Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   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.
   5 *
   6 *		Implementation of the Transmission Control Protocol(TCP).
   7 *
   8 * Authors:	Ross Biro
   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>
  19 */
  20
  21/*
  22 * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
  23 *				:	Fragmentation on mtu decrease
  24 *				:	Segment collapse on retransmit
  25 *				:	AF independence
  26 *
  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
  34 *
  35 */
  36
  37#define pr_fmt(fmt) "TCP: " fmt
  38
  39#include <net/tcp.h>
 
  40
  41#include <linux/compiler.h>
  42#include <linux/gfp.h>
  43#include <linux/module.h>
  44#include <linux/static_key.h>
  45
  46#include <trace/events/tcp.h>
  47
 
 
 
 
 
 
 
 
 
 
 
  48static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  49			   int push_one, gfp_t gfp);
  50
  51/* Account for new data that has been sent to the network. */
  52static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  53{
  54	struct inet_connection_sock *icsk = inet_csk(sk);
  55	struct tcp_sock *tp = tcp_sk(sk);
  56	unsigned int prior_packets = tp->packets_out;
  57
  58	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  59
  60	__skb_unlink(skb, &sk->sk_write_queue);
  61	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  62
 
 
 
  63	tp->packets_out += tcp_skb_pcount(skb);
  64	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  65		tcp_rearm_rto(sk);
  66
  67	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  68		      tcp_skb_pcount(skb));
  69}
  70
  71/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  72 * window scaling factor due to loss of precision.
  73 * If window has been shrunk, what should we make? It is not clear at all.
  74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  76 * invalid. OK, let's make this for now:
  77 */
  78static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  79{
  80	const struct tcp_sock *tp = tcp_sk(sk);
  81
  82	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
  83	    (tp->rx_opt.wscale_ok &&
  84	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
  85		return tp->snd_nxt;
  86	else
  87		return tcp_wnd_end(tp);
  88}
  89
  90/* Calculate mss to advertise in SYN segment.
  91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
  92 *
  93 * 1. It is independent of path mtu.
  94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
  95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
  96 *    attached devices, because some buggy hosts are confused by
  97 *    large MSS.
  98 * 4. We do not make 3, we advertise MSS, calculated from first
  99 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 100 *    This may be overridden via information stored in routing table.
 101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 102 *    probably even Jumbo".
 103 */
 104static __u16 tcp_advertise_mss(struct sock *sk)
 105{
 106	struct tcp_sock *tp = tcp_sk(sk);
 107	const struct dst_entry *dst = __sk_dst_get(sk);
 108	int mss = tp->advmss;
 109
 110	if (dst) {
 111		unsigned int metric = dst_metric_advmss(dst);
 112
 113		if (metric < mss) {
 114			mss = metric;
 115			tp->advmss = mss;
 116		}
 117	}
 118
 119	return (__u16)mss;
 120}
 121
 122/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 123 * This is the first part of cwnd validation mechanism.
 124 */
 125void tcp_cwnd_restart(struct sock *sk, s32 delta)
 126{
 127	struct tcp_sock *tp = tcp_sk(sk);
 128	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 129	u32 cwnd = tp->snd_cwnd;
 130
 131	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 132
 133	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 134	restart_cwnd = min(restart_cwnd, cwnd);
 135
 136	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 137		cwnd >>= 1;
 138	tp->snd_cwnd = max(cwnd, restart_cwnd);
 139	tp->snd_cwnd_stamp = tcp_jiffies32;
 140	tp->snd_cwnd_used = 0;
 141}
 142
 143/* Congestion state accounting after a packet has been sent. */
 144static void tcp_event_data_sent(struct tcp_sock *tp,
 145				struct sock *sk)
 146{
 147	struct inet_connection_sock *icsk = inet_csk(sk);
 148	const u32 now = tcp_jiffies32;
 149
 150	if (tcp_packets_in_flight(tp) == 0)
 151		tcp_ca_event(sk, CA_EVENT_TX_START);
 152
 153	tp->lsndtime = now;
 
 
 
 
 
 
 
 154
 155	/* If it is a reply for ato after last received
 156	 * packet, enter pingpong mode.
 157	 */
 158	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 159		icsk->icsk_ack.pingpong = 1;
 160}
 161
 162/* Account for an ACK we sent. */
 163static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
 
 164{
 165	tcp_dec_quickack_mode(sk, pkts);
 166	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 167}
 168
 
 
 
 
 
 
 
 169
 170u32 tcp_default_init_rwnd(u32 mss)
 171{
 172	/* Initial receive window should be twice of TCP_INIT_CWND to
 173	 * enable proper sending of new unsent data during fast recovery
 174	 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
 175	 * limit when mss is larger than 1460.
 176	 */
 177	u32 init_rwnd = TCP_INIT_CWND * 2;
 178
 179	if (mss > 1460)
 180		init_rwnd = max((1460 * init_rwnd) / mss, 2U);
 181	return init_rwnd;
 182}
 183
 184/* Determine a window scaling and initial window to offer.
 185 * Based on the assumption that the given amount of space
 186 * will be offered. Store the results in the tp structure.
 187 * NOTE: for smooth operation initial space offering should
 188 * be a multiple of mss if possible. We assume here that mss >= 1.
 189 * This MUST be enforced by all callers.
 190 */
 191void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 192			       __u32 *rcv_wnd, __u32 *window_clamp,
 193			       int wscale_ok, __u8 *rcv_wscale,
 194			       __u32 init_rcv_wnd)
 195{
 196	unsigned int space = (__space < 0 ? 0 : __space);
 197
 198	/* If no clamp set the clamp to the max possible scaled window */
 199	if (*window_clamp == 0)
 200		(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
 201	space = min(*window_clamp, space);
 202
 203	/* Quantize space offering to a multiple of mss if possible. */
 204	if (space > mss)
 205		space = rounddown(space, mss);
 206
 207	/* NOTE: offering an initial window larger than 32767
 208	 * will break some buggy TCP stacks. If the admin tells us
 209	 * it is likely we could be speaking with such a buggy stack
 210	 * we will truncate our initial window offering to 32K-1
 211	 * unless the remote has sent us a window scaling option,
 212	 * which we interpret as a sign the remote TCP is not
 213	 * misinterpreting the window field as a signed quantity.
 214	 */
 215	if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 216		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 217	else
 218		(*rcv_wnd) = space;
 
 
 
 219
 220	(*rcv_wscale) = 0;
 221	if (wscale_ok) {
 222		/* Set window scaling on max possible window */
 223		space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
 224		space = max_t(u32, space, sysctl_rmem_max);
 225		space = min_t(u32, space, *window_clamp);
 226		while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
 227			space >>= 1;
 228			(*rcv_wscale)++;
 229		}
 230	}
 231
 232	if (mss > (1 << *rcv_wscale)) {
 233		if (!init_rcv_wnd) /* Use default unless specified otherwise */
 234			init_rcv_wnd = tcp_default_init_rwnd(mss);
 235		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 236	}
 237
 238	/* Set the clamp no higher than max representable value */
 239	(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
 240}
 241EXPORT_SYMBOL(tcp_select_initial_window);
 242
 243/* Chose a new window to advertise, update state in tcp_sock for the
 244 * socket, and return result with RFC1323 scaling applied.  The return
 245 * value can be stuffed directly into th->window for an outgoing
 246 * frame.
 247 */
 248static u16 tcp_select_window(struct sock *sk)
 249{
 250	struct tcp_sock *tp = tcp_sk(sk);
 251	u32 old_win = tp->rcv_wnd;
 252	u32 cur_win = tcp_receive_window(tp);
 253	u32 new_win = __tcp_select_window(sk);
 254
 255	/* Never shrink the offered window */
 256	if (new_win < cur_win) {
 257		/* Danger Will Robinson!
 258		 * Don't update rcv_wup/rcv_wnd here or else
 259		 * we will not be able to advertise a zero
 260		 * window in time.  --DaveM
 261		 *
 262		 * Relax Will Robinson.
 263		 */
 264		if (new_win == 0)
 265			NET_INC_STATS(sock_net(sk),
 266				      LINUX_MIB_TCPWANTZEROWINDOWADV);
 267		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 268	}
 269	tp->rcv_wnd = new_win;
 270	tp->rcv_wup = tp->rcv_nxt;
 271
 272	/* Make sure we do not exceed the maximum possible
 273	 * scaled window.
 274	 */
 275	if (!tp->rx_opt.rcv_wscale &&
 276	    sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 277		new_win = min(new_win, MAX_TCP_WINDOW);
 278	else
 279		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 280
 281	/* RFC1323 scaling applied */
 282	new_win >>= tp->rx_opt.rcv_wscale;
 283
 284	/* If we advertise zero window, disable fast path. */
 285	if (new_win == 0) {
 286		tp->pred_flags = 0;
 287		if (old_win)
 288			NET_INC_STATS(sock_net(sk),
 289				      LINUX_MIB_TCPTOZEROWINDOWADV);
 290	} else if (old_win == 0) {
 291		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
 292	}
 293
 294	return new_win;
 295}
 296
 297/* Packet ECN state for a SYN-ACK */
 298static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 299{
 300	const struct tcp_sock *tp = tcp_sk(sk);
 301
 302	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 303	if (!(tp->ecn_flags & TCP_ECN_OK))
 304		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 305	else if (tcp_ca_needs_ecn(sk) ||
 306		 tcp_bpf_ca_needs_ecn(sk))
 307		INET_ECN_xmit(sk);
 308}
 309
 310/* Packet ECN state for a SYN.  */
 311static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 312{
 313	struct tcp_sock *tp = tcp_sk(sk);
 314	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 315	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
 316		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 317
 318	if (!use_ecn) {
 319		const struct dst_entry *dst = __sk_dst_get(sk);
 320
 321		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 322			use_ecn = true;
 323	}
 324
 325	tp->ecn_flags = 0;
 326
 327	if (use_ecn) {
 328		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 329		tp->ecn_flags = TCP_ECN_OK;
 330		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 331			INET_ECN_xmit(sk);
 332	}
 333}
 334
 335static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 336{
 337	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
 338		/* tp->ecn_flags are cleared at a later point in time when
 339		 * SYN ACK is ultimatively being received.
 340		 */
 341		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 342}
 343
 344static void
 345tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 346{
 347	if (inet_rsk(req)->ecn_ok)
 348		th->ece = 1;
 349}
 350
 351/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 352 * be sent.
 353 */
 354static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 355			 struct tcphdr *th, int tcp_header_len)
 356{
 357	struct tcp_sock *tp = tcp_sk(sk);
 358
 359	if (tp->ecn_flags & TCP_ECN_OK) {
 360		/* Not-retransmitted data segment: set ECT and inject CWR. */
 361		if (skb->len != tcp_header_len &&
 362		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 363			INET_ECN_xmit(sk);
 364			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 365				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 366				th->cwr = 1;
 367				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 368			}
 369		} else if (!tcp_ca_needs_ecn(sk)) {
 370			/* ACK or retransmitted segment: clear ECT|CE */
 371			INET_ECN_dontxmit(sk);
 372		}
 373		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 374			th->ece = 1;
 375	}
 376}
 377
 378/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 379 * auto increment end seqno.
 380 */
 381static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 382{
 383	skb->ip_summed = CHECKSUM_PARTIAL;
 384
 385	TCP_SKB_CB(skb)->tcp_flags = flags;
 386	TCP_SKB_CB(skb)->sacked = 0;
 387
 388	tcp_skb_pcount_set(skb, 1);
 389
 390	TCP_SKB_CB(skb)->seq = seq;
 391	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 392		seq++;
 393	TCP_SKB_CB(skb)->end_seq = seq;
 394}
 395
 396static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 397{
 398	return tp->snd_una != tp->snd_up;
 399}
 400
 401#define OPTION_SACK_ADVERTISE	(1 << 0)
 402#define OPTION_TS		(1 << 1)
 403#define OPTION_MD5		(1 << 2)
 404#define OPTION_WSCALE		(1 << 3)
 405#define OPTION_FAST_OPEN_COOKIE	(1 << 8)
 406#define OPTION_SMC		(1 << 9)
 
 407
 408static void smc_options_write(__be32 *ptr, u16 *options)
 409{
 410#if IS_ENABLED(CONFIG_SMC)
 411	if (static_branch_unlikely(&tcp_have_smc)) {
 412		if (unlikely(OPTION_SMC & *options)) {
 413			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 414				       (TCPOPT_NOP  << 16) |
 415				       (TCPOPT_EXP <<  8) |
 416				       (TCPOLEN_EXP_SMC_BASE));
 417			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 418		}
 419	}
 420#endif
 421}
 422
 423struct tcp_out_options {
 424	u16 options;		/* bit field of OPTION_* */
 425	u16 mss;		/* 0 to disable */
 426	u8 ws;			/* window scale, 0 to disable */
 427	u8 num_sack_blocks;	/* number of SACK blocks to include */
 428	u8 hash_size;		/* bytes in hash_location */
 429	__u8 *hash_location;	/* temporary pointer, overloaded */
 430	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 431	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 
 432};
 433
 
 
 
 
 
 
 
 
 434/* Write previously computed TCP options to the packet.
 435 *
 436 * Beware: Something in the Internet is very sensitive to the ordering of
 437 * TCP options, we learned this through the hard way, so be careful here.
 438 * Luckily we can at least blame others for their non-compliance but from
 439 * inter-operability perspective it seems that we're somewhat stuck with
 440 * the ordering which we have been using if we want to keep working with
 441 * those broken things (not that it currently hurts anybody as there isn't
 442 * particular reason why the ordering would need to be changed).
 443 *
 444 * At least SACK_PERM as the first option is known to lead to a disaster
 445 * (but it may well be that other scenarios fail similarly).
 446 */
 447static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 448			      struct tcp_out_options *opts)
 449{
 450	u16 options = opts->options;	/* mungable copy */
 451
 452	if (unlikely(OPTION_MD5 & options)) {
 453		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 454			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 455		/* overload cookie hash location */
 456		opts->hash_location = (__u8 *)ptr;
 457		ptr += 4;
 458	}
 459
 460	if (unlikely(opts->mss)) {
 461		*ptr++ = htonl((TCPOPT_MSS << 24) |
 462			       (TCPOLEN_MSS << 16) |
 463			       opts->mss);
 464	}
 465
 466	if (likely(OPTION_TS & options)) {
 467		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 468			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 469				       (TCPOLEN_SACK_PERM << 16) |
 470				       (TCPOPT_TIMESTAMP << 8) |
 471				       TCPOLEN_TIMESTAMP);
 472			options &= ~OPTION_SACK_ADVERTISE;
 473		} else {
 474			*ptr++ = htonl((TCPOPT_NOP << 24) |
 475				       (TCPOPT_NOP << 16) |
 476				       (TCPOPT_TIMESTAMP << 8) |
 477				       TCPOLEN_TIMESTAMP);
 478		}
 479		*ptr++ = htonl(opts->tsval);
 480		*ptr++ = htonl(opts->tsecr);
 481	}
 482
 483	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 484		*ptr++ = htonl((TCPOPT_NOP << 24) |
 485			       (TCPOPT_NOP << 16) |
 486			       (TCPOPT_SACK_PERM << 8) |
 487			       TCPOLEN_SACK_PERM);
 488	}
 489
 490	if (unlikely(OPTION_WSCALE & options)) {
 491		*ptr++ = htonl((TCPOPT_NOP << 24) |
 492			       (TCPOPT_WINDOW << 16) |
 493			       (TCPOLEN_WINDOW << 8) |
 494			       opts->ws);
 495	}
 496
 497	if (unlikely(opts->num_sack_blocks)) {
 498		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 499			tp->duplicate_sack : tp->selective_acks;
 500		int this_sack;
 501
 502		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 503			       (TCPOPT_NOP  << 16) |
 504			       (TCPOPT_SACK <<  8) |
 505			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 506						     TCPOLEN_SACK_PERBLOCK)));
 507
 508		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 509		     ++this_sack) {
 510			*ptr++ = htonl(sp[this_sack].start_seq);
 511			*ptr++ = htonl(sp[this_sack].end_seq);
 512		}
 513
 514		tp->rx_opt.dsack = 0;
 515	}
 516
 517	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 518		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 519		u8 *p = (u8 *)ptr;
 520		u32 len; /* Fast Open option length */
 521
 522		if (foc->exp) {
 523			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 524			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 525				     TCPOPT_FASTOPEN_MAGIC);
 526			p += TCPOLEN_EXP_FASTOPEN_BASE;
 527		} else {
 528			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 529			*p++ = TCPOPT_FASTOPEN;
 530			*p++ = len;
 531		}
 532
 533		memcpy(p, foc->val, foc->len);
 534		if ((len & 3) == 2) {
 535			p[foc->len] = TCPOPT_NOP;
 536			p[foc->len + 1] = TCPOPT_NOP;
 537		}
 538		ptr += (len + 3) >> 2;
 539	}
 540
 541	smc_options_write(ptr, &options);
 
 
 542}
 543
 544static void smc_set_option(const struct tcp_sock *tp,
 545			   struct tcp_out_options *opts,
 546			   unsigned int *remaining)
 547{
 548#if IS_ENABLED(CONFIG_SMC)
 549	if (static_branch_unlikely(&tcp_have_smc)) {
 550		if (tp->syn_smc) {
 551			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 552				opts->options |= OPTION_SMC;
 553				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 554			}
 555		}
 556	}
 557#endif
 558}
 559
 560static void smc_set_option_cond(const struct tcp_sock *tp,
 561				const struct inet_request_sock *ireq,
 562				struct tcp_out_options *opts,
 563				unsigned int *remaining)
 564{
 565#if IS_ENABLED(CONFIG_SMC)
 566	if (static_branch_unlikely(&tcp_have_smc)) {
 567		if (tp->syn_smc && ireq->smc_ok) {
 568			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 569				opts->options |= OPTION_SMC;
 570				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 571			}
 572		}
 573	}
 574#endif
 575}
 576
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 577/* Compute TCP options for SYN packets. This is not the final
 578 * network wire format yet.
 579 */
 580static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 581				struct tcp_out_options *opts,
 582				struct tcp_md5sig_key **md5)
 583{
 584	struct tcp_sock *tp = tcp_sk(sk);
 585	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 586	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 587
 
 588#ifdef CONFIG_TCP_MD5SIG
 589	*md5 = tp->af_specific->md5_lookup(sk, sk);
 590	if (*md5) {
 591		opts->options |= OPTION_MD5;
 592		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 
 
 
 593	}
 594#else
 595	*md5 = NULL;
 596#endif
 597
 598	/* We always get an MSS option.  The option bytes which will be seen in
 599	 * normal data packets should timestamps be used, must be in the MSS
 600	 * advertised.  But we subtract them from tp->mss_cache so that
 601	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 602	 * fact here if necessary.  If we don't do this correctly, as a
 603	 * receiver we won't recognize data packets as being full sized when we
 604	 * should, and thus we won't abide by the delayed ACK rules correctly.
 605	 * SACKs don't matter, we never delay an ACK when we have any of those
 606	 * going out.  */
 607	opts->mss = tcp_advertise_mss(sk);
 608	remaining -= TCPOLEN_MSS_ALIGNED;
 609
 610	if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
 611		opts->options |= OPTION_TS;
 612		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
 613		opts->tsecr = tp->rx_opt.ts_recent;
 614		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 615	}
 616	if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
 617		opts->ws = tp->rx_opt.rcv_wscale;
 618		opts->options |= OPTION_WSCALE;
 619		remaining -= TCPOLEN_WSCALE_ALIGNED;
 620	}
 621	if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
 622		opts->options |= OPTION_SACK_ADVERTISE;
 623		if (unlikely(!(OPTION_TS & opts->options)))
 624			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 625	}
 626
 627	if (fastopen && fastopen->cookie.len >= 0) {
 628		u32 need = fastopen->cookie.len;
 629
 630		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 631					       TCPOLEN_FASTOPEN_BASE;
 632		need = (need + 3) & ~3U;  /* Align to 32 bits */
 633		if (remaining >= need) {
 634			opts->options |= OPTION_FAST_OPEN_COOKIE;
 635			opts->fastopen_cookie = &fastopen->cookie;
 636			remaining -= need;
 637			tp->syn_fastopen = 1;
 638			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 639		}
 640	}
 641
 642	smc_set_option(tp, opts, &remaining);
 643
 
 
 
 
 
 
 
 
 
 644	return MAX_TCP_OPTION_SPACE - remaining;
 645}
 646
 647/* Set up TCP options for SYN-ACKs. */
 648static unsigned int tcp_synack_options(const struct sock *sk,
 649				       struct request_sock *req,
 650				       unsigned int mss, struct sk_buff *skb,
 651				       struct tcp_out_options *opts,
 652				       const struct tcp_md5sig_key *md5,
 653				       struct tcp_fastopen_cookie *foc)
 
 654{
 655	struct inet_request_sock *ireq = inet_rsk(req);
 656	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 657
 658#ifdef CONFIG_TCP_MD5SIG
 659	if (md5) {
 660		opts->options |= OPTION_MD5;
 661		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 662
 663		/* We can't fit any SACK blocks in a packet with MD5 + TS
 664		 * options. There was discussion about disabling SACK
 665		 * rather than TS in order to fit in better with old,
 666		 * buggy kernels, but that was deemed to be unnecessary.
 667		 */
 668		ireq->tstamp_ok &= !ireq->sack_ok;
 
 669	}
 670#endif
 671
 672	/* We always send an MSS option. */
 673	opts->mss = mss;
 674	remaining -= TCPOLEN_MSS_ALIGNED;
 675
 676	if (likely(ireq->wscale_ok)) {
 677		opts->ws = ireq->rcv_wscale;
 678		opts->options |= OPTION_WSCALE;
 679		remaining -= TCPOLEN_WSCALE_ALIGNED;
 680	}
 681	if (likely(ireq->tstamp_ok)) {
 682		opts->options |= OPTION_TS;
 683		opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
 684		opts->tsecr = req->ts_recent;
 685		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 686	}
 687	if (likely(ireq->sack_ok)) {
 688		opts->options |= OPTION_SACK_ADVERTISE;
 689		if (unlikely(!ireq->tstamp_ok))
 690			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 691	}
 692	if (foc != NULL && foc->len >= 0) {
 693		u32 need = foc->len;
 694
 695		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 696				   TCPOLEN_FASTOPEN_BASE;
 697		need = (need + 3) & ~3U;  /* Align to 32 bits */
 698		if (remaining >= need) {
 699			opts->options |= OPTION_FAST_OPEN_COOKIE;
 700			opts->fastopen_cookie = foc;
 701			remaining -= need;
 702		}
 703	}
 704
 
 
 705	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 706
 707	return MAX_TCP_OPTION_SPACE - remaining;
 708}
 709
 710/* Compute TCP options for ESTABLISHED sockets. This is not the
 711 * final wire format yet.
 712 */
 713static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 714					struct tcp_out_options *opts,
 715					struct tcp_md5sig_key **md5)
 716{
 717	struct tcp_sock *tp = tcp_sk(sk);
 718	unsigned int size = 0;
 719	unsigned int eff_sacks;
 720
 721	opts->options = 0;
 722
 
 723#ifdef CONFIG_TCP_MD5SIG
 724	*md5 = tp->af_specific->md5_lookup(sk, sk);
 725	if (unlikely(*md5)) {
 726		opts->options |= OPTION_MD5;
 727		size += TCPOLEN_MD5SIG_ALIGNED;
 
 
 
 728	}
 729#else
 730	*md5 = NULL;
 731#endif
 732
 733	if (likely(tp->rx_opt.tstamp_ok)) {
 734		opts->options |= OPTION_TS;
 735		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
 736		opts->tsecr = tp->rx_opt.ts_recent;
 737		size += TCPOLEN_TSTAMP_ALIGNED;
 738	}
 739
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 740	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 741	if (unlikely(eff_sacks)) {
 742		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 
 
 
 
 743		opts->num_sack_blocks =
 744			min_t(unsigned int, eff_sacks,
 745			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 746			      TCPOLEN_SACK_PERBLOCK);
 
 747		size += TCPOLEN_SACK_BASE_ALIGNED +
 748			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 749	}
 750
 751	return size;
 752}
 753
 754
 755/* TCP SMALL QUEUES (TSQ)
 756 *
 757 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 758 * to reduce RTT and bufferbloat.
 759 * We do this using a special skb destructor (tcp_wfree).
 760 *
 761 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 762 * needs to be reallocated in a driver.
 763 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
 764 *
 765 * Since transmit from skb destructor is forbidden, we use a tasklet
 766 * to process all sockets that eventually need to send more skbs.
 767 * We use one tasklet per cpu, with its own queue of sockets.
 768 */
 769struct tsq_tasklet {
 770	struct tasklet_struct	tasklet;
 771	struct list_head	head; /* queue of tcp sockets */
 772};
 773static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
 774
 775static void tcp_tsq_handler(struct sock *sk)
 776{
 777	if ((1 << sk->sk_state) &
 778	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
 779	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
 780		struct tcp_sock *tp = tcp_sk(sk);
 781
 782		if (tp->lost_out > tp->retrans_out &&
 783		    tp->snd_cwnd > tcp_packets_in_flight(tp)) {
 784			tcp_mstamp_refresh(tp);
 785			tcp_xmit_retransmit_queue(sk);
 786		}
 787
 788		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
 789			       0, GFP_ATOMIC);
 790	}
 791}
 
 
 
 
 
 
 
 
 
 
 792/*
 793 * One tasklet per cpu tries to send more skbs.
 794 * We run in tasklet context but need to disable irqs when
 795 * transferring tsq->head because tcp_wfree() might
 796 * interrupt us (non NAPI drivers)
 797 */
 798static void tcp_tasklet_func(unsigned long data)
 799{
 800	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
 801	LIST_HEAD(list);
 802	unsigned long flags;
 803	struct list_head *q, *n;
 804	struct tcp_sock *tp;
 805	struct sock *sk;
 806
 807	local_irq_save(flags);
 808	list_splice_init(&tsq->head, &list);
 809	local_irq_restore(flags);
 810
 811	list_for_each_safe(q, n, &list) {
 812		tp = list_entry(q, struct tcp_sock, tsq_node);
 813		list_del(&tp->tsq_node);
 814
 815		sk = (struct sock *)tp;
 816		smp_mb__before_atomic();
 817		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
 818
 819		if (!sk->sk_lock.owned &&
 820		    test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
 821			bh_lock_sock(sk);
 822			if (!sock_owned_by_user(sk)) {
 823				clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
 824				tcp_tsq_handler(sk);
 825			}
 826			bh_unlock_sock(sk);
 827		}
 828
 829		sk_free(sk);
 830	}
 831}
 832
 833#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
 834			  TCPF_WRITE_TIMER_DEFERRED |	\
 835			  TCPF_DELACK_TIMER_DEFERRED |	\
 836			  TCPF_MTU_REDUCED_DEFERRED)
 837/**
 838 * tcp_release_cb - tcp release_sock() callback
 839 * @sk: socket
 840 *
 841 * called from release_sock() to perform protocol dependent
 842 * actions before socket release.
 843 */
 844void tcp_release_cb(struct sock *sk)
 845{
 846	unsigned long flags, nflags;
 847
 848	/* perform an atomic operation only if at least one flag is set */
 849	do {
 850		flags = sk->sk_tsq_flags;
 851		if (!(flags & TCP_DEFERRED_ALL))
 852			return;
 853		nflags = flags & ~TCP_DEFERRED_ALL;
 854	} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
 855
 856	if (flags & TCPF_TSQ_DEFERRED)
 857		tcp_tsq_handler(sk);
 858
 
 859	/* Here begins the tricky part :
 860	 * We are called from release_sock() with :
 861	 * 1) BH disabled
 862	 * 2) sk_lock.slock spinlock held
 863	 * 3) socket owned by us (sk->sk_lock.owned == 1)
 864	 *
 865	 * But following code is meant to be called from BH handlers,
 866	 * so we should keep BH disabled, but early release socket ownership
 867	 */
 868	sock_release_ownership(sk);
 869
 870	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 871		tcp_write_timer_handler(sk);
 872		__sock_put(sk);
 873	}
 874	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
 875		tcp_delack_timer_handler(sk);
 876		__sock_put(sk);
 877	}
 878	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
 879		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 880		__sock_put(sk);
 881	}
 882}
 883EXPORT_SYMBOL(tcp_release_cb);
 884
 885void __init tcp_tasklet_init(void)
 886{
 887	int i;
 888
 889	for_each_possible_cpu(i) {
 890		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
 891
 892		INIT_LIST_HEAD(&tsq->head);
 893		tasklet_init(&tsq->tasklet,
 894			     tcp_tasklet_func,
 895			     (unsigned long)tsq);
 896	}
 897}
 898
 899/*
 900 * Write buffer destructor automatically called from kfree_skb.
 901 * We can't xmit new skbs from this context, as we might already
 902 * hold qdisc lock.
 903 */
 904void tcp_wfree(struct sk_buff *skb)
 905{
 906	struct sock *sk = skb->sk;
 907	struct tcp_sock *tp = tcp_sk(sk);
 908	unsigned long flags, nval, oval;
 909
 910	/* Keep one reference on sk_wmem_alloc.
 911	 * Will be released by sk_free() from here or tcp_tasklet_func()
 912	 */
 913	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
 914
 915	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
 916	 * Wait until our queues (qdisc + devices) are drained.
 917	 * This gives :
 918	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
 919	 * - chance for incoming ACK (processed by another cpu maybe)
 920	 *   to migrate this flow (skb->ooo_okay will be eventually set)
 921	 */
 922	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
 923		goto out;
 924
 925	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
 926		struct tsq_tasklet *tsq;
 927		bool empty;
 928
 929		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
 930			goto out;
 931
 932		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
 933		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
 934		if (nval != oval)
 935			continue;
 936
 937		/* queue this socket to tasklet queue */
 938		local_irq_save(flags);
 939		tsq = this_cpu_ptr(&tsq_tasklet);
 940		empty = list_empty(&tsq->head);
 941		list_add(&tp->tsq_node, &tsq->head);
 942		if (empty)
 943			tasklet_schedule(&tsq->tasklet);
 944		local_irq_restore(flags);
 945		return;
 946	}
 947out:
 948	sk_free(sk);
 949}
 950
 951/* Note: Called under hard irq.
 952 * We can not call TCP stack right away.
 953 */
 954enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
 955{
 956	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
 957	struct sock *sk = (struct sock *)tp;
 958	unsigned long nval, oval;
 959
 960	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
 961		struct tsq_tasklet *tsq;
 962		bool empty;
 963
 964		if (oval & TSQF_QUEUED)
 965			break;
 966
 967		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
 968		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
 969		if (nval != oval)
 970			continue;
 971
 972		if (!refcount_inc_not_zero(&sk->sk_wmem_alloc))
 973			break;
 974		/* queue this socket to tasklet queue */
 975		tsq = this_cpu_ptr(&tsq_tasklet);
 976		empty = list_empty(&tsq->head);
 977		list_add(&tp->tsq_node, &tsq->head);
 978		if (empty)
 979			tasklet_schedule(&tsq->tasklet);
 980		break;
 981	}
 982	return HRTIMER_NORESTART;
 983}
 984
 985/* BBR congestion control needs pacing.
 986 * Same remark for SO_MAX_PACING_RATE.
 987 * sch_fq packet scheduler is efficiently handling pacing,
 988 * but is not always installed/used.
 989 * Return true if TCP stack should pace packets itself.
 990 */
 991static bool tcp_needs_internal_pacing(const struct sock *sk)
 992{
 993	return smp_load_acquire(&sk->sk_pacing_status) == SK_PACING_NEEDED;
 994}
 995
 996static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
 997{
 998	u64 len_ns;
 999	u32 rate;
1000
1001	if (!tcp_needs_internal_pacing(sk))
1002		return;
1003	rate = sk->sk_pacing_rate;
1004	if (!rate || rate == ~0U)
1005		return;
1006
1007	/* Should account for header sizes as sch_fq does,
1008	 * but lets make things simple.
1009	 */
1010	len_ns = (u64)skb->len * NSEC_PER_SEC;
1011	do_div(len_ns, rate);
1012	hrtimer_start(&tcp_sk(sk)->pacing_timer,
1013		      ktime_add_ns(ktime_get(), len_ns),
1014		      HRTIMER_MODE_ABS_PINNED);
1015}
1016
1017static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1018{
1019	skb->skb_mstamp = tp->tcp_mstamp;
 
 
1020	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1021}
1022
 
 
 
 
1023/* This routine actually transmits TCP packets queued in by
1024 * tcp_do_sendmsg().  This is used by both the initial
1025 * transmission and possible later retransmissions.
1026 * All SKB's seen here are completely headerless.  It is our
1027 * job to build the TCP header, and pass the packet down to
1028 * IP so it can do the same plus pass the packet off to the
1029 * device.
1030 *
1031 * We are working here with either a clone of the original
1032 * SKB, or a fresh unique copy made by the retransmit engine.
1033 */
1034static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1035			    gfp_t gfp_mask)
1036{
1037	const struct inet_connection_sock *icsk = inet_csk(sk);
1038	struct inet_sock *inet;
1039	struct tcp_sock *tp;
1040	struct tcp_skb_cb *tcb;
1041	struct tcp_out_options opts;
1042	unsigned int tcp_options_size, tcp_header_size;
1043	struct sk_buff *oskb = NULL;
1044	struct tcp_md5sig_key *md5;
1045	struct tcphdr *th;
 
1046	int err;
1047
1048	BUG_ON(!skb || !tcp_skb_pcount(skb));
1049	tp = tcp_sk(sk);
1050
 
 
1051	if (clone_it) {
1052		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1053			- tp->snd_una;
1054		oskb = skb;
1055
1056		tcp_skb_tsorted_save(oskb) {
1057			if (unlikely(skb_cloned(oskb)))
1058				skb = pskb_copy(oskb, gfp_mask);
1059			else
1060				skb = skb_clone(oskb, gfp_mask);
1061		} tcp_skb_tsorted_restore(oskb);
1062
1063		if (unlikely(!skb))
1064			return -ENOBUFS;
 
 
 
 
1065	}
1066	skb->skb_mstamp = tp->tcp_mstamp;
1067
1068	inet = inet_sk(sk);
1069	tcb = TCP_SKB_CB(skb);
1070	memset(&opts, 0, sizeof(opts));
1071
1072	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1073		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1074	else
1075		tcp_options_size = tcp_established_options(sk, skb, &opts,
1076							   &md5);
 
 
 
 
 
 
 
 
 
 
 
1077	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1078
1079	/* if no packet is in qdisc/device queue, then allow XPS to select
1080	 * another queue. We can be called from tcp_tsq_handler()
1081	 * which holds one reference to sk_wmem_alloc.
1082	 *
1083	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1084	 * One way to get this would be to set skb->truesize = 2 on them.
1085	 */
1086	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1087
1088	/* If we had to use memory reserve to allocate this skb,
1089	 * this might cause drops if packet is looped back :
1090	 * Other socket might not have SOCK_MEMALLOC.
1091	 * Packets not looped back do not care about pfmemalloc.
1092	 */
1093	skb->pfmemalloc = 0;
1094
1095	skb_push(skb, tcp_header_size);
1096	skb_reset_transport_header(skb);
1097
1098	skb_orphan(skb);
1099	skb->sk = sk;
1100	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1101	skb_set_hash_from_sk(skb, sk);
1102	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1103
1104	skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1105
1106	/* Build TCP header and checksum it. */
1107	th = (struct tcphdr *)skb->data;
1108	th->source		= inet->inet_sport;
1109	th->dest		= inet->inet_dport;
1110	th->seq			= htonl(tcb->seq);
1111	th->ack_seq		= htonl(tp->rcv_nxt);
1112	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1113					tcb->tcp_flags);
1114
1115	th->check		= 0;
1116	th->urg_ptr		= 0;
1117
1118	/* The urg_mode check is necessary during a below snd_una win probe */
1119	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1120		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1121			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1122			th->urg = 1;
1123		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1124			th->urg_ptr = htons(0xFFFF);
1125			th->urg = 1;
1126		}
1127	}
1128
1129	tcp_options_write((__be32 *)(th + 1), tp, &opts);
1130	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1131	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1132		th->window      = htons(tcp_select_window(sk));
1133		tcp_ecn_send(sk, skb, th, tcp_header_size);
1134	} else {
1135		/* RFC1323: The window in SYN & SYN/ACK segments
1136		 * is never scaled.
1137		 */
1138		th->window	= htons(min(tp->rcv_wnd, 65535U));
1139	}
1140#ifdef CONFIG_TCP_MD5SIG
1141	/* Calculate the MD5 hash, as we have all we need now */
1142	if (md5) {
1143		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1144		tp->af_specific->calc_md5_hash(opts.hash_location,
1145					       md5, sk, skb);
1146	}
1147#endif
1148
1149	icsk->icsk_af_ops->send_check(sk, skb);
 
 
1150
1151	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1152		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1153
1154	if (skb->len != tcp_header_size) {
1155		tcp_event_data_sent(tp, sk);
1156		tp->data_segs_out += tcp_skb_pcount(skb);
1157		tcp_internal_pacing(sk, skb);
1158	}
1159
1160	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1161		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1162			      tcp_skb_pcount(skb));
1163
1164	tp->segs_out += tcp_skb_pcount(skb);
1165	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1166	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1167	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1168
1169	/* Our usage of tstamp should remain private */
1170	skb->tstamp = 0;
1171
1172	/* Cleanup our debris for IP stacks */
1173	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1174			       sizeof(struct inet6_skb_parm)));
1175
1176	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
 
 
 
 
1177
1178	if (unlikely(err > 0)) {
1179		tcp_enter_cwr(sk);
1180		err = net_xmit_eval(err);
1181	}
1182	if (!err && oskb) {
1183		tcp_update_skb_after_send(tp, oskb);
1184		tcp_rate_skb_sent(sk, oskb);
1185	}
1186	return err;
1187}
1188
 
 
 
 
 
 
 
1189/* This routine just queues the buffer for sending.
1190 *
1191 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1192 * otherwise socket can stall.
1193 */
1194static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1195{
1196	struct tcp_sock *tp = tcp_sk(sk);
1197
1198	/* Advance write_seq and place onto the write_queue. */
1199	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1200	__skb_header_release(skb);
1201	tcp_add_write_queue_tail(sk, skb);
1202	sk->sk_wmem_queued += skb->truesize;
1203	sk_mem_charge(sk, skb->truesize);
1204}
1205
1206/* Initialize TSO segments for a packet. */
1207static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1208{
1209	if (skb->len <= mss_now) {
1210		/* Avoid the costly divide in the normal
1211		 * non-TSO case.
1212		 */
1213		tcp_skb_pcount_set(skb, 1);
1214		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1215	} else {
1216		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1217		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1218	}
1219}
1220
1221/* Pcount in the middle of the write queue got changed, we need to do various
1222 * tweaks to fix counters
1223 */
1224static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1225{
1226	struct tcp_sock *tp = tcp_sk(sk);
1227
1228	tp->packets_out -= decr;
1229
1230	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1231		tp->sacked_out -= decr;
1232	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1233		tp->retrans_out -= decr;
1234	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1235		tp->lost_out -= decr;
1236
1237	/* Reno case is special. Sigh... */
1238	if (tcp_is_reno(tp) && decr > 0)
1239		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1240
1241	if (tp->lost_skb_hint &&
1242	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1243	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1244		tp->lost_cnt_hint -= decr;
1245
1246	tcp_verify_left_out(tp);
1247}
1248
1249static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1250{
1251	return TCP_SKB_CB(skb)->txstamp_ack ||
1252		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1253}
1254
1255static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1256{
1257	struct skb_shared_info *shinfo = skb_shinfo(skb);
1258
1259	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1260	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1261		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1262		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1263
1264		shinfo->tx_flags &= ~tsflags;
1265		shinfo2->tx_flags |= tsflags;
1266		swap(shinfo->tskey, shinfo2->tskey);
1267		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1268		TCP_SKB_CB(skb)->txstamp_ack = 0;
1269	}
1270}
1271
1272static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1273{
1274	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1275	TCP_SKB_CB(skb)->eor = 0;
1276}
1277
1278/* Insert buff after skb on the write or rtx queue of sk.  */
1279static void tcp_insert_write_queue_after(struct sk_buff *skb,
1280					 struct sk_buff *buff,
1281					 struct sock *sk,
1282					 enum tcp_queue tcp_queue)
1283{
1284	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1285		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1286	else
1287		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1288}
1289
1290/* Function to create two new TCP segments.  Shrinks the given segment
1291 * to the specified size and appends a new segment with the rest of the
1292 * packet to the list.  This won't be called frequently, I hope.
1293 * Remember, these are still headerless SKBs at this point.
1294 */
1295int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1296		 struct sk_buff *skb, u32 len,
1297		 unsigned int mss_now, gfp_t gfp)
1298{
1299	struct tcp_sock *tp = tcp_sk(sk);
1300	struct sk_buff *buff;
1301	int nsize, old_factor;
 
1302	int nlen;
1303	u8 flags;
1304
1305	if (WARN_ON(len > skb->len))
1306		return -EINVAL;
1307
1308	nsize = skb_headlen(skb) - len;
1309	if (nsize < 0)
1310		nsize = 0;
1311
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1312	if (skb_unclone(skb, gfp))
1313		return -ENOMEM;
1314
1315	/* Get a new skb... force flag on. */
1316	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1317	if (!buff)
1318		return -ENOMEM; /* We'll just try again later. */
 
1319
1320	sk->sk_wmem_queued += buff->truesize;
1321	sk_mem_charge(sk, buff->truesize);
1322	nlen = skb->len - len - nsize;
1323	buff->truesize += nlen;
1324	skb->truesize -= nlen;
1325
1326	/* Correct the sequence numbers. */
1327	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1328	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1329	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1330
1331	/* PSH and FIN should only be set in the second packet. */
1332	flags = TCP_SKB_CB(skb)->tcp_flags;
1333	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1334	TCP_SKB_CB(buff)->tcp_flags = flags;
1335	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1336	tcp_skb_fragment_eor(skb, buff);
1337
1338	skb_split(skb, buff, len);
1339
1340	buff->ip_summed = CHECKSUM_PARTIAL;
1341
1342	buff->tstamp = skb->tstamp;
1343	tcp_fragment_tstamp(skb, buff);
1344
1345	old_factor = tcp_skb_pcount(skb);
1346
1347	/* Fix up tso_factor for both original and new SKB.  */
1348	tcp_set_skb_tso_segs(skb, mss_now);
1349	tcp_set_skb_tso_segs(buff, mss_now);
1350
1351	/* Update delivered info for the new segment */
1352	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1353
1354	/* If this packet has been sent out already, we must
1355	 * adjust the various packet counters.
1356	 */
1357	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1358		int diff = old_factor - tcp_skb_pcount(skb) -
1359			tcp_skb_pcount(buff);
1360
1361		if (diff)
1362			tcp_adjust_pcount(sk, skb, diff);
1363	}
1364
1365	/* Link BUFF into the send queue. */
1366	__skb_header_release(buff);
1367	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1368	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1369		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1370
1371	return 0;
1372}
1373
1374/* This is similar to __pskb_pull_tail(). The difference is that pulled
1375 * data is not copied, but immediately discarded.
1376 */
1377static int __pskb_trim_head(struct sk_buff *skb, int len)
1378{
1379	struct skb_shared_info *shinfo;
1380	int i, k, eat;
1381
1382	eat = min_t(int, len, skb_headlen(skb));
1383	if (eat) {
1384		__skb_pull(skb, eat);
1385		len -= eat;
1386		if (!len)
1387			return 0;
1388	}
1389	eat = len;
1390	k = 0;
1391	shinfo = skb_shinfo(skb);
1392	for (i = 0; i < shinfo->nr_frags; i++) {
1393		int size = skb_frag_size(&shinfo->frags[i]);
1394
1395		if (size <= eat) {
1396			skb_frag_unref(skb, i);
1397			eat -= size;
1398		} else {
1399			shinfo->frags[k] = shinfo->frags[i];
1400			if (eat) {
1401				shinfo->frags[k].page_offset += eat;
1402				skb_frag_size_sub(&shinfo->frags[k], eat);
1403				eat = 0;
1404			}
1405			k++;
1406		}
1407	}
1408	shinfo->nr_frags = k;
1409
1410	skb->data_len -= len;
1411	skb->len = skb->data_len;
1412	return len;
1413}
1414
1415/* Remove acked data from a packet in the transmit queue. */
1416int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1417{
1418	u32 delta_truesize;
1419
1420	if (skb_unclone(skb, GFP_ATOMIC))
1421		return -ENOMEM;
1422
1423	delta_truesize = __pskb_trim_head(skb, len);
1424
1425	TCP_SKB_CB(skb)->seq += len;
1426	skb->ip_summed = CHECKSUM_PARTIAL;
1427
1428	if (delta_truesize) {
1429		skb->truesize	   -= delta_truesize;
1430		sk->sk_wmem_queued -= delta_truesize;
1431		sk_mem_uncharge(sk, delta_truesize);
1432		sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1433	}
1434
1435	/* Any change of skb->len requires recalculation of tso factor. */
1436	if (tcp_skb_pcount(skb) > 1)
1437		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1438
1439	return 0;
1440}
1441
1442/* Calculate MSS not accounting any TCP options.  */
1443static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1444{
1445	const struct tcp_sock *tp = tcp_sk(sk);
1446	const struct inet_connection_sock *icsk = inet_csk(sk);
1447	int mss_now;
1448
1449	/* Calculate base mss without TCP options:
1450	   It is MMS_S - sizeof(tcphdr) of rfc1122
1451	 */
1452	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1453
1454	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1455	if (icsk->icsk_af_ops->net_frag_header_len) {
1456		const struct dst_entry *dst = __sk_dst_get(sk);
1457
1458		if (dst && dst_allfrag(dst))
1459			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1460	}
1461
1462	/* Clamp it (mss_clamp does not include tcp options) */
1463	if (mss_now > tp->rx_opt.mss_clamp)
1464		mss_now = tp->rx_opt.mss_clamp;
1465
1466	/* Now subtract optional transport overhead */
1467	mss_now -= icsk->icsk_ext_hdr_len;
1468
1469	/* Then reserve room for full set of TCP options and 8 bytes of data */
1470	if (mss_now < 48)
1471		mss_now = 48;
1472	return mss_now;
1473}
1474
1475/* Calculate MSS. Not accounting for SACKs here.  */
1476int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1477{
1478	/* Subtract TCP options size, not including SACKs */
1479	return __tcp_mtu_to_mss(sk, pmtu) -
1480	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1481}
1482
1483/* Inverse of above */
1484int tcp_mss_to_mtu(struct sock *sk, int mss)
1485{
1486	const struct tcp_sock *tp = tcp_sk(sk);
1487	const struct inet_connection_sock *icsk = inet_csk(sk);
1488	int mtu;
1489
1490	mtu = mss +
1491	      tp->tcp_header_len +
1492	      icsk->icsk_ext_hdr_len +
1493	      icsk->icsk_af_ops->net_header_len;
1494
1495	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1496	if (icsk->icsk_af_ops->net_frag_header_len) {
1497		const struct dst_entry *dst = __sk_dst_get(sk);
1498
1499		if (dst && dst_allfrag(dst))
1500			mtu += icsk->icsk_af_ops->net_frag_header_len;
1501	}
1502	return mtu;
1503}
1504EXPORT_SYMBOL(tcp_mss_to_mtu);
1505
1506/* MTU probing init per socket */
1507void tcp_mtup_init(struct sock *sk)
1508{
1509	struct tcp_sock *tp = tcp_sk(sk);
1510	struct inet_connection_sock *icsk = inet_csk(sk);
1511	struct net *net = sock_net(sk);
1512
1513	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1514	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1515			       icsk->icsk_af_ops->net_header_len;
1516	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1517	icsk->icsk_mtup.probe_size = 0;
1518	if (icsk->icsk_mtup.enabled)
1519		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1520}
1521EXPORT_SYMBOL(tcp_mtup_init);
1522
1523/* This function synchronize snd mss to current pmtu/exthdr set.
1524
1525   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1526   for TCP options, but includes only bare TCP header.
1527
1528   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1529   It is minimum of user_mss and mss received with SYN.
1530   It also does not include TCP options.
1531
1532   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1533
1534   tp->mss_cache is current effective sending mss, including
1535   all tcp options except for SACKs. It is evaluated,
1536   taking into account current pmtu, but never exceeds
1537   tp->rx_opt.mss_clamp.
1538
1539   NOTE1. rfc1122 clearly states that advertised MSS
1540   DOES NOT include either tcp or ip options.
1541
1542   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1543   are READ ONLY outside this function.		--ANK (980731)
1544 */
1545unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1546{
1547	struct tcp_sock *tp = tcp_sk(sk);
1548	struct inet_connection_sock *icsk = inet_csk(sk);
1549	int mss_now;
1550
1551	if (icsk->icsk_mtup.search_high > pmtu)
1552		icsk->icsk_mtup.search_high = pmtu;
1553
1554	mss_now = tcp_mtu_to_mss(sk, pmtu);
1555	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1556
1557	/* And store cached results */
1558	icsk->icsk_pmtu_cookie = pmtu;
1559	if (icsk->icsk_mtup.enabled)
1560		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1561	tp->mss_cache = mss_now;
1562
1563	return mss_now;
1564}
1565EXPORT_SYMBOL(tcp_sync_mss);
1566
1567/* Compute the current effective MSS, taking SACKs and IP options,
1568 * and even PMTU discovery events into account.
1569 */
1570unsigned int tcp_current_mss(struct sock *sk)
1571{
1572	const struct tcp_sock *tp = tcp_sk(sk);
1573	const struct dst_entry *dst = __sk_dst_get(sk);
1574	u32 mss_now;
1575	unsigned int header_len;
1576	struct tcp_out_options opts;
1577	struct tcp_md5sig_key *md5;
1578
1579	mss_now = tp->mss_cache;
1580
1581	if (dst) {
1582		u32 mtu = dst_mtu(dst);
1583		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1584			mss_now = tcp_sync_mss(sk, mtu);
1585	}
1586
1587	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1588		     sizeof(struct tcphdr);
1589	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1590	 * some common options. If this is an odd packet (because we have SACK
1591	 * blocks etc) then our calculated header_len will be different, and
1592	 * we have to adjust mss_now correspondingly */
1593	if (header_len != tp->tcp_header_len) {
1594		int delta = (int) header_len - tp->tcp_header_len;
1595		mss_now -= delta;
1596	}
1597
1598	return mss_now;
1599}
1600
1601/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1602 * As additional protections, we do not touch cwnd in retransmission phases,
1603 * and if application hit its sndbuf limit recently.
1604 */
1605static void tcp_cwnd_application_limited(struct sock *sk)
1606{
1607	struct tcp_sock *tp = tcp_sk(sk);
1608
1609	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1610	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1611		/* Limited by application or receiver window. */
1612		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1613		u32 win_used = max(tp->snd_cwnd_used, init_win);
1614		if (win_used < tp->snd_cwnd) {
1615			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1616			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1617		}
1618		tp->snd_cwnd_used = 0;
1619	}
1620	tp->snd_cwnd_stamp = tcp_jiffies32;
1621}
1622
1623static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1624{
1625	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1626	struct tcp_sock *tp = tcp_sk(sk);
1627
1628	/* Track the maximum number of outstanding packets in each
1629	 * window, and remember whether we were cwnd-limited then.
1630	 */
1631	if (!before(tp->snd_una, tp->max_packets_seq) ||
1632	    tp->packets_out > tp->max_packets_out) {
1633		tp->max_packets_out = tp->packets_out;
1634		tp->max_packets_seq = tp->snd_nxt;
1635		tp->is_cwnd_limited = is_cwnd_limited;
1636	}
1637
1638	if (tcp_is_cwnd_limited(sk)) {
1639		/* Network is feed fully. */
1640		tp->snd_cwnd_used = 0;
1641		tp->snd_cwnd_stamp = tcp_jiffies32;
1642	} else {
1643		/* Network starves. */
1644		if (tp->packets_out > tp->snd_cwnd_used)
1645			tp->snd_cwnd_used = tp->packets_out;
1646
1647		if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1648		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1649		    !ca_ops->cong_control)
1650			tcp_cwnd_application_limited(sk);
1651
1652		/* The following conditions together indicate the starvation
1653		 * is caused by insufficient sender buffer:
1654		 * 1) just sent some data (see tcp_write_xmit)
1655		 * 2) not cwnd limited (this else condition)
1656		 * 3) no more data to send (tcp_write_queue_empty())
1657		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1658		 */
1659		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1660		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1661		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1662			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1663	}
1664}
1665
1666/* Minshall's variant of the Nagle send check. */
1667static bool tcp_minshall_check(const struct tcp_sock *tp)
1668{
1669	return after(tp->snd_sml, tp->snd_una) &&
1670		!after(tp->snd_sml, tp->snd_nxt);
1671}
1672
1673/* Update snd_sml if this skb is under mss
1674 * Note that a TSO packet might end with a sub-mss segment
1675 * The test is really :
1676 * if ((skb->len % mss) != 0)
1677 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1678 * But we can avoid doing the divide again given we already have
1679 *  skb_pcount = skb->len / mss_now
1680 */
1681static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1682				const struct sk_buff *skb)
1683{
1684	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1685		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1686}
1687
1688/* Return false, if packet can be sent now without violation Nagle's rules:
1689 * 1. It is full sized. (provided by caller in %partial bool)
1690 * 2. Or it contains FIN. (already checked by caller)
1691 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1692 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1693 *    With Minshall's modification: all sent small packets are ACKed.
1694 */
1695static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1696			    int nonagle)
1697{
1698	return partial &&
1699		((nonagle & TCP_NAGLE_CORK) ||
1700		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1701}
1702
1703/* Return how many segs we'd like on a TSO packet,
1704 * to send one TSO packet per ms
1705 */
1706static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1707			    int min_tso_segs)
1708{
1709	u32 bytes, segs;
1710
1711	bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1712		    sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
 
1713
1714	/* Goal is to send at least one packet per ms,
1715	 * not one big TSO packet every 100 ms.
1716	 * This preserves ACK clocking and is consistent
1717	 * with tcp_tso_should_defer() heuristic.
1718	 */
1719	segs = max_t(u32, bytes / mss_now, min_tso_segs);
1720
1721	return segs;
1722}
1723
1724/* Return the number of segments we want in the skb we are transmitting.
1725 * See if congestion control module wants to decide; otherwise, autosize.
1726 */
1727static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1728{
1729	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1730	u32 min_tso, tso_segs;
1731
1732	min_tso = ca_ops->min_tso_segs ?
1733			ca_ops->min_tso_segs(sk) :
1734			sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1735
1736	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1737	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1738}
1739
1740/* Returns the portion of skb which can be sent right away */
1741static unsigned int tcp_mss_split_point(const struct sock *sk,
1742					const struct sk_buff *skb,
1743					unsigned int mss_now,
1744					unsigned int max_segs,
1745					int nonagle)
1746{
1747	const struct tcp_sock *tp = tcp_sk(sk);
1748	u32 partial, needed, window, max_len;
1749
1750	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1751	max_len = mss_now * max_segs;
1752
1753	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1754		return max_len;
1755
1756	needed = min(skb->len, window);
1757
1758	if (max_len <= needed)
1759		return max_len;
1760
1761	partial = needed % mss_now;
1762	/* If last segment is not a full MSS, check if Nagle rules allow us
1763	 * to include this last segment in this skb.
1764	 * Otherwise, we'll split the skb at last MSS boundary
1765	 */
1766	if (tcp_nagle_check(partial != 0, tp, nonagle))
1767		return needed - partial;
1768
1769	return needed;
1770}
1771
1772/* Can at least one segment of SKB be sent right now, according to the
1773 * congestion window rules?  If so, return how many segments are allowed.
1774 */
1775static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1776					 const struct sk_buff *skb)
1777{
1778	u32 in_flight, cwnd, halfcwnd;
1779
1780	/* Don't be strict about the congestion window for the final FIN.  */
1781	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1782	    tcp_skb_pcount(skb) == 1)
1783		return 1;
1784
1785	in_flight = tcp_packets_in_flight(tp);
1786	cwnd = tp->snd_cwnd;
1787	if (in_flight >= cwnd)
1788		return 0;
1789
1790	/* For better scheduling, ensure we have at least
1791	 * 2 GSO packets in flight.
1792	 */
1793	halfcwnd = max(cwnd >> 1, 1U);
1794	return min(halfcwnd, cwnd - in_flight);
1795}
1796
1797/* Initialize TSO state of a skb.
1798 * This must be invoked the first time we consider transmitting
1799 * SKB onto the wire.
1800 */
1801static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1802{
1803	int tso_segs = tcp_skb_pcount(skb);
1804
1805	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1806		tcp_set_skb_tso_segs(skb, mss_now);
1807		tso_segs = tcp_skb_pcount(skb);
1808	}
1809	return tso_segs;
1810}
1811
1812
1813/* Return true if the Nagle test allows this packet to be
1814 * sent now.
1815 */
1816static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1817				  unsigned int cur_mss, int nonagle)
1818{
1819	/* Nagle rule does not apply to frames, which sit in the middle of the
1820	 * write_queue (they have no chances to get new data).
1821	 *
1822	 * This is implemented in the callers, where they modify the 'nonagle'
1823	 * argument based upon the location of SKB in the send queue.
1824	 */
1825	if (nonagle & TCP_NAGLE_PUSH)
1826		return true;
1827
1828	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1829	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1830		return true;
1831
1832	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1833		return true;
1834
1835	return false;
1836}
1837
1838/* Does at least the first segment of SKB fit into the send window? */
1839static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1840			     const struct sk_buff *skb,
1841			     unsigned int cur_mss)
1842{
1843	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1844
1845	if (skb->len > cur_mss)
1846		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1847
1848	return !after(end_seq, tcp_wnd_end(tp));
1849}
1850
1851/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1852 * which is put after SKB on the list.  It is very much like
1853 * tcp_fragment() except that it may make several kinds of assumptions
1854 * in order to speed up the splitting operation.  In particular, we
1855 * know that all the data is in scatter-gather pages, and that the
1856 * packet has never been sent out before (and thus is not cloned).
1857 */
1858static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1859			struct sk_buff *skb, unsigned int len,
1860			unsigned int mss_now, gfp_t gfp)
1861{
1862	struct sk_buff *buff;
1863	int nlen = skb->len - len;
 
1864	u8 flags;
1865
1866	/* All of a TSO frame must be composed of paged data.  */
1867	if (skb->len != skb->data_len)
1868		return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
 
1869
1870	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1871	if (unlikely(!buff))
1872		return -ENOMEM;
 
1873
1874	sk->sk_wmem_queued += buff->truesize;
1875	sk_mem_charge(sk, buff->truesize);
1876	buff->truesize += nlen;
1877	skb->truesize -= nlen;
1878
1879	/* Correct the sequence numbers. */
1880	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1881	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1882	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1883
1884	/* PSH and FIN should only be set in the second packet. */
1885	flags = TCP_SKB_CB(skb)->tcp_flags;
1886	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1887	TCP_SKB_CB(buff)->tcp_flags = flags;
1888
1889	/* This packet was never sent out yet, so no SACK bits. */
1890	TCP_SKB_CB(buff)->sacked = 0;
1891
1892	tcp_skb_fragment_eor(skb, buff);
1893
1894	buff->ip_summed = CHECKSUM_PARTIAL;
1895	skb_split(skb, buff, len);
1896	tcp_fragment_tstamp(skb, buff);
1897
1898	/* Fix up tso_factor for both original and new SKB.  */
1899	tcp_set_skb_tso_segs(skb, mss_now);
1900	tcp_set_skb_tso_segs(buff, mss_now);
1901
1902	/* Link BUFF into the send queue. */
1903	__skb_header_release(buff);
1904	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1905
1906	return 0;
1907}
1908
1909/* Try to defer sending, if possible, in order to minimize the amount
1910 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1911 *
1912 * This algorithm is from John Heffner.
1913 */
1914static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1915				 bool *is_cwnd_limited, u32 max_segs)
 
 
1916{
1917	const struct inet_connection_sock *icsk = inet_csk(sk);
1918	u32 age, send_win, cong_win, limit, in_flight;
1919	struct tcp_sock *tp = tcp_sk(sk);
1920	struct sk_buff *head;
1921	int win_divisor;
1922
1923	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1924		goto send_now;
1925
1926	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1927		goto send_now;
1928
1929	/* Avoid bursty behavior by allowing defer
1930	 * only if the last write was recent.
 
 
1931	 */
1932	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
 
1933		goto send_now;
1934
1935	in_flight = tcp_packets_in_flight(tp);
1936
1937	BUG_ON(tcp_skb_pcount(skb) <= 1);
1938	BUG_ON(tp->snd_cwnd <= in_flight);
1939
1940	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1941
1942	/* From in_flight test above, we know that cwnd > in_flight.  */
1943	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1944
1945	limit = min(send_win, cong_win);
1946
1947	/* If a full-sized TSO skb can be sent, do it. */
1948	if (limit >= max_segs * tp->mss_cache)
1949		goto send_now;
1950
1951	/* Middle in queue won't get any more data, full sendable already? */
1952	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1953		goto send_now;
1954
1955	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1956	if (win_divisor) {
1957		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1958
1959		/* If at least some fraction of a window is available,
1960		 * just use it.
1961		 */
1962		chunk /= win_divisor;
1963		if (limit >= chunk)
1964			goto send_now;
1965	} else {
1966		/* Different approach, try not to defer past a single
1967		 * ACK.  Receiver should ACK every other full sized
1968		 * frame, so if we have space for more than 3 frames
1969		 * then send now.
1970		 */
1971		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1972			goto send_now;
1973	}
1974
1975	/* TODO : use tsorted_sent_queue ? */
1976	head = tcp_rtx_queue_head(sk);
1977	if (!head)
1978		goto send_now;
1979	age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1980	/* If next ACK is likely to come too late (half srtt), do not defer */
1981	if (age < (tp->srtt_us >> 4))
1982		goto send_now;
1983
1984	/* Ok, it looks like it is advisable to defer. */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1985
1986	if (cong_win < send_win && cong_win <= skb->len)
1987		*is_cwnd_limited = true;
 
 
1988
1989	return true;
1990
1991send_now:
1992	return false;
1993}
1994
1995static inline void tcp_mtu_check_reprobe(struct sock *sk)
1996{
1997	struct inet_connection_sock *icsk = inet_csk(sk);
1998	struct tcp_sock *tp = tcp_sk(sk);
1999	struct net *net = sock_net(sk);
2000	u32 interval;
2001	s32 delta;
2002
2003	interval = net->ipv4.sysctl_tcp_probe_interval;
2004	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2005	if (unlikely(delta >= interval * HZ)) {
2006		int mss = tcp_current_mss(sk);
2007
2008		/* Update current search range */
2009		icsk->icsk_mtup.probe_size = 0;
2010		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2011			sizeof(struct tcphdr) +
2012			icsk->icsk_af_ops->net_header_len;
2013		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2014
2015		/* Update probe time stamp */
2016		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2017	}
2018}
2019
2020static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2021{
2022	struct sk_buff *skb, *next;
2023
2024	skb = tcp_send_head(sk);
2025	tcp_for_write_queue_from_safe(skb, next, sk) {
2026		if (len <= skb->len)
2027			break;
2028
2029		if (unlikely(TCP_SKB_CB(skb)->eor))
2030			return false;
2031
2032		len -= skb->len;
2033	}
2034
2035	return true;
2036}
2037
2038/* Create a new MTU probe if we are ready.
2039 * MTU probe is regularly attempting to increase the path MTU by
2040 * deliberately sending larger packets.  This discovers routing
2041 * changes resulting in larger path MTUs.
2042 *
2043 * Returns 0 if we should wait to probe (no cwnd available),
2044 *         1 if a probe was sent,
2045 *         -1 otherwise
2046 */
2047static int tcp_mtu_probe(struct sock *sk)
2048{
2049	struct inet_connection_sock *icsk = inet_csk(sk);
2050	struct tcp_sock *tp = tcp_sk(sk);
2051	struct sk_buff *skb, *nskb, *next;
2052	struct net *net = sock_net(sk);
2053	int probe_size;
2054	int size_needed;
2055	int copy, len;
2056	int mss_now;
2057	int interval;
2058
2059	/* Not currently probing/verifying,
2060	 * not in recovery,
2061	 * have enough cwnd, and
2062	 * not SACKing (the variable headers throw things off)
2063	 */
2064	if (likely(!icsk->icsk_mtup.enabled ||
2065		   icsk->icsk_mtup.probe_size ||
2066		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2067		   tp->snd_cwnd < 11 ||
2068		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2069		return -1;
2070
2071	/* Use binary search for probe_size between tcp_mss_base,
2072	 * and current mss_clamp. if (search_high - search_low)
2073	 * smaller than a threshold, backoff from probing.
2074	 */
2075	mss_now = tcp_current_mss(sk);
2076	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2077				    icsk->icsk_mtup.search_low) >> 1);
2078	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2079	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2080	/* When misfortune happens, we are reprobing actively,
2081	 * and then reprobe timer has expired. We stick with current
2082	 * probing process by not resetting search range to its orignal.
2083	 */
2084	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2085		interval < net->ipv4.sysctl_tcp_probe_threshold) {
2086		/* Check whether enough time has elaplased for
2087		 * another round of probing.
2088		 */
2089		tcp_mtu_check_reprobe(sk);
2090		return -1;
2091	}
2092
2093	/* Have enough data in the send queue to probe? */
2094	if (tp->write_seq - tp->snd_nxt < size_needed)
2095		return -1;
2096
2097	if (tp->snd_wnd < size_needed)
2098		return -1;
2099	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2100		return 0;
2101
2102	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2103	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2104		if (!tcp_packets_in_flight(tp))
2105			return -1;
2106		else
2107			return 0;
2108	}
2109
2110	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2111		return -1;
2112
2113	/* We're allowed to probe.  Build it now. */
2114	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2115	if (!nskb)
2116		return -1;
2117	sk->sk_wmem_queued += nskb->truesize;
2118	sk_mem_charge(sk, nskb->truesize);
2119
2120	skb = tcp_send_head(sk);
 
2121
2122	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2123	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2124	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2125	TCP_SKB_CB(nskb)->sacked = 0;
2126	nskb->csum = 0;
2127	nskb->ip_summed = CHECKSUM_PARTIAL;
2128
2129	tcp_insert_write_queue_before(nskb, skb, sk);
2130	tcp_highest_sack_replace(sk, skb, nskb);
2131
2132	len = 0;
2133	tcp_for_write_queue_from_safe(skb, next, sk) {
2134		copy = min_t(int, skb->len, probe_size - len);
2135		skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2136
2137		if (skb->len <= copy) {
2138			/* We've eaten all the data from this skb.
2139			 * Throw it away. */
2140			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2141			/* If this is the last SKB we copy and eor is set
2142			 * we need to propagate it to the new skb.
2143			 */
2144			TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
 
2145			tcp_unlink_write_queue(skb, sk);
2146			sk_wmem_free_skb(sk, skb);
2147		} else {
2148			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2149						   ~(TCPHDR_FIN|TCPHDR_PSH);
2150			if (!skb_shinfo(skb)->nr_frags) {
2151				skb_pull(skb, copy);
2152			} else {
2153				__pskb_trim_head(skb, copy);
2154				tcp_set_skb_tso_segs(skb, mss_now);
2155			}
2156			TCP_SKB_CB(skb)->seq += copy;
2157		}
2158
2159		len += copy;
2160
2161		if (len >= probe_size)
2162			break;
2163	}
2164	tcp_init_tso_segs(nskb, nskb->len);
2165
2166	/* We're ready to send.  If this fails, the probe will
2167	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2168	 */
2169	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2170		/* Decrement cwnd here because we are sending
2171		 * effectively two packets. */
2172		tp->snd_cwnd--;
2173		tcp_event_new_data_sent(sk, nskb);
2174
2175		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2176		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2177		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2178
2179		return 1;
2180	}
2181
2182	return -1;
2183}
2184
2185static bool tcp_pacing_check(const struct sock *sk)
2186{
2187	return tcp_needs_internal_pacing(sk) &&
2188	       hrtimer_active(&tcp_sk(sk)->pacing_timer);
 
 
 
 
 
 
 
 
 
 
 
 
 
2189}
2190
2191/* TCP Small Queues :
2192 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2193 * (These limits are doubled for retransmits)
2194 * This allows for :
2195 *  - better RTT estimation and ACK scheduling
2196 *  - faster recovery
2197 *  - high rates
2198 * Alas, some drivers / subsystems require a fair amount
2199 * of queued bytes to ensure line rate.
2200 * One example is wifi aggregation (802.11 AMPDU)
2201 */
2202static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2203				  unsigned int factor)
2204{
2205	unsigned int limit;
2206
2207	limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2208	limit = min_t(u32, limit,
2209		      sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
 
 
 
2210	limit <<= factor;
2211
 
 
 
 
 
 
 
 
 
 
 
 
2212	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2213		/* Always send skb if rtx queue is empty.
2214		 * No need to wait for TX completion to call us back,
2215		 * after softirq/tasklet schedule.
2216		 * This helps when TX completions are delayed too much.
2217		 */
2218		if (tcp_rtx_queue_empty(sk))
2219			return false;
2220
2221		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2222		/* It is possible TX completion already happened
2223		 * before we set TSQ_THROTTLED, so we must
2224		 * test again the condition.
2225		 */
2226		smp_mb__after_atomic();
2227		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2228			return true;
2229	}
2230	return false;
2231}
2232
2233static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2234{
2235	const u32 now = tcp_jiffies32;
2236	enum tcp_chrono old = tp->chrono_type;
2237
2238	if (old > TCP_CHRONO_UNSPEC)
2239		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2240	tp->chrono_start = now;
2241	tp->chrono_type = new;
2242}
2243
2244void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2245{
2246	struct tcp_sock *tp = tcp_sk(sk);
2247
2248	/* If there are multiple conditions worthy of tracking in a
2249	 * chronograph then the highest priority enum takes precedence
2250	 * over the other conditions. So that if something "more interesting"
2251	 * starts happening, stop the previous chrono and start a new one.
2252	 */
2253	if (type > tp->chrono_type)
2254		tcp_chrono_set(tp, type);
2255}
2256
2257void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2258{
2259	struct tcp_sock *tp = tcp_sk(sk);
2260
2261
2262	/* There are multiple conditions worthy of tracking in a
2263	 * chronograph, so that the highest priority enum takes
2264	 * precedence over the other conditions (see tcp_chrono_start).
2265	 * If a condition stops, we only stop chrono tracking if
2266	 * it's the "most interesting" or current chrono we are
2267	 * tracking and starts busy chrono if we have pending data.
2268	 */
2269	if (tcp_rtx_and_write_queues_empty(sk))
2270		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2271	else if (type == tp->chrono_type)
2272		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2273}
2274
2275/* This routine writes packets to the network.  It advances the
2276 * send_head.  This happens as incoming acks open up the remote
2277 * window for us.
2278 *
2279 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2280 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2281 * account rare use of URG, this is not a big flaw.
2282 *
2283 * Send at most one packet when push_one > 0. Temporarily ignore
2284 * cwnd limit to force at most one packet out when push_one == 2.
2285
2286 * Returns true, if no segments are in flight and we have queued segments,
2287 * but cannot send anything now because of SWS or another problem.
2288 */
2289static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2290			   int push_one, gfp_t gfp)
2291{
2292	struct tcp_sock *tp = tcp_sk(sk);
2293	struct sk_buff *skb;
2294	unsigned int tso_segs, sent_pkts;
2295	int cwnd_quota;
2296	int result;
2297	bool is_cwnd_limited = false, is_rwnd_limited = false;
2298	u32 max_segs;
2299
2300	sent_pkts = 0;
2301
2302	tcp_mstamp_refresh(tp);
2303	if (!push_one) {
2304		/* Do MTU probing. */
2305		result = tcp_mtu_probe(sk);
2306		if (!result) {
2307			return false;
2308		} else if (result > 0) {
2309			sent_pkts = 1;
2310		}
2311	}
2312
2313	max_segs = tcp_tso_segs(sk, mss_now);
2314	while ((skb = tcp_send_head(sk))) {
2315		unsigned int limit;
2316
 
 
 
 
 
 
 
 
2317		if (tcp_pacing_check(sk))
2318			break;
2319
2320		tso_segs = tcp_init_tso_segs(skb, mss_now);
2321		BUG_ON(!tso_segs);
2322
2323		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2324			/* "skb_mstamp" is used as a start point for the retransmit timer */
2325			tcp_update_skb_after_send(tp, skb);
2326			goto repair; /* Skip network transmission */
2327		}
2328
2329		cwnd_quota = tcp_cwnd_test(tp, skb);
2330		if (!cwnd_quota) {
2331			if (push_one == 2)
2332				/* Force out a loss probe pkt. */
2333				cwnd_quota = 1;
2334			else
2335				break;
2336		}
2337
2338		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2339			is_rwnd_limited = true;
2340			break;
2341		}
2342
2343		if (tso_segs == 1) {
2344			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2345						     (tcp_skb_is_last(sk, skb) ?
2346						      nonagle : TCP_NAGLE_PUSH))))
2347				break;
2348		} else {
2349			if (!push_one &&
2350			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2351						 max_segs))
2352				break;
2353		}
2354
2355		limit = mss_now;
2356		if (tso_segs > 1 && !tcp_urg_mode(tp))
2357			limit = tcp_mss_split_point(sk, skb, mss_now,
2358						    min_t(unsigned int,
2359							  cwnd_quota,
2360							  max_segs),
2361						    nonagle);
2362
2363		if (skb->len > limit &&
2364		    unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2365					  skb, limit, mss_now, gfp)))
2366			break;
2367
2368		if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2369			clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2370		if (tcp_small_queue_check(sk, skb, 0))
2371			break;
2372
 
 
 
 
 
 
 
 
2373		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2374			break;
2375
2376repair:
2377		/* Advance the send_head.  This one is sent out.
2378		 * This call will increment packets_out.
2379		 */
2380		tcp_event_new_data_sent(sk, skb);
2381
2382		tcp_minshall_update(tp, mss_now, skb);
2383		sent_pkts += tcp_skb_pcount(skb);
2384
2385		if (push_one)
2386			break;
2387	}
2388
2389	if (is_rwnd_limited)
2390		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2391	else
2392		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2393
2394	if (likely(sent_pkts)) {
2395		if (tcp_in_cwnd_reduction(sk))
2396			tp->prr_out += sent_pkts;
2397
2398		/* Send one loss probe per tail loss episode. */
2399		if (push_one != 2)
2400			tcp_schedule_loss_probe(sk, false);
2401		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2402		tcp_cwnd_validate(sk, is_cwnd_limited);
2403		return false;
2404	}
2405	return !tp->packets_out && !tcp_write_queue_empty(sk);
2406}
2407
2408bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2409{
2410	struct inet_connection_sock *icsk = inet_csk(sk);
2411	struct tcp_sock *tp = tcp_sk(sk);
2412	u32 timeout, rto_delta_us;
2413	int early_retrans;
2414
2415	/* Don't do any loss probe on a Fast Open connection before 3WHS
2416	 * finishes.
2417	 */
2418	if (tp->fastopen_rsk)
2419		return false;
2420
2421	early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2422	/* Schedule a loss probe in 2*RTT for SACK capable connections
2423	 * not in loss recovery, that are either limited by cwnd or application.
2424	 */
2425	if ((early_retrans != 3 && early_retrans != 4) ||
2426	    !tp->packets_out || !tcp_is_sack(tp) ||
2427	    (icsk->icsk_ca_state != TCP_CA_Open &&
2428	     icsk->icsk_ca_state != TCP_CA_CWR))
2429		return false;
2430
2431	/* Probe timeout is 2*rtt. Add minimum RTO to account
2432	 * for delayed ack when there's one outstanding packet. If no RTT
2433	 * sample is available then probe after TCP_TIMEOUT_INIT.
2434	 */
2435	if (tp->srtt_us) {
2436		timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2437		if (tp->packets_out == 1)
2438			timeout += TCP_RTO_MIN;
2439		else
2440			timeout += TCP_TIMEOUT_MIN;
2441	} else {
2442		timeout = TCP_TIMEOUT_INIT;
2443	}
2444
2445	/* If the RTO formula yields an earlier time, then use that time. */
2446	rto_delta_us = advancing_rto ?
2447			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2448			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2449	if (rto_delta_us > 0)
2450		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2451
2452	inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2453				  TCP_RTO_MAX);
2454	return true;
2455}
2456
2457/* Thanks to skb fast clones, we can detect if a prior transmit of
2458 * a packet is still in a qdisc or driver queue.
2459 * In this case, there is very little point doing a retransmit !
2460 */
2461static bool skb_still_in_host_queue(const struct sock *sk,
2462				    const struct sk_buff *skb)
2463{
2464	if (unlikely(skb_fclone_busy(sk, skb))) {
2465		NET_INC_STATS(sock_net(sk),
2466			      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2467		return true;
2468	}
2469	return false;
2470}
2471
2472/* When probe timeout (PTO) fires, try send a new segment if possible, else
2473 * retransmit the last segment.
2474 */
2475void tcp_send_loss_probe(struct sock *sk)
2476{
2477	struct tcp_sock *tp = tcp_sk(sk);
2478	struct sk_buff *skb;
2479	int pcount;
2480	int mss = tcp_current_mss(sk);
2481
 
 
 
 
 
2482	skb = tcp_send_head(sk);
2483	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2484		pcount = tp->packets_out;
2485		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2486		if (tp->packets_out > pcount)
2487			goto probe_sent;
2488		goto rearm_timer;
2489	}
2490	skb = skb_rb_last(&sk->tcp_rtx_queue);
2491
2492	/* At most one outstanding TLP retransmission. */
2493	if (tp->tlp_high_seq)
2494		goto rearm_timer;
2495
2496	/* Retransmit last segment. */
2497	if (WARN_ON(!skb))
2498		goto rearm_timer;
2499
2500	if (skb_still_in_host_queue(sk, skb))
2501		goto rearm_timer;
2502
2503	pcount = tcp_skb_pcount(skb);
2504	if (WARN_ON(!pcount))
2505		goto rearm_timer;
2506
2507	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2508		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2509					  (pcount - 1) * mss, mss,
2510					  GFP_ATOMIC)))
2511			goto rearm_timer;
2512		skb = skb_rb_next(skb);
2513	}
2514
2515	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2516		goto rearm_timer;
2517
2518	if (__tcp_retransmit_skb(sk, skb, 1))
2519		goto rearm_timer;
2520
 
 
 
2521	/* Record snd_nxt for loss detection. */
2522	tp->tlp_high_seq = tp->snd_nxt;
2523
2524probe_sent:
2525	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2526	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2527	inet_csk(sk)->icsk_pending = 0;
2528rearm_timer:
2529	tcp_rearm_rto(sk);
2530}
2531
2532/* Push out any pending frames which were held back due to
2533 * TCP_CORK or attempt at coalescing tiny packets.
2534 * The socket must be locked by the caller.
2535 */
2536void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2537			       int nonagle)
2538{
2539	/* If we are closed, the bytes will have to remain here.
2540	 * In time closedown will finish, we empty the write queue and
2541	 * all will be happy.
2542	 */
2543	if (unlikely(sk->sk_state == TCP_CLOSE))
2544		return;
2545
2546	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2547			   sk_gfp_mask(sk, GFP_ATOMIC)))
2548		tcp_check_probe_timer(sk);
2549}
2550
2551/* Send _single_ skb sitting at the send head. This function requires
2552 * true push pending frames to setup probe timer etc.
2553 */
2554void tcp_push_one(struct sock *sk, unsigned int mss_now)
2555{
2556	struct sk_buff *skb = tcp_send_head(sk);
2557
2558	BUG_ON(!skb || skb->len < mss_now);
2559
2560	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2561}
2562
2563/* This function returns the amount that we can raise the
2564 * usable window based on the following constraints
2565 *
2566 * 1. The window can never be shrunk once it is offered (RFC 793)
2567 * 2. We limit memory per socket
2568 *
2569 * RFC 1122:
2570 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2571 *  RECV.NEXT + RCV.WIN fixed until:
2572 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2573 *
2574 * i.e. don't raise the right edge of the window until you can raise
2575 * it at least MSS bytes.
2576 *
2577 * Unfortunately, the recommended algorithm breaks header prediction,
2578 * since header prediction assumes th->window stays fixed.
2579 *
2580 * Strictly speaking, keeping th->window fixed violates the receiver
2581 * side SWS prevention criteria. The problem is that under this rule
2582 * a stream of single byte packets will cause the right side of the
2583 * window to always advance by a single byte.
2584 *
2585 * Of course, if the sender implements sender side SWS prevention
2586 * then this will not be a problem.
2587 *
2588 * BSD seems to make the following compromise:
2589 *
2590 *	If the free space is less than the 1/4 of the maximum
2591 *	space available and the free space is less than 1/2 mss,
2592 *	then set the window to 0.
2593 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2594 *	Otherwise, just prevent the window from shrinking
2595 *	and from being larger than the largest representable value.
2596 *
2597 * This prevents incremental opening of the window in the regime
2598 * where TCP is limited by the speed of the reader side taking
2599 * data out of the TCP receive queue. It does nothing about
2600 * those cases where the window is constrained on the sender side
2601 * because the pipeline is full.
2602 *
2603 * BSD also seems to "accidentally" limit itself to windows that are a
2604 * multiple of MSS, at least until the free space gets quite small.
2605 * This would appear to be a side effect of the mbuf implementation.
2606 * Combining these two algorithms results in the observed behavior
2607 * of having a fixed window size at almost all times.
2608 *
2609 * Below we obtain similar behavior by forcing the offered window to
2610 * a multiple of the mss when it is feasible to do so.
2611 *
2612 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2613 * Regular options like TIMESTAMP are taken into account.
2614 */
2615u32 __tcp_select_window(struct sock *sk)
2616{
2617	struct inet_connection_sock *icsk = inet_csk(sk);
2618	struct tcp_sock *tp = tcp_sk(sk);
2619	/* MSS for the peer's data.  Previous versions used mss_clamp
2620	 * here.  I don't know if the value based on our guesses
2621	 * of peer's MSS is better for the performance.  It's more correct
2622	 * but may be worse for the performance because of rcv_mss
2623	 * fluctuations.  --SAW  1998/11/1
2624	 */
2625	int mss = icsk->icsk_ack.rcv_mss;
2626	int free_space = tcp_space(sk);
2627	int allowed_space = tcp_full_space(sk);
2628	int full_space = min_t(int, tp->window_clamp, allowed_space);
2629	int window;
 
 
 
 
2630
2631	if (unlikely(mss > full_space)) {
2632		mss = full_space;
2633		if (mss <= 0)
2634			return 0;
2635	}
2636	if (free_space < (full_space >> 1)) {
2637		icsk->icsk_ack.quick = 0;
2638
2639		if (tcp_under_memory_pressure(sk))
2640			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2641					       4U * tp->advmss);
2642
2643		/* free_space might become our new window, make sure we don't
2644		 * increase it due to wscale.
2645		 */
2646		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2647
2648		/* if free space is less than mss estimate, or is below 1/16th
2649		 * of the maximum allowed, try to move to zero-window, else
2650		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2651		 * new incoming data is dropped due to memory limits.
2652		 * With large window, mss test triggers way too late in order
2653		 * to announce zero window in time before rmem limit kicks in.
2654		 */
2655		if (free_space < (allowed_space >> 4) || free_space < mss)
2656			return 0;
2657	}
2658
2659	if (free_space > tp->rcv_ssthresh)
2660		free_space = tp->rcv_ssthresh;
2661
2662	/* Don't do rounding if we are using window scaling, since the
2663	 * scaled window will not line up with the MSS boundary anyway.
2664	 */
2665	if (tp->rx_opt.rcv_wscale) {
2666		window = free_space;
2667
2668		/* Advertise enough space so that it won't get scaled away.
2669		 * Import case: prevent zero window announcement if
2670		 * 1<<rcv_wscale > mss.
2671		 */
2672		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2673	} else {
2674		window = tp->rcv_wnd;
2675		/* Get the largest window that is a nice multiple of mss.
2676		 * Window clamp already applied above.
2677		 * If our current window offering is within 1 mss of the
2678		 * free space we just keep it. This prevents the divide
2679		 * and multiply from happening most of the time.
2680		 * We also don't do any window rounding when the free space
2681		 * is too small.
2682		 */
2683		if (window <= free_space - mss || window > free_space)
2684			window = rounddown(free_space, mss);
2685		else if (mss == full_space &&
2686			 free_space > window + (full_space >> 1))
2687			window = free_space;
2688	}
2689
2690	return window;
2691}
2692
2693void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2694			     const struct sk_buff *next_skb)
2695{
2696	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2697		const struct skb_shared_info *next_shinfo =
2698			skb_shinfo(next_skb);
2699		struct skb_shared_info *shinfo = skb_shinfo(skb);
2700
2701		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2702		shinfo->tskey = next_shinfo->tskey;
2703		TCP_SKB_CB(skb)->txstamp_ack |=
2704			TCP_SKB_CB(next_skb)->txstamp_ack;
2705	}
2706}
2707
2708/* Collapses two adjacent SKB's during retransmission. */
2709static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2710{
2711	struct tcp_sock *tp = tcp_sk(sk);
2712	struct sk_buff *next_skb = skb_rb_next(skb);
2713	int skb_size, next_skb_size;
2714
2715	skb_size = skb->len;
2716	next_skb_size = next_skb->len;
2717
2718	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2719
2720	if (next_skb_size) {
2721		if (next_skb_size <= skb_availroom(skb))
2722			skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2723				      next_skb_size);
2724		else if (!skb_shift(skb, next_skb, next_skb_size))
2725			return false;
2726	}
2727	tcp_highest_sack_replace(sk, next_skb, skb);
2728
2729	/* Update sequence range on original skb. */
2730	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2731
2732	/* Merge over control information. This moves PSH/FIN etc. over */
2733	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2734
2735	/* All done, get rid of second SKB and account for it so
2736	 * packet counting does not break.
2737	 */
2738	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2739	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2740
2741	/* changed transmit queue under us so clear hints */
2742	tcp_clear_retrans_hints_partial(tp);
2743	if (next_skb == tp->retransmit_skb_hint)
2744		tp->retransmit_skb_hint = skb;
2745
2746	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2747
2748	tcp_skb_collapse_tstamp(skb, next_skb);
2749
2750	tcp_rtx_queue_unlink_and_free(next_skb, sk);
2751	return true;
2752}
2753
2754/* Check if coalescing SKBs is legal. */
2755static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2756{
2757	if (tcp_skb_pcount(skb) > 1)
2758		return false;
2759	if (skb_cloned(skb))
2760		return false;
2761	/* Some heuristics for collapsing over SACK'd could be invented */
2762	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2763		return false;
2764
2765	return true;
2766}
2767
2768/* Collapse packets in the retransmit queue to make to create
2769 * less packets on the wire. This is only done on retransmission.
2770 */
2771static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2772				     int space)
2773{
2774	struct tcp_sock *tp = tcp_sk(sk);
2775	struct sk_buff *skb = to, *tmp;
2776	bool first = true;
2777
2778	if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2779		return;
2780	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2781		return;
2782
2783	skb_rbtree_walk_from_safe(skb, tmp) {
2784		if (!tcp_can_collapse(sk, skb))
2785			break;
2786
2787		if (!tcp_skb_can_collapse_to(to))
2788			break;
2789
2790		space -= skb->len;
2791
2792		if (first) {
2793			first = false;
2794			continue;
2795		}
2796
2797		if (space < 0)
2798			break;
2799
2800		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2801			break;
2802
2803		if (!tcp_collapse_retrans(sk, to))
2804			break;
2805	}
2806}
2807
2808/* This retransmits one SKB.  Policy decisions and retransmit queue
2809 * state updates are done by the caller.  Returns non-zero if an
2810 * error occurred which prevented the send.
2811 */
2812int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2813{
2814	struct inet_connection_sock *icsk = inet_csk(sk);
2815	struct tcp_sock *tp = tcp_sk(sk);
2816	unsigned int cur_mss;
2817	int diff, len, err;
2818
2819
2820	/* Inconclusive MTU probe */
2821	if (icsk->icsk_mtup.probe_size)
2822		icsk->icsk_mtup.probe_size = 0;
2823
2824	/* Do not sent more than we queued. 1/4 is reserved for possible
2825	 * copying overhead: fragmentation, tunneling, mangling etc.
2826	 */
2827	if (refcount_read(&sk->sk_wmem_alloc) >
2828	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2829		  sk->sk_sndbuf))
2830		return -EAGAIN;
2831
2832	if (skb_still_in_host_queue(sk, skb))
2833		return -EBUSY;
2834
2835	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2836		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2837			WARN_ON_ONCE(1);
2838			return -EINVAL;
2839		}
2840		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2841			return -ENOMEM;
2842	}
2843
2844	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2845		return -EHOSTUNREACH; /* Routing failure or similar. */
2846
2847	cur_mss = tcp_current_mss(sk);
2848
2849	/* If receiver has shrunk his window, and skb is out of
2850	 * new window, do not retransmit it. The exception is the
2851	 * case, when window is shrunk to zero. In this case
2852	 * our retransmit serves as a zero window probe.
2853	 */
2854	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2855	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2856		return -EAGAIN;
2857
2858	len = cur_mss * segs;
2859	if (skb->len > len) {
2860		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2861				 cur_mss, GFP_ATOMIC))
2862			return -ENOMEM; /* We'll try again later. */
2863	} else {
2864		if (skb_unclone(skb, GFP_ATOMIC))
2865			return -ENOMEM;
2866
2867		diff = tcp_skb_pcount(skb);
2868		tcp_set_skb_tso_segs(skb, cur_mss);
2869		diff -= tcp_skb_pcount(skb);
2870		if (diff)
2871			tcp_adjust_pcount(sk, skb, diff);
2872		if (skb->len < cur_mss)
2873			tcp_retrans_try_collapse(sk, skb, cur_mss);
2874	}
2875
2876	/* RFC3168, section 6.1.1.1. ECN fallback */
2877	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2878		tcp_ecn_clear_syn(sk, skb);
2879
2880	/* Update global and local TCP statistics. */
2881	segs = tcp_skb_pcount(skb);
2882	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2883	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2884		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2885	tp->total_retrans += segs;
 
2886
2887	/* make sure skb->data is aligned on arches that require it
2888	 * and check if ack-trimming & collapsing extended the headroom
2889	 * beyond what csum_start can cover.
2890	 */
2891	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2892		     skb_headroom(skb) >= 0xFFFF)) {
2893		struct sk_buff *nskb;
2894
2895		tcp_skb_tsorted_save(skb) {
2896			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2897			err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2898				     -ENOBUFS;
 
 
 
 
2899		} tcp_skb_tsorted_restore(skb);
2900
2901		if (!err) {
2902			tcp_update_skb_after_send(tp, skb);
2903			tcp_rate_skb_sent(sk, skb);
2904		}
2905	} else {
2906		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2907	}
2908
 
 
 
 
 
2909	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2910		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2911				  TCP_SKB_CB(skb)->seq, segs, err);
2912
2913	if (likely(!err)) {
2914		TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2915		trace_tcp_retransmit_skb(sk, skb);
2916	} else if (err != -EBUSY) {
2917		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2918	}
2919	return err;
2920}
2921
2922int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2923{
2924	struct tcp_sock *tp = tcp_sk(sk);
2925	int err = __tcp_retransmit_skb(sk, skb, segs);
2926
2927	if (err == 0) {
2928#if FASTRETRANS_DEBUG > 0
2929		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2930			net_dbg_ratelimited("retrans_out leaked\n");
2931		}
2932#endif
2933		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2934		tp->retrans_out += tcp_skb_pcount(skb);
2935
2936		/* Save stamp of the first retransmit. */
2937		if (!tp->retrans_stamp)
2938			tp->retrans_stamp = tcp_skb_timestamp(skb);
2939
2940	}
2941
 
 
 
 
2942	if (tp->undo_retrans < 0)
2943		tp->undo_retrans = 0;
2944	tp->undo_retrans += tcp_skb_pcount(skb);
2945	return err;
2946}
2947
2948/* This gets called after a retransmit timeout, and the initially
2949 * retransmitted data is acknowledged.  It tries to continue
2950 * resending the rest of the retransmit queue, until either
2951 * we've sent it all or the congestion window limit is reached.
2952 */
2953void tcp_xmit_retransmit_queue(struct sock *sk)
2954{
2955	const struct inet_connection_sock *icsk = inet_csk(sk);
2956	struct sk_buff *skb, *rtx_head, *hole = NULL;
2957	struct tcp_sock *tp = tcp_sk(sk);
 
2958	u32 max_segs;
2959	int mib_idx;
2960
2961	if (!tp->packets_out)
2962		return;
2963
2964	rtx_head = tcp_rtx_queue_head(sk);
2965	skb = tp->retransmit_skb_hint ?: rtx_head;
2966	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2967	skb_rbtree_walk_from(skb) {
2968		__u8 sacked;
2969		int segs;
2970
2971		if (tcp_pacing_check(sk))
2972			break;
2973
2974		/* we could do better than to assign each time */
2975		if (!hole)
2976			tp->retransmit_skb_hint = skb;
2977
2978		segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2979		if (segs <= 0)
2980			return;
2981		sacked = TCP_SKB_CB(skb)->sacked;
2982		/* In case tcp_shift_skb_data() have aggregated large skbs,
2983		 * we need to make sure not sending too bigs TSO packets
2984		 */
2985		segs = min_t(int, segs, max_segs);
2986
2987		if (tp->retrans_out >= tp->lost_out) {
2988			break;
2989		} else if (!(sacked & TCPCB_LOST)) {
2990			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2991				hole = skb;
2992			continue;
2993
2994		} else {
2995			if (icsk->icsk_ca_state != TCP_CA_Loss)
2996				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2997			else
2998				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2999		}
3000
3001		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3002			continue;
3003
3004		if (tcp_small_queue_check(sk, skb, 1))
3005			return;
3006
3007		if (tcp_retransmit_skb(sk, skb, segs))
3008			return;
3009
3010		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3011
3012		if (tcp_in_cwnd_reduction(sk))
3013			tp->prr_out += tcp_skb_pcount(skb);
3014
3015		if (skb == rtx_head &&
3016		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3017			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3018						  inet_csk(sk)->icsk_rto,
3019						  TCP_RTO_MAX);
3020	}
 
 
 
 
3021}
3022
3023/* We allow to exceed memory limits for FIN packets to expedite
3024 * connection tear down and (memory) recovery.
3025 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3026 * or even be forced to close flow without any FIN.
3027 * In general, we want to allow one skb per socket to avoid hangs
3028 * with edge trigger epoll()
3029 */
3030void sk_forced_mem_schedule(struct sock *sk, int size)
3031{
3032	int amt;
3033
3034	if (size <= sk->sk_forward_alloc)
3035		return;
3036	amt = sk_mem_pages(size);
3037	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3038	sk_memory_allocated_add(sk, amt);
3039
3040	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3041		mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3042}
3043
3044/* Send a FIN. The caller locks the socket for us.
3045 * We should try to send a FIN packet really hard, but eventually give up.
3046 */
3047void tcp_send_fin(struct sock *sk)
3048{
3049	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3050	struct tcp_sock *tp = tcp_sk(sk);
3051
3052	/* Optimization, tack on the FIN if we have one skb in write queue and
3053	 * this skb was not yet sent, or we are under memory pressure.
3054	 * Note: in the latter case, FIN packet will be sent after a timeout,
3055	 * as TCP stack thinks it has already been transmitted.
3056	 */
 
3057	if (!tskb && tcp_under_memory_pressure(sk))
3058		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3059
3060	if (tskb) {
3061coalesce:
3062		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3063		TCP_SKB_CB(tskb)->end_seq++;
3064		tp->write_seq++;
3065		if (tcp_write_queue_empty(sk)) {
3066			/* This means tskb was already sent.
3067			 * Pretend we included the FIN on previous transmit.
3068			 * We need to set tp->snd_nxt to the value it would have
3069			 * if FIN had been sent. This is because retransmit path
3070			 * does not change tp->snd_nxt.
3071			 */
3072			tp->snd_nxt++;
3073			return;
3074		}
3075	} else {
3076		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3077		if (unlikely(!skb)) {
3078			if (tskb)
3079				goto coalesce;
3080			return;
3081		}
3082		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3083		skb_reserve(skb, MAX_TCP_HEADER);
3084		sk_forced_mem_schedule(sk, skb->truesize);
3085		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3086		tcp_init_nondata_skb(skb, tp->write_seq,
3087				     TCPHDR_ACK | TCPHDR_FIN);
3088		tcp_queue_skb(sk, skb);
3089	}
3090	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3091}
3092
3093/* We get here when a process closes a file descriptor (either due to
3094 * an explicit close() or as a byproduct of exit()'ing) and there
3095 * was unread data in the receive queue.  This behavior is recommended
3096 * by RFC 2525, section 2.17.  -DaveM
3097 */
3098void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3099{
3100	struct sk_buff *skb;
3101
3102	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3103
3104	/* NOTE: No TCP options attached and we never retransmit this. */
3105	skb = alloc_skb(MAX_TCP_HEADER, priority);
3106	if (!skb) {
3107		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3108		return;
3109	}
3110
3111	/* Reserve space for headers and prepare control bits. */
3112	skb_reserve(skb, MAX_TCP_HEADER);
3113	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3114			     TCPHDR_ACK | TCPHDR_RST);
3115	tcp_mstamp_refresh(tcp_sk(sk));
3116	/* Send it off. */
3117	if (tcp_transmit_skb(sk, skb, 0, priority))
3118		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3119
3120	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3121	 * skb here is different to the troublesome skb, so use NULL
3122	 */
3123	trace_tcp_send_reset(sk, NULL);
3124}
3125
3126/* Send a crossed SYN-ACK during socket establishment.
3127 * WARNING: This routine must only be called when we have already sent
3128 * a SYN packet that crossed the incoming SYN that caused this routine
3129 * to get called. If this assumption fails then the initial rcv_wnd
3130 * and rcv_wscale values will not be correct.
3131 */
3132int tcp_send_synack(struct sock *sk)
3133{
3134	struct sk_buff *skb;
3135
3136	skb = tcp_rtx_queue_head(sk);
3137	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3138		pr_err("%s: wrong queue state\n", __func__);
3139		return -EFAULT;
3140	}
3141	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3142		if (skb_cloned(skb)) {
3143			struct sk_buff *nskb;
3144
3145			tcp_skb_tsorted_save(skb) {
3146				nskb = skb_copy(skb, GFP_ATOMIC);
3147			} tcp_skb_tsorted_restore(skb);
3148			if (!nskb)
3149				return -ENOMEM;
3150			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
 
3151			tcp_rtx_queue_unlink_and_free(skb, sk);
3152			__skb_header_release(nskb);
3153			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3154			sk->sk_wmem_queued += nskb->truesize;
3155			sk_mem_charge(sk, nskb->truesize);
3156			skb = nskb;
3157		}
3158
3159		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3160		tcp_ecn_send_synack(sk, skb);
3161	}
3162	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3163}
3164
3165/**
3166 * tcp_make_synack - Prepare a SYN-ACK.
3167 * sk: listener socket
3168 * dst: dst entry attached to the SYNACK
3169 * req: request_sock pointer
 
 
3170 *
3171 * Allocate one skb and build a SYNACK packet.
3172 * @dst is consumed : Caller should not use it again.
3173 */
3174struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3175				struct request_sock *req,
3176				struct tcp_fastopen_cookie *foc,
3177				enum tcp_synack_type synack_type)
3178{
3179	struct inet_request_sock *ireq = inet_rsk(req);
3180	const struct tcp_sock *tp = tcp_sk(sk);
3181	struct tcp_md5sig_key *md5 = NULL;
3182	struct tcp_out_options opts;
3183	struct sk_buff *skb;
3184	int tcp_header_size;
3185	struct tcphdr *th;
3186	int mss;
 
3187
3188	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3189	if (unlikely(!skb)) {
3190		dst_release(dst);
3191		return NULL;
3192	}
3193	/* Reserve space for headers. */
3194	skb_reserve(skb, MAX_TCP_HEADER);
3195
3196	switch (synack_type) {
3197	case TCP_SYNACK_NORMAL:
3198		skb_set_owner_w(skb, req_to_sk(req));
3199		break;
3200	case TCP_SYNACK_COOKIE:
3201		/* Under synflood, we do not attach skb to a socket,
3202		 * to avoid false sharing.
3203		 */
3204		break;
3205	case TCP_SYNACK_FASTOPEN:
3206		/* sk is a const pointer, because we want to express multiple
3207		 * cpu might call us concurrently.
3208		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3209		 */
3210		skb_set_owner_w(skb, (struct sock *)sk);
3211		break;
3212	}
3213	skb_dst_set(skb, dst);
3214
3215	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3216
3217	memset(&opts, 0, sizeof(opts));
 
3218#ifdef CONFIG_SYN_COOKIES
3219	if (unlikely(req->cookie_ts))
3220		skb->skb_mstamp = cookie_init_timestamp(req);
3221	else
3222#endif
3223		skb->skb_mstamp = tcp_clock_us();
 
 
 
 
3224
3225#ifdef CONFIG_TCP_MD5SIG
3226	rcu_read_lock();
3227	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3228#endif
3229	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3230	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3231					     foc) + sizeof(*th);
3232
3233	skb_push(skb, tcp_header_size);
3234	skb_reset_transport_header(skb);
3235
3236	th = (struct tcphdr *)skb->data;
3237	memset(th, 0, sizeof(struct tcphdr));
3238	th->syn = 1;
3239	th->ack = 1;
3240	tcp_ecn_make_synack(req, th);
3241	th->source = htons(ireq->ir_num);
3242	th->dest = ireq->ir_rmt_port;
3243	skb->mark = ireq->ir_mark;
3244	skb->ip_summed = CHECKSUM_PARTIAL;
3245	th->seq = htonl(tcp_rsk(req)->snt_isn);
3246	/* XXX data is queued and acked as is. No buffer/window check */
3247	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3248
3249	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3250	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3251	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3252	th->doff = (tcp_header_size >> 2);
3253	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3254
3255#ifdef CONFIG_TCP_MD5SIG
3256	/* Okay, we have all we need - do the md5 hash if needed */
3257	if (md5)
3258		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3259					       md5, req_to_sk(req), skb);
3260	rcu_read_unlock();
3261#endif
3262
3263	/* Do not fool tcpdump (if any), clean our debris */
3264	skb->tstamp = 0;
 
3265	return skb;
3266}
3267EXPORT_SYMBOL(tcp_make_synack);
3268
3269static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3270{
3271	struct inet_connection_sock *icsk = inet_csk(sk);
3272	const struct tcp_congestion_ops *ca;
3273	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3274
3275	if (ca_key == TCP_CA_UNSPEC)
3276		return;
3277
3278	rcu_read_lock();
3279	ca = tcp_ca_find_key(ca_key);
3280	if (likely(ca && try_module_get(ca->owner))) {
3281		module_put(icsk->icsk_ca_ops->owner);
3282		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3283		icsk->icsk_ca_ops = ca;
3284	}
3285	rcu_read_unlock();
3286}
3287
3288/* Do all connect socket setups that can be done AF independent. */
3289static void tcp_connect_init(struct sock *sk)
3290{
3291	const struct dst_entry *dst = __sk_dst_get(sk);
3292	struct tcp_sock *tp = tcp_sk(sk);
3293	__u8 rcv_wscale;
3294	u32 rcv_wnd;
3295
3296	/* We'll fix this up when we get a response from the other end.
3297	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3298	 */
3299	tp->tcp_header_len = sizeof(struct tcphdr);
3300	if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3301		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3302
3303#ifdef CONFIG_TCP_MD5SIG
3304	if (tp->af_specific->md5_lookup(sk, sk))
3305		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3306#endif
3307
3308	/* If user gave his TCP_MAXSEG, record it to clamp */
3309	if (tp->rx_opt.user_mss)
3310		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3311	tp->max_window = 0;
3312	tcp_mtup_init(sk);
3313	tcp_sync_mss(sk, dst_mtu(dst));
3314
3315	tcp_ca_dst_init(sk, dst);
3316
3317	if (!tp->window_clamp)
3318		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3319	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3320
3321	tcp_initialize_rcv_mss(sk);
3322
3323	/* limit the window selection if the user enforce a smaller rx buffer */
3324	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3325	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3326		tp->window_clamp = tcp_full_space(sk);
3327
3328	rcv_wnd = tcp_rwnd_init_bpf(sk);
3329	if (rcv_wnd == 0)
3330		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3331
3332	tcp_select_initial_window(sk, tcp_full_space(sk),
3333				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3334				  &tp->rcv_wnd,
3335				  &tp->window_clamp,
3336				  sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3337				  &rcv_wscale,
3338				  rcv_wnd);
3339
3340	tp->rx_opt.rcv_wscale = rcv_wscale;
3341	tp->rcv_ssthresh = tp->rcv_wnd;
3342
3343	sk->sk_err = 0;
3344	sock_reset_flag(sk, SOCK_DONE);
3345	tp->snd_wnd = 0;
3346	tcp_init_wl(tp, 0);
3347	tcp_write_queue_purge(sk);
3348	tp->snd_una = tp->write_seq;
3349	tp->snd_sml = tp->write_seq;
3350	tp->snd_up = tp->write_seq;
3351	tp->snd_nxt = tp->write_seq;
3352
3353	if (likely(!tp->repair))
3354		tp->rcv_nxt = 0;
3355	else
3356		tp->rcv_tstamp = tcp_jiffies32;
3357	tp->rcv_wup = tp->rcv_nxt;
3358	tp->copied_seq = tp->rcv_nxt;
3359
3360	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3361	inet_csk(sk)->icsk_retransmits = 0;
3362	tcp_clear_retrans(tp);
3363}
3364
3365static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3366{
3367	struct tcp_sock *tp = tcp_sk(sk);
3368	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3369
3370	tcb->end_seq += skb->len;
3371	__skb_header_release(skb);
3372	sk->sk_wmem_queued += skb->truesize;
3373	sk_mem_charge(sk, skb->truesize);
3374	tp->write_seq = tcb->end_seq;
3375	tp->packets_out += tcp_skb_pcount(skb);
3376}
3377
3378/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3379 * queue a data-only packet after the regular SYN, such that regular SYNs
3380 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3381 * only the SYN sequence, the data are retransmitted in the first ACK.
3382 * If cookie is not cached or other error occurs, falls back to send a
3383 * regular SYN with Fast Open cookie request option.
3384 */
3385static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3386{
3387	struct tcp_sock *tp = tcp_sk(sk);
3388	struct tcp_fastopen_request *fo = tp->fastopen_req;
3389	int space, err = 0;
3390	struct sk_buff *syn_data;
3391
3392	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3393	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3394		goto fallback;
3395
3396	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3397	 * user-MSS. Reserve maximum option space for middleboxes that add
3398	 * private TCP options. The cost is reduced data space in SYN :(
3399	 */
3400	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3401
3402	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3403		MAX_TCP_OPTION_SPACE;
3404
3405	space = min_t(size_t, space, fo->size);
3406
3407	/* limit to order-0 allocations */
3408	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3409
3410	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3411	if (!syn_data)
3412		goto fallback;
3413	syn_data->ip_summed = CHECKSUM_PARTIAL;
3414	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3415	if (space) {
3416		int copied = copy_from_iter(skb_put(syn_data, space), space,
3417					    &fo->data->msg_iter);
3418		if (unlikely(!copied)) {
3419			tcp_skb_tsorted_anchor_cleanup(syn_data);
3420			kfree_skb(syn_data);
3421			goto fallback;
3422		}
3423		if (copied != space) {
3424			skb_trim(syn_data, copied);
3425			space = copied;
3426		}
 
3427	}
3428	/* No more data pending in inet_wait_for_connect() */
3429	if (space == fo->size)
3430		fo->data = NULL;
3431	fo->copied = space;
3432
3433	tcp_connect_queue_skb(sk, syn_data);
3434	if (syn_data->len)
3435		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3436
3437	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3438
3439	syn->skb_mstamp = syn_data->skb_mstamp;
3440
3441	/* Now full SYN+DATA was cloned and sent (or not),
3442	 * remove the SYN from the original skb (syn_data)
3443	 * we keep in write queue in case of a retransmit, as we
3444	 * also have the SYN packet (with no data) in the same queue.
3445	 */
3446	TCP_SKB_CB(syn_data)->seq++;
3447	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3448	if (!err) {
3449		tp->syn_data = (fo->copied > 0);
3450		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3451		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3452		goto done;
3453	}
3454
3455	/* data was not sent, put it in write_queue */
3456	__skb_queue_tail(&sk->sk_write_queue, syn_data);
3457	tp->packets_out -= tcp_skb_pcount(syn_data);
3458
3459fallback:
3460	/* Send a regular SYN with Fast Open cookie request option */
3461	if (fo->cookie.len > 0)
3462		fo->cookie.len = 0;
3463	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3464	if (err)
3465		tp->syn_fastopen = 0;
3466done:
3467	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3468	return err;
3469}
3470
3471/* Build a SYN and send it off. */
3472int tcp_connect(struct sock *sk)
3473{
3474	struct tcp_sock *tp = tcp_sk(sk);
3475	struct sk_buff *buff;
3476	int err;
3477
3478	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3479
3480	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3481		return -EHOSTUNREACH; /* Routing failure or similar. */
3482
3483	tcp_connect_init(sk);
3484
3485	if (unlikely(tp->repair)) {
3486		tcp_finish_connect(sk, NULL);
3487		return 0;
3488	}
3489
3490	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3491	if (unlikely(!buff))
3492		return -ENOBUFS;
3493
3494	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3495	tcp_mstamp_refresh(tp);
3496	tp->retrans_stamp = tcp_time_stamp(tp);
3497	tcp_connect_queue_skb(sk, buff);
3498	tcp_ecn_send_syn(sk, buff);
3499	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3500
3501	/* Send off SYN; include data in Fast Open. */
3502	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3503	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3504	if (err == -ECONNREFUSED)
3505		return err;
3506
3507	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3508	 * in order to make this packet get counted in tcpOutSegs.
3509	 */
3510	tp->snd_nxt = tp->write_seq;
3511	tp->pushed_seq = tp->write_seq;
3512	buff = tcp_send_head(sk);
3513	if (unlikely(buff)) {
3514		tp->snd_nxt	= TCP_SKB_CB(buff)->seq;
3515		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
3516	}
3517	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3518
3519	/* Timer for repeating the SYN until an answer. */
3520	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3521				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3522	return 0;
3523}
3524EXPORT_SYMBOL(tcp_connect);
3525
3526/* Send out a delayed ack, the caller does the policy checking
3527 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3528 * for details.
3529 */
3530void tcp_send_delayed_ack(struct sock *sk)
3531{
3532	struct inet_connection_sock *icsk = inet_csk(sk);
3533	int ato = icsk->icsk_ack.ato;
3534	unsigned long timeout;
3535
3536	tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3537
3538	if (ato > TCP_DELACK_MIN) {
3539		const struct tcp_sock *tp = tcp_sk(sk);
3540		int max_ato = HZ / 2;
3541
3542		if (icsk->icsk_ack.pingpong ||
3543		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3544			max_ato = TCP_DELACK_MAX;
3545
3546		/* Slow path, intersegment interval is "high". */
3547
3548		/* If some rtt estimate is known, use it to bound delayed ack.
3549		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3550		 * directly.
3551		 */
3552		if (tp->srtt_us) {
3553			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3554					TCP_DELACK_MIN);
3555
3556			if (rtt < max_ato)
3557				max_ato = rtt;
3558		}
3559
3560		ato = min(ato, max_ato);
3561	}
3562
3563	/* Stay within the limit we were given */
3564	timeout = jiffies + ato;
3565
3566	/* Use new timeout only if there wasn't a older one earlier. */
3567	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3568		/* If delack timer was blocked or is about to expire,
3569		 * send ACK now.
3570		 */
3571		if (icsk->icsk_ack.blocked ||
3572		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3573			tcp_send_ack(sk);
3574			return;
3575		}
3576
3577		if (!time_before(timeout, icsk->icsk_ack.timeout))
3578			timeout = icsk->icsk_ack.timeout;
3579	}
3580	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3581	icsk->icsk_ack.timeout = timeout;
3582	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3583}
3584
3585/* This routine sends an ack and also updates the window. */
3586void tcp_send_ack(struct sock *sk)
3587{
3588	struct sk_buff *buff;
3589
3590	/* If we have been reset, we may not send again. */
3591	if (sk->sk_state == TCP_CLOSE)
3592		return;
3593
3594	tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3595
3596	/* We are not putting this on the write queue, so
3597	 * tcp_transmit_skb() will set the ownership to this
3598	 * sock.
3599	 */
3600	buff = alloc_skb(MAX_TCP_HEADER,
3601			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3602	if (unlikely(!buff)) {
3603		inet_csk_schedule_ack(sk);
3604		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3605		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3606					  TCP_DELACK_MAX, TCP_RTO_MAX);
3607		return;
3608	}
3609
3610	/* Reserve space for headers and prepare control bits. */
3611	skb_reserve(buff, MAX_TCP_HEADER);
3612	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3613
3614	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3615	 * too much.
3616	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3617	 */
3618	skb_set_tcp_pure_ack(buff);
3619
3620	/* Send it off, this clears delayed acks for us. */
3621	tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
 
 
 
 
 
 
3622}
3623EXPORT_SYMBOL_GPL(tcp_send_ack);
3624
3625/* This routine sends a packet with an out of date sequence
3626 * number. It assumes the other end will try to ack it.
3627 *
3628 * Question: what should we make while urgent mode?
3629 * 4.4BSD forces sending single byte of data. We cannot send
3630 * out of window data, because we have SND.NXT==SND.MAX...
3631 *
3632 * Current solution: to send TWO zero-length segments in urgent mode:
3633 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3634 * out-of-date with SND.UNA-1 to probe window.
3635 */
3636static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3637{
3638	struct tcp_sock *tp = tcp_sk(sk);
3639	struct sk_buff *skb;
3640
3641	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3642	skb = alloc_skb(MAX_TCP_HEADER,
3643			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3644	if (!skb)
3645		return -1;
3646
3647	/* Reserve space for headers and set control bits. */
3648	skb_reserve(skb, MAX_TCP_HEADER);
3649	/* Use a previous sequence.  This should cause the other
3650	 * end to send an ack.  Don't queue or clone SKB, just
3651	 * send it.
3652	 */
3653	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3654	NET_INC_STATS(sock_net(sk), mib);
3655	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3656}
3657
3658/* Called from setsockopt( ... TCP_REPAIR ) */
3659void tcp_send_window_probe(struct sock *sk)
3660{
3661	if (sk->sk_state == TCP_ESTABLISHED) {
3662		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3663		tcp_mstamp_refresh(tcp_sk(sk));
3664		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3665	}
3666}
3667
3668/* Initiate keepalive or window probe from timer. */
3669int tcp_write_wakeup(struct sock *sk, int mib)
3670{
3671	struct tcp_sock *tp = tcp_sk(sk);
3672	struct sk_buff *skb;
3673
3674	if (sk->sk_state == TCP_CLOSE)
3675		return -1;
3676
3677	skb = tcp_send_head(sk);
3678	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3679		int err;
3680		unsigned int mss = tcp_current_mss(sk);
3681		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3682
3683		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3684			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3685
3686		/* We are probing the opening of a window
3687		 * but the window size is != 0
3688		 * must have been a result SWS avoidance ( sender )
3689		 */
3690		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3691		    skb->len > mss) {
3692			seg_size = min(seg_size, mss);
3693			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3694			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3695					 skb, seg_size, mss, GFP_ATOMIC))
3696				return -1;
3697		} else if (!tcp_skb_pcount(skb))
3698			tcp_set_skb_tso_segs(skb, mss);
3699
3700		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3701		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3702		if (!err)
3703			tcp_event_new_data_sent(sk, skb);
3704		return err;
3705	} else {
3706		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3707			tcp_xmit_probe_skb(sk, 1, mib);
3708		return tcp_xmit_probe_skb(sk, 0, mib);
3709	}
3710}
3711
3712/* A window probe timeout has occurred.  If window is not closed send
3713 * a partial packet else a zero probe.
3714 */
3715void tcp_send_probe0(struct sock *sk)
3716{
3717	struct inet_connection_sock *icsk = inet_csk(sk);
3718	struct tcp_sock *tp = tcp_sk(sk);
3719	struct net *net = sock_net(sk);
3720	unsigned long probe_max;
3721	int err;
3722
3723	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3724
3725	if (tp->packets_out || tcp_write_queue_empty(sk)) {
3726		/* Cancel probe timer, if it is not required. */
3727		icsk->icsk_probes_out = 0;
3728		icsk->icsk_backoff = 0;
3729		return;
3730	}
3731
 
3732	if (err <= 0) {
3733		if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3734			icsk->icsk_backoff++;
3735		icsk->icsk_probes_out++;
3736		probe_max = TCP_RTO_MAX;
3737	} else {
3738		/* If packet was not sent due to local congestion,
3739		 * do not backoff and do not remember icsk_probes_out.
3740		 * Let local senders to fight for local resources.
3741		 *
3742		 * Use accumulated backoff yet.
3743		 */
3744		if (!icsk->icsk_probes_out)
3745			icsk->icsk_probes_out = 1;
3746		probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3747	}
3748	inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3749				  tcp_probe0_when(sk, probe_max),
3750				  TCP_RTO_MAX);
3751}
3752
3753int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3754{
3755	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3756	struct flowi fl;
3757	int res;
3758
3759	tcp_rsk(req)->txhash = net_tx_rndhash();
3760	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3761	if (!res) {
3762		__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3763		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3764		if (unlikely(tcp_passive_fastopen(sk)))
3765			tcp_sk(sk)->total_retrans++;
3766		trace_tcp_retransmit_synack(sk, req);
3767	}
3768	return res;
3769}
3770EXPORT_SYMBOL(tcp_rtx_synack);
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * INET		An implementation of the TCP/IP protocol suite for the LINUX
   4 *		operating system.  INET is implemented using the  BSD Socket
   5 *		interface as the means of communication with the user level.
   6 *
   7 *		Implementation of the Transmission Control Protocol(TCP).
   8 *
   9 * Authors:	Ross Biro
  10 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  11 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
  12 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
  13 *		Florian La Roche, <flla@stud.uni-sb.de>
  14 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  15 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
  16 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
  17 *		Matthew Dillon, <dillon@apollo.west.oic.com>
  18 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  19 *		Jorge Cwik, <jorge@laser.satlink.net>
  20 */
  21
  22/*
  23 * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
  24 *				:	Fragmentation on mtu decrease
  25 *				:	Segment collapse on retransmit
  26 *				:	AF independence
  27 *
  28 *		Linus Torvalds	:	send_delayed_ack
  29 *		David S. Miller	:	Charge memory using the right skb
  30 *					during syn/ack processing.
  31 *		David S. Miller :	Output engine completely rewritten.
  32 *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
  33 *		Cacophonix Gaul :	draft-minshall-nagle-01
  34 *		J Hadi Salim	:	ECN support
  35 *
  36 */
  37
  38#define pr_fmt(fmt) "TCP: " fmt
  39
  40#include <net/tcp.h>
  41#include <net/mptcp.h>
  42
  43#include <linux/compiler.h>
  44#include <linux/gfp.h>
  45#include <linux/module.h>
  46#include <linux/static_key.h>
  47
  48#include <trace/events/tcp.h>
  49
  50/* Refresh clocks of a TCP socket,
  51 * ensuring monotically increasing values.
  52 */
  53void tcp_mstamp_refresh(struct tcp_sock *tp)
  54{
  55	u64 val = tcp_clock_ns();
  56
  57	tp->tcp_clock_cache = val;
  58	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  59}
  60
  61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  62			   int push_one, gfp_t gfp);
  63
  64/* Account for new data that has been sent to the network. */
  65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  66{
  67	struct inet_connection_sock *icsk = inet_csk(sk);
  68	struct tcp_sock *tp = tcp_sk(sk);
  69	unsigned int prior_packets = tp->packets_out;
  70
  71	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
  72
  73	__skb_unlink(skb, &sk->sk_write_queue);
  74	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  75
  76	if (tp->highest_sack == NULL)
  77		tp->highest_sack = skb;
  78
  79	tp->packets_out += tcp_skb_pcount(skb);
  80	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  81		tcp_rearm_rto(sk);
  82
  83	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  84		      tcp_skb_pcount(skb));
  85}
  86
  87/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  88 * window scaling factor due to loss of precision.
  89 * If window has been shrunk, what should we make? It is not clear at all.
  90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  92 * invalid. OK, let's make this for now:
  93 */
  94static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  95{
  96	const struct tcp_sock *tp = tcp_sk(sk);
  97
  98	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
  99	    (tp->rx_opt.wscale_ok &&
 100	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
 101		return tp->snd_nxt;
 102	else
 103		return tcp_wnd_end(tp);
 104}
 105
 106/* Calculate mss to advertise in SYN segment.
 107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 108 *
 109 * 1. It is independent of path mtu.
 110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 112 *    attached devices, because some buggy hosts are confused by
 113 *    large MSS.
 114 * 4. We do not make 3, we advertise MSS, calculated from first
 115 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 116 *    This may be overridden via information stored in routing table.
 117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 118 *    probably even Jumbo".
 119 */
 120static __u16 tcp_advertise_mss(struct sock *sk)
 121{
 122	struct tcp_sock *tp = tcp_sk(sk);
 123	const struct dst_entry *dst = __sk_dst_get(sk);
 124	int mss = tp->advmss;
 125
 126	if (dst) {
 127		unsigned int metric = dst_metric_advmss(dst);
 128
 129		if (metric < mss) {
 130			mss = metric;
 131			tp->advmss = mss;
 132		}
 133	}
 134
 135	return (__u16)mss;
 136}
 137
 138/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 139 * This is the first part of cwnd validation mechanism.
 140 */
 141void tcp_cwnd_restart(struct sock *sk, s32 delta)
 142{
 143	struct tcp_sock *tp = tcp_sk(sk);
 144	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 145	u32 cwnd = tp->snd_cwnd;
 146
 147	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 148
 149	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 150	restart_cwnd = min(restart_cwnd, cwnd);
 151
 152	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 153		cwnd >>= 1;
 154	tp->snd_cwnd = max(cwnd, restart_cwnd);
 155	tp->snd_cwnd_stamp = tcp_jiffies32;
 156	tp->snd_cwnd_used = 0;
 157}
 158
 159/* Congestion state accounting after a packet has been sent. */
 160static void tcp_event_data_sent(struct tcp_sock *tp,
 161				struct sock *sk)
 162{
 163	struct inet_connection_sock *icsk = inet_csk(sk);
 164	const u32 now = tcp_jiffies32;
 165
 166	if (tcp_packets_in_flight(tp) == 0)
 167		tcp_ca_event(sk, CA_EVENT_TX_START);
 168
 169	/* If this is the first data packet sent in response to the
 170	 * previous received data,
 171	 * and it is a reply for ato after last received packet,
 172	 * increase pingpong count.
 173	 */
 174	if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
 175	    (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 176		inet_csk_inc_pingpong_cnt(sk);
 177
 178	tp->lsndtime = now;
 
 
 
 
 179}
 180
 181/* Account for an ACK we sent. */
 182static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
 183				      u32 rcv_nxt)
 184{
 185	struct tcp_sock *tp = tcp_sk(sk);
 
 
 186
 187	if (unlikely(tp->compressed_ack)) {
 188		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 189			      tp->compressed_ack);
 190		tp->compressed_ack = 0;
 191		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 192			__sock_put(sk);
 193	}
 194
 195	if (unlikely(rcv_nxt != tp->rcv_nxt))
 196		return;  /* Special ACK sent by DCTCP to reflect ECN */
 197	tcp_dec_quickack_mode(sk, pkts);
 198	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 
 
 
 
 
 
 
 
 199}
 200
 201/* Determine a window scaling and initial window to offer.
 202 * Based on the assumption that the given amount of space
 203 * will be offered. Store the results in the tp structure.
 204 * NOTE: for smooth operation initial space offering should
 205 * be a multiple of mss if possible. We assume here that mss >= 1.
 206 * This MUST be enforced by all callers.
 207 */
 208void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 209			       __u32 *rcv_wnd, __u32 *window_clamp,
 210			       int wscale_ok, __u8 *rcv_wscale,
 211			       __u32 init_rcv_wnd)
 212{
 213	unsigned int space = (__space < 0 ? 0 : __space);
 214
 215	/* If no clamp set the clamp to the max possible scaled window */
 216	if (*window_clamp == 0)
 217		(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
 218	space = min(*window_clamp, space);
 219
 220	/* Quantize space offering to a multiple of mss if possible. */
 221	if (space > mss)
 222		space = rounddown(space, mss);
 223
 224	/* NOTE: offering an initial window larger than 32767
 225	 * will break some buggy TCP stacks. If the admin tells us
 226	 * it is likely we could be speaking with such a buggy stack
 227	 * we will truncate our initial window offering to 32K-1
 228	 * unless the remote has sent us a window scaling option,
 229	 * which we interpret as a sign the remote TCP is not
 230	 * misinterpreting the window field as a signed quantity.
 231	 */
 232	if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 233		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 234	else
 235		(*rcv_wnd) = min_t(u32, space, U16_MAX);
 236
 237	if (init_rcv_wnd)
 238		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 239
 240	*rcv_wscale = 0;
 241	if (wscale_ok) {
 242		/* Set window scaling on max possible window */
 243		space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
 244		space = max_t(u32, space, sysctl_rmem_max);
 245		space = min_t(u32, space, *window_clamp);
 246		*rcv_wscale = clamp_t(int, ilog2(space) - 15,
 247				      0, TCP_MAX_WSCALE);
 
 
 
 
 
 
 
 
 248	}
 
 249	/* Set the clamp no higher than max representable value */
 250	(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
 251}
 252EXPORT_SYMBOL(tcp_select_initial_window);
 253
 254/* Chose a new window to advertise, update state in tcp_sock for the
 255 * socket, and return result with RFC1323 scaling applied.  The return
 256 * value can be stuffed directly into th->window for an outgoing
 257 * frame.
 258 */
 259static u16 tcp_select_window(struct sock *sk)
 260{
 261	struct tcp_sock *tp = tcp_sk(sk);
 262	u32 old_win = tp->rcv_wnd;
 263	u32 cur_win = tcp_receive_window(tp);
 264	u32 new_win = __tcp_select_window(sk);
 265
 266	/* Never shrink the offered window */
 267	if (new_win < cur_win) {
 268		/* Danger Will Robinson!
 269		 * Don't update rcv_wup/rcv_wnd here or else
 270		 * we will not be able to advertise a zero
 271		 * window in time.  --DaveM
 272		 *
 273		 * Relax Will Robinson.
 274		 */
 275		if (new_win == 0)
 276			NET_INC_STATS(sock_net(sk),
 277				      LINUX_MIB_TCPWANTZEROWINDOWADV);
 278		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 279	}
 280	tp->rcv_wnd = new_win;
 281	tp->rcv_wup = tp->rcv_nxt;
 282
 283	/* Make sure we do not exceed the maximum possible
 284	 * scaled window.
 285	 */
 286	if (!tp->rx_opt.rcv_wscale &&
 287	    sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 288		new_win = min(new_win, MAX_TCP_WINDOW);
 289	else
 290		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 291
 292	/* RFC1323 scaling applied */
 293	new_win >>= tp->rx_opt.rcv_wscale;
 294
 295	/* If we advertise zero window, disable fast path. */
 296	if (new_win == 0) {
 297		tp->pred_flags = 0;
 298		if (old_win)
 299			NET_INC_STATS(sock_net(sk),
 300				      LINUX_MIB_TCPTOZEROWINDOWADV);
 301	} else if (old_win == 0) {
 302		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
 303	}
 304
 305	return new_win;
 306}
 307
 308/* Packet ECN state for a SYN-ACK */
 309static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 310{
 311	const struct tcp_sock *tp = tcp_sk(sk);
 312
 313	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 314	if (!(tp->ecn_flags & TCP_ECN_OK))
 315		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 316	else if (tcp_ca_needs_ecn(sk) ||
 317		 tcp_bpf_ca_needs_ecn(sk))
 318		INET_ECN_xmit(sk);
 319}
 320
 321/* Packet ECN state for a SYN.  */
 322static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 323{
 324	struct tcp_sock *tp = tcp_sk(sk);
 325	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 326	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
 327		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 328
 329	if (!use_ecn) {
 330		const struct dst_entry *dst = __sk_dst_get(sk);
 331
 332		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 333			use_ecn = true;
 334	}
 335
 336	tp->ecn_flags = 0;
 337
 338	if (use_ecn) {
 339		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 340		tp->ecn_flags = TCP_ECN_OK;
 341		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 342			INET_ECN_xmit(sk);
 343	}
 344}
 345
 346static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 347{
 348	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
 349		/* tp->ecn_flags are cleared at a later point in time when
 350		 * SYN ACK is ultimatively being received.
 351		 */
 352		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 353}
 354
 355static void
 356tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 357{
 358	if (inet_rsk(req)->ecn_ok)
 359		th->ece = 1;
 360}
 361
 362/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 363 * be sent.
 364 */
 365static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 366			 struct tcphdr *th, int tcp_header_len)
 367{
 368	struct tcp_sock *tp = tcp_sk(sk);
 369
 370	if (tp->ecn_flags & TCP_ECN_OK) {
 371		/* Not-retransmitted data segment: set ECT and inject CWR. */
 372		if (skb->len != tcp_header_len &&
 373		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 374			INET_ECN_xmit(sk);
 375			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 376				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 377				th->cwr = 1;
 378				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 379			}
 380		} else if (!tcp_ca_needs_ecn(sk)) {
 381			/* ACK or retransmitted segment: clear ECT|CE */
 382			INET_ECN_dontxmit(sk);
 383		}
 384		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 385			th->ece = 1;
 386	}
 387}
 388
 389/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 390 * auto increment end seqno.
 391 */
 392static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 393{
 394	skb->ip_summed = CHECKSUM_PARTIAL;
 395
 396	TCP_SKB_CB(skb)->tcp_flags = flags;
 397	TCP_SKB_CB(skb)->sacked = 0;
 398
 399	tcp_skb_pcount_set(skb, 1);
 400
 401	TCP_SKB_CB(skb)->seq = seq;
 402	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 403		seq++;
 404	TCP_SKB_CB(skb)->end_seq = seq;
 405}
 406
 407static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 408{
 409	return tp->snd_una != tp->snd_up;
 410}
 411
 412#define OPTION_SACK_ADVERTISE	(1 << 0)
 413#define OPTION_TS		(1 << 1)
 414#define OPTION_MD5		(1 << 2)
 415#define OPTION_WSCALE		(1 << 3)
 416#define OPTION_FAST_OPEN_COOKIE	(1 << 8)
 417#define OPTION_SMC		(1 << 9)
 418#define OPTION_MPTCP		(1 << 10)
 419
 420static void smc_options_write(__be32 *ptr, u16 *options)
 421{
 422#if IS_ENABLED(CONFIG_SMC)
 423	if (static_branch_unlikely(&tcp_have_smc)) {
 424		if (unlikely(OPTION_SMC & *options)) {
 425			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 426				       (TCPOPT_NOP  << 16) |
 427				       (TCPOPT_EXP <<  8) |
 428				       (TCPOLEN_EXP_SMC_BASE));
 429			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 430		}
 431	}
 432#endif
 433}
 434
 435struct tcp_out_options {
 436	u16 options;		/* bit field of OPTION_* */
 437	u16 mss;		/* 0 to disable */
 438	u8 ws;			/* window scale, 0 to disable */
 439	u8 num_sack_blocks;	/* number of SACK blocks to include */
 440	u8 hash_size;		/* bytes in hash_location */
 441	__u8 *hash_location;	/* temporary pointer, overloaded */
 442	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 443	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 444	struct mptcp_out_options mptcp;
 445};
 446
 447static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts)
 448{
 449#if IS_ENABLED(CONFIG_MPTCP)
 450	if (unlikely(OPTION_MPTCP & opts->options))
 451		mptcp_write_options(ptr, &opts->mptcp);
 452#endif
 453}
 454
 455/* Write previously computed TCP options to the packet.
 456 *
 457 * Beware: Something in the Internet is very sensitive to the ordering of
 458 * TCP options, we learned this through the hard way, so be careful here.
 459 * Luckily we can at least blame others for their non-compliance but from
 460 * inter-operability perspective it seems that we're somewhat stuck with
 461 * the ordering which we have been using if we want to keep working with
 462 * those broken things (not that it currently hurts anybody as there isn't
 463 * particular reason why the ordering would need to be changed).
 464 *
 465 * At least SACK_PERM as the first option is known to lead to a disaster
 466 * (but it may well be that other scenarios fail similarly).
 467 */
 468static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 469			      struct tcp_out_options *opts)
 470{
 471	u16 options = opts->options;	/* mungable copy */
 472
 473	if (unlikely(OPTION_MD5 & options)) {
 474		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 475			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 476		/* overload cookie hash location */
 477		opts->hash_location = (__u8 *)ptr;
 478		ptr += 4;
 479	}
 480
 481	if (unlikely(opts->mss)) {
 482		*ptr++ = htonl((TCPOPT_MSS << 24) |
 483			       (TCPOLEN_MSS << 16) |
 484			       opts->mss);
 485	}
 486
 487	if (likely(OPTION_TS & options)) {
 488		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 489			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 490				       (TCPOLEN_SACK_PERM << 16) |
 491				       (TCPOPT_TIMESTAMP << 8) |
 492				       TCPOLEN_TIMESTAMP);
 493			options &= ~OPTION_SACK_ADVERTISE;
 494		} else {
 495			*ptr++ = htonl((TCPOPT_NOP << 24) |
 496				       (TCPOPT_NOP << 16) |
 497				       (TCPOPT_TIMESTAMP << 8) |
 498				       TCPOLEN_TIMESTAMP);
 499		}
 500		*ptr++ = htonl(opts->tsval);
 501		*ptr++ = htonl(opts->tsecr);
 502	}
 503
 504	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 505		*ptr++ = htonl((TCPOPT_NOP << 24) |
 506			       (TCPOPT_NOP << 16) |
 507			       (TCPOPT_SACK_PERM << 8) |
 508			       TCPOLEN_SACK_PERM);
 509	}
 510
 511	if (unlikely(OPTION_WSCALE & options)) {
 512		*ptr++ = htonl((TCPOPT_NOP << 24) |
 513			       (TCPOPT_WINDOW << 16) |
 514			       (TCPOLEN_WINDOW << 8) |
 515			       opts->ws);
 516	}
 517
 518	if (unlikely(opts->num_sack_blocks)) {
 519		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 520			tp->duplicate_sack : tp->selective_acks;
 521		int this_sack;
 522
 523		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 524			       (TCPOPT_NOP  << 16) |
 525			       (TCPOPT_SACK <<  8) |
 526			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 527						     TCPOLEN_SACK_PERBLOCK)));
 528
 529		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 530		     ++this_sack) {
 531			*ptr++ = htonl(sp[this_sack].start_seq);
 532			*ptr++ = htonl(sp[this_sack].end_seq);
 533		}
 534
 535		tp->rx_opt.dsack = 0;
 536	}
 537
 538	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 539		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 540		u8 *p = (u8 *)ptr;
 541		u32 len; /* Fast Open option length */
 542
 543		if (foc->exp) {
 544			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 545			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 546				     TCPOPT_FASTOPEN_MAGIC);
 547			p += TCPOLEN_EXP_FASTOPEN_BASE;
 548		} else {
 549			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 550			*p++ = TCPOPT_FASTOPEN;
 551			*p++ = len;
 552		}
 553
 554		memcpy(p, foc->val, foc->len);
 555		if ((len & 3) == 2) {
 556			p[foc->len] = TCPOPT_NOP;
 557			p[foc->len + 1] = TCPOPT_NOP;
 558		}
 559		ptr += (len + 3) >> 2;
 560	}
 561
 562	smc_options_write(ptr, &options);
 563
 564	mptcp_options_write(ptr, opts);
 565}
 566
 567static void smc_set_option(const struct tcp_sock *tp,
 568			   struct tcp_out_options *opts,
 569			   unsigned int *remaining)
 570{
 571#if IS_ENABLED(CONFIG_SMC)
 572	if (static_branch_unlikely(&tcp_have_smc)) {
 573		if (tp->syn_smc) {
 574			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 575				opts->options |= OPTION_SMC;
 576				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 577			}
 578		}
 579	}
 580#endif
 581}
 582
 583static void smc_set_option_cond(const struct tcp_sock *tp,
 584				const struct inet_request_sock *ireq,
 585				struct tcp_out_options *opts,
 586				unsigned int *remaining)
 587{
 588#if IS_ENABLED(CONFIG_SMC)
 589	if (static_branch_unlikely(&tcp_have_smc)) {
 590		if (tp->syn_smc && ireq->smc_ok) {
 591			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 592				opts->options |= OPTION_SMC;
 593				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 594			}
 595		}
 596	}
 597#endif
 598}
 599
 600static void mptcp_set_option_cond(const struct request_sock *req,
 601				  struct tcp_out_options *opts,
 602				  unsigned int *remaining)
 603{
 604	if (rsk_is_mptcp(req)) {
 605		unsigned int size;
 606
 607		if (mptcp_synack_options(req, &size, &opts->mptcp)) {
 608			if (*remaining >= size) {
 609				opts->options |= OPTION_MPTCP;
 610				*remaining -= size;
 611			}
 612		}
 613	}
 614}
 615
 616/* Compute TCP options for SYN packets. This is not the final
 617 * network wire format yet.
 618 */
 619static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 620				struct tcp_out_options *opts,
 621				struct tcp_md5sig_key **md5)
 622{
 623	struct tcp_sock *tp = tcp_sk(sk);
 624	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 625	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 626
 627	*md5 = NULL;
 628#ifdef CONFIG_TCP_MD5SIG
 629	if (static_branch_unlikely(&tcp_md5_needed) &&
 630	    rcu_access_pointer(tp->md5sig_info)) {
 631		*md5 = tp->af_specific->md5_lookup(sk, sk);
 632		if (*md5) {
 633			opts->options |= OPTION_MD5;
 634			remaining -= TCPOLEN_MD5SIG_ALIGNED;
 635		}
 636	}
 
 
 637#endif
 638
 639	/* We always get an MSS option.  The option bytes which will be seen in
 640	 * normal data packets should timestamps be used, must be in the MSS
 641	 * advertised.  But we subtract them from tp->mss_cache so that
 642	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 643	 * fact here if necessary.  If we don't do this correctly, as a
 644	 * receiver we won't recognize data packets as being full sized when we
 645	 * should, and thus we won't abide by the delayed ACK rules correctly.
 646	 * SACKs don't matter, we never delay an ACK when we have any of those
 647	 * going out.  */
 648	opts->mss = tcp_advertise_mss(sk);
 649	remaining -= TCPOLEN_MSS_ALIGNED;
 650
 651	if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
 652		opts->options |= OPTION_TS;
 653		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
 654		opts->tsecr = tp->rx_opt.ts_recent;
 655		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 656	}
 657	if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
 658		opts->ws = tp->rx_opt.rcv_wscale;
 659		opts->options |= OPTION_WSCALE;
 660		remaining -= TCPOLEN_WSCALE_ALIGNED;
 661	}
 662	if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
 663		opts->options |= OPTION_SACK_ADVERTISE;
 664		if (unlikely(!(OPTION_TS & opts->options)))
 665			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 666	}
 667
 668	if (fastopen && fastopen->cookie.len >= 0) {
 669		u32 need = fastopen->cookie.len;
 670
 671		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 672					       TCPOLEN_FASTOPEN_BASE;
 673		need = (need + 3) & ~3U;  /* Align to 32 bits */
 674		if (remaining >= need) {
 675			opts->options |= OPTION_FAST_OPEN_COOKIE;
 676			opts->fastopen_cookie = &fastopen->cookie;
 677			remaining -= need;
 678			tp->syn_fastopen = 1;
 679			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 680		}
 681	}
 682
 683	smc_set_option(tp, opts, &remaining);
 684
 685	if (sk_is_mptcp(sk)) {
 686		unsigned int size;
 687
 688		if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
 689			opts->options |= OPTION_MPTCP;
 690			remaining -= size;
 691		}
 692	}
 693
 694	return MAX_TCP_OPTION_SPACE - remaining;
 695}
 696
 697/* Set up TCP options for SYN-ACKs. */
 698static unsigned int tcp_synack_options(const struct sock *sk,
 699				       struct request_sock *req,
 700				       unsigned int mss, struct sk_buff *skb,
 701				       struct tcp_out_options *opts,
 702				       const struct tcp_md5sig_key *md5,
 703				       struct tcp_fastopen_cookie *foc,
 704				       enum tcp_synack_type synack_type)
 705{
 706	struct inet_request_sock *ireq = inet_rsk(req);
 707	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 708
 709#ifdef CONFIG_TCP_MD5SIG
 710	if (md5) {
 711		opts->options |= OPTION_MD5;
 712		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 713
 714		/* We can't fit any SACK blocks in a packet with MD5 + TS
 715		 * options. There was discussion about disabling SACK
 716		 * rather than TS in order to fit in better with old,
 717		 * buggy kernels, but that was deemed to be unnecessary.
 718		 */
 719		if (synack_type != TCP_SYNACK_COOKIE)
 720			ireq->tstamp_ok &= !ireq->sack_ok;
 721	}
 722#endif
 723
 724	/* We always send an MSS option. */
 725	opts->mss = mss;
 726	remaining -= TCPOLEN_MSS_ALIGNED;
 727
 728	if (likely(ireq->wscale_ok)) {
 729		opts->ws = ireq->rcv_wscale;
 730		opts->options |= OPTION_WSCALE;
 731		remaining -= TCPOLEN_WSCALE_ALIGNED;
 732	}
 733	if (likely(ireq->tstamp_ok)) {
 734		opts->options |= OPTION_TS;
 735		opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
 736		opts->tsecr = req->ts_recent;
 737		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 738	}
 739	if (likely(ireq->sack_ok)) {
 740		opts->options |= OPTION_SACK_ADVERTISE;
 741		if (unlikely(!ireq->tstamp_ok))
 742			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 743	}
 744	if (foc != NULL && foc->len >= 0) {
 745		u32 need = foc->len;
 746
 747		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 748				   TCPOLEN_FASTOPEN_BASE;
 749		need = (need + 3) & ~3U;  /* Align to 32 bits */
 750		if (remaining >= need) {
 751			opts->options |= OPTION_FAST_OPEN_COOKIE;
 752			opts->fastopen_cookie = foc;
 753			remaining -= need;
 754		}
 755	}
 756
 757	mptcp_set_option_cond(req, opts, &remaining);
 758
 759	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 760
 761	return MAX_TCP_OPTION_SPACE - remaining;
 762}
 763
 764/* Compute TCP options for ESTABLISHED sockets. This is not the
 765 * final wire format yet.
 766 */
 767static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 768					struct tcp_out_options *opts,
 769					struct tcp_md5sig_key **md5)
 770{
 771	struct tcp_sock *tp = tcp_sk(sk);
 772	unsigned int size = 0;
 773	unsigned int eff_sacks;
 774
 775	opts->options = 0;
 776
 777	*md5 = NULL;
 778#ifdef CONFIG_TCP_MD5SIG
 779	if (static_branch_unlikely(&tcp_md5_needed) &&
 780	    rcu_access_pointer(tp->md5sig_info)) {
 781		*md5 = tp->af_specific->md5_lookup(sk, sk);
 782		if (*md5) {
 783			opts->options |= OPTION_MD5;
 784			size += TCPOLEN_MD5SIG_ALIGNED;
 785		}
 786	}
 
 
 787#endif
 788
 789	if (likely(tp->rx_opt.tstamp_ok)) {
 790		opts->options |= OPTION_TS;
 791		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
 792		opts->tsecr = tp->rx_opt.ts_recent;
 793		size += TCPOLEN_TSTAMP_ALIGNED;
 794	}
 795
 796	/* MPTCP options have precedence over SACK for the limited TCP
 797	 * option space because a MPTCP connection would be forced to
 798	 * fall back to regular TCP if a required multipath option is
 799	 * missing. SACK still gets a chance to use whatever space is
 800	 * left.
 801	 */
 802	if (sk_is_mptcp(sk)) {
 803		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 804		unsigned int opt_size = 0;
 805
 806		if (mptcp_established_options(sk, skb, &opt_size, remaining,
 807					      &opts->mptcp)) {
 808			opts->options |= OPTION_MPTCP;
 809			size += opt_size;
 810		}
 811	}
 812
 813	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 814	if (unlikely(eff_sacks)) {
 815		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 816		if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
 817					 TCPOLEN_SACK_PERBLOCK))
 818			return size;
 819
 820		opts->num_sack_blocks =
 821			min_t(unsigned int, eff_sacks,
 822			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 823			      TCPOLEN_SACK_PERBLOCK);
 824
 825		size += TCPOLEN_SACK_BASE_ALIGNED +
 826			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 827	}
 828
 829	return size;
 830}
 831
 832
 833/* TCP SMALL QUEUES (TSQ)
 834 *
 835 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 836 * to reduce RTT and bufferbloat.
 837 * We do this using a special skb destructor (tcp_wfree).
 838 *
 839 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 840 * needs to be reallocated in a driver.
 841 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
 842 *
 843 * Since transmit from skb destructor is forbidden, we use a tasklet
 844 * to process all sockets that eventually need to send more skbs.
 845 * We use one tasklet per cpu, with its own queue of sockets.
 846 */
 847struct tsq_tasklet {
 848	struct tasklet_struct	tasklet;
 849	struct list_head	head; /* queue of tcp sockets */
 850};
 851static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
 852
 853static void tcp_tsq_write(struct sock *sk)
 854{
 855	if ((1 << sk->sk_state) &
 856	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
 857	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
 858		struct tcp_sock *tp = tcp_sk(sk);
 859
 860		if (tp->lost_out > tp->retrans_out &&
 861		    tp->snd_cwnd > tcp_packets_in_flight(tp)) {
 862			tcp_mstamp_refresh(tp);
 863			tcp_xmit_retransmit_queue(sk);
 864		}
 865
 866		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
 867			       0, GFP_ATOMIC);
 868	}
 869}
 870
 871static void tcp_tsq_handler(struct sock *sk)
 872{
 873	bh_lock_sock(sk);
 874	if (!sock_owned_by_user(sk))
 875		tcp_tsq_write(sk);
 876	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
 877		sock_hold(sk);
 878	bh_unlock_sock(sk);
 879}
 880/*
 881 * One tasklet per cpu tries to send more skbs.
 882 * We run in tasklet context but need to disable irqs when
 883 * transferring tsq->head because tcp_wfree() might
 884 * interrupt us (non NAPI drivers)
 885 */
 886static void tcp_tasklet_func(unsigned long data)
 887{
 888	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
 889	LIST_HEAD(list);
 890	unsigned long flags;
 891	struct list_head *q, *n;
 892	struct tcp_sock *tp;
 893	struct sock *sk;
 894
 895	local_irq_save(flags);
 896	list_splice_init(&tsq->head, &list);
 897	local_irq_restore(flags);
 898
 899	list_for_each_safe(q, n, &list) {
 900		tp = list_entry(q, struct tcp_sock, tsq_node);
 901		list_del(&tp->tsq_node);
 902
 903		sk = (struct sock *)tp;
 904		smp_mb__before_atomic();
 905		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
 906
 907		tcp_tsq_handler(sk);
 
 
 
 
 
 
 
 
 
 908		sk_free(sk);
 909	}
 910}
 911
 912#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
 913			  TCPF_WRITE_TIMER_DEFERRED |	\
 914			  TCPF_DELACK_TIMER_DEFERRED |	\
 915			  TCPF_MTU_REDUCED_DEFERRED)
 916/**
 917 * tcp_release_cb - tcp release_sock() callback
 918 * @sk: socket
 919 *
 920 * called from release_sock() to perform protocol dependent
 921 * actions before socket release.
 922 */
 923void tcp_release_cb(struct sock *sk)
 924{
 925	unsigned long flags, nflags;
 926
 927	/* perform an atomic operation only if at least one flag is set */
 928	do {
 929		flags = sk->sk_tsq_flags;
 930		if (!(flags & TCP_DEFERRED_ALL))
 931			return;
 932		nflags = flags & ~TCP_DEFERRED_ALL;
 933	} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
 934
 935	if (flags & TCPF_TSQ_DEFERRED) {
 936		tcp_tsq_write(sk);
 937		__sock_put(sk);
 938	}
 939	/* Here begins the tricky part :
 940	 * We are called from release_sock() with :
 941	 * 1) BH disabled
 942	 * 2) sk_lock.slock spinlock held
 943	 * 3) socket owned by us (sk->sk_lock.owned == 1)
 944	 *
 945	 * But following code is meant to be called from BH handlers,
 946	 * so we should keep BH disabled, but early release socket ownership
 947	 */
 948	sock_release_ownership(sk);
 949
 950	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 951		tcp_write_timer_handler(sk);
 952		__sock_put(sk);
 953	}
 954	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
 955		tcp_delack_timer_handler(sk);
 956		__sock_put(sk);
 957	}
 958	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
 959		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 960		__sock_put(sk);
 961	}
 962}
 963EXPORT_SYMBOL(tcp_release_cb);
 964
 965void __init tcp_tasklet_init(void)
 966{
 967	int i;
 968
 969	for_each_possible_cpu(i) {
 970		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
 971
 972		INIT_LIST_HEAD(&tsq->head);
 973		tasklet_init(&tsq->tasklet,
 974			     tcp_tasklet_func,
 975			     (unsigned long)tsq);
 976	}
 977}
 978
 979/*
 980 * Write buffer destructor automatically called from kfree_skb.
 981 * We can't xmit new skbs from this context, as we might already
 982 * hold qdisc lock.
 983 */
 984void tcp_wfree(struct sk_buff *skb)
 985{
 986	struct sock *sk = skb->sk;
 987	struct tcp_sock *tp = tcp_sk(sk);
 988	unsigned long flags, nval, oval;
 989
 990	/* Keep one reference on sk_wmem_alloc.
 991	 * Will be released by sk_free() from here or tcp_tasklet_func()
 992	 */
 993	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
 994
 995	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
 996	 * Wait until our queues (qdisc + devices) are drained.
 997	 * This gives :
 998	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
 999	 * - chance for incoming ACK (processed by another cpu maybe)
1000	 *   to migrate this flow (skb->ooo_okay will be eventually set)
1001	 */
1002	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1003		goto out;
1004
1005	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1006		struct tsq_tasklet *tsq;
1007		bool empty;
1008
1009		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1010			goto out;
1011
1012		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1013		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1014		if (nval != oval)
1015			continue;
1016
1017		/* queue this socket to tasklet queue */
1018		local_irq_save(flags);
1019		tsq = this_cpu_ptr(&tsq_tasklet);
1020		empty = list_empty(&tsq->head);
1021		list_add(&tp->tsq_node, &tsq->head);
1022		if (empty)
1023			tasklet_schedule(&tsq->tasklet);
1024		local_irq_restore(flags);
1025		return;
1026	}
1027out:
1028	sk_free(sk);
1029}
1030
1031/* Note: Called under soft irq.
1032 * We can call TCP stack right away, unless socket is owned by user.
1033 */
1034enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1035{
1036	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1037	struct sock *sk = (struct sock *)tp;
 
 
 
 
 
1038
1039	tcp_tsq_handler(sk);
1040	sock_put(sk);
1041
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1042	return HRTIMER_NORESTART;
1043}
1044
1045static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1046				      u64 prior_wstamp)
 
 
 
 
 
 
 
 
 
 
1047{
1048	struct tcp_sock *tp = tcp_sk(sk);
 
1049
1050	if (sk->sk_pacing_status != SK_PACING_NONE) {
1051		unsigned long rate = sk->sk_pacing_rate;
 
 
 
1052
1053		/* Original sch_fq does not pace first 10 MSS
1054		 * Note that tp->data_segs_out overflows after 2^32 packets,
1055		 * this is a minor annoyance.
1056		 */
1057		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1058			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1059			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
 
 
1060
1061			/* take into account OS jitter */
1062			len_ns -= min_t(u64, len_ns / 2, credit);
1063			tp->tcp_wstamp_ns += len_ns;
1064		}
1065	}
1066	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1067}
1068
1069INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1070INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1071INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1072
1073/* This routine actually transmits TCP packets queued in by
1074 * tcp_do_sendmsg().  This is used by both the initial
1075 * transmission and possible later retransmissions.
1076 * All SKB's seen here are completely headerless.  It is our
1077 * job to build the TCP header, and pass the packet down to
1078 * IP so it can do the same plus pass the packet off to the
1079 * device.
1080 *
1081 * We are working here with either a clone of the original
1082 * SKB, or a fresh unique copy made by the retransmit engine.
1083 */
1084static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1085			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1086{
1087	const struct inet_connection_sock *icsk = inet_csk(sk);
1088	struct inet_sock *inet;
1089	struct tcp_sock *tp;
1090	struct tcp_skb_cb *tcb;
1091	struct tcp_out_options opts;
1092	unsigned int tcp_options_size, tcp_header_size;
1093	struct sk_buff *oskb = NULL;
1094	struct tcp_md5sig_key *md5;
1095	struct tcphdr *th;
1096	u64 prior_wstamp;
1097	int err;
1098
1099	BUG_ON(!skb || !tcp_skb_pcount(skb));
1100	tp = tcp_sk(sk);
1101	prior_wstamp = tp->tcp_wstamp_ns;
1102	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1103	skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1104	if (clone_it) {
1105		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1106			- tp->snd_una;
1107		oskb = skb;
1108
1109		tcp_skb_tsorted_save(oskb) {
1110			if (unlikely(skb_cloned(oskb)))
1111				skb = pskb_copy(oskb, gfp_mask);
1112			else
1113				skb = skb_clone(oskb, gfp_mask);
1114		} tcp_skb_tsorted_restore(oskb);
1115
1116		if (unlikely(!skb))
1117			return -ENOBUFS;
1118		/* retransmit skbs might have a non zero value in skb->dev
1119		 * because skb->dev is aliased with skb->rbnode.rb_left
1120		 */
1121		skb->dev = NULL;
1122	}
 
1123
1124	inet = inet_sk(sk);
1125	tcb = TCP_SKB_CB(skb);
1126	memset(&opts, 0, sizeof(opts));
1127
1128	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1129		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1130	} else {
1131		tcp_options_size = tcp_established_options(sk, skb, &opts,
1132							   &md5);
1133		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1134		 * at receiver : This slightly improve GRO performance.
1135		 * Note that we do not force the PSH flag for non GSO packets,
1136		 * because they might be sent under high congestion events,
1137		 * and in this case it is better to delay the delivery of 1-MSS
1138		 * packets and thus the corresponding ACK packet that would
1139		 * release the following packet.
1140		 */
1141		if (tcp_skb_pcount(skb) > 1)
1142			tcb->tcp_flags |= TCPHDR_PSH;
1143	}
1144	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1145
1146	/* if no packet is in qdisc/device queue, then allow XPS to select
1147	 * another queue. We can be called from tcp_tsq_handler()
1148	 * which holds one reference to sk.
1149	 *
1150	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1151	 * One way to get this would be to set skb->truesize = 2 on them.
1152	 */
1153	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1154
1155	/* If we had to use memory reserve to allocate this skb,
1156	 * this might cause drops if packet is looped back :
1157	 * Other socket might not have SOCK_MEMALLOC.
1158	 * Packets not looped back do not care about pfmemalloc.
1159	 */
1160	skb->pfmemalloc = 0;
1161
1162	skb_push(skb, tcp_header_size);
1163	skb_reset_transport_header(skb);
1164
1165	skb_orphan(skb);
1166	skb->sk = sk;
1167	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1168	skb_set_hash_from_sk(skb, sk);
1169	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1170
1171	skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1172
1173	/* Build TCP header and checksum it. */
1174	th = (struct tcphdr *)skb->data;
1175	th->source		= inet->inet_sport;
1176	th->dest		= inet->inet_dport;
1177	th->seq			= htonl(tcb->seq);
1178	th->ack_seq		= htonl(rcv_nxt);
1179	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1180					tcb->tcp_flags);
1181
1182	th->check		= 0;
1183	th->urg_ptr		= 0;
1184
1185	/* The urg_mode check is necessary during a below snd_una win probe */
1186	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1187		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1188			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1189			th->urg = 1;
1190		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1191			th->urg_ptr = htons(0xFFFF);
1192			th->urg = 1;
1193		}
1194	}
1195
1196	tcp_options_write((__be32 *)(th + 1), tp, &opts);
1197	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1198	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1199		th->window      = htons(tcp_select_window(sk));
1200		tcp_ecn_send(sk, skb, th, tcp_header_size);
1201	} else {
1202		/* RFC1323: The window in SYN & SYN/ACK segments
1203		 * is never scaled.
1204		 */
1205		th->window	= htons(min(tp->rcv_wnd, 65535U));
1206	}
1207#ifdef CONFIG_TCP_MD5SIG
1208	/* Calculate the MD5 hash, as we have all we need now */
1209	if (md5) {
1210		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1211		tp->af_specific->calc_md5_hash(opts.hash_location,
1212					       md5, sk, skb);
1213	}
1214#endif
1215
1216	INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1217			   tcp_v6_send_check, tcp_v4_send_check,
1218			   sk, skb);
1219
1220	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1221		tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1222
1223	if (skb->len != tcp_header_size) {
1224		tcp_event_data_sent(tp, sk);
1225		tp->data_segs_out += tcp_skb_pcount(skb);
1226		tp->bytes_sent += skb->len - tcp_header_size;
1227	}
1228
1229	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1230		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1231			      tcp_skb_pcount(skb));
1232
1233	tp->segs_out += tcp_skb_pcount(skb);
1234	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1235	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1236	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1237
1238	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
 
1239
1240	/* Cleanup our debris for IP stacks */
1241	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1242			       sizeof(struct inet6_skb_parm)));
1243
1244	tcp_add_tx_delay(skb, tp);
1245
1246	err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1247				 inet6_csk_xmit, ip_queue_xmit,
1248				 sk, skb, &inet->cork.fl);
1249
1250	if (unlikely(err > 0)) {
1251		tcp_enter_cwr(sk);
1252		err = net_xmit_eval(err);
1253	}
1254	if (!err && oskb) {
1255		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1256		tcp_rate_skb_sent(sk, oskb);
1257	}
1258	return err;
1259}
1260
1261static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1262			    gfp_t gfp_mask)
1263{
1264	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1265				  tcp_sk(sk)->rcv_nxt);
1266}
1267
1268/* This routine just queues the buffer for sending.
1269 *
1270 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1271 * otherwise socket can stall.
1272 */
1273static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1274{
1275	struct tcp_sock *tp = tcp_sk(sk);
1276
1277	/* Advance write_seq and place onto the write_queue. */
1278	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1279	__skb_header_release(skb);
1280	tcp_add_write_queue_tail(sk, skb);
1281	sk_wmem_queued_add(sk, skb->truesize);
1282	sk_mem_charge(sk, skb->truesize);
1283}
1284
1285/* Initialize TSO segments for a packet. */
1286static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1287{
1288	if (skb->len <= mss_now) {
1289		/* Avoid the costly divide in the normal
1290		 * non-TSO case.
1291		 */
1292		tcp_skb_pcount_set(skb, 1);
1293		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1294	} else {
1295		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1296		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1297	}
1298}
1299
1300/* Pcount in the middle of the write queue got changed, we need to do various
1301 * tweaks to fix counters
1302 */
1303static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1304{
1305	struct tcp_sock *tp = tcp_sk(sk);
1306
1307	tp->packets_out -= decr;
1308
1309	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1310		tp->sacked_out -= decr;
1311	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1312		tp->retrans_out -= decr;
1313	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1314		tp->lost_out -= decr;
1315
1316	/* Reno case is special. Sigh... */
1317	if (tcp_is_reno(tp) && decr > 0)
1318		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1319
1320	if (tp->lost_skb_hint &&
1321	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1322	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1323		tp->lost_cnt_hint -= decr;
1324
1325	tcp_verify_left_out(tp);
1326}
1327
1328static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1329{
1330	return TCP_SKB_CB(skb)->txstamp_ack ||
1331		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1332}
1333
1334static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1335{
1336	struct skb_shared_info *shinfo = skb_shinfo(skb);
1337
1338	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1339	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1340		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1341		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1342
1343		shinfo->tx_flags &= ~tsflags;
1344		shinfo2->tx_flags |= tsflags;
1345		swap(shinfo->tskey, shinfo2->tskey);
1346		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1347		TCP_SKB_CB(skb)->txstamp_ack = 0;
1348	}
1349}
1350
1351static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1352{
1353	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1354	TCP_SKB_CB(skb)->eor = 0;
1355}
1356
1357/* Insert buff after skb on the write or rtx queue of sk.  */
1358static void tcp_insert_write_queue_after(struct sk_buff *skb,
1359					 struct sk_buff *buff,
1360					 struct sock *sk,
1361					 enum tcp_queue tcp_queue)
1362{
1363	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1364		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1365	else
1366		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1367}
1368
1369/* Function to create two new TCP segments.  Shrinks the given segment
1370 * to the specified size and appends a new segment with the rest of the
1371 * packet to the list.  This won't be called frequently, I hope.
1372 * Remember, these are still headerless SKBs at this point.
1373 */
1374int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1375		 struct sk_buff *skb, u32 len,
1376		 unsigned int mss_now, gfp_t gfp)
1377{
1378	struct tcp_sock *tp = tcp_sk(sk);
1379	struct sk_buff *buff;
1380	int nsize, old_factor;
1381	long limit;
1382	int nlen;
1383	u8 flags;
1384
1385	if (WARN_ON(len > skb->len))
1386		return -EINVAL;
1387
1388	nsize = skb_headlen(skb) - len;
1389	if (nsize < 0)
1390		nsize = 0;
1391
1392	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1393	 * We need some allowance to not penalize applications setting small
1394	 * SO_SNDBUF values.
1395	 * Also allow first and last skb in retransmit queue to be split.
1396	 */
1397	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1398	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1399		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1400		     skb != tcp_rtx_queue_head(sk) &&
1401		     skb != tcp_rtx_queue_tail(sk))) {
1402		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1403		return -ENOMEM;
1404	}
1405
1406	if (skb_unclone(skb, gfp))
1407		return -ENOMEM;
1408
1409	/* Get a new skb... force flag on. */
1410	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1411	if (!buff)
1412		return -ENOMEM; /* We'll just try again later. */
1413	skb_copy_decrypted(buff, skb);
1414
1415	sk_wmem_queued_add(sk, buff->truesize);
1416	sk_mem_charge(sk, buff->truesize);
1417	nlen = skb->len - len - nsize;
1418	buff->truesize += nlen;
1419	skb->truesize -= nlen;
1420
1421	/* Correct the sequence numbers. */
1422	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1423	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1424	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1425
1426	/* PSH and FIN should only be set in the second packet. */
1427	flags = TCP_SKB_CB(skb)->tcp_flags;
1428	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1429	TCP_SKB_CB(buff)->tcp_flags = flags;
1430	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1431	tcp_skb_fragment_eor(skb, buff);
1432
1433	skb_split(skb, buff, len);
1434
1435	buff->ip_summed = CHECKSUM_PARTIAL;
1436
1437	buff->tstamp = skb->tstamp;
1438	tcp_fragment_tstamp(skb, buff);
1439
1440	old_factor = tcp_skb_pcount(skb);
1441
1442	/* Fix up tso_factor for both original and new SKB.  */
1443	tcp_set_skb_tso_segs(skb, mss_now);
1444	tcp_set_skb_tso_segs(buff, mss_now);
1445
1446	/* Update delivered info for the new segment */
1447	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1448
1449	/* If this packet has been sent out already, we must
1450	 * adjust the various packet counters.
1451	 */
1452	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1453		int diff = old_factor - tcp_skb_pcount(skb) -
1454			tcp_skb_pcount(buff);
1455
1456		if (diff)
1457			tcp_adjust_pcount(sk, skb, diff);
1458	}
1459
1460	/* Link BUFF into the send queue. */
1461	__skb_header_release(buff);
1462	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1463	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1464		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1465
1466	return 0;
1467}
1468
1469/* This is similar to __pskb_pull_tail(). The difference is that pulled
1470 * data is not copied, but immediately discarded.
1471 */
1472static int __pskb_trim_head(struct sk_buff *skb, int len)
1473{
1474	struct skb_shared_info *shinfo;
1475	int i, k, eat;
1476
1477	eat = min_t(int, len, skb_headlen(skb));
1478	if (eat) {
1479		__skb_pull(skb, eat);
1480		len -= eat;
1481		if (!len)
1482			return 0;
1483	}
1484	eat = len;
1485	k = 0;
1486	shinfo = skb_shinfo(skb);
1487	for (i = 0; i < shinfo->nr_frags; i++) {
1488		int size = skb_frag_size(&shinfo->frags[i]);
1489
1490		if (size <= eat) {
1491			skb_frag_unref(skb, i);
1492			eat -= size;
1493		} else {
1494			shinfo->frags[k] = shinfo->frags[i];
1495			if (eat) {
1496				skb_frag_off_add(&shinfo->frags[k], eat);
1497				skb_frag_size_sub(&shinfo->frags[k], eat);
1498				eat = 0;
1499			}
1500			k++;
1501		}
1502	}
1503	shinfo->nr_frags = k;
1504
1505	skb->data_len -= len;
1506	skb->len = skb->data_len;
1507	return len;
1508}
1509
1510/* Remove acked data from a packet in the transmit queue. */
1511int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1512{
1513	u32 delta_truesize;
1514
1515	if (skb_unclone(skb, GFP_ATOMIC))
1516		return -ENOMEM;
1517
1518	delta_truesize = __pskb_trim_head(skb, len);
1519
1520	TCP_SKB_CB(skb)->seq += len;
1521	skb->ip_summed = CHECKSUM_PARTIAL;
1522
1523	if (delta_truesize) {
1524		skb->truesize	   -= delta_truesize;
1525		sk_wmem_queued_add(sk, -delta_truesize);
1526		sk_mem_uncharge(sk, delta_truesize);
1527		sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1528	}
1529
1530	/* Any change of skb->len requires recalculation of tso factor. */
1531	if (tcp_skb_pcount(skb) > 1)
1532		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1533
1534	return 0;
1535}
1536
1537/* Calculate MSS not accounting any TCP options.  */
1538static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1539{
1540	const struct tcp_sock *tp = tcp_sk(sk);
1541	const struct inet_connection_sock *icsk = inet_csk(sk);
1542	int mss_now;
1543
1544	/* Calculate base mss without TCP options:
1545	   It is MMS_S - sizeof(tcphdr) of rfc1122
1546	 */
1547	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1548
1549	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1550	if (icsk->icsk_af_ops->net_frag_header_len) {
1551		const struct dst_entry *dst = __sk_dst_get(sk);
1552
1553		if (dst && dst_allfrag(dst))
1554			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1555	}
1556
1557	/* Clamp it (mss_clamp does not include tcp options) */
1558	if (mss_now > tp->rx_opt.mss_clamp)
1559		mss_now = tp->rx_opt.mss_clamp;
1560
1561	/* Now subtract optional transport overhead */
1562	mss_now -= icsk->icsk_ext_hdr_len;
1563
1564	/* Then reserve room for full set of TCP options and 8 bytes of data */
1565	mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
 
1566	return mss_now;
1567}
1568
1569/* Calculate MSS. Not accounting for SACKs here.  */
1570int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1571{
1572	/* Subtract TCP options size, not including SACKs */
1573	return __tcp_mtu_to_mss(sk, pmtu) -
1574	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1575}
1576
1577/* Inverse of above */
1578int tcp_mss_to_mtu(struct sock *sk, int mss)
1579{
1580	const struct tcp_sock *tp = tcp_sk(sk);
1581	const struct inet_connection_sock *icsk = inet_csk(sk);
1582	int mtu;
1583
1584	mtu = mss +
1585	      tp->tcp_header_len +
1586	      icsk->icsk_ext_hdr_len +
1587	      icsk->icsk_af_ops->net_header_len;
1588
1589	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1590	if (icsk->icsk_af_ops->net_frag_header_len) {
1591		const struct dst_entry *dst = __sk_dst_get(sk);
1592
1593		if (dst && dst_allfrag(dst))
1594			mtu += icsk->icsk_af_ops->net_frag_header_len;
1595	}
1596	return mtu;
1597}
1598EXPORT_SYMBOL(tcp_mss_to_mtu);
1599
1600/* MTU probing init per socket */
1601void tcp_mtup_init(struct sock *sk)
1602{
1603	struct tcp_sock *tp = tcp_sk(sk);
1604	struct inet_connection_sock *icsk = inet_csk(sk);
1605	struct net *net = sock_net(sk);
1606
1607	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1608	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1609			       icsk->icsk_af_ops->net_header_len;
1610	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1611	icsk->icsk_mtup.probe_size = 0;
1612	if (icsk->icsk_mtup.enabled)
1613		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1614}
1615EXPORT_SYMBOL(tcp_mtup_init);
1616
1617/* This function synchronize snd mss to current pmtu/exthdr set.
1618
1619   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1620   for TCP options, but includes only bare TCP header.
1621
1622   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1623   It is minimum of user_mss and mss received with SYN.
1624   It also does not include TCP options.
1625
1626   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1627
1628   tp->mss_cache is current effective sending mss, including
1629   all tcp options except for SACKs. It is evaluated,
1630   taking into account current pmtu, but never exceeds
1631   tp->rx_opt.mss_clamp.
1632
1633   NOTE1. rfc1122 clearly states that advertised MSS
1634   DOES NOT include either tcp or ip options.
1635
1636   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1637   are READ ONLY outside this function.		--ANK (980731)
1638 */
1639unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1640{
1641	struct tcp_sock *tp = tcp_sk(sk);
1642	struct inet_connection_sock *icsk = inet_csk(sk);
1643	int mss_now;
1644
1645	if (icsk->icsk_mtup.search_high > pmtu)
1646		icsk->icsk_mtup.search_high = pmtu;
1647
1648	mss_now = tcp_mtu_to_mss(sk, pmtu);
1649	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1650
1651	/* And store cached results */
1652	icsk->icsk_pmtu_cookie = pmtu;
1653	if (icsk->icsk_mtup.enabled)
1654		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1655	tp->mss_cache = mss_now;
1656
1657	return mss_now;
1658}
1659EXPORT_SYMBOL(tcp_sync_mss);
1660
1661/* Compute the current effective MSS, taking SACKs and IP options,
1662 * and even PMTU discovery events into account.
1663 */
1664unsigned int tcp_current_mss(struct sock *sk)
1665{
1666	const struct tcp_sock *tp = tcp_sk(sk);
1667	const struct dst_entry *dst = __sk_dst_get(sk);
1668	u32 mss_now;
1669	unsigned int header_len;
1670	struct tcp_out_options opts;
1671	struct tcp_md5sig_key *md5;
1672
1673	mss_now = tp->mss_cache;
1674
1675	if (dst) {
1676		u32 mtu = dst_mtu(dst);
1677		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1678			mss_now = tcp_sync_mss(sk, mtu);
1679	}
1680
1681	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1682		     sizeof(struct tcphdr);
1683	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1684	 * some common options. If this is an odd packet (because we have SACK
1685	 * blocks etc) then our calculated header_len will be different, and
1686	 * we have to adjust mss_now correspondingly */
1687	if (header_len != tp->tcp_header_len) {
1688		int delta = (int) header_len - tp->tcp_header_len;
1689		mss_now -= delta;
1690	}
1691
1692	return mss_now;
1693}
1694
1695/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1696 * As additional protections, we do not touch cwnd in retransmission phases,
1697 * and if application hit its sndbuf limit recently.
1698 */
1699static void tcp_cwnd_application_limited(struct sock *sk)
1700{
1701	struct tcp_sock *tp = tcp_sk(sk);
1702
1703	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1704	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1705		/* Limited by application or receiver window. */
1706		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1707		u32 win_used = max(tp->snd_cwnd_used, init_win);
1708		if (win_used < tp->snd_cwnd) {
1709			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1710			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1711		}
1712		tp->snd_cwnd_used = 0;
1713	}
1714	tp->snd_cwnd_stamp = tcp_jiffies32;
1715}
1716
1717static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1718{
1719	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1720	struct tcp_sock *tp = tcp_sk(sk);
1721
1722	/* Track the maximum number of outstanding packets in each
1723	 * window, and remember whether we were cwnd-limited then.
1724	 */
1725	if (!before(tp->snd_una, tp->max_packets_seq) ||
1726	    tp->packets_out > tp->max_packets_out) {
1727		tp->max_packets_out = tp->packets_out;
1728		tp->max_packets_seq = tp->snd_nxt;
1729		tp->is_cwnd_limited = is_cwnd_limited;
1730	}
1731
1732	if (tcp_is_cwnd_limited(sk)) {
1733		/* Network is feed fully. */
1734		tp->snd_cwnd_used = 0;
1735		tp->snd_cwnd_stamp = tcp_jiffies32;
1736	} else {
1737		/* Network starves. */
1738		if (tp->packets_out > tp->snd_cwnd_used)
1739			tp->snd_cwnd_used = tp->packets_out;
1740
1741		if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1742		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1743		    !ca_ops->cong_control)
1744			tcp_cwnd_application_limited(sk);
1745
1746		/* The following conditions together indicate the starvation
1747		 * is caused by insufficient sender buffer:
1748		 * 1) just sent some data (see tcp_write_xmit)
1749		 * 2) not cwnd limited (this else condition)
1750		 * 3) no more data to send (tcp_write_queue_empty())
1751		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1752		 */
1753		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1754		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1755		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1756			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1757	}
1758}
1759
1760/* Minshall's variant of the Nagle send check. */
1761static bool tcp_minshall_check(const struct tcp_sock *tp)
1762{
1763	return after(tp->snd_sml, tp->snd_una) &&
1764		!after(tp->snd_sml, tp->snd_nxt);
1765}
1766
1767/* Update snd_sml if this skb is under mss
1768 * Note that a TSO packet might end with a sub-mss segment
1769 * The test is really :
1770 * if ((skb->len % mss) != 0)
1771 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1772 * But we can avoid doing the divide again given we already have
1773 *  skb_pcount = skb->len / mss_now
1774 */
1775static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1776				const struct sk_buff *skb)
1777{
1778	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1779		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1780}
1781
1782/* Return false, if packet can be sent now without violation Nagle's rules:
1783 * 1. It is full sized. (provided by caller in %partial bool)
1784 * 2. Or it contains FIN. (already checked by caller)
1785 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1786 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1787 *    With Minshall's modification: all sent small packets are ACKed.
1788 */
1789static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1790			    int nonagle)
1791{
1792	return partial &&
1793		((nonagle & TCP_NAGLE_CORK) ||
1794		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1795}
1796
1797/* Return how many segs we'd like on a TSO packet,
1798 * to send one TSO packet per ms
1799 */
1800static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1801			    int min_tso_segs)
1802{
1803	u32 bytes, segs;
1804
1805	bytes = min_t(unsigned long,
1806		      sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1807		      sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1808
1809	/* Goal is to send at least one packet per ms,
1810	 * not one big TSO packet every 100 ms.
1811	 * This preserves ACK clocking and is consistent
1812	 * with tcp_tso_should_defer() heuristic.
1813	 */
1814	segs = max_t(u32, bytes / mss_now, min_tso_segs);
1815
1816	return segs;
1817}
1818
1819/* Return the number of segments we want in the skb we are transmitting.
1820 * See if congestion control module wants to decide; otherwise, autosize.
1821 */
1822static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1823{
1824	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1825	u32 min_tso, tso_segs;
1826
1827	min_tso = ca_ops->min_tso_segs ?
1828			ca_ops->min_tso_segs(sk) :
1829			sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1830
1831	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1832	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1833}
1834
1835/* Returns the portion of skb which can be sent right away */
1836static unsigned int tcp_mss_split_point(const struct sock *sk,
1837					const struct sk_buff *skb,
1838					unsigned int mss_now,
1839					unsigned int max_segs,
1840					int nonagle)
1841{
1842	const struct tcp_sock *tp = tcp_sk(sk);
1843	u32 partial, needed, window, max_len;
1844
1845	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1846	max_len = mss_now * max_segs;
1847
1848	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1849		return max_len;
1850
1851	needed = min(skb->len, window);
1852
1853	if (max_len <= needed)
1854		return max_len;
1855
1856	partial = needed % mss_now;
1857	/* If last segment is not a full MSS, check if Nagle rules allow us
1858	 * to include this last segment in this skb.
1859	 * Otherwise, we'll split the skb at last MSS boundary
1860	 */
1861	if (tcp_nagle_check(partial != 0, tp, nonagle))
1862		return needed - partial;
1863
1864	return needed;
1865}
1866
1867/* Can at least one segment of SKB be sent right now, according to the
1868 * congestion window rules?  If so, return how many segments are allowed.
1869 */
1870static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1871					 const struct sk_buff *skb)
1872{
1873	u32 in_flight, cwnd, halfcwnd;
1874
1875	/* Don't be strict about the congestion window for the final FIN.  */
1876	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1877	    tcp_skb_pcount(skb) == 1)
1878		return 1;
1879
1880	in_flight = tcp_packets_in_flight(tp);
1881	cwnd = tp->snd_cwnd;
1882	if (in_flight >= cwnd)
1883		return 0;
1884
1885	/* For better scheduling, ensure we have at least
1886	 * 2 GSO packets in flight.
1887	 */
1888	halfcwnd = max(cwnd >> 1, 1U);
1889	return min(halfcwnd, cwnd - in_flight);
1890}
1891
1892/* Initialize TSO state of a skb.
1893 * This must be invoked the first time we consider transmitting
1894 * SKB onto the wire.
1895 */
1896static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1897{
1898	int tso_segs = tcp_skb_pcount(skb);
1899
1900	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1901		tcp_set_skb_tso_segs(skb, mss_now);
1902		tso_segs = tcp_skb_pcount(skb);
1903	}
1904	return tso_segs;
1905}
1906
1907
1908/* Return true if the Nagle test allows this packet to be
1909 * sent now.
1910 */
1911static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1912				  unsigned int cur_mss, int nonagle)
1913{
1914	/* Nagle rule does not apply to frames, which sit in the middle of the
1915	 * write_queue (they have no chances to get new data).
1916	 *
1917	 * This is implemented in the callers, where they modify the 'nonagle'
1918	 * argument based upon the location of SKB in the send queue.
1919	 */
1920	if (nonagle & TCP_NAGLE_PUSH)
1921		return true;
1922
1923	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1924	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1925		return true;
1926
1927	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1928		return true;
1929
1930	return false;
1931}
1932
1933/* Does at least the first segment of SKB fit into the send window? */
1934static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1935			     const struct sk_buff *skb,
1936			     unsigned int cur_mss)
1937{
1938	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1939
1940	if (skb->len > cur_mss)
1941		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1942
1943	return !after(end_seq, tcp_wnd_end(tp));
1944}
1945
1946/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1947 * which is put after SKB on the list.  It is very much like
1948 * tcp_fragment() except that it may make several kinds of assumptions
1949 * in order to speed up the splitting operation.  In particular, we
1950 * know that all the data is in scatter-gather pages, and that the
1951 * packet has never been sent out before (and thus is not cloned).
1952 */
1953static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
 
1954			unsigned int mss_now, gfp_t gfp)
1955{
 
1956	int nlen = skb->len - len;
1957	struct sk_buff *buff;
1958	u8 flags;
1959
1960	/* All of a TSO frame must be composed of paged data.  */
1961	if (skb->len != skb->data_len)
1962		return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1963				    skb, len, mss_now, gfp);
1964
1965	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1966	if (unlikely(!buff))
1967		return -ENOMEM;
1968	skb_copy_decrypted(buff, skb);
1969
1970	sk_wmem_queued_add(sk, buff->truesize);
1971	sk_mem_charge(sk, buff->truesize);
1972	buff->truesize += nlen;
1973	skb->truesize -= nlen;
1974
1975	/* Correct the sequence numbers. */
1976	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1977	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1978	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1979
1980	/* PSH and FIN should only be set in the second packet. */
1981	flags = TCP_SKB_CB(skb)->tcp_flags;
1982	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1983	TCP_SKB_CB(buff)->tcp_flags = flags;
1984
1985	/* This packet was never sent out yet, so no SACK bits. */
1986	TCP_SKB_CB(buff)->sacked = 0;
1987
1988	tcp_skb_fragment_eor(skb, buff);
1989
1990	buff->ip_summed = CHECKSUM_PARTIAL;
1991	skb_split(skb, buff, len);
1992	tcp_fragment_tstamp(skb, buff);
1993
1994	/* Fix up tso_factor for both original and new SKB.  */
1995	tcp_set_skb_tso_segs(skb, mss_now);
1996	tcp_set_skb_tso_segs(buff, mss_now);
1997
1998	/* Link BUFF into the send queue. */
1999	__skb_header_release(buff);
2000	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2001
2002	return 0;
2003}
2004
2005/* Try to defer sending, if possible, in order to minimize the amount
2006 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
2007 *
2008 * This algorithm is from John Heffner.
2009 */
2010static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2011				 bool *is_cwnd_limited,
2012				 bool *is_rwnd_limited,
2013				 u32 max_segs)
2014{
2015	const struct inet_connection_sock *icsk = inet_csk(sk);
2016	u32 send_win, cong_win, limit, in_flight;
2017	struct tcp_sock *tp = tcp_sk(sk);
2018	struct sk_buff *head;
2019	int win_divisor;
2020	s64 delta;
 
 
2021
2022	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2023		goto send_now;
2024
2025	/* Avoid bursty behavior by allowing defer
2026	 * only if the last write was recent (1 ms).
2027	 * Note that tp->tcp_wstamp_ns can be in the future if we have
2028	 * packets waiting in a qdisc or device for EDT delivery.
2029	 */
2030	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2031	if (delta > 0)
2032		goto send_now;
2033
2034	in_flight = tcp_packets_in_flight(tp);
2035
2036	BUG_ON(tcp_skb_pcount(skb) <= 1);
2037	BUG_ON(tp->snd_cwnd <= in_flight);
2038
2039	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2040
2041	/* From in_flight test above, we know that cwnd > in_flight.  */
2042	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2043
2044	limit = min(send_win, cong_win);
2045
2046	/* If a full-sized TSO skb can be sent, do it. */
2047	if (limit >= max_segs * tp->mss_cache)
2048		goto send_now;
2049
2050	/* Middle in queue won't get any more data, full sendable already? */
2051	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2052		goto send_now;
2053
2054	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2055	if (win_divisor) {
2056		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2057
2058		/* If at least some fraction of a window is available,
2059		 * just use it.
2060		 */
2061		chunk /= win_divisor;
2062		if (limit >= chunk)
2063			goto send_now;
2064	} else {
2065		/* Different approach, try not to defer past a single
2066		 * ACK.  Receiver should ACK every other full sized
2067		 * frame, so if we have space for more than 3 frames
2068		 * then send now.
2069		 */
2070		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2071			goto send_now;
2072	}
2073
2074	/* TODO : use tsorted_sent_queue ? */
2075	head = tcp_rtx_queue_head(sk);
2076	if (!head)
2077		goto send_now;
2078	delta = tp->tcp_clock_cache - head->tstamp;
2079	/* If next ACK is likely to come too late (half srtt), do not defer */
2080	if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2081		goto send_now;
2082
2083	/* Ok, it looks like it is advisable to defer.
2084	 * Three cases are tracked :
2085	 * 1) We are cwnd-limited
2086	 * 2) We are rwnd-limited
2087	 * 3) We are application limited.
2088	 */
2089	if (cong_win < send_win) {
2090		if (cong_win <= skb->len) {
2091			*is_cwnd_limited = true;
2092			return true;
2093		}
2094	} else {
2095		if (send_win <= skb->len) {
2096			*is_rwnd_limited = true;
2097			return true;
2098		}
2099	}
2100
2101	/* If this packet won't get more data, do not wait. */
2102	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2103	    TCP_SKB_CB(skb)->eor)
2104		goto send_now;
2105
2106	return true;
2107
2108send_now:
2109	return false;
2110}
2111
2112static inline void tcp_mtu_check_reprobe(struct sock *sk)
2113{
2114	struct inet_connection_sock *icsk = inet_csk(sk);
2115	struct tcp_sock *tp = tcp_sk(sk);
2116	struct net *net = sock_net(sk);
2117	u32 interval;
2118	s32 delta;
2119
2120	interval = net->ipv4.sysctl_tcp_probe_interval;
2121	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2122	if (unlikely(delta >= interval * HZ)) {
2123		int mss = tcp_current_mss(sk);
2124
2125		/* Update current search range */
2126		icsk->icsk_mtup.probe_size = 0;
2127		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2128			sizeof(struct tcphdr) +
2129			icsk->icsk_af_ops->net_header_len;
2130		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2131
2132		/* Update probe time stamp */
2133		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2134	}
2135}
2136
2137static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2138{
2139	struct sk_buff *skb, *next;
2140
2141	skb = tcp_send_head(sk);
2142	tcp_for_write_queue_from_safe(skb, next, sk) {
2143		if (len <= skb->len)
2144			break;
2145
2146		if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2147			return false;
2148
2149		len -= skb->len;
2150	}
2151
2152	return true;
2153}
2154
2155/* Create a new MTU probe if we are ready.
2156 * MTU probe is regularly attempting to increase the path MTU by
2157 * deliberately sending larger packets.  This discovers routing
2158 * changes resulting in larger path MTUs.
2159 *
2160 * Returns 0 if we should wait to probe (no cwnd available),
2161 *         1 if a probe was sent,
2162 *         -1 otherwise
2163 */
2164static int tcp_mtu_probe(struct sock *sk)
2165{
2166	struct inet_connection_sock *icsk = inet_csk(sk);
2167	struct tcp_sock *tp = tcp_sk(sk);
2168	struct sk_buff *skb, *nskb, *next;
2169	struct net *net = sock_net(sk);
2170	int probe_size;
2171	int size_needed;
2172	int copy, len;
2173	int mss_now;
2174	int interval;
2175
2176	/* Not currently probing/verifying,
2177	 * not in recovery,
2178	 * have enough cwnd, and
2179	 * not SACKing (the variable headers throw things off)
2180	 */
2181	if (likely(!icsk->icsk_mtup.enabled ||
2182		   icsk->icsk_mtup.probe_size ||
2183		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2184		   tp->snd_cwnd < 11 ||
2185		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2186		return -1;
2187
2188	/* Use binary search for probe_size between tcp_mss_base,
2189	 * and current mss_clamp. if (search_high - search_low)
2190	 * smaller than a threshold, backoff from probing.
2191	 */
2192	mss_now = tcp_current_mss(sk);
2193	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2194				    icsk->icsk_mtup.search_low) >> 1);
2195	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2196	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2197	/* When misfortune happens, we are reprobing actively,
2198	 * and then reprobe timer has expired. We stick with current
2199	 * probing process by not resetting search range to its orignal.
2200	 */
2201	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2202		interval < net->ipv4.sysctl_tcp_probe_threshold) {
2203		/* Check whether enough time has elaplased for
2204		 * another round of probing.
2205		 */
2206		tcp_mtu_check_reprobe(sk);
2207		return -1;
2208	}
2209
2210	/* Have enough data in the send queue to probe? */
2211	if (tp->write_seq - tp->snd_nxt < size_needed)
2212		return -1;
2213
2214	if (tp->snd_wnd < size_needed)
2215		return -1;
2216	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2217		return 0;
2218
2219	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2220	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2221		if (!tcp_packets_in_flight(tp))
2222			return -1;
2223		else
2224			return 0;
2225	}
2226
2227	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2228		return -1;
2229
2230	/* We're allowed to probe.  Build it now. */
2231	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2232	if (!nskb)
2233		return -1;
2234	sk_wmem_queued_add(sk, nskb->truesize);
2235	sk_mem_charge(sk, nskb->truesize);
2236
2237	skb = tcp_send_head(sk);
2238	skb_copy_decrypted(nskb, skb);
2239
2240	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2241	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2242	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2243	TCP_SKB_CB(nskb)->sacked = 0;
2244	nskb->csum = 0;
2245	nskb->ip_summed = CHECKSUM_PARTIAL;
2246
2247	tcp_insert_write_queue_before(nskb, skb, sk);
2248	tcp_highest_sack_replace(sk, skb, nskb);
2249
2250	len = 0;
2251	tcp_for_write_queue_from_safe(skb, next, sk) {
2252		copy = min_t(int, skb->len, probe_size - len);
2253		skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2254
2255		if (skb->len <= copy) {
2256			/* We've eaten all the data from this skb.
2257			 * Throw it away. */
2258			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2259			/* If this is the last SKB we copy and eor is set
2260			 * we need to propagate it to the new skb.
2261			 */
2262			TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2263			tcp_skb_collapse_tstamp(nskb, skb);
2264			tcp_unlink_write_queue(skb, sk);
2265			sk_wmem_free_skb(sk, skb);
2266		} else {
2267			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2268						   ~(TCPHDR_FIN|TCPHDR_PSH);
2269			if (!skb_shinfo(skb)->nr_frags) {
2270				skb_pull(skb, copy);
2271			} else {
2272				__pskb_trim_head(skb, copy);
2273				tcp_set_skb_tso_segs(skb, mss_now);
2274			}
2275			TCP_SKB_CB(skb)->seq += copy;
2276		}
2277
2278		len += copy;
2279
2280		if (len >= probe_size)
2281			break;
2282	}
2283	tcp_init_tso_segs(nskb, nskb->len);
2284
2285	/* We're ready to send.  If this fails, the probe will
2286	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2287	 */
2288	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2289		/* Decrement cwnd here because we are sending
2290		 * effectively two packets. */
2291		tp->snd_cwnd--;
2292		tcp_event_new_data_sent(sk, nskb);
2293
2294		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2295		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2296		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2297
2298		return 1;
2299	}
2300
2301	return -1;
2302}
2303
2304static bool tcp_pacing_check(struct sock *sk)
2305{
2306	struct tcp_sock *tp = tcp_sk(sk);
2307
2308	if (!tcp_needs_internal_pacing(sk))
2309		return false;
2310
2311	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2312		return false;
2313
2314	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2315		hrtimer_start(&tp->pacing_timer,
2316			      ns_to_ktime(tp->tcp_wstamp_ns),
2317			      HRTIMER_MODE_ABS_PINNED_SOFT);
2318		sock_hold(sk);
2319	}
2320	return true;
2321}
2322
2323/* TCP Small Queues :
2324 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2325 * (These limits are doubled for retransmits)
2326 * This allows for :
2327 *  - better RTT estimation and ACK scheduling
2328 *  - faster recovery
2329 *  - high rates
2330 * Alas, some drivers / subsystems require a fair amount
2331 * of queued bytes to ensure line rate.
2332 * One example is wifi aggregation (802.11 AMPDU)
2333 */
2334static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2335				  unsigned int factor)
2336{
2337	unsigned long limit;
2338
2339	limit = max_t(unsigned long,
2340		      2 * skb->truesize,
2341		      sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2342	if (sk->sk_pacing_status == SK_PACING_NONE)
2343		limit = min_t(unsigned long, limit,
2344			      sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2345	limit <<= factor;
2346
2347	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2348	    tcp_sk(sk)->tcp_tx_delay) {
2349		u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2350
2351		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2352		 * approximate our needs assuming an ~100% skb->truesize overhead.
2353		 * USEC_PER_SEC is approximated by 2^20.
2354		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2355		 */
2356		extra_bytes >>= (20 - 1);
2357		limit += extra_bytes;
2358	}
2359	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2360		/* Always send skb if rtx queue is empty.
2361		 * No need to wait for TX completion to call us back,
2362		 * after softirq/tasklet schedule.
2363		 * This helps when TX completions are delayed too much.
2364		 */
2365		if (tcp_rtx_queue_empty(sk))
2366			return false;
2367
2368		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2369		/* It is possible TX completion already happened
2370		 * before we set TSQ_THROTTLED, so we must
2371		 * test again the condition.
2372		 */
2373		smp_mb__after_atomic();
2374		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2375			return true;
2376	}
2377	return false;
2378}
2379
2380static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2381{
2382	const u32 now = tcp_jiffies32;
2383	enum tcp_chrono old = tp->chrono_type;
2384
2385	if (old > TCP_CHRONO_UNSPEC)
2386		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2387	tp->chrono_start = now;
2388	tp->chrono_type = new;
2389}
2390
2391void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2392{
2393	struct tcp_sock *tp = tcp_sk(sk);
2394
2395	/* If there are multiple conditions worthy of tracking in a
2396	 * chronograph then the highest priority enum takes precedence
2397	 * over the other conditions. So that if something "more interesting"
2398	 * starts happening, stop the previous chrono and start a new one.
2399	 */
2400	if (type > tp->chrono_type)
2401		tcp_chrono_set(tp, type);
2402}
2403
2404void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2405{
2406	struct tcp_sock *tp = tcp_sk(sk);
2407
2408
2409	/* There are multiple conditions worthy of tracking in a
2410	 * chronograph, so that the highest priority enum takes
2411	 * precedence over the other conditions (see tcp_chrono_start).
2412	 * If a condition stops, we only stop chrono tracking if
2413	 * it's the "most interesting" or current chrono we are
2414	 * tracking and starts busy chrono if we have pending data.
2415	 */
2416	if (tcp_rtx_and_write_queues_empty(sk))
2417		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2418	else if (type == tp->chrono_type)
2419		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2420}
2421
2422/* This routine writes packets to the network.  It advances the
2423 * send_head.  This happens as incoming acks open up the remote
2424 * window for us.
2425 *
2426 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2427 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2428 * account rare use of URG, this is not a big flaw.
2429 *
2430 * Send at most one packet when push_one > 0. Temporarily ignore
2431 * cwnd limit to force at most one packet out when push_one == 2.
2432
2433 * Returns true, if no segments are in flight and we have queued segments,
2434 * but cannot send anything now because of SWS or another problem.
2435 */
2436static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2437			   int push_one, gfp_t gfp)
2438{
2439	struct tcp_sock *tp = tcp_sk(sk);
2440	struct sk_buff *skb;
2441	unsigned int tso_segs, sent_pkts;
2442	int cwnd_quota;
2443	int result;
2444	bool is_cwnd_limited = false, is_rwnd_limited = false;
2445	u32 max_segs;
2446
2447	sent_pkts = 0;
2448
2449	tcp_mstamp_refresh(tp);
2450	if (!push_one) {
2451		/* Do MTU probing. */
2452		result = tcp_mtu_probe(sk);
2453		if (!result) {
2454			return false;
2455		} else if (result > 0) {
2456			sent_pkts = 1;
2457		}
2458	}
2459
2460	max_segs = tcp_tso_segs(sk, mss_now);
2461	while ((skb = tcp_send_head(sk))) {
2462		unsigned int limit;
2463
2464		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2465			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2466			skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2467			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2468			tcp_init_tso_segs(skb, mss_now);
2469			goto repair; /* Skip network transmission */
2470		}
2471
2472		if (tcp_pacing_check(sk))
2473			break;
2474
2475		tso_segs = tcp_init_tso_segs(skb, mss_now);
2476		BUG_ON(!tso_segs);
2477
 
 
 
 
 
 
2478		cwnd_quota = tcp_cwnd_test(tp, skb);
2479		if (!cwnd_quota) {
2480			if (push_one == 2)
2481				/* Force out a loss probe pkt. */
2482				cwnd_quota = 1;
2483			else
2484				break;
2485		}
2486
2487		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2488			is_rwnd_limited = true;
2489			break;
2490		}
2491
2492		if (tso_segs == 1) {
2493			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2494						     (tcp_skb_is_last(sk, skb) ?
2495						      nonagle : TCP_NAGLE_PUSH))))
2496				break;
2497		} else {
2498			if (!push_one &&
2499			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2500						 &is_rwnd_limited, max_segs))
2501				break;
2502		}
2503
2504		limit = mss_now;
2505		if (tso_segs > 1 && !tcp_urg_mode(tp))
2506			limit = tcp_mss_split_point(sk, skb, mss_now,
2507						    min_t(unsigned int,
2508							  cwnd_quota,
2509							  max_segs),
2510						    nonagle);
2511
2512		if (skb->len > limit &&
2513		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
 
2514			break;
2515
 
 
2516		if (tcp_small_queue_check(sk, skb, 0))
2517			break;
2518
2519		/* Argh, we hit an empty skb(), presumably a thread
2520		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2521		 * We do not want to send a pure-ack packet and have
2522		 * a strange looking rtx queue with empty packet(s).
2523		 */
2524		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2525			break;
2526
2527		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2528			break;
2529
2530repair:
2531		/* Advance the send_head.  This one is sent out.
2532		 * This call will increment packets_out.
2533		 */
2534		tcp_event_new_data_sent(sk, skb);
2535
2536		tcp_minshall_update(tp, mss_now, skb);
2537		sent_pkts += tcp_skb_pcount(skb);
2538
2539		if (push_one)
2540			break;
2541	}
2542
2543	if (is_rwnd_limited)
2544		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2545	else
2546		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2547
2548	if (likely(sent_pkts)) {
2549		if (tcp_in_cwnd_reduction(sk))
2550			tp->prr_out += sent_pkts;
2551
2552		/* Send one loss probe per tail loss episode. */
2553		if (push_one != 2)
2554			tcp_schedule_loss_probe(sk, false);
2555		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2556		tcp_cwnd_validate(sk, is_cwnd_limited);
2557		return false;
2558	}
2559	return !tp->packets_out && !tcp_write_queue_empty(sk);
2560}
2561
2562bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2563{
2564	struct inet_connection_sock *icsk = inet_csk(sk);
2565	struct tcp_sock *tp = tcp_sk(sk);
2566	u32 timeout, rto_delta_us;
2567	int early_retrans;
2568
2569	/* Don't do any loss probe on a Fast Open connection before 3WHS
2570	 * finishes.
2571	 */
2572	if (rcu_access_pointer(tp->fastopen_rsk))
2573		return false;
2574
2575	early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2576	/* Schedule a loss probe in 2*RTT for SACK capable connections
2577	 * not in loss recovery, that are either limited by cwnd or application.
2578	 */
2579	if ((early_retrans != 3 && early_retrans != 4) ||
2580	    !tp->packets_out || !tcp_is_sack(tp) ||
2581	    (icsk->icsk_ca_state != TCP_CA_Open &&
2582	     icsk->icsk_ca_state != TCP_CA_CWR))
2583		return false;
2584
2585	/* Probe timeout is 2*rtt. Add minimum RTO to account
2586	 * for delayed ack when there's one outstanding packet. If no RTT
2587	 * sample is available then probe after TCP_TIMEOUT_INIT.
2588	 */
2589	if (tp->srtt_us) {
2590		timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2591		if (tp->packets_out == 1)
2592			timeout += TCP_RTO_MIN;
2593		else
2594			timeout += TCP_TIMEOUT_MIN;
2595	} else {
2596		timeout = TCP_TIMEOUT_INIT;
2597	}
2598
2599	/* If the RTO formula yields an earlier time, then use that time. */
2600	rto_delta_us = advancing_rto ?
2601			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2602			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2603	if (rto_delta_us > 0)
2604		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2605
2606	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
 
2607	return true;
2608}
2609
2610/* Thanks to skb fast clones, we can detect if a prior transmit of
2611 * a packet is still in a qdisc or driver queue.
2612 * In this case, there is very little point doing a retransmit !
2613 */
2614static bool skb_still_in_host_queue(const struct sock *sk,
2615				    const struct sk_buff *skb)
2616{
2617	if (unlikely(skb_fclone_busy(sk, skb))) {
2618		NET_INC_STATS(sock_net(sk),
2619			      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2620		return true;
2621	}
2622	return false;
2623}
2624
2625/* When probe timeout (PTO) fires, try send a new segment if possible, else
2626 * retransmit the last segment.
2627 */
2628void tcp_send_loss_probe(struct sock *sk)
2629{
2630	struct tcp_sock *tp = tcp_sk(sk);
2631	struct sk_buff *skb;
2632	int pcount;
2633	int mss = tcp_current_mss(sk);
2634
2635	/* At most one outstanding TLP */
2636	if (tp->tlp_high_seq)
2637		goto rearm_timer;
2638
2639	tp->tlp_retrans = 0;
2640	skb = tcp_send_head(sk);
2641	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2642		pcount = tp->packets_out;
2643		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2644		if (tp->packets_out > pcount)
2645			goto probe_sent;
2646		goto rearm_timer;
2647	}
2648	skb = skb_rb_last(&sk->tcp_rtx_queue);
2649	if (unlikely(!skb)) {
2650		WARN_ONCE(tp->packets_out,
2651			  "invalid inflight: %u state %u cwnd %u mss %d\n",
2652			  tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2653		inet_csk(sk)->icsk_pending = 0;
2654		return;
2655	}
 
2656
2657	if (skb_still_in_host_queue(sk, skb))
2658		goto rearm_timer;
2659
2660	pcount = tcp_skb_pcount(skb);
2661	if (WARN_ON(!pcount))
2662		goto rearm_timer;
2663
2664	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2665		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2666					  (pcount - 1) * mss, mss,
2667					  GFP_ATOMIC)))
2668			goto rearm_timer;
2669		skb = skb_rb_next(skb);
2670	}
2671
2672	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2673		goto rearm_timer;
2674
2675	if (__tcp_retransmit_skb(sk, skb, 1))
2676		goto rearm_timer;
2677
2678	tp->tlp_retrans = 1;
2679
2680probe_sent:
2681	/* Record snd_nxt for loss detection. */
2682	tp->tlp_high_seq = tp->snd_nxt;
2683
 
2684	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2685	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2686	inet_csk(sk)->icsk_pending = 0;
2687rearm_timer:
2688	tcp_rearm_rto(sk);
2689}
2690
2691/* Push out any pending frames which were held back due to
2692 * TCP_CORK or attempt at coalescing tiny packets.
2693 * The socket must be locked by the caller.
2694 */
2695void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2696			       int nonagle)
2697{
2698	/* If we are closed, the bytes will have to remain here.
2699	 * In time closedown will finish, we empty the write queue and
2700	 * all will be happy.
2701	 */
2702	if (unlikely(sk->sk_state == TCP_CLOSE))
2703		return;
2704
2705	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2706			   sk_gfp_mask(sk, GFP_ATOMIC)))
2707		tcp_check_probe_timer(sk);
2708}
2709
2710/* Send _single_ skb sitting at the send head. This function requires
2711 * true push pending frames to setup probe timer etc.
2712 */
2713void tcp_push_one(struct sock *sk, unsigned int mss_now)
2714{
2715	struct sk_buff *skb = tcp_send_head(sk);
2716
2717	BUG_ON(!skb || skb->len < mss_now);
2718
2719	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2720}
2721
2722/* This function returns the amount that we can raise the
2723 * usable window based on the following constraints
2724 *
2725 * 1. The window can never be shrunk once it is offered (RFC 793)
2726 * 2. We limit memory per socket
2727 *
2728 * RFC 1122:
2729 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2730 *  RECV.NEXT + RCV.WIN fixed until:
2731 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2732 *
2733 * i.e. don't raise the right edge of the window until you can raise
2734 * it at least MSS bytes.
2735 *
2736 * Unfortunately, the recommended algorithm breaks header prediction,
2737 * since header prediction assumes th->window stays fixed.
2738 *
2739 * Strictly speaking, keeping th->window fixed violates the receiver
2740 * side SWS prevention criteria. The problem is that under this rule
2741 * a stream of single byte packets will cause the right side of the
2742 * window to always advance by a single byte.
2743 *
2744 * Of course, if the sender implements sender side SWS prevention
2745 * then this will not be a problem.
2746 *
2747 * BSD seems to make the following compromise:
2748 *
2749 *	If the free space is less than the 1/4 of the maximum
2750 *	space available and the free space is less than 1/2 mss,
2751 *	then set the window to 0.
2752 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2753 *	Otherwise, just prevent the window from shrinking
2754 *	and from being larger than the largest representable value.
2755 *
2756 * This prevents incremental opening of the window in the regime
2757 * where TCP is limited by the speed of the reader side taking
2758 * data out of the TCP receive queue. It does nothing about
2759 * those cases where the window is constrained on the sender side
2760 * because the pipeline is full.
2761 *
2762 * BSD also seems to "accidentally" limit itself to windows that are a
2763 * multiple of MSS, at least until the free space gets quite small.
2764 * This would appear to be a side effect of the mbuf implementation.
2765 * Combining these two algorithms results in the observed behavior
2766 * of having a fixed window size at almost all times.
2767 *
2768 * Below we obtain similar behavior by forcing the offered window to
2769 * a multiple of the mss when it is feasible to do so.
2770 *
2771 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2772 * Regular options like TIMESTAMP are taken into account.
2773 */
2774u32 __tcp_select_window(struct sock *sk)
2775{
2776	struct inet_connection_sock *icsk = inet_csk(sk);
2777	struct tcp_sock *tp = tcp_sk(sk);
2778	/* MSS for the peer's data.  Previous versions used mss_clamp
2779	 * here.  I don't know if the value based on our guesses
2780	 * of peer's MSS is better for the performance.  It's more correct
2781	 * but may be worse for the performance because of rcv_mss
2782	 * fluctuations.  --SAW  1998/11/1
2783	 */
2784	int mss = icsk->icsk_ack.rcv_mss;
2785	int free_space = tcp_space(sk);
2786	int allowed_space = tcp_full_space(sk);
2787	int full_space, window;
2788
2789	if (sk_is_mptcp(sk))
2790		mptcp_space(sk, &free_space, &allowed_space);
2791
2792	full_space = min_t(int, tp->window_clamp, allowed_space);
2793
2794	if (unlikely(mss > full_space)) {
2795		mss = full_space;
2796		if (mss <= 0)
2797			return 0;
2798	}
2799	if (free_space < (full_space >> 1)) {
2800		icsk->icsk_ack.quick = 0;
2801
2802		if (tcp_under_memory_pressure(sk))
2803			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2804					       4U * tp->advmss);
2805
2806		/* free_space might become our new window, make sure we don't
2807		 * increase it due to wscale.
2808		 */
2809		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2810
2811		/* if free space is less than mss estimate, or is below 1/16th
2812		 * of the maximum allowed, try to move to zero-window, else
2813		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2814		 * new incoming data is dropped due to memory limits.
2815		 * With large window, mss test triggers way too late in order
2816		 * to announce zero window in time before rmem limit kicks in.
2817		 */
2818		if (free_space < (allowed_space >> 4) || free_space < mss)
2819			return 0;
2820	}
2821
2822	if (free_space > tp->rcv_ssthresh)
2823		free_space = tp->rcv_ssthresh;
2824
2825	/* Don't do rounding if we are using window scaling, since the
2826	 * scaled window will not line up with the MSS boundary anyway.
2827	 */
2828	if (tp->rx_opt.rcv_wscale) {
2829		window = free_space;
2830
2831		/* Advertise enough space so that it won't get scaled away.
2832		 * Import case: prevent zero window announcement if
2833		 * 1<<rcv_wscale > mss.
2834		 */
2835		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2836	} else {
2837		window = tp->rcv_wnd;
2838		/* Get the largest window that is a nice multiple of mss.
2839		 * Window clamp already applied above.
2840		 * If our current window offering is within 1 mss of the
2841		 * free space we just keep it. This prevents the divide
2842		 * and multiply from happening most of the time.
2843		 * We also don't do any window rounding when the free space
2844		 * is too small.
2845		 */
2846		if (window <= free_space - mss || window > free_space)
2847			window = rounddown(free_space, mss);
2848		else if (mss == full_space &&
2849			 free_space > window + (full_space >> 1))
2850			window = free_space;
2851	}
2852
2853	return window;
2854}
2855
2856void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2857			     const struct sk_buff *next_skb)
2858{
2859	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2860		const struct skb_shared_info *next_shinfo =
2861			skb_shinfo(next_skb);
2862		struct skb_shared_info *shinfo = skb_shinfo(skb);
2863
2864		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2865		shinfo->tskey = next_shinfo->tskey;
2866		TCP_SKB_CB(skb)->txstamp_ack |=
2867			TCP_SKB_CB(next_skb)->txstamp_ack;
2868	}
2869}
2870
2871/* Collapses two adjacent SKB's during retransmission. */
2872static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2873{
2874	struct tcp_sock *tp = tcp_sk(sk);
2875	struct sk_buff *next_skb = skb_rb_next(skb);
2876	int next_skb_size;
2877
 
2878	next_skb_size = next_skb->len;
2879
2880	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2881
2882	if (next_skb_size) {
2883		if (next_skb_size <= skb_availroom(skb))
2884			skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2885				      next_skb_size);
2886		else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2887			return false;
2888	}
2889	tcp_highest_sack_replace(sk, next_skb, skb);
2890
2891	/* Update sequence range on original skb. */
2892	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2893
2894	/* Merge over control information. This moves PSH/FIN etc. over */
2895	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2896
2897	/* All done, get rid of second SKB and account for it so
2898	 * packet counting does not break.
2899	 */
2900	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2901	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2902
2903	/* changed transmit queue under us so clear hints */
2904	tcp_clear_retrans_hints_partial(tp);
2905	if (next_skb == tp->retransmit_skb_hint)
2906		tp->retransmit_skb_hint = skb;
2907
2908	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2909
2910	tcp_skb_collapse_tstamp(skb, next_skb);
2911
2912	tcp_rtx_queue_unlink_and_free(next_skb, sk);
2913	return true;
2914}
2915
2916/* Check if coalescing SKBs is legal. */
2917static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2918{
2919	if (tcp_skb_pcount(skb) > 1)
2920		return false;
2921	if (skb_cloned(skb))
2922		return false;
2923	/* Some heuristics for collapsing over SACK'd could be invented */
2924	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2925		return false;
2926
2927	return true;
2928}
2929
2930/* Collapse packets in the retransmit queue to make to create
2931 * less packets on the wire. This is only done on retransmission.
2932 */
2933static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2934				     int space)
2935{
2936	struct tcp_sock *tp = tcp_sk(sk);
2937	struct sk_buff *skb = to, *tmp;
2938	bool first = true;
2939
2940	if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2941		return;
2942	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2943		return;
2944
2945	skb_rbtree_walk_from_safe(skb, tmp) {
2946		if (!tcp_can_collapse(sk, skb))
2947			break;
2948
2949		if (!tcp_skb_can_collapse(to, skb))
2950			break;
2951
2952		space -= skb->len;
2953
2954		if (first) {
2955			first = false;
2956			continue;
2957		}
2958
2959		if (space < 0)
2960			break;
2961
2962		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2963			break;
2964
2965		if (!tcp_collapse_retrans(sk, to))
2966			break;
2967	}
2968}
2969
2970/* This retransmits one SKB.  Policy decisions and retransmit queue
2971 * state updates are done by the caller.  Returns non-zero if an
2972 * error occurred which prevented the send.
2973 */
2974int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2975{
2976	struct inet_connection_sock *icsk = inet_csk(sk);
2977	struct tcp_sock *tp = tcp_sk(sk);
2978	unsigned int cur_mss;
2979	int diff, len, err;
2980
2981
2982	/* Inconclusive MTU probe */
2983	if (icsk->icsk_mtup.probe_size)
2984		icsk->icsk_mtup.probe_size = 0;
2985
2986	/* Do not sent more than we queued. 1/4 is reserved for possible
2987	 * copying overhead: fragmentation, tunneling, mangling etc.
2988	 */
2989	if (refcount_read(&sk->sk_wmem_alloc) >
2990	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2991		  sk->sk_sndbuf))
2992		return -EAGAIN;
2993
2994	if (skb_still_in_host_queue(sk, skb))
2995		return -EBUSY;
2996
2997	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2998		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2999			WARN_ON_ONCE(1);
3000			return -EINVAL;
3001		}
3002		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3003			return -ENOMEM;
3004	}
3005
3006	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3007		return -EHOSTUNREACH; /* Routing failure or similar. */
3008
3009	cur_mss = tcp_current_mss(sk);
3010
3011	/* If receiver has shrunk his window, and skb is out of
3012	 * new window, do not retransmit it. The exception is the
3013	 * case, when window is shrunk to zero. In this case
3014	 * our retransmit serves as a zero window probe.
3015	 */
3016	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3017	    TCP_SKB_CB(skb)->seq != tp->snd_una)
3018		return -EAGAIN;
3019
3020	len = cur_mss * segs;
3021	if (skb->len > len) {
3022		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3023				 cur_mss, GFP_ATOMIC))
3024			return -ENOMEM; /* We'll try again later. */
3025	} else {
3026		if (skb_unclone(skb, GFP_ATOMIC))
3027			return -ENOMEM;
3028
3029		diff = tcp_skb_pcount(skb);
3030		tcp_set_skb_tso_segs(skb, cur_mss);
3031		diff -= tcp_skb_pcount(skb);
3032		if (diff)
3033			tcp_adjust_pcount(sk, skb, diff);
3034		if (skb->len < cur_mss)
3035			tcp_retrans_try_collapse(sk, skb, cur_mss);
3036	}
3037
3038	/* RFC3168, section 6.1.1.1. ECN fallback */
3039	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3040		tcp_ecn_clear_syn(sk, skb);
3041
3042	/* Update global and local TCP statistics. */
3043	segs = tcp_skb_pcount(skb);
3044	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3045	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3046		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3047	tp->total_retrans += segs;
3048	tp->bytes_retrans += skb->len;
3049
3050	/* make sure skb->data is aligned on arches that require it
3051	 * and check if ack-trimming & collapsing extended the headroom
3052	 * beyond what csum_start can cover.
3053	 */
3054	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3055		     skb_headroom(skb) >= 0xFFFF)) {
3056		struct sk_buff *nskb;
3057
3058		tcp_skb_tsorted_save(skb) {
3059			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3060			if (nskb) {
3061				nskb->dev = NULL;
3062				err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3063			} else {
3064				err = -ENOBUFS;
3065			}
3066		} tcp_skb_tsorted_restore(skb);
3067
3068		if (!err) {
3069			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3070			tcp_rate_skb_sent(sk, skb);
3071		}
3072	} else {
3073		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3074	}
3075
3076	/* To avoid taking spuriously low RTT samples based on a timestamp
3077	 * for a transmit that never happened, always mark EVER_RETRANS
3078	 */
3079	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3080
3081	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3082		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3083				  TCP_SKB_CB(skb)->seq, segs, err);
3084
3085	if (likely(!err)) {
 
3086		trace_tcp_retransmit_skb(sk, skb);
3087	} else if (err != -EBUSY) {
3088		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3089	}
3090	return err;
3091}
3092
3093int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3094{
3095	struct tcp_sock *tp = tcp_sk(sk);
3096	int err = __tcp_retransmit_skb(sk, skb, segs);
3097
3098	if (err == 0) {
3099#if FASTRETRANS_DEBUG > 0
3100		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3101			net_dbg_ratelimited("retrans_out leaked\n");
3102		}
3103#endif
3104		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3105		tp->retrans_out += tcp_skb_pcount(skb);
 
 
 
 
 
3106	}
3107
3108	/* Save stamp of the first (attempted) retransmit. */
3109	if (!tp->retrans_stamp)
3110		tp->retrans_stamp = tcp_skb_timestamp(skb);
3111
3112	if (tp->undo_retrans < 0)
3113		tp->undo_retrans = 0;
3114	tp->undo_retrans += tcp_skb_pcount(skb);
3115	return err;
3116}
3117
3118/* This gets called after a retransmit timeout, and the initially
3119 * retransmitted data is acknowledged.  It tries to continue
3120 * resending the rest of the retransmit queue, until either
3121 * we've sent it all or the congestion window limit is reached.
3122 */
3123void tcp_xmit_retransmit_queue(struct sock *sk)
3124{
3125	const struct inet_connection_sock *icsk = inet_csk(sk);
3126	struct sk_buff *skb, *rtx_head, *hole = NULL;
3127	struct tcp_sock *tp = tcp_sk(sk);
3128	bool rearm_timer = false;
3129	u32 max_segs;
3130	int mib_idx;
3131
3132	if (!tp->packets_out)
3133		return;
3134
3135	rtx_head = tcp_rtx_queue_head(sk);
3136	skb = tp->retransmit_skb_hint ?: rtx_head;
3137	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3138	skb_rbtree_walk_from(skb) {
3139		__u8 sacked;
3140		int segs;
3141
3142		if (tcp_pacing_check(sk))
3143			break;
3144
3145		/* we could do better than to assign each time */
3146		if (!hole)
3147			tp->retransmit_skb_hint = skb;
3148
3149		segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3150		if (segs <= 0)
3151			break;
3152		sacked = TCP_SKB_CB(skb)->sacked;
3153		/* In case tcp_shift_skb_data() have aggregated large skbs,
3154		 * we need to make sure not sending too bigs TSO packets
3155		 */
3156		segs = min_t(int, segs, max_segs);
3157
3158		if (tp->retrans_out >= tp->lost_out) {
3159			break;
3160		} else if (!(sacked & TCPCB_LOST)) {
3161			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3162				hole = skb;
3163			continue;
3164
3165		} else {
3166			if (icsk->icsk_ca_state != TCP_CA_Loss)
3167				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3168			else
3169				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3170		}
3171
3172		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3173			continue;
3174
3175		if (tcp_small_queue_check(sk, skb, 1))
3176			break;
3177
3178		if (tcp_retransmit_skb(sk, skb, segs))
3179			break;
3180
3181		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3182
3183		if (tcp_in_cwnd_reduction(sk))
3184			tp->prr_out += tcp_skb_pcount(skb);
3185
3186		if (skb == rtx_head &&
3187		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3188			rearm_timer = true;
3189
 
3190	}
3191	if (rearm_timer)
3192		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3193				     inet_csk(sk)->icsk_rto,
3194				     TCP_RTO_MAX);
3195}
3196
3197/* We allow to exceed memory limits for FIN packets to expedite
3198 * connection tear down and (memory) recovery.
3199 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3200 * or even be forced to close flow without any FIN.
3201 * In general, we want to allow one skb per socket to avoid hangs
3202 * with edge trigger epoll()
3203 */
3204void sk_forced_mem_schedule(struct sock *sk, int size)
3205{
3206	int amt;
3207
3208	if (size <= sk->sk_forward_alloc)
3209		return;
3210	amt = sk_mem_pages(size);
3211	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3212	sk_memory_allocated_add(sk, amt);
3213
3214	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3215		mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3216}
3217
3218/* Send a FIN. The caller locks the socket for us.
3219 * We should try to send a FIN packet really hard, but eventually give up.
3220 */
3221void tcp_send_fin(struct sock *sk)
3222{
3223	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3224	struct tcp_sock *tp = tcp_sk(sk);
3225
3226	/* Optimization, tack on the FIN if we have one skb in write queue and
3227	 * this skb was not yet sent, or we are under memory pressure.
3228	 * Note: in the latter case, FIN packet will be sent after a timeout,
3229	 * as TCP stack thinks it has already been transmitted.
3230	 */
3231	tskb = tail;
3232	if (!tskb && tcp_under_memory_pressure(sk))
3233		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3234
3235	if (tskb) {
 
3236		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3237		TCP_SKB_CB(tskb)->end_seq++;
3238		tp->write_seq++;
3239		if (!tail) {
3240			/* This means tskb was already sent.
3241			 * Pretend we included the FIN on previous transmit.
3242			 * We need to set tp->snd_nxt to the value it would have
3243			 * if FIN had been sent. This is because retransmit path
3244			 * does not change tp->snd_nxt.
3245			 */
3246			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3247			return;
3248		}
3249	} else {
3250		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3251		if (unlikely(!skb))
 
 
3252			return;
3253
3254		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3255		skb_reserve(skb, MAX_TCP_HEADER);
3256		sk_forced_mem_schedule(sk, skb->truesize);
3257		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3258		tcp_init_nondata_skb(skb, tp->write_seq,
3259				     TCPHDR_ACK | TCPHDR_FIN);
3260		tcp_queue_skb(sk, skb);
3261	}
3262	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3263}
3264
3265/* We get here when a process closes a file descriptor (either due to
3266 * an explicit close() or as a byproduct of exit()'ing) and there
3267 * was unread data in the receive queue.  This behavior is recommended
3268 * by RFC 2525, section 2.17.  -DaveM
3269 */
3270void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3271{
3272	struct sk_buff *skb;
3273
3274	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3275
3276	/* NOTE: No TCP options attached and we never retransmit this. */
3277	skb = alloc_skb(MAX_TCP_HEADER, priority);
3278	if (!skb) {
3279		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3280		return;
3281	}
3282
3283	/* Reserve space for headers and prepare control bits. */
3284	skb_reserve(skb, MAX_TCP_HEADER);
3285	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3286			     TCPHDR_ACK | TCPHDR_RST);
3287	tcp_mstamp_refresh(tcp_sk(sk));
3288	/* Send it off. */
3289	if (tcp_transmit_skb(sk, skb, 0, priority))
3290		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3291
3292	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3293	 * skb here is different to the troublesome skb, so use NULL
3294	 */
3295	trace_tcp_send_reset(sk, NULL);
3296}
3297
3298/* Send a crossed SYN-ACK during socket establishment.
3299 * WARNING: This routine must only be called when we have already sent
3300 * a SYN packet that crossed the incoming SYN that caused this routine
3301 * to get called. If this assumption fails then the initial rcv_wnd
3302 * and rcv_wscale values will not be correct.
3303 */
3304int tcp_send_synack(struct sock *sk)
3305{
3306	struct sk_buff *skb;
3307
3308	skb = tcp_rtx_queue_head(sk);
3309	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3310		pr_err("%s: wrong queue state\n", __func__);
3311		return -EFAULT;
3312	}
3313	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3314		if (skb_cloned(skb)) {
3315			struct sk_buff *nskb;
3316
3317			tcp_skb_tsorted_save(skb) {
3318				nskb = skb_copy(skb, GFP_ATOMIC);
3319			} tcp_skb_tsorted_restore(skb);
3320			if (!nskb)
3321				return -ENOMEM;
3322			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3323			tcp_highest_sack_replace(sk, skb, nskb);
3324			tcp_rtx_queue_unlink_and_free(skb, sk);
3325			__skb_header_release(nskb);
3326			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3327			sk_wmem_queued_add(sk, nskb->truesize);
3328			sk_mem_charge(sk, nskb->truesize);
3329			skb = nskb;
3330		}
3331
3332		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3333		tcp_ecn_send_synack(sk, skb);
3334	}
3335	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3336}
3337
3338/**
3339 * tcp_make_synack - Prepare a SYN-ACK.
3340 * sk: listener socket
3341 * dst: dst entry attached to the SYNACK
3342 * req: request_sock pointer
3343 * foc: cookie for tcp fast open
3344 * synack_type: Type of synback to prepare
3345 *
3346 * Allocate one skb and build a SYNACK packet.
3347 * @dst is consumed : Caller should not use it again.
3348 */
3349struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3350				struct request_sock *req,
3351				struct tcp_fastopen_cookie *foc,
3352				enum tcp_synack_type synack_type)
3353{
3354	struct inet_request_sock *ireq = inet_rsk(req);
3355	const struct tcp_sock *tp = tcp_sk(sk);
3356	struct tcp_md5sig_key *md5 = NULL;
3357	struct tcp_out_options opts;
3358	struct sk_buff *skb;
3359	int tcp_header_size;
3360	struct tcphdr *th;
3361	int mss;
3362	u64 now;
3363
3364	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3365	if (unlikely(!skb)) {
3366		dst_release(dst);
3367		return NULL;
3368	}
3369	/* Reserve space for headers. */
3370	skb_reserve(skb, MAX_TCP_HEADER);
3371
3372	switch (synack_type) {
3373	case TCP_SYNACK_NORMAL:
3374		skb_set_owner_w(skb, req_to_sk(req));
3375		break;
3376	case TCP_SYNACK_COOKIE:
3377		/* Under synflood, we do not attach skb to a socket,
3378		 * to avoid false sharing.
3379		 */
3380		break;
3381	case TCP_SYNACK_FASTOPEN:
3382		/* sk is a const pointer, because we want to express multiple
3383		 * cpu might call us concurrently.
3384		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3385		 */
3386		skb_set_owner_w(skb, (struct sock *)sk);
3387		break;
3388	}
3389	skb_dst_set(skb, dst);
3390
3391	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3392
3393	memset(&opts, 0, sizeof(opts));
3394	now = tcp_clock_ns();
3395#ifdef CONFIG_SYN_COOKIES
3396	if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3397		skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3398	else
3399#endif
3400	{
3401		skb->skb_mstamp_ns = now;
3402		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3403			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3404	}
3405
3406#ifdef CONFIG_TCP_MD5SIG
3407	rcu_read_lock();
3408	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3409#endif
3410	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3411	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3412					     foc, synack_type) + sizeof(*th);
3413
3414	skb_push(skb, tcp_header_size);
3415	skb_reset_transport_header(skb);
3416
3417	th = (struct tcphdr *)skb->data;
3418	memset(th, 0, sizeof(struct tcphdr));
3419	th->syn = 1;
3420	th->ack = 1;
3421	tcp_ecn_make_synack(req, th);
3422	th->source = htons(ireq->ir_num);
3423	th->dest = ireq->ir_rmt_port;
3424	skb->mark = ireq->ir_mark;
3425	skb->ip_summed = CHECKSUM_PARTIAL;
3426	th->seq = htonl(tcp_rsk(req)->snt_isn);
3427	/* XXX data is queued and acked as is. No buffer/window check */
3428	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3429
3430	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3431	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3432	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3433	th->doff = (tcp_header_size >> 2);
3434	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3435
3436#ifdef CONFIG_TCP_MD5SIG
3437	/* Okay, we have all we need - do the md5 hash if needed */
3438	if (md5)
3439		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3440					       md5, req_to_sk(req), skb);
3441	rcu_read_unlock();
3442#endif
3443
3444	skb->skb_mstamp_ns = now;
3445	tcp_add_tx_delay(skb, tp);
3446
3447	return skb;
3448}
3449EXPORT_SYMBOL(tcp_make_synack);
3450
3451static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3452{
3453	struct inet_connection_sock *icsk = inet_csk(sk);
3454	const struct tcp_congestion_ops *ca;
3455	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3456
3457	if (ca_key == TCP_CA_UNSPEC)
3458		return;
3459
3460	rcu_read_lock();
3461	ca = tcp_ca_find_key(ca_key);
3462	if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3463		bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3464		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3465		icsk->icsk_ca_ops = ca;
3466	}
3467	rcu_read_unlock();
3468}
3469
3470/* Do all connect socket setups that can be done AF independent. */
3471static void tcp_connect_init(struct sock *sk)
3472{
3473	const struct dst_entry *dst = __sk_dst_get(sk);
3474	struct tcp_sock *tp = tcp_sk(sk);
3475	__u8 rcv_wscale;
3476	u32 rcv_wnd;
3477
3478	/* We'll fix this up when we get a response from the other end.
3479	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3480	 */
3481	tp->tcp_header_len = sizeof(struct tcphdr);
3482	if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3483		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3484
3485#ifdef CONFIG_TCP_MD5SIG
3486	if (tp->af_specific->md5_lookup(sk, sk))
3487		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3488#endif
3489
3490	/* If user gave his TCP_MAXSEG, record it to clamp */
3491	if (tp->rx_opt.user_mss)
3492		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3493	tp->max_window = 0;
3494	tcp_mtup_init(sk);
3495	tcp_sync_mss(sk, dst_mtu(dst));
3496
3497	tcp_ca_dst_init(sk, dst);
3498
3499	if (!tp->window_clamp)
3500		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3501	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3502
3503	tcp_initialize_rcv_mss(sk);
3504
3505	/* limit the window selection if the user enforce a smaller rx buffer */
3506	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3507	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3508		tp->window_clamp = tcp_full_space(sk);
3509
3510	rcv_wnd = tcp_rwnd_init_bpf(sk);
3511	if (rcv_wnd == 0)
3512		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3513
3514	tcp_select_initial_window(sk, tcp_full_space(sk),
3515				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3516				  &tp->rcv_wnd,
3517				  &tp->window_clamp,
3518				  sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3519				  &rcv_wscale,
3520				  rcv_wnd);
3521
3522	tp->rx_opt.rcv_wscale = rcv_wscale;
3523	tp->rcv_ssthresh = tp->rcv_wnd;
3524
3525	sk->sk_err = 0;
3526	sock_reset_flag(sk, SOCK_DONE);
3527	tp->snd_wnd = 0;
3528	tcp_init_wl(tp, 0);
3529	tcp_write_queue_purge(sk);
3530	tp->snd_una = tp->write_seq;
3531	tp->snd_sml = tp->write_seq;
3532	tp->snd_up = tp->write_seq;
3533	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3534
3535	if (likely(!tp->repair))
3536		tp->rcv_nxt = 0;
3537	else
3538		tp->rcv_tstamp = tcp_jiffies32;
3539	tp->rcv_wup = tp->rcv_nxt;
3540	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3541
3542	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3543	inet_csk(sk)->icsk_retransmits = 0;
3544	tcp_clear_retrans(tp);
3545}
3546
3547static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3548{
3549	struct tcp_sock *tp = tcp_sk(sk);
3550	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3551
3552	tcb->end_seq += skb->len;
3553	__skb_header_release(skb);
3554	sk_wmem_queued_add(sk, skb->truesize);
3555	sk_mem_charge(sk, skb->truesize);
3556	WRITE_ONCE(tp->write_seq, tcb->end_seq);
3557	tp->packets_out += tcp_skb_pcount(skb);
3558}
3559
3560/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3561 * queue a data-only packet after the regular SYN, such that regular SYNs
3562 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3563 * only the SYN sequence, the data are retransmitted in the first ACK.
3564 * If cookie is not cached or other error occurs, falls back to send a
3565 * regular SYN with Fast Open cookie request option.
3566 */
3567static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3568{
3569	struct tcp_sock *tp = tcp_sk(sk);
3570	struct tcp_fastopen_request *fo = tp->fastopen_req;
3571	int space, err = 0;
3572	struct sk_buff *syn_data;
3573
3574	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3575	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3576		goto fallback;
3577
3578	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3579	 * user-MSS. Reserve maximum option space for middleboxes that add
3580	 * private TCP options. The cost is reduced data space in SYN :(
3581	 */
3582	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3583
3584	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3585		MAX_TCP_OPTION_SPACE;
3586
3587	space = min_t(size_t, space, fo->size);
3588
3589	/* limit to order-0 allocations */
3590	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3591
3592	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3593	if (!syn_data)
3594		goto fallback;
3595	syn_data->ip_summed = CHECKSUM_PARTIAL;
3596	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3597	if (space) {
3598		int copied = copy_from_iter(skb_put(syn_data, space), space,
3599					    &fo->data->msg_iter);
3600		if (unlikely(!copied)) {
3601			tcp_skb_tsorted_anchor_cleanup(syn_data);
3602			kfree_skb(syn_data);
3603			goto fallback;
3604		}
3605		if (copied != space) {
3606			skb_trim(syn_data, copied);
3607			space = copied;
3608		}
3609		skb_zcopy_set(syn_data, fo->uarg, NULL);
3610	}
3611	/* No more data pending in inet_wait_for_connect() */
3612	if (space == fo->size)
3613		fo->data = NULL;
3614	fo->copied = space;
3615
3616	tcp_connect_queue_skb(sk, syn_data);
3617	if (syn_data->len)
3618		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3619
3620	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3621
3622	syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3623
3624	/* Now full SYN+DATA was cloned and sent (or not),
3625	 * remove the SYN from the original skb (syn_data)
3626	 * we keep in write queue in case of a retransmit, as we
3627	 * also have the SYN packet (with no data) in the same queue.
3628	 */
3629	TCP_SKB_CB(syn_data)->seq++;
3630	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3631	if (!err) {
3632		tp->syn_data = (fo->copied > 0);
3633		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3634		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3635		goto done;
3636	}
3637
3638	/* data was not sent, put it in write_queue */
3639	__skb_queue_tail(&sk->sk_write_queue, syn_data);
3640	tp->packets_out -= tcp_skb_pcount(syn_data);
3641
3642fallback:
3643	/* Send a regular SYN with Fast Open cookie request option */
3644	if (fo->cookie.len > 0)
3645		fo->cookie.len = 0;
3646	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3647	if (err)
3648		tp->syn_fastopen = 0;
3649done:
3650	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3651	return err;
3652}
3653
3654/* Build a SYN and send it off. */
3655int tcp_connect(struct sock *sk)
3656{
3657	struct tcp_sock *tp = tcp_sk(sk);
3658	struct sk_buff *buff;
3659	int err;
3660
3661	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3662
3663	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3664		return -EHOSTUNREACH; /* Routing failure or similar. */
3665
3666	tcp_connect_init(sk);
3667
3668	if (unlikely(tp->repair)) {
3669		tcp_finish_connect(sk, NULL);
3670		return 0;
3671	}
3672
3673	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3674	if (unlikely(!buff))
3675		return -ENOBUFS;
3676
3677	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3678	tcp_mstamp_refresh(tp);
3679	tp->retrans_stamp = tcp_time_stamp(tp);
3680	tcp_connect_queue_skb(sk, buff);
3681	tcp_ecn_send_syn(sk, buff);
3682	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3683
3684	/* Send off SYN; include data in Fast Open. */
3685	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3686	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3687	if (err == -ECONNREFUSED)
3688		return err;
3689
3690	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3691	 * in order to make this packet get counted in tcpOutSegs.
3692	 */
3693	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3694	tp->pushed_seq = tp->write_seq;
3695	buff = tcp_send_head(sk);
3696	if (unlikely(buff)) {
3697		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3698		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
3699	}
3700	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3701
3702	/* Timer for repeating the SYN until an answer. */
3703	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3704				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3705	return 0;
3706}
3707EXPORT_SYMBOL(tcp_connect);
3708
3709/* Send out a delayed ack, the caller does the policy checking
3710 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3711 * for details.
3712 */
3713void tcp_send_delayed_ack(struct sock *sk)
3714{
3715	struct inet_connection_sock *icsk = inet_csk(sk);
3716	int ato = icsk->icsk_ack.ato;
3717	unsigned long timeout;
3718
 
 
3719	if (ato > TCP_DELACK_MIN) {
3720		const struct tcp_sock *tp = tcp_sk(sk);
3721		int max_ato = HZ / 2;
3722
3723		if (inet_csk_in_pingpong_mode(sk) ||
3724		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3725			max_ato = TCP_DELACK_MAX;
3726
3727		/* Slow path, intersegment interval is "high". */
3728
3729		/* If some rtt estimate is known, use it to bound delayed ack.
3730		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3731		 * directly.
3732		 */
3733		if (tp->srtt_us) {
3734			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3735					TCP_DELACK_MIN);
3736
3737			if (rtt < max_ato)
3738				max_ato = rtt;
3739		}
3740
3741		ato = min(ato, max_ato);
3742	}
3743
3744	/* Stay within the limit we were given */
3745	timeout = jiffies + ato;
3746
3747	/* Use new timeout only if there wasn't a older one earlier. */
3748	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3749		/* If delack timer was blocked or is about to expire,
3750		 * send ACK now.
3751		 */
3752		if (icsk->icsk_ack.blocked ||
3753		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3754			tcp_send_ack(sk);
3755			return;
3756		}
3757
3758		if (!time_before(timeout, icsk->icsk_ack.timeout))
3759			timeout = icsk->icsk_ack.timeout;
3760	}
3761	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3762	icsk->icsk_ack.timeout = timeout;
3763	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3764}
3765
3766/* This routine sends an ack and also updates the window. */
3767void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3768{
3769	struct sk_buff *buff;
3770
3771	/* If we have been reset, we may not send again. */
3772	if (sk->sk_state == TCP_CLOSE)
3773		return;
3774
 
 
3775	/* We are not putting this on the write queue, so
3776	 * tcp_transmit_skb() will set the ownership to this
3777	 * sock.
3778	 */
3779	buff = alloc_skb(MAX_TCP_HEADER,
3780			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3781	if (unlikely(!buff)) {
3782		inet_csk_schedule_ack(sk);
3783		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3784		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3785					  TCP_DELACK_MAX, TCP_RTO_MAX);
3786		return;
3787	}
3788
3789	/* Reserve space for headers and prepare control bits. */
3790	skb_reserve(buff, MAX_TCP_HEADER);
3791	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3792
3793	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3794	 * too much.
3795	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3796	 */
3797	skb_set_tcp_pure_ack(buff);
3798
3799	/* Send it off, this clears delayed acks for us. */
3800	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3801}
3802EXPORT_SYMBOL_GPL(__tcp_send_ack);
3803
3804void tcp_send_ack(struct sock *sk)
3805{
3806	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3807}
 
3808
3809/* This routine sends a packet with an out of date sequence
3810 * number. It assumes the other end will try to ack it.
3811 *
3812 * Question: what should we make while urgent mode?
3813 * 4.4BSD forces sending single byte of data. We cannot send
3814 * out of window data, because we have SND.NXT==SND.MAX...
3815 *
3816 * Current solution: to send TWO zero-length segments in urgent mode:
3817 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3818 * out-of-date with SND.UNA-1 to probe window.
3819 */
3820static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3821{
3822	struct tcp_sock *tp = tcp_sk(sk);
3823	struct sk_buff *skb;
3824
3825	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3826	skb = alloc_skb(MAX_TCP_HEADER,
3827			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3828	if (!skb)
3829		return -1;
3830
3831	/* Reserve space for headers and set control bits. */
3832	skb_reserve(skb, MAX_TCP_HEADER);
3833	/* Use a previous sequence.  This should cause the other
3834	 * end to send an ack.  Don't queue or clone SKB, just
3835	 * send it.
3836	 */
3837	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3838	NET_INC_STATS(sock_net(sk), mib);
3839	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3840}
3841
3842/* Called from setsockopt( ... TCP_REPAIR ) */
3843void tcp_send_window_probe(struct sock *sk)
3844{
3845	if (sk->sk_state == TCP_ESTABLISHED) {
3846		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3847		tcp_mstamp_refresh(tcp_sk(sk));
3848		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3849	}
3850}
3851
3852/* Initiate keepalive or window probe from timer. */
3853int tcp_write_wakeup(struct sock *sk, int mib)
3854{
3855	struct tcp_sock *tp = tcp_sk(sk);
3856	struct sk_buff *skb;
3857
3858	if (sk->sk_state == TCP_CLOSE)
3859		return -1;
3860
3861	skb = tcp_send_head(sk);
3862	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3863		int err;
3864		unsigned int mss = tcp_current_mss(sk);
3865		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3866
3867		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3868			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3869
3870		/* We are probing the opening of a window
3871		 * but the window size is != 0
3872		 * must have been a result SWS avoidance ( sender )
3873		 */
3874		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3875		    skb->len > mss) {
3876			seg_size = min(seg_size, mss);
3877			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3878			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3879					 skb, seg_size, mss, GFP_ATOMIC))
3880				return -1;
3881		} else if (!tcp_skb_pcount(skb))
3882			tcp_set_skb_tso_segs(skb, mss);
3883
3884		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3885		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3886		if (!err)
3887			tcp_event_new_data_sent(sk, skb);
3888		return err;
3889	} else {
3890		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3891			tcp_xmit_probe_skb(sk, 1, mib);
3892		return tcp_xmit_probe_skb(sk, 0, mib);
3893	}
3894}
3895
3896/* A window probe timeout has occurred.  If window is not closed send
3897 * a partial packet else a zero probe.
3898 */
3899void tcp_send_probe0(struct sock *sk)
3900{
3901	struct inet_connection_sock *icsk = inet_csk(sk);
3902	struct tcp_sock *tp = tcp_sk(sk);
3903	struct net *net = sock_net(sk);
3904	unsigned long timeout;
3905	int err;
3906
3907	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3908
3909	if (tp->packets_out || tcp_write_queue_empty(sk)) {
3910		/* Cancel probe timer, if it is not required. */
3911		icsk->icsk_probes_out = 0;
3912		icsk->icsk_backoff = 0;
3913		return;
3914	}
3915
3916	icsk->icsk_probes_out++;
3917	if (err <= 0) {
3918		if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3919			icsk->icsk_backoff++;
3920		timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
 
3921	} else {
3922		/* If packet was not sent due to local congestion,
3923		 * Let senders fight for local resources conservatively.
 
 
 
3924		 */
3925		timeout = TCP_RESOURCE_PROBE_INTERVAL;
3926	}
3927	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
 
 
 
 
3928}
3929
3930int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3931{
3932	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3933	struct flowi fl;
3934	int res;
3935
3936	tcp_rsk(req)->txhash = net_tx_rndhash();
3937	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3938	if (!res) {
3939		__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3940		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3941		if (unlikely(tcp_passive_fastopen(sk)))
3942			tcp_sk(sk)->total_retrans++;
3943		trace_tcp_retransmit_synack(sk, req);
3944	}
3945	return res;
3946}
3947EXPORT_SYMBOL(tcp_rtx_synack);