2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
44 #include <linux/static_key.h>
46 #include <trace/events/tcp.h>
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
57 int push_one, gfp_t gfp);
59 /* Account for new data that has been sent to the network. */
60 static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
62 struct inet_connection_sock *icsk = inet_csk(sk);
63 struct tcp_sock *tp = tcp_sk(sk);
64 unsigned int prior_packets = tp->packets_out;
66 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
68 __skb_unlink(skb, &sk->sk_write_queue);
69 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
71 tp->packets_out += tcp_skb_pcount(skb);
72 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
75 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
79 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
80 * window scaling factor due to loss of precision.
81 * If window has been shrunk, what should we make? It is not clear at all.
82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
84 * invalid. OK, let's make this for now:
86 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
88 const struct tcp_sock *tp = tcp_sk(sk);
90 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
91 (tp->rx_opt.wscale_ok &&
92 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
95 return tcp_wnd_end(tp);
98 /* Calculate mss to advertise in SYN segment.
99 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
101 * 1. It is independent of path mtu.
102 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
103 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
104 * attached devices, because some buggy hosts are confused by
106 * 4. We do not make 3, we advertise MSS, calculated from first
107 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
108 * This may be overridden via information stored in routing table.
109 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
110 * probably even Jumbo".
112 static __u16 tcp_advertise_mss(struct sock *sk)
114 struct tcp_sock *tp = tcp_sk(sk);
115 const struct dst_entry *dst = __sk_dst_get(sk);
116 int mss = tp->advmss;
119 unsigned int metric = dst_metric_advmss(dst);
130 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
131 * This is the first part of cwnd validation mechanism.
133 void tcp_cwnd_restart(struct sock *sk, s32 delta)
135 struct tcp_sock *tp = tcp_sk(sk);
136 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
137 u32 cwnd = tp->snd_cwnd;
139 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
141 tp->snd_ssthresh = tcp_current_ssthresh(sk);
142 restart_cwnd = min(restart_cwnd, cwnd);
144 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
146 tp->snd_cwnd = max(cwnd, restart_cwnd);
147 tp->snd_cwnd_stamp = tcp_jiffies32;
148 tp->snd_cwnd_used = 0;
151 /* Congestion state accounting after a packet has been sent. */
152 static void tcp_event_data_sent(struct tcp_sock *tp,
155 struct inet_connection_sock *icsk = inet_csk(sk);
156 const u32 now = tcp_jiffies32;
158 if (tcp_packets_in_flight(tp) == 0)
159 tcp_ca_event(sk, CA_EVENT_TX_START);
163 /* If it is a reply for ato after last received
164 * packet, enter pingpong mode.
166 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
167 icsk->icsk_ack.pingpong = 1;
170 /* Account for an ACK we sent. */
171 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
173 tcp_dec_quickack_mode(sk, pkts);
174 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
178 u32 tcp_default_init_rwnd(u32 mss)
180 /* Initial receive window should be twice of TCP_INIT_CWND to
181 * enable proper sending of new unsent data during fast recovery
182 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
183 * limit when mss is larger than 1460.
185 u32 init_rwnd = TCP_INIT_CWND * 2;
188 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
192 /* Determine a window scaling and initial window to offer.
193 * Based on the assumption that the given amount of space
194 * will be offered. Store the results in the tp structure.
195 * NOTE: for smooth operation initial space offering should
196 * be a multiple of mss if possible. We assume here that mss >= 1.
197 * This MUST be enforced by all callers.
199 void tcp_select_initial_window(int __space, __u32 mss,
200 __u32 *rcv_wnd, __u32 *window_clamp,
201 int wscale_ok, __u8 *rcv_wscale,
204 unsigned int space = (__space < 0 ? 0 : __space);
206 /* If no clamp set the clamp to the max possible scaled window */
207 if (*window_clamp == 0)
208 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
209 space = min(*window_clamp, space);
211 /* Quantize space offering to a multiple of mss if possible. */
213 space = rounddown(space, mss);
215 /* NOTE: offering an initial window larger than 32767
216 * will break some buggy TCP stacks. If the admin tells us
217 * it is likely we could be speaking with such a buggy stack
218 * we will truncate our initial window offering to 32K-1
219 * unless the remote has sent us a window scaling option,
220 * which we interpret as a sign the remote TCP is not
221 * misinterpreting the window field as a signed quantity.
223 if (sysctl_tcp_workaround_signed_windows)
224 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
230 /* Set window scaling on max possible window */
231 space = max_t(u32, space, sysctl_tcp_rmem[2]);
232 space = max_t(u32, space, sysctl_rmem_max);
233 space = min_t(u32, space, *window_clamp);
234 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
240 if (mss > (1 << *rcv_wscale)) {
241 if (!init_rcv_wnd) /* Use default unless specified otherwise */
242 init_rcv_wnd = tcp_default_init_rwnd(mss);
243 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249 EXPORT_SYMBOL(tcp_select_initial_window);
251 /* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
256 static u16 tcp_select_window(struct sock *sk)
258 struct tcp_sock *tp = tcp_sk(sk);
259 u32 old_win = tp->rcv_wnd;
260 u32 cur_win = tcp_receive_window(tp);
261 u32 new_win = __tcp_select_window(sk);
263 /* Never shrink the offered window */
264 if (new_win < cur_win) {
265 /* Danger Will Robinson!
266 * Don't update rcv_wup/rcv_wnd here or else
267 * we will not be able to advertise a zero
268 * window in time. --DaveM
270 * Relax Will Robinson.
273 NET_INC_STATS(sock_net(sk),
274 LINUX_MIB_TCPWANTZEROWINDOWADV);
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
280 /* Make sure we do not exceed the maximum possible
283 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
291 /* If we advertise zero window, disable fast path. */
295 NET_INC_STATS(sock_net(sk),
296 LINUX_MIB_TCPTOZEROWINDOWADV);
297 } else if (old_win == 0) {
298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
304 /* Packet ECN state for a SYN-ACK */
305 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
307 const struct tcp_sock *tp = tcp_sk(sk);
309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
310 if (!(tp->ecn_flags & TCP_ECN_OK))
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
312 else if (tcp_ca_needs_ecn(sk) ||
313 tcp_bpf_ca_needs_ecn(sk))
317 /* Packet ECN state for a SYN. */
318 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
320 struct tcp_sock *tp = tcp_sk(sk);
321 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
322 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
323 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
326 const struct dst_entry *dst = __sk_dst_get(sk);
328 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
337 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
342 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
344 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
345 /* tp->ecn_flags are cleared at a later point in time when
346 * SYN ACK is ultimatively being received.
348 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
352 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
354 if (inet_rsk(req)->ecn_ok)
358 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
362 struct tcphdr *th, int tcp_header_len)
364 struct tcp_sock *tp = tcp_sk(sk);
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
376 } else if (!tcp_ca_needs_ecn(sk)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
385 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
388 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
390 skb->ip_summed = CHECKSUM_PARTIAL;
393 TCP_SKB_CB(skb)->tcp_flags = flags;
394 TCP_SKB_CB(skb)->sacked = 0;
396 tcp_skb_pcount_set(skb, 1);
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
401 TCP_SKB_CB(skb)->end_seq = seq;
404 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
406 return tp->snd_una != tp->snd_up;
409 #define OPTION_SACK_ADVERTISE (1 << 0)
410 #define OPTION_TS (1 << 1)
411 #define OPTION_MD5 (1 << 2)
412 #define OPTION_WSCALE (1 << 3)
413 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
414 #define OPTION_SMC (1 << 9)
416 static void smc_options_write(__be32 *ptr, u16 *options)
418 #if IS_ENABLED(CONFIG_SMC)
419 if (static_branch_unlikely(&tcp_have_smc)) {
420 if (unlikely(OPTION_SMC & *options)) {
421 *ptr++ = htonl((TCPOPT_NOP << 24) |
424 (TCPOLEN_EXP_SMC_BASE));
425 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
431 struct tcp_out_options {
432 u16 options; /* bit field of OPTION_* */
433 u16 mss; /* 0 to disable */
434 u8 ws; /* window scale, 0 to disable */
435 u8 num_sack_blocks; /* number of SACK blocks to include */
436 u8 hash_size; /* bytes in hash_location */
437 __u8 *hash_location; /* temporary pointer, overloaded */
438 __u32 tsval, tsecr; /* need to include OPTION_TS */
439 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 /* Write previously computed TCP options to the packet.
444 * Beware: Something in the Internet is very sensitive to the ordering of
445 * TCP options, we learned this through the hard way, so be careful here.
446 * Luckily we can at least blame others for their non-compliance but from
447 * inter-operability perspective it seems that we're somewhat stuck with
448 * the ordering which we have been using if we want to keep working with
449 * those broken things (not that it currently hurts anybody as there isn't
450 * particular reason why the ordering would need to be changed).
452 * At least SACK_PERM as the first option is known to lead to a disaster
453 * (but it may well be that other scenarios fail similarly).
455 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
456 struct tcp_out_options *opts)
458 u16 options = opts->options; /* mungable copy */
460 if (unlikely(OPTION_MD5 & options)) {
461 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
462 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
463 /* overload cookie hash location */
464 opts->hash_location = (__u8 *)ptr;
468 if (unlikely(opts->mss)) {
469 *ptr++ = htonl((TCPOPT_MSS << 24) |
470 (TCPOLEN_MSS << 16) |
474 if (likely(OPTION_TS & options)) {
475 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
476 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
477 (TCPOLEN_SACK_PERM << 16) |
478 (TCPOPT_TIMESTAMP << 8) |
480 options &= ~OPTION_SACK_ADVERTISE;
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
484 (TCPOPT_TIMESTAMP << 8) |
487 *ptr++ = htonl(opts->tsval);
488 *ptr++ = htonl(opts->tsecr);
491 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
492 *ptr++ = htonl((TCPOPT_NOP << 24) |
494 (TCPOPT_SACK_PERM << 8) |
498 if (unlikely(OPTION_WSCALE & options)) {
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
500 (TCPOPT_WINDOW << 16) |
501 (TCPOLEN_WINDOW << 8) |
505 if (unlikely(opts->num_sack_blocks)) {
506 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
507 tp->duplicate_sack : tp->selective_acks;
510 *ptr++ = htonl((TCPOPT_NOP << 24) |
513 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
514 TCPOLEN_SACK_PERBLOCK)));
516 for (this_sack = 0; this_sack < opts->num_sack_blocks;
518 *ptr++ = htonl(sp[this_sack].start_seq);
519 *ptr++ = htonl(sp[this_sack].end_seq);
522 tp->rx_opt.dsack = 0;
525 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
526 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
528 u32 len; /* Fast Open option length */
531 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
532 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
533 TCPOPT_FASTOPEN_MAGIC);
534 p += TCPOLEN_EXP_FASTOPEN_BASE;
536 len = TCPOLEN_FASTOPEN_BASE + foc->len;
537 *p++ = TCPOPT_FASTOPEN;
541 memcpy(p, foc->val, foc->len);
542 if ((len & 3) == 2) {
543 p[foc->len] = TCPOPT_NOP;
544 p[foc->len + 1] = TCPOPT_NOP;
546 ptr += (len + 3) >> 2;
549 smc_options_write(ptr, &options);
552 static void smc_set_option(const struct tcp_sock *tp,
553 struct tcp_out_options *opts,
554 unsigned int *remaining)
556 #if IS_ENABLED(CONFIG_SMC)
557 if (static_branch_unlikely(&tcp_have_smc)) {
559 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
560 opts->options |= OPTION_SMC;
561 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
568 static void smc_set_option_cond(const struct tcp_sock *tp,
569 const struct inet_request_sock *ireq,
570 struct tcp_out_options *opts,
571 unsigned int *remaining)
573 #if IS_ENABLED(CONFIG_SMC)
574 if (static_branch_unlikely(&tcp_have_smc)) {
575 if (tp->syn_smc && ireq->smc_ok) {
576 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
577 opts->options |= OPTION_SMC;
578 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
585 /* Compute TCP options for SYN packets. This is not the final
586 * network wire format yet.
588 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
589 struct tcp_out_options *opts,
590 struct tcp_md5sig_key **md5)
592 struct tcp_sock *tp = tcp_sk(sk);
593 unsigned int remaining = MAX_TCP_OPTION_SPACE;
594 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
596 #ifdef CONFIG_TCP_MD5SIG
597 *md5 = tp->af_specific->md5_lookup(sk, sk);
599 opts->options |= OPTION_MD5;
600 remaining -= TCPOLEN_MD5SIG_ALIGNED;
606 /* We always get an MSS option. The option bytes which will be seen in
607 * normal data packets should timestamps be used, must be in the MSS
608 * advertised. But we subtract them from tp->mss_cache so that
609 * calculations in tcp_sendmsg are simpler etc. So account for this
610 * fact here if necessary. If we don't do this correctly, as a
611 * receiver we won't recognize data packets as being full sized when we
612 * should, and thus we won't abide by the delayed ACK rules correctly.
613 * SACKs don't matter, we never delay an ACK when we have any of those
615 opts->mss = tcp_advertise_mss(sk);
616 remaining -= TCPOLEN_MSS_ALIGNED;
618 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
619 opts->options |= OPTION_TS;
620 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
621 opts->tsecr = tp->rx_opt.ts_recent;
622 remaining -= TCPOLEN_TSTAMP_ALIGNED;
624 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
625 opts->ws = tp->rx_opt.rcv_wscale;
626 opts->options |= OPTION_WSCALE;
627 remaining -= TCPOLEN_WSCALE_ALIGNED;
629 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
630 opts->options |= OPTION_SACK_ADVERTISE;
631 if (unlikely(!(OPTION_TS & opts->options)))
632 remaining -= TCPOLEN_SACKPERM_ALIGNED;
635 if (fastopen && fastopen->cookie.len >= 0) {
636 u32 need = fastopen->cookie.len;
638 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
639 TCPOLEN_FASTOPEN_BASE;
640 need = (need + 3) & ~3U; /* Align to 32 bits */
641 if (remaining >= need) {
642 opts->options |= OPTION_FAST_OPEN_COOKIE;
643 opts->fastopen_cookie = &fastopen->cookie;
645 tp->syn_fastopen = 1;
646 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
650 smc_set_option(tp, opts, &remaining);
652 return MAX_TCP_OPTION_SPACE - remaining;
655 /* Set up TCP options for SYN-ACKs. */
656 static unsigned int tcp_synack_options(const struct sock *sk,
657 struct request_sock *req,
658 unsigned int mss, struct sk_buff *skb,
659 struct tcp_out_options *opts,
660 const struct tcp_md5sig_key *md5,
661 struct tcp_fastopen_cookie *foc)
663 struct inet_request_sock *ireq = inet_rsk(req);
664 unsigned int remaining = MAX_TCP_OPTION_SPACE;
666 #ifdef CONFIG_TCP_MD5SIG
668 opts->options |= OPTION_MD5;
669 remaining -= TCPOLEN_MD5SIG_ALIGNED;
671 /* We can't fit any SACK blocks in a packet with MD5 + TS
672 * options. There was discussion about disabling SACK
673 * rather than TS in order to fit in better with old,
674 * buggy kernels, but that was deemed to be unnecessary.
676 ireq->tstamp_ok &= !ireq->sack_ok;
680 /* We always send an MSS option. */
682 remaining -= TCPOLEN_MSS_ALIGNED;
684 if (likely(ireq->wscale_ok)) {
685 opts->ws = ireq->rcv_wscale;
686 opts->options |= OPTION_WSCALE;
687 remaining -= TCPOLEN_WSCALE_ALIGNED;
689 if (likely(ireq->tstamp_ok)) {
690 opts->options |= OPTION_TS;
691 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
692 opts->tsecr = req->ts_recent;
693 remaining -= TCPOLEN_TSTAMP_ALIGNED;
695 if (likely(ireq->sack_ok)) {
696 opts->options |= OPTION_SACK_ADVERTISE;
697 if (unlikely(!ireq->tstamp_ok))
698 remaining -= TCPOLEN_SACKPERM_ALIGNED;
700 if (foc != NULL && foc->len >= 0) {
703 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
704 TCPOLEN_FASTOPEN_BASE;
705 need = (need + 3) & ~3U; /* Align to 32 bits */
706 if (remaining >= need) {
707 opts->options |= OPTION_FAST_OPEN_COOKIE;
708 opts->fastopen_cookie = foc;
713 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
715 return MAX_TCP_OPTION_SPACE - remaining;
718 /* Compute TCP options for ESTABLISHED sockets. This is not the
719 * final wire format yet.
721 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
722 struct tcp_out_options *opts,
723 struct tcp_md5sig_key **md5)
725 struct tcp_sock *tp = tcp_sk(sk);
726 unsigned int size = 0;
727 unsigned int eff_sacks;
731 #ifdef CONFIG_TCP_MD5SIG
732 *md5 = tp->af_specific->md5_lookup(sk, sk);
733 if (unlikely(*md5)) {
734 opts->options |= OPTION_MD5;
735 size += TCPOLEN_MD5SIG_ALIGNED;
741 if (likely(tp->rx_opt.tstamp_ok)) {
742 opts->options |= OPTION_TS;
743 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
744 opts->tsecr = tp->rx_opt.ts_recent;
745 size += TCPOLEN_TSTAMP_ALIGNED;
748 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
749 if (unlikely(eff_sacks)) {
750 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
751 opts->num_sack_blocks =
752 min_t(unsigned int, eff_sacks,
753 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
754 TCPOLEN_SACK_PERBLOCK);
755 size += TCPOLEN_SACK_BASE_ALIGNED +
756 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
763 /* TCP SMALL QUEUES (TSQ)
765 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
766 * to reduce RTT and bufferbloat.
767 * We do this using a special skb destructor (tcp_wfree).
769 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
770 * needs to be reallocated in a driver.
771 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
773 * Since transmit from skb destructor is forbidden, we use a tasklet
774 * to process all sockets that eventually need to send more skbs.
775 * We use one tasklet per cpu, with its own queue of sockets.
778 struct tasklet_struct tasklet;
779 struct list_head head; /* queue of tcp sockets */
781 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
783 static void tcp_tsq_handler(struct sock *sk)
785 if ((1 << sk->sk_state) &
786 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
787 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
788 struct tcp_sock *tp = tcp_sk(sk);
790 if (tp->lost_out > tp->retrans_out &&
791 tp->snd_cwnd > tcp_packets_in_flight(tp))
792 tcp_xmit_retransmit_queue(sk);
794 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
799 * One tasklet per cpu tries to send more skbs.
800 * We run in tasklet context but need to disable irqs when
801 * transferring tsq->head because tcp_wfree() might
802 * interrupt us (non NAPI drivers)
804 static void tcp_tasklet_func(unsigned long data)
806 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
809 struct list_head *q, *n;
813 local_irq_save(flags);
814 list_splice_init(&tsq->head, &list);
815 local_irq_restore(flags);
817 list_for_each_safe(q, n, &list) {
818 tp = list_entry(q, struct tcp_sock, tsq_node);
819 list_del(&tp->tsq_node);
821 sk = (struct sock *)tp;
822 smp_mb__before_atomic();
823 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
825 if (!sk->sk_lock.owned &&
826 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
828 if (!sock_owned_by_user(sk)) {
829 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
839 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
840 TCPF_WRITE_TIMER_DEFERRED | \
841 TCPF_DELACK_TIMER_DEFERRED | \
842 TCPF_MTU_REDUCED_DEFERRED)
844 * tcp_release_cb - tcp release_sock() callback
847 * called from release_sock() to perform protocol dependent
848 * actions before socket release.
850 void tcp_release_cb(struct sock *sk)
852 unsigned long flags, nflags;
854 /* perform an atomic operation only if at least one flag is set */
856 flags = sk->sk_tsq_flags;
857 if (!(flags & TCP_DEFERRED_ALL))
859 nflags = flags & ~TCP_DEFERRED_ALL;
860 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
862 if (flags & TCPF_TSQ_DEFERRED)
865 /* Here begins the tricky part :
866 * We are called from release_sock() with :
868 * 2) sk_lock.slock spinlock held
869 * 3) socket owned by us (sk->sk_lock.owned == 1)
871 * But following code is meant to be called from BH handlers,
872 * so we should keep BH disabled, but early release socket ownership
874 sock_release_ownership(sk);
876 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
877 tcp_write_timer_handler(sk);
880 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
881 tcp_delack_timer_handler(sk);
884 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
885 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
889 EXPORT_SYMBOL(tcp_release_cb);
891 void __init tcp_tasklet_init(void)
895 for_each_possible_cpu(i) {
896 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
898 INIT_LIST_HEAD(&tsq->head);
899 tasklet_init(&tsq->tasklet,
906 * Write buffer destructor automatically called from kfree_skb.
907 * We can't xmit new skbs from this context, as we might already
910 void tcp_wfree(struct sk_buff *skb)
912 struct sock *sk = skb->sk;
913 struct tcp_sock *tp = tcp_sk(sk);
914 unsigned long flags, nval, oval;
916 /* Keep one reference on sk_wmem_alloc.
917 * Will be released by sk_free() from here or tcp_tasklet_func()
919 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
921 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
922 * Wait until our queues (qdisc + devices) are drained.
924 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
925 * - chance for incoming ACK (processed by another cpu maybe)
926 * to migrate this flow (skb->ooo_okay will be eventually set)
928 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
931 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
932 struct tsq_tasklet *tsq;
935 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
938 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
939 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
943 /* queue this socket to tasklet queue */
944 local_irq_save(flags);
945 tsq = this_cpu_ptr(&tsq_tasklet);
946 empty = list_empty(&tsq->head);
947 list_add(&tp->tsq_node, &tsq->head);
949 tasklet_schedule(&tsq->tasklet);
950 local_irq_restore(flags);
957 /* Note: Called under hard irq.
958 * We can not call TCP stack right away.
960 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
962 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
963 struct sock *sk = (struct sock *)tp;
964 unsigned long nval, oval;
966 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
967 struct tsq_tasklet *tsq;
970 if (oval & TSQF_QUEUED)
973 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
974 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
978 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
980 /* queue this socket to tasklet queue */
981 tsq = this_cpu_ptr(&tsq_tasklet);
982 empty = list_empty(&tsq->head);
983 list_add(&tp->tsq_node, &tsq->head);
985 tasklet_schedule(&tsq->tasklet);
988 return HRTIMER_NORESTART;
991 /* BBR congestion control needs pacing.
992 * Same remark for SO_MAX_PACING_RATE.
993 * sch_fq packet scheduler is efficiently handling pacing,
994 * but is not always installed/used.
995 * Return true if TCP stack should pace packets itself.
997 static bool tcp_needs_internal_pacing(const struct sock *sk)
999 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
1002 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
1007 if (!tcp_needs_internal_pacing(sk))
1009 rate = sk->sk_pacing_rate;
1010 if (!rate || rate == ~0U)
1013 /* Should account for header sizes as sch_fq does,
1014 * but lets make things simple.
1016 len_ns = (u64)skb->len * NSEC_PER_SEC;
1017 do_div(len_ns, rate);
1018 hrtimer_start(&tcp_sk(sk)->pacing_timer,
1019 ktime_add_ns(ktime_get(), len_ns),
1020 HRTIMER_MODE_ABS_PINNED);
1023 static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1025 skb->skb_mstamp = tp->tcp_mstamp;
1026 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1029 /* This routine actually transmits TCP packets queued in by
1030 * tcp_do_sendmsg(). This is used by both the initial
1031 * transmission and possible later retransmissions.
1032 * All SKB's seen here are completely headerless. It is our
1033 * job to build the TCP header, and pass the packet down to
1034 * IP so it can do the same plus pass the packet off to the
1037 * We are working here with either a clone of the original
1038 * SKB, or a fresh unique copy made by the retransmit engine.
1040 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1043 const struct inet_connection_sock *icsk = inet_csk(sk);
1044 struct inet_sock *inet;
1045 struct tcp_sock *tp;
1046 struct tcp_skb_cb *tcb;
1047 struct tcp_out_options opts;
1048 unsigned int tcp_options_size, tcp_header_size;
1049 struct sk_buff *oskb = NULL;
1050 struct tcp_md5sig_key *md5;
1054 BUG_ON(!skb || !tcp_skb_pcount(skb));
1058 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1062 tcp_skb_tsorted_save(oskb) {
1063 if (unlikely(skb_cloned(oskb)))
1064 skb = pskb_copy(oskb, gfp_mask);
1066 skb = skb_clone(oskb, gfp_mask);
1067 } tcp_skb_tsorted_restore(oskb);
1072 skb->skb_mstamp = tp->tcp_mstamp;
1075 tcb = TCP_SKB_CB(skb);
1076 memset(&opts, 0, sizeof(opts));
1078 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1079 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1081 tcp_options_size = tcp_established_options(sk, skb, &opts,
1083 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1085 /* if no packet is in qdisc/device queue, then allow XPS to select
1086 * another queue. We can be called from tcp_tsq_handler()
1087 * which holds one reference to sk_wmem_alloc.
1089 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1090 * One way to get this would be to set skb->truesize = 2 on them.
1092 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1094 /* If we had to use memory reserve to allocate this skb,
1095 * this might cause drops if packet is looped back :
1096 * Other socket might not have SOCK_MEMALLOC.
1097 * Packets not looped back do not care about pfmemalloc.
1099 skb->pfmemalloc = 0;
1101 skb_push(skb, tcp_header_size);
1102 skb_reset_transport_header(skb);
1106 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1107 skb_set_hash_from_sk(skb, sk);
1108 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1110 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1112 /* Build TCP header and checksum it. */
1113 th = (struct tcphdr *)skb->data;
1114 th->source = inet->inet_sport;
1115 th->dest = inet->inet_dport;
1116 th->seq = htonl(tcb->seq);
1117 th->ack_seq = htonl(tp->rcv_nxt);
1118 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1124 /* The urg_mode check is necessary during a below snd_una win probe */
1125 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1126 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1127 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1129 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1130 th->urg_ptr = htons(0xFFFF);
1135 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1136 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1137 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1138 th->window = htons(tcp_select_window(sk));
1139 tcp_ecn_send(sk, skb, th, tcp_header_size);
1141 /* RFC1323: The window in SYN & SYN/ACK segments
1144 th->window = htons(min(tp->rcv_wnd, 65535U));
1146 #ifdef CONFIG_TCP_MD5SIG
1147 /* Calculate the MD5 hash, as we have all we need now */
1149 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1150 tp->af_specific->calc_md5_hash(opts.hash_location,
1155 icsk->icsk_af_ops->send_check(sk, skb);
1157 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1158 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1160 if (skb->len != tcp_header_size) {
1161 tcp_event_data_sent(tp, sk);
1162 tp->data_segs_out += tcp_skb_pcount(skb);
1163 tcp_internal_pacing(sk, skb);
1166 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1167 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1168 tcp_skb_pcount(skb));
1170 tp->segs_out += tcp_skb_pcount(skb);
1171 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1172 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1173 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1175 /* Our usage of tstamp should remain private */
1178 /* Cleanup our debris for IP stacks */
1179 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1180 sizeof(struct inet6_skb_parm)));
1182 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1184 if (unlikely(err > 0)) {
1186 err = net_xmit_eval(err);
1189 tcp_update_skb_after_send(tp, oskb);
1190 tcp_rate_skb_sent(sk, oskb);
1195 /* This routine just queues the buffer for sending.
1197 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1198 * otherwise socket can stall.
1200 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1202 struct tcp_sock *tp = tcp_sk(sk);
1204 /* Advance write_seq and place onto the write_queue. */
1205 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1206 __skb_header_release(skb);
1207 tcp_add_write_queue_tail(sk, skb);
1208 sk->sk_wmem_queued += skb->truesize;
1209 sk_mem_charge(sk, skb->truesize);
1212 /* Initialize TSO segments for a packet. */
1213 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1215 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1216 /* Avoid the costly divide in the normal
1219 tcp_skb_pcount_set(skb, 1);
1220 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1222 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1223 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1227 /* When a modification to fackets out becomes necessary, we need to check
1228 * skb is counted to fackets_out or not.
1230 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1233 struct tcp_sock *tp = tcp_sk(sk);
1235 if (!tp->sacked_out || tcp_is_reno(tp))
1238 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1239 tp->fackets_out -= decr;
1242 /* Pcount in the middle of the write queue got changed, we need to do various
1243 * tweaks to fix counters
1245 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1247 struct tcp_sock *tp = tcp_sk(sk);
1249 tp->packets_out -= decr;
1251 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1252 tp->sacked_out -= decr;
1253 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1254 tp->retrans_out -= decr;
1255 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1256 tp->lost_out -= decr;
1258 /* Reno case is special. Sigh... */
1259 if (tcp_is_reno(tp) && decr > 0)
1260 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1262 tcp_adjust_fackets_out(sk, skb, decr);
1264 if (tp->lost_skb_hint &&
1265 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1266 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1267 tp->lost_cnt_hint -= decr;
1269 tcp_verify_left_out(tp);
1272 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1274 return TCP_SKB_CB(skb)->txstamp_ack ||
1275 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1278 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1280 struct skb_shared_info *shinfo = skb_shinfo(skb);
1282 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1283 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1284 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1285 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1287 shinfo->tx_flags &= ~tsflags;
1288 shinfo2->tx_flags |= tsflags;
1289 swap(shinfo->tskey, shinfo2->tskey);
1290 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1291 TCP_SKB_CB(skb)->txstamp_ack = 0;
1295 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1297 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1298 TCP_SKB_CB(skb)->eor = 0;
1301 /* Insert buff after skb on the write or rtx queue of sk. */
1302 static void tcp_insert_write_queue_after(struct sk_buff *skb,
1303 struct sk_buff *buff,
1305 enum tcp_queue tcp_queue)
1307 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1308 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1310 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1313 /* Function to create two new TCP segments. Shrinks the given segment
1314 * to the specified size and appends a new segment with the rest of the
1315 * packet to the list. This won't be called frequently, I hope.
1316 * Remember, these are still headerless SKBs at this point.
1318 int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1319 struct sk_buff *skb, u32 len,
1320 unsigned int mss_now, gfp_t gfp)
1322 struct tcp_sock *tp = tcp_sk(sk);
1323 struct sk_buff *buff;
1324 int nsize, old_factor;
1328 if (WARN_ON(len > skb->len))
1331 nsize = skb_headlen(skb) - len;
1335 if (skb_unclone(skb, gfp))
1338 /* Get a new skb... force flag on. */
1339 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1341 return -ENOMEM; /* We'll just try again later. */
1343 sk->sk_wmem_queued += buff->truesize;
1344 sk_mem_charge(sk, buff->truesize);
1345 nlen = skb->len - len - nsize;
1346 buff->truesize += nlen;
1347 skb->truesize -= nlen;
1349 /* Correct the sequence numbers. */
1350 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1351 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1352 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1354 /* PSH and FIN should only be set in the second packet. */
1355 flags = TCP_SKB_CB(skb)->tcp_flags;
1356 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1357 TCP_SKB_CB(buff)->tcp_flags = flags;
1358 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1359 tcp_skb_fragment_eor(skb, buff);
1361 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1362 /* Copy and checksum data tail into the new buffer. */
1363 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1364 skb_put(buff, nsize),
1369 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1371 skb->ip_summed = CHECKSUM_PARTIAL;
1372 skb_split(skb, buff, len);
1375 buff->ip_summed = skb->ip_summed;
1377 buff->tstamp = skb->tstamp;
1378 tcp_fragment_tstamp(skb, buff);
1380 old_factor = tcp_skb_pcount(skb);
1382 /* Fix up tso_factor for both original and new SKB. */
1383 tcp_set_skb_tso_segs(skb, mss_now);
1384 tcp_set_skb_tso_segs(buff, mss_now);
1386 /* Update delivered info for the new segment */
1387 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1389 /* If this packet has been sent out already, we must
1390 * adjust the various packet counters.
1392 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1393 int diff = old_factor - tcp_skb_pcount(skb) -
1394 tcp_skb_pcount(buff);
1397 tcp_adjust_pcount(sk, skb, diff);
1400 /* Link BUFF into the send queue. */
1401 __skb_header_release(buff);
1402 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1403 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1408 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1409 * data is not copied, but immediately discarded.
1411 static int __pskb_trim_head(struct sk_buff *skb, int len)
1413 struct skb_shared_info *shinfo;
1416 eat = min_t(int, len, skb_headlen(skb));
1418 __skb_pull(skb, eat);
1425 shinfo = skb_shinfo(skb);
1426 for (i = 0; i < shinfo->nr_frags; i++) {
1427 int size = skb_frag_size(&shinfo->frags[i]);
1430 skb_frag_unref(skb, i);
1433 shinfo->frags[k] = shinfo->frags[i];
1435 shinfo->frags[k].page_offset += eat;
1436 skb_frag_size_sub(&shinfo->frags[k], eat);
1442 shinfo->nr_frags = k;
1444 skb->data_len -= len;
1445 skb->len = skb->data_len;
1449 /* Remove acked data from a packet in the transmit queue. */
1450 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1454 if (skb_unclone(skb, GFP_ATOMIC))
1457 delta_truesize = __pskb_trim_head(skb, len);
1459 TCP_SKB_CB(skb)->seq += len;
1460 skb->ip_summed = CHECKSUM_PARTIAL;
1462 if (delta_truesize) {
1463 skb->truesize -= delta_truesize;
1464 sk->sk_wmem_queued -= delta_truesize;
1465 sk_mem_uncharge(sk, delta_truesize);
1466 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1469 /* Any change of skb->len requires recalculation of tso factor. */
1470 if (tcp_skb_pcount(skb) > 1)
1471 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1476 /* Calculate MSS not accounting any TCP options. */
1477 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1479 const struct tcp_sock *tp = tcp_sk(sk);
1480 const struct inet_connection_sock *icsk = inet_csk(sk);
1483 /* Calculate base mss without TCP options:
1484 It is MMS_S - sizeof(tcphdr) of rfc1122
1486 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1488 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1489 if (icsk->icsk_af_ops->net_frag_header_len) {
1490 const struct dst_entry *dst = __sk_dst_get(sk);
1492 if (dst && dst_allfrag(dst))
1493 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1496 /* Clamp it (mss_clamp does not include tcp options) */
1497 if (mss_now > tp->rx_opt.mss_clamp)
1498 mss_now = tp->rx_opt.mss_clamp;
1500 /* Now subtract optional transport overhead */
1501 mss_now -= icsk->icsk_ext_hdr_len;
1503 /* Then reserve room for full set of TCP options and 8 bytes of data */
1509 /* Calculate MSS. Not accounting for SACKs here. */
1510 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1512 /* Subtract TCP options size, not including SACKs */
1513 return __tcp_mtu_to_mss(sk, pmtu) -
1514 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1517 /* Inverse of above */
1518 int tcp_mss_to_mtu(struct sock *sk, int mss)
1520 const struct tcp_sock *tp = tcp_sk(sk);
1521 const struct inet_connection_sock *icsk = inet_csk(sk);
1525 tp->tcp_header_len +
1526 icsk->icsk_ext_hdr_len +
1527 icsk->icsk_af_ops->net_header_len;
1529 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1530 if (icsk->icsk_af_ops->net_frag_header_len) {
1531 const struct dst_entry *dst = __sk_dst_get(sk);
1533 if (dst && dst_allfrag(dst))
1534 mtu += icsk->icsk_af_ops->net_frag_header_len;
1538 EXPORT_SYMBOL(tcp_mss_to_mtu);
1540 /* MTU probing init per socket */
1541 void tcp_mtup_init(struct sock *sk)
1543 struct tcp_sock *tp = tcp_sk(sk);
1544 struct inet_connection_sock *icsk = inet_csk(sk);
1545 struct net *net = sock_net(sk);
1547 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1548 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1549 icsk->icsk_af_ops->net_header_len;
1550 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1551 icsk->icsk_mtup.probe_size = 0;
1552 if (icsk->icsk_mtup.enabled)
1553 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1555 EXPORT_SYMBOL(tcp_mtup_init);
1557 /* This function synchronize snd mss to current pmtu/exthdr set.
1559 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1560 for TCP options, but includes only bare TCP header.
1562 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1563 It is minimum of user_mss and mss received with SYN.
1564 It also does not include TCP options.
1566 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1568 tp->mss_cache is current effective sending mss, including
1569 all tcp options except for SACKs. It is evaluated,
1570 taking into account current pmtu, but never exceeds
1571 tp->rx_opt.mss_clamp.
1573 NOTE1. rfc1122 clearly states that advertised MSS
1574 DOES NOT include either tcp or ip options.
1576 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1577 are READ ONLY outside this function. --ANK (980731)
1579 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1581 struct tcp_sock *tp = tcp_sk(sk);
1582 struct inet_connection_sock *icsk = inet_csk(sk);
1585 if (icsk->icsk_mtup.search_high > pmtu)
1586 icsk->icsk_mtup.search_high = pmtu;
1588 mss_now = tcp_mtu_to_mss(sk, pmtu);
1589 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1591 /* And store cached results */
1592 icsk->icsk_pmtu_cookie = pmtu;
1593 if (icsk->icsk_mtup.enabled)
1594 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1595 tp->mss_cache = mss_now;
1599 EXPORT_SYMBOL(tcp_sync_mss);
1601 /* Compute the current effective MSS, taking SACKs and IP options,
1602 * and even PMTU discovery events into account.
1604 unsigned int tcp_current_mss(struct sock *sk)
1606 const struct tcp_sock *tp = tcp_sk(sk);
1607 const struct dst_entry *dst = __sk_dst_get(sk);
1609 unsigned int header_len;
1610 struct tcp_out_options opts;
1611 struct tcp_md5sig_key *md5;
1613 mss_now = tp->mss_cache;
1616 u32 mtu = dst_mtu(dst);
1617 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1618 mss_now = tcp_sync_mss(sk, mtu);
1621 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1622 sizeof(struct tcphdr);
1623 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1624 * some common options. If this is an odd packet (because we have SACK
1625 * blocks etc) then our calculated header_len will be different, and
1626 * we have to adjust mss_now correspondingly */
1627 if (header_len != tp->tcp_header_len) {
1628 int delta = (int) header_len - tp->tcp_header_len;
1635 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1636 * As additional protections, we do not touch cwnd in retransmission phases,
1637 * and if application hit its sndbuf limit recently.
1639 static void tcp_cwnd_application_limited(struct sock *sk)
1641 struct tcp_sock *tp = tcp_sk(sk);
1643 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1644 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1645 /* Limited by application or receiver window. */
1646 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1647 u32 win_used = max(tp->snd_cwnd_used, init_win);
1648 if (win_used < tp->snd_cwnd) {
1649 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1650 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1652 tp->snd_cwnd_used = 0;
1654 tp->snd_cwnd_stamp = tcp_jiffies32;
1657 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1659 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1660 struct tcp_sock *tp = tcp_sk(sk);
1662 /* Track the maximum number of outstanding packets in each
1663 * window, and remember whether we were cwnd-limited then.
1665 if (!before(tp->snd_una, tp->max_packets_seq) ||
1666 tp->packets_out > tp->max_packets_out) {
1667 tp->max_packets_out = tp->packets_out;
1668 tp->max_packets_seq = tp->snd_nxt;
1669 tp->is_cwnd_limited = is_cwnd_limited;
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 (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 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1745 bytes = min(sk->sk_pacing_rate >> 10,
1746 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1748 /* Goal is to send at least one packet per ms,
1749 * not one big TSO packet every 100 ms.
1750 * This preserves ACK clocking and is consistent
1751 * with tcp_tso_should_defer() heuristic.
1753 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1755 return min_t(u32, segs, sk->sk_gso_max_segs);
1757 EXPORT_SYMBOL(tcp_tso_autosize);
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 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1768 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1771 /* Returns the portion of skb which can be sent right away */
1772 static unsigned int tcp_mss_split_point(const struct sock *sk,
1773 const struct sk_buff *skb,
1774 unsigned int mss_now,
1775 unsigned int max_segs,
1778 const struct tcp_sock *tp = tcp_sk(sk);
1779 u32 partial, needed, window, max_len;
1781 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1782 max_len = mss_now * max_segs;
1784 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1787 needed = min(skb->len, window);
1789 if (max_len <= needed)
1792 partial = needed % mss_now;
1793 /* If last segment is not a full MSS, check if Nagle rules allow us
1794 * to include this last segment in this skb.
1795 * Otherwise, we'll split the skb at last MSS boundary
1797 if (tcp_nagle_check(partial != 0, tp, nonagle))
1798 return needed - partial;
1803 /* Can at least one segment of SKB be sent right now, according to the
1804 * congestion window rules? If so, return how many segments are allowed.
1806 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1807 const struct sk_buff *skb)
1809 u32 in_flight, cwnd, halfcwnd;
1811 /* Don't be strict about the congestion window for the final FIN. */
1812 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1813 tcp_skb_pcount(skb) == 1)
1816 in_flight = tcp_packets_in_flight(tp);
1817 cwnd = tp->snd_cwnd;
1818 if (in_flight >= cwnd)
1821 /* For better scheduling, ensure we have at least
1822 * 2 GSO packets in flight.
1824 halfcwnd = max(cwnd >> 1, 1U);
1825 return min(halfcwnd, cwnd - in_flight);
1828 /* Initialize TSO state of a skb.
1829 * This must be invoked the first time we consider transmitting
1830 * SKB onto the wire.
1832 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1834 int tso_segs = tcp_skb_pcount(skb);
1836 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1837 tcp_set_skb_tso_segs(skb, mss_now);
1838 tso_segs = tcp_skb_pcount(skb);
1844 /* Return true if the Nagle test allows this packet to be
1847 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1848 unsigned int cur_mss, int nonagle)
1850 /* Nagle rule does not apply to frames, which sit in the middle of the
1851 * write_queue (they have no chances to get new data).
1853 * This is implemented in the callers, where they modify the 'nonagle'
1854 * argument based upon the location of SKB in the send queue.
1856 if (nonagle & TCP_NAGLE_PUSH)
1859 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1860 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1863 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1869 /* Does at least the first segment of SKB fit into the send window? */
1870 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1871 const struct sk_buff *skb,
1872 unsigned int cur_mss)
1874 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1876 if (skb->len > cur_mss)
1877 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1879 return !after(end_seq, tcp_wnd_end(tp));
1882 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1883 * which is put after SKB on the list. It is very much like
1884 * tcp_fragment() except that it may make several kinds of assumptions
1885 * in order to speed up the splitting operation. In particular, we
1886 * know that all the data is in scatter-gather pages, and that the
1887 * packet has never been sent out before (and thus is not cloned).
1889 static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1890 struct sk_buff *skb, unsigned int len,
1891 unsigned int mss_now, gfp_t gfp)
1893 struct sk_buff *buff;
1894 int nlen = skb->len - len;
1897 /* All of a TSO frame must be composed of paged data. */
1898 if (skb->len != skb->data_len)
1899 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1901 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1902 if (unlikely(!buff))
1905 sk->sk_wmem_queued += buff->truesize;
1906 sk_mem_charge(sk, buff->truesize);
1907 buff->truesize += nlen;
1908 skb->truesize -= nlen;
1910 /* Correct the sequence numbers. */
1911 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1912 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1913 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1915 /* PSH and FIN should only be set in the second packet. */
1916 flags = TCP_SKB_CB(skb)->tcp_flags;
1917 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1918 TCP_SKB_CB(buff)->tcp_flags = flags;
1920 /* This packet was never sent out yet, so no SACK bits. */
1921 TCP_SKB_CB(buff)->sacked = 0;
1923 tcp_skb_fragment_eor(skb, buff);
1925 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1926 skb_split(skb, buff, len);
1927 tcp_fragment_tstamp(skb, buff);
1929 /* Fix up tso_factor for both original and new SKB. */
1930 tcp_set_skb_tso_segs(skb, mss_now);
1931 tcp_set_skb_tso_segs(buff, mss_now);
1933 /* Link BUFF into the send queue. */
1934 __skb_header_release(buff);
1935 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1940 /* Try to defer sending, if possible, in order to minimize the amount
1941 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1943 * This algorithm is from John Heffner.
1945 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1946 bool *is_cwnd_limited, u32 max_segs)
1948 const struct inet_connection_sock *icsk = inet_csk(sk);
1949 u32 age, send_win, cong_win, limit, in_flight;
1950 struct tcp_sock *tp = tcp_sk(sk);
1951 struct sk_buff *head;
1954 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1957 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1960 /* Avoid bursty behavior by allowing defer
1961 * only if the last write was recent.
1963 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1966 in_flight = tcp_packets_in_flight(tp);
1968 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1970 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1972 /* From in_flight test above, we know that cwnd > in_flight. */
1973 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1975 limit = min(send_win, cong_win);
1977 /* If a full-sized TSO skb can be sent, do it. */
1978 if (limit >= max_segs * tp->mss_cache)
1981 /* Middle in queue won't get any more data, full sendable already? */
1982 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1985 win_divisor = ACCESS_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1987 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1989 /* If at least some fraction of a window is available,
1992 chunk /= win_divisor;
1996 /* Different approach, try not to defer past a single
1997 * ACK. Receiver should ACK every other full sized
1998 * frame, so if we have space for more than 3 frames
2001 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2005 /* TODO : use tsorted_sent_queue ? */
2006 head = tcp_rtx_queue_head(sk);
2009 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
2010 /* If next ACK is likely to come too late (half srtt), do not defer */
2011 if (age < (tp->srtt_us >> 4))
2014 /* Ok, it looks like it is advisable to defer. */
2016 if (cong_win < send_win && cong_win <= skb->len)
2017 *is_cwnd_limited = true;
2025 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2027 struct inet_connection_sock *icsk = inet_csk(sk);
2028 struct tcp_sock *tp = tcp_sk(sk);
2029 struct net *net = sock_net(sk);
2033 interval = net->ipv4.sysctl_tcp_probe_interval;
2034 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2035 if (unlikely(delta >= interval * HZ)) {
2036 int mss = tcp_current_mss(sk);
2038 /* Update current search range */
2039 icsk->icsk_mtup.probe_size = 0;
2040 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2041 sizeof(struct tcphdr) +
2042 icsk->icsk_af_ops->net_header_len;
2043 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2045 /* Update probe time stamp */
2046 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2050 /* Create a new MTU probe if we are ready.
2051 * MTU probe is regularly attempting to increase the path MTU by
2052 * deliberately sending larger packets. This discovers routing
2053 * changes resulting in larger path MTUs.
2055 * Returns 0 if we should wait to probe (no cwnd available),
2056 * 1 if a probe was sent,
2059 static int tcp_mtu_probe(struct sock *sk)
2061 struct inet_connection_sock *icsk = inet_csk(sk);
2062 struct tcp_sock *tp = tcp_sk(sk);
2063 struct sk_buff *skb, *nskb, *next;
2064 struct net *net = sock_net(sk);
2071 /* Not currently probing/verifying,
2073 * have enough cwnd, and
2074 * not SACKing (the variable headers throw things off)
2076 if (likely(!icsk->icsk_mtup.enabled ||
2077 icsk->icsk_mtup.probe_size ||
2078 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2079 tp->snd_cwnd < 11 ||
2080 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2083 /* Use binary search for probe_size between tcp_mss_base,
2084 * and current mss_clamp. if (search_high - search_low)
2085 * smaller than a threshold, backoff from probing.
2087 mss_now = tcp_current_mss(sk);
2088 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2089 icsk->icsk_mtup.search_low) >> 1);
2090 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2091 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2092 /* When misfortune happens, we are reprobing actively,
2093 * and then reprobe timer has expired. We stick with current
2094 * probing process by not resetting search range to its orignal.
2096 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2097 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2098 /* Check whether enough time has elaplased for
2099 * another round of probing.
2101 tcp_mtu_check_reprobe(sk);
2105 /* Have enough data in the send queue to probe? */
2106 if (tp->write_seq - tp->snd_nxt < size_needed)
2109 if (tp->snd_wnd < size_needed)
2111 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2114 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2115 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2116 if (!tcp_packets_in_flight(tp))
2122 /* We're allowed to probe. Build it now. */
2123 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2126 sk->sk_wmem_queued += nskb->truesize;
2127 sk_mem_charge(sk, nskb->truesize);
2129 skb = tcp_send_head(sk);
2131 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2132 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2133 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2134 TCP_SKB_CB(nskb)->sacked = 0;
2136 nskb->ip_summed = skb->ip_summed;
2138 tcp_insert_write_queue_before(nskb, skb, sk);
2141 tcp_for_write_queue_from_safe(skb, next, sk) {
2142 copy = min_t(int, skb->len, probe_size - len);
2143 if (nskb->ip_summed) {
2144 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2146 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2147 skb_put(nskb, copy),
2149 nskb->csum = csum_block_add(nskb->csum, csum, len);
2152 if (skb->len <= copy) {
2153 /* We've eaten all the data from this skb.
2155 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2156 tcp_unlink_write_queue(skb, sk);
2157 sk_wmem_free_skb(sk, skb);
2159 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2160 ~(TCPHDR_FIN|TCPHDR_PSH);
2161 if (!skb_shinfo(skb)->nr_frags) {
2162 skb_pull(skb, copy);
2163 if (skb->ip_summed != CHECKSUM_PARTIAL)
2164 skb->csum = csum_partial(skb->data,
2167 __pskb_trim_head(skb, copy);
2168 tcp_set_skb_tso_segs(skb, mss_now);
2170 TCP_SKB_CB(skb)->seq += copy;
2175 if (len >= probe_size)
2178 tcp_init_tso_segs(nskb, nskb->len);
2180 /* We're ready to send. If this fails, the probe will
2181 * be resegmented into mss-sized pieces by tcp_write_xmit().
2183 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2184 /* Decrement cwnd here because we are sending
2185 * effectively two packets. */
2187 tcp_event_new_data_sent(sk, nskb);
2189 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2190 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2191 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2199 static bool tcp_pacing_check(const struct sock *sk)
2201 return tcp_needs_internal_pacing(sk) &&
2202 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2205 /* TCP Small Queues :
2206 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2207 * (These limits are doubled for retransmits)
2209 * - better RTT estimation and ACK scheduling
2212 * Alas, some drivers / subsystems require a fair amount
2213 * of queued bytes to ensure line rate.
2214 * One example is wifi aggregation (802.11 AMPDU)
2216 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2217 unsigned int factor)
2221 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2222 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2225 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2226 /* Always send skb if rtx queue is empty.
2227 * No need to wait for TX completion to call us back,
2228 * after softirq/tasklet schedule.
2229 * This helps when TX completions are delayed too much.
2231 if (tcp_rtx_queue_empty(sk))
2234 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2235 /* It is possible TX completion already happened
2236 * before we set TSQ_THROTTLED, so we must
2237 * test again the condition.
2239 smp_mb__after_atomic();
2240 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2246 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2248 const u32 now = tcp_jiffies32;
2249 enum tcp_chrono old = tp->chrono_type;
2251 if (old > TCP_CHRONO_UNSPEC)
2252 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2253 tp->chrono_start = now;
2254 tp->chrono_type = new;
2257 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2259 struct tcp_sock *tp = tcp_sk(sk);
2261 /* If there are multiple conditions worthy of tracking in a
2262 * chronograph then the highest priority enum takes precedence
2263 * over the other conditions. So that if something "more interesting"
2264 * starts happening, stop the previous chrono and start a new one.
2266 if (type > tp->chrono_type)
2267 tcp_chrono_set(tp, type);
2270 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2272 struct tcp_sock *tp = tcp_sk(sk);
2275 /* There are multiple conditions worthy of tracking in a
2276 * chronograph, so that the highest priority enum takes
2277 * precedence over the other conditions (see tcp_chrono_start).
2278 * If a condition stops, we only stop chrono tracking if
2279 * it's the "most interesting" or current chrono we are
2280 * tracking and starts busy chrono if we have pending data.
2282 if (tcp_rtx_and_write_queues_empty(sk))
2283 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2284 else if (type == tp->chrono_type)
2285 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2288 /* This routine writes packets to the network. It advances the
2289 * send_head. This happens as incoming acks open up the remote
2292 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2293 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2294 * account rare use of URG, this is not a big flaw.
2296 * Send at most one packet when push_one > 0. Temporarily ignore
2297 * cwnd limit to force at most one packet out when push_one == 2.
2299 * Returns true, if no segments are in flight and we have queued segments,
2300 * but cannot send anything now because of SWS or another problem.
2302 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2303 int push_one, gfp_t gfp)
2305 struct tcp_sock *tp = tcp_sk(sk);
2306 struct sk_buff *skb;
2307 unsigned int tso_segs, sent_pkts;
2310 bool is_cwnd_limited = false, is_rwnd_limited = false;
2316 /* Do MTU probing. */
2317 result = tcp_mtu_probe(sk);
2320 } else if (result > 0) {
2325 max_segs = tcp_tso_segs(sk, mss_now);
2326 tcp_mstamp_refresh(tp);
2327 while ((skb = tcp_send_head(sk))) {
2330 if (tcp_pacing_check(sk))
2333 tso_segs = tcp_init_tso_segs(skb, mss_now);
2336 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2337 /* "skb_mstamp" is used as a start point for the retransmit timer */
2338 tcp_update_skb_after_send(tp, skb);
2339 goto repair; /* Skip network transmission */
2342 cwnd_quota = tcp_cwnd_test(tp, skb);
2345 /* Force out a loss probe pkt. */
2351 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2352 is_rwnd_limited = true;
2356 if (tso_segs == 1) {
2357 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2358 (tcp_skb_is_last(sk, skb) ?
2359 nonagle : TCP_NAGLE_PUSH))))
2363 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2369 if (tso_segs > 1 && !tcp_urg_mode(tp))
2370 limit = tcp_mss_split_point(sk, skb, mss_now,
2376 if (skb->len > limit &&
2377 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2378 skb, limit, mss_now, gfp)))
2381 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2382 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2383 if (tcp_small_queue_check(sk, skb, 0))
2386 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2390 /* Advance the send_head. This one is sent out.
2391 * This call will increment packets_out.
2393 tcp_event_new_data_sent(sk, skb);
2395 tcp_minshall_update(tp, mss_now, skb);
2396 sent_pkts += tcp_skb_pcount(skb);
2402 if (is_rwnd_limited)
2403 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2405 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2407 if (likely(sent_pkts)) {
2408 if (tcp_in_cwnd_reduction(sk))
2409 tp->prr_out += sent_pkts;
2411 /* Send one loss probe per tail loss episode. */
2413 tcp_schedule_loss_probe(sk);
2414 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2415 tcp_cwnd_validate(sk, is_cwnd_limited);
2418 return !tp->packets_out && !tcp_write_queue_empty(sk);
2421 bool tcp_schedule_loss_probe(struct sock *sk)
2423 struct inet_connection_sock *icsk = inet_csk(sk);
2424 struct tcp_sock *tp = tcp_sk(sk);
2425 u32 timeout, rto_delta_us;
2428 /* Don't do any loss probe on a Fast Open connection before 3WHS
2431 if (tp->fastopen_rsk)
2434 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2435 /* Schedule a loss probe in 2*RTT for SACK capable connections
2436 * in Open state, that are either limited by cwnd or application.
2438 if ((early_retrans != 3 && early_retrans != 4) ||
2439 !tp->packets_out || !tcp_is_sack(tp) ||
2440 icsk->icsk_ca_state != TCP_CA_Open)
2443 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2444 !tcp_write_queue_empty(sk))
2447 /* Probe timeout is 2*rtt. Add minimum RTO to account
2448 * for delayed ack when there's one outstanding packet. If no RTT
2449 * sample is available then probe after TCP_TIMEOUT_INIT.
2452 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2453 if (tp->packets_out == 1)
2454 timeout += TCP_RTO_MIN;
2456 timeout += TCP_TIMEOUT_MIN;
2458 timeout = TCP_TIMEOUT_INIT;
2461 /* If the RTO formula yields an earlier time, then use that time. */
2462 rto_delta_us = tcp_rto_delta_us(sk); /* How far in future is RTO? */
2463 if (rto_delta_us > 0)
2464 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2466 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2471 /* Thanks to skb fast clones, we can detect if a prior transmit of
2472 * a packet is still in a qdisc or driver queue.
2473 * In this case, there is very little point doing a retransmit !
2475 static bool skb_still_in_host_queue(const struct sock *sk,
2476 const struct sk_buff *skb)
2478 if (unlikely(skb_fclone_busy(sk, skb))) {
2479 NET_INC_STATS(sock_net(sk),
2480 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2486 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2487 * retransmit the last segment.
2489 void tcp_send_loss_probe(struct sock *sk)
2491 struct tcp_sock *tp = tcp_sk(sk);
2492 struct sk_buff *skb;
2494 int mss = tcp_current_mss(sk);
2496 skb = tcp_send_head(sk);
2497 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2498 pcount = tp->packets_out;
2499 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2500 if (tp->packets_out > pcount)
2504 skb = skb_rb_last(&sk->tcp_rtx_queue);
2506 /* At most one outstanding TLP retransmission. */
2507 if (tp->tlp_high_seq)
2510 /* Retransmit last segment. */
2514 if (skb_still_in_host_queue(sk, skb))
2517 pcount = tcp_skb_pcount(skb);
2518 if (WARN_ON(!pcount))
2521 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2522 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2523 (pcount - 1) * mss, mss,
2526 skb = skb_rb_next(skb);
2529 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2532 if (__tcp_retransmit_skb(sk, skb, 1))
2535 /* Record snd_nxt for loss detection. */
2536 tp->tlp_high_seq = tp->snd_nxt;
2539 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2540 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2541 inet_csk(sk)->icsk_pending = 0;
2546 /* Push out any pending frames which were held back due to
2547 * TCP_CORK or attempt at coalescing tiny packets.
2548 * The socket must be locked by the caller.
2550 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2553 /* If we are closed, the bytes will have to remain here.
2554 * In time closedown will finish, we empty the write queue and
2555 * all will be happy.
2557 if (unlikely(sk->sk_state == TCP_CLOSE))
2560 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2561 sk_gfp_mask(sk, GFP_ATOMIC)))
2562 tcp_check_probe_timer(sk);
2565 /* Send _single_ skb sitting at the send head. This function requires
2566 * true push pending frames to setup probe timer etc.
2568 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2570 struct sk_buff *skb = tcp_send_head(sk);
2572 BUG_ON(!skb || skb->len < mss_now);
2574 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2577 /* This function returns the amount that we can raise the
2578 * usable window based on the following constraints
2580 * 1. The window can never be shrunk once it is offered (RFC 793)
2581 * 2. We limit memory per socket
2584 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2585 * RECV.NEXT + RCV.WIN fixed until:
2586 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2588 * i.e. don't raise the right edge of the window until you can raise
2589 * it at least MSS bytes.
2591 * Unfortunately, the recommended algorithm breaks header prediction,
2592 * since header prediction assumes th->window stays fixed.
2594 * Strictly speaking, keeping th->window fixed violates the receiver
2595 * side SWS prevention criteria. The problem is that under this rule
2596 * a stream of single byte packets will cause the right side of the
2597 * window to always advance by a single byte.
2599 * Of course, if the sender implements sender side SWS prevention
2600 * then this will not be a problem.
2602 * BSD seems to make the following compromise:
2604 * If the free space is less than the 1/4 of the maximum
2605 * space available and the free space is less than 1/2 mss,
2606 * then set the window to 0.
2607 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2608 * Otherwise, just prevent the window from shrinking
2609 * and from being larger than the largest representable value.
2611 * This prevents incremental opening of the window in the regime
2612 * where TCP is limited by the speed of the reader side taking
2613 * data out of the TCP receive queue. It does nothing about
2614 * those cases where the window is constrained on the sender side
2615 * because the pipeline is full.
2617 * BSD also seems to "accidentally" limit itself to windows that are a
2618 * multiple of MSS, at least until the free space gets quite small.
2619 * This would appear to be a side effect of the mbuf implementation.
2620 * Combining these two algorithms results in the observed behavior
2621 * of having a fixed window size at almost all times.
2623 * Below we obtain similar behavior by forcing the offered window to
2624 * a multiple of the mss when it is feasible to do so.
2626 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2627 * Regular options like TIMESTAMP are taken into account.
2629 u32 __tcp_select_window(struct sock *sk)
2631 struct inet_connection_sock *icsk = inet_csk(sk);
2632 struct tcp_sock *tp = tcp_sk(sk);
2633 /* MSS for the peer's data. Previous versions used mss_clamp
2634 * here. I don't know if the value based on our guesses
2635 * of peer's MSS is better for the performance. It's more correct
2636 * but may be worse for the performance because of rcv_mss
2637 * fluctuations. --SAW 1998/11/1
2639 int mss = icsk->icsk_ack.rcv_mss;
2640 int free_space = tcp_space(sk);
2641 int allowed_space = tcp_full_space(sk);
2642 int full_space = min_t(int, tp->window_clamp, allowed_space);
2645 if (unlikely(mss > full_space)) {
2650 if (free_space < (full_space >> 1)) {
2651 icsk->icsk_ack.quick = 0;
2653 if (tcp_under_memory_pressure(sk))
2654 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2657 /* free_space might become our new window, make sure we don't
2658 * increase it due to wscale.
2660 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2662 /* if free space is less than mss estimate, or is below 1/16th
2663 * of the maximum allowed, try to move to zero-window, else
2664 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2665 * new incoming data is dropped due to memory limits.
2666 * With large window, mss test triggers way too late in order
2667 * to announce zero window in time before rmem limit kicks in.
2669 if (free_space < (allowed_space >> 4) || free_space < mss)
2673 if (free_space > tp->rcv_ssthresh)
2674 free_space = tp->rcv_ssthresh;
2676 /* Don't do rounding if we are using window scaling, since the
2677 * scaled window will not line up with the MSS boundary anyway.
2679 if (tp->rx_opt.rcv_wscale) {
2680 window = free_space;
2682 /* Advertise enough space so that it won't get scaled away.
2683 * Import case: prevent zero window announcement if
2684 * 1<<rcv_wscale > mss.
2686 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2688 window = tp->rcv_wnd;
2689 /* Get the largest window that is a nice multiple of mss.
2690 * Window clamp already applied above.
2691 * If our current window offering is within 1 mss of the
2692 * free space we just keep it. This prevents the divide
2693 * and multiply from happening most of the time.
2694 * We also don't do any window rounding when the free space
2697 if (window <= free_space - mss || window > free_space)
2698 window = rounddown(free_space, mss);
2699 else if (mss == full_space &&
2700 free_space > window + (full_space >> 1))
2701 window = free_space;
2707 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2708 const struct sk_buff *next_skb)
2710 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2711 const struct skb_shared_info *next_shinfo =
2712 skb_shinfo(next_skb);
2713 struct skb_shared_info *shinfo = skb_shinfo(skb);
2715 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2716 shinfo->tskey = next_shinfo->tskey;
2717 TCP_SKB_CB(skb)->txstamp_ack |=
2718 TCP_SKB_CB(next_skb)->txstamp_ack;
2722 /* Collapses two adjacent SKB's during retransmission. */
2723 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2725 struct tcp_sock *tp = tcp_sk(sk);
2726 struct sk_buff *next_skb = skb_rb_next(skb);
2727 int skb_size, next_skb_size;
2729 skb_size = skb->len;
2730 next_skb_size = next_skb->len;
2732 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2734 if (next_skb_size) {
2735 if (next_skb_size <= skb_availroom(skb))
2736 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2738 else if (!skb_shift(skb, next_skb, next_skb_size))
2741 tcp_highest_sack_combine(sk, next_skb, skb);
2743 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2744 skb->ip_summed = CHECKSUM_PARTIAL;
2746 if (skb->ip_summed != CHECKSUM_PARTIAL)
2747 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2749 /* Update sequence range on original skb. */
2750 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2752 /* Merge over control information. This moves PSH/FIN etc. over */
2753 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2755 /* All done, get rid of second SKB and account for it so
2756 * packet counting does not break.
2758 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2759 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2761 /* changed transmit queue under us so clear hints */
2762 tcp_clear_retrans_hints_partial(tp);
2763 if (next_skb == tp->retransmit_skb_hint)
2764 tp->retransmit_skb_hint = skb;
2766 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2768 tcp_skb_collapse_tstamp(skb, next_skb);
2770 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2774 /* Check if coalescing SKBs is legal. */
2775 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2777 if (tcp_skb_pcount(skb) > 1)
2779 if (skb_cloned(skb))
2781 /* Some heuristics for collapsing over SACK'd could be invented */
2782 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2788 /* Collapse packets in the retransmit queue to make to create
2789 * less packets on the wire. This is only done on retransmission.
2791 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2794 struct tcp_sock *tp = tcp_sk(sk);
2795 struct sk_buff *skb = to, *tmp;
2798 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2800 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2803 skb_rbtree_walk_from_safe(skb, tmp) {
2804 if (!tcp_can_collapse(sk, skb))
2807 if (!tcp_skb_can_collapse_to(to))
2820 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2823 if (!tcp_collapse_retrans(sk, to))
2828 /* This retransmits one SKB. Policy decisions and retransmit queue
2829 * state updates are done by the caller. Returns non-zero if an
2830 * error occurred which prevented the send.
2832 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2834 struct inet_connection_sock *icsk = inet_csk(sk);
2835 struct tcp_sock *tp = tcp_sk(sk);
2836 unsigned int cur_mss;
2840 /* Inconclusive MTU probe */
2841 if (icsk->icsk_mtup.probe_size)
2842 icsk->icsk_mtup.probe_size = 0;
2844 /* Do not sent more than we queued. 1/4 is reserved for possible
2845 * copying overhead: fragmentation, tunneling, mangling etc.
2847 if (refcount_read(&sk->sk_wmem_alloc) >
2848 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2852 if (skb_still_in_host_queue(sk, skb))
2855 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2856 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2858 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2862 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2863 return -EHOSTUNREACH; /* Routing failure or similar. */
2865 cur_mss = tcp_current_mss(sk);
2867 /* If receiver has shrunk his window, and skb is out of
2868 * new window, do not retransmit it. The exception is the
2869 * case, when window is shrunk to zero. In this case
2870 * our retransmit serves as a zero window probe.
2872 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2873 TCP_SKB_CB(skb)->seq != tp->snd_una)
2876 len = cur_mss * segs;
2877 if (skb->len > len) {
2878 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2879 cur_mss, GFP_ATOMIC))
2880 return -ENOMEM; /* We'll try again later. */
2882 if (skb_unclone(skb, GFP_ATOMIC))
2885 diff = tcp_skb_pcount(skb);
2886 tcp_set_skb_tso_segs(skb, cur_mss);
2887 diff -= tcp_skb_pcount(skb);
2889 tcp_adjust_pcount(sk, skb, diff);
2890 if (skb->len < cur_mss)
2891 tcp_retrans_try_collapse(sk, skb, cur_mss);
2894 /* RFC3168, section 6.1.1.1. ECN fallback */
2895 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2896 tcp_ecn_clear_syn(sk, skb);
2898 /* Update global and local TCP statistics. */
2899 segs = tcp_skb_pcount(skb);
2900 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2901 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2902 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2903 tp->total_retrans += segs;
2905 /* make sure skb->data is aligned on arches that require it
2906 * and check if ack-trimming & collapsing extended the headroom
2907 * beyond what csum_start can cover.
2909 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2910 skb_headroom(skb) >= 0xFFFF)) {
2911 struct sk_buff *nskb;
2913 tcp_skb_tsorted_save(skb) {
2914 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2915 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2917 } tcp_skb_tsorted_restore(skb);
2920 tcp_update_skb_after_send(tp, skb);
2922 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2926 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2927 trace_tcp_retransmit_skb(sk, skb);
2928 } else if (err != -EBUSY) {
2929 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2934 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 int err = __tcp_retransmit_skb(sk, skb, segs);
2940 #if FASTRETRANS_DEBUG > 0
2941 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2942 net_dbg_ratelimited("retrans_out leaked\n");
2945 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2946 tp->retrans_out += tcp_skb_pcount(skb);
2948 /* Save stamp of the first retransmit. */
2949 if (!tp->retrans_stamp)
2950 tp->retrans_stamp = tcp_skb_timestamp(skb);
2954 if (tp->undo_retrans < 0)
2955 tp->undo_retrans = 0;
2956 tp->undo_retrans += tcp_skb_pcount(skb);
2960 /* This gets called after a retransmit timeout, and the initially
2961 * retransmitted data is acknowledged. It tries to continue
2962 * resending the rest of the retransmit queue, until either
2963 * we've sent it all or the congestion window limit is reached.
2964 * If doing SACK, the first ACK which comes back for a timeout
2965 * based retransmit packet might feed us FACK information again.
2966 * If so, we use it to avoid unnecessarily retransmissions.
2968 void tcp_xmit_retransmit_queue(struct sock *sk)
2970 const struct inet_connection_sock *icsk = inet_csk(sk);
2971 struct sk_buff *skb, *rtx_head, *hole = NULL;
2972 struct tcp_sock *tp = tcp_sk(sk);
2976 if (!tp->packets_out)
2979 rtx_head = tcp_rtx_queue_head(sk);
2980 skb = tp->retransmit_skb_hint ?: rtx_head;
2981 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2982 skb_rbtree_walk_from(skb) {
2986 if (tcp_pacing_check(sk))
2989 /* we could do better than to assign each time */
2991 tp->retransmit_skb_hint = skb;
2993 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2996 sacked = TCP_SKB_CB(skb)->sacked;
2997 /* In case tcp_shift_skb_data() have aggregated large skbs,
2998 * we need to make sure not sending too bigs TSO packets
3000 segs = min_t(int, segs, max_segs);
3002 if (tp->retrans_out >= tp->lost_out) {
3004 } else if (!(sacked & TCPCB_LOST)) {
3005 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3010 if (icsk->icsk_ca_state != TCP_CA_Loss)
3011 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3013 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3016 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3019 if (tcp_small_queue_check(sk, skb, 1))
3022 if (tcp_retransmit_skb(sk, skb, segs))
3025 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3027 if (tcp_in_cwnd_reduction(sk))
3028 tp->prr_out += tcp_skb_pcount(skb);
3030 if (skb == rtx_head &&
3031 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3032 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3033 inet_csk(sk)->icsk_rto,
3038 /* We allow to exceed memory limits for FIN packets to expedite
3039 * connection tear down and (memory) recovery.
3040 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3041 * or even be forced to close flow without any FIN.
3042 * In general, we want to allow one skb per socket to avoid hangs
3043 * with edge trigger epoll()
3045 void sk_forced_mem_schedule(struct sock *sk, int size)
3049 if (size <= sk->sk_forward_alloc)
3051 amt = sk_mem_pages(size);
3052 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3053 sk_memory_allocated_add(sk, amt);
3055 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3056 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3059 /* Send a FIN. The caller locks the socket for us.
3060 * We should try to send a FIN packet really hard, but eventually give up.
3062 void tcp_send_fin(struct sock *sk)
3064 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3065 struct tcp_sock *tp = tcp_sk(sk);
3067 /* Optimization, tack on the FIN if we have one skb in write queue and
3068 * this skb was not yet sent, or we are under memory pressure.
3069 * Note: in the latter case, FIN packet will be sent after a timeout,
3070 * as TCP stack thinks it has already been transmitted.
3072 if (!tskb && tcp_under_memory_pressure(sk))
3073 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3077 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3078 TCP_SKB_CB(tskb)->end_seq++;
3080 if (tcp_write_queue_empty(sk)) {
3081 /* This means tskb was already sent.
3082 * Pretend we included the FIN on previous transmit.
3083 * We need to set tp->snd_nxt to the value it would have
3084 * if FIN had been sent. This is because retransmit path
3085 * does not change tp->snd_nxt.
3091 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3092 if (unlikely(!skb)) {
3097 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3098 skb_reserve(skb, MAX_TCP_HEADER);
3099 sk_forced_mem_schedule(sk, skb->truesize);
3100 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3101 tcp_init_nondata_skb(skb, tp->write_seq,
3102 TCPHDR_ACK | TCPHDR_FIN);
3103 tcp_queue_skb(sk, skb);
3105 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3108 /* We get here when a process closes a file descriptor (either due to
3109 * an explicit close() or as a byproduct of exit()'ing) and there
3110 * was unread data in the receive queue. This behavior is recommended
3111 * by RFC 2525, section 2.17. -DaveM
3113 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3115 struct sk_buff *skb;
3117 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3119 /* NOTE: No TCP options attached and we never retransmit this. */
3120 skb = alloc_skb(MAX_TCP_HEADER, priority);
3122 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3126 /* Reserve space for headers and prepare control bits. */
3127 skb_reserve(skb, MAX_TCP_HEADER);
3128 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3129 TCPHDR_ACK | TCPHDR_RST);
3130 tcp_mstamp_refresh(tcp_sk(sk));
3132 if (tcp_transmit_skb(sk, skb, 0, priority))
3133 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3135 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3136 * skb here is different to the troublesome skb, so use NULL
3138 trace_tcp_send_reset(sk, NULL);
3141 /* Send a crossed SYN-ACK during socket establishment.
3142 * WARNING: This routine must only be called when we have already sent
3143 * a SYN packet that crossed the incoming SYN that caused this routine
3144 * to get called. If this assumption fails then the initial rcv_wnd
3145 * and rcv_wscale values will not be correct.
3147 int tcp_send_synack(struct sock *sk)
3149 struct sk_buff *skb;
3151 skb = tcp_rtx_queue_head(sk);
3152 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3153 pr_err("%s: wrong queue state\n", __func__);
3156 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3157 if (skb_cloned(skb)) {
3158 struct sk_buff *nskb;
3160 tcp_skb_tsorted_save(skb) {
3161 nskb = skb_copy(skb, GFP_ATOMIC);
3162 } tcp_skb_tsorted_restore(skb);
3165 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3166 tcp_rtx_queue_unlink_and_free(skb, sk);
3167 __skb_header_release(nskb);
3168 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3169 sk->sk_wmem_queued += nskb->truesize;
3170 sk_mem_charge(sk, nskb->truesize);
3174 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3175 tcp_ecn_send_synack(sk, skb);
3177 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3181 * tcp_make_synack - Prepare a SYN-ACK.
3182 * sk: listener socket
3183 * dst: dst entry attached to the SYNACK
3184 * req: request_sock pointer
3186 * Allocate one skb and build a SYNACK packet.
3187 * @dst is consumed : Caller should not use it again.
3189 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3190 struct request_sock *req,
3191 struct tcp_fastopen_cookie *foc,
3192 enum tcp_synack_type synack_type)
3194 struct inet_request_sock *ireq = inet_rsk(req);
3195 const struct tcp_sock *tp = tcp_sk(sk);
3196 struct tcp_md5sig_key *md5 = NULL;
3197 struct tcp_out_options opts;
3198 struct sk_buff *skb;
3199 int tcp_header_size;
3203 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3204 if (unlikely(!skb)) {
3208 /* Reserve space for headers. */
3209 skb_reserve(skb, MAX_TCP_HEADER);
3211 switch (synack_type) {
3212 case TCP_SYNACK_NORMAL:
3213 skb_set_owner_w(skb, req_to_sk(req));
3215 case TCP_SYNACK_COOKIE:
3216 /* Under synflood, we do not attach skb to a socket,
3217 * to avoid false sharing.
3220 case TCP_SYNACK_FASTOPEN:
3221 /* sk is a const pointer, because we want to express multiple
3222 * cpu might call us concurrently.
3223 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3225 skb_set_owner_w(skb, (struct sock *)sk);
3228 skb_dst_set(skb, dst);
3230 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3232 memset(&opts, 0, sizeof(opts));
3233 #ifdef CONFIG_SYN_COOKIES
3234 if (unlikely(req->cookie_ts))
3235 skb->skb_mstamp = cookie_init_timestamp(req);
3238 skb->skb_mstamp = tcp_clock_us();
3240 #ifdef CONFIG_TCP_MD5SIG
3242 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3244 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3245 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3248 skb_push(skb, tcp_header_size);
3249 skb_reset_transport_header(skb);
3251 th = (struct tcphdr *)skb->data;
3252 memset(th, 0, sizeof(struct tcphdr));
3255 tcp_ecn_make_synack(req, th);
3256 th->source = htons(ireq->ir_num);
3257 th->dest = ireq->ir_rmt_port;
3258 skb->mark = ireq->ir_mark;
3259 /* Setting of flags are superfluous here for callers (and ECE is
3260 * not even correctly set)
3262 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3263 TCPHDR_SYN | TCPHDR_ACK);
3265 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3266 /* XXX data is queued and acked as is. No buffer/window check */
3267 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3269 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3270 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3271 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3272 th->doff = (tcp_header_size >> 2);
3273 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3275 #ifdef CONFIG_TCP_MD5SIG
3276 /* Okay, we have all we need - do the md5 hash if needed */
3278 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3279 md5, req_to_sk(req), skb);
3283 /* Do not fool tcpdump (if any), clean our debris */
3287 EXPORT_SYMBOL(tcp_make_synack);
3289 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3291 struct inet_connection_sock *icsk = inet_csk(sk);
3292 const struct tcp_congestion_ops *ca;
3293 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3295 if (ca_key == TCP_CA_UNSPEC)
3299 ca = tcp_ca_find_key(ca_key);
3300 if (likely(ca && try_module_get(ca->owner))) {
3301 module_put(icsk->icsk_ca_ops->owner);
3302 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3303 icsk->icsk_ca_ops = ca;
3308 /* Do all connect socket setups that can be done AF independent. */
3309 static void tcp_connect_init(struct sock *sk)
3311 const struct dst_entry *dst = __sk_dst_get(sk);
3312 struct tcp_sock *tp = tcp_sk(sk);
3316 /* We'll fix this up when we get a response from the other end.
3317 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3319 tp->tcp_header_len = sizeof(struct tcphdr);
3320 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3321 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3323 #ifdef CONFIG_TCP_MD5SIG
3324 if (tp->af_specific->md5_lookup(sk, sk))
3325 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3328 /* If user gave his TCP_MAXSEG, record it to clamp */
3329 if (tp->rx_opt.user_mss)
3330 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3333 tcp_sync_mss(sk, dst_mtu(dst));
3335 tcp_ca_dst_init(sk, dst);
3337 if (!tp->window_clamp)
3338 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3339 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3341 tcp_initialize_rcv_mss(sk);
3343 /* limit the window selection if the user enforce a smaller rx buffer */
3344 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3345 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3346 tp->window_clamp = tcp_full_space(sk);
3348 rcv_wnd = tcp_rwnd_init_bpf(sk);
3350 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3352 tcp_select_initial_window(tcp_full_space(sk),
3353 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3356 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3360 tp->rx_opt.rcv_wscale = rcv_wscale;
3361 tp->rcv_ssthresh = tp->rcv_wnd;
3364 sock_reset_flag(sk, SOCK_DONE);
3367 tp->snd_una = tp->write_seq;
3368 tp->snd_sml = tp->write_seq;
3369 tp->snd_up = tp->write_seq;
3370 tp->snd_nxt = tp->write_seq;
3372 if (likely(!tp->repair))
3375 tp->rcv_tstamp = tcp_jiffies32;
3376 tp->rcv_wup = tp->rcv_nxt;
3377 tp->copied_seq = tp->rcv_nxt;
3379 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3380 inet_csk(sk)->icsk_retransmits = 0;
3381 tcp_clear_retrans(tp);
3384 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3386 struct tcp_sock *tp = tcp_sk(sk);
3387 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3389 tcb->end_seq += skb->len;
3390 __skb_header_release(skb);
3391 sk->sk_wmem_queued += skb->truesize;
3392 sk_mem_charge(sk, skb->truesize);
3393 tp->write_seq = tcb->end_seq;
3394 tp->packets_out += tcp_skb_pcount(skb);
3397 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3398 * queue a data-only packet after the regular SYN, such that regular SYNs
3399 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3400 * only the SYN sequence, the data are retransmitted in the first ACK.
3401 * If cookie is not cached or other error occurs, falls back to send a
3402 * regular SYN with Fast Open cookie request option.
3404 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3406 struct tcp_sock *tp = tcp_sk(sk);
3407 struct tcp_fastopen_request *fo = tp->fastopen_req;
3409 struct sk_buff *syn_data;
3411 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3412 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3415 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3416 * user-MSS. Reserve maximum option space for middleboxes that add
3417 * private TCP options. The cost is reduced data space in SYN :(
3419 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3421 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3422 MAX_TCP_OPTION_SPACE;
3424 space = min_t(size_t, space, fo->size);
3426 /* limit to order-0 allocations */
3427 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3429 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3432 syn_data->ip_summed = CHECKSUM_PARTIAL;
3433 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3435 int copied = copy_from_iter(skb_put(syn_data, space), space,
3436 &fo->data->msg_iter);
3437 if (unlikely(!copied)) {
3438 tcp_skb_tsorted_anchor_cleanup(syn_data);
3439 kfree_skb(syn_data);
3442 if (copied != space) {
3443 skb_trim(syn_data, copied);
3447 /* No more data pending in inet_wait_for_connect() */
3448 if (space == fo->size)
3452 tcp_connect_queue_skb(sk, syn_data);
3454 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3456 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3458 syn->skb_mstamp = syn_data->skb_mstamp;
3460 /* Now full SYN+DATA was cloned and sent (or not),
3461 * remove the SYN from the original skb (syn_data)
3462 * we keep in write queue in case of a retransmit, as we
3463 * also have the SYN packet (with no data) in the same queue.
3465 TCP_SKB_CB(syn_data)->seq++;
3466 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3468 tp->syn_data = (fo->copied > 0);
3469 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3470 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3474 /* data was not sent, put it in write_queue */
3475 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3476 tp->packets_out -= tcp_skb_pcount(syn_data);
3479 /* Send a regular SYN with Fast Open cookie request option */
3480 if (fo->cookie.len > 0)
3482 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3484 tp->syn_fastopen = 0;
3486 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3490 /* Build a SYN and send it off. */
3491 int tcp_connect(struct sock *sk)
3493 struct tcp_sock *tp = tcp_sk(sk);
3494 struct sk_buff *buff;
3497 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3499 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3500 return -EHOSTUNREACH; /* Routing failure or similar. */
3502 tcp_connect_init(sk);
3504 if (unlikely(tp->repair)) {
3505 tcp_finish_connect(sk, NULL);
3509 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3510 if (unlikely(!buff))
3513 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3514 tcp_mstamp_refresh(tp);
3515 tp->retrans_stamp = tcp_time_stamp(tp);
3516 tcp_connect_queue_skb(sk, buff);
3517 tcp_ecn_send_syn(sk, buff);
3518 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3520 /* Send off SYN; include data in Fast Open. */
3521 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3522 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3523 if (err == -ECONNREFUSED)
3526 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3527 * in order to make this packet get counted in tcpOutSegs.
3529 tp->snd_nxt = tp->write_seq;
3530 tp->pushed_seq = tp->write_seq;
3531 buff = tcp_send_head(sk);
3532 if (unlikely(buff)) {
3533 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3534 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3536 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3538 /* Timer for repeating the SYN until an answer. */
3539 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3540 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3543 EXPORT_SYMBOL(tcp_connect);
3545 /* Send out a delayed ack, the caller does the policy checking
3546 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3549 void tcp_send_delayed_ack(struct sock *sk)
3551 struct inet_connection_sock *icsk = inet_csk(sk);
3552 int ato = icsk->icsk_ack.ato;
3553 unsigned long timeout;
3555 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3557 if (ato > TCP_DELACK_MIN) {
3558 const struct tcp_sock *tp = tcp_sk(sk);
3559 int max_ato = HZ / 2;
3561 if (icsk->icsk_ack.pingpong ||
3562 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3563 max_ato = TCP_DELACK_MAX;
3565 /* Slow path, intersegment interval is "high". */
3567 /* If some rtt estimate is known, use it to bound delayed ack.
3568 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3572 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3579 ato = min(ato, max_ato);
3582 /* Stay within the limit we were given */
3583 timeout = jiffies + ato;
3585 /* Use new timeout only if there wasn't a older one earlier. */
3586 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3587 /* If delack timer was blocked or is about to expire,
3590 if (icsk->icsk_ack.blocked ||
3591 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3596 if (!time_before(timeout, icsk->icsk_ack.timeout))
3597 timeout = icsk->icsk_ack.timeout;
3599 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3600 icsk->icsk_ack.timeout = timeout;
3601 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3604 /* This routine sends an ack and also updates the window. */
3605 void tcp_send_ack(struct sock *sk)
3607 struct sk_buff *buff;
3609 /* If we have been reset, we may not send again. */
3610 if (sk->sk_state == TCP_CLOSE)
3613 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3615 /* We are not putting this on the write queue, so
3616 * tcp_transmit_skb() will set the ownership to this
3619 buff = alloc_skb(MAX_TCP_HEADER,
3620 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3621 if (unlikely(!buff)) {
3622 inet_csk_schedule_ack(sk);
3623 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3624 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3625 TCP_DELACK_MAX, TCP_RTO_MAX);
3629 /* Reserve space for headers and prepare control bits. */
3630 skb_reserve(buff, MAX_TCP_HEADER);
3631 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3633 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3635 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3637 skb_set_tcp_pure_ack(buff);
3639 /* Send it off, this clears delayed acks for us. */
3640 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3642 EXPORT_SYMBOL_GPL(tcp_send_ack);
3644 /* This routine sends a packet with an out of date sequence
3645 * number. It assumes the other end will try to ack it.
3647 * Question: what should we make while urgent mode?
3648 * 4.4BSD forces sending single byte of data. We cannot send
3649 * out of window data, because we have SND.NXT==SND.MAX...
3651 * Current solution: to send TWO zero-length segments in urgent mode:
3652 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3653 * out-of-date with SND.UNA-1 to probe window.
3655 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3657 struct tcp_sock *tp = tcp_sk(sk);
3658 struct sk_buff *skb;
3660 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3661 skb = alloc_skb(MAX_TCP_HEADER,
3662 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3666 /* Reserve space for headers and set control bits. */
3667 skb_reserve(skb, MAX_TCP_HEADER);
3668 /* Use a previous sequence. This should cause the other
3669 * end to send an ack. Don't queue or clone SKB, just
3672 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3673 NET_INC_STATS(sock_net(sk), mib);
3674 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3677 /* Called from setsockopt( ... TCP_REPAIR ) */
3678 void tcp_send_window_probe(struct sock *sk)
3680 if (sk->sk_state == TCP_ESTABLISHED) {
3681 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3682 tcp_mstamp_refresh(tcp_sk(sk));
3683 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3687 /* Initiate keepalive or window probe from timer. */
3688 int tcp_write_wakeup(struct sock *sk, int mib)
3690 struct tcp_sock *tp = tcp_sk(sk);
3691 struct sk_buff *skb;
3693 if (sk->sk_state == TCP_CLOSE)
3696 skb = tcp_send_head(sk);
3697 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3699 unsigned int mss = tcp_current_mss(sk);
3700 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3702 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3703 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3705 /* We are probing the opening of a window
3706 * but the window size is != 0
3707 * must have been a result SWS avoidance ( sender )
3709 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3711 seg_size = min(seg_size, mss);
3712 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3713 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3714 skb, seg_size, mss, GFP_ATOMIC))
3716 } else if (!tcp_skb_pcount(skb))
3717 tcp_set_skb_tso_segs(skb, mss);
3719 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3720 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3722 tcp_event_new_data_sent(sk, skb);
3725 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3726 tcp_xmit_probe_skb(sk, 1, mib);
3727 return tcp_xmit_probe_skb(sk, 0, mib);
3731 /* A window probe timeout has occurred. If window is not closed send
3732 * a partial packet else a zero probe.
3734 void tcp_send_probe0(struct sock *sk)
3736 struct inet_connection_sock *icsk = inet_csk(sk);
3737 struct tcp_sock *tp = tcp_sk(sk);
3738 struct net *net = sock_net(sk);
3739 unsigned long probe_max;
3742 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3744 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3745 /* Cancel probe timer, if it is not required. */
3746 icsk->icsk_probes_out = 0;
3747 icsk->icsk_backoff = 0;
3752 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3753 icsk->icsk_backoff++;
3754 icsk->icsk_probes_out++;
3755 probe_max = TCP_RTO_MAX;
3757 /* If packet was not sent due to local congestion,
3758 * do not backoff and do not remember icsk_probes_out.
3759 * Let local senders to fight for local resources.
3761 * Use accumulated backoff yet.
3763 if (!icsk->icsk_probes_out)
3764 icsk->icsk_probes_out = 1;
3765 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3767 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3768 tcp_probe0_when(sk, probe_max),
3772 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3774 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3778 tcp_rsk(req)->txhash = net_tx_rndhash();
3779 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3781 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3782 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3783 if (unlikely(tcp_passive_fastopen(sk)))
3784 tcp_sk(sk)->total_retrans++;
3788 EXPORT_SYMBOL(tcp_rtx_synack);