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>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 int push_one, gfp_t gfp);
51 /* Account for new data that has been sent to the network. */
52 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
54 struct inet_connection_sock *icsk = inet_csk(sk);
55 struct tcp_sock *tp = tcp_sk(sk);
56 unsigned int prior_packets = tp->packets_out;
58 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
60 __skb_unlink(skb, &sk->sk_write_queue);
61 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
63 if (tp->highest_sack == NULL)
64 tp->highest_sack = skb;
66 tp->packets_out += tcp_skb_pcount(skb);
67 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
70 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
74 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
75 * window scaling factor due to loss of precision.
76 * If window has been shrunk, what should we make? It is not clear at all.
77 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
78 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
79 * invalid. OK, let's make this for now:
81 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
83 const struct tcp_sock *tp = tcp_sk(sk);
85 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
86 (tp->rx_opt.wscale_ok &&
87 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
90 return tcp_wnd_end(tp);
93 /* Calculate mss to advertise in SYN segment.
94 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
96 * 1. It is independent of path mtu.
97 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
98 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
99 * attached devices, because some buggy hosts are confused by
101 * 4. We do not make 3, we advertise MSS, calculated from first
102 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
103 * This may be overridden via information stored in routing table.
104 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
105 * probably even Jumbo".
107 static __u16 tcp_advertise_mss(struct sock *sk)
109 struct tcp_sock *tp = tcp_sk(sk);
110 const struct dst_entry *dst = __sk_dst_get(sk);
111 int mss = tp->advmss;
114 unsigned int metric = dst_metric_advmss(dst);
125 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
126 * This is the first part of cwnd validation mechanism.
128 void tcp_cwnd_restart(struct sock *sk, s32 delta)
130 struct tcp_sock *tp = tcp_sk(sk);
131 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
132 u32 cwnd = tp->snd_cwnd;
134 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
136 tp->snd_ssthresh = tcp_current_ssthresh(sk);
137 restart_cwnd = min(restart_cwnd, cwnd);
139 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
141 tp->snd_cwnd = max(cwnd, restart_cwnd);
142 tp->snd_cwnd_stamp = tcp_jiffies32;
143 tp->snd_cwnd_used = 0;
146 /* Congestion state accounting after a packet has been sent. */
147 static void tcp_event_data_sent(struct tcp_sock *tp,
150 struct inet_connection_sock *icsk = inet_csk(sk);
151 const u32 now = tcp_jiffies32;
153 if (tcp_packets_in_flight(tp) == 0)
154 tcp_ca_event(sk, CA_EVENT_TX_START);
158 /* If it is a reply for ato after last received
159 * packet, enter pingpong mode.
161 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
162 icsk->icsk_ack.pingpong = 1;
165 /* Account for an ACK we sent. */
166 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
169 struct tcp_sock *tp = tcp_sk(sk);
171 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
172 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
173 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
174 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
175 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
179 if (unlikely(rcv_nxt != tp->rcv_nxt))
180 return; /* Special ACK sent by DCTCP to reflect ECN */
181 tcp_dec_quickack_mode(sk, pkts);
182 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
185 /* Determine a window scaling and initial window to offer.
186 * Based on the assumption that the given amount of space
187 * will be offered. Store the results in the tp structure.
188 * NOTE: for smooth operation initial space offering should
189 * be a multiple of mss if possible. We assume here that mss >= 1.
190 * This MUST be enforced by all callers.
192 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
193 __u32 *rcv_wnd, __u32 *window_clamp,
194 int wscale_ok, __u8 *rcv_wscale,
197 unsigned int space = (__space < 0 ? 0 : __space);
199 /* If no clamp set the clamp to the max possible scaled window */
200 if (*window_clamp == 0)
201 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
202 space = min(*window_clamp, space);
204 /* Quantize space offering to a multiple of mss if possible. */
206 space = rounddown(space, mss);
208 /* NOTE: offering an initial window larger than 32767
209 * will break some buggy TCP stacks. If the admin tells us
210 * it is likely we could be speaking with such a buggy stack
211 * we will truncate our initial window offering to 32K-1
212 * unless the remote has sent us a window scaling option,
213 * which we interpret as a sign the remote TCP is not
214 * misinterpreting the window field as a signed quantity.
216 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
217 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
219 (*rcv_wnd) = min_t(u32, space, U16_MAX);
222 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
226 /* Set window scaling on max possible window */
227 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
228 space = max_t(u32, space, sysctl_rmem_max);
229 space = min_t(u32, space, *window_clamp);
230 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
235 /* Set the clamp no higher than max representable value */
236 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
238 EXPORT_SYMBOL(tcp_select_initial_window);
240 /* Chose a new window to advertise, update state in tcp_sock for the
241 * socket, and return result with RFC1323 scaling applied. The return
242 * value can be stuffed directly into th->window for an outgoing
245 static u16 tcp_select_window(struct sock *sk)
247 struct tcp_sock *tp = tcp_sk(sk);
248 u32 old_win = tp->rcv_wnd;
249 u32 cur_win = tcp_receive_window(tp);
250 u32 new_win = __tcp_select_window(sk);
252 /* Never shrink the offered window */
253 if (new_win < cur_win) {
254 /* Danger Will Robinson!
255 * Don't update rcv_wup/rcv_wnd here or else
256 * we will not be able to advertise a zero
257 * window in time. --DaveM
259 * Relax Will Robinson.
262 NET_INC_STATS(sock_net(sk),
263 LINUX_MIB_TCPWANTZEROWINDOWADV);
264 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
266 tp->rcv_wnd = new_win;
267 tp->rcv_wup = tp->rcv_nxt;
269 /* Make sure we do not exceed the maximum possible
272 if (!tp->rx_opt.rcv_wscale &&
273 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
274 new_win = min(new_win, MAX_TCP_WINDOW);
276 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
278 /* RFC1323 scaling applied */
279 new_win >>= tp->rx_opt.rcv_wscale;
281 /* If we advertise zero window, disable fast path. */
285 NET_INC_STATS(sock_net(sk),
286 LINUX_MIB_TCPTOZEROWINDOWADV);
287 } else if (old_win == 0) {
288 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
294 /* Packet ECN state for a SYN-ACK */
295 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
297 const struct tcp_sock *tp = tcp_sk(sk);
299 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
300 if (!(tp->ecn_flags & TCP_ECN_OK))
301 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
302 else if (tcp_ca_needs_ecn(sk) ||
303 tcp_bpf_ca_needs_ecn(sk))
307 /* Packet ECN state for a SYN. */
308 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
310 struct tcp_sock *tp = tcp_sk(sk);
311 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
312 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
313 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
316 const struct dst_entry *dst = __sk_dst_get(sk);
318 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
325 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
326 tp->ecn_flags = TCP_ECN_OK;
327 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
332 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
334 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
335 /* tp->ecn_flags are cleared at a later point in time when
336 * SYN ACK is ultimatively being received.
338 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
342 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
344 if (inet_rsk(req)->ecn_ok)
348 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
351 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
352 struct tcphdr *th, int tcp_header_len)
354 struct tcp_sock *tp = tcp_sk(sk);
356 if (tp->ecn_flags & TCP_ECN_OK) {
357 /* Not-retransmitted data segment: set ECT and inject CWR. */
358 if (skb->len != tcp_header_len &&
359 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
361 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
362 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
364 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
366 } else if (!tcp_ca_needs_ecn(sk)) {
367 /* ACK or retransmitted segment: clear ECT|CE */
368 INET_ECN_dontxmit(sk);
370 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
375 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
376 * auto increment end seqno.
378 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
380 skb->ip_summed = CHECKSUM_PARTIAL;
382 TCP_SKB_CB(skb)->tcp_flags = flags;
383 TCP_SKB_CB(skb)->sacked = 0;
385 tcp_skb_pcount_set(skb, 1);
387 TCP_SKB_CB(skb)->seq = seq;
388 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
390 TCP_SKB_CB(skb)->end_seq = seq;
393 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
395 return tp->snd_una != tp->snd_up;
398 #define OPTION_SACK_ADVERTISE (1 << 0)
399 #define OPTION_TS (1 << 1)
400 #define OPTION_MD5 (1 << 2)
401 #define OPTION_WSCALE (1 << 3)
402 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
403 #define OPTION_SMC (1 << 9)
405 static void smc_options_write(__be32 *ptr, u16 *options)
407 #if IS_ENABLED(CONFIG_SMC)
408 if (static_branch_unlikely(&tcp_have_smc)) {
409 if (unlikely(OPTION_SMC & *options)) {
410 *ptr++ = htonl((TCPOPT_NOP << 24) |
413 (TCPOLEN_EXP_SMC_BASE));
414 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
420 struct tcp_out_options {
421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
425 u8 hash_size; /* bytes in hash_location */
426 __u8 *hash_location; /* temporary pointer, overloaded */
427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
431 /* Write previously computed TCP options to the packet.
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
447 u16 options = opts->options; /* mungable copy */
449 if (unlikely(OPTION_MD5 & options)) {
450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
469 options &= ~OPTION_SACK_ADVERTISE;
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
473 (TCPOPT_TIMESTAMP << 8) |
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_SACK_PERM << 8) |
487 if (unlikely(OPTION_WSCALE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503 TCPOLEN_SACK_PERBLOCK)));
505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
511 tp->rx_opt.dsack = 0;
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517 u32 len; /* Fast Open option length */
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
535 ptr += (len + 3) >> 2;
538 smc_options_write(ptr, &options);
541 static void smc_set_option(const struct tcp_sock *tp,
542 struct tcp_out_options *opts,
543 unsigned int *remaining)
545 #if IS_ENABLED(CONFIG_SMC)
546 if (static_branch_unlikely(&tcp_have_smc)) {
548 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
549 opts->options |= OPTION_SMC;
550 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
557 static void smc_set_option_cond(const struct tcp_sock *tp,
558 const struct inet_request_sock *ireq,
559 struct tcp_out_options *opts,
560 unsigned int *remaining)
562 #if IS_ENABLED(CONFIG_SMC)
563 if (static_branch_unlikely(&tcp_have_smc)) {
564 if (tp->syn_smc && ireq->smc_ok) {
565 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
566 opts->options |= OPTION_SMC;
567 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
574 /* Compute TCP options for SYN packets. This is not the final
575 * network wire format yet.
577 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
578 struct tcp_out_options *opts,
579 struct tcp_md5sig_key **md5)
581 struct tcp_sock *tp = tcp_sk(sk);
582 unsigned int remaining = MAX_TCP_OPTION_SPACE;
583 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
586 #ifdef CONFIG_TCP_MD5SIG
587 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
588 *md5 = tp->af_specific->md5_lookup(sk, sk);
590 opts->options |= OPTION_MD5;
591 remaining -= TCPOLEN_MD5SIG_ALIGNED;
596 /* We always get an MSS option. The option bytes which will be seen in
597 * normal data packets should timestamps be used, must be in the MSS
598 * advertised. But we subtract them from tp->mss_cache so that
599 * calculations in tcp_sendmsg are simpler etc. So account for this
600 * fact here if necessary. If we don't do this correctly, as a
601 * receiver we won't recognize data packets as being full sized when we
602 * should, and thus we won't abide by the delayed ACK rules correctly.
603 * SACKs don't matter, we never delay an ACK when we have any of those
605 opts->mss = tcp_advertise_mss(sk);
606 remaining -= TCPOLEN_MSS_ALIGNED;
608 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
609 opts->options |= OPTION_TS;
610 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
611 opts->tsecr = tp->rx_opt.ts_recent;
612 remaining -= TCPOLEN_TSTAMP_ALIGNED;
614 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
615 opts->ws = tp->rx_opt.rcv_wscale;
616 opts->options |= OPTION_WSCALE;
617 remaining -= TCPOLEN_WSCALE_ALIGNED;
619 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
620 opts->options |= OPTION_SACK_ADVERTISE;
621 if (unlikely(!(OPTION_TS & opts->options)))
622 remaining -= TCPOLEN_SACKPERM_ALIGNED;
625 if (fastopen && fastopen->cookie.len >= 0) {
626 u32 need = fastopen->cookie.len;
628 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
629 TCPOLEN_FASTOPEN_BASE;
630 need = (need + 3) & ~3U; /* Align to 32 bits */
631 if (remaining >= need) {
632 opts->options |= OPTION_FAST_OPEN_COOKIE;
633 opts->fastopen_cookie = &fastopen->cookie;
635 tp->syn_fastopen = 1;
636 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
640 smc_set_option(tp, opts, &remaining);
642 return MAX_TCP_OPTION_SPACE - remaining;
645 /* Set up TCP options for SYN-ACKs. */
646 static unsigned int tcp_synack_options(const struct sock *sk,
647 struct request_sock *req,
648 unsigned int mss, struct sk_buff *skb,
649 struct tcp_out_options *opts,
650 const struct tcp_md5sig_key *md5,
651 struct tcp_fastopen_cookie *foc,
652 enum tcp_synack_type synack_type)
654 struct inet_request_sock *ireq = inet_rsk(req);
655 unsigned int remaining = MAX_TCP_OPTION_SPACE;
657 #ifdef CONFIG_TCP_MD5SIG
659 opts->options |= OPTION_MD5;
660 remaining -= TCPOLEN_MD5SIG_ALIGNED;
662 /* We can't fit any SACK blocks in a packet with MD5 + TS
663 * options. There was discussion about disabling SACK
664 * rather than TS in order to fit in better with old,
665 * buggy kernels, but that was deemed to be unnecessary.
667 if (synack_type != TCP_SYNACK_COOKIE)
668 ireq->tstamp_ok &= !ireq->sack_ok;
672 /* We always send an MSS option. */
674 remaining -= TCPOLEN_MSS_ALIGNED;
676 if (likely(ireq->wscale_ok)) {
677 opts->ws = ireq->rcv_wscale;
678 opts->options |= OPTION_WSCALE;
679 remaining -= TCPOLEN_WSCALE_ALIGNED;
681 if (likely(ireq->tstamp_ok)) {
682 opts->options |= OPTION_TS;
683 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
684 opts->tsecr = req->ts_recent;
685 remaining -= TCPOLEN_TSTAMP_ALIGNED;
687 if (likely(ireq->sack_ok)) {
688 opts->options |= OPTION_SACK_ADVERTISE;
689 if (unlikely(!ireq->tstamp_ok))
690 remaining -= TCPOLEN_SACKPERM_ALIGNED;
692 if (foc != NULL && foc->len >= 0) {
695 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
696 TCPOLEN_FASTOPEN_BASE;
697 need = (need + 3) & ~3U; /* Align to 32 bits */
698 if (remaining >= need) {
699 opts->options |= OPTION_FAST_OPEN_COOKIE;
700 opts->fastopen_cookie = foc;
705 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
707 return MAX_TCP_OPTION_SPACE - remaining;
710 /* Compute TCP options for ESTABLISHED sockets. This is not the
711 * final wire format yet.
713 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
714 struct tcp_out_options *opts,
715 struct tcp_md5sig_key **md5)
717 struct tcp_sock *tp = tcp_sk(sk);
718 unsigned int size = 0;
719 unsigned int eff_sacks;
724 #ifdef CONFIG_TCP_MD5SIG
725 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
726 *md5 = tp->af_specific->md5_lookup(sk, sk);
728 opts->options |= OPTION_MD5;
729 size += TCPOLEN_MD5SIG_ALIGNED;
734 if (likely(tp->rx_opt.tstamp_ok)) {
735 opts->options |= OPTION_TS;
736 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
737 opts->tsecr = tp->rx_opt.ts_recent;
738 size += TCPOLEN_TSTAMP_ALIGNED;
741 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
742 if (unlikely(eff_sacks)) {
743 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
744 opts->num_sack_blocks =
745 min_t(unsigned int, eff_sacks,
746 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
747 TCPOLEN_SACK_PERBLOCK);
748 if (likely(opts->num_sack_blocks))
749 size += TCPOLEN_SACK_BASE_ALIGNED +
750 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
757 /* TCP SMALL QUEUES (TSQ)
759 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
760 * to reduce RTT and bufferbloat.
761 * We do this using a special skb destructor (tcp_wfree).
763 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
764 * needs to be reallocated in a driver.
765 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
767 * Since transmit from skb destructor is forbidden, we use a tasklet
768 * to process all sockets that eventually need to send more skbs.
769 * We use one tasklet per cpu, with its own queue of sockets.
772 struct tasklet_struct tasklet;
773 struct list_head head; /* queue of tcp sockets */
775 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
777 static void tcp_tsq_write(struct sock *sk)
779 if ((1 << sk->sk_state) &
780 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
781 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
782 struct tcp_sock *tp = tcp_sk(sk);
784 if (tp->lost_out > tp->retrans_out &&
785 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
786 tcp_mstamp_refresh(tp);
787 tcp_xmit_retransmit_queue(sk);
790 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
795 static void tcp_tsq_handler(struct sock *sk)
798 if (!sock_owned_by_user(sk))
800 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
805 * One tasklet per cpu tries to send more skbs.
806 * We run in tasklet context but need to disable irqs when
807 * transferring tsq->head because tcp_wfree() might
808 * interrupt us (non NAPI drivers)
810 static void tcp_tasklet_func(unsigned long data)
812 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
815 struct list_head *q, *n;
819 local_irq_save(flags);
820 list_splice_init(&tsq->head, &list);
821 local_irq_restore(flags);
823 list_for_each_safe(q, n, &list) {
824 tp = list_entry(q, struct tcp_sock, tsq_node);
825 list_del(&tp->tsq_node);
827 sk = (struct sock *)tp;
828 smp_mb__before_atomic();
829 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
836 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
837 TCPF_WRITE_TIMER_DEFERRED | \
838 TCPF_DELACK_TIMER_DEFERRED | \
839 TCPF_MTU_REDUCED_DEFERRED)
841 * tcp_release_cb - tcp release_sock() callback
844 * called from release_sock() to perform protocol dependent
845 * actions before socket release.
847 void tcp_release_cb(struct sock *sk)
849 unsigned long flags, nflags;
851 /* perform an atomic operation only if at least one flag is set */
853 flags = sk->sk_tsq_flags;
854 if (!(flags & TCP_DEFERRED_ALL))
856 nflags = flags & ~TCP_DEFERRED_ALL;
857 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
859 if (flags & TCPF_TSQ_DEFERRED) {
863 /* Here begins the tricky part :
864 * We are called from release_sock() with :
866 * 2) sk_lock.slock spinlock held
867 * 3) socket owned by us (sk->sk_lock.owned == 1)
869 * But following code is meant to be called from BH handlers,
870 * so we should keep BH disabled, but early release socket ownership
872 sock_release_ownership(sk);
874 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
875 tcp_write_timer_handler(sk);
878 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
879 tcp_delack_timer_handler(sk);
882 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
883 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
887 EXPORT_SYMBOL(tcp_release_cb);
889 void __init tcp_tasklet_init(void)
893 for_each_possible_cpu(i) {
894 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
896 INIT_LIST_HEAD(&tsq->head);
897 tasklet_init(&tsq->tasklet,
904 * Write buffer destructor automatically called from kfree_skb.
905 * We can't xmit new skbs from this context, as we might already
908 void tcp_wfree(struct sk_buff *skb)
910 struct sock *sk = skb->sk;
911 struct tcp_sock *tp = tcp_sk(sk);
912 unsigned long flags, nval, oval;
914 /* Keep one reference on sk_wmem_alloc.
915 * Will be released by sk_free() from here or tcp_tasklet_func()
917 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
919 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
920 * Wait until our queues (qdisc + devices) are drained.
922 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
923 * - chance for incoming ACK (processed by another cpu maybe)
924 * to migrate this flow (skb->ooo_okay will be eventually set)
926 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
929 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
930 struct tsq_tasklet *tsq;
933 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
936 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
937 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
941 /* queue this socket to tasklet queue */
942 local_irq_save(flags);
943 tsq = this_cpu_ptr(&tsq_tasklet);
944 empty = list_empty(&tsq->head);
945 list_add(&tp->tsq_node, &tsq->head);
947 tasklet_schedule(&tsq->tasklet);
948 local_irq_restore(flags);
955 /* Note: Called under soft irq.
956 * We can call TCP stack right away, unless socket is owned by user.
958 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
960 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
961 struct sock *sk = (struct sock *)tp;
966 return HRTIMER_NORESTART;
969 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
974 if (!tcp_needs_internal_pacing(sk))
976 rate = sk->sk_pacing_rate;
977 if (!rate || rate == ~0U)
980 len_ns = (u64)skb->len * NSEC_PER_SEC;
981 do_div(len_ns, rate);
982 hrtimer_start(&tcp_sk(sk)->pacing_timer,
983 ktime_add_ns(ktime_get(), len_ns),
984 HRTIMER_MODE_ABS_PINNED_SOFT);
988 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
990 skb->skb_mstamp = tp->tcp_mstamp;
991 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
994 /* This routine actually transmits TCP packets queued in by
995 * tcp_do_sendmsg(). This is used by both the initial
996 * transmission and possible later retransmissions.
997 * All SKB's seen here are completely headerless. It is our
998 * job to build the TCP header, and pass the packet down to
999 * IP so it can do the same plus pass the packet off to the
1002 * We are working here with either a clone of the original
1003 * SKB, or a fresh unique copy made by the retransmit engine.
1005 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1006 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1008 const struct inet_connection_sock *icsk = inet_csk(sk);
1009 struct inet_sock *inet;
1010 struct tcp_sock *tp;
1011 struct tcp_skb_cb *tcb;
1012 struct tcp_out_options opts;
1013 unsigned int tcp_options_size, tcp_header_size;
1014 struct sk_buff *oskb = NULL;
1015 struct tcp_md5sig_key *md5;
1019 BUG_ON(!skb || !tcp_skb_pcount(skb));
1023 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1027 tcp_skb_tsorted_save(oskb) {
1028 if (unlikely(skb_cloned(oskb)))
1029 skb = pskb_copy(oskb, gfp_mask);
1031 skb = skb_clone(oskb, gfp_mask);
1032 } tcp_skb_tsorted_restore(oskb);
1037 skb->skb_mstamp = tp->tcp_mstamp;
1040 tcb = TCP_SKB_CB(skb);
1041 memset(&opts, 0, sizeof(opts));
1043 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1044 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1046 tcp_options_size = tcp_established_options(sk, skb, &opts,
1048 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1050 /* if no packet is in qdisc/device queue, then allow XPS to select
1051 * another queue. We can be called from tcp_tsq_handler()
1052 * which holds one reference to sk.
1054 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1055 * One way to get this would be to set skb->truesize = 2 on them.
1057 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1059 /* If we had to use memory reserve to allocate this skb,
1060 * this might cause drops if packet is looped back :
1061 * Other socket might not have SOCK_MEMALLOC.
1062 * Packets not looped back do not care about pfmemalloc.
1064 skb->pfmemalloc = 0;
1066 skb_push(skb, tcp_header_size);
1067 skb_reset_transport_header(skb);
1071 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1072 skb_set_hash_from_sk(skb, sk);
1073 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1075 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1077 /* Build TCP header and checksum it. */
1078 th = (struct tcphdr *)skb->data;
1079 th->source = inet->inet_sport;
1080 th->dest = inet->inet_dport;
1081 th->seq = htonl(tcb->seq);
1082 th->ack_seq = htonl(rcv_nxt);
1083 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1089 /* The urg_mode check is necessary during a below snd_una win probe */
1090 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1091 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1092 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1094 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1095 th->urg_ptr = htons(0xFFFF);
1100 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1101 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1102 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1103 th->window = htons(tcp_select_window(sk));
1104 tcp_ecn_send(sk, skb, th, tcp_header_size);
1106 /* RFC1323: The window in SYN & SYN/ACK segments
1109 th->window = htons(min(tp->rcv_wnd, 65535U));
1111 #ifdef CONFIG_TCP_MD5SIG
1112 /* Calculate the MD5 hash, as we have all we need now */
1114 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1115 tp->af_specific->calc_md5_hash(opts.hash_location,
1120 icsk->icsk_af_ops->send_check(sk, skb);
1122 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1123 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1125 if (skb->len != tcp_header_size) {
1126 tcp_event_data_sent(tp, sk);
1127 tp->data_segs_out += tcp_skb_pcount(skb);
1128 tp->bytes_sent += skb->len - tcp_header_size;
1129 tcp_internal_pacing(sk, skb);
1132 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1133 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1134 tcp_skb_pcount(skb));
1136 tp->segs_out += tcp_skb_pcount(skb);
1137 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1138 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1139 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1141 /* Our usage of tstamp should remain private */
1144 /* Cleanup our debris for IP stacks */
1145 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1146 sizeof(struct inet6_skb_parm)));
1148 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1150 if (unlikely(err > 0)) {
1152 err = net_xmit_eval(err);
1155 tcp_update_skb_after_send(tp, oskb);
1156 tcp_rate_skb_sent(sk, oskb);
1161 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1164 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1165 tcp_sk(sk)->rcv_nxt);
1168 /* This routine just queues the buffer for sending.
1170 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1171 * otherwise socket can stall.
1173 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1175 struct tcp_sock *tp = tcp_sk(sk);
1177 /* Advance write_seq and place onto the write_queue. */
1178 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1179 __skb_header_release(skb);
1180 tcp_add_write_queue_tail(sk, skb);
1181 sk->sk_wmem_queued += skb->truesize;
1182 sk_mem_charge(sk, skb->truesize);
1185 /* Initialize TSO segments for a packet. */
1186 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1188 if (skb->len <= mss_now) {
1189 /* Avoid the costly divide in the normal
1192 tcp_skb_pcount_set(skb, 1);
1193 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1195 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1196 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1200 /* Pcount in the middle of the write queue got changed, we need to do various
1201 * tweaks to fix counters
1203 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1205 struct tcp_sock *tp = tcp_sk(sk);
1207 tp->packets_out -= decr;
1209 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1210 tp->sacked_out -= decr;
1211 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1212 tp->retrans_out -= decr;
1213 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1214 tp->lost_out -= decr;
1216 /* Reno case is special. Sigh... */
1217 if (tcp_is_reno(tp) && decr > 0)
1218 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1220 if (tp->lost_skb_hint &&
1221 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1222 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1223 tp->lost_cnt_hint -= decr;
1225 tcp_verify_left_out(tp);
1228 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1230 return TCP_SKB_CB(skb)->txstamp_ack ||
1231 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1234 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1236 struct skb_shared_info *shinfo = skb_shinfo(skb);
1238 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1239 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1240 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1241 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1243 shinfo->tx_flags &= ~tsflags;
1244 shinfo2->tx_flags |= tsflags;
1245 swap(shinfo->tskey, shinfo2->tskey);
1246 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1247 TCP_SKB_CB(skb)->txstamp_ack = 0;
1251 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1253 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1254 TCP_SKB_CB(skb)->eor = 0;
1257 /* Insert buff after skb on the write or rtx queue of sk. */
1258 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1259 struct sk_buff *buff,
1261 enum tcp_queue tcp_queue)
1263 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1264 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1266 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1269 /* Function to create two new TCP segments. Shrinks the given segment
1270 * to the specified size and appends a new segment with the rest of the
1271 * packet to the list. This won't be called frequently, I hope.
1272 * Remember, these are still headerless SKBs at this point.
1274 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1275 struct sk_buff *skb, u32 len,
1276 unsigned int mss_now, gfp_t gfp)
1278 struct tcp_sock *tp = tcp_sk(sk);
1279 struct sk_buff *buff;
1280 int nsize, old_factor;
1285 if (WARN_ON(len > skb->len))
1288 nsize = skb_headlen(skb) - len;
1292 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1293 * We need some allowance to not penalize applications setting small
1295 * Also allow first and last skb in retransmit queue to be split.
1297 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1298 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1299 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1300 skb != tcp_rtx_queue_head(sk) &&
1301 skb != tcp_rtx_queue_tail(sk))) {
1302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1306 if (skb_unclone(skb, gfp))
1309 /* Get a new skb... force flag on. */
1310 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1312 return -ENOMEM; /* We'll just try again later. */
1314 sk->sk_wmem_queued += buff->truesize;
1315 sk_mem_charge(sk, buff->truesize);
1316 nlen = skb->len - len - nsize;
1317 buff->truesize += nlen;
1318 skb->truesize -= nlen;
1320 /* Correct the sequence numbers. */
1321 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1322 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1323 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1325 /* PSH and FIN should only be set in the second packet. */
1326 flags = TCP_SKB_CB(skb)->tcp_flags;
1327 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1328 TCP_SKB_CB(buff)->tcp_flags = flags;
1329 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1330 tcp_skb_fragment_eor(skb, buff);
1332 skb_split(skb, buff, len);
1334 buff->ip_summed = CHECKSUM_PARTIAL;
1336 buff->tstamp = skb->tstamp;
1337 tcp_fragment_tstamp(skb, buff);
1339 old_factor = tcp_skb_pcount(skb);
1341 /* Fix up tso_factor for both original and new SKB. */
1342 tcp_set_skb_tso_segs(skb, mss_now);
1343 tcp_set_skb_tso_segs(buff, mss_now);
1345 /* Update delivered info for the new segment */
1346 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1348 /* If this packet has been sent out already, we must
1349 * adjust the various packet counters.
1351 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1352 int diff = old_factor - tcp_skb_pcount(skb) -
1353 tcp_skb_pcount(buff);
1356 tcp_adjust_pcount(sk, skb, diff);
1359 /* Link BUFF into the send queue. */
1360 __skb_header_release(buff);
1361 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1362 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1363 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1368 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1369 * data is not copied, but immediately discarded.
1371 static int __pskb_trim_head(struct sk_buff *skb, int len)
1373 struct skb_shared_info *shinfo;
1376 eat = min_t(int, len, skb_headlen(skb));
1378 __skb_pull(skb, eat);
1385 shinfo = skb_shinfo(skb);
1386 for (i = 0; i < shinfo->nr_frags; i++) {
1387 int size = skb_frag_size(&shinfo->frags[i]);
1390 skb_frag_unref(skb, i);
1393 shinfo->frags[k] = shinfo->frags[i];
1395 shinfo->frags[k].page_offset += eat;
1396 skb_frag_size_sub(&shinfo->frags[k], eat);
1402 shinfo->nr_frags = k;
1404 skb->data_len -= len;
1405 skb->len = skb->data_len;
1409 /* Remove acked data from a packet in the transmit queue. */
1410 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1414 if (skb_unclone(skb, GFP_ATOMIC))
1417 delta_truesize = __pskb_trim_head(skb, len);
1419 TCP_SKB_CB(skb)->seq += len;
1420 skb->ip_summed = CHECKSUM_PARTIAL;
1422 if (delta_truesize) {
1423 skb->truesize -= delta_truesize;
1424 sk->sk_wmem_queued -= delta_truesize;
1425 sk_mem_uncharge(sk, delta_truesize);
1426 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1429 /* Any change of skb->len requires recalculation of tso factor. */
1430 if (tcp_skb_pcount(skb) > 1)
1431 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1436 /* Calculate MSS not accounting any TCP options. */
1437 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1439 const struct tcp_sock *tp = tcp_sk(sk);
1440 const struct inet_connection_sock *icsk = inet_csk(sk);
1443 /* Calculate base mss without TCP options:
1444 It is MMS_S - sizeof(tcphdr) of rfc1122
1446 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1448 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1449 if (icsk->icsk_af_ops->net_frag_header_len) {
1450 const struct dst_entry *dst = __sk_dst_get(sk);
1452 if (dst && dst_allfrag(dst))
1453 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1456 /* Clamp it (mss_clamp does not include tcp options) */
1457 if (mss_now > tp->rx_opt.mss_clamp)
1458 mss_now = tp->rx_opt.mss_clamp;
1460 /* Now subtract optional transport overhead */
1461 mss_now -= icsk->icsk_ext_hdr_len;
1463 /* Then reserve room for full set of TCP options and 8 bytes of data */
1464 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1468 /* Calculate MSS. Not accounting for SACKs here. */
1469 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1471 /* Subtract TCP options size, not including SACKs */
1472 return __tcp_mtu_to_mss(sk, pmtu) -
1473 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1475 EXPORT_SYMBOL(tcp_mtu_to_mss);
1477 /* Inverse of above */
1478 int tcp_mss_to_mtu(struct sock *sk, int mss)
1480 const struct tcp_sock *tp = tcp_sk(sk);
1481 const struct inet_connection_sock *icsk = inet_csk(sk);
1485 tp->tcp_header_len +
1486 icsk->icsk_ext_hdr_len +
1487 icsk->icsk_af_ops->net_header_len;
1489 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1490 if (icsk->icsk_af_ops->net_frag_header_len) {
1491 const struct dst_entry *dst = __sk_dst_get(sk);
1493 if (dst && dst_allfrag(dst))
1494 mtu += icsk->icsk_af_ops->net_frag_header_len;
1498 EXPORT_SYMBOL(tcp_mss_to_mtu);
1500 /* MTU probing init per socket */
1501 void tcp_mtup_init(struct sock *sk)
1503 struct tcp_sock *tp = tcp_sk(sk);
1504 struct inet_connection_sock *icsk = inet_csk(sk);
1505 struct net *net = sock_net(sk);
1507 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1508 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1509 icsk->icsk_af_ops->net_header_len;
1510 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1511 icsk->icsk_mtup.probe_size = 0;
1512 if (icsk->icsk_mtup.enabled)
1513 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1515 EXPORT_SYMBOL(tcp_mtup_init);
1517 /* This function synchronize snd mss to current pmtu/exthdr set.
1519 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1520 for TCP options, but includes only bare TCP header.
1522 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1523 It is minimum of user_mss and mss received with SYN.
1524 It also does not include TCP options.
1526 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1528 tp->mss_cache is current effective sending mss, including
1529 all tcp options except for SACKs. It is evaluated,
1530 taking into account current pmtu, but never exceeds
1531 tp->rx_opt.mss_clamp.
1533 NOTE1. rfc1122 clearly states that advertised MSS
1534 DOES NOT include either tcp or ip options.
1536 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1537 are READ ONLY outside this function. --ANK (980731)
1539 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1541 struct tcp_sock *tp = tcp_sk(sk);
1542 struct inet_connection_sock *icsk = inet_csk(sk);
1545 if (icsk->icsk_mtup.search_high > pmtu)
1546 icsk->icsk_mtup.search_high = pmtu;
1548 mss_now = tcp_mtu_to_mss(sk, pmtu);
1549 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1551 /* And store cached results */
1552 icsk->icsk_pmtu_cookie = pmtu;
1553 if (icsk->icsk_mtup.enabled)
1554 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1555 tp->mss_cache = mss_now;
1559 EXPORT_SYMBOL(tcp_sync_mss);
1561 /* Compute the current effective MSS, taking SACKs and IP options,
1562 * and even PMTU discovery events into account.
1564 unsigned int tcp_current_mss(struct sock *sk)
1566 const struct tcp_sock *tp = tcp_sk(sk);
1567 const struct dst_entry *dst = __sk_dst_get(sk);
1569 unsigned int header_len;
1570 struct tcp_out_options opts;
1571 struct tcp_md5sig_key *md5;
1573 mss_now = tp->mss_cache;
1576 u32 mtu = dst_mtu(dst);
1577 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1578 mss_now = tcp_sync_mss(sk, mtu);
1581 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1582 sizeof(struct tcphdr);
1583 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1584 * some common options. If this is an odd packet (because we have SACK
1585 * blocks etc) then our calculated header_len will be different, and
1586 * we have to adjust mss_now correspondingly */
1587 if (header_len != tp->tcp_header_len) {
1588 int delta = (int) header_len - tp->tcp_header_len;
1595 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1596 * As additional protections, we do not touch cwnd in retransmission phases,
1597 * and if application hit its sndbuf limit recently.
1599 static void tcp_cwnd_application_limited(struct sock *sk)
1601 struct tcp_sock *tp = tcp_sk(sk);
1603 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1604 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1605 /* Limited by application or receiver window. */
1606 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1607 u32 win_used = max(tp->snd_cwnd_used, init_win);
1608 if (win_used < tp->snd_cwnd) {
1609 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1610 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1612 tp->snd_cwnd_used = 0;
1614 tp->snd_cwnd_stamp = tcp_jiffies32;
1617 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1619 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1620 struct tcp_sock *tp = tcp_sk(sk);
1622 /* Track the maximum number of outstanding packets in each
1623 * window, and remember whether we were cwnd-limited then.
1625 if (!before(tp->snd_una, tp->max_packets_seq) ||
1626 tp->packets_out > tp->max_packets_out ||
1628 tp->max_packets_out = tp->packets_out;
1629 tp->max_packets_seq = tp->snd_nxt;
1630 tp->is_cwnd_limited = is_cwnd_limited;
1633 if (tcp_is_cwnd_limited(sk)) {
1634 /* Network is feed fully. */
1635 tp->snd_cwnd_used = 0;
1636 tp->snd_cwnd_stamp = tcp_jiffies32;
1638 /* Network starves. */
1639 if (tp->packets_out > tp->snd_cwnd_used)
1640 tp->snd_cwnd_used = tp->packets_out;
1642 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1643 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1644 !ca_ops->cong_control)
1645 tcp_cwnd_application_limited(sk);
1647 /* The following conditions together indicate the starvation
1648 * is caused by insufficient sender buffer:
1649 * 1) just sent some data (see tcp_write_xmit)
1650 * 2) not cwnd limited (this else condition)
1651 * 3) no more data to send (tcp_write_queue_empty())
1652 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1654 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1655 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1656 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1657 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1661 /* Minshall's variant of the Nagle send check. */
1662 static bool tcp_minshall_check(const struct tcp_sock *tp)
1664 return after(tp->snd_sml, tp->snd_una) &&
1665 !after(tp->snd_sml, tp->snd_nxt);
1668 /* Update snd_sml if this skb is under mss
1669 * Note that a TSO packet might end with a sub-mss segment
1670 * The test is really :
1671 * if ((skb->len % mss) != 0)
1672 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1673 * But we can avoid doing the divide again given we already have
1674 * skb_pcount = skb->len / mss_now
1676 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1677 const struct sk_buff *skb)
1679 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1680 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1683 /* Return false, if packet can be sent now without violation Nagle's rules:
1684 * 1. It is full sized. (provided by caller in %partial bool)
1685 * 2. Or it contains FIN. (already checked by caller)
1686 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1687 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1688 * With Minshall's modification: all sent small packets are ACKed.
1690 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1694 ((nonagle & TCP_NAGLE_CORK) ||
1695 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1698 /* Return how many segs we'd like on a TSO packet,
1699 * to send one TSO packet per ms
1701 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1706 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1707 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1709 /* Goal is to send at least one packet per ms,
1710 * not one big TSO packet every 100 ms.
1711 * This preserves ACK clocking and is consistent
1712 * with tcp_tso_should_defer() heuristic.
1714 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1719 /* Return the number of segments we want in the skb we are transmitting.
1720 * See if congestion control module wants to decide; otherwise, autosize.
1722 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1724 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1725 u32 min_tso, tso_segs;
1727 min_tso = ca_ops->min_tso_segs ?
1728 ca_ops->min_tso_segs(sk) :
1729 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1731 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1732 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1735 /* Returns the portion of skb which can be sent right away */
1736 static unsigned int tcp_mss_split_point(const struct sock *sk,
1737 const struct sk_buff *skb,
1738 unsigned int mss_now,
1739 unsigned int max_segs,
1742 const struct tcp_sock *tp = tcp_sk(sk);
1743 u32 partial, needed, window, max_len;
1745 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1746 max_len = mss_now * max_segs;
1748 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1751 needed = min(skb->len, window);
1753 if (max_len <= needed)
1756 partial = needed % mss_now;
1757 /* If last segment is not a full MSS, check if Nagle rules allow us
1758 * to include this last segment in this skb.
1759 * Otherwise, we'll split the skb at last MSS boundary
1761 if (tcp_nagle_check(partial != 0, tp, nonagle))
1762 return needed - partial;
1767 /* Can at least one segment of SKB be sent right now, according to the
1768 * congestion window rules? If so, return how many segments are allowed.
1770 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1771 const struct sk_buff *skb)
1773 u32 in_flight, cwnd, halfcwnd;
1775 /* Don't be strict about the congestion window for the final FIN. */
1776 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1777 tcp_skb_pcount(skb) == 1)
1780 in_flight = tcp_packets_in_flight(tp);
1781 cwnd = tp->snd_cwnd;
1782 if (in_flight >= cwnd)
1785 /* For better scheduling, ensure we have at least
1786 * 2 GSO packets in flight.
1788 halfcwnd = max(cwnd >> 1, 1U);
1789 return min(halfcwnd, cwnd - in_flight);
1792 /* Initialize TSO state of a skb.
1793 * This must be invoked the first time we consider transmitting
1794 * SKB onto the wire.
1796 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1798 int tso_segs = tcp_skb_pcount(skb);
1800 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1801 tcp_set_skb_tso_segs(skb, mss_now);
1802 tso_segs = tcp_skb_pcount(skb);
1808 /* Return true if the Nagle test allows this packet to be
1811 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1812 unsigned int cur_mss, int nonagle)
1814 /* Nagle rule does not apply to frames, which sit in the middle of the
1815 * write_queue (they have no chances to get new data).
1817 * This is implemented in the callers, where they modify the 'nonagle'
1818 * argument based upon the location of SKB in the send queue.
1820 if (nonagle & TCP_NAGLE_PUSH)
1823 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1824 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1827 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1833 /* Does at least the first segment of SKB fit into the send window? */
1834 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1835 const struct sk_buff *skb,
1836 unsigned int cur_mss)
1838 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1840 if (skb->len > cur_mss)
1841 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1843 return !after(end_seq, tcp_wnd_end(tp));
1846 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1847 * which is put after SKB on the list. It is very much like
1848 * tcp_fragment() except that it may make several kinds of assumptions
1849 * in order to speed up the splitting operation. In particular, we
1850 * know that all the data is in scatter-gather pages, and that the
1851 * packet has never been sent out before (and thus is not cloned).
1853 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1854 struct sk_buff *skb, unsigned int len,
1855 unsigned int mss_now, gfp_t gfp)
1857 struct sk_buff *buff;
1858 int nlen = skb->len - len;
1861 /* All of a TSO frame must be composed of paged data. */
1862 if (skb->len != skb->data_len)
1863 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1865 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1866 if (unlikely(!buff))
1869 sk->sk_wmem_queued += buff->truesize;
1870 sk_mem_charge(sk, buff->truesize);
1871 buff->truesize += nlen;
1872 skb->truesize -= nlen;
1874 /* Correct the sequence numbers. */
1875 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1876 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1877 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1879 /* PSH and FIN should only be set in the second packet. */
1880 flags = TCP_SKB_CB(skb)->tcp_flags;
1881 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1882 TCP_SKB_CB(buff)->tcp_flags = flags;
1884 /* This packet was never sent out yet, so no SACK bits. */
1885 TCP_SKB_CB(buff)->sacked = 0;
1887 tcp_skb_fragment_eor(skb, buff);
1889 buff->ip_summed = CHECKSUM_PARTIAL;
1890 skb_split(skb, buff, len);
1891 tcp_fragment_tstamp(skb, buff);
1893 /* Fix up tso_factor for both original and new SKB. */
1894 tcp_set_skb_tso_segs(skb, mss_now);
1895 tcp_set_skb_tso_segs(buff, mss_now);
1897 /* Link BUFF into the send queue. */
1898 __skb_header_release(buff);
1899 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1904 /* Try to defer sending, if possible, in order to minimize the amount
1905 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1907 * This algorithm is from John Heffner.
1909 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1910 bool *is_cwnd_limited,
1911 bool *is_rwnd_limited,
1914 const struct inet_connection_sock *icsk = inet_csk(sk);
1915 u32 age, send_win, cong_win, limit, in_flight;
1916 struct tcp_sock *tp = tcp_sk(sk);
1917 struct sk_buff *head;
1920 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1923 /* Avoid bursty behavior by allowing defer
1924 * only if the last write was recent.
1926 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1929 in_flight = tcp_packets_in_flight(tp);
1931 BUG_ON(tcp_skb_pcount(skb) <= 1);
1932 BUG_ON(tp->snd_cwnd <= in_flight);
1934 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1936 /* From in_flight test above, we know that cwnd > in_flight. */
1937 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1939 limit = min(send_win, cong_win);
1941 /* If a full-sized TSO skb can be sent, do it. */
1942 if (limit >= max_segs * tp->mss_cache)
1945 /* Middle in queue won't get any more data, full sendable already? */
1946 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1949 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1951 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1953 /* If at least some fraction of a window is available,
1956 chunk /= win_divisor;
1960 /* Different approach, try not to defer past a single
1961 * ACK. Receiver should ACK every other full sized
1962 * frame, so if we have space for more than 3 frames
1965 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1969 /* TODO : use tsorted_sent_queue ? */
1970 head = tcp_rtx_queue_head(sk);
1973 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1974 /* If next ACK is likely to come too late (half srtt), do not defer */
1975 if (age < (tp->srtt_us >> 4))
1978 /* Ok, it looks like it is advisable to defer.
1979 * Three cases are tracked :
1980 * 1) We are cwnd-limited
1981 * 2) We are rwnd-limited
1982 * 3) We are application limited.
1984 if (cong_win < send_win) {
1985 if (cong_win <= skb->len) {
1986 *is_cwnd_limited = true;
1990 if (send_win <= skb->len) {
1991 *is_rwnd_limited = true;
1996 /* If this packet won't get more data, do not wait. */
1997 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2006 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2008 struct inet_connection_sock *icsk = inet_csk(sk);
2009 struct tcp_sock *tp = tcp_sk(sk);
2010 struct net *net = sock_net(sk);
2014 interval = net->ipv4.sysctl_tcp_probe_interval;
2015 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2016 if (unlikely(delta >= interval * HZ)) {
2017 int mss = tcp_current_mss(sk);
2019 /* Update current search range */
2020 icsk->icsk_mtup.probe_size = 0;
2021 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2022 sizeof(struct tcphdr) +
2023 icsk->icsk_af_ops->net_header_len;
2024 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2026 /* Update probe time stamp */
2027 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2031 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2033 struct sk_buff *skb, *next;
2035 skb = tcp_send_head(sk);
2036 tcp_for_write_queue_from_safe(skb, next, sk) {
2037 if (len <= skb->len)
2040 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2049 /* Create a new MTU probe if we are ready.
2050 * MTU probe is regularly attempting to increase the path MTU by
2051 * deliberately sending larger packets. This discovers routing
2052 * changes resulting in larger path MTUs.
2054 * Returns 0 if we should wait to probe (no cwnd available),
2055 * 1 if a probe was sent,
2058 static int tcp_mtu_probe(struct sock *sk)
2060 struct inet_connection_sock *icsk = inet_csk(sk);
2061 struct tcp_sock *tp = tcp_sk(sk);
2062 struct sk_buff *skb, *nskb, *next;
2063 struct net *net = sock_net(sk);
2070 /* Not currently probing/verifying,
2072 * have enough cwnd, and
2073 * not SACKing (the variable headers throw things off)
2075 if (likely(!icsk->icsk_mtup.enabled ||
2076 icsk->icsk_mtup.probe_size ||
2077 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2078 tp->snd_cwnd < 11 ||
2079 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2082 /* Use binary search for probe_size between tcp_mss_base,
2083 * and current mss_clamp. if (search_high - search_low)
2084 * smaller than a threshold, backoff from probing.
2086 mss_now = tcp_current_mss(sk);
2087 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2088 icsk->icsk_mtup.search_low) >> 1);
2089 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2090 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2091 /* When misfortune happens, we are reprobing actively,
2092 * and then reprobe timer has expired. We stick with current
2093 * probing process by not resetting search range to its orignal.
2095 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2096 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2097 /* Check whether enough time has elaplased for
2098 * another round of probing.
2100 tcp_mtu_check_reprobe(sk);
2104 /* Have enough data in the send queue to probe? */
2105 if (tp->write_seq - tp->snd_nxt < size_needed)
2108 if (tp->snd_wnd < size_needed)
2110 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2113 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2114 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2115 if (!tcp_packets_in_flight(tp))
2121 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2124 /* We're allowed to probe. Build it now. */
2125 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2128 sk->sk_wmem_queued += nskb->truesize;
2129 sk_mem_charge(sk, nskb->truesize);
2131 skb = tcp_send_head(sk);
2133 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2134 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2135 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2136 TCP_SKB_CB(nskb)->sacked = 0;
2138 nskb->ip_summed = CHECKSUM_PARTIAL;
2140 tcp_insert_write_queue_before(nskb, skb, sk);
2141 tcp_highest_sack_replace(sk, skb, nskb);
2144 tcp_for_write_queue_from_safe(skb, next, sk) {
2145 copy = min_t(int, skb->len, probe_size - len);
2146 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2148 if (skb->len <= copy) {
2149 /* We've eaten all the data from this skb.
2151 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2152 /* If this is the last SKB we copy and eor is set
2153 * we need to propagate it to the new skb.
2155 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2156 tcp_skb_collapse_tstamp(nskb, skb);
2157 tcp_unlink_write_queue(skb, sk);
2158 sk_wmem_free_skb(sk, skb);
2160 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2161 ~(TCPHDR_FIN|TCPHDR_PSH);
2162 if (!skb_shinfo(skb)->nr_frags) {
2163 skb_pull(skb, copy);
2165 __pskb_trim_head(skb, copy);
2166 tcp_set_skb_tso_segs(skb, mss_now);
2168 TCP_SKB_CB(skb)->seq += copy;
2173 if (len >= probe_size)
2176 tcp_init_tso_segs(nskb, nskb->len);
2178 /* We're ready to send. If this fails, the probe will
2179 * be resegmented into mss-sized pieces by tcp_write_xmit().
2181 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2182 /* Decrement cwnd here because we are sending
2183 * effectively two packets. */
2185 tcp_event_new_data_sent(sk, nskb);
2187 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2188 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2189 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2197 static bool tcp_pacing_check(const struct sock *sk)
2199 return tcp_needs_internal_pacing(sk) &&
2200 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2203 /* TCP Small Queues :
2204 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2205 * (These limits are doubled for retransmits)
2207 * - better RTT estimation and ACK scheduling
2210 * Alas, some drivers / subsystems require a fair amount
2211 * of queued bytes to ensure line rate.
2212 * One example is wifi aggregation (802.11 AMPDU)
2214 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2215 unsigned int factor)
2219 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2220 limit = min_t(u32, limit,
2221 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2224 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2225 /* Always send skb if rtx queue is empty.
2226 * No need to wait for TX completion to call us back,
2227 * after softirq/tasklet schedule.
2228 * This helps when TX completions are delayed too much.
2230 if (tcp_rtx_queue_empty(sk))
2233 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2234 /* It is possible TX completion already happened
2235 * before we set TSQ_THROTTLED, so we must
2236 * test again the condition.
2238 smp_mb__after_atomic();
2239 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2245 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2247 const u32 now = tcp_jiffies32;
2248 enum tcp_chrono old = tp->chrono_type;
2250 if (old > TCP_CHRONO_UNSPEC)
2251 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2252 tp->chrono_start = now;
2253 tp->chrono_type = new;
2256 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2258 struct tcp_sock *tp = tcp_sk(sk);
2260 /* If there are multiple conditions worthy of tracking in a
2261 * chronograph then the highest priority enum takes precedence
2262 * over the other conditions. So that if something "more interesting"
2263 * starts happening, stop the previous chrono and start a new one.
2265 if (type > tp->chrono_type)
2266 tcp_chrono_set(tp, type);
2269 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2271 struct tcp_sock *tp = tcp_sk(sk);
2274 /* There are multiple conditions worthy of tracking in a
2275 * chronograph, so that the highest priority enum takes
2276 * precedence over the other conditions (see tcp_chrono_start).
2277 * If a condition stops, we only stop chrono tracking if
2278 * it's the "most interesting" or current chrono we are
2279 * tracking and starts busy chrono if we have pending data.
2281 if (tcp_rtx_and_write_queues_empty(sk))
2282 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2283 else if (type == tp->chrono_type)
2284 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2287 /* This routine writes packets to the network. It advances the
2288 * send_head. This happens as incoming acks open up the remote
2291 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2292 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2293 * account rare use of URG, this is not a big flaw.
2295 * Send at most one packet when push_one > 0. Temporarily ignore
2296 * cwnd limit to force at most one packet out when push_one == 2.
2298 * Returns true, if no segments are in flight and we have queued segments,
2299 * but cannot send anything now because of SWS or another problem.
2301 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2302 int push_one, gfp_t gfp)
2304 struct tcp_sock *tp = tcp_sk(sk);
2305 struct sk_buff *skb;
2306 unsigned int tso_segs, sent_pkts;
2309 bool is_cwnd_limited = false, is_rwnd_limited = false;
2314 tcp_mstamp_refresh(tp);
2316 /* Do MTU probing. */
2317 result = tcp_mtu_probe(sk);
2320 } else if (result > 0) {
2325 max_segs = tcp_tso_segs(sk, mss_now);
2326 while ((skb = tcp_send_head(sk))) {
2329 if (tcp_pacing_check(sk))
2332 tso_segs = tcp_init_tso_segs(skb, mss_now);
2335 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2336 /* "skb_mstamp" is used as a start point for the retransmit timer */
2337 tcp_update_skb_after_send(tp, skb);
2338 goto repair; /* Skip network transmission */
2341 cwnd_quota = tcp_cwnd_test(tp, skb);
2344 /* Force out a loss probe pkt. */
2350 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2351 is_rwnd_limited = true;
2355 if (tso_segs == 1) {
2356 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2357 (tcp_skb_is_last(sk, skb) ?
2358 nonagle : TCP_NAGLE_PUSH))))
2362 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2363 &is_rwnd_limited, max_segs))
2368 if (tso_segs > 1 && !tcp_urg_mode(tp))
2369 limit = tcp_mss_split_point(sk, skb, mss_now,
2375 if (skb->len > limit &&
2376 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2377 skb, limit, mss_now, gfp)))
2380 if (tcp_small_queue_check(sk, skb, 0))
2383 /* Argh, we hit an empty skb(), presumably a thread
2384 * is sleeping in sendmsg()/sk_stream_wait_memory().
2385 * We do not want to send a pure-ack packet and have
2386 * a strange looking rtx queue with empty packet(s).
2388 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2391 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2395 /* Advance the send_head. This one is sent out.
2396 * This call will increment packets_out.
2398 tcp_event_new_data_sent(sk, skb);
2400 tcp_minshall_update(tp, mss_now, skb);
2401 sent_pkts += tcp_skb_pcount(skb);
2407 if (is_rwnd_limited)
2408 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2410 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2412 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2413 if (likely(sent_pkts || is_cwnd_limited))
2414 tcp_cwnd_validate(sk, is_cwnd_limited);
2416 if (likely(sent_pkts)) {
2417 if (tcp_in_cwnd_reduction(sk))
2418 tp->prr_out += sent_pkts;
2420 /* Send one loss probe per tail loss episode. */
2422 tcp_schedule_loss_probe(sk, false);
2425 return !tp->packets_out && !tcp_write_queue_empty(sk);
2428 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2430 struct inet_connection_sock *icsk = inet_csk(sk);
2431 struct tcp_sock *tp = tcp_sk(sk);
2432 u32 timeout, rto_delta_us;
2435 /* Don't do any loss probe on a Fast Open connection before 3WHS
2438 if (tp->fastopen_rsk)
2441 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2442 /* Schedule a loss probe in 2*RTT for SACK capable connections
2443 * not in loss recovery, that are either limited by cwnd or application.
2445 if ((early_retrans != 3 && early_retrans != 4) ||
2446 !tp->packets_out || !tcp_is_sack(tp) ||
2447 (icsk->icsk_ca_state != TCP_CA_Open &&
2448 icsk->icsk_ca_state != TCP_CA_CWR))
2451 /* Probe timeout is 2*rtt. Add minimum RTO to account
2452 * for delayed ack when there's one outstanding packet. If no RTT
2453 * sample is available then probe after TCP_TIMEOUT_INIT.
2456 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2457 if (tp->packets_out == 1)
2458 timeout += TCP_RTO_MIN;
2460 timeout += TCP_TIMEOUT_MIN;
2462 timeout = TCP_TIMEOUT_INIT;
2465 /* If the RTO formula yields an earlier time, then use that time. */
2466 rto_delta_us = advancing_rto ?
2467 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2468 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2469 if (rto_delta_us > 0)
2470 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2472 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2477 /* Thanks to skb fast clones, we can detect if a prior transmit of
2478 * a packet is still in a qdisc or driver queue.
2479 * In this case, there is very little point doing a retransmit !
2481 static bool skb_still_in_host_queue(const struct sock *sk,
2482 const struct sk_buff *skb)
2484 if (unlikely(skb_fclone_busy(sk, skb))) {
2485 NET_INC_STATS(sock_net(sk),
2486 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2492 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2493 * retransmit the last segment.
2495 void tcp_send_loss_probe(struct sock *sk)
2497 struct tcp_sock *tp = tcp_sk(sk);
2498 struct sk_buff *skb;
2500 int mss = tcp_current_mss(sk);
2502 /* At most one outstanding TLP */
2503 if (tp->tlp_high_seq)
2506 tp->tlp_retrans = 0;
2507 skb = tcp_send_head(sk);
2508 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2509 pcount = tp->packets_out;
2510 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2511 if (tp->packets_out > pcount)
2515 skb = skb_rb_last(&sk->tcp_rtx_queue);
2516 if (unlikely(!skb)) {
2517 WARN_ONCE(tp->packets_out,
2518 "invalid inflight: %u state %u cwnd %u mss %d\n",
2519 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2520 inet_csk(sk)->icsk_pending = 0;
2524 if (skb_still_in_host_queue(sk, skb))
2527 pcount = tcp_skb_pcount(skb);
2528 if (WARN_ON(!pcount))
2531 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2532 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2533 (pcount - 1) * mss, mss,
2536 skb = skb_rb_next(skb);
2539 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2542 if (__tcp_retransmit_skb(sk, skb, 1))
2545 tp->tlp_retrans = 1;
2548 /* Record snd_nxt for loss detection. */
2549 tp->tlp_high_seq = tp->snd_nxt;
2551 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2552 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2553 inet_csk(sk)->icsk_pending = 0;
2558 /* Push out any pending frames which were held back due to
2559 * TCP_CORK or attempt at coalescing tiny packets.
2560 * The socket must be locked by the caller.
2562 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2565 /* If we are closed, the bytes will have to remain here.
2566 * In time closedown will finish, we empty the write queue and
2567 * all will be happy.
2569 if (unlikely(sk->sk_state == TCP_CLOSE))
2572 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2573 sk_gfp_mask(sk, GFP_ATOMIC)))
2574 tcp_check_probe_timer(sk);
2577 /* Send _single_ skb sitting at the send head. This function requires
2578 * true push pending frames to setup probe timer etc.
2580 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2582 struct sk_buff *skb = tcp_send_head(sk);
2584 BUG_ON(!skb || skb->len < mss_now);
2586 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2589 /* This function returns the amount that we can raise the
2590 * usable window based on the following constraints
2592 * 1. The window can never be shrunk once it is offered (RFC 793)
2593 * 2. We limit memory per socket
2596 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2597 * RECV.NEXT + RCV.WIN fixed until:
2598 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2600 * i.e. don't raise the right edge of the window until you can raise
2601 * it at least MSS bytes.
2603 * Unfortunately, the recommended algorithm breaks header prediction,
2604 * since header prediction assumes th->window stays fixed.
2606 * Strictly speaking, keeping th->window fixed violates the receiver
2607 * side SWS prevention criteria. The problem is that under this rule
2608 * a stream of single byte packets will cause the right side of the
2609 * window to always advance by a single byte.
2611 * Of course, if the sender implements sender side SWS prevention
2612 * then this will not be a problem.
2614 * BSD seems to make the following compromise:
2616 * If the free space is less than the 1/4 of the maximum
2617 * space available and the free space is less than 1/2 mss,
2618 * then set the window to 0.
2619 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2620 * Otherwise, just prevent the window from shrinking
2621 * and from being larger than the largest representable value.
2623 * This prevents incremental opening of the window in the regime
2624 * where TCP is limited by the speed of the reader side taking
2625 * data out of the TCP receive queue. It does nothing about
2626 * those cases where the window is constrained on the sender side
2627 * because the pipeline is full.
2629 * BSD also seems to "accidentally" limit itself to windows that are a
2630 * multiple of MSS, at least until the free space gets quite small.
2631 * This would appear to be a side effect of the mbuf implementation.
2632 * Combining these two algorithms results in the observed behavior
2633 * of having a fixed window size at almost all times.
2635 * Below we obtain similar behavior by forcing the offered window to
2636 * a multiple of the mss when it is feasible to do so.
2638 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2639 * Regular options like TIMESTAMP are taken into account.
2641 u32 __tcp_select_window(struct sock *sk)
2643 struct inet_connection_sock *icsk = inet_csk(sk);
2644 struct tcp_sock *tp = tcp_sk(sk);
2645 /* MSS for the peer's data. Previous versions used mss_clamp
2646 * here. I don't know if the value based on our guesses
2647 * of peer's MSS is better for the performance. It's more correct
2648 * but may be worse for the performance because of rcv_mss
2649 * fluctuations. --SAW 1998/11/1
2651 int mss = icsk->icsk_ack.rcv_mss;
2652 int free_space = tcp_space(sk);
2653 int allowed_space = tcp_full_space(sk);
2654 int full_space = min_t(int, tp->window_clamp, allowed_space);
2657 if (unlikely(mss > full_space)) {
2662 if (free_space < (full_space >> 1)) {
2663 icsk->icsk_ack.quick = 0;
2665 if (tcp_under_memory_pressure(sk))
2666 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2669 /* free_space might become our new window, make sure we don't
2670 * increase it due to wscale.
2672 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2674 /* if free space is less than mss estimate, or is below 1/16th
2675 * of the maximum allowed, try to move to zero-window, else
2676 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2677 * new incoming data is dropped due to memory limits.
2678 * With large window, mss test triggers way too late in order
2679 * to announce zero window in time before rmem limit kicks in.
2681 if (free_space < (allowed_space >> 4) || free_space < mss)
2685 if (free_space > tp->rcv_ssthresh)
2686 free_space = tp->rcv_ssthresh;
2688 /* Don't do rounding if we are using window scaling, since the
2689 * scaled window will not line up with the MSS boundary anyway.
2691 if (tp->rx_opt.rcv_wscale) {
2692 window = free_space;
2694 /* Advertise enough space so that it won't get scaled away.
2695 * Import case: prevent zero window announcement if
2696 * 1<<rcv_wscale > mss.
2698 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2700 window = tp->rcv_wnd;
2701 /* Get the largest window that is a nice multiple of mss.
2702 * Window clamp already applied above.
2703 * If our current window offering is within 1 mss of the
2704 * free space we just keep it. This prevents the divide
2705 * and multiply from happening most of the time.
2706 * We also don't do any window rounding when the free space
2709 if (window <= free_space - mss || window > free_space)
2710 window = rounddown(free_space, mss);
2711 else if (mss == full_space &&
2712 free_space > window + (full_space >> 1))
2713 window = free_space;
2719 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2720 const struct sk_buff *next_skb)
2722 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2723 const struct skb_shared_info *next_shinfo =
2724 skb_shinfo(next_skb);
2725 struct skb_shared_info *shinfo = skb_shinfo(skb);
2727 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2728 shinfo->tskey = next_shinfo->tskey;
2729 TCP_SKB_CB(skb)->txstamp_ack |=
2730 TCP_SKB_CB(next_skb)->txstamp_ack;
2734 /* Collapses two adjacent SKB's during retransmission. */
2735 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2737 struct tcp_sock *tp = tcp_sk(sk);
2738 struct sk_buff *next_skb = skb_rb_next(skb);
2741 next_skb_size = next_skb->len;
2743 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2745 if (next_skb_size) {
2746 if (next_skb_size <= skb_availroom(skb))
2747 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2749 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2752 tcp_highest_sack_replace(sk, next_skb, skb);
2754 /* Update sequence range on original skb. */
2755 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2757 /* Merge over control information. This moves PSH/FIN etc. over */
2758 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2760 /* All done, get rid of second SKB and account for it so
2761 * packet counting does not break.
2763 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2764 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2766 /* changed transmit queue under us so clear hints */
2767 tcp_clear_retrans_hints_partial(tp);
2768 if (next_skb == tp->retransmit_skb_hint)
2769 tp->retransmit_skb_hint = skb;
2771 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2773 tcp_skb_collapse_tstamp(skb, next_skb);
2775 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2779 /* Check if coalescing SKBs is legal. */
2780 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2782 if (tcp_skb_pcount(skb) > 1)
2784 if (skb_cloned(skb))
2786 /* Some heuristics for collapsing over SACK'd could be invented */
2787 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2793 /* Collapse packets in the retransmit queue to make to create
2794 * less packets on the wire. This is only done on retransmission.
2796 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2799 struct tcp_sock *tp = tcp_sk(sk);
2800 struct sk_buff *skb = to, *tmp;
2803 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2805 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2808 skb_rbtree_walk_from_safe(skb, tmp) {
2809 if (!tcp_can_collapse(sk, skb))
2812 if (!tcp_skb_can_collapse_to(to))
2825 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2828 if (!tcp_collapse_retrans(sk, to))
2833 /* This retransmits one SKB. Policy decisions and retransmit queue
2834 * state updates are done by the caller. Returns non-zero if an
2835 * error occurred which prevented the send.
2837 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2839 struct inet_connection_sock *icsk = inet_csk(sk);
2840 struct tcp_sock *tp = tcp_sk(sk);
2841 unsigned int cur_mss;
2845 /* Inconclusive MTU probe */
2846 if (icsk->icsk_mtup.probe_size)
2847 icsk->icsk_mtup.probe_size = 0;
2849 /* Do not sent more than we queued. 1/4 is reserved for possible
2850 * copying overhead: fragmentation, tunneling, mangling etc.
2852 if (refcount_read(&sk->sk_wmem_alloc) >
2853 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2857 if (skb_still_in_host_queue(sk, skb))
2860 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2861 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2865 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2869 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2870 return -EHOSTUNREACH; /* Routing failure or similar. */
2872 cur_mss = tcp_current_mss(sk);
2874 /* If receiver has shrunk his window, and skb is out of
2875 * new window, do not retransmit it. The exception is the
2876 * case, when window is shrunk to zero. In this case
2877 * our retransmit serves as a zero window probe.
2879 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2880 TCP_SKB_CB(skb)->seq != tp->snd_una)
2883 len = cur_mss * segs;
2884 if (skb->len > len) {
2885 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2886 cur_mss, GFP_ATOMIC))
2887 return -ENOMEM; /* We'll try again later. */
2889 if (skb_unclone(skb, GFP_ATOMIC))
2892 diff = tcp_skb_pcount(skb);
2893 tcp_set_skb_tso_segs(skb, cur_mss);
2894 diff -= tcp_skb_pcount(skb);
2896 tcp_adjust_pcount(sk, skb, diff);
2897 if (skb->len < cur_mss)
2898 tcp_retrans_try_collapse(sk, skb, cur_mss);
2901 /* RFC3168, section 6.1.1.1. ECN fallback */
2902 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2903 tcp_ecn_clear_syn(sk, skb);
2905 /* Update global and local TCP statistics. */
2906 segs = tcp_skb_pcount(skb);
2907 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2908 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2909 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2910 tp->total_retrans += segs;
2911 tp->bytes_retrans += skb->len;
2913 /* make sure skb->data is aligned on arches that require it
2914 * and check if ack-trimming & collapsing extended the headroom
2915 * beyond what csum_start can cover.
2917 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2918 skb_headroom(skb) >= 0xFFFF)) {
2919 struct sk_buff *nskb;
2921 tcp_skb_tsorted_save(skb) {
2922 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2923 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2925 } tcp_skb_tsorted_restore(skb);
2928 tcp_update_skb_after_send(tp, skb);
2929 tcp_rate_skb_sent(sk, skb);
2932 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2935 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2936 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2937 TCP_SKB_CB(skb)->seq, segs, err);
2940 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2941 trace_tcp_retransmit_skb(sk, skb);
2942 } else if (err != -EBUSY) {
2943 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2948 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2950 struct tcp_sock *tp = tcp_sk(sk);
2951 int err = __tcp_retransmit_skb(sk, skb, segs);
2954 #if FASTRETRANS_DEBUG > 0
2955 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2956 net_dbg_ratelimited("retrans_out leaked\n");
2959 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2960 tp->retrans_out += tcp_skb_pcount(skb);
2962 /* Save stamp of the first retransmit. */
2963 if (!tp->retrans_stamp)
2964 tp->retrans_stamp = tcp_skb_timestamp(skb);
2968 if (tp->undo_retrans < 0)
2969 tp->undo_retrans = 0;
2970 tp->undo_retrans += tcp_skb_pcount(skb);
2974 /* This gets called after a retransmit timeout, and the initially
2975 * retransmitted data is acknowledged. It tries to continue
2976 * resending the rest of the retransmit queue, until either
2977 * we've sent it all or the congestion window limit is reached.
2979 void tcp_xmit_retransmit_queue(struct sock *sk)
2981 const struct inet_connection_sock *icsk = inet_csk(sk);
2982 struct sk_buff *skb, *rtx_head, *hole = NULL;
2983 struct tcp_sock *tp = tcp_sk(sk);
2987 if (!tp->packets_out)
2990 rtx_head = tcp_rtx_queue_head(sk);
2991 skb = tp->retransmit_skb_hint ?: rtx_head;
2992 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2993 skb_rbtree_walk_from(skb) {
2997 if (tcp_pacing_check(sk))
3000 /* we could do better than to assign each time */
3002 tp->retransmit_skb_hint = skb;
3004 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3007 sacked = TCP_SKB_CB(skb)->sacked;
3008 /* In case tcp_shift_skb_data() have aggregated large skbs,
3009 * we need to make sure not sending too bigs TSO packets
3011 segs = min_t(int, segs, max_segs);
3013 if (tp->retrans_out >= tp->lost_out) {
3015 } else if (!(sacked & TCPCB_LOST)) {
3016 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3021 if (icsk->icsk_ca_state != TCP_CA_Loss)
3022 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3024 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3027 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3030 if (tcp_small_queue_check(sk, skb, 1))
3033 if (tcp_retransmit_skb(sk, skb, segs))
3036 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3038 if (tcp_in_cwnd_reduction(sk))
3039 tp->prr_out += tcp_skb_pcount(skb);
3041 if (skb == rtx_head &&
3042 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3043 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3044 inet_csk(sk)->icsk_rto,
3049 /* We allow to exceed memory limits for FIN packets to expedite
3050 * connection tear down and (memory) recovery.
3051 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3052 * or even be forced to close flow without any FIN.
3053 * In general, we want to allow one skb per socket to avoid hangs
3054 * with edge trigger epoll()
3056 void sk_forced_mem_schedule(struct sock *sk, int size)
3060 if (size <= sk->sk_forward_alloc)
3062 amt = sk_mem_pages(size);
3063 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3064 sk_memory_allocated_add(sk, amt);
3066 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3067 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3070 /* Send a FIN. The caller locks the socket for us.
3071 * We should try to send a FIN packet really hard, but eventually give up.
3073 void tcp_send_fin(struct sock *sk)
3075 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3076 struct tcp_sock *tp = tcp_sk(sk);
3078 /* Optimization, tack on the FIN if we have one skb in write queue and
3079 * this skb was not yet sent, or we are under memory pressure.
3080 * Note: in the latter case, FIN packet will be sent after a timeout,
3081 * as TCP stack thinks it has already been transmitted.
3083 if (!tskb && tcp_under_memory_pressure(sk))
3084 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3088 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3089 TCP_SKB_CB(tskb)->end_seq++;
3091 if (tcp_write_queue_empty(sk)) {
3092 /* This means tskb was already sent.
3093 * Pretend we included the FIN on previous transmit.
3094 * We need to set tp->snd_nxt to the value it would have
3095 * if FIN had been sent. This is because retransmit path
3096 * does not change tp->snd_nxt.
3102 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3103 if (unlikely(!skb)) {
3108 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3109 skb_reserve(skb, MAX_TCP_HEADER);
3110 sk_forced_mem_schedule(sk, skb->truesize);
3111 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3112 tcp_init_nondata_skb(skb, tp->write_seq,
3113 TCPHDR_ACK | TCPHDR_FIN);
3114 tcp_queue_skb(sk, skb);
3116 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3119 /* We get here when a process closes a file descriptor (either due to
3120 * an explicit close() or as a byproduct of exit()'ing) and there
3121 * was unread data in the receive queue. This behavior is recommended
3122 * by RFC 2525, section 2.17. -DaveM
3124 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3126 struct sk_buff *skb;
3128 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3130 /* NOTE: No TCP options attached and we never retransmit this. */
3131 skb = alloc_skb(MAX_TCP_HEADER, priority);
3133 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3137 /* Reserve space for headers and prepare control bits. */
3138 skb_reserve(skb, MAX_TCP_HEADER);
3139 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3140 TCPHDR_ACK | TCPHDR_RST);
3141 tcp_mstamp_refresh(tcp_sk(sk));
3143 if (tcp_transmit_skb(sk, skb, 0, priority))
3144 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3146 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3147 * skb here is different to the troublesome skb, so use NULL
3149 trace_tcp_send_reset(sk, NULL);
3152 /* Send a crossed SYN-ACK during socket establishment.
3153 * WARNING: This routine must only be called when we have already sent
3154 * a SYN packet that crossed the incoming SYN that caused this routine
3155 * to get called. If this assumption fails then the initial rcv_wnd
3156 * and rcv_wscale values will not be correct.
3158 int tcp_send_synack(struct sock *sk)
3160 struct sk_buff *skb;
3162 skb = tcp_rtx_queue_head(sk);
3163 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3164 pr_err("%s: wrong queue state\n", __func__);
3167 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3168 if (skb_cloned(skb)) {
3169 struct sk_buff *nskb;
3171 tcp_skb_tsorted_save(skb) {
3172 nskb = skb_copy(skb, GFP_ATOMIC);
3173 } tcp_skb_tsorted_restore(skb);
3176 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3177 tcp_highest_sack_replace(sk, skb, nskb);
3178 tcp_rtx_queue_unlink_and_free(skb, sk);
3179 __skb_header_release(nskb);
3180 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3181 sk->sk_wmem_queued += nskb->truesize;
3182 sk_mem_charge(sk, nskb->truesize);
3186 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3187 tcp_ecn_send_synack(sk, skb);
3189 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3193 * tcp_make_synack - Prepare a SYN-ACK.
3194 * sk: listener socket
3195 * dst: dst entry attached to the SYNACK
3196 * req: request_sock pointer
3198 * Allocate one skb and build a SYNACK packet.
3199 * @dst is consumed : Caller should not use it again.
3201 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3202 struct request_sock *req,
3203 struct tcp_fastopen_cookie *foc,
3204 enum tcp_synack_type synack_type)
3206 struct inet_request_sock *ireq = inet_rsk(req);
3207 const struct tcp_sock *tp = tcp_sk(sk);
3208 struct tcp_md5sig_key *md5 = NULL;
3209 struct tcp_out_options opts;
3210 struct sk_buff *skb;
3211 int tcp_header_size;
3215 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3216 if (unlikely(!skb)) {
3220 /* Reserve space for headers. */
3221 skb_reserve(skb, MAX_TCP_HEADER);
3223 switch (synack_type) {
3224 case TCP_SYNACK_NORMAL:
3225 skb_set_owner_w(skb, req_to_sk(req));
3227 case TCP_SYNACK_COOKIE:
3228 /* Under synflood, we do not attach skb to a socket,
3229 * to avoid false sharing.
3232 case TCP_SYNACK_FASTOPEN:
3233 /* sk is a const pointer, because we want to express multiple
3234 * cpu might call us concurrently.
3235 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3237 skb_set_owner_w(skb, (struct sock *)sk);
3240 skb_dst_set(skb, dst);
3242 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3244 memset(&opts, 0, sizeof(opts));
3245 #ifdef CONFIG_SYN_COOKIES
3246 if (unlikely(req->cookie_ts))
3247 skb->skb_mstamp = cookie_init_timestamp(req);
3250 skb->skb_mstamp = tcp_clock_us();
3252 #ifdef CONFIG_TCP_MD5SIG
3254 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3256 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3257 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3258 foc, synack_type) + sizeof(*th);
3260 skb_push(skb, tcp_header_size);
3261 skb_reset_transport_header(skb);
3263 th = (struct tcphdr *)skb->data;
3264 memset(th, 0, sizeof(struct tcphdr));
3267 tcp_ecn_make_synack(req, th);
3268 th->source = htons(ireq->ir_num);
3269 th->dest = ireq->ir_rmt_port;
3270 skb->mark = ireq->ir_mark;
3271 skb->ip_summed = CHECKSUM_PARTIAL;
3272 th->seq = htonl(tcp_rsk(req)->snt_isn);
3273 /* XXX data is queued and acked as is. No buffer/window check */
3274 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3276 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3277 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3278 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3279 th->doff = (tcp_header_size >> 2);
3280 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3282 #ifdef CONFIG_TCP_MD5SIG
3283 /* Okay, we have all we need - do the md5 hash if needed */
3285 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3286 md5, req_to_sk(req), skb);
3290 /* Do not fool tcpdump (if any), clean our debris */
3294 EXPORT_SYMBOL(tcp_make_synack);
3296 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3298 struct inet_connection_sock *icsk = inet_csk(sk);
3299 const struct tcp_congestion_ops *ca;
3300 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3302 if (ca_key == TCP_CA_UNSPEC)
3306 ca = tcp_ca_find_key(ca_key);
3307 if (likely(ca && try_module_get(ca->owner))) {
3308 module_put(icsk->icsk_ca_ops->owner);
3309 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3310 icsk->icsk_ca_ops = ca;
3315 /* Do all connect socket setups that can be done AF independent. */
3316 static void tcp_connect_init(struct sock *sk)
3318 const struct dst_entry *dst = __sk_dst_get(sk);
3319 struct tcp_sock *tp = tcp_sk(sk);
3323 /* We'll fix this up when we get a response from the other end.
3324 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3326 tp->tcp_header_len = sizeof(struct tcphdr);
3327 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3328 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3330 #ifdef CONFIG_TCP_MD5SIG
3331 if (tp->af_specific->md5_lookup(sk, sk))
3332 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3335 /* If user gave his TCP_MAXSEG, record it to clamp */
3336 if (tp->rx_opt.user_mss)
3337 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3340 tcp_sync_mss(sk, dst_mtu(dst));
3342 tcp_ca_dst_init(sk, dst);
3344 if (!tp->window_clamp)
3345 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3346 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3348 tcp_initialize_rcv_mss(sk);
3350 /* limit the window selection if the user enforce a smaller rx buffer */
3351 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3352 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3353 tp->window_clamp = tcp_full_space(sk);
3355 rcv_wnd = tcp_rwnd_init_bpf(sk);
3357 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3359 tcp_select_initial_window(sk, tcp_full_space(sk),
3360 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3363 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3367 tp->rx_opt.rcv_wscale = rcv_wscale;
3368 tp->rcv_ssthresh = tp->rcv_wnd;
3371 sock_reset_flag(sk, SOCK_DONE);
3374 tcp_write_queue_purge(sk);
3375 tp->snd_una = tp->write_seq;
3376 tp->snd_sml = tp->write_seq;
3377 tp->snd_up = tp->write_seq;
3378 tp->snd_nxt = tp->write_seq;
3380 if (likely(!tp->repair))
3383 tp->rcv_tstamp = tcp_jiffies32;
3384 tp->rcv_wup = tp->rcv_nxt;
3385 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3387 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3388 inet_csk(sk)->icsk_retransmits = 0;
3389 tcp_clear_retrans(tp);
3392 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3394 struct tcp_sock *tp = tcp_sk(sk);
3395 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3397 tcb->end_seq += skb->len;
3398 __skb_header_release(skb);
3399 sk->sk_wmem_queued += skb->truesize;
3400 sk_mem_charge(sk, skb->truesize);
3401 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3402 tp->packets_out += tcp_skb_pcount(skb);
3405 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3406 * queue a data-only packet after the regular SYN, such that regular SYNs
3407 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3408 * only the SYN sequence, the data are retransmitted in the first ACK.
3409 * If cookie is not cached or other error occurs, falls back to send a
3410 * regular SYN with Fast Open cookie request option.
3412 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3414 struct tcp_sock *tp = tcp_sk(sk);
3415 struct tcp_fastopen_request *fo = tp->fastopen_req;
3417 struct sk_buff *syn_data;
3419 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3420 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3423 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3424 * user-MSS. Reserve maximum option space for middleboxes that add
3425 * private TCP options. The cost is reduced data space in SYN :(
3427 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3429 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3430 MAX_TCP_OPTION_SPACE;
3432 space = min_t(size_t, space, fo->size);
3434 /* limit to order-0 allocations */
3435 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3437 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3440 syn_data->ip_summed = CHECKSUM_PARTIAL;
3441 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3443 int copied = copy_from_iter(skb_put(syn_data, space), space,
3444 &fo->data->msg_iter);
3445 if (unlikely(!copied)) {
3446 tcp_skb_tsorted_anchor_cleanup(syn_data);
3447 kfree_skb(syn_data);
3450 if (copied != space) {
3451 skb_trim(syn_data, copied);
3455 /* No more data pending in inet_wait_for_connect() */
3456 if (space == fo->size)
3460 tcp_connect_queue_skb(sk, syn_data);
3462 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3464 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3466 syn->skb_mstamp = syn_data->skb_mstamp;
3468 /* Now full SYN+DATA was cloned and sent (or not),
3469 * remove the SYN from the original skb (syn_data)
3470 * we keep in write queue in case of a retransmit, as we
3471 * also have the SYN packet (with no data) in the same queue.
3473 TCP_SKB_CB(syn_data)->seq++;
3474 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3476 tp->syn_data = (fo->copied > 0);
3477 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3478 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3482 /* data was not sent, put it in write_queue */
3483 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3484 tp->packets_out -= tcp_skb_pcount(syn_data);
3487 /* Send a regular SYN with Fast Open cookie request option */
3488 if (fo->cookie.len > 0)
3490 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3492 tp->syn_fastopen = 0;
3494 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3498 /* Build a SYN and send it off. */
3499 int tcp_connect(struct sock *sk)
3501 struct tcp_sock *tp = tcp_sk(sk);
3502 struct sk_buff *buff;
3505 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3507 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3508 return -EHOSTUNREACH; /* Routing failure or similar. */
3510 tcp_connect_init(sk);
3512 if (unlikely(tp->repair)) {
3513 tcp_finish_connect(sk, NULL);
3517 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3518 if (unlikely(!buff))
3521 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3522 tcp_mstamp_refresh(tp);
3523 tp->retrans_stamp = tcp_time_stamp(tp);
3524 tcp_connect_queue_skb(sk, buff);
3525 tcp_ecn_send_syn(sk, buff);
3526 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3528 /* Send off SYN; include data in Fast Open. */
3529 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3530 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3531 if (err == -ECONNREFUSED)
3534 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3535 * in order to make this packet get counted in tcpOutSegs.
3537 tp->snd_nxt = tp->write_seq;
3538 tp->pushed_seq = tp->write_seq;
3539 buff = tcp_send_head(sk);
3540 if (unlikely(buff)) {
3541 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3542 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3544 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3546 /* Timer for repeating the SYN until an answer. */
3547 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3548 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3551 EXPORT_SYMBOL(tcp_connect);
3553 /* Send out a delayed ack, the caller does the policy checking
3554 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3557 void tcp_send_delayed_ack(struct sock *sk)
3559 struct inet_connection_sock *icsk = inet_csk(sk);
3560 int ato = icsk->icsk_ack.ato;
3561 unsigned long timeout;
3563 if (ato > TCP_DELACK_MIN) {
3564 const struct tcp_sock *tp = tcp_sk(sk);
3565 int max_ato = HZ / 2;
3567 if (icsk->icsk_ack.pingpong ||
3568 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3569 max_ato = TCP_DELACK_MAX;
3571 /* Slow path, intersegment interval is "high". */
3573 /* If some rtt estimate is known, use it to bound delayed ack.
3574 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3578 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3585 ato = min(ato, max_ato);
3588 /* Stay within the limit we were given */
3589 timeout = jiffies + ato;
3591 /* Use new timeout only if there wasn't a older one earlier. */
3592 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3593 /* If delack timer was blocked or is about to expire,
3596 if (icsk->icsk_ack.blocked ||
3597 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3602 if (!time_before(timeout, icsk->icsk_ack.timeout))
3603 timeout = icsk->icsk_ack.timeout;
3605 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3606 icsk->icsk_ack.timeout = timeout;
3607 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3610 /* This routine sends an ack and also updates the window. */
3611 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3613 struct sk_buff *buff;
3615 /* If we have been reset, we may not send again. */
3616 if (sk->sk_state == TCP_CLOSE)
3619 /* We are not putting this on the write queue, so
3620 * tcp_transmit_skb() will set the ownership to this
3623 buff = alloc_skb(MAX_TCP_HEADER,
3624 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3625 if (unlikely(!buff)) {
3626 inet_csk_schedule_ack(sk);
3627 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3628 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3629 TCP_DELACK_MAX, TCP_RTO_MAX);
3633 /* Reserve space for headers and prepare control bits. */
3634 skb_reserve(buff, MAX_TCP_HEADER);
3635 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3637 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3639 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3641 skb_set_tcp_pure_ack(buff);
3643 /* Send it off, this clears delayed acks for us. */
3644 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3646 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3648 void tcp_send_ack(struct sock *sk)
3650 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3653 /* This routine sends a packet with an out of date sequence
3654 * number. It assumes the other end will try to ack it.
3656 * Question: what should we make while urgent mode?
3657 * 4.4BSD forces sending single byte of data. We cannot send
3658 * out of window data, because we have SND.NXT==SND.MAX...
3660 * Current solution: to send TWO zero-length segments in urgent mode:
3661 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3662 * out-of-date with SND.UNA-1 to probe window.
3664 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3666 struct tcp_sock *tp = tcp_sk(sk);
3667 struct sk_buff *skb;
3669 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3670 skb = alloc_skb(MAX_TCP_HEADER,
3671 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3675 /* Reserve space for headers and set control bits. */
3676 skb_reserve(skb, MAX_TCP_HEADER);
3677 /* Use a previous sequence. This should cause the other
3678 * end to send an ack. Don't queue or clone SKB, just
3681 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3682 NET_INC_STATS(sock_net(sk), mib);
3683 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3686 /* Called from setsockopt( ... TCP_REPAIR ) */
3687 void tcp_send_window_probe(struct sock *sk)
3689 if (sk->sk_state == TCP_ESTABLISHED) {
3690 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3691 tcp_mstamp_refresh(tcp_sk(sk));
3692 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3696 /* Initiate keepalive or window probe from timer. */
3697 int tcp_write_wakeup(struct sock *sk, int mib)
3699 struct tcp_sock *tp = tcp_sk(sk);
3700 struct sk_buff *skb;
3702 if (sk->sk_state == TCP_CLOSE)
3705 skb = tcp_send_head(sk);
3706 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3708 unsigned int mss = tcp_current_mss(sk);
3709 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3711 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3712 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3714 /* We are probing the opening of a window
3715 * but the window size is != 0
3716 * must have been a result SWS avoidance ( sender )
3718 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3720 seg_size = min(seg_size, mss);
3721 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3722 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3723 skb, seg_size, mss, GFP_ATOMIC))
3725 } else if (!tcp_skb_pcount(skb))
3726 tcp_set_skb_tso_segs(skb, mss);
3728 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3729 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3731 tcp_event_new_data_sent(sk, skb);
3734 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3735 tcp_xmit_probe_skb(sk, 1, mib);
3736 return tcp_xmit_probe_skb(sk, 0, mib);
3740 /* A window probe timeout has occurred. If window is not closed send
3741 * a partial packet else a zero probe.
3743 void tcp_send_probe0(struct sock *sk)
3745 struct inet_connection_sock *icsk = inet_csk(sk);
3746 struct tcp_sock *tp = tcp_sk(sk);
3747 struct net *net = sock_net(sk);
3748 unsigned long probe_max;
3751 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3753 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3754 /* Cancel probe timer, if it is not required. */
3755 icsk->icsk_probes_out = 0;
3756 icsk->icsk_backoff = 0;
3761 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3762 icsk->icsk_backoff++;
3763 icsk->icsk_probes_out++;
3764 probe_max = TCP_RTO_MAX;
3766 /* If packet was not sent due to local congestion,
3767 * do not backoff and do not remember icsk_probes_out.
3768 * Let local senders to fight for local resources.
3770 * Use accumulated backoff yet.
3772 if (!icsk->icsk_probes_out)
3773 icsk->icsk_probes_out = 1;
3774 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3776 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3777 tcp_probe0_when(sk, probe_max),
3781 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3783 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3787 tcp_rsk(req)->txhash = net_tx_rndhash();
3788 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3790 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3791 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3792 if (unlikely(tcp_passive_fastopen(sk)))
3793 tcp_sk(sk)->total_retrans++;
3794 trace_tcp_retransmit_synack(sk, req);
3798 EXPORT_SYMBOL(tcp_rtx_synack);