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) {
91 /* SND.NXT, if window was not shrunk.
92 * If window has been shrunk, what should we make? It is not clear at all.
93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
95 * invalid. OK, let's make this for now:
97 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
99 const struct tcp_sock *tp = tcp_sk(sk);
101 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
104 return tcp_wnd_end(tp);
107 /* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
121 static __u16 tcp_advertise_mss(struct sock *sk)
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
128 unsigned int metric = dst_metric_advmss(dst);
139 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism. */
141 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
143 struct tcp_sock *tp = tcp_sk(sk);
144 s32 delta = tcp_time_stamp - tp->lsndtime;
145 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
146 u32 cwnd = tp->snd_cwnd;
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 tp->snd_cwnd = max(cwnd, restart_cwnd);
156 tp->snd_cwnd_stamp = tcp_time_stamp;
157 tp->snd_cwnd_used = 0;
160 /* Congestion state accounting after a packet has been sent. */
161 static void tcp_event_data_sent(struct tcp_sock *tp,
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_time_stamp;
166 const struct dst_entry *dst = __sk_dst_get(sk);
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
178 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
179 icsk->icsk_ack.pingpong = 1;
182 /* Account for an ACK we sent. */
183 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 u32 tcp_default_init_rwnd(u32 mss)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd = TCP_INIT_CWND * 2;
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
213 int wscale_ok, __u8 *rcv_wscale,
216 unsigned int space = (__space < 0 ? 0 : __space);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (65535 << 14);
221 space = min(*window_clamp, space);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space = (space / mss) * mss;
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
242 /* Set window scaling on max possible window
243 * See RFC1323 for an explanation of the limit to 14
245 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
246 space = min_t(u32, space, *window_clamp);
247 while (space > 65535 && (*rcv_wscale) < 14) {
253 if (mss > (1 << *rcv_wscale)) {
254 if (!init_rcv_wnd) /* Use default unless specified otherwise */
255 init_rcv_wnd = tcp_default_init_rwnd(mss);
256 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
259 /* Set the clamp no higher than max representable value */
260 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
262 EXPORT_SYMBOL(tcp_select_initial_window);
264 /* Chose a new window to advertise, update state in tcp_sock for the
265 * socket, and return result with RFC1323 scaling applied. The return
266 * value can be stuffed directly into th->window for an outgoing
269 static u16 tcp_select_window(struct sock *sk)
271 struct tcp_sock *tp = tcp_sk(sk);
272 u32 cur_win = tcp_receive_window(tp);
273 u32 new_win = __tcp_select_window(sk);
275 /* Never shrink the offered window */
276 if (new_win < cur_win) {
277 /* Danger Will Robinson!
278 * Don't update rcv_wup/rcv_wnd here or else
279 * we will not be able to advertise a zero
280 * window in time. --DaveM
282 * Relax Will Robinson.
284 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
286 tp->rcv_wnd = new_win;
287 tp->rcv_wup = tp->rcv_nxt;
289 /* Make sure we do not exceed the maximum possible
292 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
293 new_win = min(new_win, MAX_TCP_WINDOW);
295 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
297 /* RFC1323 scaling applied */
298 new_win >>= tp->rx_opt.rcv_wscale;
300 /* If we advertise zero window, disable fast path. */
307 /* Packet ECN state for a SYN-ACK */
308 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
310 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
311 if (!(tp->ecn_flags & TCP_ECN_OK))
312 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
315 /* Packet ECN state for a SYN. */
316 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
318 struct tcp_sock *tp = tcp_sk(sk);
321 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
322 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
323 tp->ecn_flags = TCP_ECN_OK;
327 static __inline__ void
328 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
330 if (inet_rsk(req)->ecn_ok)
334 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
337 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
340 struct tcp_sock *tp = tcp_sk(sk);
342 if (tp->ecn_flags & TCP_ECN_OK) {
343 /* Not-retransmitted data segment: set ECT and inject CWR. */
344 if (skb->len != tcp_header_len &&
345 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
347 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
348 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
349 tcp_hdr(skb)->cwr = 1;
350 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
353 /* ACK or retransmitted segment: clear ECT|CE */
354 INET_ECN_dontxmit(sk);
356 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
357 tcp_hdr(skb)->ece = 1;
361 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
362 * auto increment end seqno.
364 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
366 skb->ip_summed = CHECKSUM_PARTIAL;
369 TCP_SKB_CB(skb)->tcp_flags = flags;
370 TCP_SKB_CB(skb)->sacked = 0;
372 skb_shinfo(skb)->gso_segs = 1;
373 skb_shinfo(skb)->gso_size = 0;
374 skb_shinfo(skb)->gso_type = 0;
376 TCP_SKB_CB(skb)->seq = seq;
377 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
379 TCP_SKB_CB(skb)->end_seq = seq;
382 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
384 return tp->snd_una != tp->snd_up;
387 #define OPTION_SACK_ADVERTISE (1 << 0)
388 #define OPTION_TS (1 << 1)
389 #define OPTION_MD5 (1 << 2)
390 #define OPTION_WSCALE (1 << 3)
391 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
393 struct tcp_out_options {
394 u16 options; /* bit field of OPTION_* */
395 u16 mss; /* 0 to disable */
396 u8 ws; /* window scale, 0 to disable */
397 u8 num_sack_blocks; /* number of SACK blocks to include */
398 u8 hash_size; /* bytes in hash_location */
399 __u8 *hash_location; /* temporary pointer, overloaded */
400 __u32 tsval, tsecr; /* need to include OPTION_TS */
401 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
404 /* Write previously computed TCP options to the packet.
406 * Beware: Something in the Internet is very sensitive to the ordering of
407 * TCP options, we learned this through the hard way, so be careful here.
408 * Luckily we can at least blame others for their non-compliance but from
409 * inter-operatibility perspective it seems that we're somewhat stuck with
410 * the ordering which we have been using if we want to keep working with
411 * those broken things (not that it currently hurts anybody as there isn't
412 * particular reason why the ordering would need to be changed).
414 * At least SACK_PERM as the first option is known to lead to a disaster
415 * (but it may well be that other scenarios fail similarly).
417 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
418 struct tcp_out_options *opts)
420 u16 options = opts->options; /* mungable copy */
422 if (unlikely(OPTION_MD5 & options)) {
423 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
424 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
425 /* overload cookie hash location */
426 opts->hash_location = (__u8 *)ptr;
430 if (unlikely(opts->mss)) {
431 *ptr++ = htonl((TCPOPT_MSS << 24) |
432 (TCPOLEN_MSS << 16) |
436 if (likely(OPTION_TS & options)) {
437 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
438 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
439 (TCPOLEN_SACK_PERM << 16) |
440 (TCPOPT_TIMESTAMP << 8) |
442 options &= ~OPTION_SACK_ADVERTISE;
444 *ptr++ = htonl((TCPOPT_NOP << 24) |
446 (TCPOPT_TIMESTAMP << 8) |
449 *ptr++ = htonl(opts->tsval);
450 *ptr++ = htonl(opts->tsecr);
453 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
454 *ptr++ = htonl((TCPOPT_NOP << 24) |
456 (TCPOPT_SACK_PERM << 8) |
460 if (unlikely(OPTION_WSCALE & options)) {
461 *ptr++ = htonl((TCPOPT_NOP << 24) |
462 (TCPOPT_WINDOW << 16) |
463 (TCPOLEN_WINDOW << 8) |
467 if (unlikely(opts->num_sack_blocks)) {
468 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
469 tp->duplicate_sack : tp->selective_acks;
472 *ptr++ = htonl((TCPOPT_NOP << 24) |
475 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
476 TCPOLEN_SACK_PERBLOCK)));
478 for (this_sack = 0; this_sack < opts->num_sack_blocks;
480 *ptr++ = htonl(sp[this_sack].start_seq);
481 *ptr++ = htonl(sp[this_sack].end_seq);
484 tp->rx_opt.dsack = 0;
487 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
488 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
490 *ptr++ = htonl((TCPOPT_EXP << 24) |
491 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
492 TCPOPT_FASTOPEN_MAGIC);
494 memcpy(ptr, foc->val, foc->len);
495 if ((foc->len & 3) == 2) {
496 u8 *align = ((u8 *)ptr) + foc->len;
497 align[0] = align[1] = TCPOPT_NOP;
499 ptr += (foc->len + 3) >> 2;
503 /* Compute TCP options for SYN packets. This is not the final
504 * network wire format yet.
506 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
507 struct tcp_out_options *opts,
508 struct tcp_md5sig_key **md5)
510 struct tcp_sock *tp = tcp_sk(sk);
511 unsigned int remaining = MAX_TCP_OPTION_SPACE;
512 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
514 #ifdef CONFIG_TCP_MD5SIG
515 *md5 = tp->af_specific->md5_lookup(sk, sk);
517 opts->options |= OPTION_MD5;
518 remaining -= TCPOLEN_MD5SIG_ALIGNED;
524 /* We always get an MSS option. The option bytes which will be seen in
525 * normal data packets should timestamps be used, must be in the MSS
526 * advertised. But we subtract them from tp->mss_cache so that
527 * calculations in tcp_sendmsg are simpler etc. So account for this
528 * fact here if necessary. If we don't do this correctly, as a
529 * receiver we won't recognize data packets as being full sized when we
530 * should, and thus we won't abide by the delayed ACK rules correctly.
531 * SACKs don't matter, we never delay an ACK when we have any of those
533 opts->mss = tcp_advertise_mss(sk);
534 remaining -= TCPOLEN_MSS_ALIGNED;
536 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
537 opts->options |= OPTION_TS;
538 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
539 opts->tsecr = tp->rx_opt.ts_recent;
540 remaining -= TCPOLEN_TSTAMP_ALIGNED;
542 if (likely(sysctl_tcp_window_scaling)) {
543 opts->ws = tp->rx_opt.rcv_wscale;
544 opts->options |= OPTION_WSCALE;
545 remaining -= TCPOLEN_WSCALE_ALIGNED;
547 if (likely(sysctl_tcp_sack)) {
548 opts->options |= OPTION_SACK_ADVERTISE;
549 if (unlikely(!(OPTION_TS & opts->options)))
550 remaining -= TCPOLEN_SACKPERM_ALIGNED;
553 if (fastopen && fastopen->cookie.len >= 0) {
554 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
555 need = (need + 3) & ~3U; /* Align to 32 bits */
556 if (remaining >= need) {
557 opts->options |= OPTION_FAST_OPEN_COOKIE;
558 opts->fastopen_cookie = &fastopen->cookie;
560 tp->syn_fastopen = 1;
564 return MAX_TCP_OPTION_SPACE - remaining;
567 /* Set up TCP options for SYN-ACKs. */
568 static unsigned int tcp_synack_options(struct sock *sk,
569 struct request_sock *req,
570 unsigned int mss, struct sk_buff *skb,
571 struct tcp_out_options *opts,
572 struct tcp_md5sig_key **md5,
573 struct tcp_fastopen_cookie *foc)
575 struct inet_request_sock *ireq = inet_rsk(req);
576 unsigned int remaining = MAX_TCP_OPTION_SPACE;
578 #ifdef CONFIG_TCP_MD5SIG
579 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
581 opts->options |= OPTION_MD5;
582 remaining -= TCPOLEN_MD5SIG_ALIGNED;
584 /* We can't fit any SACK blocks in a packet with MD5 + TS
585 * options. There was discussion about disabling SACK
586 * rather than TS in order to fit in better with old,
587 * buggy kernels, but that was deemed to be unnecessary.
589 ireq->tstamp_ok &= !ireq->sack_ok;
595 /* We always send an MSS option. */
597 remaining -= TCPOLEN_MSS_ALIGNED;
599 if (likely(ireq->wscale_ok)) {
600 opts->ws = ireq->rcv_wscale;
601 opts->options |= OPTION_WSCALE;
602 remaining -= TCPOLEN_WSCALE_ALIGNED;
604 if (likely(ireq->tstamp_ok)) {
605 opts->options |= OPTION_TS;
606 opts->tsval = TCP_SKB_CB(skb)->when;
607 opts->tsecr = req->ts_recent;
608 remaining -= TCPOLEN_TSTAMP_ALIGNED;
610 if (likely(ireq->sack_ok)) {
611 opts->options |= OPTION_SACK_ADVERTISE;
612 if (unlikely(!ireq->tstamp_ok))
613 remaining -= TCPOLEN_SACKPERM_ALIGNED;
616 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
617 need = (need + 3) & ~3U; /* Align to 32 bits */
618 if (remaining >= need) {
619 opts->options |= OPTION_FAST_OPEN_COOKIE;
620 opts->fastopen_cookie = foc;
625 return MAX_TCP_OPTION_SPACE - remaining;
628 /* Compute TCP options for ESTABLISHED sockets. This is not the
629 * final wire format yet.
631 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
632 struct tcp_out_options *opts,
633 struct tcp_md5sig_key **md5)
635 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
636 struct tcp_sock *tp = tcp_sk(sk);
637 unsigned int size = 0;
638 unsigned int eff_sacks;
640 #ifdef CONFIG_TCP_MD5SIG
641 *md5 = tp->af_specific->md5_lookup(sk, sk);
642 if (unlikely(*md5)) {
643 opts->options |= OPTION_MD5;
644 size += TCPOLEN_MD5SIG_ALIGNED;
650 if (likely(tp->rx_opt.tstamp_ok)) {
651 opts->options |= OPTION_TS;
652 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
653 opts->tsecr = tp->rx_opt.ts_recent;
654 size += TCPOLEN_TSTAMP_ALIGNED;
657 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
658 if (unlikely(eff_sacks)) {
659 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
660 opts->num_sack_blocks =
661 min_t(unsigned int, eff_sacks,
662 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
663 TCPOLEN_SACK_PERBLOCK);
664 size += TCPOLEN_SACK_BASE_ALIGNED +
665 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
672 /* TCP SMALL QUEUES (TSQ)
674 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
675 * to reduce RTT and bufferbloat.
676 * We do this using a special skb destructor (tcp_wfree).
678 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
679 * needs to be reallocated in a driver.
680 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
682 * Since transmit from skb destructor is forbidden, we use a tasklet
683 * to process all sockets that eventually need to send more skbs.
684 * We use one tasklet per cpu, with its own queue of sockets.
687 struct tasklet_struct tasklet;
688 struct list_head head; /* queue of tcp sockets */
690 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
692 static void tcp_tsq_handler(struct sock *sk)
694 if ((1 << sk->sk_state) &
695 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
696 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
697 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
700 * One tasklest per cpu tries to send more skbs.
701 * We run in tasklet context but need to disable irqs when
702 * transfering tsq->head because tcp_wfree() might
703 * interrupt us (non NAPI drivers)
705 static void tcp_tasklet_func(unsigned long data)
707 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
710 struct list_head *q, *n;
714 local_irq_save(flags);
715 list_splice_init(&tsq->head, &list);
716 local_irq_restore(flags);
718 list_for_each_safe(q, n, &list) {
719 tp = list_entry(q, struct tcp_sock, tsq_node);
720 list_del(&tp->tsq_node);
722 sk = (struct sock *)tp;
725 if (!sock_owned_by_user(sk)) {
728 /* defer the work to tcp_release_cb() */
729 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
733 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
738 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
739 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
740 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
741 (1UL << TCP_MTU_REDUCED_DEFERRED))
743 * tcp_release_cb - tcp release_sock() callback
746 * called from release_sock() to perform protocol dependent
747 * actions before socket release.
749 void tcp_release_cb(struct sock *sk)
751 struct tcp_sock *tp = tcp_sk(sk);
752 unsigned long flags, nflags;
754 /* perform an atomic operation only if at least one flag is set */
756 flags = tp->tsq_flags;
757 if (!(flags & TCP_DEFERRED_ALL))
759 nflags = flags & ~TCP_DEFERRED_ALL;
760 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
762 if (flags & (1UL << TCP_TSQ_DEFERRED))
765 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
766 tcp_write_timer_handler(sk);
769 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
770 tcp_delack_timer_handler(sk);
773 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
774 sk->sk_prot->mtu_reduced(sk);
778 EXPORT_SYMBOL(tcp_release_cb);
780 void __init tcp_tasklet_init(void)
784 for_each_possible_cpu(i) {
785 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
787 INIT_LIST_HEAD(&tsq->head);
788 tasklet_init(&tsq->tasklet,
795 * Write buffer destructor automatically called from kfree_skb.
796 * We cant xmit new skbs from this context, as we might already
799 void tcp_wfree(struct sk_buff *skb)
801 struct sock *sk = skb->sk;
802 struct tcp_sock *tp = tcp_sk(sk);
804 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
805 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
807 struct tsq_tasklet *tsq;
809 /* Keep a ref on socket.
810 * This last ref will be released in tcp_tasklet_func()
812 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
814 /* queue this socket to tasklet queue */
815 local_irq_save(flags);
816 tsq = &__get_cpu_var(tsq_tasklet);
817 list_add(&tp->tsq_node, &tsq->head);
818 tasklet_schedule(&tsq->tasklet);
819 local_irq_restore(flags);
825 /* This routine actually transmits TCP packets queued in by
826 * tcp_do_sendmsg(). This is used by both the initial
827 * transmission and possible later retransmissions.
828 * All SKB's seen here are completely headerless. It is our
829 * job to build the TCP header, and pass the packet down to
830 * IP so it can do the same plus pass the packet off to the
833 * We are working here with either a clone of the original
834 * SKB, or a fresh unique copy made by the retransmit engine.
836 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
839 const struct inet_connection_sock *icsk = inet_csk(sk);
840 struct inet_sock *inet;
842 struct tcp_skb_cb *tcb;
843 struct tcp_out_options opts;
844 unsigned int tcp_options_size, tcp_header_size;
845 struct tcp_md5sig_key *md5;
849 BUG_ON(!skb || !tcp_skb_pcount(skb));
851 /* If congestion control is doing timestamping, we must
852 * take such a timestamp before we potentially clone/copy.
854 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
855 __net_timestamp(skb);
857 if (likely(clone_it)) {
858 const struct sk_buff *fclone = skb + 1;
860 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
861 fclone->fclone == SKB_FCLONE_CLONE))
862 NET_INC_STATS_BH(sock_net(sk),
863 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
865 if (unlikely(skb_cloned(skb)))
866 skb = pskb_copy(skb, gfp_mask);
868 skb = skb_clone(skb, gfp_mask);
875 tcb = TCP_SKB_CB(skb);
876 memset(&opts, 0, sizeof(opts));
878 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
879 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
881 tcp_options_size = tcp_established_options(sk, skb, &opts,
883 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
885 if (tcp_packets_in_flight(tp) == 0)
886 tcp_ca_event(sk, CA_EVENT_TX_START);
888 /* if no packet is in qdisc/device queue, then allow XPS to select
891 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
893 skb_push(skb, tcp_header_size);
894 skb_reset_transport_header(skb);
898 skb->destructor = tcp_wfree;
899 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
901 /* Build TCP header and checksum it. */
903 th->source = inet->inet_sport;
904 th->dest = inet->inet_dport;
905 th->seq = htonl(tcb->seq);
906 th->ack_seq = htonl(tp->rcv_nxt);
907 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
910 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
911 /* RFC1323: The window in SYN & SYN/ACK segments
914 th->window = htons(min(tp->rcv_wnd, 65535U));
916 th->window = htons(tcp_select_window(sk));
921 /* The urg_mode check is necessary during a below snd_una win probe */
922 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
923 if (before(tp->snd_up, tcb->seq + 0x10000)) {
924 th->urg_ptr = htons(tp->snd_up - tcb->seq);
926 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
927 th->urg_ptr = htons(0xFFFF);
932 tcp_options_write((__be32 *)(th + 1), tp, &opts);
933 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
934 TCP_ECN_send(sk, skb, tcp_header_size);
936 #ifdef CONFIG_TCP_MD5SIG
937 /* Calculate the MD5 hash, as we have all we need now */
939 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
940 tp->af_specific->calc_md5_hash(opts.hash_location,
945 icsk->icsk_af_ops->send_check(sk, skb);
947 if (likely(tcb->tcp_flags & TCPHDR_ACK))
948 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
950 if (skb->len != tcp_header_size)
951 tcp_event_data_sent(tp, sk);
953 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
954 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
955 tcp_skb_pcount(skb));
957 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
958 if (likely(err <= 0))
961 tcp_enter_cwr(sk, 1);
963 return net_xmit_eval(err);
966 /* This routine just queues the buffer for sending.
968 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
969 * otherwise socket can stall.
971 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
973 struct tcp_sock *tp = tcp_sk(sk);
975 /* Advance write_seq and place onto the write_queue. */
976 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
977 skb_header_release(skb);
978 tcp_add_write_queue_tail(sk, skb);
979 sk->sk_wmem_queued += skb->truesize;
980 sk_mem_charge(sk, skb->truesize);
983 /* Initialize TSO segments for a packet. */
984 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
985 unsigned int mss_now)
987 if (skb->len <= mss_now || !sk_can_gso(sk) ||
988 skb->ip_summed == CHECKSUM_NONE) {
989 /* Avoid the costly divide in the normal
992 skb_shinfo(skb)->gso_segs = 1;
993 skb_shinfo(skb)->gso_size = 0;
994 skb_shinfo(skb)->gso_type = 0;
996 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
997 skb_shinfo(skb)->gso_size = mss_now;
998 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1002 /* When a modification to fackets out becomes necessary, we need to check
1003 * skb is counted to fackets_out or not.
1005 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1008 struct tcp_sock *tp = tcp_sk(sk);
1010 if (!tp->sacked_out || tcp_is_reno(tp))
1013 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1014 tp->fackets_out -= decr;
1017 /* Pcount in the middle of the write queue got changed, we need to do various
1018 * tweaks to fix counters
1020 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1022 struct tcp_sock *tp = tcp_sk(sk);
1024 tp->packets_out -= decr;
1026 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1027 tp->sacked_out -= decr;
1028 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1029 tp->retrans_out -= decr;
1030 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1031 tp->lost_out -= decr;
1033 /* Reno case is special. Sigh... */
1034 if (tcp_is_reno(tp) && decr > 0)
1035 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1037 tcp_adjust_fackets_out(sk, skb, decr);
1039 if (tp->lost_skb_hint &&
1040 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1041 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1042 tp->lost_cnt_hint -= decr;
1044 tcp_verify_left_out(tp);
1047 /* Function to create two new TCP segments. Shrinks the given segment
1048 * to the specified size and appends a new segment with the rest of the
1049 * packet to the list. This won't be called frequently, I hope.
1050 * Remember, these are still headerless SKBs at this point.
1052 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1053 unsigned int mss_now)
1055 struct tcp_sock *tp = tcp_sk(sk);
1056 struct sk_buff *buff;
1057 int nsize, old_factor;
1061 if (WARN_ON(len > skb->len))
1064 nsize = skb_headlen(skb) - len;
1068 if (skb_cloned(skb) &&
1069 skb_is_nonlinear(skb) &&
1070 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1073 /* Get a new skb... force flag on. */
1074 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1076 return -ENOMEM; /* We'll just try again later. */
1078 sk->sk_wmem_queued += buff->truesize;
1079 sk_mem_charge(sk, buff->truesize);
1080 nlen = skb->len - len - nsize;
1081 buff->truesize += nlen;
1082 skb->truesize -= nlen;
1084 /* Correct the sequence numbers. */
1085 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1086 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1087 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1089 /* PSH and FIN should only be set in the second packet. */
1090 flags = TCP_SKB_CB(skb)->tcp_flags;
1091 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1092 TCP_SKB_CB(buff)->tcp_flags = flags;
1093 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1095 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1096 /* Copy and checksum data tail into the new buffer. */
1097 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1098 skb_put(buff, nsize),
1103 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1105 skb->ip_summed = CHECKSUM_PARTIAL;
1106 skb_split(skb, buff, len);
1109 buff->ip_summed = skb->ip_summed;
1111 /* Looks stupid, but our code really uses when of
1112 * skbs, which it never sent before. --ANK
1114 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1115 buff->tstamp = skb->tstamp;
1117 old_factor = tcp_skb_pcount(skb);
1119 /* Fix up tso_factor for both original and new SKB. */
1120 tcp_set_skb_tso_segs(sk, skb, mss_now);
1121 tcp_set_skb_tso_segs(sk, buff, mss_now);
1123 /* If this packet has been sent out already, we must
1124 * adjust the various packet counters.
1126 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1127 int diff = old_factor - tcp_skb_pcount(skb) -
1128 tcp_skb_pcount(buff);
1131 tcp_adjust_pcount(sk, skb, diff);
1134 /* Link BUFF into the send queue. */
1135 skb_header_release(buff);
1136 tcp_insert_write_queue_after(skb, buff, sk);
1141 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1142 * eventually). The difference is that pulled data not copied, but
1143 * immediately discarded.
1145 static void __pskb_trim_head(struct sk_buff *skb, int len)
1149 eat = min_t(int, len, skb_headlen(skb));
1151 __skb_pull(skb, eat);
1158 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1159 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1162 skb_frag_unref(skb, i);
1165 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1167 skb_shinfo(skb)->frags[k].page_offset += eat;
1168 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1174 skb_shinfo(skb)->nr_frags = k;
1176 skb_reset_tail_pointer(skb);
1177 skb->data_len -= len;
1178 skb->len = skb->data_len;
1181 /* Remove acked data from a packet in the transmit queue. */
1182 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1184 if (skb_unclone(skb, GFP_ATOMIC))
1187 __pskb_trim_head(skb, len);
1189 TCP_SKB_CB(skb)->seq += len;
1190 skb->ip_summed = CHECKSUM_PARTIAL;
1192 skb->truesize -= len;
1193 sk->sk_wmem_queued -= len;
1194 sk_mem_uncharge(sk, len);
1195 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1197 /* Any change of skb->len requires recalculation of tso factor. */
1198 if (tcp_skb_pcount(skb) > 1)
1199 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1204 /* Calculate MSS not accounting any TCP options. */
1205 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1207 const struct tcp_sock *tp = tcp_sk(sk);
1208 const struct inet_connection_sock *icsk = inet_csk(sk);
1211 /* Calculate base mss without TCP options:
1212 It is MMS_S - sizeof(tcphdr) of rfc1122
1214 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1216 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1217 if (icsk->icsk_af_ops->net_frag_header_len) {
1218 const struct dst_entry *dst = __sk_dst_get(sk);
1220 if (dst && dst_allfrag(dst))
1221 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1224 /* Clamp it (mss_clamp does not include tcp options) */
1225 if (mss_now > tp->rx_opt.mss_clamp)
1226 mss_now = tp->rx_opt.mss_clamp;
1228 /* Now subtract optional transport overhead */
1229 mss_now -= icsk->icsk_ext_hdr_len;
1231 /* Then reserve room for full set of TCP options and 8 bytes of data */
1237 /* Calculate MSS. Not accounting for SACKs here. */
1238 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1240 /* Subtract TCP options size, not including SACKs */
1241 return __tcp_mtu_to_mss(sk, pmtu) -
1242 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1245 /* Inverse of above */
1246 int tcp_mss_to_mtu(struct sock *sk, int mss)
1248 const struct tcp_sock *tp = tcp_sk(sk);
1249 const struct inet_connection_sock *icsk = inet_csk(sk);
1253 tp->tcp_header_len +
1254 icsk->icsk_ext_hdr_len +
1255 icsk->icsk_af_ops->net_header_len;
1257 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1258 if (icsk->icsk_af_ops->net_frag_header_len) {
1259 const struct dst_entry *dst = __sk_dst_get(sk);
1261 if (dst && dst_allfrag(dst))
1262 mtu += icsk->icsk_af_ops->net_frag_header_len;
1267 /* MTU probing init per socket */
1268 void tcp_mtup_init(struct sock *sk)
1270 struct tcp_sock *tp = tcp_sk(sk);
1271 struct inet_connection_sock *icsk = inet_csk(sk);
1273 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1274 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1275 icsk->icsk_af_ops->net_header_len;
1276 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1277 icsk->icsk_mtup.probe_size = 0;
1279 EXPORT_SYMBOL(tcp_mtup_init);
1281 /* This function synchronize snd mss to current pmtu/exthdr set.
1283 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1284 for TCP options, but includes only bare TCP header.
1286 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1287 It is minimum of user_mss and mss received with SYN.
1288 It also does not include TCP options.
1290 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1292 tp->mss_cache is current effective sending mss, including
1293 all tcp options except for SACKs. It is evaluated,
1294 taking into account current pmtu, but never exceeds
1295 tp->rx_opt.mss_clamp.
1297 NOTE1. rfc1122 clearly states that advertised MSS
1298 DOES NOT include either tcp or ip options.
1300 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1301 are READ ONLY outside this function. --ANK (980731)
1303 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1305 struct tcp_sock *tp = tcp_sk(sk);
1306 struct inet_connection_sock *icsk = inet_csk(sk);
1309 if (icsk->icsk_mtup.search_high > pmtu)
1310 icsk->icsk_mtup.search_high = pmtu;
1312 mss_now = tcp_mtu_to_mss(sk, pmtu);
1313 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1315 /* And store cached results */
1316 icsk->icsk_pmtu_cookie = pmtu;
1317 if (icsk->icsk_mtup.enabled)
1318 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1319 tp->mss_cache = mss_now;
1323 EXPORT_SYMBOL(tcp_sync_mss);
1325 /* Compute the current effective MSS, taking SACKs and IP options,
1326 * and even PMTU discovery events into account.
1328 unsigned int tcp_current_mss(struct sock *sk)
1330 const struct tcp_sock *tp = tcp_sk(sk);
1331 const struct dst_entry *dst = __sk_dst_get(sk);
1333 unsigned int header_len;
1334 struct tcp_out_options opts;
1335 struct tcp_md5sig_key *md5;
1337 mss_now = tp->mss_cache;
1340 u32 mtu = dst_mtu(dst);
1341 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1342 mss_now = tcp_sync_mss(sk, mtu);
1345 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1346 sizeof(struct tcphdr);
1347 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1348 * some common options. If this is an odd packet (because we have SACK
1349 * blocks etc) then our calculated header_len will be different, and
1350 * we have to adjust mss_now correspondingly */
1351 if (header_len != tp->tcp_header_len) {
1352 int delta = (int) header_len - tp->tcp_header_len;
1359 /* Congestion window validation. (RFC2861) */
1360 static void tcp_cwnd_validate(struct sock *sk)
1362 struct tcp_sock *tp = tcp_sk(sk);
1364 if (tp->packets_out >= tp->snd_cwnd) {
1365 /* Network is feed fully. */
1366 tp->snd_cwnd_used = 0;
1367 tp->snd_cwnd_stamp = tcp_time_stamp;
1369 /* Network starves. */
1370 if (tp->packets_out > tp->snd_cwnd_used)
1371 tp->snd_cwnd_used = tp->packets_out;
1373 if (sysctl_tcp_slow_start_after_idle &&
1374 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1375 tcp_cwnd_application_limited(sk);
1379 /* Returns the portion of skb which can be sent right away without
1380 * introducing MSS oddities to segment boundaries. In rare cases where
1381 * mss_now != mss_cache, we will request caller to create a small skb
1382 * per input skb which could be mostly avoided here (if desired).
1384 * We explicitly want to create a request for splitting write queue tail
1385 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1386 * thus all the complexity (cwnd_len is always MSS multiple which we
1387 * return whenever allowed by the other factors). Basically we need the
1388 * modulo only when the receiver window alone is the limiting factor or
1389 * when we would be allowed to send the split-due-to-Nagle skb fully.
1391 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1392 unsigned int mss_now, unsigned int max_segs)
1394 const struct tcp_sock *tp = tcp_sk(sk);
1395 u32 needed, window, max_len;
1397 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1398 max_len = mss_now * max_segs;
1400 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1403 needed = min(skb->len, window);
1405 if (max_len <= needed)
1408 return needed - needed % mss_now;
1411 /* Can at least one segment of SKB be sent right now, according to the
1412 * congestion window rules? If so, return how many segments are allowed.
1414 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1415 const struct sk_buff *skb)
1417 u32 in_flight, cwnd;
1419 /* Don't be strict about the congestion window for the final FIN. */
1420 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1421 tcp_skb_pcount(skb) == 1)
1424 in_flight = tcp_packets_in_flight(tp);
1425 cwnd = tp->snd_cwnd;
1426 if (in_flight < cwnd)
1427 return (cwnd - in_flight);
1432 /* Initialize TSO state of a skb.
1433 * This must be invoked the first time we consider transmitting
1434 * SKB onto the wire.
1436 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1437 unsigned int mss_now)
1439 int tso_segs = tcp_skb_pcount(skb);
1441 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1442 tcp_set_skb_tso_segs(sk, skb, mss_now);
1443 tso_segs = tcp_skb_pcount(skb);
1448 /* Minshall's variant of the Nagle send check. */
1449 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1451 return after(tp->snd_sml, tp->snd_una) &&
1452 !after(tp->snd_sml, tp->snd_nxt);
1455 /* Return false, if packet can be sent now without violation Nagle's rules:
1456 * 1. It is full sized.
1457 * 2. Or it contains FIN. (already checked by caller)
1458 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1459 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1460 * With Minshall's modification: all sent small packets are ACKed.
1462 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1463 const struct sk_buff *skb,
1464 unsigned int mss_now, int nonagle)
1466 return skb->len < mss_now &&
1467 ((nonagle & TCP_NAGLE_CORK) ||
1468 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1471 /* Return true if the Nagle test allows this packet to be
1474 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1475 unsigned int cur_mss, int nonagle)
1477 /* Nagle rule does not apply to frames, which sit in the middle of the
1478 * write_queue (they have no chances to get new data).
1480 * This is implemented in the callers, where they modify the 'nonagle'
1481 * argument based upon the location of SKB in the send queue.
1483 if (nonagle & TCP_NAGLE_PUSH)
1486 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1487 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1490 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1496 /* Does at least the first segment of SKB fit into the send window? */
1497 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1498 const struct sk_buff *skb,
1499 unsigned int cur_mss)
1501 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1503 if (skb->len > cur_mss)
1504 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1506 return !after(end_seq, tcp_wnd_end(tp));
1509 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1510 * should be put on the wire right now. If so, it returns the number of
1511 * packets allowed by the congestion window.
1513 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1514 unsigned int cur_mss, int nonagle)
1516 const struct tcp_sock *tp = tcp_sk(sk);
1517 unsigned int cwnd_quota;
1519 tcp_init_tso_segs(sk, skb, cur_mss);
1521 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1524 cwnd_quota = tcp_cwnd_test(tp, skb);
1525 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1531 /* Test if sending is allowed right now. */
1532 bool tcp_may_send_now(struct sock *sk)
1534 const struct tcp_sock *tp = tcp_sk(sk);
1535 struct sk_buff *skb = tcp_send_head(sk);
1538 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1539 (tcp_skb_is_last(sk, skb) ?
1540 tp->nonagle : TCP_NAGLE_PUSH));
1543 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1544 * which is put after SKB on the list. It is very much like
1545 * tcp_fragment() except that it may make several kinds of assumptions
1546 * in order to speed up the splitting operation. In particular, we
1547 * know that all the data is in scatter-gather pages, and that the
1548 * packet has never been sent out before (and thus is not cloned).
1550 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1551 unsigned int mss_now, gfp_t gfp)
1553 struct sk_buff *buff;
1554 int nlen = skb->len - len;
1557 /* All of a TSO frame must be composed of paged data. */
1558 if (skb->len != skb->data_len)
1559 return tcp_fragment(sk, skb, len, mss_now);
1561 buff = sk_stream_alloc_skb(sk, 0, gfp);
1562 if (unlikely(buff == NULL))
1565 sk->sk_wmem_queued += buff->truesize;
1566 sk_mem_charge(sk, buff->truesize);
1567 buff->truesize += nlen;
1568 skb->truesize -= nlen;
1570 /* Correct the sequence numbers. */
1571 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1572 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1573 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1575 /* PSH and FIN should only be set in the second packet. */
1576 flags = TCP_SKB_CB(skb)->tcp_flags;
1577 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1578 TCP_SKB_CB(buff)->tcp_flags = flags;
1580 /* This packet was never sent out yet, so no SACK bits. */
1581 TCP_SKB_CB(buff)->sacked = 0;
1583 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1584 skb_split(skb, buff, len);
1586 /* Fix up tso_factor for both original and new SKB. */
1587 tcp_set_skb_tso_segs(sk, skb, mss_now);
1588 tcp_set_skb_tso_segs(sk, buff, mss_now);
1590 /* Link BUFF into the send queue. */
1591 skb_header_release(buff);
1592 tcp_insert_write_queue_after(skb, buff, sk);
1597 /* Try to defer sending, if possible, in order to minimize the amount
1598 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1600 * This algorithm is from John Heffner.
1602 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1604 struct tcp_sock *tp = tcp_sk(sk);
1605 const struct inet_connection_sock *icsk = inet_csk(sk);
1606 u32 send_win, cong_win, limit, in_flight;
1609 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1612 if (icsk->icsk_ca_state != TCP_CA_Open)
1615 /* Defer for less than two clock ticks. */
1616 if (tp->tso_deferred &&
1617 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1620 in_flight = tcp_packets_in_flight(tp);
1622 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1624 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1626 /* From in_flight test above, we know that cwnd > in_flight. */
1627 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1629 limit = min(send_win, cong_win);
1631 /* If a full-sized TSO skb can be sent, do it. */
1632 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1633 tp->xmit_size_goal_segs * tp->mss_cache))
1636 /* Middle in queue won't get any more data, full sendable already? */
1637 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1640 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1642 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1644 /* If at least some fraction of a window is available,
1647 chunk /= win_divisor;
1651 /* Different approach, try not to defer past a single
1652 * ACK. Receiver should ACK every other full sized
1653 * frame, so if we have space for more than 3 frames
1656 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1660 /* Ok, it looks like it is advisable to defer.
1661 * Do not rearm the timer if already set to not break TCP ACK clocking.
1663 if (!tp->tso_deferred)
1664 tp->tso_deferred = 1 | (jiffies << 1);
1669 tp->tso_deferred = 0;
1673 /* Create a new MTU probe if we are ready.
1674 * MTU probe is regularly attempting to increase the path MTU by
1675 * deliberately sending larger packets. This discovers routing
1676 * changes resulting in larger path MTUs.
1678 * Returns 0 if we should wait to probe (no cwnd available),
1679 * 1 if a probe was sent,
1682 static int tcp_mtu_probe(struct sock *sk)
1684 struct tcp_sock *tp = tcp_sk(sk);
1685 struct inet_connection_sock *icsk = inet_csk(sk);
1686 struct sk_buff *skb, *nskb, *next;
1693 /* Not currently probing/verifying,
1695 * have enough cwnd, and
1696 * not SACKing (the variable headers throw things off) */
1697 if (!icsk->icsk_mtup.enabled ||
1698 icsk->icsk_mtup.probe_size ||
1699 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1700 tp->snd_cwnd < 11 ||
1701 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1704 /* Very simple search strategy: just double the MSS. */
1705 mss_now = tcp_current_mss(sk);
1706 probe_size = 2 * tp->mss_cache;
1707 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1708 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1709 /* TODO: set timer for probe_converge_event */
1713 /* Have enough data in the send queue to probe? */
1714 if (tp->write_seq - tp->snd_nxt < size_needed)
1717 if (tp->snd_wnd < size_needed)
1719 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1722 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1723 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1724 if (!tcp_packets_in_flight(tp))
1730 /* We're allowed to probe. Build it now. */
1731 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1733 sk->sk_wmem_queued += nskb->truesize;
1734 sk_mem_charge(sk, nskb->truesize);
1736 skb = tcp_send_head(sk);
1738 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1739 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1740 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1741 TCP_SKB_CB(nskb)->sacked = 0;
1743 nskb->ip_summed = skb->ip_summed;
1745 tcp_insert_write_queue_before(nskb, skb, sk);
1748 tcp_for_write_queue_from_safe(skb, next, sk) {
1749 copy = min_t(int, skb->len, probe_size - len);
1750 if (nskb->ip_summed)
1751 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1753 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1754 skb_put(nskb, copy),
1757 if (skb->len <= copy) {
1758 /* We've eaten all the data from this skb.
1760 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1761 tcp_unlink_write_queue(skb, sk);
1762 sk_wmem_free_skb(sk, skb);
1764 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1765 ~(TCPHDR_FIN|TCPHDR_PSH);
1766 if (!skb_shinfo(skb)->nr_frags) {
1767 skb_pull(skb, copy);
1768 if (skb->ip_summed != CHECKSUM_PARTIAL)
1769 skb->csum = csum_partial(skb->data,
1772 __pskb_trim_head(skb, copy);
1773 tcp_set_skb_tso_segs(sk, skb, mss_now);
1775 TCP_SKB_CB(skb)->seq += copy;
1780 if (len >= probe_size)
1783 tcp_init_tso_segs(sk, nskb, nskb->len);
1785 /* We're ready to send. If this fails, the probe will
1786 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1787 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1788 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1789 /* Decrement cwnd here because we are sending
1790 * effectively two packets. */
1792 tcp_event_new_data_sent(sk, nskb);
1794 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1795 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1796 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1804 /* This routine writes packets to the network. It advances the
1805 * send_head. This happens as incoming acks open up the remote
1808 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1809 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1810 * account rare use of URG, this is not a big flaw.
1812 * Send at most one packet when push_one > 0. Temporarily ignore
1813 * cwnd limit to force at most one packet out when push_one == 2.
1815 * Returns true, if no segments are in flight and we have queued segments,
1816 * but cannot send anything now because of SWS or another problem.
1818 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1819 int push_one, gfp_t gfp)
1821 struct tcp_sock *tp = tcp_sk(sk);
1822 struct sk_buff *skb;
1823 unsigned int tso_segs, sent_pkts;
1830 /* Do MTU probing. */
1831 result = tcp_mtu_probe(sk);
1834 } else if (result > 0) {
1839 while ((skb = tcp_send_head(sk))) {
1842 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1845 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1846 goto repair; /* Skip network transmission */
1848 cwnd_quota = tcp_cwnd_test(tp, skb);
1851 /* Force out a loss probe pkt. */
1857 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1860 if (tso_segs == 1) {
1861 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1862 (tcp_skb_is_last(sk, skb) ?
1863 nonagle : TCP_NAGLE_PUSH))))
1866 if (!push_one && tcp_tso_should_defer(sk, skb))
1870 /* TCP Small Queues :
1871 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1873 * - better RTT estimation and ACK scheduling
1877 limit = max(skb->truesize, sk->sk_pacing_rate >> 10);
1879 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1880 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1885 if (tso_segs > 1 && !tcp_urg_mode(tp))
1886 limit = tcp_mss_split_point(sk, skb, mss_now,
1889 sk->sk_gso_max_segs));
1891 if (skb->len > limit &&
1892 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1895 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1897 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1901 /* Advance the send_head. This one is sent out.
1902 * This call will increment packets_out.
1904 tcp_event_new_data_sent(sk, skb);
1906 tcp_minshall_update(tp, mss_now, skb);
1907 sent_pkts += tcp_skb_pcount(skb);
1913 if (likely(sent_pkts)) {
1914 if (tcp_in_cwnd_reduction(sk))
1915 tp->prr_out += sent_pkts;
1917 /* Send one loss probe per tail loss episode. */
1919 tcp_schedule_loss_probe(sk);
1920 tcp_cwnd_validate(sk);
1923 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1926 bool tcp_schedule_loss_probe(struct sock *sk)
1928 struct inet_connection_sock *icsk = inet_csk(sk);
1929 struct tcp_sock *tp = tcp_sk(sk);
1930 u32 timeout, tlp_time_stamp, rto_time_stamp;
1931 u32 rtt = tp->srtt >> 3;
1933 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1935 /* No consecutive loss probes. */
1936 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1940 /* Don't do any loss probe on a Fast Open connection before 3WHS
1943 if (sk->sk_state == TCP_SYN_RECV)
1946 /* TLP is only scheduled when next timer event is RTO. */
1947 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
1950 /* Schedule a loss probe in 2*RTT for SACK capable connections
1951 * in Open state, that are either limited by cwnd or application.
1953 if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out ||
1954 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
1957 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
1961 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
1962 * for delayed ack when there's one outstanding packet.
1965 if (tp->packets_out == 1)
1966 timeout = max_t(u32, timeout,
1967 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
1968 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
1970 /* If RTO is shorter, just schedule TLP in its place. */
1971 tlp_time_stamp = tcp_time_stamp + timeout;
1972 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
1973 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
1974 s32 delta = rto_time_stamp - tcp_time_stamp;
1979 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
1984 /* When probe timeout (PTO) fires, send a new segment if one exists, else
1985 * retransmit the last segment.
1987 void tcp_send_loss_probe(struct sock *sk)
1989 struct tcp_sock *tp = tcp_sk(sk);
1990 struct sk_buff *skb;
1992 int mss = tcp_current_mss(sk);
1995 if (tcp_send_head(sk) != NULL) {
1996 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2000 /* At most one outstanding TLP retransmission. */
2001 if (tp->tlp_high_seq)
2004 /* Retransmit last segment. */
2005 skb = tcp_write_queue_tail(sk);
2009 pcount = tcp_skb_pcount(skb);
2010 if (WARN_ON(!pcount))
2013 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2014 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2016 skb = tcp_write_queue_tail(sk);
2019 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2022 /* Probe with zero data doesn't trigger fast recovery. */
2024 err = __tcp_retransmit_skb(sk, skb);
2026 /* Record snd_nxt for loss detection. */
2028 tp->tlp_high_seq = tp->snd_nxt;
2031 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2032 inet_csk(sk)->icsk_rto,
2036 NET_INC_STATS_BH(sock_net(sk),
2037 LINUX_MIB_TCPLOSSPROBES);
2041 /* Push out any pending frames which were held back due to
2042 * TCP_CORK or attempt at coalescing tiny packets.
2043 * The socket must be locked by the caller.
2045 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2048 /* If we are closed, the bytes will have to remain here.
2049 * In time closedown will finish, we empty the write queue and
2050 * all will be happy.
2052 if (unlikely(sk->sk_state == TCP_CLOSE))
2055 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2056 sk_gfp_atomic(sk, GFP_ATOMIC)))
2057 tcp_check_probe_timer(sk);
2060 /* Send _single_ skb sitting at the send head. This function requires
2061 * true push pending frames to setup probe timer etc.
2063 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2065 struct sk_buff *skb = tcp_send_head(sk);
2067 BUG_ON(!skb || skb->len < mss_now);
2069 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2072 /* This function returns the amount that we can raise the
2073 * usable window based on the following constraints
2075 * 1. The window can never be shrunk once it is offered (RFC 793)
2076 * 2. We limit memory per socket
2079 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2080 * RECV.NEXT + RCV.WIN fixed until:
2081 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2083 * i.e. don't raise the right edge of the window until you can raise
2084 * it at least MSS bytes.
2086 * Unfortunately, the recommended algorithm breaks header prediction,
2087 * since header prediction assumes th->window stays fixed.
2089 * Strictly speaking, keeping th->window fixed violates the receiver
2090 * side SWS prevention criteria. The problem is that under this rule
2091 * a stream of single byte packets will cause the right side of the
2092 * window to always advance by a single byte.
2094 * Of course, if the sender implements sender side SWS prevention
2095 * then this will not be a problem.
2097 * BSD seems to make the following compromise:
2099 * If the free space is less than the 1/4 of the maximum
2100 * space available and the free space is less than 1/2 mss,
2101 * then set the window to 0.
2102 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2103 * Otherwise, just prevent the window from shrinking
2104 * and from being larger than the largest representable value.
2106 * This prevents incremental opening of the window in the regime
2107 * where TCP is limited by the speed of the reader side taking
2108 * data out of the TCP receive queue. It does nothing about
2109 * those cases where the window is constrained on the sender side
2110 * because the pipeline is full.
2112 * BSD also seems to "accidentally" limit itself to windows that are a
2113 * multiple of MSS, at least until the free space gets quite small.
2114 * This would appear to be a side effect of the mbuf implementation.
2115 * Combining these two algorithms results in the observed behavior
2116 * of having a fixed window size at almost all times.
2118 * Below we obtain similar behavior by forcing the offered window to
2119 * a multiple of the mss when it is feasible to do so.
2121 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2122 * Regular options like TIMESTAMP are taken into account.
2124 u32 __tcp_select_window(struct sock *sk)
2126 struct inet_connection_sock *icsk = inet_csk(sk);
2127 struct tcp_sock *tp = tcp_sk(sk);
2128 /* MSS for the peer's data. Previous versions used mss_clamp
2129 * here. I don't know if the value based on our guesses
2130 * of peer's MSS is better for the performance. It's more correct
2131 * but may be worse for the performance because of rcv_mss
2132 * fluctuations. --SAW 1998/11/1
2134 int mss = icsk->icsk_ack.rcv_mss;
2135 int free_space = tcp_space(sk);
2136 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2139 if (mss > full_space)
2142 if (free_space < (full_space >> 1)) {
2143 icsk->icsk_ack.quick = 0;
2145 if (sk_under_memory_pressure(sk))
2146 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2149 if (free_space < mss)
2153 if (free_space > tp->rcv_ssthresh)
2154 free_space = tp->rcv_ssthresh;
2156 /* Don't do rounding if we are using window scaling, since the
2157 * scaled window will not line up with the MSS boundary anyway.
2159 window = tp->rcv_wnd;
2160 if (tp->rx_opt.rcv_wscale) {
2161 window = free_space;
2163 /* Advertise enough space so that it won't get scaled away.
2164 * Import case: prevent zero window announcement if
2165 * 1<<rcv_wscale > mss.
2167 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2168 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2169 << tp->rx_opt.rcv_wscale);
2171 /* Get the largest window that is a nice multiple of mss.
2172 * Window clamp already applied above.
2173 * If our current window offering is within 1 mss of the
2174 * free space we just keep it. This prevents the divide
2175 * and multiply from happening most of the time.
2176 * We also don't do any window rounding when the free space
2179 if (window <= free_space - mss || window > free_space)
2180 window = (free_space / mss) * mss;
2181 else if (mss == full_space &&
2182 free_space > window + (full_space >> 1))
2183 window = free_space;
2189 /* Collapses two adjacent SKB's during retransmission. */
2190 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2194 int skb_size, next_skb_size;
2196 skb_size = skb->len;
2197 next_skb_size = next_skb->len;
2199 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2201 tcp_highest_sack_combine(sk, next_skb, skb);
2203 tcp_unlink_write_queue(next_skb, sk);
2205 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2208 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2209 skb->ip_summed = CHECKSUM_PARTIAL;
2211 if (skb->ip_summed != CHECKSUM_PARTIAL)
2212 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2214 /* Update sequence range on original skb. */
2215 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2217 /* Merge over control information. This moves PSH/FIN etc. over */
2218 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2220 /* All done, get rid of second SKB and account for it so
2221 * packet counting does not break.
2223 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2225 /* changed transmit queue under us so clear hints */
2226 tcp_clear_retrans_hints_partial(tp);
2227 if (next_skb == tp->retransmit_skb_hint)
2228 tp->retransmit_skb_hint = skb;
2230 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2232 sk_wmem_free_skb(sk, next_skb);
2235 /* Check if coalescing SKBs is legal. */
2236 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2238 if (tcp_skb_pcount(skb) > 1)
2240 /* TODO: SACK collapsing could be used to remove this condition */
2241 if (skb_shinfo(skb)->nr_frags != 0)
2243 if (skb_cloned(skb))
2245 if (skb == tcp_send_head(sk))
2247 /* Some heurestics for collapsing over SACK'd could be invented */
2248 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2254 /* Collapse packets in the retransmit queue to make to create
2255 * less packets on the wire. This is only done on retransmission.
2257 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2260 struct tcp_sock *tp = tcp_sk(sk);
2261 struct sk_buff *skb = to, *tmp;
2264 if (!sysctl_tcp_retrans_collapse)
2266 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2269 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2270 if (!tcp_can_collapse(sk, skb))
2282 /* Punt if not enough space exists in the first SKB for
2283 * the data in the second
2285 if (skb->len > skb_availroom(to))
2288 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2291 tcp_collapse_retrans(sk, to);
2295 /* This retransmits one SKB. Policy decisions and retransmit queue
2296 * state updates are done by the caller. Returns non-zero if an
2297 * error occurred which prevented the send.
2299 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2301 struct tcp_sock *tp = tcp_sk(sk);
2302 struct inet_connection_sock *icsk = inet_csk(sk);
2303 unsigned int cur_mss;
2305 /* Inconslusive MTU probe */
2306 if (icsk->icsk_mtup.probe_size) {
2307 icsk->icsk_mtup.probe_size = 0;
2310 /* Do not sent more than we queued. 1/4 is reserved for possible
2311 * copying overhead: fragmentation, tunneling, mangling etc.
2313 if (atomic_read(&sk->sk_wmem_alloc) >
2314 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2317 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2318 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2320 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2324 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2325 return -EHOSTUNREACH; /* Routing failure or similar. */
2327 cur_mss = tcp_current_mss(sk);
2329 /* If receiver has shrunk his window, and skb is out of
2330 * new window, do not retransmit it. The exception is the
2331 * case, when window is shrunk to zero. In this case
2332 * our retransmit serves as a zero window probe.
2334 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2335 TCP_SKB_CB(skb)->seq != tp->snd_una)
2338 if (skb->len > cur_mss) {
2339 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2340 return -ENOMEM; /* We'll try again later. */
2342 int oldpcount = tcp_skb_pcount(skb);
2344 if (unlikely(oldpcount > 1)) {
2345 tcp_init_tso_segs(sk, skb, cur_mss);
2346 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2350 tcp_retrans_try_collapse(sk, skb, cur_mss);
2352 /* Some Solaris stacks overoptimize and ignore the FIN on a
2353 * retransmit when old data is attached. So strip it off
2354 * since it is cheap to do so and saves bytes on the network.
2357 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2358 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2359 if (!pskb_trim(skb, 0)) {
2360 /* Reuse, even though it does some unnecessary work */
2361 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2362 TCP_SKB_CB(skb)->tcp_flags);
2363 skb->ip_summed = CHECKSUM_NONE;
2367 /* Make a copy, if the first transmission SKB clone we made
2368 * is still in somebody's hands, else make a clone.
2370 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2372 /* make sure skb->data is aligned on arches that require it
2373 * and check if ack-trimming & collapsing extended the headroom
2374 * beyond what csum_start can cover.
2376 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2377 skb_headroom(skb) >= 0xFFFF)) {
2378 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2380 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2383 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2387 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2389 struct tcp_sock *tp = tcp_sk(sk);
2390 int err = __tcp_retransmit_skb(sk, skb);
2393 /* Update global TCP statistics. */
2394 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2396 tp->total_retrans++;
2398 #if FASTRETRANS_DEBUG > 0
2399 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2400 net_dbg_ratelimited("retrans_out leaked\n");
2403 if (!tp->retrans_out)
2404 tp->lost_retrans_low = tp->snd_nxt;
2405 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2406 tp->retrans_out += tcp_skb_pcount(skb);
2408 /* Save stamp of the first retransmit. */
2409 if (!tp->retrans_stamp)
2410 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2412 tp->undo_retrans += tcp_skb_pcount(skb);
2414 /* snd_nxt is stored to detect loss of retransmitted segment,
2415 * see tcp_input.c tcp_sacktag_write_queue().
2417 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2419 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2424 /* Check if we forward retransmits are possible in the current
2425 * window/congestion state.
2427 static bool tcp_can_forward_retransmit(struct sock *sk)
2429 const struct inet_connection_sock *icsk = inet_csk(sk);
2430 const struct tcp_sock *tp = tcp_sk(sk);
2432 /* Forward retransmissions are possible only during Recovery. */
2433 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2436 /* No forward retransmissions in Reno are possible. */
2437 if (tcp_is_reno(tp))
2440 /* Yeah, we have to make difficult choice between forward transmission
2441 * and retransmission... Both ways have their merits...
2443 * For now we do not retransmit anything, while we have some new
2444 * segments to send. In the other cases, follow rule 3 for
2445 * NextSeg() specified in RFC3517.
2448 if (tcp_may_send_now(sk))
2454 /* This gets called after a retransmit timeout, and the initially
2455 * retransmitted data is acknowledged. It tries to continue
2456 * resending the rest of the retransmit queue, until either
2457 * we've sent it all or the congestion window limit is reached.
2458 * If doing SACK, the first ACK which comes back for a timeout
2459 * based retransmit packet might feed us FACK information again.
2460 * If so, we use it to avoid unnecessarily retransmissions.
2462 void tcp_xmit_retransmit_queue(struct sock *sk)
2464 const struct inet_connection_sock *icsk = inet_csk(sk);
2465 struct tcp_sock *tp = tcp_sk(sk);
2466 struct sk_buff *skb;
2467 struct sk_buff *hole = NULL;
2470 int fwd_rexmitting = 0;
2472 if (!tp->packets_out)
2476 tp->retransmit_high = tp->snd_una;
2478 if (tp->retransmit_skb_hint) {
2479 skb = tp->retransmit_skb_hint;
2480 last_lost = TCP_SKB_CB(skb)->end_seq;
2481 if (after(last_lost, tp->retransmit_high))
2482 last_lost = tp->retransmit_high;
2484 skb = tcp_write_queue_head(sk);
2485 last_lost = tp->snd_una;
2488 tcp_for_write_queue_from(skb, sk) {
2489 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2491 if (skb == tcp_send_head(sk))
2493 /* we could do better than to assign each time */
2495 tp->retransmit_skb_hint = skb;
2497 /* Assume this retransmit will generate
2498 * only one packet for congestion window
2499 * calculation purposes. This works because
2500 * tcp_retransmit_skb() will chop up the
2501 * packet to be MSS sized and all the
2502 * packet counting works out.
2504 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2507 if (fwd_rexmitting) {
2509 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2511 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2513 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2514 tp->retransmit_high = last_lost;
2515 if (!tcp_can_forward_retransmit(sk))
2517 /* Backtrack if necessary to non-L'ed skb */
2525 } else if (!(sacked & TCPCB_LOST)) {
2526 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2531 last_lost = TCP_SKB_CB(skb)->end_seq;
2532 if (icsk->icsk_ca_state != TCP_CA_Loss)
2533 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2535 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2538 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2541 if (tcp_retransmit_skb(sk, skb))
2544 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2546 if (tcp_in_cwnd_reduction(sk))
2547 tp->prr_out += tcp_skb_pcount(skb);
2549 if (skb == tcp_write_queue_head(sk))
2550 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2551 inet_csk(sk)->icsk_rto,
2556 /* Send a fin. The caller locks the socket for us. This cannot be
2557 * allowed to fail queueing a FIN frame under any circumstances.
2559 void tcp_send_fin(struct sock *sk)
2561 struct tcp_sock *tp = tcp_sk(sk);
2562 struct sk_buff *skb = tcp_write_queue_tail(sk);
2565 /* Optimization, tack on the FIN if we have a queue of
2566 * unsent frames. But be careful about outgoing SACKS
2569 mss_now = tcp_current_mss(sk);
2571 if (tcp_send_head(sk) != NULL) {
2572 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2573 TCP_SKB_CB(skb)->end_seq++;
2576 /* Socket is locked, keep trying until memory is available. */
2578 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2585 /* Reserve space for headers and prepare control bits. */
2586 skb_reserve(skb, MAX_TCP_HEADER);
2587 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2588 tcp_init_nondata_skb(skb, tp->write_seq,
2589 TCPHDR_ACK | TCPHDR_FIN);
2590 tcp_queue_skb(sk, skb);
2592 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2595 /* We get here when a process closes a file descriptor (either due to
2596 * an explicit close() or as a byproduct of exit()'ing) and there
2597 * was unread data in the receive queue. This behavior is recommended
2598 * by RFC 2525, section 2.17. -DaveM
2600 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2602 struct sk_buff *skb;
2604 /* NOTE: No TCP options attached and we never retransmit this. */
2605 skb = alloc_skb(MAX_TCP_HEADER, priority);
2607 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2611 /* Reserve space for headers and prepare control bits. */
2612 skb_reserve(skb, MAX_TCP_HEADER);
2613 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2614 TCPHDR_ACK | TCPHDR_RST);
2616 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2617 if (tcp_transmit_skb(sk, skb, 0, priority))
2618 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2620 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2623 /* Send a crossed SYN-ACK during socket establishment.
2624 * WARNING: This routine must only be called when we have already sent
2625 * a SYN packet that crossed the incoming SYN that caused this routine
2626 * to get called. If this assumption fails then the initial rcv_wnd
2627 * and rcv_wscale values will not be correct.
2629 int tcp_send_synack(struct sock *sk)
2631 struct sk_buff *skb;
2633 skb = tcp_write_queue_head(sk);
2634 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2635 pr_debug("%s: wrong queue state\n", __func__);
2638 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2639 if (skb_cloned(skb)) {
2640 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2643 tcp_unlink_write_queue(skb, sk);
2644 skb_header_release(nskb);
2645 __tcp_add_write_queue_head(sk, nskb);
2646 sk_wmem_free_skb(sk, skb);
2647 sk->sk_wmem_queued += nskb->truesize;
2648 sk_mem_charge(sk, nskb->truesize);
2652 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2653 TCP_ECN_send_synack(tcp_sk(sk), skb);
2655 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2656 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2660 * tcp_make_synack - Prepare a SYN-ACK.
2661 * sk: listener socket
2662 * dst: dst entry attached to the SYNACK
2663 * req: request_sock pointer
2665 * Allocate one skb and build a SYNACK packet.
2666 * @dst is consumed : Caller should not use it again.
2668 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2669 struct request_sock *req,
2670 struct tcp_fastopen_cookie *foc)
2672 struct tcp_out_options opts;
2673 struct inet_request_sock *ireq = inet_rsk(req);
2674 struct tcp_sock *tp = tcp_sk(sk);
2676 struct sk_buff *skb;
2677 struct tcp_md5sig_key *md5;
2678 int tcp_header_size;
2681 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC);
2682 if (unlikely(!skb)) {
2686 /* Reserve space for headers. */
2687 skb_reserve(skb, MAX_TCP_HEADER);
2689 skb_dst_set(skb, dst);
2690 security_skb_owned_by(skb, sk);
2692 mss = dst_metric_advmss(dst);
2693 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2694 mss = tp->rx_opt.user_mss;
2696 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2698 /* Set this up on the first call only */
2699 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2701 /* limit the window selection if the user enforce a smaller rx buffer */
2702 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2703 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2704 req->window_clamp = tcp_full_space(sk);
2706 /* tcp_full_space because it is guaranteed to be the first packet */
2707 tcp_select_initial_window(tcp_full_space(sk),
2708 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2713 dst_metric(dst, RTAX_INITRWND));
2714 ireq->rcv_wscale = rcv_wscale;
2717 memset(&opts, 0, sizeof(opts));
2718 #ifdef CONFIG_SYN_COOKIES
2719 if (unlikely(req->cookie_ts))
2720 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2723 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2724 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2727 skb_push(skb, tcp_header_size);
2728 skb_reset_transport_header(skb);
2731 memset(th, 0, sizeof(struct tcphdr));
2734 TCP_ECN_make_synack(req, th);
2735 th->source = ireq->loc_port;
2736 th->dest = ireq->rmt_port;
2737 /* Setting of flags are superfluous here for callers (and ECE is
2738 * not even correctly set)
2740 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2741 TCPHDR_SYN | TCPHDR_ACK);
2743 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2744 /* XXX data is queued and acked as is. No buffer/window check */
2745 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2747 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2748 th->window = htons(min(req->rcv_wnd, 65535U));
2749 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2750 th->doff = (tcp_header_size >> 2);
2751 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2753 #ifdef CONFIG_TCP_MD5SIG
2754 /* Okay, we have all we need - do the md5 hash if needed */
2756 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2757 md5, NULL, req, skb);
2763 EXPORT_SYMBOL(tcp_make_synack);
2765 /* Do all connect socket setups that can be done AF independent. */
2766 void tcp_connect_init(struct sock *sk)
2768 const struct dst_entry *dst = __sk_dst_get(sk);
2769 struct tcp_sock *tp = tcp_sk(sk);
2772 /* We'll fix this up when we get a response from the other end.
2773 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2775 tp->tcp_header_len = sizeof(struct tcphdr) +
2776 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2778 #ifdef CONFIG_TCP_MD5SIG
2779 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2780 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2783 /* If user gave his TCP_MAXSEG, record it to clamp */
2784 if (tp->rx_opt.user_mss)
2785 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2788 tcp_sync_mss(sk, dst_mtu(dst));
2790 if (!tp->window_clamp)
2791 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2792 tp->advmss = dst_metric_advmss(dst);
2793 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2794 tp->advmss = tp->rx_opt.user_mss;
2796 tcp_initialize_rcv_mss(sk);
2798 /* limit the window selection if the user enforce a smaller rx buffer */
2799 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2800 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2801 tp->window_clamp = tcp_full_space(sk);
2803 tcp_select_initial_window(tcp_full_space(sk),
2804 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2807 sysctl_tcp_window_scaling,
2809 dst_metric(dst, RTAX_INITRWND));
2811 tp->rx_opt.rcv_wscale = rcv_wscale;
2812 tp->rcv_ssthresh = tp->rcv_wnd;
2815 sock_reset_flag(sk, SOCK_DONE);
2818 tp->snd_una = tp->write_seq;
2819 tp->snd_sml = tp->write_seq;
2820 tp->snd_up = tp->write_seq;
2821 tp->snd_nxt = tp->write_seq;
2823 if (likely(!tp->repair))
2826 tp->rcv_tstamp = tcp_time_stamp;
2827 tp->rcv_wup = tp->rcv_nxt;
2828 tp->copied_seq = tp->rcv_nxt;
2830 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2831 inet_csk(sk)->icsk_retransmits = 0;
2832 tcp_clear_retrans(tp);
2835 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2837 struct tcp_sock *tp = tcp_sk(sk);
2838 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2840 tcb->end_seq += skb->len;
2841 skb_header_release(skb);
2842 __tcp_add_write_queue_tail(sk, skb);
2843 sk->sk_wmem_queued += skb->truesize;
2844 sk_mem_charge(sk, skb->truesize);
2845 tp->write_seq = tcb->end_seq;
2846 tp->packets_out += tcp_skb_pcount(skb);
2849 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2850 * queue a data-only packet after the regular SYN, such that regular SYNs
2851 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2852 * only the SYN sequence, the data are retransmitted in the first ACK.
2853 * If cookie is not cached or other error occurs, falls back to send a
2854 * regular SYN with Fast Open cookie request option.
2856 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2858 struct tcp_sock *tp = tcp_sk(sk);
2859 struct tcp_fastopen_request *fo = tp->fastopen_req;
2860 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2861 struct sk_buff *syn_data = NULL, *data;
2862 unsigned long last_syn_loss = 0;
2864 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2865 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2866 &syn_loss, &last_syn_loss);
2867 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2869 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2870 fo->cookie.len = -1;
2874 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2875 fo->cookie.len = -1;
2876 else if (fo->cookie.len <= 0)
2879 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2880 * user-MSS. Reserve maximum option space for middleboxes that add
2881 * private TCP options. The cost is reduced data space in SYN :(
2883 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2884 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2885 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2886 MAX_TCP_OPTION_SPACE;
2888 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2890 if (syn_data == NULL)
2893 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2894 struct iovec *iov = &fo->data->msg_iov[i];
2895 unsigned char __user *from = iov->iov_base;
2896 int len = iov->iov_len;
2898 if (syn_data->len + len > space)
2899 len = space - syn_data->len;
2900 else if (i + 1 == iovlen)
2901 /* No more data pending in inet_wait_for_connect() */
2904 if (skb_add_data(syn_data, from, len))
2908 /* Queue a data-only packet after the regular SYN for retransmission */
2909 data = pskb_copy(syn_data, sk->sk_allocation);
2912 TCP_SKB_CB(data)->seq++;
2913 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2914 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2915 tcp_connect_queue_skb(sk, data);
2916 fo->copied = data->len;
2918 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2919 tp->syn_data = (fo->copied > 0);
2920 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2926 /* Send a regular SYN with Fast Open cookie request option */
2927 if (fo->cookie.len > 0)
2929 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2931 tp->syn_fastopen = 0;
2932 kfree_skb(syn_data);
2934 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2938 /* Build a SYN and send it off. */
2939 int tcp_connect(struct sock *sk)
2941 struct tcp_sock *tp = tcp_sk(sk);
2942 struct sk_buff *buff;
2945 tcp_connect_init(sk);
2947 if (unlikely(tp->repair)) {
2948 tcp_finish_connect(sk, NULL);
2952 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2953 if (unlikely(buff == NULL))
2956 /* Reserve space for headers. */
2957 skb_reserve(buff, MAX_TCP_HEADER);
2959 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2960 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
2961 tcp_connect_queue_skb(sk, buff);
2962 TCP_ECN_send_syn(sk, buff);
2964 /* Send off SYN; include data in Fast Open. */
2965 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
2966 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2967 if (err == -ECONNREFUSED)
2970 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2971 * in order to make this packet get counted in tcpOutSegs.
2973 tp->snd_nxt = tp->write_seq;
2974 tp->pushed_seq = tp->write_seq;
2975 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2977 /* Timer for repeating the SYN until an answer. */
2978 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2979 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2982 EXPORT_SYMBOL(tcp_connect);
2984 /* Send out a delayed ack, the caller does the policy checking
2985 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2988 void tcp_send_delayed_ack(struct sock *sk)
2990 struct inet_connection_sock *icsk = inet_csk(sk);
2991 int ato = icsk->icsk_ack.ato;
2992 unsigned long timeout;
2994 if (ato > TCP_DELACK_MIN) {
2995 const struct tcp_sock *tp = tcp_sk(sk);
2996 int max_ato = HZ / 2;
2998 if (icsk->icsk_ack.pingpong ||
2999 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3000 max_ato = TCP_DELACK_MAX;
3002 /* Slow path, intersegment interval is "high". */
3004 /* If some rtt estimate is known, use it to bound delayed ack.
3005 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3009 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3015 ato = min(ato, max_ato);
3018 /* Stay within the limit we were given */
3019 timeout = jiffies + ato;
3021 /* Use new timeout only if there wasn't a older one earlier. */
3022 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3023 /* If delack timer was blocked or is about to expire,
3026 if (icsk->icsk_ack.blocked ||
3027 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3032 if (!time_before(timeout, icsk->icsk_ack.timeout))
3033 timeout = icsk->icsk_ack.timeout;
3035 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3036 icsk->icsk_ack.timeout = timeout;
3037 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3040 /* This routine sends an ack and also updates the window. */
3041 void tcp_send_ack(struct sock *sk)
3043 struct sk_buff *buff;
3045 /* If we have been reset, we may not send again. */
3046 if (sk->sk_state == TCP_CLOSE)
3049 /* We are not putting this on the write queue, so
3050 * tcp_transmit_skb() will set the ownership to this
3053 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3055 inet_csk_schedule_ack(sk);
3056 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3057 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3058 TCP_DELACK_MAX, TCP_RTO_MAX);
3062 /* Reserve space for headers and prepare control bits. */
3063 skb_reserve(buff, MAX_TCP_HEADER);
3064 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3066 /* Send it off, this clears delayed acks for us. */
3067 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3068 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3071 /* This routine sends a packet with an out of date sequence
3072 * number. It assumes the other end will try to ack it.
3074 * Question: what should we make while urgent mode?
3075 * 4.4BSD forces sending single byte of data. We cannot send
3076 * out of window data, because we have SND.NXT==SND.MAX...
3078 * Current solution: to send TWO zero-length segments in urgent mode:
3079 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3080 * out-of-date with SND.UNA-1 to probe window.
3082 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3084 struct tcp_sock *tp = tcp_sk(sk);
3085 struct sk_buff *skb;
3087 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3088 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3092 /* Reserve space for headers and set control bits. */
3093 skb_reserve(skb, MAX_TCP_HEADER);
3094 /* Use a previous sequence. This should cause the other
3095 * end to send an ack. Don't queue or clone SKB, just
3098 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3099 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3100 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3103 void tcp_send_window_probe(struct sock *sk)
3105 if (sk->sk_state == TCP_ESTABLISHED) {
3106 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3107 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3108 tcp_xmit_probe_skb(sk, 0);
3112 /* Initiate keepalive or window probe from timer. */
3113 int tcp_write_wakeup(struct sock *sk)
3115 struct tcp_sock *tp = tcp_sk(sk);
3116 struct sk_buff *skb;
3118 if (sk->sk_state == TCP_CLOSE)
3121 if ((skb = tcp_send_head(sk)) != NULL &&
3122 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3124 unsigned int mss = tcp_current_mss(sk);
3125 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3127 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3128 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3130 /* We are probing the opening of a window
3131 * but the window size is != 0
3132 * must have been a result SWS avoidance ( sender )
3134 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3136 seg_size = min(seg_size, mss);
3137 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3138 if (tcp_fragment(sk, skb, seg_size, mss))
3140 } else if (!tcp_skb_pcount(skb))
3141 tcp_set_skb_tso_segs(sk, skb, mss);
3143 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3144 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3145 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3147 tcp_event_new_data_sent(sk, skb);
3150 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3151 tcp_xmit_probe_skb(sk, 1);
3152 return tcp_xmit_probe_skb(sk, 0);
3156 /* A window probe timeout has occurred. If window is not closed send
3157 * a partial packet else a zero probe.
3159 void tcp_send_probe0(struct sock *sk)
3161 struct inet_connection_sock *icsk = inet_csk(sk);
3162 struct tcp_sock *tp = tcp_sk(sk);
3165 err = tcp_write_wakeup(sk);
3167 if (tp->packets_out || !tcp_send_head(sk)) {
3168 /* Cancel probe timer, if it is not required. */
3169 icsk->icsk_probes_out = 0;
3170 icsk->icsk_backoff = 0;
3175 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3176 icsk->icsk_backoff++;
3177 icsk->icsk_probes_out++;
3178 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3179 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3182 /* If packet was not sent due to local congestion,
3183 * do not backoff and do not remember icsk_probes_out.
3184 * Let local senders to fight for local resources.
3186 * Use accumulated backoff yet.
3188 if (!icsk->icsk_probes_out)
3189 icsk->icsk_probes_out = 1;
3190 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3191 min(icsk->icsk_rto << icsk->icsk_backoff,
3192 TCP_RESOURCE_PROBE_INTERVAL),