2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
6 * Implementation of the Transmission Control Protocol(TCP).
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
37 #define pr_fmt(fmt) "TCP: " fmt
41 #include <linux/compiler.h>
42 #include <linux/gfp.h>
43 #include <linux/module.h>
45 /* People can turn this off for buggy TCP's found in printers etc. */
46 int sysctl_tcp_retrans_collapse __read_mostly = 1;
48 /* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
53 /* Default TSQ limit of four TSO segments */
54 int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
56 /* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
62 /* By default, RFC2861 behavior. */
63 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
65 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
68 /* Account for new data that has been sent to the network. */
69 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
82 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
86 /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
87 * window scaling factor due to loss of precision.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
93 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95 const struct tcp_sock *tp = tcp_sk(sk);
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
98 (tp->rx_opt.wscale_ok &&
99 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tcp_wnd_end(tp);
105 /* Calculate mss to advertise in SYN segment.
106 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 * 1. It is independent of path mtu.
109 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
110 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
111 * attached devices, because some buggy hosts are confused by
113 * 4. We do not make 3, we advertise MSS, calculated from first
114 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
115 * This may be overridden via information stored in routing table.
116 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
117 * probably even Jumbo".
119 static __u16 tcp_advertise_mss(struct sock *sk)
121 struct tcp_sock *tp = tcp_sk(sk);
122 const struct dst_entry *dst = __sk_dst_get(sk);
123 int mss = tp->advmss;
126 unsigned int metric = dst_metric_advmss(dst);
137 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
138 * This is the first part of cwnd validation mechanism.
140 void tcp_cwnd_restart(struct sock *sk, s32 delta)
142 struct tcp_sock *tp = tcp_sk(sk);
143 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
144 u32 cwnd = tp->snd_cwnd;
146 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
148 tp->snd_ssthresh = tcp_current_ssthresh(sk);
149 restart_cwnd = min(restart_cwnd, cwnd);
151 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
153 tp->snd_cwnd = max(cwnd, restart_cwnd);
154 tp->snd_cwnd_stamp = tcp_jiffies32;
155 tp->snd_cwnd_used = 0;
158 /* Congestion state accounting after a packet has been sent. */
159 static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct inet_connection_sock *icsk = inet_csk(sk);
163 const u32 now = tcp_jiffies32;
165 if (tcp_packets_in_flight(tp) == 0)
166 tcp_ca_event(sk, CA_EVENT_TX_START);
170 /* If it is a reply for ato after last received
171 * packet, enter pingpong mode.
173 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
174 icsk->icsk_ack.pingpong = 1;
177 /* Account for an ACK we sent. */
178 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 struct tcp_sock *tp = tcp_sk(sk);
183 if (unlikely(rcv_nxt != tp->rcv_nxt))
184 return; /* Special ACK sent by DCTCP to reflect ECN */
185 tcp_dec_quickack_mode(sk, pkts);
186 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190 u32 tcp_default_init_rwnd(u32 mss)
192 /* Initial receive window should be twice of TCP_INIT_CWND to
193 * enable proper sending of new unsent data during fast recovery
194 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
195 * limit when mss is larger than 1460.
197 u32 init_rwnd = TCP_INIT_CWND * 2;
200 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 /* Determine a window scaling and initial window to offer.
205 * Based on the assumption that the given amount of space
206 * will be offered. Store the results in the tp structure.
207 * NOTE: for smooth operation initial space offering should
208 * be a multiple of mss if possible. We assume here that mss >= 1.
209 * This MUST be enforced by all callers.
211 void tcp_select_initial_window(int __space, __u32 mss,
212 __u32 *rcv_wnd, __u32 *window_clamp,
213 int wscale_ok, __u8 *rcv_wscale,
216 unsigned int space = (__space < 0 ? 0 : __space);
218 /* If no clamp set the clamp to the max possible scaled window */
219 if (*window_clamp == 0)
220 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
221 space = min(*window_clamp, space);
223 /* Quantize space offering to a multiple of mss if possible. */
225 space = rounddown(space, mss);
227 /* NOTE: offering an initial window larger than 32767
228 * will break some buggy TCP stacks. If the admin tells us
229 * it is likely we could be speaking with such a buggy stack
230 * we will truncate our initial window offering to 32K-1
231 * unless the remote has sent us a window scaling option,
232 * which we interpret as a sign the remote TCP is not
233 * misinterpreting the window field as a signed quantity.
235 if (sysctl_tcp_workaround_signed_windows)
236 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
252 if (mss > (1 << *rcv_wscale)) {
253 if (!init_rcv_wnd) /* Use default unless specified otherwise */
254 init_rcv_wnd = tcp_default_init_rwnd(mss);
255 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
258 /* Set the clamp no higher than max representable value */
259 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
261 EXPORT_SYMBOL(tcp_select_initial_window);
263 /* Chose a new window to advertise, update state in tcp_sock for the
264 * socket, and return result with RFC1323 scaling applied. The return
265 * value can be stuffed directly into th->window for an outgoing
268 static u16 tcp_select_window(struct sock *sk)
270 struct tcp_sock *tp = tcp_sk(sk);
271 u32 old_win = tp->rcv_wnd;
272 u32 cur_win = tcp_receive_window(tp);
273 u32 new_win = __tcp_select_window(sk);
275 /* Never shrink the offered window */
276 if (new_win < cur_win) {
277 /* Danger Will Robinson!
278 * Don't update rcv_wup/rcv_wnd here or else
279 * we will not be able to advertise a zero
280 * window in time. --DaveM
282 * Relax Will Robinson.
285 NET_INC_STATS(sock_net(sk),
286 LINUX_MIB_TCPWANTZEROWINDOWADV);
287 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
289 tp->rcv_wnd = new_win;
290 tp->rcv_wup = tp->rcv_nxt;
292 /* Make sure we do not exceed the maximum possible
295 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
296 new_win = min(new_win, MAX_TCP_WINDOW);
298 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
300 /* RFC1323 scaling applied */
301 new_win >>= tp->rx_opt.rcv_wscale;
303 /* If we advertise zero window, disable fast path. */
307 NET_INC_STATS(sock_net(sk),
308 LINUX_MIB_TCPTOZEROWINDOWADV);
309 } else if (old_win == 0) {
310 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
316 /* Packet ECN state for a SYN-ACK */
317 static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
319 const struct tcp_sock *tp = tcp_sk(sk);
321 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
322 if (!(tp->ecn_flags & TCP_ECN_OK))
323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
324 else if (tcp_ca_needs_ecn(sk) ||
325 tcp_bpf_ca_needs_ecn(sk))
329 /* Packet ECN state for a SYN. */
330 static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
332 struct tcp_sock *tp = tcp_sk(sk);
333 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
334 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
335 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
338 const struct dst_entry *dst = __sk_dst_get(sk);
340 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
347 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
348 tp->ecn_flags = TCP_ECN_OK;
349 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
354 static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
356 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
357 /* tp->ecn_flags are cleared at a later point in time when
358 * SYN ACK is ultimatively being received.
360 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
364 tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
366 if (inet_rsk(req)->ecn_ok)
370 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
373 static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
374 struct tcphdr *th, int tcp_header_len)
376 struct tcp_sock *tp = tcp_sk(sk);
378 if (tp->ecn_flags & TCP_ECN_OK) {
379 /* Not-retransmitted data segment: set ECT and inject CWR. */
380 if (skb->len != tcp_header_len &&
381 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
383 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
384 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
386 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
388 } else if (!tcp_ca_needs_ecn(sk)) {
389 /* ACK or retransmitted segment: clear ECT|CE */
390 INET_ECN_dontxmit(sk);
392 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
397 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
398 * auto increment end seqno.
400 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
402 skb->ip_summed = CHECKSUM_PARTIAL;
405 TCP_SKB_CB(skb)->tcp_flags = flags;
406 TCP_SKB_CB(skb)->sacked = 0;
408 tcp_skb_pcount_set(skb, 1);
410 TCP_SKB_CB(skb)->seq = seq;
411 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
413 TCP_SKB_CB(skb)->end_seq = seq;
416 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
418 return tp->snd_una != tp->snd_up;
421 #define OPTION_SACK_ADVERTISE (1 << 0)
422 #define OPTION_TS (1 << 1)
423 #define OPTION_MD5 (1 << 2)
424 #define OPTION_WSCALE (1 << 3)
425 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
427 struct tcp_out_options {
428 u16 options; /* bit field of OPTION_* */
429 u16 mss; /* 0 to disable */
430 u8 ws; /* window scale, 0 to disable */
431 u8 num_sack_blocks; /* number of SACK blocks to include */
432 u8 hash_size; /* bytes in hash_location */
433 __u8 *hash_location; /* temporary pointer, overloaded */
434 __u32 tsval, tsecr; /* need to include OPTION_TS */
435 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
438 /* Write previously computed TCP options to the packet.
440 * Beware: Something in the Internet is very sensitive to the ordering of
441 * TCP options, we learned this through the hard way, so be careful here.
442 * Luckily we can at least blame others for their non-compliance but from
443 * inter-operability perspective it seems that we're somewhat stuck with
444 * the ordering which we have been using if we want to keep working with
445 * those broken things (not that it currently hurts anybody as there isn't
446 * particular reason why the ordering would need to be changed).
448 * At least SACK_PERM as the first option is known to lead to a disaster
449 * (but it may well be that other scenarios fail similarly).
451 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
452 struct tcp_out_options *opts)
454 u16 options = opts->options; /* mungable copy */
456 if (unlikely(OPTION_MD5 & options)) {
457 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
458 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
459 /* overload cookie hash location */
460 opts->hash_location = (__u8 *)ptr;
464 if (unlikely(opts->mss)) {
465 *ptr++ = htonl((TCPOPT_MSS << 24) |
466 (TCPOLEN_MSS << 16) |
470 if (likely(OPTION_TS & options)) {
471 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
472 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
473 (TCPOLEN_SACK_PERM << 16) |
474 (TCPOPT_TIMESTAMP << 8) |
476 options &= ~OPTION_SACK_ADVERTISE;
478 *ptr++ = htonl((TCPOPT_NOP << 24) |
480 (TCPOPT_TIMESTAMP << 8) |
483 *ptr++ = htonl(opts->tsval);
484 *ptr++ = htonl(opts->tsecr);
487 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_SACK_PERM << 8) |
494 if (unlikely(OPTION_WSCALE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
496 (TCPOPT_WINDOW << 16) |
497 (TCPOLEN_WINDOW << 8) |
501 if (unlikely(opts->num_sack_blocks)) {
502 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
503 tp->duplicate_sack : tp->selective_acks;
506 *ptr++ = htonl((TCPOPT_NOP << 24) |
509 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
510 TCPOLEN_SACK_PERBLOCK)));
512 for (this_sack = 0; this_sack < opts->num_sack_blocks;
514 *ptr++ = htonl(sp[this_sack].start_seq);
515 *ptr++ = htonl(sp[this_sack].end_seq);
518 tp->rx_opt.dsack = 0;
521 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
522 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
524 u32 len; /* Fast Open option length */
527 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
528 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
529 TCPOPT_FASTOPEN_MAGIC);
530 p += TCPOLEN_EXP_FASTOPEN_BASE;
532 len = TCPOLEN_FASTOPEN_BASE + foc->len;
533 *p++ = TCPOPT_FASTOPEN;
537 memcpy(p, foc->val, foc->len);
538 if ((len & 3) == 2) {
539 p[foc->len] = TCPOPT_NOP;
540 p[foc->len + 1] = TCPOPT_NOP;
542 ptr += (len + 3) >> 2;
546 /* Compute TCP options for SYN packets. This is not the final
547 * network wire format yet.
549 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
550 struct tcp_out_options *opts,
551 struct tcp_md5sig_key **md5)
553 struct tcp_sock *tp = tcp_sk(sk);
554 unsigned int remaining = MAX_TCP_OPTION_SPACE;
555 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
557 #ifdef CONFIG_TCP_MD5SIG
558 *md5 = tp->af_specific->md5_lookup(sk, sk);
560 opts->options |= OPTION_MD5;
561 remaining -= TCPOLEN_MD5SIG_ALIGNED;
567 /* We always get an MSS option. The option bytes which will be seen in
568 * normal data packets should timestamps be used, must be in the MSS
569 * advertised. But we subtract them from tp->mss_cache so that
570 * calculations in tcp_sendmsg are simpler etc. So account for this
571 * fact here if necessary. If we don't do this correctly, as a
572 * receiver we won't recognize data packets as being full sized when we
573 * should, and thus we won't abide by the delayed ACK rules correctly.
574 * SACKs don't matter, we never delay an ACK when we have any of those
576 opts->mss = tcp_advertise_mss(sk);
577 remaining -= TCPOLEN_MSS_ALIGNED;
579 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
580 opts->options |= OPTION_TS;
581 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
582 opts->tsecr = tp->rx_opt.ts_recent;
583 remaining -= TCPOLEN_TSTAMP_ALIGNED;
585 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
586 opts->ws = tp->rx_opt.rcv_wscale;
587 opts->options |= OPTION_WSCALE;
588 remaining -= TCPOLEN_WSCALE_ALIGNED;
590 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
591 opts->options |= OPTION_SACK_ADVERTISE;
592 if (unlikely(!(OPTION_TS & opts->options)))
593 remaining -= TCPOLEN_SACKPERM_ALIGNED;
596 if (fastopen && fastopen->cookie.len >= 0) {
597 u32 need = fastopen->cookie.len;
599 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
600 TCPOLEN_FASTOPEN_BASE;
601 need = (need + 3) & ~3U; /* Align to 32 bits */
602 if (remaining >= need) {
603 opts->options |= OPTION_FAST_OPEN_COOKIE;
604 opts->fastopen_cookie = &fastopen->cookie;
606 tp->syn_fastopen = 1;
607 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
611 return MAX_TCP_OPTION_SPACE - remaining;
614 /* Set up TCP options for SYN-ACKs. */
615 static unsigned int tcp_synack_options(struct request_sock *req,
616 unsigned int mss, struct sk_buff *skb,
617 struct tcp_out_options *opts,
618 const struct tcp_md5sig_key *md5,
619 struct tcp_fastopen_cookie *foc,
620 enum tcp_synack_type synack_type)
622 struct inet_request_sock *ireq = inet_rsk(req);
623 unsigned int remaining = MAX_TCP_OPTION_SPACE;
625 #ifdef CONFIG_TCP_MD5SIG
627 opts->options |= OPTION_MD5;
628 remaining -= TCPOLEN_MD5SIG_ALIGNED;
630 /* We can't fit any SACK blocks in a packet with MD5 + TS
631 * options. There was discussion about disabling SACK
632 * rather than TS in order to fit in better with old,
633 * buggy kernels, but that was deemed to be unnecessary.
635 if (synack_type != TCP_SYNACK_COOKIE)
636 ireq->tstamp_ok &= !ireq->sack_ok;
640 /* We always send an MSS option. */
642 remaining -= TCPOLEN_MSS_ALIGNED;
644 if (likely(ireq->wscale_ok)) {
645 opts->ws = ireq->rcv_wscale;
646 opts->options |= OPTION_WSCALE;
647 remaining -= TCPOLEN_WSCALE_ALIGNED;
649 if (likely(ireq->tstamp_ok)) {
650 opts->options |= OPTION_TS;
651 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
652 opts->tsecr = req->ts_recent;
653 remaining -= TCPOLEN_TSTAMP_ALIGNED;
655 if (likely(ireq->sack_ok)) {
656 opts->options |= OPTION_SACK_ADVERTISE;
657 if (unlikely(!ireq->tstamp_ok))
658 remaining -= TCPOLEN_SACKPERM_ALIGNED;
660 if (foc != NULL && foc->len >= 0) {
663 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
664 TCPOLEN_FASTOPEN_BASE;
665 need = (need + 3) & ~3U; /* Align to 32 bits */
666 if (remaining >= need) {
667 opts->options |= OPTION_FAST_OPEN_COOKIE;
668 opts->fastopen_cookie = foc;
673 return MAX_TCP_OPTION_SPACE - remaining;
676 /* Compute TCP options for ESTABLISHED sockets. This is not the
677 * final wire format yet.
679 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
680 struct tcp_out_options *opts,
681 struct tcp_md5sig_key **md5)
683 struct tcp_sock *tp = tcp_sk(sk);
684 unsigned int size = 0;
685 unsigned int eff_sacks;
689 #ifdef CONFIG_TCP_MD5SIG
690 *md5 = tp->af_specific->md5_lookup(sk, sk);
691 if (unlikely(*md5)) {
692 opts->options |= OPTION_MD5;
693 size += TCPOLEN_MD5SIG_ALIGNED;
699 if (likely(tp->rx_opt.tstamp_ok)) {
700 opts->options |= OPTION_TS;
701 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
702 opts->tsecr = tp->rx_opt.ts_recent;
703 size += TCPOLEN_TSTAMP_ALIGNED;
706 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
707 if (unlikely(eff_sacks)) {
708 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
709 opts->num_sack_blocks =
710 min_t(unsigned int, eff_sacks,
711 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
712 TCPOLEN_SACK_PERBLOCK);
713 if (likely(opts->num_sack_blocks))
714 size += TCPOLEN_SACK_BASE_ALIGNED +
715 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
722 /* TCP SMALL QUEUES (TSQ)
724 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
725 * to reduce RTT and bufferbloat.
726 * We do this using a special skb destructor (tcp_wfree).
728 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
729 * needs to be reallocated in a driver.
730 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
732 * Since transmit from skb destructor is forbidden, we use a tasklet
733 * to process all sockets that eventually need to send more skbs.
734 * We use one tasklet per cpu, with its own queue of sockets.
737 struct tasklet_struct tasklet;
738 struct list_head head; /* queue of tcp sockets */
740 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
742 static void tcp_tsq_handler(struct sock *sk)
744 if ((1 << sk->sk_state) &
745 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
746 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
747 struct tcp_sock *tp = tcp_sk(sk);
749 if (tp->lost_out > tp->retrans_out &&
750 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
751 tcp_mstamp_refresh(tp);
752 tcp_xmit_retransmit_queue(sk);
755 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
760 * One tasklet per cpu tries to send more skbs.
761 * We run in tasklet context but need to disable irqs when
762 * transferring tsq->head because tcp_wfree() might
763 * interrupt us (non NAPI drivers)
765 static void tcp_tasklet_func(unsigned long data)
767 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
770 struct list_head *q, *n;
774 local_irq_save(flags);
775 list_splice_init(&tsq->head, &list);
776 local_irq_restore(flags);
778 list_for_each_safe(q, n, &list) {
779 tp = list_entry(q, struct tcp_sock, tsq_node);
780 list_del(&tp->tsq_node);
782 sk = (struct sock *)tp;
783 smp_mb__before_atomic();
784 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
786 if (!sk->sk_lock.owned &&
787 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
789 if (!sock_owned_by_user(sk)) {
790 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
800 #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
801 TCPF_WRITE_TIMER_DEFERRED | \
802 TCPF_DELACK_TIMER_DEFERRED | \
803 TCPF_MTU_REDUCED_DEFERRED)
805 * tcp_release_cb - tcp release_sock() callback
808 * called from release_sock() to perform protocol dependent
809 * actions before socket release.
811 void tcp_release_cb(struct sock *sk)
813 unsigned long flags, nflags;
815 /* perform an atomic operation only if at least one flag is set */
817 flags = sk->sk_tsq_flags;
818 if (!(flags & TCP_DEFERRED_ALL))
820 nflags = flags & ~TCP_DEFERRED_ALL;
821 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
823 if (flags & TCPF_TSQ_DEFERRED)
826 /* Here begins the tricky part :
827 * We are called from release_sock() with :
829 * 2) sk_lock.slock spinlock held
830 * 3) socket owned by us (sk->sk_lock.owned == 1)
832 * But following code is meant to be called from BH handlers,
833 * so we should keep BH disabled, but early release socket ownership
835 sock_release_ownership(sk);
837 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
838 tcp_write_timer_handler(sk);
841 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
842 tcp_delack_timer_handler(sk);
845 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
846 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
850 EXPORT_SYMBOL(tcp_release_cb);
852 void __init tcp_tasklet_init(void)
856 for_each_possible_cpu(i) {
857 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
859 INIT_LIST_HEAD(&tsq->head);
860 tasklet_init(&tsq->tasklet,
867 * Write buffer destructor automatically called from kfree_skb.
868 * We can't xmit new skbs from this context, as we might already
871 void tcp_wfree(struct sk_buff *skb)
873 struct sock *sk = skb->sk;
874 struct tcp_sock *tp = tcp_sk(sk);
875 unsigned long flags, nval, oval;
877 /* Keep one reference on sk_wmem_alloc.
878 * Will be released by sk_free() from here or tcp_tasklet_func()
880 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
882 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
883 * Wait until our queues (qdisc + devices) are drained.
885 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
886 * - chance for incoming ACK (processed by another cpu maybe)
887 * to migrate this flow (skb->ooo_okay will be eventually set)
889 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
892 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
893 struct tsq_tasklet *tsq;
896 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
899 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
900 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
904 /* queue this socket to tasklet queue */
905 local_irq_save(flags);
906 tsq = this_cpu_ptr(&tsq_tasklet);
907 empty = list_empty(&tsq->head);
908 list_add(&tp->tsq_node, &tsq->head);
910 tasklet_schedule(&tsq->tasklet);
911 local_irq_restore(flags);
918 /* Note: Called under hard irq.
919 * We can not call TCP stack right away.
921 enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
923 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
924 struct sock *sk = (struct sock *)tp;
925 unsigned long nval, oval;
927 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
928 struct tsq_tasklet *tsq;
931 if (oval & TSQF_QUEUED)
934 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
935 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
939 if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
941 /* queue this socket to tasklet queue */
942 tsq = this_cpu_ptr(&tsq_tasklet);
943 empty = list_empty(&tsq->head);
944 list_add(&tp->tsq_node, &tsq->head);
946 tasklet_schedule(&tsq->tasklet);
949 return HRTIMER_NORESTART;
952 /* BBR congestion control needs pacing.
953 * Same remark for SO_MAX_PACING_RATE.
954 * sch_fq packet scheduler is efficiently handling pacing,
955 * but is not always installed/used.
956 * Return true if TCP stack should pace packets itself.
958 static bool tcp_needs_internal_pacing(const struct sock *sk)
960 return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
963 static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
968 if (!tcp_needs_internal_pacing(sk))
970 rate = sk->sk_pacing_rate;
971 if (!rate || rate == ~0U)
974 /* Should account for header sizes as sch_fq does,
975 * but lets make things simple.
977 len_ns = (u64)skb->len * NSEC_PER_SEC;
978 do_div(len_ns, rate);
979 hrtimer_start(&tcp_sk(sk)->pacing_timer,
980 ktime_add_ns(ktime_get(), len_ns),
981 HRTIMER_MODE_ABS_PINNED);
984 /* This routine actually transmits TCP packets queued in by
985 * tcp_do_sendmsg(). This is used by both the initial
986 * transmission and possible later retransmissions.
987 * All SKB's seen here are completely headerless. It is our
988 * job to build the TCP header, and pass the packet down to
989 * IP so it can do the same plus pass the packet off to the
992 * We are working here with either a clone of the original
993 * SKB, or a fresh unique copy made by the retransmit engine.
995 static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
996 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
998 const struct inet_connection_sock *icsk = inet_csk(sk);
999 struct inet_sock *inet;
1000 struct tcp_sock *tp;
1001 struct tcp_skb_cb *tcb;
1002 struct tcp_out_options opts;
1003 unsigned int tcp_options_size, tcp_header_size;
1004 struct sk_buff *oskb = NULL;
1005 struct tcp_md5sig_key *md5;
1009 BUG_ON(!skb || !tcp_skb_pcount(skb));
1013 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1016 if (unlikely(skb_cloned(skb)))
1017 skb = pskb_copy(skb, gfp_mask);
1019 skb = skb_clone(skb, gfp_mask);
1023 skb->skb_mstamp = tp->tcp_mstamp;
1026 tcb = TCP_SKB_CB(skb);
1027 memset(&opts, 0, sizeof(opts));
1029 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1030 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1032 tcp_options_size = tcp_established_options(sk, skb, &opts,
1034 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1036 /* if no packet is in qdisc/device queue, then allow XPS to select
1037 * another queue. We can be called from tcp_tsq_handler()
1038 * which holds one reference to sk_wmem_alloc.
1040 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1041 * One way to get this would be to set skb->truesize = 2 on them.
1043 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1045 /* If we had to use memory reserve to allocate this skb,
1046 * this might cause drops if packet is looped back :
1047 * Other socket might not have SOCK_MEMALLOC.
1048 * Packets not looped back do not care about pfmemalloc.
1050 skb->pfmemalloc = 0;
1052 skb_push(skb, tcp_header_size);
1053 skb_reset_transport_header(skb);
1057 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1058 skb_set_hash_from_sk(skb, sk);
1059 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1061 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1063 /* Build TCP header and checksum it. */
1064 th = (struct tcphdr *)skb->data;
1065 th->source = inet->inet_sport;
1066 th->dest = inet->inet_dport;
1067 th->seq = htonl(tcb->seq);
1068 th->ack_seq = htonl(rcv_nxt);
1069 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1075 /* The urg_mode check is necessary during a below snd_una win probe */
1076 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1077 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1078 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1080 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1081 th->urg_ptr = htons(0xFFFF);
1086 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1087 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1088 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1089 th->window = htons(tcp_select_window(sk));
1090 tcp_ecn_send(sk, skb, th, tcp_header_size);
1092 /* RFC1323: The window in SYN & SYN/ACK segments
1095 th->window = htons(min(tp->rcv_wnd, 65535U));
1097 #ifdef CONFIG_TCP_MD5SIG
1098 /* Calculate the MD5 hash, as we have all we need now */
1100 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1101 tp->af_specific->calc_md5_hash(opts.hash_location,
1106 icsk->icsk_af_ops->send_check(sk, skb);
1108 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1109 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1111 if (skb->len != tcp_header_size) {
1112 tcp_event_data_sent(tp, sk);
1113 tp->data_segs_out += tcp_skb_pcount(skb);
1114 tcp_internal_pacing(sk, skb);
1117 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1118 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1119 tcp_skb_pcount(skb));
1121 tp->segs_out += tcp_skb_pcount(skb);
1122 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1123 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1124 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1126 /* Our usage of tstamp should remain private */
1129 /* Cleanup our debris for IP stacks */
1130 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1131 sizeof(struct inet6_skb_parm)));
1133 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1135 if (unlikely(err > 0)) {
1137 err = net_xmit_eval(err);
1140 oskb->skb_mstamp = tp->tcp_mstamp;
1141 tcp_rate_skb_sent(sk, oskb);
1146 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1149 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1150 tcp_sk(sk)->rcv_nxt);
1153 /* This routine just queues the buffer for sending.
1155 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1156 * otherwise socket can stall.
1158 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1160 struct tcp_sock *tp = tcp_sk(sk);
1162 /* Advance write_seq and place onto the write_queue. */
1163 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1164 __skb_header_release(skb);
1165 tcp_add_write_queue_tail(sk, skb);
1166 sk->sk_wmem_queued += skb->truesize;
1167 sk_mem_charge(sk, skb->truesize);
1170 /* Initialize TSO segments for a packet. */
1171 static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1173 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1174 /* Avoid the costly divide in the normal
1177 tcp_skb_pcount_set(skb, 1);
1178 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1180 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1181 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1185 /* When a modification to fackets out becomes necessary, we need to check
1186 * skb is counted to fackets_out or not.
1188 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1191 struct tcp_sock *tp = tcp_sk(sk);
1193 if (!tp->sacked_out || tcp_is_reno(tp))
1196 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1197 tp->fackets_out -= decr;
1200 /* Pcount in the middle of the write queue got changed, we need to do various
1201 * tweaks to fix counters
1203 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1205 struct tcp_sock *tp = tcp_sk(sk);
1207 tp->packets_out -= decr;
1209 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1210 tp->sacked_out -= decr;
1211 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1212 tp->retrans_out -= decr;
1213 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1214 tp->lost_out -= decr;
1216 /* Reno case is special. Sigh... */
1217 if (tcp_is_reno(tp) && decr > 0)
1218 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1220 tcp_adjust_fackets_out(sk, skb, decr);
1222 if (tp->lost_skb_hint &&
1223 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1224 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1225 tp->lost_cnt_hint -= decr;
1227 tcp_verify_left_out(tp);
1230 static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1232 return TCP_SKB_CB(skb)->txstamp_ack ||
1233 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1236 static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1238 struct skb_shared_info *shinfo = skb_shinfo(skb);
1240 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1241 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1242 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1243 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1245 shinfo->tx_flags &= ~tsflags;
1246 shinfo2->tx_flags |= tsflags;
1247 swap(shinfo->tskey, shinfo2->tskey);
1248 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1249 TCP_SKB_CB(skb)->txstamp_ack = 0;
1253 static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1255 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1256 TCP_SKB_CB(skb)->eor = 0;
1259 /* Function to create two new TCP segments. Shrinks the given segment
1260 * to the specified size and appends a new segment with the rest of the
1261 * packet to the list. This won't be called frequently, I hope.
1262 * Remember, these are still headerless SKBs at this point.
1264 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1265 unsigned int mss_now, gfp_t gfp)
1267 struct tcp_sock *tp = tcp_sk(sk);
1268 struct sk_buff *buff;
1269 int nsize, old_factor;
1274 if (WARN_ON(len > skb->len))
1277 nsize = skb_headlen(skb) - len;
1281 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1282 * We need some allowance to not penalize applications setting small
1284 * Also allow first and last skb in retransmit queue to be split.
1286 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1287 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1288 skb != tcp_rtx_queue_head(sk) &&
1289 skb != tcp_rtx_queue_tail(sk))) {
1290 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1294 if (skb_unclone(skb, gfp))
1297 /* Get a new skb... force flag on. */
1298 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1300 return -ENOMEM; /* We'll just try again later. */
1302 sk->sk_wmem_queued += buff->truesize;
1303 sk_mem_charge(sk, buff->truesize);
1304 nlen = skb->len - len - nsize;
1305 buff->truesize += nlen;
1306 skb->truesize -= nlen;
1308 /* Correct the sequence numbers. */
1309 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1310 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1311 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1313 /* PSH and FIN should only be set in the second packet. */
1314 flags = TCP_SKB_CB(skb)->tcp_flags;
1315 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1316 TCP_SKB_CB(buff)->tcp_flags = flags;
1317 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1318 tcp_skb_fragment_eor(skb, buff);
1320 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1321 /* Copy and checksum data tail into the new buffer. */
1322 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1323 skb_put(buff, nsize),
1328 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1330 skb->ip_summed = CHECKSUM_PARTIAL;
1331 skb_split(skb, buff, len);
1334 buff->ip_summed = skb->ip_summed;
1336 buff->tstamp = skb->tstamp;
1337 tcp_fragment_tstamp(skb, buff);
1339 old_factor = tcp_skb_pcount(skb);
1341 /* Fix up tso_factor for both original and new SKB. */
1342 tcp_set_skb_tso_segs(skb, mss_now);
1343 tcp_set_skb_tso_segs(buff, mss_now);
1345 /* Update delivered info for the new segment */
1346 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1348 /* If this packet has been sent out already, we must
1349 * adjust the various packet counters.
1351 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1352 int diff = old_factor - tcp_skb_pcount(skb) -
1353 tcp_skb_pcount(buff);
1356 tcp_adjust_pcount(sk, skb, diff);
1359 /* Link BUFF into the send queue. */
1360 __skb_header_release(buff);
1361 tcp_insert_write_queue_after(skb, buff, sk);
1366 /* This is similar to __pskb_pull_tail(). The difference is that pulled
1367 * data is not copied, but immediately discarded.
1369 static int __pskb_trim_head(struct sk_buff *skb, int len)
1371 struct skb_shared_info *shinfo;
1374 eat = min_t(int, len, skb_headlen(skb));
1376 __skb_pull(skb, eat);
1383 shinfo = skb_shinfo(skb);
1384 for (i = 0; i < shinfo->nr_frags; i++) {
1385 int size = skb_frag_size(&shinfo->frags[i]);
1388 skb_frag_unref(skb, i);
1391 shinfo->frags[k] = shinfo->frags[i];
1393 shinfo->frags[k].page_offset += eat;
1394 skb_frag_size_sub(&shinfo->frags[k], eat);
1400 shinfo->nr_frags = k;
1402 skb->data_len -= len;
1403 skb->len = skb->data_len;
1407 /* Remove acked data from a packet in the transmit queue. */
1408 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1412 if (skb_unclone(skb, GFP_ATOMIC))
1415 delta_truesize = __pskb_trim_head(skb, len);
1417 TCP_SKB_CB(skb)->seq += len;
1418 skb->ip_summed = CHECKSUM_PARTIAL;
1420 if (delta_truesize) {
1421 skb->truesize -= delta_truesize;
1422 sk->sk_wmem_queued -= delta_truesize;
1423 sk_mem_uncharge(sk, delta_truesize);
1424 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1427 /* Any change of skb->len requires recalculation of tso factor. */
1428 if (tcp_skb_pcount(skb) > 1)
1429 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1434 /* Calculate MSS not accounting any TCP options. */
1435 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1437 const struct tcp_sock *tp = tcp_sk(sk);
1438 const struct inet_connection_sock *icsk = inet_csk(sk);
1441 /* Calculate base mss without TCP options:
1442 It is MMS_S - sizeof(tcphdr) of rfc1122
1444 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1446 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1447 if (icsk->icsk_af_ops->net_frag_header_len) {
1448 const struct dst_entry *dst = __sk_dst_get(sk);
1450 if (dst && dst_allfrag(dst))
1451 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1454 /* Clamp it (mss_clamp does not include tcp options) */
1455 if (mss_now > tp->rx_opt.mss_clamp)
1456 mss_now = tp->rx_opt.mss_clamp;
1458 /* Now subtract optional transport overhead */
1459 mss_now -= icsk->icsk_ext_hdr_len;
1461 /* Then reserve room for full set of TCP options and 8 bytes of data */
1462 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1466 /* Calculate MSS. Not accounting for SACKs here. */
1467 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1469 /* Subtract TCP options size, not including SACKs */
1470 return __tcp_mtu_to_mss(sk, pmtu) -
1471 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1473 EXPORT_SYMBOL(tcp_mtu_to_mss);
1475 /* Inverse of above */
1476 int tcp_mss_to_mtu(struct sock *sk, int mss)
1478 const struct tcp_sock *tp = tcp_sk(sk);
1479 const struct inet_connection_sock *icsk = inet_csk(sk);
1483 tp->tcp_header_len +
1484 icsk->icsk_ext_hdr_len +
1485 icsk->icsk_af_ops->net_header_len;
1487 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1488 if (icsk->icsk_af_ops->net_frag_header_len) {
1489 const struct dst_entry *dst = __sk_dst_get(sk);
1491 if (dst && dst_allfrag(dst))
1492 mtu += icsk->icsk_af_ops->net_frag_header_len;
1496 EXPORT_SYMBOL(tcp_mss_to_mtu);
1498 /* MTU probing init per socket */
1499 void tcp_mtup_init(struct sock *sk)
1501 struct tcp_sock *tp = tcp_sk(sk);
1502 struct inet_connection_sock *icsk = inet_csk(sk);
1503 struct net *net = sock_net(sk);
1505 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1506 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1507 icsk->icsk_af_ops->net_header_len;
1508 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1509 icsk->icsk_mtup.probe_size = 0;
1510 if (icsk->icsk_mtup.enabled)
1511 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1513 EXPORT_SYMBOL(tcp_mtup_init);
1515 /* This function synchronize snd mss to current pmtu/exthdr set.
1517 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1518 for TCP options, but includes only bare TCP header.
1520 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1521 It is minimum of user_mss and mss received with SYN.
1522 It also does not include TCP options.
1524 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1526 tp->mss_cache is current effective sending mss, including
1527 all tcp options except for SACKs. It is evaluated,
1528 taking into account current pmtu, but never exceeds
1529 tp->rx_opt.mss_clamp.
1531 NOTE1. rfc1122 clearly states that advertised MSS
1532 DOES NOT include either tcp or ip options.
1534 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1535 are READ ONLY outside this function. --ANK (980731)
1537 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1539 struct tcp_sock *tp = tcp_sk(sk);
1540 struct inet_connection_sock *icsk = inet_csk(sk);
1543 if (icsk->icsk_mtup.search_high > pmtu)
1544 icsk->icsk_mtup.search_high = pmtu;
1546 mss_now = tcp_mtu_to_mss(sk, pmtu);
1547 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1549 /* And store cached results */
1550 icsk->icsk_pmtu_cookie = pmtu;
1551 if (icsk->icsk_mtup.enabled)
1552 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1553 tp->mss_cache = mss_now;
1557 EXPORT_SYMBOL(tcp_sync_mss);
1559 /* Compute the current effective MSS, taking SACKs and IP options,
1560 * and even PMTU discovery events into account.
1562 unsigned int tcp_current_mss(struct sock *sk)
1564 const struct tcp_sock *tp = tcp_sk(sk);
1565 const struct dst_entry *dst = __sk_dst_get(sk);
1567 unsigned int header_len;
1568 struct tcp_out_options opts;
1569 struct tcp_md5sig_key *md5;
1571 mss_now = tp->mss_cache;
1574 u32 mtu = dst_mtu(dst);
1575 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1576 mss_now = tcp_sync_mss(sk, mtu);
1579 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1580 sizeof(struct tcphdr);
1581 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1582 * some common options. If this is an odd packet (because we have SACK
1583 * blocks etc) then our calculated header_len will be different, and
1584 * we have to adjust mss_now correspondingly */
1585 if (header_len != tp->tcp_header_len) {
1586 int delta = (int) header_len - tp->tcp_header_len;
1593 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1594 * As additional protections, we do not touch cwnd in retransmission phases,
1595 * and if application hit its sndbuf limit recently.
1597 static void tcp_cwnd_application_limited(struct sock *sk)
1599 struct tcp_sock *tp = tcp_sk(sk);
1601 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1602 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1603 /* Limited by application or receiver window. */
1604 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1605 u32 win_used = max(tp->snd_cwnd_used, init_win);
1606 if (win_used < tp->snd_cwnd) {
1607 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1608 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1610 tp->snd_cwnd_used = 0;
1612 tp->snd_cwnd_stamp = tcp_jiffies32;
1615 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1617 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1618 struct tcp_sock *tp = tcp_sk(sk);
1620 /* Track the maximum number of outstanding packets in each
1621 * window, and remember whether we were cwnd-limited then.
1623 if (!before(tp->snd_una, tp->max_packets_seq) ||
1624 tp->packets_out > tp->max_packets_out ||
1626 tp->max_packets_out = tp->packets_out;
1627 tp->max_packets_seq = tp->snd_nxt;
1628 tp->is_cwnd_limited = is_cwnd_limited;
1631 if (tcp_is_cwnd_limited(sk)) {
1632 /* Network is feed fully. */
1633 tp->snd_cwnd_used = 0;
1634 tp->snd_cwnd_stamp = tcp_jiffies32;
1636 /* Network starves. */
1637 if (tp->packets_out > tp->snd_cwnd_used)
1638 tp->snd_cwnd_used = tp->packets_out;
1640 if (sysctl_tcp_slow_start_after_idle &&
1641 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1642 !ca_ops->cong_control)
1643 tcp_cwnd_application_limited(sk);
1645 /* The following conditions together indicate the starvation
1646 * is caused by insufficient sender buffer:
1647 * 1) just sent some data (see tcp_write_xmit)
1648 * 2) not cwnd limited (this else condition)
1649 * 3) no more data to send (null tcp_send_head )
1650 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1652 if (!tcp_send_head(sk) && sk->sk_socket &&
1653 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1654 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1655 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1659 /* Minshall's variant of the Nagle send check. */
1660 static bool tcp_minshall_check(const struct tcp_sock *tp)
1662 return after(tp->snd_sml, tp->snd_una) &&
1663 !after(tp->snd_sml, tp->snd_nxt);
1666 /* Update snd_sml if this skb is under mss
1667 * Note that a TSO packet might end with a sub-mss segment
1668 * The test is really :
1669 * if ((skb->len % mss) != 0)
1670 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1671 * But we can avoid doing the divide again given we already have
1672 * skb_pcount = skb->len / mss_now
1674 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1675 const struct sk_buff *skb)
1677 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1678 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1681 /* Return false, if packet can be sent now without violation Nagle's rules:
1682 * 1. It is full sized. (provided by caller in %partial bool)
1683 * 2. Or it contains FIN. (already checked by caller)
1684 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1685 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1686 * With Minshall's modification: all sent small packets are ACKed.
1688 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1692 ((nonagle & TCP_NAGLE_CORK) ||
1693 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1696 /* Return how many segs we'd like on a TSO packet,
1697 * to send one TSO packet per ms
1699 u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1704 bytes = min(sk->sk_pacing_rate >> 10,
1705 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1707 /* Goal is to send at least one packet per ms,
1708 * not one big TSO packet every 100 ms.
1709 * This preserves ACK clocking and is consistent
1710 * with tcp_tso_should_defer() heuristic.
1712 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1716 EXPORT_SYMBOL(tcp_tso_autosize);
1718 /* Return the number of segments we want in the skb we are transmitting.
1719 * See if congestion control module wants to decide; otherwise, autosize.
1721 static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1723 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1724 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1727 tso_segs = tcp_tso_autosize(sk, mss_now,
1728 sysctl_tcp_min_tso_segs);
1729 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1732 /* Returns the portion of skb which can be sent right away */
1733 static unsigned int tcp_mss_split_point(const struct sock *sk,
1734 const struct sk_buff *skb,
1735 unsigned int mss_now,
1736 unsigned int max_segs,
1739 const struct tcp_sock *tp = tcp_sk(sk);
1740 u32 partial, needed, window, max_len;
1742 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1743 max_len = mss_now * max_segs;
1745 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1748 needed = min(skb->len, window);
1750 if (max_len <= needed)
1753 partial = needed % mss_now;
1754 /* If last segment is not a full MSS, check if Nagle rules allow us
1755 * to include this last segment in this skb.
1756 * Otherwise, we'll split the skb at last MSS boundary
1758 if (tcp_nagle_check(partial != 0, tp, nonagle))
1759 return needed - partial;
1764 /* Can at least one segment of SKB be sent right now, according to the
1765 * congestion window rules? If so, return how many segments are allowed.
1767 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1768 const struct sk_buff *skb)
1770 u32 in_flight, cwnd, halfcwnd;
1772 /* Don't be strict about the congestion window for the final FIN. */
1773 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1774 tcp_skb_pcount(skb) == 1)
1777 in_flight = tcp_packets_in_flight(tp);
1778 cwnd = tp->snd_cwnd;
1779 if (in_flight >= cwnd)
1782 /* For better scheduling, ensure we have at least
1783 * 2 GSO packets in flight.
1785 halfcwnd = max(cwnd >> 1, 1U);
1786 return min(halfcwnd, cwnd - in_flight);
1789 /* Initialize TSO state of a skb.
1790 * This must be invoked the first time we consider transmitting
1791 * SKB onto the wire.
1793 static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1795 int tso_segs = tcp_skb_pcount(skb);
1797 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1798 tcp_set_skb_tso_segs(skb, mss_now);
1799 tso_segs = tcp_skb_pcount(skb);
1805 /* Return true if the Nagle test allows this packet to be
1808 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1809 unsigned int cur_mss, int nonagle)
1811 /* Nagle rule does not apply to frames, which sit in the middle of the
1812 * write_queue (they have no chances to get new data).
1814 * This is implemented in the callers, where they modify the 'nonagle'
1815 * argument based upon the location of SKB in the send queue.
1817 if (nonagle & TCP_NAGLE_PUSH)
1820 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1821 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1824 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1830 /* Does at least the first segment of SKB fit into the send window? */
1831 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1832 const struct sk_buff *skb,
1833 unsigned int cur_mss)
1835 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1837 if (skb->len > cur_mss)
1838 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1840 return !after(end_seq, tcp_wnd_end(tp));
1843 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1844 * which is put after SKB on the list. It is very much like
1845 * tcp_fragment() except that it may make several kinds of assumptions
1846 * in order to speed up the splitting operation. In particular, we
1847 * know that all the data is in scatter-gather pages, and that the
1848 * packet has never been sent out before (and thus is not cloned).
1850 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1851 unsigned int mss_now, gfp_t gfp)
1853 struct sk_buff *buff;
1854 int nlen = skb->len - len;
1857 /* All of a TSO frame must be composed of paged data. */
1858 if (skb->len != skb->data_len)
1859 return tcp_fragment(sk, skb, len, mss_now, gfp);
1861 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1862 if (unlikely(!buff))
1865 sk->sk_wmem_queued += buff->truesize;
1866 sk_mem_charge(sk, buff->truesize);
1867 buff->truesize += nlen;
1868 skb->truesize -= nlen;
1870 /* Correct the sequence numbers. */
1871 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1872 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1873 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1875 /* PSH and FIN should only be set in the second packet. */
1876 flags = TCP_SKB_CB(skb)->tcp_flags;
1877 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1878 TCP_SKB_CB(buff)->tcp_flags = flags;
1880 /* This packet was never sent out yet, so no SACK bits. */
1881 TCP_SKB_CB(buff)->sacked = 0;
1883 tcp_skb_fragment_eor(skb, buff);
1885 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1886 skb_split(skb, buff, len);
1887 tcp_fragment_tstamp(skb, buff);
1889 /* Fix up tso_factor for both original and new SKB. */
1890 tcp_set_skb_tso_segs(skb, mss_now);
1891 tcp_set_skb_tso_segs(buff, mss_now);
1893 /* Link BUFF into the send queue. */
1894 __skb_header_release(buff);
1895 tcp_insert_write_queue_after(skb, buff, sk);
1900 /* Try to defer sending, if possible, in order to minimize the amount
1901 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1903 * This algorithm is from John Heffner.
1905 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1906 bool *is_cwnd_limited,
1907 bool *is_rwnd_limited,
1910 const struct inet_connection_sock *icsk = inet_csk(sk);
1911 u32 age, send_win, cong_win, limit, in_flight;
1912 struct tcp_sock *tp = tcp_sk(sk);
1913 struct sk_buff *head;
1916 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1919 /* Avoid bursty behavior by allowing defer
1920 * only if the last write was recent.
1922 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1925 in_flight = tcp_packets_in_flight(tp);
1927 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1929 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1931 /* From in_flight test above, we know that cwnd > in_flight. */
1932 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1934 limit = min(send_win, cong_win);
1936 /* If a full-sized TSO skb can be sent, do it. */
1937 if (limit >= max_segs * tp->mss_cache)
1940 /* Middle in queue won't get any more data, full sendable already? */
1941 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1944 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1946 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1948 /* If at least some fraction of a window is available,
1951 chunk /= win_divisor;
1955 /* Different approach, try not to defer past a single
1956 * ACK. Receiver should ACK every other full sized
1957 * frame, so if we have space for more than 3 frames
1960 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1964 head = tcp_write_queue_head(sk);
1966 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1967 /* If next ACK is likely to come too late (half srtt), do not defer */
1968 if (age < (tp->srtt_us >> 4))
1971 /* Ok, it looks like it is advisable to defer.
1972 * Three cases are tracked :
1973 * 1) We are cwnd-limited
1974 * 2) We are rwnd-limited
1975 * 3) We are application limited.
1977 if (cong_win < send_win) {
1978 if (cong_win <= skb->len) {
1979 *is_cwnd_limited = true;
1983 if (send_win <= skb->len) {
1984 *is_rwnd_limited = true;
1989 /* If this packet won't get more data, do not wait. */
1990 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1999 static inline void tcp_mtu_check_reprobe(struct sock *sk)
2001 struct inet_connection_sock *icsk = inet_csk(sk);
2002 struct tcp_sock *tp = tcp_sk(sk);
2003 struct net *net = sock_net(sk);
2007 interval = net->ipv4.sysctl_tcp_probe_interval;
2008 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2009 if (unlikely(delta >= interval * HZ)) {
2010 int mss = tcp_current_mss(sk);
2012 /* Update current search range */
2013 icsk->icsk_mtup.probe_size = 0;
2014 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2015 sizeof(struct tcphdr) +
2016 icsk->icsk_af_ops->net_header_len;
2017 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2019 /* Update probe time stamp */
2020 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2024 static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2026 struct sk_buff *skb, *next;
2028 skb = tcp_send_head(sk);
2029 tcp_for_write_queue_from_safe(skb, next, sk) {
2030 if (len <= skb->len)
2033 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2042 /* Create a new MTU probe if we are ready.
2043 * MTU probe is regularly attempting to increase the path MTU by
2044 * deliberately sending larger packets. This discovers routing
2045 * changes resulting in larger path MTUs.
2047 * Returns 0 if we should wait to probe (no cwnd available),
2048 * 1 if a probe was sent,
2051 static int tcp_mtu_probe(struct sock *sk)
2053 struct inet_connection_sock *icsk = inet_csk(sk);
2054 struct tcp_sock *tp = tcp_sk(sk);
2055 struct sk_buff *skb, *nskb, *next;
2056 struct net *net = sock_net(sk);
2063 /* Not currently probing/verifying,
2065 * have enough cwnd, and
2066 * not SACKing (the variable headers throw things off)
2068 if (likely(!icsk->icsk_mtup.enabled ||
2069 icsk->icsk_mtup.probe_size ||
2070 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2071 tp->snd_cwnd < 11 ||
2072 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2075 /* Use binary search for probe_size between tcp_mss_base,
2076 * and current mss_clamp. if (search_high - search_low)
2077 * smaller than a threshold, backoff from probing.
2079 mss_now = tcp_current_mss(sk);
2080 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2081 icsk->icsk_mtup.search_low) >> 1);
2082 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2083 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2084 /* When misfortune happens, we are reprobing actively,
2085 * and then reprobe timer has expired. We stick with current
2086 * probing process by not resetting search range to its orignal.
2088 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2089 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2090 /* Check whether enough time has elaplased for
2091 * another round of probing.
2093 tcp_mtu_check_reprobe(sk);
2097 /* Have enough data in the send queue to probe? */
2098 if (tp->write_seq - tp->snd_nxt < size_needed)
2101 if (tp->snd_wnd < size_needed)
2103 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2106 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2107 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2108 if (!tcp_packets_in_flight(tp))
2114 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2117 /* We're allowed to probe. Build it now. */
2118 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2121 sk->sk_wmem_queued += nskb->truesize;
2122 sk_mem_charge(sk, nskb->truesize);
2124 skb = tcp_send_head(sk);
2126 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2127 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2128 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2129 TCP_SKB_CB(nskb)->sacked = 0;
2131 nskb->ip_summed = skb->ip_summed;
2133 tcp_insert_write_queue_before(nskb, skb, sk);
2134 tcp_highest_sack_replace(sk, skb, nskb);
2137 tcp_for_write_queue_from_safe(skb, next, sk) {
2138 copy = min_t(int, skb->len, probe_size - len);
2139 if (nskb->ip_summed) {
2140 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2142 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2143 skb_put(nskb, copy),
2145 nskb->csum = csum_block_add(nskb->csum, csum, len);
2148 if (skb->len <= copy) {
2149 /* We've eaten all the data from this skb.
2151 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2152 /* If this is the last SKB we copy and eor is set
2153 * we need to propagate it to the new skb.
2155 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2156 tcp_skb_collapse_tstamp(nskb, skb);
2157 tcp_unlink_write_queue(skb, sk);
2158 sk_wmem_free_skb(sk, skb);
2160 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2161 ~(TCPHDR_FIN|TCPHDR_PSH);
2162 if (!skb_shinfo(skb)->nr_frags) {
2163 skb_pull(skb, copy);
2164 if (skb->ip_summed != CHECKSUM_PARTIAL)
2165 skb->csum = csum_partial(skb->data,
2168 __pskb_trim_head(skb, copy);
2169 tcp_set_skb_tso_segs(skb, mss_now);
2171 TCP_SKB_CB(skb)->seq += copy;
2176 if (len >= probe_size)
2179 tcp_init_tso_segs(nskb, nskb->len);
2181 /* We're ready to send. If this fails, the probe will
2182 * be resegmented into mss-sized pieces by tcp_write_xmit().
2184 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2185 /* Decrement cwnd here because we are sending
2186 * effectively two packets. */
2188 tcp_event_new_data_sent(sk, nskb);
2190 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2191 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2192 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2200 static bool tcp_pacing_check(const struct sock *sk)
2202 return tcp_needs_internal_pacing(sk) &&
2203 hrtimer_active(&tcp_sk(sk)->pacing_timer);
2206 /* TCP Small Queues :
2207 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2208 * (These limits are doubled for retransmits)
2210 * - better RTT estimation and ACK scheduling
2213 * Alas, some drivers / subsystems require a fair amount
2214 * of queued bytes to ensure line rate.
2215 * One example is wifi aggregation (802.11 AMPDU)
2217 static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2218 unsigned int factor)
2222 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2223 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2226 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2227 /* Always send the 1st or 2nd skb in write queue.
2228 * No need to wait for TX completion to call us back,
2229 * after softirq/tasklet schedule.
2230 * This helps when TX completions are delayed too much.
2232 if (skb == sk->sk_write_queue.next ||
2233 skb->prev == sk->sk_write_queue.next)
2236 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2237 /* It is possible TX completion already happened
2238 * before we set TSQ_THROTTLED, so we must
2239 * test again the condition.
2241 smp_mb__after_atomic();
2242 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2248 static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2250 const u32 now = tcp_jiffies32;
2251 enum tcp_chrono old = tp->chrono_type;
2253 if (old > TCP_CHRONO_UNSPEC)
2254 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2255 tp->chrono_start = now;
2256 tp->chrono_type = new;
2259 void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2261 struct tcp_sock *tp = tcp_sk(sk);
2263 /* If there are multiple conditions worthy of tracking in a
2264 * chronograph then the highest priority enum takes precedence
2265 * over the other conditions. So that if something "more interesting"
2266 * starts happening, stop the previous chrono and start a new one.
2268 if (type > tp->chrono_type)
2269 tcp_chrono_set(tp, type);
2272 void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2274 struct tcp_sock *tp = tcp_sk(sk);
2277 /* There are multiple conditions worthy of tracking in a
2278 * chronograph, so that the highest priority enum takes
2279 * precedence over the other conditions (see tcp_chrono_start).
2280 * If a condition stops, we only stop chrono tracking if
2281 * it's the "most interesting" or current chrono we are
2282 * tracking and starts busy chrono if we have pending data.
2284 if (tcp_write_queue_empty(sk))
2285 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2286 else if (type == tp->chrono_type)
2287 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2290 /* This routine writes packets to the network. It advances the
2291 * send_head. This happens as incoming acks open up the remote
2294 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2295 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2296 * account rare use of URG, this is not a big flaw.
2298 * Send at most one packet when push_one > 0. Temporarily ignore
2299 * cwnd limit to force at most one packet out when push_one == 2.
2301 * Returns true, if no segments are in flight and we have queued segments,
2302 * but cannot send anything now because of SWS or another problem.
2304 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2305 int push_one, gfp_t gfp)
2307 struct tcp_sock *tp = tcp_sk(sk);
2308 struct sk_buff *skb;
2309 unsigned int tso_segs, sent_pkts;
2312 bool is_cwnd_limited = false, is_rwnd_limited = false;
2317 tcp_mstamp_refresh(tp);
2319 /* Do MTU probing. */
2320 result = tcp_mtu_probe(sk);
2323 } else if (result > 0) {
2328 max_segs = tcp_tso_segs(sk, mss_now);
2329 while ((skb = tcp_send_head(sk))) {
2332 if (tcp_pacing_check(sk))
2335 tso_segs = tcp_init_tso_segs(skb, mss_now);
2338 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2339 /* "skb_mstamp" is used as a start point for the retransmit timer */
2340 skb->skb_mstamp = tp->tcp_mstamp;
2341 goto repair; /* Skip network transmission */
2344 cwnd_quota = tcp_cwnd_test(tp, skb);
2347 /* Force out a loss probe pkt. */
2353 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2354 is_rwnd_limited = true;
2358 if (tso_segs == 1) {
2359 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2360 (tcp_skb_is_last(sk, skb) ?
2361 nonagle : TCP_NAGLE_PUSH))))
2365 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2366 &is_rwnd_limited, max_segs))
2371 if (tso_segs > 1 && !tcp_urg_mode(tp))
2372 limit = tcp_mss_split_point(sk, skb, mss_now,
2378 if (skb->len > limit &&
2379 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2382 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2383 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2384 if (tcp_small_queue_check(sk, skb, 0))
2387 /* Argh, we hit an empty skb(), presumably a thread
2388 * is sleeping in sendmsg()/sk_stream_wait_memory().
2389 * We do not want to send a pure-ack packet and have
2390 * a strange looking rtx queue with empty packet(s).
2392 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2395 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2399 /* Advance the send_head. This one is sent out.
2400 * This call will increment packets_out.
2402 tcp_event_new_data_sent(sk, skb);
2404 tcp_minshall_update(tp, mss_now, skb);
2405 sent_pkts += tcp_skb_pcount(skb);
2411 if (is_rwnd_limited)
2412 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2414 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2416 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2417 if (likely(sent_pkts || is_cwnd_limited))
2418 tcp_cwnd_validate(sk, is_cwnd_limited);
2420 if (likely(sent_pkts)) {
2421 if (tcp_in_cwnd_reduction(sk))
2422 tp->prr_out += sent_pkts;
2424 /* Send one loss probe per tail loss episode. */
2426 tcp_schedule_loss_probe(sk, false);
2429 return !tp->packets_out && tcp_send_head(sk);
2432 bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2434 struct inet_connection_sock *icsk = inet_csk(sk);
2435 struct tcp_sock *tp = tcp_sk(sk);
2436 u32 timeout, rto_delta_us;
2438 /* Don't do any loss probe on a Fast Open connection before 3WHS
2441 if (tp->fastopen_rsk)
2444 /* Schedule a loss probe in 2*RTT for SACK capable connections
2445 * in Open state, that are either limited by cwnd or application.
2447 if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) ||
2448 !tp->packets_out || !tcp_is_sack(tp) ||
2449 icsk->icsk_ca_state != TCP_CA_Open)
2452 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2456 /* Probe timeout is 2*rtt. Add minimum RTO to account
2457 * for delayed ack when there's one outstanding packet. If no RTT
2458 * sample is available then probe after TCP_TIMEOUT_INIT.
2461 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2462 if (tp->packets_out == 1)
2463 timeout += TCP_RTO_MIN;
2465 timeout += TCP_TIMEOUT_MIN;
2467 timeout = TCP_TIMEOUT_INIT;
2470 /* If the RTO formula yields an earlier time, then use that time. */
2471 rto_delta_us = advancing_rto ?
2472 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2473 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2474 if (rto_delta_us > 0)
2475 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2477 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2482 /* Thanks to skb fast clones, we can detect if a prior transmit of
2483 * a packet is still in a qdisc or driver queue.
2484 * In this case, there is very little point doing a retransmit !
2486 static bool skb_still_in_host_queue(const struct sock *sk,
2487 const struct sk_buff *skb)
2489 if (unlikely(skb_fclone_busy(sk, skb))) {
2490 NET_INC_STATS(sock_net(sk),
2491 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2497 /* When probe timeout (PTO) fires, try send a new segment if possible, else
2498 * retransmit the last segment.
2500 void tcp_send_loss_probe(struct sock *sk)
2502 struct tcp_sock *tp = tcp_sk(sk);
2503 struct sk_buff *skb;
2505 int mss = tcp_current_mss(sk);
2507 /* At most one outstanding TLP */
2508 if (tp->tlp_high_seq)
2511 tp->tlp_retrans = 0;
2512 skb = tcp_send_head(sk);
2514 if (tcp_snd_wnd_test(tp, skb, mss)) {
2515 pcount = tp->packets_out;
2516 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2517 if (tp->packets_out > pcount)
2521 skb = tcp_write_queue_prev(sk, skb);
2523 skb = tcp_write_queue_tail(sk);
2526 if (unlikely(!skb)) {
2527 WARN_ONCE(tp->packets_out,
2528 "invalid inflight: %u state %u cwnd %u mss %d\n",
2529 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2530 inet_csk(sk)->icsk_pending = 0;
2534 if (skb_still_in_host_queue(sk, skb))
2537 pcount = tcp_skb_pcount(skb);
2538 if (WARN_ON(!pcount))
2541 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2542 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2545 skb = tcp_write_queue_next(sk, skb);
2548 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2551 if (__tcp_retransmit_skb(sk, skb, 1))
2554 tp->tlp_retrans = 1;
2557 /* Record snd_nxt for loss detection. */
2558 tp->tlp_high_seq = tp->snd_nxt;
2560 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2561 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2562 inet_csk(sk)->icsk_pending = 0;
2567 /* Push out any pending frames which were held back due to
2568 * TCP_CORK or attempt at coalescing tiny packets.
2569 * The socket must be locked by the caller.
2571 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2574 /* If we are closed, the bytes will have to remain here.
2575 * In time closedown will finish, we empty the write queue and
2576 * all will be happy.
2578 if (unlikely(sk->sk_state == TCP_CLOSE))
2581 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2582 sk_gfp_mask(sk, GFP_ATOMIC)))
2583 tcp_check_probe_timer(sk);
2586 /* Send _single_ skb sitting at the send head. This function requires
2587 * true push pending frames to setup probe timer etc.
2589 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2591 struct sk_buff *skb = tcp_send_head(sk);
2593 BUG_ON(!skb || skb->len < mss_now);
2595 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2598 /* This function returns the amount that we can raise the
2599 * usable window based on the following constraints
2601 * 1. The window can never be shrunk once it is offered (RFC 793)
2602 * 2. We limit memory per socket
2605 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2606 * RECV.NEXT + RCV.WIN fixed until:
2607 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2609 * i.e. don't raise the right edge of the window until you can raise
2610 * it at least MSS bytes.
2612 * Unfortunately, the recommended algorithm breaks header prediction,
2613 * since header prediction assumes th->window stays fixed.
2615 * Strictly speaking, keeping th->window fixed violates the receiver
2616 * side SWS prevention criteria. The problem is that under this rule
2617 * a stream of single byte packets will cause the right side of the
2618 * window to always advance by a single byte.
2620 * Of course, if the sender implements sender side SWS prevention
2621 * then this will not be a problem.
2623 * BSD seems to make the following compromise:
2625 * If the free space is less than the 1/4 of the maximum
2626 * space available and the free space is less than 1/2 mss,
2627 * then set the window to 0.
2628 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2629 * Otherwise, just prevent the window from shrinking
2630 * and from being larger than the largest representable value.
2632 * This prevents incremental opening of the window in the regime
2633 * where TCP is limited by the speed of the reader side taking
2634 * data out of the TCP receive queue. It does nothing about
2635 * those cases where the window is constrained on the sender side
2636 * because the pipeline is full.
2638 * BSD also seems to "accidentally" limit itself to windows that are a
2639 * multiple of MSS, at least until the free space gets quite small.
2640 * This would appear to be a side effect of the mbuf implementation.
2641 * Combining these two algorithms results in the observed behavior
2642 * of having a fixed window size at almost all times.
2644 * Below we obtain similar behavior by forcing the offered window to
2645 * a multiple of the mss when it is feasible to do so.
2647 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2648 * Regular options like TIMESTAMP are taken into account.
2650 u32 __tcp_select_window(struct sock *sk)
2652 struct inet_connection_sock *icsk = inet_csk(sk);
2653 struct tcp_sock *tp = tcp_sk(sk);
2654 /* MSS for the peer's data. Previous versions used mss_clamp
2655 * here. I don't know if the value based on our guesses
2656 * of peer's MSS is better for the performance. It's more correct
2657 * but may be worse for the performance because of rcv_mss
2658 * fluctuations. --SAW 1998/11/1
2660 int mss = icsk->icsk_ack.rcv_mss;
2661 int free_space = tcp_space(sk);
2662 int allowed_space = tcp_full_space(sk);
2663 int full_space = min_t(int, tp->window_clamp, allowed_space);
2666 if (unlikely(mss > full_space)) {
2671 if (free_space < (full_space >> 1)) {
2672 icsk->icsk_ack.quick = 0;
2674 if (tcp_under_memory_pressure(sk))
2675 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2678 /* free_space might become our new window, make sure we don't
2679 * increase it due to wscale.
2681 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2683 /* if free space is less than mss estimate, or is below 1/16th
2684 * of the maximum allowed, try to move to zero-window, else
2685 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2686 * new incoming data is dropped due to memory limits.
2687 * With large window, mss test triggers way too late in order
2688 * to announce zero window in time before rmem limit kicks in.
2690 if (free_space < (allowed_space >> 4) || free_space < mss)
2694 if (free_space > tp->rcv_ssthresh)
2695 free_space = tp->rcv_ssthresh;
2697 /* Don't do rounding if we are using window scaling, since the
2698 * scaled window will not line up with the MSS boundary anyway.
2700 if (tp->rx_opt.rcv_wscale) {
2701 window = free_space;
2703 /* Advertise enough space so that it won't get scaled away.
2704 * Import case: prevent zero window announcement if
2705 * 1<<rcv_wscale > mss.
2707 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2709 window = tp->rcv_wnd;
2710 /* Get the largest window that is a nice multiple of mss.
2711 * Window clamp already applied above.
2712 * If our current window offering is within 1 mss of the
2713 * free space we just keep it. This prevents the divide
2714 * and multiply from happening most of the time.
2715 * We also don't do any window rounding when the free space
2718 if (window <= free_space - mss || window > free_space)
2719 window = rounddown(free_space, mss);
2720 else if (mss == full_space &&
2721 free_space > window + (full_space >> 1))
2722 window = free_space;
2728 void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2729 const struct sk_buff *next_skb)
2731 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2732 const struct skb_shared_info *next_shinfo =
2733 skb_shinfo(next_skb);
2734 struct skb_shared_info *shinfo = skb_shinfo(skb);
2736 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2737 shinfo->tskey = next_shinfo->tskey;
2738 TCP_SKB_CB(skb)->txstamp_ack |=
2739 TCP_SKB_CB(next_skb)->txstamp_ack;
2743 /* Collapses two adjacent SKB's during retransmission. */
2744 static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2746 struct tcp_sock *tp = tcp_sk(sk);
2747 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2748 int skb_size, next_skb_size;
2750 skb_size = skb->len;
2751 next_skb_size = next_skb->len;
2753 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2755 if (next_skb_size) {
2756 if (next_skb_size <= skb_availroom(skb))
2757 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2759 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2762 tcp_highest_sack_replace(sk, next_skb, skb);
2764 tcp_unlink_write_queue(next_skb, sk);
2766 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2767 skb->ip_summed = CHECKSUM_PARTIAL;
2769 if (skb->ip_summed != CHECKSUM_PARTIAL)
2770 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2772 /* Update sequence range on original skb. */
2773 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2775 /* Merge over control information. This moves PSH/FIN etc. over */
2776 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2778 /* All done, get rid of second SKB and account for it so
2779 * packet counting does not break.
2781 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2782 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2784 /* changed transmit queue under us so clear hints */
2785 tcp_clear_retrans_hints_partial(tp);
2786 if (next_skb == tp->retransmit_skb_hint)
2787 tp->retransmit_skb_hint = skb;
2789 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2791 tcp_skb_collapse_tstamp(skb, next_skb);
2793 sk_wmem_free_skb(sk, next_skb);
2797 /* Check if coalescing SKBs is legal. */
2798 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2800 if (tcp_skb_pcount(skb) > 1)
2802 if (skb_cloned(skb))
2804 if (skb == tcp_send_head(sk))
2806 /* Some heuristics for collapsing over SACK'd could be invented */
2807 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2813 /* Collapse packets in the retransmit queue to make to create
2814 * less packets on the wire. This is only done on retransmission.
2816 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2819 struct tcp_sock *tp = tcp_sk(sk);
2820 struct sk_buff *skb = to, *tmp;
2823 if (!sysctl_tcp_retrans_collapse)
2825 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2828 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2829 if (!tcp_can_collapse(sk, skb))
2832 if (!tcp_skb_can_collapse_to(to))
2845 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2848 if (!tcp_collapse_retrans(sk, to))
2853 /* This retransmits one SKB. Policy decisions and retransmit queue
2854 * state updates are done by the caller. Returns non-zero if an
2855 * error occurred which prevented the send.
2857 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2859 struct inet_connection_sock *icsk = inet_csk(sk);
2860 struct tcp_sock *tp = tcp_sk(sk);
2861 unsigned int cur_mss;
2865 /* Inconclusive MTU probe */
2866 if (icsk->icsk_mtup.probe_size)
2867 icsk->icsk_mtup.probe_size = 0;
2869 /* Do not sent more than we queued. 1/4 is reserved for possible
2870 * copying overhead: fragmentation, tunneling, mangling etc.
2872 if (refcount_read(&sk->sk_wmem_alloc) >
2873 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2877 if (skb_still_in_host_queue(sk, skb))
2880 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2881 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2885 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2889 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2890 return -EHOSTUNREACH; /* Routing failure or similar. */
2892 cur_mss = tcp_current_mss(sk);
2894 /* If receiver has shrunk his window, and skb is out of
2895 * new window, do not retransmit it. The exception is the
2896 * case, when window is shrunk to zero. In this case
2897 * our retransmit serves as a zero window probe.
2899 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2900 TCP_SKB_CB(skb)->seq != tp->snd_una)
2903 len = cur_mss * segs;
2904 if (skb->len > len) {
2905 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2906 return -ENOMEM; /* We'll try again later. */
2908 if (skb_unclone(skb, GFP_ATOMIC))
2911 diff = tcp_skb_pcount(skb);
2912 tcp_set_skb_tso_segs(skb, cur_mss);
2913 diff -= tcp_skb_pcount(skb);
2915 tcp_adjust_pcount(sk, skb, diff);
2916 if (skb->len < cur_mss)
2917 tcp_retrans_try_collapse(sk, skb, cur_mss);
2920 /* RFC3168, section 6.1.1.1. ECN fallback */
2921 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2922 tcp_ecn_clear_syn(sk, skb);
2924 /* Update global and local TCP statistics. */
2925 segs = tcp_skb_pcount(skb);
2926 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2927 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2928 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2929 tp->total_retrans += segs;
2931 /* make sure skb->data is aligned on arches that require it
2932 * and check if ack-trimming & collapsing extended the headroom
2933 * beyond what csum_start can cover.
2935 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2936 skb_headroom(skb) >= 0xFFFF)) {
2937 struct sk_buff *nskb;
2939 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2940 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2943 skb->skb_mstamp = tp->tcp_mstamp;
2944 tcp_rate_skb_sent(sk, skb);
2947 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2951 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2952 } else if (err != -EBUSY) {
2953 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2958 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2960 struct tcp_sock *tp = tcp_sk(sk);
2961 int err = __tcp_retransmit_skb(sk, skb, segs);
2964 #if FASTRETRANS_DEBUG > 0
2965 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2966 net_dbg_ratelimited("retrans_out leaked\n");
2969 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2970 tp->retrans_out += tcp_skb_pcount(skb);
2972 /* Save stamp of the first retransmit. */
2973 if (!tp->retrans_stamp)
2974 tp->retrans_stamp = tcp_skb_timestamp(skb);
2978 if (tp->undo_retrans < 0)
2979 tp->undo_retrans = 0;
2980 tp->undo_retrans += tcp_skb_pcount(skb);
2984 /* This gets called after a retransmit timeout, and the initially
2985 * retransmitted data is acknowledged. It tries to continue
2986 * resending the rest of the retransmit queue, until either
2987 * we've sent it all or the congestion window limit is reached.
2988 * If doing SACK, the first ACK which comes back for a timeout
2989 * based retransmit packet might feed us FACK information again.
2990 * If so, we use it to avoid unnecessarily retransmissions.
2992 void tcp_xmit_retransmit_queue(struct sock *sk)
2994 const struct inet_connection_sock *icsk = inet_csk(sk);
2995 struct tcp_sock *tp = tcp_sk(sk);
2996 struct sk_buff *skb;
2997 struct sk_buff *hole = NULL;
3001 if (!tp->packets_out)
3004 if (tp->retransmit_skb_hint) {
3005 skb = tp->retransmit_skb_hint;
3007 skb = tcp_write_queue_head(sk);
3010 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3011 tcp_for_write_queue_from(skb, sk) {
3015 if (skb == tcp_send_head(sk))
3018 if (tcp_pacing_check(sk))
3021 /* we could do better than to assign each time */
3023 tp->retransmit_skb_hint = skb;
3025 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3028 sacked = TCP_SKB_CB(skb)->sacked;
3029 /* In case tcp_shift_skb_data() have aggregated large skbs,
3030 * we need to make sure not sending too bigs TSO packets
3032 segs = min_t(int, segs, max_segs);
3034 if (tp->retrans_out >= tp->lost_out) {
3036 } else if (!(sacked & TCPCB_LOST)) {
3037 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3042 if (icsk->icsk_ca_state != TCP_CA_Loss)
3043 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3045 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3048 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3051 if (tcp_small_queue_check(sk, skb, 1))
3054 if (tcp_retransmit_skb(sk, skb, segs))
3057 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3059 if (tcp_in_cwnd_reduction(sk))
3060 tp->prr_out += tcp_skb_pcount(skb);
3062 if (skb == tcp_write_queue_head(sk) &&
3063 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3064 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3065 inet_csk(sk)->icsk_rto,
3070 /* We allow to exceed memory limits for FIN packets to expedite
3071 * connection tear down and (memory) recovery.
3072 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3073 * or even be forced to close flow without any FIN.
3074 * In general, we want to allow one skb per socket to avoid hangs
3075 * with edge trigger epoll()
3077 void sk_forced_mem_schedule(struct sock *sk, int size)
3081 if (size <= sk->sk_forward_alloc)
3083 amt = sk_mem_pages(size);
3084 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3085 sk_memory_allocated_add(sk, amt);
3087 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3088 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3091 /* Send a FIN. The caller locks the socket for us.
3092 * We should try to send a FIN packet really hard, but eventually give up.
3094 void tcp_send_fin(struct sock *sk)
3096 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3097 struct tcp_sock *tp = tcp_sk(sk);
3099 /* Optimization, tack on the FIN if we have one skb in write queue and
3100 * this skb was not yet sent, or we are under memory pressure.
3101 * Note: in the latter case, FIN packet will be sent after a timeout,
3102 * as TCP stack thinks it has already been transmitted.
3104 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3106 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3107 TCP_SKB_CB(tskb)->end_seq++;
3109 if (!tcp_send_head(sk)) {
3110 /* This means tskb was already sent.
3111 * Pretend we included the FIN on previous transmit.
3112 * We need to set tp->snd_nxt to the value it would have
3113 * if FIN had been sent. This is because retransmit path
3114 * does not change tp->snd_nxt.
3120 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3121 if (unlikely(!skb)) {
3126 skb_reserve(skb, MAX_TCP_HEADER);
3127 sk_forced_mem_schedule(sk, skb->truesize);
3128 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3129 tcp_init_nondata_skb(skb, tp->write_seq,
3130 TCPHDR_ACK | TCPHDR_FIN);
3131 tcp_queue_skb(sk, skb);
3133 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3136 /* We get here when a process closes a file descriptor (either due to
3137 * an explicit close() or as a byproduct of exit()'ing) and there
3138 * was unread data in the receive queue. This behavior is recommended
3139 * by RFC 2525, section 2.17. -DaveM
3141 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3143 struct sk_buff *skb;
3145 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3147 /* NOTE: No TCP options attached and we never retransmit this. */
3148 skb = alloc_skb(MAX_TCP_HEADER, priority);
3150 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3154 /* Reserve space for headers and prepare control bits. */
3155 skb_reserve(skb, MAX_TCP_HEADER);
3156 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3157 TCPHDR_ACK | TCPHDR_RST);
3158 tcp_mstamp_refresh(tcp_sk(sk));
3160 if (tcp_transmit_skb(sk, skb, 0, priority))
3161 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3164 /* Send a crossed SYN-ACK during socket establishment.
3165 * WARNING: This routine must only be called when we have already sent
3166 * a SYN packet that crossed the incoming SYN that caused this routine
3167 * to get called. If this assumption fails then the initial rcv_wnd
3168 * and rcv_wscale values will not be correct.
3170 int tcp_send_synack(struct sock *sk)
3172 struct sk_buff *skb;
3174 skb = tcp_write_queue_head(sk);
3175 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3176 pr_debug("%s: wrong queue state\n", __func__);
3179 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3180 if (skb_cloned(skb)) {
3181 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3184 tcp_unlink_write_queue(skb, sk);
3185 __skb_header_release(nskb);
3186 __tcp_add_write_queue_head(sk, nskb);
3187 sk_wmem_free_skb(sk, skb);
3188 sk->sk_wmem_queued += nskb->truesize;
3189 sk_mem_charge(sk, nskb->truesize);
3193 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3194 tcp_ecn_send_synack(sk, skb);
3196 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3200 * tcp_make_synack - Prepare a SYN-ACK.
3201 * sk: listener socket
3202 * dst: dst entry attached to the SYNACK
3203 * req: request_sock pointer
3205 * Allocate one skb and build a SYNACK packet.
3206 * @dst is consumed : Caller should not use it again.
3208 struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3209 struct request_sock *req,
3210 struct tcp_fastopen_cookie *foc,
3211 enum tcp_synack_type synack_type)
3213 struct inet_request_sock *ireq = inet_rsk(req);
3214 const struct tcp_sock *tp = tcp_sk(sk);
3215 struct tcp_md5sig_key *md5 = NULL;
3216 struct tcp_out_options opts;
3217 struct sk_buff *skb;
3218 int tcp_header_size;
3222 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3223 if (unlikely(!skb)) {
3227 /* Reserve space for headers. */
3228 skb_reserve(skb, MAX_TCP_HEADER);
3230 switch (synack_type) {
3231 case TCP_SYNACK_NORMAL:
3232 skb_set_owner_w(skb, req_to_sk(req));
3234 case TCP_SYNACK_COOKIE:
3235 /* Under synflood, we do not attach skb to a socket,
3236 * to avoid false sharing.
3239 case TCP_SYNACK_FASTOPEN:
3240 /* sk is a const pointer, because we want to express multiple
3241 * cpu might call us concurrently.
3242 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3244 skb_set_owner_w(skb, (struct sock *)sk);
3247 skb_dst_set(skb, dst);
3249 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3251 memset(&opts, 0, sizeof(opts));
3252 #ifdef CONFIG_SYN_COOKIES
3253 if (unlikely(req->cookie_ts))
3254 skb->skb_mstamp = cookie_init_timestamp(req);
3257 skb->skb_mstamp = tcp_clock_us();
3259 #ifdef CONFIG_TCP_MD5SIG
3261 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3263 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3264 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5,
3265 foc, synack_type) + sizeof(*th);
3267 skb_push(skb, tcp_header_size);
3268 skb_reset_transport_header(skb);
3270 th = (struct tcphdr *)skb->data;
3271 memset(th, 0, sizeof(struct tcphdr));
3274 tcp_ecn_make_synack(req, th);
3275 th->source = htons(ireq->ir_num);
3276 th->dest = ireq->ir_rmt_port;
3277 skb->mark = ireq->ir_mark;
3278 skb->ip_summed = CHECKSUM_PARTIAL;
3279 th->seq = htonl(tcp_rsk(req)->snt_isn);
3280 /* XXX data is queued and acked as is. No buffer/window check */
3281 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3283 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3284 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3285 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3286 th->doff = (tcp_header_size >> 2);
3287 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3289 #ifdef CONFIG_TCP_MD5SIG
3290 /* Okay, we have all we need - do the md5 hash if needed */
3292 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3293 md5, req_to_sk(req), skb);
3297 /* Do not fool tcpdump (if any), clean our debris */
3301 EXPORT_SYMBOL(tcp_make_synack);
3303 static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3305 struct inet_connection_sock *icsk = inet_csk(sk);
3306 const struct tcp_congestion_ops *ca;
3307 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3309 if (ca_key == TCP_CA_UNSPEC)
3313 ca = tcp_ca_find_key(ca_key);
3314 if (likely(ca && try_module_get(ca->owner))) {
3315 module_put(icsk->icsk_ca_ops->owner);
3316 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3317 icsk->icsk_ca_ops = ca;
3322 /* Do all connect socket setups that can be done AF independent. */
3323 static void tcp_connect_init(struct sock *sk)
3325 const struct dst_entry *dst = __sk_dst_get(sk);
3326 struct tcp_sock *tp = tcp_sk(sk);
3330 /* We'll fix this up when we get a response from the other end.
3331 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3333 tp->tcp_header_len = sizeof(struct tcphdr);
3334 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3335 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3337 #ifdef CONFIG_TCP_MD5SIG
3338 if (tp->af_specific->md5_lookup(sk, sk))
3339 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3342 /* If user gave his TCP_MAXSEG, record it to clamp */
3343 if (tp->rx_opt.user_mss)
3344 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3347 tcp_sync_mss(sk, dst_mtu(dst));
3349 tcp_ca_dst_init(sk, dst);
3351 if (!tp->window_clamp)
3352 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3353 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3355 tcp_initialize_rcv_mss(sk);
3357 /* limit the window selection if the user enforce a smaller rx buffer */
3358 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3359 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3360 tp->window_clamp = tcp_full_space(sk);
3362 rcv_wnd = tcp_rwnd_init_bpf(sk);
3364 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3366 tcp_select_initial_window(tcp_full_space(sk),
3367 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3370 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3374 tp->rx_opt.rcv_wscale = rcv_wscale;
3375 tp->rcv_ssthresh = tp->rcv_wnd;
3378 sock_reset_flag(sk, SOCK_DONE);
3381 tcp_write_queue_purge(sk);
3382 tp->snd_una = tp->write_seq;
3383 tp->snd_sml = tp->write_seq;
3384 tp->snd_up = tp->write_seq;
3385 tp->snd_nxt = tp->write_seq;
3387 if (likely(!tp->repair))
3390 tp->rcv_tstamp = tcp_jiffies32;
3391 tp->rcv_wup = tp->rcv_nxt;
3392 tp->copied_seq = tp->rcv_nxt;
3394 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3395 inet_csk(sk)->icsk_retransmits = 0;
3396 tcp_clear_retrans(tp);
3399 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3401 struct tcp_sock *tp = tcp_sk(sk);
3402 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3404 tcb->end_seq += skb->len;
3405 __skb_header_release(skb);
3406 __tcp_add_write_queue_tail(sk, skb);
3407 sk->sk_wmem_queued += skb->truesize;
3408 sk_mem_charge(sk, skb->truesize);
3409 tp->write_seq = tcb->end_seq;
3410 tp->packets_out += tcp_skb_pcount(skb);
3413 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
3414 * queue a data-only packet after the regular SYN, such that regular SYNs
3415 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3416 * only the SYN sequence, the data are retransmitted in the first ACK.
3417 * If cookie is not cached or other error occurs, falls back to send a
3418 * regular SYN with Fast Open cookie request option.
3420 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3422 struct tcp_sock *tp = tcp_sk(sk);
3423 struct tcp_fastopen_request *fo = tp->fastopen_req;
3425 struct sk_buff *syn_data;
3427 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3428 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3431 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3432 * user-MSS. Reserve maximum option space for middleboxes that add
3433 * private TCP options. The cost is reduced data space in SYN :(
3435 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3437 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3438 MAX_TCP_OPTION_SPACE;
3440 space = min_t(size_t, space, fo->size);
3442 /* limit to order-0 allocations */
3443 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3445 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3448 syn_data->ip_summed = CHECKSUM_PARTIAL;
3449 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3451 int copied = copy_from_iter(skb_put(syn_data, space), space,
3452 &fo->data->msg_iter);
3453 if (unlikely(!copied)) {
3454 kfree_skb(syn_data);
3457 if (copied != space) {
3458 skb_trim(syn_data, copied);
3462 /* No more data pending in inet_wait_for_connect() */
3463 if (space == fo->size)
3467 tcp_connect_queue_skb(sk, syn_data);
3469 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3471 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3473 syn->skb_mstamp = syn_data->skb_mstamp;
3475 /* Now full SYN+DATA was cloned and sent (or not),
3476 * remove the SYN from the original skb (syn_data)
3477 * we keep in write queue in case of a retransmit, as we
3478 * also have the SYN packet (with no data) in the same queue.
3480 TCP_SKB_CB(syn_data)->seq++;
3481 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3483 tp->syn_data = (fo->copied > 0);
3484 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3488 /* data was not sent, this is our new send_head */
3489 sk->sk_send_head = syn_data;
3490 tp->packets_out -= tcp_skb_pcount(syn_data);
3493 /* Send a regular SYN with Fast Open cookie request option */
3494 if (fo->cookie.len > 0)
3496 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3498 tp->syn_fastopen = 0;
3500 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3504 /* Build a SYN and send it off. */
3505 int tcp_connect(struct sock *sk)
3507 struct tcp_sock *tp = tcp_sk(sk);
3508 struct sk_buff *buff;
3511 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB);
3513 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3514 return -EHOSTUNREACH; /* Routing failure or similar. */
3516 tcp_connect_init(sk);
3518 if (unlikely(tp->repair)) {
3519 tcp_finish_connect(sk, NULL);
3523 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3524 if (unlikely(!buff))
3527 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3528 tcp_mstamp_refresh(tp);
3529 tp->retrans_stamp = tcp_time_stamp(tp);
3530 tcp_connect_queue_skb(sk, buff);
3531 tcp_ecn_send_syn(sk, buff);
3533 /* Send off SYN; include data in Fast Open. */
3534 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3535 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3536 if (err == -ECONNREFUSED)
3539 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3540 * in order to make this packet get counted in tcpOutSegs.
3542 tp->snd_nxt = tp->write_seq;
3543 tp->pushed_seq = tp->write_seq;
3544 buff = tcp_send_head(sk);
3545 if (unlikely(buff)) {
3546 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3547 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3549 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3551 /* Timer for repeating the SYN until an answer. */
3552 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3553 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3556 EXPORT_SYMBOL(tcp_connect);
3558 /* Send out a delayed ack, the caller does the policy checking
3559 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3562 void tcp_send_delayed_ack(struct sock *sk)
3564 struct inet_connection_sock *icsk = inet_csk(sk);
3565 int ato = icsk->icsk_ack.ato;
3566 unsigned long timeout;
3568 if (ato > TCP_DELACK_MIN) {
3569 const struct tcp_sock *tp = tcp_sk(sk);
3570 int max_ato = HZ / 2;
3572 if (icsk->icsk_ack.pingpong ||
3573 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3574 max_ato = TCP_DELACK_MAX;
3576 /* Slow path, intersegment interval is "high". */
3578 /* If some rtt estimate is known, use it to bound delayed ack.
3579 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3583 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3590 ato = min(ato, max_ato);
3593 /* Stay within the limit we were given */
3594 timeout = jiffies + ato;
3596 /* Use new timeout only if there wasn't a older one earlier. */
3597 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3598 /* If delack timer was blocked or is about to expire,
3601 if (icsk->icsk_ack.blocked ||
3602 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3607 if (!time_before(timeout, icsk->icsk_ack.timeout))
3608 timeout = icsk->icsk_ack.timeout;
3610 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3611 icsk->icsk_ack.timeout = timeout;
3612 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3615 /* This routine sends an ack and also updates the window. */
3616 void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3618 struct sk_buff *buff;
3620 /* If we have been reset, we may not send again. */
3621 if (sk->sk_state == TCP_CLOSE)
3624 /* We are not putting this on the write queue, so
3625 * tcp_transmit_skb() will set the ownership to this
3628 buff = alloc_skb(MAX_TCP_HEADER,
3629 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3630 if (unlikely(!buff)) {
3631 inet_csk_schedule_ack(sk);
3632 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3633 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3634 TCP_DELACK_MAX, TCP_RTO_MAX);
3638 /* Reserve space for headers and prepare control bits. */
3639 skb_reserve(buff, MAX_TCP_HEADER);
3640 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3642 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3644 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3646 skb_set_tcp_pure_ack(buff);
3648 /* Send it off, this clears delayed acks for us. */
3649 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3651 EXPORT_SYMBOL_GPL(__tcp_send_ack);
3653 void tcp_send_ack(struct sock *sk)
3655 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3658 /* This routine sends a packet with an out of date sequence
3659 * number. It assumes the other end will try to ack it.
3661 * Question: what should we make while urgent mode?
3662 * 4.4BSD forces sending single byte of data. We cannot send
3663 * out of window data, because we have SND.NXT==SND.MAX...
3665 * Current solution: to send TWO zero-length segments in urgent mode:
3666 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3667 * out-of-date with SND.UNA-1 to probe window.
3669 static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3671 struct tcp_sock *tp = tcp_sk(sk);
3672 struct sk_buff *skb;
3674 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3675 skb = alloc_skb(MAX_TCP_HEADER,
3676 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3680 /* Reserve space for headers and set control bits. */
3681 skb_reserve(skb, MAX_TCP_HEADER);
3682 /* Use a previous sequence. This should cause the other
3683 * end to send an ack. Don't queue or clone SKB, just
3686 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3687 NET_INC_STATS(sock_net(sk), mib);
3688 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3691 /* Called from setsockopt( ... TCP_REPAIR ) */
3692 void tcp_send_window_probe(struct sock *sk)
3694 if (sk->sk_state == TCP_ESTABLISHED) {
3695 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3696 tcp_mstamp_refresh(tcp_sk(sk));
3697 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3701 /* Initiate keepalive or window probe from timer. */
3702 int tcp_write_wakeup(struct sock *sk, int mib)
3704 struct tcp_sock *tp = tcp_sk(sk);
3705 struct sk_buff *skb;
3707 if (sk->sk_state == TCP_CLOSE)
3710 skb = tcp_send_head(sk);
3711 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3713 unsigned int mss = tcp_current_mss(sk);
3714 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3716 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3717 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3719 /* We are probing the opening of a window
3720 * but the window size is != 0
3721 * must have been a result SWS avoidance ( sender )
3723 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3725 seg_size = min(seg_size, mss);
3726 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3727 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3729 } else if (!tcp_skb_pcount(skb))
3730 tcp_set_skb_tso_segs(skb, mss);
3732 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3733 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3735 tcp_event_new_data_sent(sk, skb);
3738 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3739 tcp_xmit_probe_skb(sk, 1, mib);
3740 return tcp_xmit_probe_skb(sk, 0, mib);
3744 /* A window probe timeout has occurred. If window is not closed send
3745 * a partial packet else a zero probe.
3747 void tcp_send_probe0(struct sock *sk)
3749 struct inet_connection_sock *icsk = inet_csk(sk);
3750 struct tcp_sock *tp = tcp_sk(sk);
3751 struct net *net = sock_net(sk);
3752 unsigned long probe_max;
3755 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3757 if (tp->packets_out || !tcp_send_head(sk)) {
3758 /* Cancel probe timer, if it is not required. */
3759 icsk->icsk_probes_out = 0;
3760 icsk->icsk_backoff = 0;
3765 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3766 icsk->icsk_backoff++;
3767 icsk->icsk_probes_out++;
3768 probe_max = TCP_RTO_MAX;
3770 /* If packet was not sent due to local congestion,
3771 * do not backoff and do not remember icsk_probes_out.
3772 * Let local senders to fight for local resources.
3774 * Use accumulated backoff yet.
3776 if (!icsk->icsk_probes_out)
3777 icsk->icsk_probes_out = 1;
3778 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3780 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3781 tcp_probe0_when(sk, probe_max),
3785 int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3787 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3791 tcp_rsk(req)->txhash = net_tx_rndhash();
3792 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3794 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3795 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3796 if (unlikely(tcp_passive_fastopen(sk)))
3797 tcp_sk(sk)->total_retrans++;
3801 EXPORT_SYMBOL(tcp_rtx_synack);