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;
977 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
979 if (likely(err <= 0))
984 return net_xmit_eval(err);
987 /* This routine just queues the buffer for sending.
989 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
990 * otherwise socket can stall.
992 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
994 struct tcp_sock *tp = tcp_sk(sk);
996 /* Advance write_seq and place onto the write_queue. */
997 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
998 skb_header_release(skb);
999 tcp_add_write_queue_tail(sk, skb);
1000 sk->sk_wmem_queued += skb->truesize;
1001 sk_mem_charge(sk, skb->truesize);
1004 /* Initialize TSO segments for a packet. */
1005 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1006 unsigned int mss_now)
1008 struct skb_shared_info *shinfo = skb_shinfo(skb);
1010 /* Make sure we own this skb before messing gso_size/gso_segs */
1011 WARN_ON_ONCE(skb_cloned(skb));
1013 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1014 /* Avoid the costly divide in the normal
1017 shinfo->gso_segs = 1;
1018 shinfo->gso_size = 0;
1019 shinfo->gso_type = 0;
1021 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1022 shinfo->gso_size = mss_now;
1023 shinfo->gso_type = sk->sk_gso_type;
1027 /* When a modification to fackets out becomes necessary, we need to check
1028 * skb is counted to fackets_out or not.
1030 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1033 struct tcp_sock *tp = tcp_sk(sk);
1035 if (!tp->sacked_out || tcp_is_reno(tp))
1038 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1039 tp->fackets_out -= decr;
1042 /* Pcount in the middle of the write queue got changed, we need to do various
1043 * tweaks to fix counters
1045 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1047 struct tcp_sock *tp = tcp_sk(sk);
1049 tp->packets_out -= decr;
1051 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1052 tp->sacked_out -= decr;
1053 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1054 tp->retrans_out -= decr;
1055 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1056 tp->lost_out -= decr;
1058 /* Reno case is special. Sigh... */
1059 if (tcp_is_reno(tp) && decr > 0)
1060 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1062 tcp_adjust_fackets_out(sk, skb, decr);
1064 if (tp->lost_skb_hint &&
1065 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1066 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1067 tp->lost_cnt_hint -= decr;
1069 tcp_verify_left_out(tp);
1072 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1074 struct skb_shared_info *shinfo = skb_shinfo(skb);
1076 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1077 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1078 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1079 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1081 shinfo->tx_flags &= ~tsflags;
1082 shinfo2->tx_flags |= tsflags;
1083 swap(shinfo->tskey, shinfo2->tskey);
1087 /* Function to create two new TCP segments. Shrinks the given segment
1088 * to the specified size and appends a new segment with the rest of the
1089 * packet to the list. This won't be called frequently, I hope.
1090 * Remember, these are still headerless SKBs at this point.
1092 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1093 unsigned int mss_now, gfp_t gfp)
1095 struct tcp_sock *tp = tcp_sk(sk);
1096 struct sk_buff *buff;
1097 int nsize, old_factor;
1101 if (WARN_ON(len > skb->len))
1104 nsize = skb_headlen(skb) - len;
1108 if (skb_unclone(skb, gfp))
1111 /* Get a new skb... force flag on. */
1112 buff = sk_stream_alloc_skb(sk, nsize, gfp);
1114 return -ENOMEM; /* We'll just try again later. */
1116 sk->sk_wmem_queued += buff->truesize;
1117 sk_mem_charge(sk, buff->truesize);
1118 nlen = skb->len - len - nsize;
1119 buff->truesize += nlen;
1120 skb->truesize -= nlen;
1122 /* Correct the sequence numbers. */
1123 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1124 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1125 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1127 /* PSH and FIN should only be set in the second packet. */
1128 flags = TCP_SKB_CB(skb)->tcp_flags;
1129 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1130 TCP_SKB_CB(buff)->tcp_flags = flags;
1131 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1133 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1134 /* Copy and checksum data tail into the new buffer. */
1135 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1136 skb_put(buff, nsize),
1141 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1143 skb->ip_summed = CHECKSUM_PARTIAL;
1144 skb_split(skb, buff, len);
1147 buff->ip_summed = skb->ip_summed;
1149 buff->tstamp = skb->tstamp;
1150 tcp_fragment_tstamp(skb, buff);
1152 old_factor = tcp_skb_pcount(skb);
1154 /* Fix up tso_factor for both original and new SKB. */
1155 tcp_set_skb_tso_segs(sk, skb, mss_now);
1156 tcp_set_skb_tso_segs(sk, buff, mss_now);
1158 /* If this packet has been sent out already, we must
1159 * adjust the various packet counters.
1161 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1162 int diff = old_factor - tcp_skb_pcount(skb) -
1163 tcp_skb_pcount(buff);
1166 tcp_adjust_pcount(sk, skb, diff);
1169 /* Link BUFF into the send queue. */
1170 skb_header_release(buff);
1171 tcp_insert_write_queue_after(skb, buff, sk);
1176 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1177 * eventually). The difference is that pulled data not copied, but
1178 * immediately discarded.
1180 static void __pskb_trim_head(struct sk_buff *skb, int len)
1182 struct skb_shared_info *shinfo;
1185 eat = min_t(int, len, skb_headlen(skb));
1187 __skb_pull(skb, eat);
1194 shinfo = skb_shinfo(skb);
1195 for (i = 0; i < shinfo->nr_frags; i++) {
1196 int size = skb_frag_size(&shinfo->frags[i]);
1199 skb_frag_unref(skb, i);
1202 shinfo->frags[k] = shinfo->frags[i];
1204 shinfo->frags[k].page_offset += eat;
1205 skb_frag_size_sub(&shinfo->frags[k], eat);
1211 shinfo->nr_frags = k;
1213 skb_reset_tail_pointer(skb);
1214 skb->data_len -= len;
1215 skb->len = skb->data_len;
1218 /* Remove acked data from a packet in the transmit queue. */
1219 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1221 if (skb_unclone(skb, GFP_ATOMIC))
1224 __pskb_trim_head(skb, len);
1226 TCP_SKB_CB(skb)->seq += len;
1227 skb->ip_summed = CHECKSUM_PARTIAL;
1229 skb->truesize -= len;
1230 sk->sk_wmem_queued -= len;
1231 sk_mem_uncharge(sk, len);
1232 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1234 /* Any change of skb->len requires recalculation of tso factor. */
1235 if (tcp_skb_pcount(skb) > 1)
1236 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1241 /* Calculate MSS not accounting any TCP options. */
1242 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1244 const struct tcp_sock *tp = tcp_sk(sk);
1245 const struct inet_connection_sock *icsk = inet_csk(sk);
1248 /* Calculate base mss without TCP options:
1249 It is MMS_S - sizeof(tcphdr) of rfc1122
1251 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1253 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1254 if (icsk->icsk_af_ops->net_frag_header_len) {
1255 const struct dst_entry *dst = __sk_dst_get(sk);
1257 if (dst && dst_allfrag(dst))
1258 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1261 /* Clamp it (mss_clamp does not include tcp options) */
1262 if (mss_now > tp->rx_opt.mss_clamp)
1263 mss_now = tp->rx_opt.mss_clamp;
1265 /* Now subtract optional transport overhead */
1266 mss_now -= icsk->icsk_ext_hdr_len;
1268 /* Then reserve room for full set of TCP options and 8 bytes of data */
1274 /* Calculate MSS. Not accounting for SACKs here. */
1275 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1277 /* Subtract TCP options size, not including SACKs */
1278 return __tcp_mtu_to_mss(sk, pmtu) -
1279 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1282 /* Inverse of above */
1283 int tcp_mss_to_mtu(struct sock *sk, int mss)
1285 const struct tcp_sock *tp = tcp_sk(sk);
1286 const struct inet_connection_sock *icsk = inet_csk(sk);
1290 tp->tcp_header_len +
1291 icsk->icsk_ext_hdr_len +
1292 icsk->icsk_af_ops->net_header_len;
1294 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1295 if (icsk->icsk_af_ops->net_frag_header_len) {
1296 const struct dst_entry *dst = __sk_dst_get(sk);
1298 if (dst && dst_allfrag(dst))
1299 mtu += icsk->icsk_af_ops->net_frag_header_len;
1304 /* MTU probing init per socket */
1305 void tcp_mtup_init(struct sock *sk)
1307 struct tcp_sock *tp = tcp_sk(sk);
1308 struct inet_connection_sock *icsk = inet_csk(sk);
1310 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1311 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1312 icsk->icsk_af_ops->net_header_len;
1313 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1314 icsk->icsk_mtup.probe_size = 0;
1316 EXPORT_SYMBOL(tcp_mtup_init);
1318 /* This function synchronize snd mss to current pmtu/exthdr set.
1320 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1321 for TCP options, but includes only bare TCP header.
1323 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1324 It is minimum of user_mss and mss received with SYN.
1325 It also does not include TCP options.
1327 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1329 tp->mss_cache is current effective sending mss, including
1330 all tcp options except for SACKs. It is evaluated,
1331 taking into account current pmtu, but never exceeds
1332 tp->rx_opt.mss_clamp.
1334 NOTE1. rfc1122 clearly states that advertised MSS
1335 DOES NOT include either tcp or ip options.
1337 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1338 are READ ONLY outside this function. --ANK (980731)
1340 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1342 struct tcp_sock *tp = tcp_sk(sk);
1343 struct inet_connection_sock *icsk = inet_csk(sk);
1346 if (icsk->icsk_mtup.search_high > pmtu)
1347 icsk->icsk_mtup.search_high = pmtu;
1349 mss_now = tcp_mtu_to_mss(sk, pmtu);
1350 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1352 /* And store cached results */
1353 icsk->icsk_pmtu_cookie = pmtu;
1354 if (icsk->icsk_mtup.enabled)
1355 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1356 tp->mss_cache = mss_now;
1360 EXPORT_SYMBOL(tcp_sync_mss);
1362 /* Compute the current effective MSS, taking SACKs and IP options,
1363 * and even PMTU discovery events into account.
1365 unsigned int tcp_current_mss(struct sock *sk)
1367 const struct tcp_sock *tp = tcp_sk(sk);
1368 const struct dst_entry *dst = __sk_dst_get(sk);
1370 unsigned int header_len;
1371 struct tcp_out_options opts;
1372 struct tcp_md5sig_key *md5;
1374 mss_now = tp->mss_cache;
1377 u32 mtu = dst_mtu(dst);
1378 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1379 mss_now = tcp_sync_mss(sk, mtu);
1382 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1383 sizeof(struct tcphdr);
1384 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1385 * some common options. If this is an odd packet (because we have SACK
1386 * blocks etc) then our calculated header_len will be different, and
1387 * we have to adjust mss_now correspondingly */
1388 if (header_len != tp->tcp_header_len) {
1389 int delta = (int) header_len - tp->tcp_header_len;
1396 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1397 * As additional protections, we do not touch cwnd in retransmission phases,
1398 * and if application hit its sndbuf limit recently.
1400 static void tcp_cwnd_application_limited(struct sock *sk)
1402 struct tcp_sock *tp = tcp_sk(sk);
1404 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1405 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1406 /* Limited by application or receiver window. */
1407 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1408 u32 win_used = max(tp->snd_cwnd_used, init_win);
1409 if (win_used < tp->snd_cwnd) {
1410 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1411 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1413 tp->snd_cwnd_used = 0;
1415 tp->snd_cwnd_stamp = tcp_time_stamp;
1418 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1420 struct tcp_sock *tp = tcp_sk(sk);
1422 /* Track the maximum number of outstanding packets in each
1423 * window, and remember whether we were cwnd-limited then.
1425 if (!before(tp->snd_una, tp->max_packets_seq) ||
1426 tp->packets_out > tp->max_packets_out) {
1427 tp->max_packets_out = tp->packets_out;
1428 tp->max_packets_seq = tp->snd_nxt;
1429 tp->is_cwnd_limited = is_cwnd_limited;
1432 if (tcp_is_cwnd_limited(sk)) {
1433 /* Network is feed fully. */
1434 tp->snd_cwnd_used = 0;
1435 tp->snd_cwnd_stamp = tcp_time_stamp;
1437 /* Network starves. */
1438 if (tp->packets_out > tp->snd_cwnd_used)
1439 tp->snd_cwnd_used = tp->packets_out;
1441 if (sysctl_tcp_slow_start_after_idle &&
1442 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1443 tcp_cwnd_application_limited(sk);
1447 /* Minshall's variant of the Nagle send check. */
1448 static bool tcp_minshall_check(const struct tcp_sock *tp)
1450 return after(tp->snd_sml, tp->snd_una) &&
1451 !after(tp->snd_sml, tp->snd_nxt);
1454 /* Update snd_sml if this skb is under mss
1455 * Note that a TSO packet might end with a sub-mss segment
1456 * The test is really :
1457 * if ((skb->len % mss) != 0)
1458 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1459 * But we can avoid doing the divide again given we already have
1460 * skb_pcount = skb->len / mss_now
1462 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1463 const struct sk_buff *skb)
1465 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1466 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1469 /* Return false, if packet can be sent now without violation Nagle's rules:
1470 * 1. It is full sized. (provided by caller in %partial bool)
1471 * 2. Or it contains FIN. (already checked by caller)
1472 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1473 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1474 * With Minshall's modification: all sent small packets are ACKed.
1476 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1480 ((nonagle & TCP_NAGLE_CORK) ||
1481 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1483 /* Returns the portion of skb which can be sent right away */
1484 static unsigned int tcp_mss_split_point(const struct sock *sk,
1485 const struct sk_buff *skb,
1486 unsigned int mss_now,
1487 unsigned int max_segs,
1490 const struct tcp_sock *tp = tcp_sk(sk);
1491 u32 partial, needed, window, max_len;
1493 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1494 max_len = mss_now * max_segs;
1496 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1499 needed = min(skb->len, window);
1501 if (max_len <= needed)
1504 partial = needed % mss_now;
1505 /* If last segment is not a full MSS, check if Nagle rules allow us
1506 * to include this last segment in this skb.
1507 * Otherwise, we'll split the skb at last MSS boundary
1509 if (tcp_nagle_check(partial != 0, tp, nonagle))
1510 return needed - partial;
1515 /* Can at least one segment of SKB be sent right now, according to the
1516 * congestion window rules? If so, return how many segments are allowed.
1518 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1519 const struct sk_buff *skb)
1521 u32 in_flight, cwnd;
1523 /* Don't be strict about the congestion window for the final FIN. */
1524 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1525 tcp_skb_pcount(skb) == 1)
1528 in_flight = tcp_packets_in_flight(tp);
1529 cwnd = tp->snd_cwnd;
1530 if (in_flight < cwnd)
1531 return (cwnd - in_flight);
1536 /* Initialize TSO state of a skb.
1537 * This must be invoked the first time we consider transmitting
1538 * SKB onto the wire.
1540 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1541 unsigned int mss_now)
1543 int tso_segs = tcp_skb_pcount(skb);
1545 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1546 tcp_set_skb_tso_segs(sk, skb, mss_now);
1547 tso_segs = tcp_skb_pcount(skb);
1553 /* Return true if the Nagle test allows this packet to be
1556 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1557 unsigned int cur_mss, int nonagle)
1559 /* Nagle rule does not apply to frames, which sit in the middle of the
1560 * write_queue (they have no chances to get new data).
1562 * This is implemented in the callers, where they modify the 'nonagle'
1563 * argument based upon the location of SKB in the send queue.
1565 if (nonagle & TCP_NAGLE_PUSH)
1568 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1569 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1572 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1578 /* Does at least the first segment of SKB fit into the send window? */
1579 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1580 const struct sk_buff *skb,
1581 unsigned int cur_mss)
1583 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1585 if (skb->len > cur_mss)
1586 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1588 return !after(end_seq, tcp_wnd_end(tp));
1591 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1592 * should be put on the wire right now. If so, it returns the number of
1593 * packets allowed by the congestion window.
1595 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1596 unsigned int cur_mss, int nonagle)
1598 const struct tcp_sock *tp = tcp_sk(sk);
1599 unsigned int cwnd_quota;
1601 tcp_init_tso_segs(sk, skb, cur_mss);
1603 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1606 cwnd_quota = tcp_cwnd_test(tp, skb);
1607 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1613 /* Test if sending is allowed right now. */
1614 bool tcp_may_send_now(struct sock *sk)
1616 const struct tcp_sock *tp = tcp_sk(sk);
1617 struct sk_buff *skb = tcp_send_head(sk);
1620 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1621 (tcp_skb_is_last(sk, skb) ?
1622 tp->nonagle : TCP_NAGLE_PUSH));
1625 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1626 * which is put after SKB on the list. It is very much like
1627 * tcp_fragment() except that it may make several kinds of assumptions
1628 * in order to speed up the splitting operation. In particular, we
1629 * know that all the data is in scatter-gather pages, and that the
1630 * packet has never been sent out before (and thus is not cloned).
1632 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1633 unsigned int mss_now, gfp_t gfp)
1635 struct sk_buff *buff;
1636 int nlen = skb->len - len;
1639 /* All of a TSO frame must be composed of paged data. */
1640 if (skb->len != skb->data_len)
1641 return tcp_fragment(sk, skb, len, mss_now, gfp);
1643 buff = sk_stream_alloc_skb(sk, 0, gfp);
1644 if (unlikely(buff == NULL))
1647 sk->sk_wmem_queued += buff->truesize;
1648 sk_mem_charge(sk, buff->truesize);
1649 buff->truesize += nlen;
1650 skb->truesize -= nlen;
1652 /* Correct the sequence numbers. */
1653 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1654 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1655 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1657 /* PSH and FIN should only be set in the second packet. */
1658 flags = TCP_SKB_CB(skb)->tcp_flags;
1659 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1660 TCP_SKB_CB(buff)->tcp_flags = flags;
1662 /* This packet was never sent out yet, so no SACK bits. */
1663 TCP_SKB_CB(buff)->sacked = 0;
1665 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1666 skb_split(skb, buff, len);
1667 tcp_fragment_tstamp(skb, buff);
1669 /* Fix up tso_factor for both original and new SKB. */
1670 tcp_set_skb_tso_segs(sk, skb, mss_now);
1671 tcp_set_skb_tso_segs(sk, buff, mss_now);
1673 /* Link BUFF into the send queue. */
1674 skb_header_release(buff);
1675 tcp_insert_write_queue_after(skb, buff, sk);
1680 /* Try to defer sending, if possible, in order to minimize the amount
1681 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1683 * This algorithm is from John Heffner.
1685 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1686 bool *is_cwnd_limited)
1688 struct tcp_sock *tp = tcp_sk(sk);
1689 const struct inet_connection_sock *icsk = inet_csk(sk);
1690 u32 send_win, cong_win, limit, in_flight;
1693 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1696 if (icsk->icsk_ca_state != TCP_CA_Open)
1699 /* Defer for less than two clock ticks. */
1700 if (tp->tso_deferred &&
1701 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1704 in_flight = tcp_packets_in_flight(tp);
1706 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1708 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1710 /* From in_flight test above, we know that cwnd > in_flight. */
1711 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1713 limit = min(send_win, cong_win);
1715 /* If a full-sized TSO skb can be sent, do it. */
1716 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1717 tp->xmit_size_goal_segs * tp->mss_cache))
1720 /* Middle in queue won't get any more data, full sendable already? */
1721 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1724 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1726 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1728 /* If at least some fraction of a window is available,
1731 chunk /= win_divisor;
1735 /* Different approach, try not to defer past a single
1736 * ACK. Receiver should ACK every other full sized
1737 * frame, so if we have space for more than 3 frames
1740 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1744 /* Ok, it looks like it is advisable to defer.
1745 * Do not rearm the timer if already set to not break TCP ACK clocking.
1747 if (!tp->tso_deferred)
1748 tp->tso_deferred = 1 | (jiffies << 1);
1750 if (cong_win < send_win && cong_win < skb->len)
1751 *is_cwnd_limited = true;
1756 tp->tso_deferred = 0;
1760 /* Create a new MTU probe if we are ready.
1761 * MTU probe is regularly attempting to increase the path MTU by
1762 * deliberately sending larger packets. This discovers routing
1763 * changes resulting in larger path MTUs.
1765 * Returns 0 if we should wait to probe (no cwnd available),
1766 * 1 if a probe was sent,
1769 static int tcp_mtu_probe(struct sock *sk)
1771 struct tcp_sock *tp = tcp_sk(sk);
1772 struct inet_connection_sock *icsk = inet_csk(sk);
1773 struct sk_buff *skb, *nskb, *next;
1780 /* Not currently probing/verifying,
1782 * have enough cwnd, and
1783 * not SACKing (the variable headers throw things off) */
1784 if (!icsk->icsk_mtup.enabled ||
1785 icsk->icsk_mtup.probe_size ||
1786 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1787 tp->snd_cwnd < 11 ||
1788 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1791 /* Very simple search strategy: just double the MSS. */
1792 mss_now = tcp_current_mss(sk);
1793 probe_size = 2 * tp->mss_cache;
1794 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1795 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1796 /* TODO: set timer for probe_converge_event */
1800 /* Have enough data in the send queue to probe? */
1801 if (tp->write_seq - tp->snd_nxt < size_needed)
1804 if (tp->snd_wnd < size_needed)
1806 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1809 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1810 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1811 if (!tcp_packets_in_flight(tp))
1817 /* We're allowed to probe. Build it now. */
1818 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1820 sk->sk_wmem_queued += nskb->truesize;
1821 sk_mem_charge(sk, nskb->truesize);
1823 skb = tcp_send_head(sk);
1825 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1826 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1827 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1828 TCP_SKB_CB(nskb)->sacked = 0;
1830 nskb->ip_summed = skb->ip_summed;
1832 tcp_insert_write_queue_before(nskb, skb, sk);
1835 tcp_for_write_queue_from_safe(skb, next, sk) {
1836 copy = min_t(int, skb->len, probe_size - len);
1837 if (nskb->ip_summed)
1838 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1840 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1841 skb_put(nskb, copy),
1844 if (skb->len <= copy) {
1845 /* We've eaten all the data from this skb.
1847 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1848 tcp_unlink_write_queue(skb, sk);
1849 sk_wmem_free_skb(sk, skb);
1851 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1852 ~(TCPHDR_FIN|TCPHDR_PSH);
1853 if (!skb_shinfo(skb)->nr_frags) {
1854 skb_pull(skb, copy);
1855 if (skb->ip_summed != CHECKSUM_PARTIAL)
1856 skb->csum = csum_partial(skb->data,
1859 __pskb_trim_head(skb, copy);
1860 tcp_set_skb_tso_segs(sk, skb, mss_now);
1862 TCP_SKB_CB(skb)->seq += copy;
1867 if (len >= probe_size)
1870 tcp_init_tso_segs(sk, nskb, nskb->len);
1872 /* We're ready to send. If this fails, the probe will
1873 * be resegmented into mss-sized pieces by tcp_write_xmit().
1875 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1876 /* Decrement cwnd here because we are sending
1877 * effectively two packets. */
1879 tcp_event_new_data_sent(sk, nskb);
1881 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1882 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1883 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1891 /* This routine writes packets to the network. It advances the
1892 * send_head. This happens as incoming acks open up the remote
1895 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1896 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1897 * account rare use of URG, this is not a big flaw.
1899 * Send at most one packet when push_one > 0. Temporarily ignore
1900 * cwnd limit to force at most one packet out when push_one == 2.
1902 * Returns true, if no segments are in flight and we have queued segments,
1903 * but cannot send anything now because of SWS or another problem.
1905 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1906 int push_one, gfp_t gfp)
1908 struct tcp_sock *tp = tcp_sk(sk);
1909 struct sk_buff *skb;
1910 unsigned int tso_segs, sent_pkts;
1913 bool is_cwnd_limited = false;
1918 /* Do MTU probing. */
1919 result = tcp_mtu_probe(sk);
1922 } else if (result > 0) {
1927 while ((skb = tcp_send_head(sk))) {
1930 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1933 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
1934 /* "skb_mstamp" is used as a start point for the retransmit timer */
1935 skb_mstamp_get(&skb->skb_mstamp);
1936 goto repair; /* Skip network transmission */
1939 cwnd_quota = tcp_cwnd_test(tp, skb);
1941 is_cwnd_limited = true;
1943 /* Force out a loss probe pkt. */
1949 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1952 if (tso_segs == 1) {
1953 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1954 (tcp_skb_is_last(sk, skb) ?
1955 nonagle : TCP_NAGLE_PUSH))))
1959 tcp_tso_should_defer(sk, skb, &is_cwnd_limited))
1963 /* TCP Small Queues :
1964 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1966 * - better RTT estimation and ACK scheduling
1969 * Alas, some drivers / subsystems require a fair amount
1970 * of queued bytes to ensure line rate.
1971 * One example is wifi aggregation (802.11 AMPDU)
1973 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1974 sk->sk_pacing_rate >> 10);
1976 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1977 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1978 /* It is possible TX completion already happened
1979 * before we set TSQ_THROTTLED, so we must
1980 * test again the condition.
1982 smp_mb__after_atomic();
1983 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1988 if (tso_segs > 1 && !tcp_urg_mode(tp))
1989 limit = tcp_mss_split_point(sk, skb, mss_now,
1992 sk->sk_gso_max_segs),
1995 if (skb->len > limit &&
1996 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1999 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2003 /* Advance the send_head. This one is sent out.
2004 * This call will increment packets_out.
2006 tcp_event_new_data_sent(sk, skb);
2008 tcp_minshall_update(tp, mss_now, skb);
2009 sent_pkts += tcp_skb_pcount(skb);
2015 if (likely(sent_pkts)) {
2016 if (tcp_in_cwnd_reduction(sk))
2017 tp->prr_out += sent_pkts;
2019 /* Send one loss probe per tail loss episode. */
2021 tcp_schedule_loss_probe(sk);
2022 tcp_cwnd_validate(sk, is_cwnd_limited);
2025 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
2028 bool tcp_schedule_loss_probe(struct sock *sk)
2030 struct inet_connection_sock *icsk = inet_csk(sk);
2031 struct tcp_sock *tp = tcp_sk(sk);
2032 u32 timeout, tlp_time_stamp, rto_time_stamp;
2033 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2035 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2037 /* No consecutive loss probes. */
2038 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2042 /* Don't do any loss probe on a Fast Open connection before 3WHS
2045 if (sk->sk_state == TCP_SYN_RECV)
2048 /* TLP is only scheduled when next timer event is RTO. */
2049 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2052 /* Schedule a loss probe in 2*RTT for SACK capable connections
2053 * in Open state, that are either limited by cwnd or application.
2055 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2056 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2059 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2063 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2064 * for delayed ack when there's one outstanding packet.
2067 if (tp->packets_out == 1)
2068 timeout = max_t(u32, timeout,
2069 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2070 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2072 /* If RTO is shorter, just schedule TLP in its place. */
2073 tlp_time_stamp = tcp_time_stamp + timeout;
2074 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2075 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2076 s32 delta = rto_time_stamp - tcp_time_stamp;
2081 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2086 /* Thanks to skb fast clones, we can detect if a prior transmit of
2087 * a packet is still in a qdisc or driver queue.
2088 * In this case, there is very little point doing a retransmit !
2089 * Note: This is called from BH context only.
2091 static bool skb_still_in_host_queue(const struct sock *sk,
2092 const struct sk_buff *skb)
2094 const struct sk_buff *fclone = skb + 1;
2096 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
2097 fclone->fclone == SKB_FCLONE_CLONE)) {
2098 NET_INC_STATS_BH(sock_net(sk),
2099 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2105 /* When probe timeout (PTO) fires, send a new segment if one exists, else
2106 * retransmit the last segment.
2108 void tcp_send_loss_probe(struct sock *sk)
2110 struct tcp_sock *tp = tcp_sk(sk);
2111 struct sk_buff *skb;
2113 int mss = tcp_current_mss(sk);
2116 if (tcp_send_head(sk) != NULL) {
2117 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2121 /* At most one outstanding TLP retransmission. */
2122 if (tp->tlp_high_seq)
2125 /* Retransmit last segment. */
2126 skb = tcp_write_queue_tail(sk);
2130 if (skb_still_in_host_queue(sk, skb))
2133 pcount = tcp_skb_pcount(skb);
2134 if (WARN_ON(!pcount))
2137 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2138 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2141 skb = tcp_write_queue_tail(sk);
2144 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2147 err = __tcp_retransmit_skb(sk, skb);
2149 /* Record snd_nxt for loss detection. */
2151 tp->tlp_high_seq = tp->snd_nxt;
2154 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2155 inet_csk(sk)->icsk_rto,
2159 NET_INC_STATS_BH(sock_net(sk),
2160 LINUX_MIB_TCPLOSSPROBES);
2163 /* Push out any pending frames which were held back due to
2164 * TCP_CORK or attempt at coalescing tiny packets.
2165 * The socket must be locked by the caller.
2167 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2170 /* If we are closed, the bytes will have to remain here.
2171 * In time closedown will finish, we empty the write queue and
2172 * all will be happy.
2174 if (unlikely(sk->sk_state == TCP_CLOSE))
2177 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2178 sk_gfp_atomic(sk, GFP_ATOMIC)))
2179 tcp_check_probe_timer(sk);
2182 /* Send _single_ skb sitting at the send head. This function requires
2183 * true push pending frames to setup probe timer etc.
2185 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2187 struct sk_buff *skb = tcp_send_head(sk);
2189 BUG_ON(!skb || skb->len < mss_now);
2191 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2194 /* This function returns the amount that we can raise the
2195 * usable window based on the following constraints
2197 * 1. The window can never be shrunk once it is offered (RFC 793)
2198 * 2. We limit memory per socket
2201 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2202 * RECV.NEXT + RCV.WIN fixed until:
2203 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2205 * i.e. don't raise the right edge of the window until you can raise
2206 * it at least MSS bytes.
2208 * Unfortunately, the recommended algorithm breaks header prediction,
2209 * since header prediction assumes th->window stays fixed.
2211 * Strictly speaking, keeping th->window fixed violates the receiver
2212 * side SWS prevention criteria. The problem is that under this rule
2213 * a stream of single byte packets will cause the right side of the
2214 * window to always advance by a single byte.
2216 * Of course, if the sender implements sender side SWS prevention
2217 * then this will not be a problem.
2219 * BSD seems to make the following compromise:
2221 * If the free space is less than the 1/4 of the maximum
2222 * space available and the free space is less than 1/2 mss,
2223 * then set the window to 0.
2224 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2225 * Otherwise, just prevent the window from shrinking
2226 * and from being larger than the largest representable value.
2228 * This prevents incremental opening of the window in the regime
2229 * where TCP is limited by the speed of the reader side taking
2230 * data out of the TCP receive queue. It does nothing about
2231 * those cases where the window is constrained on the sender side
2232 * because the pipeline is full.
2234 * BSD also seems to "accidentally" limit itself to windows that are a
2235 * multiple of MSS, at least until the free space gets quite small.
2236 * This would appear to be a side effect of the mbuf implementation.
2237 * Combining these two algorithms results in the observed behavior
2238 * of having a fixed window size at almost all times.
2240 * Below we obtain similar behavior by forcing the offered window to
2241 * a multiple of the mss when it is feasible to do so.
2243 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2244 * Regular options like TIMESTAMP are taken into account.
2246 u32 __tcp_select_window(struct sock *sk)
2248 struct inet_connection_sock *icsk = inet_csk(sk);
2249 struct tcp_sock *tp = tcp_sk(sk);
2250 /* MSS for the peer's data. Previous versions used mss_clamp
2251 * here. I don't know if the value based on our guesses
2252 * of peer's MSS is better for the performance. It's more correct
2253 * but may be worse for the performance because of rcv_mss
2254 * fluctuations. --SAW 1998/11/1
2256 int mss = icsk->icsk_ack.rcv_mss;
2257 int free_space = tcp_space(sk);
2258 int allowed_space = tcp_full_space(sk);
2259 int full_space = min_t(int, tp->window_clamp, allowed_space);
2262 if (mss > full_space)
2265 if (free_space < (full_space >> 1)) {
2266 icsk->icsk_ack.quick = 0;
2268 if (sk_under_memory_pressure(sk))
2269 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2272 /* free_space might become our new window, make sure we don't
2273 * increase it due to wscale.
2275 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2277 /* if free space is less than mss estimate, or is below 1/16th
2278 * of the maximum allowed, try to move to zero-window, else
2279 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2280 * new incoming data is dropped due to memory limits.
2281 * With large window, mss test triggers way too late in order
2282 * to announce zero window in time before rmem limit kicks in.
2284 if (free_space < (allowed_space >> 4) || free_space < mss)
2288 if (free_space > tp->rcv_ssthresh)
2289 free_space = tp->rcv_ssthresh;
2291 /* Don't do rounding if we are using window scaling, since the
2292 * scaled window will not line up with the MSS boundary anyway.
2294 window = tp->rcv_wnd;
2295 if (tp->rx_opt.rcv_wscale) {
2296 window = free_space;
2298 /* Advertise enough space so that it won't get scaled away.
2299 * Import case: prevent zero window announcement if
2300 * 1<<rcv_wscale > mss.
2302 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2303 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2304 << tp->rx_opt.rcv_wscale);
2306 /* Get the largest window that is a nice multiple of mss.
2307 * Window clamp already applied above.
2308 * If our current window offering is within 1 mss of the
2309 * free space we just keep it. This prevents the divide
2310 * and multiply from happening most of the time.
2311 * We also don't do any window rounding when the free space
2314 if (window <= free_space - mss || window > free_space)
2315 window = (free_space / mss) * mss;
2316 else if (mss == full_space &&
2317 free_space > window + (full_space >> 1))
2318 window = free_space;
2324 /* Collapses two adjacent SKB's during retransmission. */
2325 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2327 struct tcp_sock *tp = tcp_sk(sk);
2328 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2329 int skb_size, next_skb_size;
2331 skb_size = skb->len;
2332 next_skb_size = next_skb->len;
2334 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2336 tcp_highest_sack_combine(sk, next_skb, skb);
2338 tcp_unlink_write_queue(next_skb, sk);
2340 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2343 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2344 skb->ip_summed = CHECKSUM_PARTIAL;
2346 if (skb->ip_summed != CHECKSUM_PARTIAL)
2347 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2349 /* Update sequence range on original skb. */
2350 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2352 /* Merge over control information. This moves PSH/FIN etc. over */
2353 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2355 /* All done, get rid of second SKB and account for it so
2356 * packet counting does not break.
2358 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2360 /* changed transmit queue under us so clear hints */
2361 tcp_clear_retrans_hints_partial(tp);
2362 if (next_skb == tp->retransmit_skb_hint)
2363 tp->retransmit_skb_hint = skb;
2365 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2367 sk_wmem_free_skb(sk, next_skb);
2370 /* Check if coalescing SKBs is legal. */
2371 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2373 if (tcp_skb_pcount(skb) > 1)
2375 /* TODO: SACK collapsing could be used to remove this condition */
2376 if (skb_shinfo(skb)->nr_frags != 0)
2378 if (skb_cloned(skb))
2380 if (skb == tcp_send_head(sk))
2382 /* Some heurestics for collapsing over SACK'd could be invented */
2383 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2389 /* Collapse packets in the retransmit queue to make to create
2390 * less packets on the wire. This is only done on retransmission.
2392 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2395 struct tcp_sock *tp = tcp_sk(sk);
2396 struct sk_buff *skb = to, *tmp;
2399 if (!sysctl_tcp_retrans_collapse)
2401 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2404 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2405 if (!tcp_can_collapse(sk, skb))
2417 /* Punt if not enough space exists in the first SKB for
2418 * the data in the second
2420 if (skb->len > skb_availroom(to))
2423 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2426 tcp_collapse_retrans(sk, to);
2430 /* This retransmits one SKB. Policy decisions and retransmit queue
2431 * state updates are done by the caller. Returns non-zero if an
2432 * error occurred which prevented the send.
2434 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2436 struct tcp_sock *tp = tcp_sk(sk);
2437 struct inet_connection_sock *icsk = inet_csk(sk);
2438 unsigned int cur_mss;
2441 /* Inconslusive MTU probe */
2442 if (icsk->icsk_mtup.probe_size) {
2443 icsk->icsk_mtup.probe_size = 0;
2446 /* Do not sent more than we queued. 1/4 is reserved for possible
2447 * copying overhead: fragmentation, tunneling, mangling etc.
2449 if (atomic_read(&sk->sk_wmem_alloc) >
2450 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2453 if (skb_still_in_host_queue(sk, skb))
2456 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2457 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2459 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2463 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2464 return -EHOSTUNREACH; /* Routing failure or similar. */
2466 cur_mss = tcp_current_mss(sk);
2468 /* If receiver has shrunk his window, and skb is out of
2469 * new window, do not retransmit it. The exception is the
2470 * case, when window is shrunk to zero. In this case
2471 * our retransmit serves as a zero window probe.
2473 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2474 TCP_SKB_CB(skb)->seq != tp->snd_una)
2477 if (skb->len > cur_mss) {
2478 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2479 return -ENOMEM; /* We'll try again later. */
2481 int oldpcount = tcp_skb_pcount(skb);
2483 if (unlikely(oldpcount > 1)) {
2484 if (skb_unclone(skb, GFP_ATOMIC))
2486 tcp_init_tso_segs(sk, skb, cur_mss);
2487 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2491 tcp_retrans_try_collapse(sk, skb, cur_mss);
2493 /* Make a copy, if the first transmission SKB clone we made
2494 * is still in somebody's hands, else make a clone.
2497 /* make sure skb->data is aligned on arches that require it
2498 * and check if ack-trimming & collapsing extended the headroom
2499 * beyond what csum_start can cover.
2501 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2502 skb_headroom(skb) >= 0xFFFF)) {
2503 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2505 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2508 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2512 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2513 /* Update global TCP statistics. */
2514 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2515 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2516 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2517 tp->total_retrans++;
2522 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2524 struct tcp_sock *tp = tcp_sk(sk);
2525 int err = __tcp_retransmit_skb(sk, skb);
2528 #if FASTRETRANS_DEBUG > 0
2529 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2530 net_dbg_ratelimited("retrans_out leaked\n");
2533 if (!tp->retrans_out)
2534 tp->lost_retrans_low = tp->snd_nxt;
2535 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2536 tp->retrans_out += tcp_skb_pcount(skb);
2538 /* Save stamp of the first retransmit. */
2539 if (!tp->retrans_stamp)
2540 tp->retrans_stamp = tcp_skb_timestamp(skb);
2542 /* snd_nxt is stored to detect loss of retransmitted segment,
2543 * see tcp_input.c tcp_sacktag_write_queue().
2545 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2546 } else if (err != -EBUSY) {
2547 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2550 if (tp->undo_retrans < 0)
2551 tp->undo_retrans = 0;
2552 tp->undo_retrans += tcp_skb_pcount(skb);
2556 /* Check if we forward retransmits are possible in the current
2557 * window/congestion state.
2559 static bool tcp_can_forward_retransmit(struct sock *sk)
2561 const struct inet_connection_sock *icsk = inet_csk(sk);
2562 const struct tcp_sock *tp = tcp_sk(sk);
2564 /* Forward retransmissions are possible only during Recovery. */
2565 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2568 /* No forward retransmissions in Reno are possible. */
2569 if (tcp_is_reno(tp))
2572 /* Yeah, we have to make difficult choice between forward transmission
2573 * and retransmission... Both ways have their merits...
2575 * For now we do not retransmit anything, while we have some new
2576 * segments to send. In the other cases, follow rule 3 for
2577 * NextSeg() specified in RFC3517.
2580 if (tcp_may_send_now(sk))
2586 /* This gets called after a retransmit timeout, and the initially
2587 * retransmitted data is acknowledged. It tries to continue
2588 * resending the rest of the retransmit queue, until either
2589 * we've sent it all or the congestion window limit is reached.
2590 * If doing SACK, the first ACK which comes back for a timeout
2591 * based retransmit packet might feed us FACK information again.
2592 * If so, we use it to avoid unnecessarily retransmissions.
2594 void tcp_xmit_retransmit_queue(struct sock *sk)
2596 const struct inet_connection_sock *icsk = inet_csk(sk);
2597 struct tcp_sock *tp = tcp_sk(sk);
2598 struct sk_buff *skb;
2599 struct sk_buff *hole = NULL;
2602 int fwd_rexmitting = 0;
2604 if (!tp->packets_out)
2608 tp->retransmit_high = tp->snd_una;
2610 if (tp->retransmit_skb_hint) {
2611 skb = tp->retransmit_skb_hint;
2612 last_lost = TCP_SKB_CB(skb)->end_seq;
2613 if (after(last_lost, tp->retransmit_high))
2614 last_lost = tp->retransmit_high;
2616 skb = tcp_write_queue_head(sk);
2617 last_lost = tp->snd_una;
2620 tcp_for_write_queue_from(skb, sk) {
2621 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2623 if (skb == tcp_send_head(sk))
2625 /* we could do better than to assign each time */
2627 tp->retransmit_skb_hint = skb;
2629 /* Assume this retransmit will generate
2630 * only one packet for congestion window
2631 * calculation purposes. This works because
2632 * tcp_retransmit_skb() will chop up the
2633 * packet to be MSS sized and all the
2634 * packet counting works out.
2636 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2639 if (fwd_rexmitting) {
2641 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2643 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2645 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2646 tp->retransmit_high = last_lost;
2647 if (!tcp_can_forward_retransmit(sk))
2649 /* Backtrack if necessary to non-L'ed skb */
2657 } else if (!(sacked & TCPCB_LOST)) {
2658 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2663 last_lost = TCP_SKB_CB(skb)->end_seq;
2664 if (icsk->icsk_ca_state != TCP_CA_Loss)
2665 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2667 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2670 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2673 if (tcp_retransmit_skb(sk, skb))
2676 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2678 if (tcp_in_cwnd_reduction(sk))
2679 tp->prr_out += tcp_skb_pcount(skb);
2681 if (skb == tcp_write_queue_head(sk))
2682 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2683 inet_csk(sk)->icsk_rto,
2688 /* Send a fin. The caller locks the socket for us. This cannot be
2689 * allowed to fail queueing a FIN frame under any circumstances.
2691 void tcp_send_fin(struct sock *sk)
2693 struct tcp_sock *tp = tcp_sk(sk);
2694 struct sk_buff *skb = tcp_write_queue_tail(sk);
2697 /* Optimization, tack on the FIN if we have a queue of
2698 * unsent frames. But be careful about outgoing SACKS
2701 mss_now = tcp_current_mss(sk);
2703 if (tcp_send_head(sk) != NULL) {
2704 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2705 TCP_SKB_CB(skb)->end_seq++;
2708 /* Socket is locked, keep trying until memory is available. */
2710 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2717 /* Reserve space for headers and prepare control bits. */
2718 skb_reserve(skb, MAX_TCP_HEADER);
2719 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2720 tcp_init_nondata_skb(skb, tp->write_seq,
2721 TCPHDR_ACK | TCPHDR_FIN);
2722 tcp_queue_skb(sk, skb);
2724 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2727 /* We get here when a process closes a file descriptor (either due to
2728 * an explicit close() or as a byproduct of exit()'ing) and there
2729 * was unread data in the receive queue. This behavior is recommended
2730 * by RFC 2525, section 2.17. -DaveM
2732 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2734 struct sk_buff *skb;
2736 /* NOTE: No TCP options attached and we never retransmit this. */
2737 skb = alloc_skb(MAX_TCP_HEADER, priority);
2739 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2743 /* Reserve space for headers and prepare control bits. */
2744 skb_reserve(skb, MAX_TCP_HEADER);
2745 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2746 TCPHDR_ACK | TCPHDR_RST);
2748 if (tcp_transmit_skb(sk, skb, 0, priority))
2749 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2751 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2754 /* Send a crossed SYN-ACK during socket establishment.
2755 * WARNING: This routine must only be called when we have already sent
2756 * a SYN packet that crossed the incoming SYN that caused this routine
2757 * to get called. If this assumption fails then the initial rcv_wnd
2758 * and rcv_wscale values will not be correct.
2760 int tcp_send_synack(struct sock *sk)
2762 struct sk_buff *skb;
2764 skb = tcp_write_queue_head(sk);
2765 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2766 pr_debug("%s: wrong queue state\n", __func__);
2769 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2770 if (skb_cloned(skb)) {
2771 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2774 tcp_unlink_write_queue(skb, sk);
2775 skb_header_release(nskb);
2776 __tcp_add_write_queue_head(sk, nskb);
2777 sk_wmem_free_skb(sk, skb);
2778 sk->sk_wmem_queued += nskb->truesize;
2779 sk_mem_charge(sk, nskb->truesize);
2783 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2784 TCP_ECN_send_synack(tcp_sk(sk), skb);
2786 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2790 * tcp_make_synack - Prepare a SYN-ACK.
2791 * sk: listener socket
2792 * dst: dst entry attached to the SYNACK
2793 * req: request_sock pointer
2795 * Allocate one skb and build a SYNACK packet.
2796 * @dst is consumed : Caller should not use it again.
2798 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2799 struct request_sock *req,
2800 struct tcp_fastopen_cookie *foc)
2802 struct tcp_out_options opts;
2803 struct inet_request_sock *ireq = inet_rsk(req);
2804 struct tcp_sock *tp = tcp_sk(sk);
2806 struct sk_buff *skb;
2807 struct tcp_md5sig_key *md5;
2808 int tcp_header_size;
2811 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2812 if (unlikely(!skb)) {
2816 /* Reserve space for headers. */
2817 skb_reserve(skb, MAX_TCP_HEADER);
2819 skb_dst_set(skb, dst);
2820 security_skb_owned_by(skb, sk);
2822 mss = dst_metric_advmss(dst);
2823 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2824 mss = tp->rx_opt.user_mss;
2826 memset(&opts, 0, sizeof(opts));
2827 #ifdef CONFIG_SYN_COOKIES
2828 if (unlikely(req->cookie_ts))
2829 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
2832 skb_mstamp_get(&skb->skb_mstamp);
2833 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2836 skb_push(skb, tcp_header_size);
2837 skb_reset_transport_header(skb);
2840 memset(th, 0, sizeof(struct tcphdr));
2843 TCP_ECN_make_synack(req, th);
2844 th->source = htons(ireq->ir_num);
2845 th->dest = ireq->ir_rmt_port;
2846 /* Setting of flags are superfluous here for callers (and ECE is
2847 * not even correctly set)
2849 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2850 TCPHDR_SYN | TCPHDR_ACK);
2852 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2853 /* XXX data is queued and acked as is. No buffer/window check */
2854 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2856 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2857 th->window = htons(min(req->rcv_wnd, 65535U));
2858 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2859 th->doff = (tcp_header_size >> 2);
2860 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2862 #ifdef CONFIG_TCP_MD5SIG
2863 /* Okay, we have all we need - do the md5 hash if needed */
2865 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2866 md5, NULL, req, skb);
2872 EXPORT_SYMBOL(tcp_make_synack);
2874 /* Do all connect socket setups that can be done AF independent. */
2875 static void tcp_connect_init(struct sock *sk)
2877 const struct dst_entry *dst = __sk_dst_get(sk);
2878 struct tcp_sock *tp = tcp_sk(sk);
2881 /* We'll fix this up when we get a response from the other end.
2882 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2884 tp->tcp_header_len = sizeof(struct tcphdr) +
2885 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2887 #ifdef CONFIG_TCP_MD5SIG
2888 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2889 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2892 /* If user gave his TCP_MAXSEG, record it to clamp */
2893 if (tp->rx_opt.user_mss)
2894 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2897 tcp_sync_mss(sk, dst_mtu(dst));
2899 if (!tp->window_clamp)
2900 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2901 tp->advmss = dst_metric_advmss(dst);
2902 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2903 tp->advmss = tp->rx_opt.user_mss;
2905 tcp_initialize_rcv_mss(sk);
2907 /* limit the window selection if the user enforce a smaller rx buffer */
2908 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2909 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2910 tp->window_clamp = tcp_full_space(sk);
2912 tcp_select_initial_window(tcp_full_space(sk),
2913 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2916 sysctl_tcp_window_scaling,
2918 dst_metric(dst, RTAX_INITRWND));
2920 tp->rx_opt.rcv_wscale = rcv_wscale;
2921 tp->rcv_ssthresh = tp->rcv_wnd;
2924 sock_reset_flag(sk, SOCK_DONE);
2927 tp->snd_una = tp->write_seq;
2928 tp->snd_sml = tp->write_seq;
2929 tp->snd_up = tp->write_seq;
2930 tp->snd_nxt = tp->write_seq;
2932 if (likely(!tp->repair))
2935 tp->rcv_tstamp = tcp_time_stamp;
2936 tp->rcv_wup = tp->rcv_nxt;
2937 tp->copied_seq = tp->rcv_nxt;
2939 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2940 inet_csk(sk)->icsk_retransmits = 0;
2941 tcp_clear_retrans(tp);
2944 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2946 struct tcp_sock *tp = tcp_sk(sk);
2947 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2949 tcb->end_seq += skb->len;
2950 skb_header_release(skb);
2951 __tcp_add_write_queue_tail(sk, skb);
2952 sk->sk_wmem_queued += skb->truesize;
2953 sk_mem_charge(sk, skb->truesize);
2954 tp->write_seq = tcb->end_seq;
2955 tp->packets_out += tcp_skb_pcount(skb);
2958 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2959 * queue a data-only packet after the regular SYN, such that regular SYNs
2960 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2961 * only the SYN sequence, the data are retransmitted in the first ACK.
2962 * If cookie is not cached or other error occurs, falls back to send a
2963 * regular SYN with Fast Open cookie request option.
2965 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2967 struct tcp_sock *tp = tcp_sk(sk);
2968 struct tcp_fastopen_request *fo = tp->fastopen_req;
2969 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2970 struct sk_buff *syn_data = NULL, *data;
2971 unsigned long last_syn_loss = 0;
2973 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2974 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2975 &syn_loss, &last_syn_loss);
2976 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2978 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2979 fo->cookie.len = -1;
2983 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2984 fo->cookie.len = -1;
2985 else if (fo->cookie.len <= 0)
2988 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2989 * user-MSS. Reserve maximum option space for middleboxes that add
2990 * private TCP options. The cost is reduced data space in SYN :(
2992 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2993 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2994 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2995 MAX_TCP_OPTION_SPACE;
2997 space = min_t(size_t, space, fo->size);
2999 /* limit to order-0 allocations */
3000 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3002 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
3004 if (syn_data == NULL)
3007 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
3008 struct iovec *iov = &fo->data->msg_iov[i];
3009 unsigned char __user *from = iov->iov_base;
3010 int len = iov->iov_len;
3012 if (syn_data->len + len > space)
3013 len = space - syn_data->len;
3014 else if (i + 1 == iovlen)
3015 /* No more data pending in inet_wait_for_connect() */
3018 if (skb_add_data(syn_data, from, len))
3022 /* Queue a data-only packet after the regular SYN for retransmission */
3023 data = pskb_copy(syn_data, sk->sk_allocation);
3026 TCP_SKB_CB(data)->seq++;
3027 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
3028 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
3029 tcp_connect_queue_skb(sk, data);
3030 fo->copied = data->len;
3032 /* syn_data is about to be sent, we need to take current time stamps
3033 * for the packets that are in write queue : SYN packet and DATA
3035 skb_mstamp_get(&syn->skb_mstamp);
3036 data->skb_mstamp = syn->skb_mstamp;
3038 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
3039 tp->syn_data = (fo->copied > 0);
3040 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3046 /* Send a regular SYN with Fast Open cookie request option */
3047 if (fo->cookie.len > 0)
3049 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3051 tp->syn_fastopen = 0;
3052 kfree_skb(syn_data);
3054 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3058 /* Build a SYN and send it off. */
3059 int tcp_connect(struct sock *sk)
3061 struct tcp_sock *tp = tcp_sk(sk);
3062 struct sk_buff *buff;
3065 tcp_connect_init(sk);
3067 if (unlikely(tp->repair)) {
3068 tcp_finish_connect(sk, NULL);
3072 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3073 if (unlikely(buff == NULL))
3076 /* Reserve space for headers. */
3077 skb_reserve(buff, MAX_TCP_HEADER);
3079 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3080 tp->retrans_stamp = tcp_time_stamp;
3081 tcp_connect_queue_skb(sk, buff);
3082 TCP_ECN_send_syn(sk, buff);
3084 /* Send off SYN; include data in Fast Open. */
3085 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3086 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3087 if (err == -ECONNREFUSED)
3090 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3091 * in order to make this packet get counted in tcpOutSegs.
3093 tp->snd_nxt = tp->write_seq;
3094 tp->pushed_seq = tp->write_seq;
3095 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3097 /* Timer for repeating the SYN until an answer. */
3098 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3099 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3102 EXPORT_SYMBOL(tcp_connect);
3104 /* Send out a delayed ack, the caller does the policy checking
3105 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3108 void tcp_send_delayed_ack(struct sock *sk)
3110 struct inet_connection_sock *icsk = inet_csk(sk);
3111 int ato = icsk->icsk_ack.ato;
3112 unsigned long timeout;
3114 if (ato > TCP_DELACK_MIN) {
3115 const struct tcp_sock *tp = tcp_sk(sk);
3116 int max_ato = HZ / 2;
3118 if (icsk->icsk_ack.pingpong ||
3119 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3120 max_ato = TCP_DELACK_MAX;
3122 /* Slow path, intersegment interval is "high". */
3124 /* If some rtt estimate is known, use it to bound delayed ack.
3125 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3129 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3136 ato = min(ato, max_ato);
3139 /* Stay within the limit we were given */
3140 timeout = jiffies + ato;
3142 /* Use new timeout only if there wasn't a older one earlier. */
3143 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3144 /* If delack timer was blocked or is about to expire,
3147 if (icsk->icsk_ack.blocked ||
3148 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3153 if (!time_before(timeout, icsk->icsk_ack.timeout))
3154 timeout = icsk->icsk_ack.timeout;
3156 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3157 icsk->icsk_ack.timeout = timeout;
3158 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3161 /* This routine sends an ack and also updates the window. */
3162 void tcp_send_ack(struct sock *sk)
3164 struct sk_buff *buff;
3166 /* If we have been reset, we may not send again. */
3167 if (sk->sk_state == TCP_CLOSE)
3170 /* We are not putting this on the write queue, so
3171 * tcp_transmit_skb() will set the ownership to this
3174 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3176 inet_csk_schedule_ack(sk);
3177 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3178 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3179 TCP_DELACK_MAX, TCP_RTO_MAX);
3183 /* Reserve space for headers and prepare control bits. */
3184 skb_reserve(buff, MAX_TCP_HEADER);
3185 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3187 /* Send it off, this clears delayed acks for us. */
3188 skb_mstamp_get(&buff->skb_mstamp);
3189 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3192 /* This routine sends a packet with an out of date sequence
3193 * number. It assumes the other end will try to ack it.
3195 * Question: what should we make while urgent mode?
3196 * 4.4BSD forces sending single byte of data. We cannot send
3197 * out of window data, because we have SND.NXT==SND.MAX...
3199 * Current solution: to send TWO zero-length segments in urgent mode:
3200 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3201 * out-of-date with SND.UNA-1 to probe window.
3203 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3205 struct tcp_sock *tp = tcp_sk(sk);
3206 struct sk_buff *skb;
3208 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3209 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3213 /* Reserve space for headers and set control bits. */
3214 skb_reserve(skb, MAX_TCP_HEADER);
3215 /* Use a previous sequence. This should cause the other
3216 * end to send an ack. Don't queue or clone SKB, just
3219 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3220 skb_mstamp_get(&skb->skb_mstamp);
3221 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3224 void tcp_send_window_probe(struct sock *sk)
3226 if (sk->sk_state == TCP_ESTABLISHED) {
3227 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3228 tcp_xmit_probe_skb(sk, 0);
3232 /* Initiate keepalive or window probe from timer. */
3233 int tcp_write_wakeup(struct sock *sk)
3235 struct tcp_sock *tp = tcp_sk(sk);
3236 struct sk_buff *skb;
3238 if (sk->sk_state == TCP_CLOSE)
3241 if ((skb = tcp_send_head(sk)) != NULL &&
3242 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3244 unsigned int mss = tcp_current_mss(sk);
3245 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3247 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3248 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3250 /* We are probing the opening of a window
3251 * but the window size is != 0
3252 * must have been a result SWS avoidance ( sender )
3254 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3256 seg_size = min(seg_size, mss);
3257 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3258 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3260 } else if (!tcp_skb_pcount(skb))
3261 tcp_set_skb_tso_segs(sk, skb, mss);
3263 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3264 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3266 tcp_event_new_data_sent(sk, skb);
3269 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3270 tcp_xmit_probe_skb(sk, 1);
3271 return tcp_xmit_probe_skb(sk, 0);
3275 /* A window probe timeout has occurred. If window is not closed send
3276 * a partial packet else a zero probe.
3278 void tcp_send_probe0(struct sock *sk)
3280 struct inet_connection_sock *icsk = inet_csk(sk);
3281 struct tcp_sock *tp = tcp_sk(sk);
3282 unsigned long probe_max;
3285 err = tcp_write_wakeup(sk);
3287 if (tp->packets_out || !tcp_send_head(sk)) {
3288 /* Cancel probe timer, if it is not required. */
3289 icsk->icsk_probes_out = 0;
3290 icsk->icsk_backoff = 0;
3295 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3296 icsk->icsk_backoff++;
3297 icsk->icsk_probes_out++;
3298 probe_max = TCP_RTO_MAX;
3300 /* If packet was not sent due to local congestion,
3301 * do not backoff and do not remember icsk_probes_out.
3302 * Let local senders to fight for local resources.
3304 * Use accumulated backoff yet.
3306 if (!icsk->icsk_probes_out)
3307 icsk->icsk_probes_out = 1;
3308 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3310 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3311 inet_csk_rto_backoff(icsk, probe_max),
3315 int tcp_rtx_synack(struct sock *sk, struct request_sock *req)
3317 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3321 res = af_ops->send_synack(sk, NULL, &fl, req, 0, NULL);
3323 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3324 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3328 EXPORT_SYMBOL(tcp_rtx_synack);