1 // SPDX-License-Identifier: GPL-2.0-only
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * Implementation of the Transmission Control Protocol(TCP).
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
38 #define pr_fmt(fmt) "TCP: " fmt
42 #include <linux/compiler.h>
43 #include <linux/gfp.h>
44 #include <linux/module.h>
45 #include <linux/static_key.h>
47 #include <trace/events/tcp.h>
49 /* Refresh clocks of a TCP socket,
50 * ensuring monotically increasing values.
52 void tcp_mstamp_refresh(struct tcp_sock *tp)
54 u64 val = tcp_clock_ns();
56 tp->tcp_clock_cache = val;
57 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
60 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
61 int push_one, gfp_t gfp);
63 /* Account for new data that has been sent to the network. */
64 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66 struct inet_connection_sock *icsk = inet_csk(sk);
67 struct tcp_sock *tp = tcp_sk(sk);
68 unsigned int prior_packets = tp->packets_out;
70 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72 __skb_unlink(skb, &sk->sk_write_queue);
73 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75 if (tp->highest_sack == NULL)
76 tp->highest_sack = skb;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
82 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
87 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
94 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96 const struct tcp_sock *tp = tcp_sk(sk);
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
103 return tcp_wnd_end(tp);
106 /* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
120 static __u16 tcp_advertise_mss(struct sock *sk)
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
127 unsigned int metric = dst_metric_advmss(dst);
138 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
141 void tcp_cwnd_restart(struct sock *sk, s32 delta)
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
159 /* Congestion state accounting after a packet has been sent. */
160 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
185 struct tcp_sock *tp = tcp_sk(sk);
187 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
189 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
190 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
195 if (unlikely(rcv_nxt != tp->rcv_nxt))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk, pkts);
198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
201 /* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
208 void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
210 int wscale_ok, __u8 *rcv_wscale,
213 unsigned int space = (__space < 0 ? 0 : __space);
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
218 space = min(*window_clamp, space);
220 /* Quantize space offering to a multiple of mss if possible. */
222 space = rounddown(space, mss);
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
235 (*rcv_wnd) = min_t(u32, space, U16_MAX);
238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
252 EXPORT_SYMBOL(tcp_select_initial_window);
254 /* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
259 static u16 tcp_select_window(struct sock *sk)
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
273 * Relax Will Robinson.
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
283 /* Make sure we do not exceed the maximum possible
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
295 /* If we advertise zero window, disable fast path. */
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
308 /* Packet ECN state for a SYN-ACK */
309 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
311 const struct tcp_sock *tp = tcp_sk(sk);
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
321 /* Packet ECN state for a SYN. */
322 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
330 const struct dst_entry *dst = __sk_dst_get(sk);
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
346 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
356 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358 if (inet_rsk(req)->ecn_ok)
362 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
368 struct tcp_sock *tp = tcp_sk(sk);
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
389 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
392 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394 skb->ip_summed = CHECKSUM_PARTIAL;
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
399 tcp_skb_pcount_set(skb, 1);
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
404 TCP_SKB_CB(skb)->end_seq = seq;
407 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
409 return tp->snd_una != tp->snd_up;
412 #define OPTION_SACK_ADVERTISE (1 << 0)
413 #define OPTION_TS (1 << 1)
414 #define OPTION_MD5 (1 << 2)
415 #define OPTION_WSCALE (1 << 3)
416 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
417 #define OPTION_SMC (1 << 9)
419 static void smc_options_write(__be32 *ptr, u16 *options)
421 #if IS_ENABLED(CONFIG_SMC)
422 if (static_branch_unlikely(&tcp_have_smc)) {
423 if (unlikely(OPTION_SMC & *options)) {
424 *ptr++ = htonl((TCPOPT_NOP << 24) |
427 (TCPOLEN_EXP_SMC_BASE));
428 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
434 struct tcp_out_options {
435 u16 options; /* bit field of OPTION_* */
436 u16 mss; /* 0 to disable */
437 u8 ws; /* window scale, 0 to disable */
438 u8 num_sack_blocks; /* number of SACK blocks to include */
439 u8 hash_size; /* bytes in hash_location */
440 __u8 *hash_location; /* temporary pointer, overloaded */
441 __u32 tsval, tsecr; /* need to include OPTION_TS */
442 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
445 /* Write previously computed TCP options to the packet.
447 * Beware: Something in the Internet is very sensitive to the ordering of
448 * TCP options, we learned this through the hard way, so be careful here.
449 * Luckily we can at least blame others for their non-compliance but from
450 * inter-operability perspective it seems that we're somewhat stuck with
451 * the ordering which we have been using if we want to keep working with
452 * those broken things (not that it currently hurts anybody as there isn't
453 * particular reason why the ordering would need to be changed).
455 * At least SACK_PERM as the first option is known to lead to a disaster
456 * (but it may well be that other scenarios fail similarly).
458 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
459 struct tcp_out_options *opts)
461 u16 options = opts->options; /* mungable copy */
463 if (unlikely(OPTION_MD5 & options)) {
464 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
465 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
466 /* overload cookie hash location */
467 opts->hash_location = (__u8 *)ptr;
471 if (unlikely(opts->mss)) {
472 *ptr++ = htonl((TCPOPT_MSS << 24) |
473 (TCPOLEN_MSS << 16) |
477 if (likely(OPTION_TS & options)) {
478 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
479 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
480 (TCPOLEN_SACK_PERM << 16) |
481 (TCPOPT_TIMESTAMP << 8) |
483 options &= ~OPTION_SACK_ADVERTISE;
485 *ptr++ = htonl((TCPOPT_NOP << 24) |
487 (TCPOPT_TIMESTAMP << 8) |
490 *ptr++ = htonl(opts->tsval);
491 *ptr++ = htonl(opts->tsecr);
494 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
497 (TCPOPT_SACK_PERM << 8) |
501 if (unlikely(OPTION_WSCALE & options)) {
502 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOPT_WINDOW << 16) |
504 (TCPOLEN_WINDOW << 8) |
508 if (unlikely(opts->num_sack_blocks)) {
509 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
510 tp->duplicate_sack : tp->selective_acks;
513 *ptr++ = htonl((TCPOPT_NOP << 24) |
516 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
517 TCPOLEN_SACK_PERBLOCK)));
519 for (this_sack = 0; this_sack < opts->num_sack_blocks;
521 *ptr++ = htonl(sp[this_sack].start_seq);
522 *ptr++ = htonl(sp[this_sack].end_seq);
525 tp->rx_opt.dsack = 0;
528 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
529 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
531 u32 len; /* Fast Open option length */
534 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
535 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
536 TCPOPT_FASTOPEN_MAGIC);
537 p += TCPOLEN_EXP_FASTOPEN_BASE;
539 len = TCPOLEN_FASTOPEN_BASE + foc->len;
540 *p++ = TCPOPT_FASTOPEN;
544 memcpy(p, foc->val, foc->len);
545 if ((len & 3) == 2) {
546 p[foc->len] = TCPOPT_NOP;
547 p[foc->len + 1] = TCPOPT_NOP;
549 ptr += (len + 3) >> 2;
552 smc_options_write(ptr, &options);
555 static void smc_set_option(const struct tcp_sock *tp,
556 struct tcp_out_options *opts,
557 unsigned int *remaining)
559 #if IS_ENABLED(CONFIG_SMC)
560 if (static_branch_unlikely(&tcp_have_smc)) {
562 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
563 opts->options |= OPTION_SMC;
564 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
571 static void smc_set_option_cond(const struct tcp_sock *tp,
572 const struct inet_request_sock *ireq,
573 struct tcp_out_options *opts,
574 unsigned int *remaining)
576 #if IS_ENABLED(CONFIG_SMC)
577 if (static_branch_unlikely(&tcp_have_smc)) {
578 if (tp->syn_smc && ireq->smc_ok) {
579 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
580 opts->options |= OPTION_SMC;
581 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
588 /* Compute TCP options for SYN packets. This is not the final
589 * network wire format yet.
591 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
592 struct tcp_out_options *opts,
593 struct tcp_md5sig_key **md5)
595 struct tcp_sock *tp = tcp_sk(sk);
596 unsigned int remaining = MAX_TCP_OPTION_SPACE;
597 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
600 #ifdef CONFIG_TCP_MD5SIG
601 if (static_branch_unlikely(&tcp_md5_needed) &&
602 rcu_access_pointer(tp->md5sig_info)) {
603 *md5 = tp->af_specific->md5_lookup(sk, sk);
605 opts->options |= OPTION_MD5;
606 remaining -= TCPOLEN_MD5SIG_ALIGNED;
611 /* We always get an MSS option. The option bytes which will be seen in
612 * normal data packets should timestamps be used, must be in the MSS
613 * advertised. But we subtract them from tp->mss_cache so that
614 * calculations in tcp_sendmsg are simpler etc. So account for this
615 * fact here if necessary. If we don't do this correctly, as a
616 * receiver we won't recognize data packets as being full sized when we
617 * should, and thus we won't abide by the delayed ACK rules correctly.
618 * SACKs don't matter, we never delay an ACK when we have any of those
620 opts->mss = tcp_advertise_mss(sk);
621 remaining -= TCPOLEN_MSS_ALIGNED;
623 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
624 opts->options |= OPTION_TS;
625 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
626 opts->tsecr = tp->rx_opt.ts_recent;
627 remaining -= TCPOLEN_TSTAMP_ALIGNED;
629 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
630 opts->ws = tp->rx_opt.rcv_wscale;
631 opts->options |= OPTION_WSCALE;
632 remaining -= TCPOLEN_WSCALE_ALIGNED;
634 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
635 opts->options |= OPTION_SACK_ADVERTISE;
636 if (unlikely(!(OPTION_TS & opts->options)))
637 remaining -= TCPOLEN_SACKPERM_ALIGNED;
640 if (fastopen && fastopen->cookie.len >= 0) {
641 u32 need = fastopen->cookie.len;
643 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
644 TCPOLEN_FASTOPEN_BASE;
645 need = (need + 3) & ~3U; /* Align to 32 bits */
646 if (remaining >= need) {
647 opts->options |= OPTION_FAST_OPEN_COOKIE;
648 opts->fastopen_cookie = &fastopen->cookie;
650 tp->syn_fastopen = 1;
651 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
655 smc_set_option(tp, opts, &remaining);
657 return MAX_TCP_OPTION_SPACE - remaining;
660 /* Set up TCP options for SYN-ACKs. */
661 static unsigned int tcp_synack_options(const struct sock *sk,
662 struct request_sock *req,
663 unsigned int mss, struct sk_buff *skb,
664 struct tcp_out_options *opts,
665 const struct tcp_md5sig_key *md5,
666 struct tcp_fastopen_cookie *foc,
667 enum tcp_synack_type synack_type)
669 struct inet_request_sock *ireq = inet_rsk(req);
670 unsigned int remaining = MAX_TCP_OPTION_SPACE;
672 #ifdef CONFIG_TCP_MD5SIG
674 opts->options |= OPTION_MD5;
675 remaining -= TCPOLEN_MD5SIG_ALIGNED;
677 /* We can't fit any SACK blocks in a packet with MD5 + TS
678 * options. There was discussion about disabling SACK
679 * rather than TS in order to fit in better with old,
680 * buggy kernels, but that was deemed to be unnecessary.
682 if (synack_type != TCP_SYNACK_COOKIE)
683 ireq->tstamp_ok &= !ireq->sack_ok;
687 /* We always send an MSS option. */
689 remaining -= TCPOLEN_MSS_ALIGNED;
691 if (likely(ireq->wscale_ok)) {
692 opts->ws = ireq->rcv_wscale;
693 opts->options |= OPTION_WSCALE;
694 remaining -= TCPOLEN_WSCALE_ALIGNED;
696 if (likely(ireq->tstamp_ok)) {
697 opts->options |= OPTION_TS;
698 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
699 opts->tsecr = req->ts_recent;
700 remaining -= TCPOLEN_TSTAMP_ALIGNED;
702 if (likely(ireq->sack_ok)) {
703 opts->options |= OPTION_SACK_ADVERTISE;
704 if (unlikely(!ireq->tstamp_ok))
705 remaining -= TCPOLEN_SACKPERM_ALIGNED;
707 if (foc != NULL && foc->len >= 0) {
710 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
711 TCPOLEN_FASTOPEN_BASE;
712 need = (need + 3) & ~3U; /* Align to 32 bits */
713 if (remaining >= need) {
714 opts->options |= OPTION_FAST_OPEN_COOKIE;
715 opts->fastopen_cookie = foc;
720 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
722 return MAX_TCP_OPTION_SPACE - remaining;
725 /* Compute TCP options for ESTABLISHED sockets. This is not the
726 * final wire format yet.
728 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
729 struct tcp_out_options *opts,
730 struct tcp_md5sig_key **md5)
732 struct tcp_sock *tp = tcp_sk(sk);
733 unsigned int size = 0;
734 unsigned int eff_sacks;
739 #ifdef CONFIG_TCP_MD5SIG
740 if (static_branch_unlikely(&tcp_md5_needed) &&
741 rcu_access_pointer(tp->md5sig_info)) {
742 *md5 = tp->af_specific->md5_lookup(sk, sk);
744 opts->options |= OPTION_MD5;
745 size += TCPOLEN_MD5SIG_ALIGNED;
750 if (likely(tp->rx_opt.tstamp_ok)) {
751 opts->options |= OPTION_TS;
752 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
753 opts->tsecr = tp->rx_opt.ts_recent;
754 size += TCPOLEN_TSTAMP_ALIGNED;
757 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
758 if (unlikely(eff_sacks)) {
759 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
760 opts->num_sack_blocks =
761 min_t(unsigned int, eff_sacks,
762 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
763 TCPOLEN_SACK_PERBLOCK);
764 if (likely(opts->num_sack_blocks))
765 size += TCPOLEN_SACK_BASE_ALIGNED +
766 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
773 /* TCP SMALL QUEUES (TSQ)
775 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
776 * to reduce RTT and bufferbloat.
777 * We do this using a special skb destructor (tcp_wfree).
779 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
780 * needs to be reallocated in a driver.
781 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
783 * Since transmit from skb destructor is forbidden, we use a tasklet
784 * to process all sockets that eventually need to send more skbs.
785 * We use one tasklet per cpu, with its own queue of sockets.
788 struct tasklet_struct tasklet;
789 struct list_head head; /* queue of tcp sockets */
791 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
793 static void tcp_tsq_write(struct sock *sk)
795 if ((1 << sk->sk_state) &
796 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
797 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
798 struct tcp_sock *tp = tcp_sk(sk);
800 if (tp->lost_out > tp->retrans_out &&
801 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
802 tcp_mstamp_refresh(tp);
803 tcp_xmit_retransmit_queue(sk);
806 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
811 static void tcp_tsq_handler(struct sock *sk)
814 if (!sock_owned_by_user(sk))
816 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
821 * One tasklet per cpu tries to send more skbs.
822 * We run in tasklet context but need to disable irqs when
823 * transferring tsq->head because tcp_wfree() might
824 * interrupt us (non NAPI drivers)
826 static void tcp_tasklet_func(unsigned long data)
828 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
831 struct list_head *q, *n;
835 local_irq_save(flags);
836 list_splice_init(&tsq->head, &list);
837 local_irq_restore(flags);
839 list_for_each_safe(q, n, &list) {
840 tp = list_entry(q, struct tcp_sock, tsq_node);
841 list_del(&tp->tsq_node);
843 sk = (struct sock *)tp;
844 smp_mb__before_atomic();
845 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
852 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
853 TCPF_WRITE_TIMER_DEFERRED | \
854 TCPF_DELACK_TIMER_DEFERRED | \
855 TCPF_MTU_REDUCED_DEFERRED)
857 * tcp_release_cb - tcp release_sock() callback
860 * called from release_sock() to perform protocol dependent
861 * actions before socket release.
863 void tcp_release_cb(struct sock *sk)
865 unsigned long flags, nflags;
867 /* perform an atomic operation only if at least one flag is set */
869 flags = sk->sk_tsq_flags;
870 if (!(flags & TCP_DEFERRED_ALL))
872 nflags = flags & ~TCP_DEFERRED_ALL;
873 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
875 if (flags & TCPF_TSQ_DEFERRED) {
879 /* Here begins the tricky part :
880 * We are called from release_sock() with :
882 * 2) sk_lock.slock spinlock held
883 * 3) socket owned by us (sk->sk_lock.owned == 1)
885 * But following code is meant to be called from BH handlers,
886 * so we should keep BH disabled, but early release socket ownership
888 sock_release_ownership(sk);
890 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
891 tcp_write_timer_handler(sk);
894 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
895 tcp_delack_timer_handler(sk);
898 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
899 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
903 EXPORT_SYMBOL(tcp_release_cb);
905 void __init tcp_tasklet_init(void)
909 for_each_possible_cpu(i) {
910 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
912 INIT_LIST_HEAD(&tsq->head);
913 tasklet_init(&tsq->tasklet,
920 * Write buffer destructor automatically called from kfree_skb.
921 * We can't xmit new skbs from this context, as we might already
924 void tcp_wfree(struct sk_buff *skb)
926 struct sock *sk = skb->sk;
927 struct tcp_sock *tp = tcp_sk(sk);
928 unsigned long flags, nval, oval;
930 /* Keep one reference on sk_wmem_alloc.
931 * Will be released by sk_free() from here or tcp_tasklet_func()
933 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
935 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
936 * Wait until our queues (qdisc + devices) are drained.
938 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
939 * - chance for incoming ACK (processed by another cpu maybe)
940 * to migrate this flow (skb->ooo_okay will be eventually set)
942 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
945 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
946 struct tsq_tasklet *tsq;
949 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
952 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
953 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
957 /* queue this socket to tasklet queue */
958 local_irq_save(flags);
959 tsq = this_cpu_ptr(&tsq_tasklet);
960 empty = list_empty(&tsq->head);
961 list_add(&tp->tsq_node, &tsq->head);
963 tasklet_schedule(&tsq->tasklet);
964 local_irq_restore(flags);
971 /* Note: Called under soft irq.
972 * We can call TCP stack right away, unless socket is owned by user.
974 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
976 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
977 struct sock *sk = (struct sock *)tp;
982 return HRTIMER_NORESTART;
985 static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
988 struct tcp_sock *tp = tcp_sk(sk);
990 if (sk->sk_pacing_status != SK_PACING_NONE) {
991 unsigned long rate = sk->sk_pacing_rate;
993 /* Original sch_fq does not pace first 10 MSS
994 * Note that tp->data_segs_out overflows after 2^32 packets,
995 * this is a minor annoyance.
997 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
998 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
999 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1001 /* take into account OS jitter */
1002 len_ns -= min_t(u64, len_ns / 2, credit);
1003 tp->tcp_wstamp_ns += len_ns;
1006 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1009 /* This routine actually transmits TCP packets queued in by
1010 * tcp_do_sendmsg(). This is used by both the initial
1011 * transmission and possible later retransmissions.
1012 * All SKB's seen here are completely headerless. It is our
1013 * job to build the TCP header, and pass the packet down to
1014 * IP so it can do the same plus pass the packet off to the
1017 * We are working here with either a clone of the original
1018 * SKB, or a fresh unique copy made by the retransmit engine.
1020 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1021 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1023 const struct inet_connection_sock *icsk = inet_csk(sk);
1024 struct inet_sock *inet;
1025 struct tcp_sock *tp;
1026 struct tcp_skb_cb *tcb;
1027 struct tcp_out_options opts;
1028 unsigned int tcp_options_size, tcp_header_size;
1029 struct sk_buff *oskb = NULL;
1030 struct tcp_md5sig_key *md5;
1035 BUG_ON(!skb || !tcp_skb_pcount(skb));
1037 prior_wstamp = tp->tcp_wstamp_ns;
1038 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1039 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1041 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1045 tcp_skb_tsorted_save(oskb) {
1046 if (unlikely(skb_cloned(oskb)))
1047 skb = pskb_copy(oskb, gfp_mask);
1049 skb = skb_clone(oskb, gfp_mask);
1050 } tcp_skb_tsorted_restore(oskb);
1054 /* retransmit skbs might have a non zero value in skb->dev
1055 * because skb->dev is aliased with skb->rbnode.rb_left
1061 tcb = TCP_SKB_CB(skb);
1062 memset(&opts, 0, sizeof(opts));
1064 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1065 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1067 tcp_options_size = tcp_established_options(sk, skb, &opts,
1069 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1070 * at receiver : This slightly improve GRO performance.
1071 * Note that we do not force the PSH flag for non GSO packets,
1072 * because they might be sent under high congestion events,
1073 * and in this case it is better to delay the delivery of 1-MSS
1074 * packets and thus the corresponding ACK packet that would
1075 * release the following packet.
1077 if (tcp_skb_pcount(skb) > 1)
1078 tcb->tcp_flags |= TCPHDR_PSH;
1080 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1082 /* if no packet is in qdisc/device queue, then allow XPS to select
1083 * another queue. We can be called from tcp_tsq_handler()
1084 * which holds one reference to sk.
1086 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1087 * One way to get this would be to set skb->truesize = 2 on them.
1089 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1091 /* If we had to use memory reserve to allocate this skb,
1092 * this might cause drops if packet is looped back :
1093 * Other socket might not have SOCK_MEMALLOC.
1094 * Packets not looped back do not care about pfmemalloc.
1096 skb->pfmemalloc = 0;
1098 skb_push(skb, tcp_header_size);
1099 skb_reset_transport_header(skb);
1103 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1104 skb_set_hash_from_sk(skb, sk);
1105 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1107 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1109 /* Build TCP header and checksum it. */
1110 th = (struct tcphdr *)skb->data;
1111 th->source = inet->inet_sport;
1112 th->dest = inet->inet_dport;
1113 th->seq = htonl(tcb->seq);
1114 th->ack_seq = htonl(rcv_nxt);
1115 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1121 /* The urg_mode check is necessary during a below snd_una win probe */
1122 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1123 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1124 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1126 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1127 th->urg_ptr = htons(0xFFFF);
1132 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1133 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1134 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1135 th->window = htons(tcp_select_window(sk));
1136 tcp_ecn_send(sk, skb, th, tcp_header_size);
1138 /* RFC1323: The window in SYN & SYN/ACK segments
1141 th->window = htons(min(tp->rcv_wnd, 65535U));
1143 #ifdef CONFIG_TCP_MD5SIG
1144 /* Calculate the MD5 hash, as we have all we need now */
1146 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1147 tp->af_specific->calc_md5_hash(opts.hash_location,
1152 icsk->icsk_af_ops->send_check(sk, skb);
1154 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1155 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1157 if (skb->len != tcp_header_size) {
1158 tcp_event_data_sent(tp, sk);
1159 tp->data_segs_out += tcp_skb_pcount(skb);
1160 tp->bytes_sent += skb->len - tcp_header_size;
1163 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1164 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1165 tcp_skb_pcount(skb));
1167 tp->segs_out += tcp_skb_pcount(skb);
1168 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1169 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1170 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1172 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1174 /* Cleanup our debris for IP stacks */
1175 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1176 sizeof(struct inet6_skb_parm)));
1178 tcp_add_tx_delay(skb, tp);
1180 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1182 if (unlikely(err > 0)) {
1184 err = net_xmit_eval(err);
1187 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1188 tcp_rate_skb_sent(sk, oskb);
1193 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1196 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1197 tcp_sk(sk)->rcv_nxt);
1200 /* This routine just queues the buffer for sending.
1202 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1203 * otherwise socket can stall.
1205 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1207 struct tcp_sock *tp = tcp_sk(sk);
1209 /* Advance write_seq and place onto the write_queue. */
1210 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1211 __skb_header_release(skb);
1212 tcp_add_write_queue_tail(sk, skb);
1213 sk_wmem_queued_add(sk, skb->truesize);
1214 sk_mem_charge(sk, skb->truesize);
1217 /* Initialize TSO segments for a packet. */
1218 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1220 if (skb->len <= mss_now) {
1221 /* Avoid the costly divide in the normal
1224 tcp_skb_pcount_set(skb, 1);
1225 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1227 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1228 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1232 /* Pcount in the middle of the write queue got changed, we need to do various
1233 * tweaks to fix counters
1235 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1237 struct tcp_sock *tp = tcp_sk(sk);
1239 tp->packets_out -= decr;
1241 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1242 tp->sacked_out -= decr;
1243 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1244 tp->retrans_out -= decr;
1245 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1246 tp->lost_out -= decr;
1248 /* Reno case is special. Sigh... */
1249 if (tcp_is_reno(tp) && decr > 0)
1250 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1252 if (tp->lost_skb_hint &&
1253 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1254 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1255 tp->lost_cnt_hint -= decr;
1257 tcp_verify_left_out(tp);
1260 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1262 return TCP_SKB_CB(skb)->txstamp_ack ||
1263 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1266 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1268 struct skb_shared_info *shinfo = skb_shinfo(skb);
1270 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1271 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1272 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1273 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1275 shinfo->tx_flags &= ~tsflags;
1276 shinfo2->tx_flags |= tsflags;
1277 swap(shinfo->tskey, shinfo2->tskey);
1278 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1279 TCP_SKB_CB(skb)->txstamp_ack = 0;
1283 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1285 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1286 TCP_SKB_CB(skb)->eor = 0;
1289 /* Insert buff after skb on the write or rtx queue of sk. */
1290 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1291 struct sk_buff *buff,
1293 enum tcp_queue tcp_queue)
1295 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1296 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1298 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1301 /* Function to create two new TCP segments. Shrinks the given segment
1302 * to the specified size and appends a new segment with the rest of the
1303 * packet to the list. This won't be called frequently, I hope.
1304 * Remember, these are still headerless SKBs at this point.
1306 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1307 struct sk_buff *skb, u32 len,
1308 unsigned int mss_now, gfp_t gfp)
1310 struct tcp_sock *tp = tcp_sk(sk);
1311 struct sk_buff *buff;
1312 int nsize, old_factor;
1317 if (WARN_ON(len > skb->len))
1320 nsize = skb_headlen(skb) - len;
1324 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1325 * We need some allowance to not penalize applications setting small
1327 * Also allow first and last skb in retransmit queue to be split.
1329 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1330 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1331 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1332 skb != tcp_rtx_queue_head(sk) &&
1333 skb != tcp_rtx_queue_tail(sk))) {
1334 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1338 if (skb_unclone(skb, gfp))
1341 /* Get a new skb... force flag on. */
1342 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1344 return -ENOMEM; /* We'll just try again later. */
1345 skb_copy_decrypted(buff, skb);
1347 sk_wmem_queued_add(sk, buff->truesize);
1348 sk_mem_charge(sk, buff->truesize);
1349 nlen = skb->len - len - nsize;
1350 buff->truesize += nlen;
1351 skb->truesize -= nlen;
1353 /* Correct the sequence numbers. */
1354 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1355 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1356 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1358 /* PSH and FIN should only be set in the second packet. */
1359 flags = TCP_SKB_CB(skb)->tcp_flags;
1360 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1361 TCP_SKB_CB(buff)->tcp_flags = flags;
1362 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1363 tcp_skb_fragment_eor(skb, buff);
1365 skb_split(skb, buff, len);
1367 buff->ip_summed = CHECKSUM_PARTIAL;
1369 buff->tstamp = skb->tstamp;
1370 tcp_fragment_tstamp(skb, buff);
1372 old_factor = tcp_skb_pcount(skb);
1374 /* Fix up tso_factor for both original and new SKB. */
1375 tcp_set_skb_tso_segs(skb, mss_now);
1376 tcp_set_skb_tso_segs(buff, mss_now);
1378 /* Update delivered info for the new segment */
1379 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1381 /* If this packet has been sent out already, we must
1382 * adjust the various packet counters.
1384 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1385 int diff = old_factor - tcp_skb_pcount(skb) -
1386 tcp_skb_pcount(buff);
1389 tcp_adjust_pcount(sk, skb, diff);
1392 /* Link BUFF into the send queue. */
1393 __skb_header_release(buff);
1394 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1395 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1396 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1401 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1402 * data is not copied, but immediately discarded.
1404 static int __pskb_trim_head(struct sk_buff *skb, int len)
1406 struct skb_shared_info *shinfo;
1409 eat = min_t(int, len, skb_headlen(skb));
1411 __skb_pull(skb, eat);
1418 shinfo = skb_shinfo(skb);
1419 for (i = 0; i < shinfo->nr_frags; i++) {
1420 int size = skb_frag_size(&shinfo->frags[i]);
1423 skb_frag_unref(skb, i);
1426 shinfo->frags[k] = shinfo->frags[i];
1428 skb_frag_off_add(&shinfo->frags[k], eat);
1429 skb_frag_size_sub(&shinfo->frags[k], eat);
1435 shinfo->nr_frags = k;
1437 skb->data_len -= len;
1438 skb->len = skb->data_len;
1442 /* Remove acked data from a packet in the transmit queue. */
1443 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1447 if (skb_unclone(skb, GFP_ATOMIC))
1450 delta_truesize = __pskb_trim_head(skb, len);
1452 TCP_SKB_CB(skb)->seq += len;
1453 skb->ip_summed = CHECKSUM_PARTIAL;
1455 if (delta_truesize) {
1456 skb->truesize -= delta_truesize;
1457 sk_wmem_queued_add(sk, -delta_truesize);
1458 sk_mem_uncharge(sk, delta_truesize);
1459 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1462 /* Any change of skb->len requires recalculation of tso factor. */
1463 if (tcp_skb_pcount(skb) > 1)
1464 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1469 /* Calculate MSS not accounting any TCP options. */
1470 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1472 const struct tcp_sock *tp = tcp_sk(sk);
1473 const struct inet_connection_sock *icsk = inet_csk(sk);
1476 /* Calculate base mss without TCP options:
1477 It is MMS_S - sizeof(tcphdr) of rfc1122
1479 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1481 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1482 if (icsk->icsk_af_ops->net_frag_header_len) {
1483 const struct dst_entry *dst = __sk_dst_get(sk);
1485 if (dst && dst_allfrag(dst))
1486 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1489 /* Clamp it (mss_clamp does not include tcp options) */
1490 if (mss_now > tp->rx_opt.mss_clamp)
1491 mss_now = tp->rx_opt.mss_clamp;
1493 /* Now subtract optional transport overhead */
1494 mss_now -= icsk->icsk_ext_hdr_len;
1496 /* Then reserve room for full set of TCP options and 8 bytes of data */
1497 mss_now = max(mss_now,
1498 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1502 /* Calculate MSS. Not accounting for SACKs here. */
1503 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1505 /* Subtract TCP options size, not including SACKs */
1506 return __tcp_mtu_to_mss(sk, pmtu) -
1507 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1509 EXPORT_SYMBOL(tcp_mtu_to_mss);
1511 /* Inverse of above */
1512 int tcp_mss_to_mtu(struct sock *sk, int mss)
1514 const struct tcp_sock *tp = tcp_sk(sk);
1515 const struct inet_connection_sock *icsk = inet_csk(sk);
1519 tp->tcp_header_len +
1520 icsk->icsk_ext_hdr_len +
1521 icsk->icsk_af_ops->net_header_len;
1523 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1524 if (icsk->icsk_af_ops->net_frag_header_len) {
1525 const struct dst_entry *dst = __sk_dst_get(sk);
1527 if (dst && dst_allfrag(dst))
1528 mtu += icsk->icsk_af_ops->net_frag_header_len;
1532 EXPORT_SYMBOL(tcp_mss_to_mtu);
1534 /* MTU probing init per socket */
1535 void tcp_mtup_init(struct sock *sk)
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct inet_connection_sock *icsk = inet_csk(sk);
1539 struct net *net = sock_net(sk);
1541 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1542 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1543 icsk->icsk_af_ops->net_header_len;
1544 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1545 icsk->icsk_mtup.probe_size = 0;
1546 if (icsk->icsk_mtup.enabled)
1547 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1549 EXPORT_SYMBOL(tcp_mtup_init);
1551 /* This function synchronize snd mss to current pmtu/exthdr set.
1553 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1554 for TCP options, but includes only bare TCP header.
1556 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1557 It is minimum of user_mss and mss received with SYN.
1558 It also does not include TCP options.
1560 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1562 tp->mss_cache is current effective sending mss, including
1563 all tcp options except for SACKs. It is evaluated,
1564 taking into account current pmtu, but never exceeds
1565 tp->rx_opt.mss_clamp.
1567 NOTE1. rfc1122 clearly states that advertised MSS
1568 DOES NOT include either tcp or ip options.
1570 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1571 are READ ONLY outside this function. --ANK (980731)
1573 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1575 struct tcp_sock *tp = tcp_sk(sk);
1576 struct inet_connection_sock *icsk = inet_csk(sk);
1579 if (icsk->icsk_mtup.search_high > pmtu)
1580 icsk->icsk_mtup.search_high = pmtu;
1582 mss_now = tcp_mtu_to_mss(sk, pmtu);
1583 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1585 /* And store cached results */
1586 icsk->icsk_pmtu_cookie = pmtu;
1587 if (icsk->icsk_mtup.enabled)
1588 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1589 tp->mss_cache = mss_now;
1593 EXPORT_SYMBOL(tcp_sync_mss);
1595 /* Compute the current effective MSS, taking SACKs and IP options,
1596 * and even PMTU discovery events into account.
1598 unsigned int tcp_current_mss(struct sock *sk)
1600 const struct tcp_sock *tp = tcp_sk(sk);
1601 const struct dst_entry *dst = __sk_dst_get(sk);
1603 unsigned int header_len;
1604 struct tcp_out_options opts;
1605 struct tcp_md5sig_key *md5;
1607 mss_now = tp->mss_cache;
1610 u32 mtu = dst_mtu(dst);
1611 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1612 mss_now = tcp_sync_mss(sk, mtu);
1615 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1616 sizeof(struct tcphdr);
1617 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1618 * some common options. If this is an odd packet (because we have SACK
1619 * blocks etc) then our calculated header_len will be different, and
1620 * we have to adjust mss_now correspondingly */
1621 if (header_len != tp->tcp_header_len) {
1622 int delta = (int) header_len - tp->tcp_header_len;
1629 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1630 * As additional protections, we do not touch cwnd in retransmission phases,
1631 * and if application hit its sndbuf limit recently.
1633 static void tcp_cwnd_application_limited(struct sock *sk)
1635 struct tcp_sock *tp = tcp_sk(sk);
1637 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1638 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1639 /* Limited by application or receiver window. */
1640 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1641 u32 win_used = max(tp->snd_cwnd_used, init_win);
1642 if (win_used < tp->snd_cwnd) {
1643 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1644 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1646 tp->snd_cwnd_used = 0;
1648 tp->snd_cwnd_stamp = tcp_jiffies32;
1651 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1653 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1654 struct tcp_sock *tp = tcp_sk(sk);
1656 /* Track the strongest available signal of the degree to which the cwnd
1657 * is fully utilized. If cwnd-limited then remember that fact for the
1658 * current window. If not cwnd-limited then track the maximum number of
1659 * outstanding packets in the current window. (If cwnd-limited then we
1660 * chose to not update tp->max_packets_out to avoid an extra else
1661 * clause with no functional impact.)
1663 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1665 (!tp->is_cwnd_limited &&
1666 tp->packets_out > tp->max_packets_out)) {
1667 tp->is_cwnd_limited = is_cwnd_limited;
1668 tp->max_packets_out = tp->packets_out;
1669 tp->cwnd_usage_seq = tp->snd_nxt;
1672 if (tcp_is_cwnd_limited(sk)) {
1673 /* Network is feed fully. */
1674 tp->snd_cwnd_used = 0;
1675 tp->snd_cwnd_stamp = tcp_jiffies32;
1677 /* Network starves. */
1678 if (tp->packets_out > tp->snd_cwnd_used)
1679 tp->snd_cwnd_used = tp->packets_out;
1681 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1682 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1683 !ca_ops->cong_control)
1684 tcp_cwnd_application_limited(sk);
1686 /* The following conditions together indicate the starvation
1687 * is caused by insufficient sender buffer:
1688 * 1) just sent some data (see tcp_write_xmit)
1689 * 2) not cwnd limited (this else condition)
1690 * 3) no more data to send (tcp_write_queue_empty())
1691 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1693 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1694 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1695 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1696 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1700 /* Minshall's variant of the Nagle send check. */
1701 static bool tcp_minshall_check(const struct tcp_sock *tp)
1703 return after(tp->snd_sml, tp->snd_una) &&
1704 !after(tp->snd_sml, tp->snd_nxt);
1707 /* Update snd_sml if this skb is under mss
1708 * Note that a TSO packet might end with a sub-mss segment
1709 * The test is really :
1710 * if ((skb->len % mss) != 0)
1711 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1712 * But we can avoid doing the divide again given we already have
1713 * skb_pcount = skb->len / mss_now
1715 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1716 const struct sk_buff *skb)
1718 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1719 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1722 /* Return false, if packet can be sent now without violation Nagle's rules:
1723 * 1. It is full sized. (provided by caller in %partial bool)
1724 * 2. Or it contains FIN. (already checked by caller)
1725 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1726 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1727 * With Minshall's modification: all sent small packets are ACKed.
1729 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1733 ((nonagle & TCP_NAGLE_CORK) ||
1734 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1737 /* Return how many segs we'd like on a TSO packet,
1738 * to send one TSO packet per ms
1740 static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1745 bytes = min_t(unsigned long,
1746 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1747 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1749 /* Goal is to send at least one packet per ms,
1750 * not one big TSO packet every 100 ms.
1751 * This preserves ACK clocking and is consistent
1752 * with tcp_tso_should_defer() heuristic.
1754 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1759 /* Return the number of segments we want in the skb we are transmitting.
1760 * See if congestion control module wants to decide; otherwise, autosize.
1762 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1764 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1765 u32 min_tso, tso_segs;
1767 min_tso = ca_ops->min_tso_segs ?
1768 ca_ops->min_tso_segs(sk) :
1769 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1771 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1772 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1775 /* Returns the portion of skb which can be sent right away */
1776 static unsigned int tcp_mss_split_point(const struct sock *sk,
1777 const struct sk_buff *skb,
1778 unsigned int mss_now,
1779 unsigned int max_segs,
1782 const struct tcp_sock *tp = tcp_sk(sk);
1783 u32 partial, needed, window, max_len;
1785 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1786 max_len = mss_now * max_segs;
1788 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1791 needed = min(skb->len, window);
1793 if (max_len <= needed)
1796 partial = needed % mss_now;
1797 /* If last segment is not a full MSS, check if Nagle rules allow us
1798 * to include this last segment in this skb.
1799 * Otherwise, we'll split the skb at last MSS boundary
1801 if (tcp_nagle_check(partial != 0, tp, nonagle))
1802 return needed - partial;
1807 /* Can at least one segment of SKB be sent right now, according to the
1808 * congestion window rules? If so, return how many segments are allowed.
1810 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1811 const struct sk_buff *skb)
1813 u32 in_flight, cwnd, halfcwnd;
1815 /* Don't be strict about the congestion window for the final FIN. */
1816 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1817 tcp_skb_pcount(skb) == 1)
1820 in_flight = tcp_packets_in_flight(tp);
1821 cwnd = tp->snd_cwnd;
1822 if (in_flight >= cwnd)
1825 /* For better scheduling, ensure we have at least
1826 * 2 GSO packets in flight.
1828 halfcwnd = max(cwnd >> 1, 1U);
1829 return min(halfcwnd, cwnd - in_flight);
1832 /* Initialize TSO state of a skb.
1833 * This must be invoked the first time we consider transmitting
1834 * SKB onto the wire.
1836 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1838 int tso_segs = tcp_skb_pcount(skb);
1840 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1841 tcp_set_skb_tso_segs(skb, mss_now);
1842 tso_segs = tcp_skb_pcount(skb);
1848 /* Return true if the Nagle test allows this packet to be
1851 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1852 unsigned int cur_mss, int nonagle)
1854 /* Nagle rule does not apply to frames, which sit in the middle of the
1855 * write_queue (they have no chances to get new data).
1857 * This is implemented in the callers, where they modify the 'nonagle'
1858 * argument based upon the location of SKB in the send queue.
1860 if (nonagle & TCP_NAGLE_PUSH)
1863 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1864 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1867 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1873 /* Does at least the first segment of SKB fit into the send window? */
1874 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1875 const struct sk_buff *skb,
1876 unsigned int cur_mss)
1878 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1880 if (skb->len > cur_mss)
1881 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1883 return !after(end_seq, tcp_wnd_end(tp));
1886 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1887 * which is put after SKB on the list. It is very much like
1888 * tcp_fragment() except that it may make several kinds of assumptions
1889 * in order to speed up the splitting operation. In particular, we
1890 * know that all the data is in scatter-gather pages, and that the
1891 * packet has never been sent out before (and thus is not cloned).
1893 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1894 unsigned int mss_now, gfp_t gfp)
1896 int nlen = skb->len - len;
1897 struct sk_buff *buff;
1900 /* All of a TSO frame must be composed of paged data. */
1901 if (skb->len != skb->data_len)
1902 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1903 skb, len, mss_now, gfp);
1905 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1906 if (unlikely(!buff))
1908 skb_copy_decrypted(buff, skb);
1910 sk_wmem_queued_add(sk, buff->truesize);
1911 sk_mem_charge(sk, buff->truesize);
1912 buff->truesize += nlen;
1913 skb->truesize -= nlen;
1915 /* Correct the sequence numbers. */
1916 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1917 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1918 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1920 /* PSH and FIN should only be set in the second packet. */
1921 flags = TCP_SKB_CB(skb)->tcp_flags;
1922 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1923 TCP_SKB_CB(buff)->tcp_flags = flags;
1925 /* This packet was never sent out yet, so no SACK bits. */
1926 TCP_SKB_CB(buff)->sacked = 0;
1928 tcp_skb_fragment_eor(skb, buff);
1930 buff->ip_summed = CHECKSUM_PARTIAL;
1931 skb_split(skb, buff, len);
1932 tcp_fragment_tstamp(skb, buff);
1934 /* Fix up tso_factor for both original and new SKB. */
1935 tcp_set_skb_tso_segs(skb, mss_now);
1936 tcp_set_skb_tso_segs(buff, mss_now);
1938 /* Link BUFF into the send queue. */
1939 __skb_header_release(buff);
1940 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1945 /* Try to defer sending, if possible, in order to minimize the amount
1946 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1948 * This algorithm is from John Heffner.
1950 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1951 bool *is_cwnd_limited,
1952 bool *is_rwnd_limited,
1955 const struct inet_connection_sock *icsk = inet_csk(sk);
1956 u32 send_win, cong_win, limit, in_flight;
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 struct sk_buff *head;
1962 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1965 /* Avoid bursty behavior by allowing defer
1966 * only if the last write was recent (1 ms).
1967 * Note that tp->tcp_wstamp_ns can be in the future if we have
1968 * packets waiting in a qdisc or device for EDT delivery.
1970 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1974 in_flight = tcp_packets_in_flight(tp);
1976 BUG_ON(tcp_skb_pcount(skb) <= 1);
1977 BUG_ON(tp->snd_cwnd <= in_flight);
1979 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1981 /* From in_flight test above, we know that cwnd > in_flight. */
1982 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1984 limit = min(send_win, cong_win);
1986 /* If a full-sized TSO skb can be sent, do it. */
1987 if (limit >= max_segs * tp->mss_cache)
1990 /* Middle in queue won't get any more data, full sendable already? */
1991 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1994 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1996 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1998 /* If at least some fraction of a window is available,
2001 chunk /= win_divisor;
2005 /* Different approach, try not to defer past a single
2006 * ACK. Receiver should ACK every other full sized
2007 * frame, so if we have space for more than 3 frames
2010 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2014 /* TODO : use tsorted_sent_queue ? */
2015 head = tcp_rtx_queue_head(sk);
2018 delta = tp->tcp_clock_cache - head->tstamp;
2019 /* If next ACK is likely to come too late (half srtt), do not defer */
2020 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2023 /* Ok, it looks like it is advisable to defer.
2024 * Three cases are tracked :
2025 * 1) We are cwnd-limited
2026 * 2) We are rwnd-limited
2027 * 3) We are application limited.
2029 if (cong_win < send_win) {
2030 if (cong_win <= skb->len) {
2031 *is_cwnd_limited = true;
2035 if (send_win <= skb->len) {
2036 *is_rwnd_limited = true;
2041 /* If this packet won't get more data, do not wait. */
2042 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2043 TCP_SKB_CB(skb)->eor)
2052 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2054 struct inet_connection_sock *icsk = inet_csk(sk);
2055 struct tcp_sock *tp = tcp_sk(sk);
2056 struct net *net = sock_net(sk);
2060 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2061 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2062 if (unlikely(delta >= interval * HZ)) {
2063 int mss = tcp_current_mss(sk);
2065 /* Update current search range */
2066 icsk->icsk_mtup.probe_size = 0;
2067 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2068 sizeof(struct tcphdr) +
2069 icsk->icsk_af_ops->net_header_len;
2070 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2072 /* Update probe time stamp */
2073 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2077 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2079 struct sk_buff *skb, *next;
2081 skb = tcp_send_head(sk);
2082 tcp_for_write_queue_from_safe(skb, next, sk) {
2083 if (len <= skb->len)
2086 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2095 /* Create a new MTU probe if we are ready.
2096 * MTU probe is regularly attempting to increase the path MTU by
2097 * deliberately sending larger packets. This discovers routing
2098 * changes resulting in larger path MTUs.
2100 * Returns 0 if we should wait to probe (no cwnd available),
2101 * 1 if a probe was sent,
2104 static int tcp_mtu_probe(struct sock *sk)
2106 struct inet_connection_sock *icsk = inet_csk(sk);
2107 struct tcp_sock *tp = tcp_sk(sk);
2108 struct sk_buff *skb, *nskb, *next;
2109 struct net *net = sock_net(sk);
2116 /* Not currently probing/verifying,
2118 * have enough cwnd, and
2119 * not SACKing (the variable headers throw things off)
2121 if (likely(!icsk->icsk_mtup.enabled ||
2122 icsk->icsk_mtup.probe_size ||
2123 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2124 tp->snd_cwnd < 11 ||
2125 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2128 /* Use binary search for probe_size between tcp_mss_base,
2129 * and current mss_clamp. if (search_high - search_low)
2130 * smaller than a threshold, backoff from probing.
2132 mss_now = tcp_current_mss(sk);
2133 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2134 icsk->icsk_mtup.search_low) >> 1);
2135 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2136 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2137 /* When misfortune happens, we are reprobing actively,
2138 * and then reprobe timer has expired. We stick with current
2139 * probing process by not resetting search range to its orignal.
2141 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2142 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2143 /* Check whether enough time has elaplased for
2144 * another round of probing.
2146 tcp_mtu_check_reprobe(sk);
2150 /* Have enough data in the send queue to probe? */
2151 if (tp->write_seq - tp->snd_nxt < size_needed)
2154 if (tp->snd_wnd < size_needed)
2156 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2159 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2160 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2161 if (!tcp_packets_in_flight(tp))
2167 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2170 /* We're allowed to probe. Build it now. */
2171 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2174 sk_wmem_queued_add(sk, nskb->truesize);
2175 sk_mem_charge(sk, nskb->truesize);
2177 skb = tcp_send_head(sk);
2178 skb_copy_decrypted(nskb, skb);
2180 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2181 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2182 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2183 TCP_SKB_CB(nskb)->sacked = 0;
2185 nskb->ip_summed = CHECKSUM_PARTIAL;
2187 tcp_insert_write_queue_before(nskb, skb, sk);
2188 tcp_highest_sack_replace(sk, skb, nskb);
2191 tcp_for_write_queue_from_safe(skb, next, sk) {
2192 copy = min_t(int, skb->len, probe_size - len);
2193 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2195 if (skb->len <= copy) {
2196 /* We've eaten all the data from this skb.
2198 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2199 /* If this is the last SKB we copy and eor is set
2200 * we need to propagate it to the new skb.
2202 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2203 tcp_skb_collapse_tstamp(nskb, skb);
2204 tcp_unlink_write_queue(skb, sk);
2205 sk_wmem_free_skb(sk, skb);
2207 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2208 ~(TCPHDR_FIN|TCPHDR_PSH);
2209 if (!skb_shinfo(skb)->nr_frags) {
2210 skb_pull(skb, copy);
2212 __pskb_trim_head(skb, copy);
2213 tcp_set_skb_tso_segs(skb, mss_now);
2215 TCP_SKB_CB(skb)->seq += copy;
2220 if (len >= probe_size)
2223 tcp_init_tso_segs(nskb, nskb->len);
2225 /* We're ready to send. If this fails, the probe will
2226 * be resegmented into mss-sized pieces by tcp_write_xmit().
2228 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2229 /* Decrement cwnd here because we are sending
2230 * effectively two packets. */
2232 tcp_event_new_data_sent(sk, nskb);
2234 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2235 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2236 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2244 static bool tcp_pacing_check(struct sock *sk)
2246 struct tcp_sock *tp = tcp_sk(sk);
2248 if (!tcp_needs_internal_pacing(sk))
2251 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2254 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2255 hrtimer_start(&tp->pacing_timer,
2256 ns_to_ktime(tp->tcp_wstamp_ns),
2257 HRTIMER_MODE_ABS_PINNED_SOFT);
2263 /* TCP Small Queues :
2264 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2265 * (These limits are doubled for retransmits)
2267 * - better RTT estimation and ACK scheduling
2270 * Alas, some drivers / subsystems require a fair amount
2271 * of queued bytes to ensure line rate.
2272 * One example is wifi aggregation (802.11 AMPDU)
2274 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2275 unsigned int factor)
2277 unsigned long limit;
2279 limit = max_t(unsigned long,
2281 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2282 if (sk->sk_pacing_status == SK_PACING_NONE)
2283 limit = min_t(unsigned long, limit,
2284 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2287 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2288 tcp_sk(sk)->tcp_tx_delay) {
2289 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2291 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2292 * approximate our needs assuming an ~100% skb->truesize overhead.
2293 * USEC_PER_SEC is approximated by 2^20.
2294 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2296 extra_bytes >>= (20 - 1);
2297 limit += extra_bytes;
2299 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2300 /* Always send skb if rtx queue is empty.
2301 * No need to wait for TX completion to call us back,
2302 * after softirq/tasklet schedule.
2303 * This helps when TX completions are delayed too much.
2305 if (tcp_rtx_queue_empty(sk))
2308 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2309 /* It is possible TX completion already happened
2310 * before we set TSQ_THROTTLED, so we must
2311 * test again the condition.
2313 smp_mb__after_atomic();
2314 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2320 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2322 const u32 now = tcp_jiffies32;
2323 enum tcp_chrono old = tp->chrono_type;
2325 if (old > TCP_CHRONO_UNSPEC)
2326 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2327 tp->chrono_start = now;
2328 tp->chrono_type = new;
2331 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2333 struct tcp_sock *tp = tcp_sk(sk);
2335 /* If there are multiple conditions worthy of tracking in a
2336 * chronograph then the highest priority enum takes precedence
2337 * over the other conditions. So that if something "more interesting"
2338 * starts happening, stop the previous chrono and start a new one.
2340 if (type > tp->chrono_type)
2341 tcp_chrono_set(tp, type);
2344 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2346 struct tcp_sock *tp = tcp_sk(sk);
2349 /* There are multiple conditions worthy of tracking in a
2350 * chronograph, so that the highest priority enum takes
2351 * precedence over the other conditions (see tcp_chrono_start).
2352 * If a condition stops, we only stop chrono tracking if
2353 * it's the "most interesting" or current chrono we are
2354 * tracking and starts busy chrono if we have pending data.
2356 if (tcp_rtx_and_write_queues_empty(sk))
2357 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2358 else if (type == tp->chrono_type)
2359 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2362 /* This routine writes packets to the network. It advances the
2363 * send_head. This happens as incoming acks open up the remote
2366 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2367 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2368 * account rare use of URG, this is not a big flaw.
2370 * Send at most one packet when push_one > 0. Temporarily ignore
2371 * cwnd limit to force at most one packet out when push_one == 2.
2373 * Returns true, if no segments are in flight and we have queued segments,
2374 * but cannot send anything now because of SWS or another problem.
2376 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2377 int push_one, gfp_t gfp)
2379 struct tcp_sock *tp = tcp_sk(sk);
2380 struct sk_buff *skb;
2381 unsigned int tso_segs, sent_pkts;
2384 bool is_cwnd_limited = false, is_rwnd_limited = false;
2389 tcp_mstamp_refresh(tp);
2391 /* Do MTU probing. */
2392 result = tcp_mtu_probe(sk);
2395 } else if (result > 0) {
2400 max_segs = tcp_tso_segs(sk, mss_now);
2401 while ((skb = tcp_send_head(sk))) {
2404 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2405 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2406 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2407 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2408 tcp_init_tso_segs(skb, mss_now);
2409 goto repair; /* Skip network transmission */
2412 if (tcp_pacing_check(sk))
2415 tso_segs = tcp_init_tso_segs(skb, mss_now);
2418 cwnd_quota = tcp_cwnd_test(tp, skb);
2421 /* Force out a loss probe pkt. */
2427 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2428 is_rwnd_limited = true;
2432 if (tso_segs == 1) {
2433 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2434 (tcp_skb_is_last(sk, skb) ?
2435 nonagle : TCP_NAGLE_PUSH))))
2439 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2440 &is_rwnd_limited, max_segs))
2445 if (tso_segs > 1 && !tcp_urg_mode(tp))
2446 limit = tcp_mss_split_point(sk, skb, mss_now,
2452 if (skb->len > limit &&
2453 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2456 if (tcp_small_queue_check(sk, skb, 0))
2459 /* Argh, we hit an empty skb(), presumably a thread
2460 * is sleeping in sendmsg()/sk_stream_wait_memory().
2461 * We do not want to send a pure-ack packet and have
2462 * a strange looking rtx queue with empty packet(s).
2464 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2467 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2471 /* Advance the send_head. This one is sent out.
2472 * This call will increment packets_out.
2474 tcp_event_new_data_sent(sk, skb);
2476 tcp_minshall_update(tp, mss_now, skb);
2477 sent_pkts += tcp_skb_pcount(skb);
2483 if (is_rwnd_limited)
2484 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2486 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2488 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2489 if (likely(sent_pkts || is_cwnd_limited))
2490 tcp_cwnd_validate(sk, is_cwnd_limited);
2492 if (likely(sent_pkts)) {
2493 if (tcp_in_cwnd_reduction(sk))
2494 tp->prr_out += sent_pkts;
2496 /* Send one loss probe per tail loss episode. */
2498 tcp_schedule_loss_probe(sk, false);
2501 return !tp->packets_out && !tcp_write_queue_empty(sk);
2504 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2506 struct inet_connection_sock *icsk = inet_csk(sk);
2507 struct tcp_sock *tp = tcp_sk(sk);
2508 u32 timeout, rto_delta_us;
2511 /* Don't do any loss probe on a Fast Open connection before 3WHS
2514 if (rcu_access_pointer(tp->fastopen_rsk))
2517 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2518 /* Schedule a loss probe in 2*RTT for SACK capable connections
2519 * not in loss recovery, that are either limited by cwnd or application.
2521 if ((early_retrans != 3 && early_retrans != 4) ||
2522 !tp->packets_out || !tcp_is_sack(tp) ||
2523 (icsk->icsk_ca_state != TCP_CA_Open &&
2524 icsk->icsk_ca_state != TCP_CA_CWR))
2527 /* Probe timeout is 2*rtt. Add minimum RTO to account
2528 * for delayed ack when there's one outstanding packet. If no RTT
2529 * sample is available then probe after TCP_TIMEOUT_INIT.
2532 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2533 if (tp->packets_out == 1)
2534 timeout += TCP_RTO_MIN;
2536 timeout += TCP_TIMEOUT_MIN;
2538 timeout = TCP_TIMEOUT_INIT;
2541 /* If the RTO formula yields an earlier time, then use that time. */
2542 rto_delta_us = advancing_rto ?
2543 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2544 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2545 if (rto_delta_us > 0)
2546 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2548 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2553 /* Thanks to skb fast clones, we can detect if a prior transmit of
2554 * a packet is still in a qdisc or driver queue.
2555 * In this case, there is very little point doing a retransmit !
2557 static bool skb_still_in_host_queue(const struct sock *sk,
2558 const struct sk_buff *skb)
2560 if (unlikely(skb_fclone_busy(sk, skb))) {
2561 NET_INC_STATS(sock_net(sk),
2562 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2568 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2569 * retransmit the last segment.
2571 void tcp_send_loss_probe(struct sock *sk)
2573 struct tcp_sock *tp = tcp_sk(sk);
2574 struct sk_buff *skb;
2576 int mss = tcp_current_mss(sk);
2578 /* At most one outstanding TLP */
2579 if (tp->tlp_high_seq)
2582 tp->tlp_retrans = 0;
2583 skb = tcp_send_head(sk);
2584 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2585 pcount = tp->packets_out;
2586 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2587 if (tp->packets_out > pcount)
2591 skb = skb_rb_last(&sk->tcp_rtx_queue);
2592 if (unlikely(!skb)) {
2593 WARN_ONCE(tp->packets_out,
2594 "invalid inflight: %u state %u cwnd %u mss %d\n",
2595 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2596 inet_csk(sk)->icsk_pending = 0;
2600 if (skb_still_in_host_queue(sk, skb))
2603 pcount = tcp_skb_pcount(skb);
2604 if (WARN_ON(!pcount))
2607 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2608 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2609 (pcount - 1) * mss, mss,
2612 skb = skb_rb_next(skb);
2615 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2618 if (__tcp_retransmit_skb(sk, skb, 1))
2621 tp->tlp_retrans = 1;
2624 /* Record snd_nxt for loss detection. */
2625 tp->tlp_high_seq = tp->snd_nxt;
2627 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2628 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2629 inet_csk(sk)->icsk_pending = 0;
2634 /* Push out any pending frames which were held back due to
2635 * TCP_CORK or attempt at coalescing tiny packets.
2636 * The socket must be locked by the caller.
2638 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2641 /* If we are closed, the bytes will have to remain here.
2642 * In time closedown will finish, we empty the write queue and
2643 * all will be happy.
2645 if (unlikely(sk->sk_state == TCP_CLOSE))
2648 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2649 sk_gfp_mask(sk, GFP_ATOMIC)))
2650 tcp_check_probe_timer(sk);
2653 /* Send _single_ skb sitting at the send head. This function requires
2654 * true push pending frames to setup probe timer etc.
2656 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2658 struct sk_buff *skb = tcp_send_head(sk);
2660 BUG_ON(!skb || skb->len < mss_now);
2662 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2665 /* This function returns the amount that we can raise the
2666 * usable window based on the following constraints
2668 * 1. The window can never be shrunk once it is offered (RFC 793)
2669 * 2. We limit memory per socket
2672 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2673 * RECV.NEXT + RCV.WIN fixed until:
2674 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2676 * i.e. don't raise the right edge of the window until you can raise
2677 * it at least MSS bytes.
2679 * Unfortunately, the recommended algorithm breaks header prediction,
2680 * since header prediction assumes th->window stays fixed.
2682 * Strictly speaking, keeping th->window fixed violates the receiver
2683 * side SWS prevention criteria. The problem is that under this rule
2684 * a stream of single byte packets will cause the right side of the
2685 * window to always advance by a single byte.
2687 * Of course, if the sender implements sender side SWS prevention
2688 * then this will not be a problem.
2690 * BSD seems to make the following compromise:
2692 * If the free space is less than the 1/4 of the maximum
2693 * space available and the free space is less than 1/2 mss,
2694 * then set the window to 0.
2695 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2696 * Otherwise, just prevent the window from shrinking
2697 * and from being larger than the largest representable value.
2699 * This prevents incremental opening of the window in the regime
2700 * where TCP is limited by the speed of the reader side taking
2701 * data out of the TCP receive queue. It does nothing about
2702 * those cases where the window is constrained on the sender side
2703 * because the pipeline is full.
2705 * BSD also seems to "accidentally" limit itself to windows that are a
2706 * multiple of MSS, at least until the free space gets quite small.
2707 * This would appear to be a side effect of the mbuf implementation.
2708 * Combining these two algorithms results in the observed behavior
2709 * of having a fixed window size at almost all times.
2711 * Below we obtain similar behavior by forcing the offered window to
2712 * a multiple of the mss when it is feasible to do so.
2714 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2715 * Regular options like TIMESTAMP are taken into account.
2717 u32 __tcp_select_window(struct sock *sk)
2719 struct inet_connection_sock *icsk = inet_csk(sk);
2720 struct tcp_sock *tp = tcp_sk(sk);
2721 /* MSS for the peer's data. Previous versions used mss_clamp
2722 * here. I don't know if the value based on our guesses
2723 * of peer's MSS is better for the performance. It's more correct
2724 * but may be worse for the performance because of rcv_mss
2725 * fluctuations. --SAW 1998/11/1
2727 int mss = icsk->icsk_ack.rcv_mss;
2728 int free_space = tcp_space(sk);
2729 int allowed_space = tcp_full_space(sk);
2730 int full_space = min_t(int, tp->window_clamp, allowed_space);
2733 if (unlikely(mss > full_space)) {
2738 if (free_space < (full_space >> 1)) {
2739 icsk->icsk_ack.quick = 0;
2741 if (tcp_under_memory_pressure(sk))
2742 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2745 /* free_space might become our new window, make sure we don't
2746 * increase it due to wscale.
2748 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2750 /* if free space is less than mss estimate, or is below 1/16th
2751 * of the maximum allowed, try to move to zero-window, else
2752 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2753 * new incoming data is dropped due to memory limits.
2754 * With large window, mss test triggers way too late in order
2755 * to announce zero window in time before rmem limit kicks in.
2757 if (free_space < (allowed_space >> 4) || free_space < mss)
2761 if (free_space > tp->rcv_ssthresh)
2762 free_space = tp->rcv_ssthresh;
2764 /* Don't do rounding if we are using window scaling, since the
2765 * scaled window will not line up with the MSS boundary anyway.
2767 if (tp->rx_opt.rcv_wscale) {
2768 window = free_space;
2770 /* Advertise enough space so that it won't get scaled away.
2771 * Import case: prevent zero window announcement if
2772 * 1<<rcv_wscale > mss.
2774 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2776 window = tp->rcv_wnd;
2777 /* Get the largest window that is a nice multiple of mss.
2778 * Window clamp already applied above.
2779 * If our current window offering is within 1 mss of the
2780 * free space we just keep it. This prevents the divide
2781 * and multiply from happening most of the time.
2782 * We also don't do any window rounding when the free space
2785 if (window <= free_space - mss || window > free_space)
2786 window = rounddown(free_space, mss);
2787 else if (mss == full_space &&
2788 free_space > window + (full_space >> 1))
2789 window = free_space;
2795 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2796 const struct sk_buff *next_skb)
2798 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2799 const struct skb_shared_info *next_shinfo =
2800 skb_shinfo(next_skb);
2801 struct skb_shared_info *shinfo = skb_shinfo(skb);
2803 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2804 shinfo->tskey = next_shinfo->tskey;
2805 TCP_SKB_CB(skb)->txstamp_ack |=
2806 TCP_SKB_CB(next_skb)->txstamp_ack;
2810 /* Collapses two adjacent SKB's during retransmission. */
2811 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2813 struct tcp_sock *tp = tcp_sk(sk);
2814 struct sk_buff *next_skb = skb_rb_next(skb);
2817 next_skb_size = next_skb->len;
2819 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2821 if (next_skb_size) {
2822 if (next_skb_size <= skb_availroom(skb))
2823 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2825 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2828 tcp_highest_sack_replace(sk, next_skb, skb);
2830 /* Update sequence range on original skb. */
2831 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2833 /* Merge over control information. This moves PSH/FIN etc. over */
2834 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2836 /* All done, get rid of second SKB and account for it so
2837 * packet counting does not break.
2839 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2840 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2842 /* changed transmit queue under us so clear hints */
2843 tcp_clear_retrans_hints_partial(tp);
2844 if (next_skb == tp->retransmit_skb_hint)
2845 tp->retransmit_skb_hint = skb;
2847 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2849 tcp_skb_collapse_tstamp(skb, next_skb);
2851 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2855 /* Check if coalescing SKBs is legal. */
2856 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2858 if (tcp_skb_pcount(skb) > 1)
2860 if (skb_cloned(skb))
2862 /* Some heuristics for collapsing over SACK'd could be invented */
2863 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2869 /* Collapse packets in the retransmit queue to make to create
2870 * less packets on the wire. This is only done on retransmission.
2872 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2875 struct tcp_sock *tp = tcp_sk(sk);
2876 struct sk_buff *skb = to, *tmp;
2879 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
2881 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2884 skb_rbtree_walk_from_safe(skb, tmp) {
2885 if (!tcp_can_collapse(sk, skb))
2888 if (!tcp_skb_can_collapse_to(to))
2901 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2904 if (!tcp_collapse_retrans(sk, to))
2909 /* This retransmits one SKB. Policy decisions and retransmit queue
2910 * state updates are done by the caller. Returns non-zero if an
2911 * error occurred which prevented the send.
2913 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2915 struct inet_connection_sock *icsk = inet_csk(sk);
2916 struct tcp_sock *tp = tcp_sk(sk);
2917 unsigned int cur_mss;
2921 /* Inconclusive MTU probe */
2922 if (icsk->icsk_mtup.probe_size)
2923 icsk->icsk_mtup.probe_size = 0;
2925 /* Do not sent more than we queued. 1/4 is reserved for possible
2926 * copying overhead: fragmentation, tunneling, mangling etc.
2928 if (refcount_read(&sk->sk_wmem_alloc) >
2929 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2933 if (skb_still_in_host_queue(sk, skb))
2936 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2937 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2941 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2945 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2946 return -EHOSTUNREACH; /* Routing failure or similar. */
2948 cur_mss = tcp_current_mss(sk);
2949 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2951 /* If receiver has shrunk his window, and skb is out of
2952 * new window, do not retransmit it. The exception is the
2953 * case, when window is shrunk to zero. In this case
2954 * our retransmit of one segment serves as a zero window probe.
2956 if (avail_wnd <= 0) {
2957 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
2959 avail_wnd = cur_mss;
2962 len = cur_mss * segs;
2963 if (len > avail_wnd) {
2964 len = rounddown(avail_wnd, cur_mss);
2968 if (skb->len > len) {
2969 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2970 cur_mss, GFP_ATOMIC))
2971 return -ENOMEM; /* We'll try again later. */
2973 if (skb_unclone(skb, GFP_ATOMIC))
2976 diff = tcp_skb_pcount(skb);
2977 tcp_set_skb_tso_segs(skb, cur_mss);
2978 diff -= tcp_skb_pcount(skb);
2980 tcp_adjust_pcount(sk, skb, diff);
2981 avail_wnd = min_t(int, avail_wnd, cur_mss);
2982 if (skb->len < avail_wnd)
2983 tcp_retrans_try_collapse(sk, skb, avail_wnd);
2986 /* RFC3168, section 6.1.1.1. ECN fallback */
2987 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2988 tcp_ecn_clear_syn(sk, skb);
2990 /* Update global and local TCP statistics. */
2991 segs = tcp_skb_pcount(skb);
2992 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2993 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2994 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2995 tp->total_retrans += segs;
2996 tp->bytes_retrans += skb->len;
2998 /* make sure skb->data is aligned on arches that require it
2999 * and check if ack-trimming & collapsing extended the headroom
3000 * beyond what csum_start can cover.
3002 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3003 skb_headroom(skb) >= 0xFFFF)) {
3004 struct sk_buff *nskb;
3006 tcp_skb_tsorted_save(skb) {
3007 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3010 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3014 } tcp_skb_tsorted_restore(skb);
3017 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3018 tcp_rate_skb_sent(sk, skb);
3021 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3024 /* To avoid taking spuriously low RTT samples based on a timestamp
3025 * for a transmit that never happened, always mark EVER_RETRANS
3027 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3029 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3030 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3031 TCP_SKB_CB(skb)->seq, segs, err);
3034 trace_tcp_retransmit_skb(sk, skb);
3035 } else if (err != -EBUSY) {
3036 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3041 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3043 struct tcp_sock *tp = tcp_sk(sk);
3044 int err = __tcp_retransmit_skb(sk, skb, segs);
3047 #if FASTRETRANS_DEBUG > 0
3048 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3049 net_dbg_ratelimited("retrans_out leaked\n");
3052 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3053 tp->retrans_out += tcp_skb_pcount(skb);
3056 /* Save stamp of the first (attempted) retransmit. */
3057 if (!tp->retrans_stamp)
3058 tp->retrans_stamp = tcp_skb_timestamp(skb);
3060 if (tp->undo_retrans < 0)
3061 tp->undo_retrans = 0;
3062 tp->undo_retrans += tcp_skb_pcount(skb);
3066 /* This gets called after a retransmit timeout, and the initially
3067 * retransmitted data is acknowledged. It tries to continue
3068 * resending the rest of the retransmit queue, until either
3069 * we've sent it all or the congestion window limit is reached.
3071 void tcp_xmit_retransmit_queue(struct sock *sk)
3073 const struct inet_connection_sock *icsk = inet_csk(sk);
3074 struct sk_buff *skb, *rtx_head, *hole = NULL;
3075 struct tcp_sock *tp = tcp_sk(sk);
3079 if (!tp->packets_out)
3082 rtx_head = tcp_rtx_queue_head(sk);
3083 skb = tp->retransmit_skb_hint ?: rtx_head;
3084 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3085 skb_rbtree_walk_from(skb) {
3089 if (tcp_pacing_check(sk))
3092 /* we could do better than to assign each time */
3094 tp->retransmit_skb_hint = skb;
3096 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3099 sacked = TCP_SKB_CB(skb)->sacked;
3100 /* In case tcp_shift_skb_data() have aggregated large skbs,
3101 * we need to make sure not sending too bigs TSO packets
3103 segs = min_t(int, segs, max_segs);
3105 if (tp->retrans_out >= tp->lost_out) {
3107 } else if (!(sacked & TCPCB_LOST)) {
3108 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3113 if (icsk->icsk_ca_state != TCP_CA_Loss)
3114 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3116 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3119 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3122 if (tcp_small_queue_check(sk, skb, 1))
3125 if (tcp_retransmit_skb(sk, skb, segs))
3128 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3130 if (tcp_in_cwnd_reduction(sk))
3131 tp->prr_out += tcp_skb_pcount(skb);
3133 if (skb == rtx_head &&
3134 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3135 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3136 inet_csk(sk)->icsk_rto,
3142 /* We allow to exceed memory limits for FIN packets to expedite
3143 * connection tear down and (memory) recovery.
3144 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3145 * or even be forced to close flow without any FIN.
3146 * In general, we want to allow one skb per socket to avoid hangs
3147 * with edge trigger epoll()
3149 void sk_forced_mem_schedule(struct sock *sk, int size)
3153 delta = size - sk->sk_forward_alloc;
3156 amt = sk_mem_pages(delta);
3157 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3158 sk_memory_allocated_add(sk, amt);
3160 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3161 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3164 /* Send a FIN. The caller locks the socket for us.
3165 * We should try to send a FIN packet really hard, but eventually give up.
3167 void tcp_send_fin(struct sock *sk)
3169 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3170 struct tcp_sock *tp = tcp_sk(sk);
3172 /* Optimization, tack on the FIN if we have one skb in write queue and
3173 * this skb was not yet sent, or we are under memory pressure.
3174 * Note: in the latter case, FIN packet will be sent after a timeout,
3175 * as TCP stack thinks it has already been transmitted.
3177 if (!tskb && tcp_under_memory_pressure(sk))
3178 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3181 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3182 TCP_SKB_CB(tskb)->end_seq++;
3184 if (tcp_write_queue_empty(sk)) {
3185 /* This means tskb was already sent.
3186 * Pretend we included the FIN on previous transmit.
3187 * We need to set tp->snd_nxt to the value it would have
3188 * if FIN had been sent. This is because retransmit path
3189 * does not change tp->snd_nxt.
3191 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3195 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3199 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3200 skb_reserve(skb, MAX_TCP_HEADER);
3201 sk_forced_mem_schedule(sk, skb->truesize);
3202 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3203 tcp_init_nondata_skb(skb, tp->write_seq,
3204 TCPHDR_ACK | TCPHDR_FIN);
3205 tcp_queue_skb(sk, skb);
3207 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3210 /* We get here when a process closes a file descriptor (either due to
3211 * an explicit close() or as a byproduct of exit()'ing) and there
3212 * was unread data in the receive queue. This behavior is recommended
3213 * by RFC 2525, section 2.17. -DaveM
3215 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3217 struct sk_buff *skb;
3219 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3221 /* NOTE: No TCP options attached and we never retransmit this. */
3222 skb = alloc_skb(MAX_TCP_HEADER, priority);
3224 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3228 /* Reserve space for headers and prepare control bits. */
3229 skb_reserve(skb, MAX_TCP_HEADER);
3230 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3231 TCPHDR_ACK | TCPHDR_RST);
3232 tcp_mstamp_refresh(tcp_sk(sk));
3234 if (tcp_transmit_skb(sk, skb, 0, priority))
3235 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3237 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3238 * skb here is different to the troublesome skb, so use NULL
3240 trace_tcp_send_reset(sk, NULL);
3243 /* Send a crossed SYN-ACK during socket establishment.
3244 * WARNING: This routine must only be called when we have already sent
3245 * a SYN packet that crossed the incoming SYN that caused this routine
3246 * to get called. If this assumption fails then the initial rcv_wnd
3247 * and rcv_wscale values will not be correct.
3249 int tcp_send_synack(struct sock *sk)
3251 struct sk_buff *skb;
3253 skb = tcp_rtx_queue_head(sk);
3254 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3255 pr_err("%s: wrong queue state\n", __func__);
3258 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3259 if (skb_cloned(skb)) {
3260 struct sk_buff *nskb;
3262 tcp_skb_tsorted_save(skb) {
3263 nskb = skb_copy(skb, GFP_ATOMIC);
3264 } tcp_skb_tsorted_restore(skb);
3267 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3268 tcp_highest_sack_replace(sk, skb, nskb);
3269 tcp_rtx_queue_unlink_and_free(skb, sk);
3270 __skb_header_release(nskb);
3271 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3272 sk_wmem_queued_add(sk, nskb->truesize);
3273 sk_mem_charge(sk, nskb->truesize);
3277 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3278 tcp_ecn_send_synack(sk, skb);
3280 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3284 * tcp_make_synack - Prepare a SYN-ACK.
3285 * sk: listener socket
3286 * dst: dst entry attached to the SYNACK
3287 * req: request_sock pointer
3289 * Allocate one skb and build a SYNACK packet.
3290 * @dst is consumed : Caller should not use it again.
3292 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3293 struct request_sock *req,
3294 struct tcp_fastopen_cookie *foc,
3295 enum tcp_synack_type synack_type)
3297 struct inet_request_sock *ireq = inet_rsk(req);
3298 const struct tcp_sock *tp = tcp_sk(sk);
3299 struct tcp_md5sig_key *md5 = NULL;
3300 struct tcp_out_options opts;
3301 struct sk_buff *skb;
3302 int tcp_header_size;
3307 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3308 if (unlikely(!skb)) {
3312 /* Reserve space for headers. */
3313 skb_reserve(skb, MAX_TCP_HEADER);
3315 switch (synack_type) {
3316 case TCP_SYNACK_NORMAL:
3317 skb_set_owner_w(skb, req_to_sk(req));
3319 case TCP_SYNACK_COOKIE:
3320 /* Under synflood, we do not attach skb to a socket,
3321 * to avoid false sharing.
3324 case TCP_SYNACK_FASTOPEN:
3325 /* sk is a const pointer, because we want to express multiple
3326 * cpu might call us concurrently.
3327 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3329 skb_set_owner_w(skb, (struct sock *)sk);
3332 skb_dst_set(skb, dst);
3334 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3336 memset(&opts, 0, sizeof(opts));
3337 now = tcp_clock_ns();
3338 #ifdef CONFIG_SYN_COOKIES
3339 if (unlikely(req->cookie_ts))
3340 skb->skb_mstamp_ns = cookie_init_timestamp(req);
3344 skb->skb_mstamp_ns = now;
3345 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3346 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3349 #ifdef CONFIG_TCP_MD5SIG
3351 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3353 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3354 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3355 foc, synack_type) + sizeof(*th);
3357 skb_push(skb, tcp_header_size);
3358 skb_reset_transport_header(skb);
3360 th = (struct tcphdr *)skb->data;
3361 memset(th, 0, sizeof(struct tcphdr));
3364 tcp_ecn_make_synack(req, th);
3365 th->source = htons(ireq->ir_num);
3366 th->dest = ireq->ir_rmt_port;
3367 skb->mark = ireq->ir_mark;
3368 skb->ip_summed = CHECKSUM_PARTIAL;
3369 th->seq = htonl(tcp_rsk(req)->snt_isn);
3370 /* XXX data is queued and acked as is. No buffer/window check */
3371 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3373 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3374 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3375 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3376 th->doff = (tcp_header_size >> 2);
3377 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3379 #ifdef CONFIG_TCP_MD5SIG
3380 /* Okay, we have all we need - do the md5 hash if needed */
3382 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3383 md5, req_to_sk(req), skb);
3387 skb->skb_mstamp_ns = now;
3388 tcp_add_tx_delay(skb, tp);
3392 EXPORT_SYMBOL(tcp_make_synack);
3394 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3396 struct inet_connection_sock *icsk = inet_csk(sk);
3397 const struct tcp_congestion_ops *ca;
3398 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3400 if (ca_key == TCP_CA_UNSPEC)
3404 ca = tcp_ca_find_key(ca_key);
3405 if (likely(ca && try_module_get(ca->owner))) {
3406 module_put(icsk->icsk_ca_ops->owner);
3407 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3408 icsk->icsk_ca_ops = ca;
3413 /* Do all connect socket setups that can be done AF independent. */
3414 static void tcp_connect_init(struct sock *sk)
3416 const struct dst_entry *dst = __sk_dst_get(sk);
3417 struct tcp_sock *tp = tcp_sk(sk);
3421 /* We'll fix this up when we get a response from the other end.
3422 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3424 tp->tcp_header_len = sizeof(struct tcphdr);
3425 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3426 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3428 #ifdef CONFIG_TCP_MD5SIG
3429 if (tp->af_specific->md5_lookup(sk, sk))
3430 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3433 /* If user gave his TCP_MAXSEG, record it to clamp */
3434 if (tp->rx_opt.user_mss)
3435 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3438 tcp_sync_mss(sk, dst_mtu(dst));
3440 tcp_ca_dst_init(sk, dst);
3442 if (!tp->window_clamp)
3443 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3444 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3446 tcp_initialize_rcv_mss(sk);
3448 /* limit the window selection if the user enforce a smaller rx buffer */
3449 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3450 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3451 tp->window_clamp = tcp_full_space(sk);
3453 rcv_wnd = tcp_rwnd_init_bpf(sk);
3455 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3457 tcp_select_initial_window(sk, tcp_full_space(sk),
3458 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3461 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3465 tp->rx_opt.rcv_wscale = rcv_wscale;
3466 tp->rcv_ssthresh = tp->rcv_wnd;
3469 sock_reset_flag(sk, SOCK_DONE);
3472 tcp_write_queue_purge(sk);
3473 tp->snd_una = tp->write_seq;
3474 tp->snd_sml = tp->write_seq;
3475 tp->snd_up = tp->write_seq;
3476 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3478 if (likely(!tp->repair))
3481 tp->rcv_tstamp = tcp_jiffies32;
3482 tp->rcv_wup = tp->rcv_nxt;
3483 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3485 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3486 inet_csk(sk)->icsk_retransmits = 0;
3487 tcp_clear_retrans(tp);
3490 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3492 struct tcp_sock *tp = tcp_sk(sk);
3493 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3495 tcb->end_seq += skb->len;
3496 __skb_header_release(skb);
3497 sk_wmem_queued_add(sk, skb->truesize);
3498 sk_mem_charge(sk, skb->truesize);
3499 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3500 tp->packets_out += tcp_skb_pcount(skb);
3503 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3504 * queue a data-only packet after the regular SYN, such that regular SYNs
3505 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3506 * only the SYN sequence, the data are retransmitted in the first ACK.
3507 * If cookie is not cached or other error occurs, falls back to send a
3508 * regular SYN with Fast Open cookie request option.
3510 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3512 struct inet_connection_sock *icsk = inet_csk(sk);
3513 struct tcp_sock *tp = tcp_sk(sk);
3514 struct tcp_fastopen_request *fo = tp->fastopen_req;
3516 struct sk_buff *syn_data;
3518 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3519 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3522 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3523 * user-MSS. Reserve maximum option space for middleboxes that add
3524 * private TCP options. The cost is reduced data space in SYN :(
3526 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3527 /* Sync mss_cache after updating the mss_clamp */
3528 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3530 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3531 MAX_TCP_OPTION_SPACE;
3533 space = min_t(size_t, space, fo->size);
3535 /* limit to order-0 allocations */
3536 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3538 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3541 syn_data->ip_summed = CHECKSUM_PARTIAL;
3542 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3544 int copied = copy_from_iter(skb_put(syn_data, space), space,
3545 &fo->data->msg_iter);
3546 if (unlikely(!copied)) {
3547 tcp_skb_tsorted_anchor_cleanup(syn_data);
3548 kfree_skb(syn_data);
3551 if (copied != space) {
3552 skb_trim(syn_data, copied);
3555 skb_zcopy_set(syn_data, fo->uarg, NULL);
3557 /* No more data pending in inet_wait_for_connect() */
3558 if (space == fo->size)
3562 tcp_connect_queue_skb(sk, syn_data);
3564 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3566 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3568 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3570 /* Now full SYN+DATA was cloned and sent (or not),
3571 * remove the SYN from the original skb (syn_data)
3572 * we keep in write queue in case of a retransmit, as we
3573 * also have the SYN packet (with no data) in the same queue.
3575 TCP_SKB_CB(syn_data)->seq++;
3576 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3578 tp->syn_data = (fo->copied > 0);
3579 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3580 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3584 /* data was not sent, put it in write_queue */
3585 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3586 tp->packets_out -= tcp_skb_pcount(syn_data);
3589 /* Send a regular SYN with Fast Open cookie request option */
3590 if (fo->cookie.len > 0)
3592 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3594 tp->syn_fastopen = 0;
3596 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3600 /* Build a SYN and send it off. */
3601 int tcp_connect(struct sock *sk)
3603 struct tcp_sock *tp = tcp_sk(sk);
3604 struct sk_buff *buff;
3607 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3609 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3610 return -EHOSTUNREACH; /* Routing failure or similar. */
3612 tcp_connect_init(sk);
3614 if (unlikely(tp->repair)) {
3615 tcp_finish_connect(sk, NULL);
3619 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3620 if (unlikely(!buff))
3623 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3624 tcp_mstamp_refresh(tp);
3625 tp->retrans_stamp = tcp_time_stamp(tp);
3626 tcp_connect_queue_skb(sk, buff);
3627 tcp_ecn_send_syn(sk, buff);
3628 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3630 /* Send off SYN; include data in Fast Open. */
3631 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3632 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3633 if (err == -ECONNREFUSED)
3636 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3637 * in order to make this packet get counted in tcpOutSegs.
3639 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3640 tp->pushed_seq = tp->write_seq;
3641 buff = tcp_send_head(sk);
3642 if (unlikely(buff)) {
3643 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3644 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3646 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3648 /* Timer for repeating the SYN until an answer. */
3649 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3650 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3653 EXPORT_SYMBOL(tcp_connect);
3655 /* Send out a delayed ack, the caller does the policy checking
3656 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3659 void tcp_send_delayed_ack(struct sock *sk)
3661 struct inet_connection_sock *icsk = inet_csk(sk);
3662 int ato = icsk->icsk_ack.ato;
3663 unsigned long timeout;
3665 if (ato > TCP_DELACK_MIN) {
3666 const struct tcp_sock *tp = tcp_sk(sk);
3667 int max_ato = HZ / 2;
3669 if (inet_csk_in_pingpong_mode(sk) ||
3670 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3671 max_ato = TCP_DELACK_MAX;
3673 /* Slow path, intersegment interval is "high". */
3675 /* If some rtt estimate is known, use it to bound delayed ack.
3676 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3680 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3687 ato = min(ato, max_ato);
3690 /* Stay within the limit we were given */
3691 timeout = jiffies + ato;
3693 /* Use new timeout only if there wasn't a older one earlier. */
3694 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3695 /* If delack timer was blocked or is about to expire,
3698 if (icsk->icsk_ack.blocked ||
3699 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3704 if (!time_before(timeout, icsk->icsk_ack.timeout))
3705 timeout = icsk->icsk_ack.timeout;
3707 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3708 icsk->icsk_ack.timeout = timeout;
3709 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3712 /* This routine sends an ack and also updates the window. */
3713 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3715 struct sk_buff *buff;
3717 /* If we have been reset, we may not send again. */
3718 if (sk->sk_state == TCP_CLOSE)
3721 /* We are not putting this on the write queue, so
3722 * tcp_transmit_skb() will set the ownership to this
3725 buff = alloc_skb(MAX_TCP_HEADER,
3726 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3727 if (unlikely(!buff)) {
3728 inet_csk_schedule_ack(sk);
3729 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3730 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3731 TCP_DELACK_MAX, TCP_RTO_MAX);
3735 /* Reserve space for headers and prepare control bits. */
3736 skb_reserve(buff, MAX_TCP_HEADER);
3737 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3739 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3741 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3743 skb_set_tcp_pure_ack(buff);
3745 /* Send it off, this clears delayed acks for us. */
3746 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3748 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3750 void tcp_send_ack(struct sock *sk)
3752 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3755 /* This routine sends a packet with an out of date sequence
3756 * number. It assumes the other end will try to ack it.
3758 * Question: what should we make while urgent mode?
3759 * 4.4BSD forces sending single byte of data. We cannot send
3760 * out of window data, because we have SND.NXT==SND.MAX...
3762 * Current solution: to send TWO zero-length segments in urgent mode:
3763 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3764 * out-of-date with SND.UNA-1 to probe window.
3766 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3768 struct tcp_sock *tp = tcp_sk(sk);
3769 struct sk_buff *skb;
3771 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3772 skb = alloc_skb(MAX_TCP_HEADER,
3773 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3777 /* Reserve space for headers and set control bits. */
3778 skb_reserve(skb, MAX_TCP_HEADER);
3779 /* Use a previous sequence. This should cause the other
3780 * end to send an ack. Don't queue or clone SKB, just
3783 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3784 NET_INC_STATS(sock_net(sk), mib);
3785 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3788 /* Called from setsockopt( ... TCP_REPAIR ) */
3789 void tcp_send_window_probe(struct sock *sk)
3791 if (sk->sk_state == TCP_ESTABLISHED) {
3792 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3793 tcp_mstamp_refresh(tcp_sk(sk));
3794 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3798 /* Initiate keepalive or window probe from timer. */
3799 int tcp_write_wakeup(struct sock *sk, int mib)
3801 struct tcp_sock *tp = tcp_sk(sk);
3802 struct sk_buff *skb;
3804 if (sk->sk_state == TCP_CLOSE)
3807 skb = tcp_send_head(sk);
3808 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3810 unsigned int mss = tcp_current_mss(sk);
3811 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3813 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3814 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3816 /* We are probing the opening of a window
3817 * but the window size is != 0
3818 * must have been a result SWS avoidance ( sender )
3820 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3822 seg_size = min(seg_size, mss);
3823 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3824 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3825 skb, seg_size, mss, GFP_ATOMIC))
3827 } else if (!tcp_skb_pcount(skb))
3828 tcp_set_skb_tso_segs(skb, mss);
3830 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3831 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3833 tcp_event_new_data_sent(sk, skb);
3836 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3837 tcp_xmit_probe_skb(sk, 1, mib);
3838 return tcp_xmit_probe_skb(sk, 0, mib);
3842 /* A window probe timeout has occurred. If window is not closed send
3843 * a partial packet else a zero probe.
3845 void tcp_send_probe0(struct sock *sk)
3847 struct inet_connection_sock *icsk = inet_csk(sk);
3848 struct tcp_sock *tp = tcp_sk(sk);
3849 struct net *net = sock_net(sk);
3850 unsigned long timeout;
3853 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3855 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3856 /* Cancel probe timer, if it is not required. */
3857 icsk->icsk_probes_out = 0;
3858 icsk->icsk_backoff = 0;
3859 icsk->icsk_probes_tstamp = 0;
3863 icsk->icsk_probes_out++;
3865 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
3866 icsk->icsk_backoff++;
3867 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3869 /* If packet was not sent due to local congestion,
3870 * Let senders fight for local resources conservatively.
3872 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3875 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
3876 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3879 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3881 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3885 tcp_rsk(req)->txhash = net_tx_rndhash();
3886 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3888 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3889 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3890 if (unlikely(tcp_passive_fastopen(sk)))
3891 tcp_sk(sk)->total_retrans++;
3892 trace_tcp_retransmit_synack(sk, req);
3896 EXPORT_SYMBOL(tcp_rtx_synack);