with Fast Open cookie request option.
2970 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2972 struct tcp_sock *tp = tcp_sk(sk);
2973 struct tcp_fastopen_request *fo = tp->fastopen_req;
2974 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2975 struct sk_buff *syn_data = NULL, *data;
2976 unsigned long last_syn_loss = 0;
2978 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2979 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2980 &syn_loss, &last_syn_loss);
2981 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2983 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2984 fo->cookie.len = -1;
2988 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2989 fo->cookie.len = -1;
2990 else if (fo->cookie.len <= 0)
2993 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2994 * user-MSS. Reserve maximum option space for middleboxes that add
2995 * private TCP options. The cost is reduced data space in SYN :(
2997 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2998 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2999 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3000 MAX_TCP_OPTION_SPACE;
3002 space = min_t(size_t, space, fo->size);
3004 /* limit to order-0 allocations */
3005 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3007 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
3009 if (syn_data == NULL)
3012 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3013 struct iovec *iov = &fo->data->msg_iov[i];
3014 unsigned char __user *from = iov->iov_base;
3015 int len = iov->iov_len;
3017 if (syn_data->len + len > space)
3018 len = space - syn_data->len;
3019 else if (i + 1 == iovlen)
3020 /* No more data pending in inet_wait_for_connect() */
3023 if (skb_add_data(syn_data, from, len))
3027 /* Queue a data-only packet after the regular SYN for retransmission */
3028 data = pskb_copy(syn_data, sk->sk_allocation);
3031 TCP_SKB_CB(data)->seq++;
3032 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3033 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3034 tcp_connect_queue_skb(sk, data);
3035 fo->copied = data->len;
3037 /* syn_data is about to be sent, we need to take current time stamps
3038 * for the packets that are in write queue : SYN packet and DATA
3040 skb_mstamp_get(&syn->skb_mstamp);
3041 data->skb_mstamp = syn->skb_mstamp;
3043 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
3044 tp->syn_data = (fo->copied > 0);
3045 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3051 /* Send a regular SYN with Fast Open cookie request option */
3052 if (fo->cookie.len > 0)
3054 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3056 tp->syn_fastopen = 0;
3057 kfree_skb(syn_data);
3059 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3063 /* Build a SYN and send it off. */
3064 int tcp_connect(struct sock *sk)
3066 struct tcp_sock *tp = tcp_sk(sk);
3067 struct sk_buff *buff;
3070 tcp_connect_init(sk);
3072 if (unlikely(tp->repair)) {
3073 tcp_finish_connect(sk, NULL);
3077 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3078 if (unlikely(buff == NULL))
3081 /* Reserve space for headers. */
3082 skb_reserve(buff, MAX_TCP_HEADER);
3084 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3085 tp->retrans_stamp = tcp_time_stamp;
3086 tcp_connect_queue_skb(sk, buff);
3087 TCP_ECN_send_syn(sk, buff);
3089 /* Send off SYN; include data in Fast Open. */
3090 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3091 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3092 if (err == -ECONNREFUSED)
3095 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3096 * in order to make this packet get counted in tcpOutSegs.
3098 tp->snd_nxt = tp->write_seq;
3099 tp->pushed_seq = tp->write_seq;
3100 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3102 /* Timer for repeating the SYN until an answer. */
3103 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3104 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3107 EXPORT_SYMBOL(tcp_connect);
3109 /* Send out a delayed ack, the caller does the policy checking
3110 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3113 void tcp_send_delayed_ack(struct sock *sk)
3115 struct inet_connection_sock *icsk = inet_csk(sk);
3116 int ato = icsk->icsk_ack.ato;
3117 unsigned long timeout;
3119 if (ato > TCP_DELACK_MIN) {
3120 const struct tcp_sock *tp = tcp_sk(sk);
3121 int max_ato = HZ / 2;
3123 if (icsk->icsk_ack.pingpong ||
3124 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3125 max_ato = TCP_DELACK_MAX;
3127 /* Slow path, intersegment interval is "high". */
3129 /* If some rtt estimate is known, use it to bound delayed ack.
3130 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3134 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3141 ato = min(ato, max_ato);
3144 /* Stay within the limit we were given */
3145 timeout = jiffies + ato;
3147 /* Use new timeout only if there wasn't a older one earlier. */
3148 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3149 /* If delack timer was blocked or is about to expire,
3152 if (icsk->icsk_ack.blocked ||
3153 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3158 if (!time_before(timeout, icsk->icsk_ack.timeout))
3159 timeout = icsk->icsk_ack.timeout;
3161 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3162 icsk->icsk_ack.timeout = timeout;
3163 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3166 /* This routine sends an ack and also updates the window. */
3167 void tcp_send_ack(struct sock *sk)
3169 struct sk_buff *buff;
3171 /* If we have been reset, we may not send again. */
3172 if (sk->sk_state == TCP_CLOSE)
3175 /* We are not putting this on the write queue, so
3176 * tcp_transmit_skb() will set the ownership to this
3179 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3181 inet_csk_schedule_ack(sk);
3182 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3183 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3184 TCP_DELACK_MAX, TCP_RTO_MAX);
3188 /* Reserve space for headers and prepare control bits. */
3189 skb_reserve(buff, MAX_TCP_HEADER);
3190 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3192 /* Send it off, this clears delayed acks for us. */
3193 skb_mstamp_get(&buff->skb_mstamp);
3194 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3197 /* This routine sends a packet with an out of date sequence
3198 * number. It assumes the other end will try to ack it.
3200 * Question: what should we make while urgent mode?
3201 * 4.4BSD forces sending single byte of data. We cannot send
3202 * out of window data, because we have SND.NXT==SND.MAX...
3204 * Current solution: to send TWO zero-length segments in urgent mode:
3205 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3206 * out-of-date with SND.UNA-1 to probe window.
3208 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3210 struct tcp_sock *tp = tcp_sk(sk);
3211 struct sk_buff *skb;
3213 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3214 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3218 /* Reserve space for headers and set control bits. */
3219 skb_reserve(skb, MAX_TCP_HEADER);
3220 /* Use a previous sequence. This should cause the other
3221 * end to send an ack. Don't queue or clone SKB, just
3224 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3225 skb_mstamp_get(&skb->skb_mstamp);
3226 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3229 void tcp_send_window_probe(struct sock *sk)
3231 if (sk->sk_state == TCP_ESTABLISHED) {
3232 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3233 tcp_xmit_probe_skb(sk, 0);
3237 /* Initiate keepalive or window probe from timer. */
3238 int tcp_write_wakeup(struct sock *sk)
3240 struct tcp_sock *tp = tcp_sk(sk);
3241 struct sk_buff *skb;
3243 if (sk->sk_state == TCP_CLOSE)
3246 if ((skb = tcp_send_head(sk)) != NULL &&
3247 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3249 unsigned int mss = tcp_current_mss(sk);
3250 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3252 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3253 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3255 /* We are probing the opening of a window
3256 * but the window size is != 0
3257 * must have been a result SWS avoidance ( sender )
3259 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3261 seg_size = min(seg_size, mss);
3262 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3263 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3265 } else if (!tcp_skb_pcount(skb))
3266 tcp_set_skb_tso_segs(sk, skb, mss);
3268 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3269 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3271 tcp_event_new_data_sent(sk, skb);
3274 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3275 tcp_xmit_probe_skb(sk, 1);
3276 return tcp_xmit_probe_skb(sk, 0);
3280 /* A window probe timeout has occurred. If window is not closed send
3281 * a partial packet else a zero probe.
3283 void tcp_send_probe0(struct sock *sk)
3285 struct inet_connection_sock *icsk = inet_csk(sk);
3286 struct tcp_sock *tp = tcp_sk(sk);
3287 unsigned long probe_max;
3290 err = tcp_write_wakeup(sk);
3292 if (tp->packets_out || !tcp_send_head(sk)) {
3293 /* Cancel probe timer, if it is not required. */
3294 icsk->icsk_probes_out = 0;
3295 icsk->icsk_backoff = 0;
3300 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3301 icsk->icsk_backoff++;
3302 icsk->icsk_probes_out++;
3303 probe_max = TCP_RTO_MAX;
3305 /* If packet was not sent due to local congestion,
3306 * do not backoff and do not remember icsk_probes_out.
3307 * Let local senders to fight for local resources.
3309 * Use accumulated backoff yet.
3311 if (!icsk->icsk_probes_out)
3312 icsk->icsk_probes_out = 1;
3313 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3315 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3316 inet_csk_rto_backoff(icsk, probe_max),
3320 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3322 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3326 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3328 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3329 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3333 EXPORT_SYMBOL(tcp_rtx_synack);