Linux Audio

Check our new training course

Loading...
v6.13.7
   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#include <net/proto_memory.h>
  43
  44#include <linux/compiler.h>
  45#include <linux/gfp.h>
  46#include <linux/module.h>
  47#include <linux/static_key.h>
  48#include <linux/skbuff_ref.h>
  49
  50#include <trace/events/tcp.h>
 
  51
  52/* Refresh clocks of a TCP socket,
  53 * ensuring monotically increasing values.
  54 */
  55void tcp_mstamp_refresh(struct tcp_sock *tp)
  56{
  57	u64 val = tcp_clock_ns();
  58
  59	tp->tcp_clock_cache = val;
  60	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  61}
 
 
 
 
 
 
 
 
 
 
 
  62
  63static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  64			   int push_one, gfp_t gfp);
  65
  66/* Account for new data that has been sent to the network. */
  67static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  68{
  69	struct inet_connection_sock *icsk = inet_csk(sk);
  70	struct tcp_sock *tp = tcp_sk(sk);
  71	unsigned int prior_packets = tp->packets_out;
  72
  73	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
  74
  75	__skb_unlink(skb, &sk->sk_write_queue);
  76	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  77
  78	if (tp->highest_sack == NULL)
  79		tp->highest_sack = skb;
 
  80
  81	tp->packets_out += tcp_skb_pcount(skb);
  82	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  83		tcp_rearm_rto(sk);
  84
  85	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  86		      tcp_skb_pcount(skb));
  87	tcp_check_space(sk);
  88}
  89
  90/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  91 * window scaling factor due to loss of precision.
  92 * If window has been shrunk, what should we make? It is not clear at all.
  93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  95 * invalid. OK, let's make this for now:
  96 */
  97static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  98{
  99	const struct tcp_sock *tp = tcp_sk(sk);
 100
 101	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
 102	    (tp->rx_opt.wscale_ok &&
 103	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
 104		return tp->snd_nxt;
 105	else
 106		return tcp_wnd_end(tp);
 107}
 108
 109/* Calculate mss to advertise in SYN segment.
 110 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 111 *
 112 * 1. It is independent of path mtu.
 113 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 114 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 115 *    attached devices, because some buggy hosts are confused by
 116 *    large MSS.
 117 * 4. We do not make 3, we advertise MSS, calculated from first
 118 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 119 *    This may be overridden via information stored in routing table.
 120 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 121 *    probably even Jumbo".
 122 */
 123static __u16 tcp_advertise_mss(struct sock *sk)
 124{
 125	struct tcp_sock *tp = tcp_sk(sk);
 126	const struct dst_entry *dst = __sk_dst_get(sk);
 127	int mss = tp->advmss;
 128
 129	if (dst) {
 130		unsigned int metric = dst_metric_advmss(dst);
 131
 132		if (metric < mss) {
 133			mss = metric;
 134			tp->advmss = mss;
 135		}
 136	}
 137
 138	return (__u16)mss;
 139}
 140
 141/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 142 * This is the first part of cwnd validation mechanism.
 143 */
 144void tcp_cwnd_restart(struct sock *sk, s32 delta)
 145{
 146	struct tcp_sock *tp = tcp_sk(sk);
 147	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 148	u32 cwnd = tcp_snd_cwnd(tp);
 
 149
 150	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 151
 152	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 153	restart_cwnd = min(restart_cwnd, cwnd);
 154
 155	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 156		cwnd >>= 1;
 157	tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
 158	tp->snd_cwnd_stamp = tcp_jiffies32;
 159	tp->snd_cwnd_used = 0;
 160}
 161
 162/* Congestion state accounting after a packet has been sent. */
 163static void tcp_event_data_sent(struct tcp_sock *tp,
 164				struct sock *sk)
 165{
 166	struct inet_connection_sock *icsk = inet_csk(sk);
 167	const u32 now = tcp_jiffies32;
 168
 169	if (tcp_packets_in_flight(tp) == 0)
 170		tcp_ca_event(sk, CA_EVENT_TX_START);
 
 171
 172	tp->lsndtime = now;
 173
 174	/* If it is a reply for ato after last received
 175	 * packet, increase pingpong count.
 176	 */
 177	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 178		inet_csk_inc_pingpong_cnt(sk);
 179}
 180
 181/* Account for an ACK we sent. */
 182static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
 183{
 184	struct tcp_sock *tp = tcp_sk(sk);
 185
 186	if (unlikely(tp->compressed_ack)) {
 187		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 188			      tp->compressed_ack);
 189		tp->compressed_ack = 0;
 190		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 191			__sock_put(sk);
 192	}
 193
 194	if (unlikely(rcv_nxt != tp->rcv_nxt))
 195		return;  /* Special ACK sent by DCTCP to reflect ECN */
 196	tcp_dec_quickack_mode(sk);
 197	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 198}
 199
 200/* Determine a window scaling and initial window to offer.
 201 * Based on the assumption that the given amount of space
 202 * will be offered. Store the results in the tp structure.
 203 * NOTE: for smooth operation initial space offering should
 204 * be a multiple of mss if possible. We assume here that mss >= 1.
 205 * This MUST be enforced by all callers.
 206 */
 207void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 208			       __u32 *rcv_wnd, __u32 *__window_clamp,
 209			       int wscale_ok, __u8 *rcv_wscale,
 210			       __u32 init_rcv_wnd)
 211{
 212	unsigned int space = (__space < 0 ? 0 : __space);
 213	u32 window_clamp = READ_ONCE(*__window_clamp);
 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 (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
 233		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 234	else
 235		(*rcv_wnd) = space;
 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, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
 244		space = max_t(u32, space, READ_ONCE(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	WRITE_ONCE(*__window_clamp,
 251		   min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp));
 252}
 253EXPORT_SYMBOL(tcp_select_initial_window);
 254
 255/* Chose a new window to advertise, update state in tcp_sock for the
 256 * socket, and return result with RFC1323 scaling applied.  The return
 257 * value can be stuffed directly into th->window for an outgoing
 258 * frame.
 259 */
 260static u16 tcp_select_window(struct sock *sk)
 261{
 262	struct tcp_sock *tp = tcp_sk(sk);
 263	struct net *net = sock_net(sk);
 264	u32 old_win = tp->rcv_wnd;
 265	u32 cur_win, new_win;
 266
 267	/* Make the window 0 if we failed to queue the data because we
 268	 * are out of memory.
 269	 */
 270	if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) {
 271		tp->pred_flags = 0;
 272		tp->rcv_wnd = 0;
 273		tp->rcv_wup = tp->rcv_nxt;
 274		return 0;
 275	}
 276
 277	cur_win = tcp_receive_window(tp);
 278	new_win = __tcp_select_window(sk);
 279	if (new_win < cur_win) {
 280		/* Danger Will Robinson!
 281		 * Don't update rcv_wup/rcv_wnd here or else
 282		 * we will not be able to advertise a zero
 283		 * window in time.  --DaveM
 284		 *
 285		 * Relax Will Robinson.
 286		 */
 287		if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
 288			/* Never shrink the offered window */
 289			if (new_win == 0)
 290				NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
 291			new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 292		}
 293	}
 294
 295	tp->rcv_wnd = new_win;
 296	tp->rcv_wup = tp->rcv_nxt;
 297
 298	/* Make sure we do not exceed the maximum possible
 299	 * scaled window.
 300	 */
 301	if (!tp->rx_opt.rcv_wscale &&
 302	    READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
 303		new_win = min(new_win, MAX_TCP_WINDOW);
 304	else
 305		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 306
 307	/* RFC1323 scaling applied */
 308	new_win >>= tp->rx_opt.rcv_wscale;
 309
 310	/* If we advertise zero window, disable fast path. */
 311	if (new_win == 0) {
 312		tp->pred_flags = 0;
 313		if (old_win)
 314			NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
 315	} else if (old_win == 0) {
 316		NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
 317	}
 318
 319	return new_win;
 320}
 321
 322/* Packet ECN state for a SYN-ACK */
 323static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 324{
 325	const struct tcp_sock *tp = tcp_sk(sk);
 326
 327	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 328	if (!(tp->ecn_flags & TCP_ECN_OK))
 329		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 330	else if (tcp_ca_needs_ecn(sk) ||
 331		 tcp_bpf_ca_needs_ecn(sk))
 332		INET_ECN_xmit(sk);
 333}
 334
 335/* Packet ECN state for a SYN.  */
 336static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 337{
 338	struct tcp_sock *tp = tcp_sk(sk);
 339	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 340	bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
 341		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 342
 343	if (!use_ecn) {
 344		const struct dst_entry *dst = __sk_dst_get(sk);
 345
 346		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 347			use_ecn = true;
 348	}
 349
 350	tp->ecn_flags = 0;
 351
 352	if (use_ecn) {
 353		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 354		tp->ecn_flags = TCP_ECN_OK;
 355		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 356			INET_ECN_xmit(sk);
 357	}
 358}
 359
 360static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 361{
 362	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
 363		/* tp->ecn_flags are cleared at a later point in time when
 364		 * SYN ACK is ultimatively being received.
 365		 */
 366		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 367}
 368
 369static void
 370tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 371{
 372	if (inet_rsk(req)->ecn_ok)
 373		th->ece = 1;
 374}
 375
 376/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 377 * be sent.
 378 */
 379static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 380			 struct tcphdr *th, int tcp_header_len)
 381{
 382	struct tcp_sock *tp = tcp_sk(sk);
 383
 384	if (tp->ecn_flags & TCP_ECN_OK) {
 385		/* Not-retransmitted data segment: set ECT and inject CWR. */
 386		if (skb->len != tcp_header_len &&
 387		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 388			INET_ECN_xmit(sk);
 389			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 390				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 391				th->cwr = 1;
 392				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 393			}
 394		} else if (!tcp_ca_needs_ecn(sk)) {
 395			/* ACK or retransmitted segment: clear ECT|CE */
 396			INET_ECN_dontxmit(sk);
 397		}
 398		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 399			th->ece = 1;
 400	}
 401}
 402
 403/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 404 * auto increment end seqno.
 405 */
 406static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 407{
 408	skb->ip_summed = CHECKSUM_PARTIAL;
 
 409
 410	TCP_SKB_CB(skb)->tcp_flags = flags;
 
 411
 412	tcp_skb_pcount_set(skb, 1);
 
 
 413
 414	TCP_SKB_CB(skb)->seq = seq;
 415	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 416		seq++;
 417	TCP_SKB_CB(skb)->end_seq = seq;
 418}
 419
 420static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 421{
 422	return tp->snd_una != tp->snd_up;
 423}
 424
 425#define OPTION_SACK_ADVERTISE	BIT(0)
 426#define OPTION_TS		BIT(1)
 427#define OPTION_MD5		BIT(2)
 428#define OPTION_WSCALE		BIT(3)
 429#define OPTION_FAST_OPEN_COOKIE	BIT(8)
 430#define OPTION_SMC		BIT(9)
 431#define OPTION_MPTCP		BIT(10)
 432#define OPTION_AO		BIT(11)
 433
 434static void smc_options_write(__be32 *ptr, u16 *options)
 435{
 436#if IS_ENABLED(CONFIG_SMC)
 437	if (static_branch_unlikely(&tcp_have_smc)) {
 438		if (unlikely(OPTION_SMC & *options)) {
 439			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 440				       (TCPOPT_NOP  << 16) |
 441				       (TCPOPT_EXP <<  8) |
 442				       (TCPOLEN_EXP_SMC_BASE));
 443			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 444		}
 445	}
 446#endif
 447}
 448
 449struct tcp_out_options {
 450	u16 options;		/* bit field of OPTION_* */
 451	u16 mss;		/* 0 to disable */
 452	u8 ws;			/* window scale, 0 to disable */
 453	u8 num_sack_blocks;	/* number of SACK blocks to include */
 454	u8 hash_size;		/* bytes in hash_location */
 455	u8 bpf_opt_len;		/* length of BPF hdr option */
 456	__u8 *hash_location;	/* temporary pointer, overloaded */
 457	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 458	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 459	struct mptcp_out_options mptcp;
 460};
 461
 462static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
 463				struct tcp_sock *tp,
 464				struct tcp_out_options *opts)
 465{
 466#if IS_ENABLED(CONFIG_MPTCP)
 467	if (unlikely(OPTION_MPTCP & opts->options))
 468		mptcp_write_options(th, ptr, tp, &opts->mptcp);
 469#endif
 470}
 471
 472#ifdef CONFIG_CGROUP_BPF
 473static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
 474					enum tcp_synack_type synack_type)
 475{
 476	if (unlikely(!skb))
 477		return BPF_WRITE_HDR_TCP_CURRENT_MSS;
 478
 479	if (unlikely(synack_type == TCP_SYNACK_COOKIE))
 480		return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
 481
 482	return 0;
 483}
 484
 485/* req, syn_skb and synack_type are used when writing synack */
 486static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 487				  struct request_sock *req,
 488				  struct sk_buff *syn_skb,
 489				  enum tcp_synack_type synack_type,
 490				  struct tcp_out_options *opts,
 491				  unsigned int *remaining)
 492{
 493	struct bpf_sock_ops_kern sock_ops;
 494	int err;
 495
 496	if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
 497					   BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
 498	    !*remaining)
 499		return;
 500
 501	/* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
 502
 503	/* init sock_ops */
 504	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 505
 506	sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
 507
 508	if (req) {
 509		/* The listen "sk" cannot be passed here because
 510		 * it is not locked.  It would not make too much
 511		 * sense to do bpf_setsockopt(listen_sk) based
 512		 * on individual connection request also.
 513		 *
 514		 * Thus, "req" is passed here and the cgroup-bpf-progs
 515		 * of the listen "sk" will be run.
 516		 *
 517		 * "req" is also used here for fastopen even the "sk" here is
 518		 * a fullsock "child" sk.  It is to keep the behavior
 519		 * consistent between fastopen and non-fastopen on
 520		 * the bpf programming side.
 521		 */
 522		sock_ops.sk = (struct sock *)req;
 523		sock_ops.syn_skb = syn_skb;
 524	} else {
 525		sock_owned_by_me(sk);
 526
 527		sock_ops.is_fullsock = 1;
 528		sock_ops.sk = sk;
 529	}
 530
 531	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 532	sock_ops.remaining_opt_len = *remaining;
 533	/* tcp_current_mss() does not pass a skb */
 534	if (skb)
 535		bpf_skops_init_skb(&sock_ops, skb, 0);
 536
 537	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 538
 539	if (err || sock_ops.remaining_opt_len == *remaining)
 540		return;
 541
 542	opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
 543	/* round up to 4 bytes */
 544	opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
 545
 546	*remaining -= opts->bpf_opt_len;
 547}
 548
 549static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 550				    struct request_sock *req,
 551				    struct sk_buff *syn_skb,
 552				    enum tcp_synack_type synack_type,
 553				    struct tcp_out_options *opts)
 554{
 555	u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
 556	struct bpf_sock_ops_kern sock_ops;
 557	int err;
 558
 559	if (likely(!max_opt_len))
 560		return;
 561
 562	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 563
 564	sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
 565
 566	if (req) {
 567		sock_ops.sk = (struct sock *)req;
 568		sock_ops.syn_skb = syn_skb;
 569	} else {
 570		sock_owned_by_me(sk);
 571
 572		sock_ops.is_fullsock = 1;
 573		sock_ops.sk = sk;
 574	}
 575
 576	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 577	sock_ops.remaining_opt_len = max_opt_len;
 578	first_opt_off = tcp_hdrlen(skb) - max_opt_len;
 579	bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
 580
 581	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 582
 583	if (err)
 584		nr_written = 0;
 585	else
 586		nr_written = max_opt_len - sock_ops.remaining_opt_len;
 587
 588	if (nr_written < max_opt_len)
 589		memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
 590		       max_opt_len - nr_written);
 591}
 592#else
 593static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 594				  struct request_sock *req,
 595				  struct sk_buff *syn_skb,
 596				  enum tcp_synack_type synack_type,
 597				  struct tcp_out_options *opts,
 598				  unsigned int *remaining)
 599{
 600}
 601
 602static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 603				    struct request_sock *req,
 604				    struct sk_buff *syn_skb,
 605				    enum tcp_synack_type synack_type,
 606				    struct tcp_out_options *opts)
 607{
 608}
 609#endif
 
 
 
 610
 611static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
 612				      const struct tcp_request_sock *tcprsk,
 613				      struct tcp_out_options *opts,
 614				      struct tcp_key *key, __be32 *ptr)
 615{
 616#ifdef CONFIG_TCP_AO
 617	u8 maclen = tcp_ao_maclen(key->ao_key);
 618
 619	if (tcprsk) {
 620		u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
 621
 622		*ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
 623			       (tcprsk->ao_keyid << 8) |
 624			       (tcprsk->ao_rcv_next));
 625	} else {
 626		struct tcp_ao_key *rnext_key;
 627		struct tcp_ao_info *ao_info;
 628
 629		ao_info = rcu_dereference_check(tp->ao_info,
 630			lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
 631		rnext_key = READ_ONCE(ao_info->rnext_key);
 632		if (WARN_ON_ONCE(!rnext_key))
 633			return ptr;
 634		*ptr++ = htonl((TCPOPT_AO << 24) |
 635			       (tcp_ao_len(key->ao_key) << 16) |
 636			       (key->ao_key->sndid << 8) |
 637			       (rnext_key->rcvid));
 638	}
 639	opts->hash_location = (__u8 *)ptr;
 640	ptr += maclen / sizeof(*ptr);
 641	if (unlikely(maclen % sizeof(*ptr))) {
 642		memset(ptr, TCPOPT_NOP, sizeof(*ptr));
 643		ptr++;
 644	}
 645#endif
 646	return ptr;
 647}
 648
 649/* Write previously computed TCP options to the packet.
 650 *
 651 * Beware: Something in the Internet is very sensitive to the ordering of
 652 * TCP options, we learned this through the hard way, so be careful here.
 653 * Luckily we can at least blame others for their non-compliance but from
 654 * inter-operability perspective it seems that we're somewhat stuck with
 655 * the ordering which we have been using if we want to keep working with
 656 * those broken things (not that it currently hurts anybody as there isn't
 657 * particular reason why the ordering would need to be changed).
 658 *
 659 * At least SACK_PERM as the first option is known to lead to a disaster
 660 * (but it may well be that other scenarios fail similarly).
 661 */
 662static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
 663			      const struct tcp_request_sock *tcprsk,
 664			      struct tcp_out_options *opts,
 665			      struct tcp_key *key)
 666{
 667	__be32 *ptr = (__be32 *)(th + 1);
 668	u16 options = opts->options;	/* mungable copy */
 669
 670	if (tcp_key_is_md5(key)) {
 671		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 672			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 673		/* overload cookie hash location */
 674		opts->hash_location = (__u8 *)ptr;
 675		ptr += 4;
 676	} else if (tcp_key_is_ao(key)) {
 677		ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
 678	}
 
 679	if (unlikely(opts->mss)) {
 680		*ptr++ = htonl((TCPOPT_MSS << 24) |
 681			       (TCPOLEN_MSS << 16) |
 682			       opts->mss);
 683	}
 684
 685	if (likely(OPTION_TS & options)) {
 686		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 687			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 688				       (TCPOLEN_SACK_PERM << 16) |
 689				       (TCPOPT_TIMESTAMP << 8) |
 690				       TCPOLEN_TIMESTAMP);
 691			options &= ~OPTION_SACK_ADVERTISE;
 692		} else {
 693			*ptr++ = htonl((TCPOPT_NOP << 24) |
 694				       (TCPOPT_NOP << 16) |
 695				       (TCPOPT_TIMESTAMP << 8) |
 696				       TCPOLEN_TIMESTAMP);
 697		}
 698		*ptr++ = htonl(opts->tsval);
 699		*ptr++ = htonl(opts->tsecr);
 700	}
 701
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 702	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 703		*ptr++ = htonl((TCPOPT_NOP << 24) |
 704			       (TCPOPT_NOP << 16) |
 705			       (TCPOPT_SACK_PERM << 8) |
 706			       TCPOLEN_SACK_PERM);
 707	}
 708
 709	if (unlikely(OPTION_WSCALE & options)) {
 710		*ptr++ = htonl((TCPOPT_NOP << 24) |
 711			       (TCPOPT_WINDOW << 16) |
 712			       (TCPOLEN_WINDOW << 8) |
 713			       opts->ws);
 714	}
 715
 716	if (unlikely(opts->num_sack_blocks)) {
 717		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 718			tp->duplicate_sack : tp->selective_acks;
 719		int this_sack;
 720
 721		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 722			       (TCPOPT_NOP  << 16) |
 723			       (TCPOPT_SACK <<  8) |
 724			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 725						     TCPOLEN_SACK_PERBLOCK)));
 726
 727		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 728		     ++this_sack) {
 729			*ptr++ = htonl(sp[this_sack].start_seq);
 730			*ptr++ = htonl(sp[this_sack].end_seq);
 731		}
 732
 733		tp->rx_opt.dsack = 0;
 734	}
 735
 736	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 737		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 738		u8 *p = (u8 *)ptr;
 739		u32 len; /* Fast Open option length */
 740
 741		if (foc->exp) {
 742			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 743			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 744				     TCPOPT_FASTOPEN_MAGIC);
 745			p += TCPOLEN_EXP_FASTOPEN_BASE;
 746		} else {
 747			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 748			*p++ = TCPOPT_FASTOPEN;
 749			*p++ = len;
 750		}
 751
 752		memcpy(p, foc->val, foc->len);
 753		if ((len & 3) == 2) {
 754			p[foc->len] = TCPOPT_NOP;
 755			p[foc->len + 1] = TCPOPT_NOP;
 756		}
 757		ptr += (len + 3) >> 2;
 758	}
 759
 760	smc_options_write(ptr, &options);
 761
 762	mptcp_options_write(th, ptr, tp, opts);
 763}
 764
 765static void smc_set_option(const struct tcp_sock *tp,
 766			   struct tcp_out_options *opts,
 767			   unsigned int *remaining)
 768{
 769#if IS_ENABLED(CONFIG_SMC)
 770	if (static_branch_unlikely(&tcp_have_smc)) {
 771		if (tp->syn_smc) {
 772			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 773				opts->options |= OPTION_SMC;
 774				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 775			}
 776		}
 777	}
 778#endif
 779}
 780
 781static void smc_set_option_cond(const struct tcp_sock *tp,
 782				const struct inet_request_sock *ireq,
 783				struct tcp_out_options *opts,
 784				unsigned int *remaining)
 785{
 786#if IS_ENABLED(CONFIG_SMC)
 787	if (static_branch_unlikely(&tcp_have_smc)) {
 788		if (tp->syn_smc && ireq->smc_ok) {
 789			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 790				opts->options |= OPTION_SMC;
 791				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 792			}
 793		}
 794	}
 795#endif
 796}
 797
 798static void mptcp_set_option_cond(const struct request_sock *req,
 799				  struct tcp_out_options *opts,
 800				  unsigned int *remaining)
 801{
 802	if (rsk_is_mptcp(req)) {
 803		unsigned int size;
 804
 805		if (mptcp_synack_options(req, &size, &opts->mptcp)) {
 806			if (*remaining >= size) {
 807				opts->options |= OPTION_MPTCP;
 808				*remaining -= size;
 809			}
 810		}
 811	}
 812}
 813
 814/* Compute TCP options for SYN packets. This is not the final
 815 * network wire format yet.
 816 */
 817static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 818				struct tcp_out_options *opts,
 819				struct tcp_key *key)
 820{
 821	struct tcp_sock *tp = tcp_sk(sk);
 822	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 823	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 824	bool timestamps;
 825
 826	/* Better than switch (key.type) as it has static branches */
 827	if (tcp_key_is_md5(key)) {
 828		timestamps = false;
 
 
 829		opts->options |= OPTION_MD5;
 830		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 831	} else {
 832		timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
 833		if (tcp_key_is_ao(key)) {
 834			opts->options |= OPTION_AO;
 835			remaining -= tcp_ao_len_aligned(key->ao_key);
 836		}
 837	}
 
 
 
 838
 839	/* We always get an MSS option.  The option bytes which will be seen in
 840	 * normal data packets should timestamps be used, must be in the MSS
 841	 * advertised.  But we subtract them from tp->mss_cache so that
 842	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 843	 * fact here if necessary.  If we don't do this correctly, as a
 844	 * receiver we won't recognize data packets as being full sized when we
 845	 * should, and thus we won't abide by the delayed ACK rules correctly.
 846	 * SACKs don't matter, we never delay an ACK when we have any of those
 847	 * going out.  */
 848	opts->mss = tcp_advertise_mss(sk);
 849	remaining -= TCPOLEN_MSS_ALIGNED;
 850
 851	if (likely(timestamps)) {
 852		opts->options |= OPTION_TS;
 853		opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset;
 854		opts->tsecr = tp->rx_opt.ts_recent;
 855		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 856	}
 857	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
 858		opts->ws = tp->rx_opt.rcv_wscale;
 859		opts->options |= OPTION_WSCALE;
 860		remaining -= TCPOLEN_WSCALE_ALIGNED;
 861	}
 862	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
 863		opts->options |= OPTION_SACK_ADVERTISE;
 864		if (unlikely(!(OPTION_TS & opts->options)))
 865			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 866	}
 867
 868	if (fastopen && fastopen->cookie.len >= 0) {
 869		u32 need = fastopen->cookie.len;
 870
 871		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 872					       TCPOLEN_FASTOPEN_BASE;
 873		need = (need + 3) & ~3U;  /* Align to 32 bits */
 874		if (remaining >= need) {
 875			opts->options |= OPTION_FAST_OPEN_COOKIE;
 876			opts->fastopen_cookie = &fastopen->cookie;
 877			remaining -= need;
 878			tp->syn_fastopen = 1;
 879			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 
 
 
 
 
 
 
 
 880		}
 881	}
 882
 883	smc_set_option(tp, opts, &remaining);
 884
 885	if (sk_is_mptcp(sk)) {
 886		unsigned int size;
 887
 888		if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
 889			if (remaining >= size) {
 890				opts->options |= OPTION_MPTCP;
 891				remaining -= size;
 
 
 
 
 
 
 
 
 
 892			}
 
 
 893		}
 894	}
 895
 896	bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
 897
 898	return MAX_TCP_OPTION_SPACE - remaining;
 899}
 900
 901/* Set up TCP options for SYN-ACKs. */
 902static unsigned int tcp_synack_options(const struct sock *sk,
 903				       struct request_sock *req,
 904				       unsigned int mss, struct sk_buff *skb,
 905				       struct tcp_out_options *opts,
 906				       const struct tcp_key *key,
 907				       struct tcp_fastopen_cookie *foc,
 908				       enum tcp_synack_type synack_type,
 909				       struct sk_buff *syn_skb)
 910{
 911	struct inet_request_sock *ireq = inet_rsk(req);
 912	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 
 
 
 913
 914	if (tcp_key_is_md5(key)) {
 
 
 915		opts->options |= OPTION_MD5;
 916		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 917
 918		/* We can't fit any SACK blocks in a packet with MD5 + TS
 919		 * options. There was discussion about disabling SACK
 920		 * rather than TS in order to fit in better with old,
 921		 * buggy kernels, but that was deemed to be unnecessary.
 922		 */
 923		if (synack_type != TCP_SYNACK_COOKIE)
 924			ireq->tstamp_ok &= !ireq->sack_ok;
 925	} else if (tcp_key_is_ao(key)) {
 926		opts->options |= OPTION_AO;
 927		remaining -= tcp_ao_len_aligned(key->ao_key);
 928		ireq->tstamp_ok &= !ireq->sack_ok;
 929	}
 
 
 
 930
 931	/* We always send an MSS option. */
 932	opts->mss = mss;
 933	remaining -= TCPOLEN_MSS_ALIGNED;
 934
 935	if (likely(ireq->wscale_ok)) {
 936		opts->ws = ireq->rcv_wscale;
 937		opts->options |= OPTION_WSCALE;
 938		remaining -= TCPOLEN_WSCALE_ALIGNED;
 939	}
 940	if (likely(ireq->tstamp_ok)) {
 941		opts->options |= OPTION_TS;
 942		opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) +
 943			      tcp_rsk(req)->ts_off;
 944		opts->tsecr = READ_ONCE(req->ts_recent);
 945		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 946	}
 947	if (likely(ireq->sack_ok)) {
 948		opts->options |= OPTION_SACK_ADVERTISE;
 949		if (unlikely(!ireq->tstamp_ok))
 950			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 951	}
 952	if (foc != NULL && foc->len >= 0) {
 953		u32 need = foc->len;
 954
 955		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 956				   TCPOLEN_FASTOPEN_BASE;
 957		need = (need + 3) & ~3U;  /* Align to 32 bits */
 958		if (remaining >= need) {
 959			opts->options |= OPTION_FAST_OPEN_COOKIE;
 960			opts->fastopen_cookie = foc;
 
 
 
 
 
 
 
 
 
 961			remaining -= need;
 
 
 
 
 962		}
 963	}
 964
 965	mptcp_set_option_cond(req, opts, &remaining);
 966
 967	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 968
 969	bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
 970			      synack_type, opts, &remaining);
 971
 972	return MAX_TCP_OPTION_SPACE - remaining;
 973}
 974
 975/* Compute TCP options for ESTABLISHED sockets. This is not the
 976 * final wire format yet.
 977 */
 978static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 979					struct tcp_out_options *opts,
 980					struct tcp_key *key)
 981{
 982	struct tcp_sock *tp = tcp_sk(sk);
 983	unsigned int size = 0;
 984	unsigned int eff_sacks;
 985
 986	opts->options = 0;
 987
 988	/* Better than switch (key.type) as it has static branches */
 989	if (tcp_key_is_md5(key)) {
 990		opts->options |= OPTION_MD5;
 991		size += TCPOLEN_MD5SIG_ALIGNED;
 992	} else if (tcp_key_is_ao(key)) {
 993		opts->options |= OPTION_AO;
 994		size += tcp_ao_len_aligned(key->ao_key);
 995	}
 
 
 
 996
 997	if (likely(tp->rx_opt.tstamp_ok)) {
 998		opts->options |= OPTION_TS;
 999		opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) +
1000				tp->tsoffset : 0;
1001		opts->tsecr = tp->rx_opt.ts_recent;
1002		size += TCPOLEN_TSTAMP_ALIGNED;
1003	}
1004
1005	/* MPTCP options have precedence over SACK for the limited TCP
1006	 * option space because a MPTCP connection would be forced to
1007	 * fall back to regular TCP if a required multipath option is
1008	 * missing. SACK still gets a chance to use whatever space is
1009	 * left.
1010	 */
1011	if (sk_is_mptcp(sk)) {
1012		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1013		unsigned int opt_size = 0;
1014
1015		if (mptcp_established_options(sk, skb, &opt_size, remaining,
1016					      &opts->mptcp)) {
1017			opts->options |= OPTION_MPTCP;
1018			size += opt_size;
1019		}
1020	}
1021
1022	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1023	if (unlikely(eff_sacks)) {
1024		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1025		if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
1026					 TCPOLEN_SACK_PERBLOCK))
1027			return size;
1028
1029		opts->num_sack_blocks =
1030			min_t(unsigned int, eff_sacks,
1031			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1032			      TCPOLEN_SACK_PERBLOCK);
1033
1034		size += TCPOLEN_SACK_BASE_ALIGNED +
1035			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1036	}
1037
1038	if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1039					    BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1040		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1041
1042		bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1043
1044		size = MAX_TCP_OPTION_SPACE - remaining;
1045	}
1046
1047	return size;
1048}
1049
1050
1051/* TCP SMALL QUEUES (TSQ)
1052 *
1053 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1054 * to reduce RTT and bufferbloat.
1055 * We do this using a special skb destructor (tcp_wfree).
1056 *
1057 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1058 * needs to be reallocated in a driver.
1059 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1060 *
1061 * Since transmit from skb destructor is forbidden, we use a tasklet
1062 * to process all sockets that eventually need to send more skbs.
1063 * We use one tasklet per cpu, with its own queue of sockets.
1064 */
1065struct tsq_tasklet {
1066	struct tasklet_struct	tasklet;
1067	struct list_head	head; /* queue of tcp sockets */
1068};
1069static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1070
1071static void tcp_tsq_write(struct sock *sk)
1072{
1073	if ((1 << sk->sk_state) &
1074	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1075	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
1076		struct tcp_sock *tp = tcp_sk(sk);
1077
1078		if (tp->lost_out > tp->retrans_out &&
1079		    tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1080			tcp_mstamp_refresh(tp);
1081			tcp_xmit_retransmit_queue(sk);
1082		}
1083
1084		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1085			       0, GFP_ATOMIC);
1086	}
1087}
1088
1089static void tcp_tsq_handler(struct sock *sk)
1090{
1091	bh_lock_sock(sk);
1092	if (!sock_owned_by_user(sk))
1093		tcp_tsq_write(sk);
1094	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1095		sock_hold(sk);
1096	bh_unlock_sock(sk);
1097}
1098/*
1099 * One tasklet per cpu tries to send more skbs.
1100 * We run in tasklet context but need to disable irqs when
1101 * transferring tsq->head because tcp_wfree() might
1102 * interrupt us (non NAPI drivers)
1103 */
1104static void tcp_tasklet_func(struct tasklet_struct *t)
1105{
1106	struct tsq_tasklet *tsq = from_tasklet(tsq,  t, tasklet);
1107	LIST_HEAD(list);
1108	unsigned long flags;
1109	struct list_head *q, *n;
1110	struct tcp_sock *tp;
1111	struct sock *sk;
1112
1113	local_irq_save(flags);
1114	list_splice_init(&tsq->head, &list);
1115	local_irq_restore(flags);
1116
1117	list_for_each_safe(q, n, &list) {
1118		tp = list_entry(q, struct tcp_sock, tsq_node);
1119		list_del(&tp->tsq_node);
1120
1121		sk = (struct sock *)tp;
1122		smp_mb__before_atomic();
1123		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1124
1125		tcp_tsq_handler(sk);
1126		sk_free(sk);
1127	}
1128}
1129
1130#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
1131			  TCPF_WRITE_TIMER_DEFERRED |	\
1132			  TCPF_DELACK_TIMER_DEFERRED |	\
1133			  TCPF_MTU_REDUCED_DEFERRED |	\
1134			  TCPF_ACK_DEFERRED)
1135/**
1136 * tcp_release_cb - tcp release_sock() callback
1137 * @sk: socket
1138 *
1139 * called from release_sock() to perform protocol dependent
1140 * actions before socket release.
1141 */
1142void tcp_release_cb(struct sock *sk)
1143{
1144	unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1145	unsigned long nflags;
1146
1147	/* perform an atomic operation only if at least one flag is set */
1148	do {
1149		if (!(flags & TCP_DEFERRED_ALL))
1150			return;
1151		nflags = flags & ~TCP_DEFERRED_ALL;
1152	} while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1153
1154	if (flags & TCPF_TSQ_DEFERRED) {
1155		tcp_tsq_write(sk);
1156		__sock_put(sk);
1157	}
1158
1159	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1160		tcp_write_timer_handler(sk);
1161		__sock_put(sk);
1162	}
1163	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1164		tcp_delack_timer_handler(sk);
1165		__sock_put(sk);
1166	}
1167	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1168		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1169		__sock_put(sk);
1170	}
1171	if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1172		tcp_send_ack(sk);
1173}
1174EXPORT_SYMBOL(tcp_release_cb);
1175
1176void __init tcp_tasklet_init(void)
1177{
1178	int i;
1179
1180	for_each_possible_cpu(i) {
1181		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1182
1183		INIT_LIST_HEAD(&tsq->head);
1184		tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1185	}
1186}
1187
1188/*
1189 * Write buffer destructor automatically called from kfree_skb.
1190 * We can't xmit new skbs from this context, as we might already
1191 * hold qdisc lock.
1192 */
1193void tcp_wfree(struct sk_buff *skb)
1194{
1195	struct sock *sk = skb->sk;
1196	struct tcp_sock *tp = tcp_sk(sk);
1197	unsigned long flags, nval, oval;
1198	struct tsq_tasklet *tsq;
1199	bool empty;
1200
1201	/* Keep one reference on sk_wmem_alloc.
1202	 * Will be released by sk_free() from here or tcp_tasklet_func()
1203	 */
1204	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1205
1206	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
1207	 * Wait until our queues (qdisc + devices) are drained.
1208	 * This gives :
1209	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1210	 * - chance for incoming ACK (processed by another cpu maybe)
1211	 *   to migrate this flow (skb->ooo_okay will be eventually set)
1212	 */
1213	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1214		goto out;
1215
1216	oval = smp_load_acquire(&sk->sk_tsq_flags);
1217	do {
1218		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1219			goto out;
1220
1221		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1222	} while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1223
1224	/* queue this socket to tasklet queue */
1225	local_irq_save(flags);
1226	tsq = this_cpu_ptr(&tsq_tasklet);
1227	empty = list_empty(&tsq->head);
1228	list_add(&tp->tsq_node, &tsq->head);
1229	if (empty)
1230		tasklet_schedule(&tsq->tasklet);
1231	local_irq_restore(flags);
1232	return;
1233out:
1234	sk_free(sk);
1235}
1236
1237/* Note: Called under soft irq.
1238 * We can call TCP stack right away, unless socket is owned by user.
1239 */
1240enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1241{
1242	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1243	struct sock *sk = (struct sock *)tp;
1244
1245	tcp_tsq_handler(sk);
1246	sock_put(sk);
1247
1248	return HRTIMER_NORESTART;
1249}
1250
1251static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1252				      u64 prior_wstamp)
1253{
1254	struct tcp_sock *tp = tcp_sk(sk);
1255
1256	if (sk->sk_pacing_status != SK_PACING_NONE) {
1257		unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1258
1259		/* Original sch_fq does not pace first 10 MSS
1260		 * Note that tp->data_segs_out overflows after 2^32 packets,
1261		 * this is a minor annoyance.
1262		 */
1263		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1264			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1265			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1266
1267			/* take into account OS jitter */
1268			len_ns -= min_t(u64, len_ns / 2, credit);
1269			tp->tcp_wstamp_ns += len_ns;
1270		}
1271	}
1272	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1273}
1274
1275INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1276INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1277INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1278
1279/* This routine actually transmits TCP packets queued in by
1280 * tcp_do_sendmsg().  This is used by both the initial
1281 * transmission and possible later retransmissions.
1282 * All SKB's seen here are completely headerless.  It is our
1283 * job to build the TCP header, and pass the packet down to
1284 * IP so it can do the same plus pass the packet off to the
1285 * device.
1286 *
1287 * We are working here with either a clone of the original
1288 * SKB, or a fresh unique copy made by the retransmit engine.
1289 */
1290static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1291			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1292{
1293	const struct inet_connection_sock *icsk = inet_csk(sk);
1294	struct inet_sock *inet;
1295	struct tcp_sock *tp;
1296	struct tcp_skb_cb *tcb;
1297	struct tcp_out_options opts;
1298	unsigned int tcp_options_size, tcp_header_size;
1299	struct sk_buff *oskb = NULL;
1300	struct tcp_key key;
1301	struct tcphdr *th;
1302	u64 prior_wstamp;
1303	int err;
1304
1305	BUG_ON(!skb || !tcp_skb_pcount(skb));
1306	tp = tcp_sk(sk);
1307	prior_wstamp = tp->tcp_wstamp_ns;
1308	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1309	skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
1310	if (clone_it) {
1311		oskb = skb;
1312
1313		tcp_skb_tsorted_save(oskb) {
1314			if (unlikely(skb_cloned(oskb)))
1315				skb = pskb_copy(oskb, gfp_mask);
1316			else
1317				skb = skb_clone(oskb, gfp_mask);
1318		} tcp_skb_tsorted_restore(oskb);
1319
 
 
 
 
 
 
 
 
 
 
 
1320		if (unlikely(!skb))
1321			return -ENOBUFS;
1322		/* retransmit skbs might have a non zero value in skb->dev
1323		 * because skb->dev is aliased with skb->rbnode.rb_left
1324		 */
1325		skb->dev = NULL;
1326	}
1327
1328	inet = inet_sk(sk);
 
1329	tcb = TCP_SKB_CB(skb);
1330	memset(&opts, 0, sizeof(opts));
1331
1332	tcp_get_current_key(sk, &key);
1333	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1334		tcp_options_size = tcp_syn_options(sk, skb, &opts, &key);
1335	} else {
1336		tcp_options_size = tcp_established_options(sk, skb, &opts, &key);
1337		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1338		 * at receiver : This slightly improve GRO performance.
1339		 * Note that we do not force the PSH flag for non GSO packets,
1340		 * because they might be sent under high congestion events,
1341		 * and in this case it is better to delay the delivery of 1-MSS
1342		 * packets and thus the corresponding ACK packet that would
1343		 * release the following packet.
1344		 */
1345		if (tcp_skb_pcount(skb) > 1)
1346			tcb->tcp_flags |= TCPHDR_PSH;
1347	}
1348	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1349
1350	/* We set skb->ooo_okay to one if this packet can select
1351	 * a different TX queue than prior packets of this flow,
1352	 * to avoid self inflicted reorders.
1353	 * The 'other' queue decision is based on current cpu number
1354	 * if XPS is enabled, or sk->sk_txhash otherwise.
1355	 * We can switch to another (and better) queue if:
1356	 * 1) No packet with payload is in qdisc/device queues.
1357	 *    Delays in TX completion can defeat the test
1358	 *    even if packets were already sent.
1359	 * 2) Or rtx queue is empty.
1360	 *    This mitigates above case if ACK packets for
1361	 *    all prior packets were already processed.
1362	 */
1363	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1364			tcp_rtx_queue_empty(sk);
1365
1366	/* If we had to use memory reserve to allocate this skb,
1367	 * this might cause drops if packet is looped back :
1368	 * Other socket might not have SOCK_MEMALLOC.
1369	 * Packets not looped back do not care about pfmemalloc.
1370	 */
1371	skb->pfmemalloc = 0;
1372
1373	skb_push(skb, tcp_header_size);
1374	skb_reset_transport_header(skb);
1375
1376	skb_orphan(skb);
1377	skb->sk = sk;
1378	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1379	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1380
1381	skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1382
1383	/* Build TCP header and checksum it. */
1384	th = (struct tcphdr *)skb->data;
1385	th->source		= inet->inet_sport;
1386	th->dest		= inet->inet_dport;
1387	th->seq			= htonl(tcb->seq);
1388	th->ack_seq		= htonl(rcv_nxt);
1389	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1390					tcb->tcp_flags);
1391
 
 
 
 
 
 
 
 
1392	th->check		= 0;
1393	th->urg_ptr		= 0;
1394
1395	/* The urg_mode check is necessary during a below snd_una win probe */
1396	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1397		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1398			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1399			th->urg = 1;
1400		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1401			th->urg_ptr = htons(0xFFFF);
1402			th->urg = 1;
1403		}
1404	}
1405
1406	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1407	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1408		th->window      = htons(tcp_select_window(sk));
1409		tcp_ecn_send(sk, skb, th, tcp_header_size);
1410	} else {
1411		/* RFC1323: The window in SYN & SYN/ACK segments
1412		 * is never scaled.
1413		 */
1414		th->window	= htons(min(tp->rcv_wnd, 65535U));
1415	}
1416
1417	tcp_options_write(th, tp, NULL, &opts, &key);
1418
1419	if (tcp_key_is_md5(&key)) {
1420#ifdef CONFIG_TCP_MD5SIG
1421		/* Calculate the MD5 hash, as we have all we need now */
1422		sk_gso_disable(sk);
 
1423		tp->af_specific->calc_md5_hash(opts.hash_location,
1424					       key.md5_key, sk, skb);
 
1425#endif
1426	} else if (tcp_key_is_ao(&key)) {
1427		int err;
1428
1429		err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th,
1430					  opts.hash_location);
1431		if (err) {
1432			kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED);
1433			return -ENOMEM;
1434		}
1435	}
1436
1437	/* BPF prog is the last one writing header option */
1438	bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1439
1440	INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1441			   tcp_v6_send_check, tcp_v4_send_check,
1442			   sk, skb);
1443
1444	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1445		tcp_event_ack_sent(sk, rcv_nxt);
1446
1447	if (skb->len != tcp_header_size) {
1448		tcp_event_data_sent(tp, sk);
1449		tp->data_segs_out += tcp_skb_pcount(skb);
1450		tp->bytes_sent += skb->len - tcp_header_size;
1451	}
1452
1453	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1454		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1455			      tcp_skb_pcount(skb));
1456
1457	tp->segs_out += tcp_skb_pcount(skb);
1458	skb_set_hash_from_sk(skb, sk);
1459	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1460	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1461	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1462
1463	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1464
1465	/* Cleanup our debris for IP stacks */
1466	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1467			       sizeof(struct inet6_skb_parm)));
1468
1469	tcp_add_tx_delay(skb, tp);
1470
1471	err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1472				 inet6_csk_xmit, ip_queue_xmit,
1473				 sk, skb, &inet->cork.fl);
1474
1475	if (unlikely(err > 0)) {
1476		tcp_enter_cwr(sk);
1477		err = net_xmit_eval(err);
1478	}
1479	if (!err && oskb) {
1480		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1481		tcp_rate_skb_sent(sk, oskb);
1482	}
1483	return err;
1484}
1485
1486static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1487			    gfp_t gfp_mask)
1488{
1489	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1490				  tcp_sk(sk)->rcv_nxt);
1491}
1492
1493/* This routine just queues the buffer for sending.
1494 *
1495 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1496 * otherwise socket can stall.
1497 */
1498static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1499{
1500	struct tcp_sock *tp = tcp_sk(sk);
1501
1502	/* Advance write_seq and place onto the write_queue. */
1503	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1504	__skb_header_release(skb);
1505	tcp_add_write_queue_tail(sk, skb);
1506	sk_wmem_queued_add(sk, skb->truesize);
1507	sk_mem_charge(sk, skb->truesize);
1508}
1509
1510/* Initialize TSO segments for a packet. */
1511static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
 
1512{
1513	int tso_segs;
1514
1515	if (skb->len <= mss_now) {
1516		/* Avoid the costly divide in the normal
1517		 * non-TSO case.
1518		 */
1519		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1520		tcp_skb_pcount_set(skb, 1);
1521		return 1;
 
 
 
 
1522	}
1523	TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1524	tso_segs = DIV_ROUND_UP(skb->len, mss_now);
1525	tcp_skb_pcount_set(skb, tso_segs);
1526	return tso_segs;
 
 
 
 
 
 
 
 
 
 
 
1527}
1528
1529/* Pcount in the middle of the write queue got changed, we need to do various
1530 * tweaks to fix counters
1531 */
1532static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1533{
1534	struct tcp_sock *tp = tcp_sk(sk);
1535
1536	tp->packets_out -= decr;
1537
1538	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1539		tp->sacked_out -= decr;
1540	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1541		tp->retrans_out -= decr;
1542	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1543		tp->lost_out -= decr;
1544
1545	/* Reno case is special. Sigh... */
1546	if (tcp_is_reno(tp) && decr > 0)
1547		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1548
 
 
1549	if (tp->lost_skb_hint &&
1550	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1551	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1552		tp->lost_cnt_hint -= decr;
1553
1554	tcp_verify_left_out(tp);
1555}
1556
1557static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1558{
1559	return TCP_SKB_CB(skb)->txstamp_ack ||
1560		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1561}
1562
1563static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1564{
1565	struct skb_shared_info *shinfo = skb_shinfo(skb);
1566
1567	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1568	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1569		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1570		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1571
1572		shinfo->tx_flags &= ~tsflags;
1573		shinfo2->tx_flags |= tsflags;
1574		swap(shinfo->tskey, shinfo2->tskey);
1575		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1576		TCP_SKB_CB(skb)->txstamp_ack = 0;
1577	}
1578}
1579
1580static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1581{
1582	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1583	TCP_SKB_CB(skb)->eor = 0;
1584}
1585
1586/* Insert buff after skb on the write or rtx queue of sk.  */
1587static void tcp_insert_write_queue_after(struct sk_buff *skb,
1588					 struct sk_buff *buff,
1589					 struct sock *sk,
1590					 enum tcp_queue tcp_queue)
1591{
1592	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1593		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1594	else
1595		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1596}
1597
1598/* Function to create two new TCP segments.  Shrinks the given segment
1599 * to the specified size and appends a new segment with the rest of the
1600 * packet to the list.  This won't be called frequently, I hope.
1601 * Remember, these are still headerless SKBs at this point.
1602 */
1603int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1604		 struct sk_buff *skb, u32 len,
1605		 unsigned int mss_now, gfp_t gfp)
1606{
1607	struct tcp_sock *tp = tcp_sk(sk);
1608	struct sk_buff *buff;
1609	int old_factor;
1610	long limit;
1611	int nlen;
1612	u8 flags;
1613
1614	if (WARN_ON(len > skb->len))
1615		return -EINVAL;
1616
1617	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1618
1619	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1620	 * We need some allowance to not penalize applications setting small
1621	 * SO_SNDBUF values.
1622	 * Also allow first and last skb in retransmit queue to be split.
1623	 */
1624	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1625	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1626		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1627		     skb != tcp_rtx_queue_head(sk) &&
1628		     skb != tcp_rtx_queue_tail(sk))) {
1629		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1630		return -ENOMEM;
1631	}
1632
1633	if (skb_unclone_keeptruesize(skb, gfp))
1634		return -ENOMEM;
1635
1636	/* Get a new skb... force flag on. */
1637	buff = tcp_stream_alloc_skb(sk, gfp, true);
1638	if (!buff)
1639		return -ENOMEM; /* We'll just try again later. */
1640	skb_copy_decrypted(buff, skb);
1641	mptcp_skb_ext_copy(buff, skb);
1642
1643	sk_wmem_queued_add(sk, buff->truesize);
1644	sk_mem_charge(sk, buff->truesize);
1645	nlen = skb->len - len;
1646	buff->truesize += nlen;
1647	skb->truesize -= nlen;
1648
1649	/* Correct the sequence numbers. */
1650	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1651	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1652	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1653
1654	/* PSH and FIN should only be set in the second packet. */
1655	flags = TCP_SKB_CB(skb)->tcp_flags;
1656	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1657	TCP_SKB_CB(buff)->tcp_flags = flags;
1658	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1659	tcp_skb_fragment_eor(skb, buff);
1660
1661	skb_split(skb, buff, len);
 
 
 
 
1662
1663	skb_set_delivery_time(buff, skb->tstamp, SKB_CLOCK_MONOTONIC);
1664	tcp_fragment_tstamp(skb, buff);
 
 
 
 
 
 
 
 
 
 
 
 
 
1665
1666	old_factor = tcp_skb_pcount(skb);
1667
1668	/* Fix up tso_factor for both original and new SKB.  */
1669	tcp_set_skb_tso_segs(skb, mss_now);
1670	tcp_set_skb_tso_segs(buff, mss_now);
1671
1672	/* Update delivered info for the new segment */
1673	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1674
1675	/* If this packet has been sent out already, we must
1676	 * adjust the various packet counters.
1677	 */
1678	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1679		int diff = old_factor - tcp_skb_pcount(skb) -
1680			tcp_skb_pcount(buff);
1681
1682		if (diff)
1683			tcp_adjust_pcount(sk, skb, diff);
1684	}
1685
1686	/* Link BUFF into the send queue. */
1687	__skb_header_release(buff);
1688	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1689	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1690		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1691
1692	return 0;
1693}
1694
1695/* This is similar to __pskb_pull_tail(). The difference is that pulled
1696 * data is not copied, but immediately discarded.
 
1697 */
1698static int __pskb_trim_head(struct sk_buff *skb, int len)
1699{
1700	struct skb_shared_info *shinfo;
1701	int i, k, eat;
1702
1703	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1704	eat = len;
1705	k = 0;
1706	shinfo = skb_shinfo(skb);
1707	for (i = 0; i < shinfo->nr_frags; i++) {
1708		int size = skb_frag_size(&shinfo->frags[i]);
1709
1710		if (size <= eat) {
1711			skb_frag_unref(skb, i);
1712			eat -= size;
1713		} else {
1714			shinfo->frags[k] = shinfo->frags[i];
1715			if (eat) {
1716				skb_frag_off_add(&shinfo->frags[k], eat);
1717				skb_frag_size_sub(&shinfo->frags[k], eat);
1718				eat = 0;
1719			}
1720			k++;
1721		}
1722	}
1723	shinfo->nr_frags = k;
1724
 
1725	skb->data_len -= len;
1726	skb->len = skb->data_len;
1727	return len;
1728}
1729
1730/* Remove acked data from a packet in the transmit queue. */
1731int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1732{
1733	u32 delta_truesize;
1734
1735	if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1736		return -ENOMEM;
1737
1738	delta_truesize = __pskb_trim_head(skb, len);
 
 
 
 
1739
1740	TCP_SKB_CB(skb)->seq += len;
 
1741
1742	skb->truesize	   -= delta_truesize;
1743	sk_wmem_queued_add(sk, -delta_truesize);
1744	if (!skb_zcopy_pure(skb))
1745		sk_mem_uncharge(sk, delta_truesize);
1746
1747	/* Any change of skb->len requires recalculation of tso factor. */
 
 
1748	if (tcp_skb_pcount(skb) > 1)
1749		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1750
1751	return 0;
1752}
1753
1754/* Calculate MSS not accounting any TCP options.  */
1755static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1756{
1757	const struct tcp_sock *tp = tcp_sk(sk);
1758	const struct inet_connection_sock *icsk = inet_csk(sk);
1759	int mss_now;
1760
1761	/* Calculate base mss without TCP options:
1762	   It is MMS_S - sizeof(tcphdr) of rfc1122
1763	 */
1764	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1765
1766	/* Clamp it (mss_clamp does not include tcp options) */
1767	if (mss_now > tp->rx_opt.mss_clamp)
1768		mss_now = tp->rx_opt.mss_clamp;
1769
1770	/* Now subtract optional transport overhead */
1771	mss_now -= icsk->icsk_ext_hdr_len;
1772
1773	/* Then reserve room for full set of TCP options and 8 bytes of data */
1774	mss_now = max(mss_now,
1775		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1776	return mss_now;
1777}
1778
1779/* Calculate MSS. Not accounting for SACKs here.  */
1780int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1781{
1782	/* Subtract TCP options size, not including SACKs */
1783	return __tcp_mtu_to_mss(sk, pmtu) -
1784	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1785}
1786EXPORT_SYMBOL(tcp_mtu_to_mss);
1787
1788/* Inverse of above */
1789int tcp_mss_to_mtu(struct sock *sk, int mss)
1790{
1791	const struct tcp_sock *tp = tcp_sk(sk);
1792	const struct inet_connection_sock *icsk = inet_csk(sk);
 
1793
1794	return mss +
1795	      tp->tcp_header_len +
1796	      icsk->icsk_ext_hdr_len +
1797	      icsk->icsk_af_ops->net_header_len;
 
 
1798}
1799EXPORT_SYMBOL(tcp_mss_to_mtu);
1800
1801/* MTU probing init per socket */
1802void tcp_mtup_init(struct sock *sk)
1803{
1804	struct tcp_sock *tp = tcp_sk(sk);
1805	struct inet_connection_sock *icsk = inet_csk(sk);
1806	struct net *net = sock_net(sk);
1807
1808	icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1809	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1810			       icsk->icsk_af_ops->net_header_len;
1811	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1812	icsk->icsk_mtup.probe_size = 0;
1813	if (icsk->icsk_mtup.enabled)
1814		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1815}
1816EXPORT_SYMBOL(tcp_mtup_init);
1817
1818/* This function synchronize snd mss to current pmtu/exthdr set.
1819
1820   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1821   for TCP options, but includes only bare TCP header.
1822
1823   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1824   It is minimum of user_mss and mss received with SYN.
1825   It also does not include TCP options.
1826
1827   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1828
1829   tp->mss_cache is current effective sending mss, including
1830   all tcp options except for SACKs. It is evaluated,
1831   taking into account current pmtu, but never exceeds
1832   tp->rx_opt.mss_clamp.
1833
1834   NOTE1. rfc1122 clearly states that advertised MSS
1835   DOES NOT include either tcp or ip options.
1836
1837   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1838   are READ ONLY outside this function.		--ANK (980731)
1839 */
1840unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1841{
1842	struct tcp_sock *tp = tcp_sk(sk);
1843	struct inet_connection_sock *icsk = inet_csk(sk);
1844	int mss_now;
1845
1846	if (icsk->icsk_mtup.search_high > pmtu)
1847		icsk->icsk_mtup.search_high = pmtu;
1848
1849	mss_now = tcp_mtu_to_mss(sk, pmtu);
1850	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1851
1852	/* And store cached results */
1853	icsk->icsk_pmtu_cookie = pmtu;
1854	if (icsk->icsk_mtup.enabled)
1855		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1856	tp->mss_cache = mss_now;
1857
1858	return mss_now;
1859}
1860EXPORT_SYMBOL(tcp_sync_mss);
1861
1862/* Compute the current effective MSS, taking SACKs and IP options,
1863 * and even PMTU discovery events into account.
1864 */
1865unsigned int tcp_current_mss(struct sock *sk)
1866{
1867	const struct tcp_sock *tp = tcp_sk(sk);
1868	const struct dst_entry *dst = __sk_dst_get(sk);
1869	u32 mss_now;
1870	unsigned int header_len;
1871	struct tcp_out_options opts;
1872	struct tcp_key key;
1873
1874	mss_now = tp->mss_cache;
1875
1876	if (dst) {
1877		u32 mtu = dst_mtu(dst);
1878		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1879			mss_now = tcp_sync_mss(sk, mtu);
1880	}
1881	tcp_get_current_key(sk, &key);
1882	header_len = tcp_established_options(sk, NULL, &opts, &key) +
1883		     sizeof(struct tcphdr);
1884	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1885	 * some common options. If this is an odd packet (because we have SACK
1886	 * blocks etc) then our calculated header_len will be different, and
1887	 * we have to adjust mss_now correspondingly */
1888	if (header_len != tp->tcp_header_len) {
1889		int delta = (int) header_len - tp->tcp_header_len;
1890		mss_now -= delta;
1891	}
1892
1893	return mss_now;
1894}
1895
1896/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1897 * As additional protections, we do not touch cwnd in retransmission phases,
1898 * and if application hit its sndbuf limit recently.
1899 */
1900static void tcp_cwnd_application_limited(struct sock *sk)
1901{
1902	struct tcp_sock *tp = tcp_sk(sk);
1903
1904	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1905	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1906		/* Limited by application or receiver window. */
1907		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1908		u32 win_used = max(tp->snd_cwnd_used, init_win);
1909		if (win_used < tcp_snd_cwnd(tp)) {
1910			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1911			tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1912		}
1913		tp->snd_cwnd_used = 0;
1914	}
1915	tp->snd_cwnd_stamp = tcp_jiffies32;
1916}
1917
1918static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1919{
1920	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1921	struct tcp_sock *tp = tcp_sk(sk);
1922
1923	/* Track the strongest available signal of the degree to which the cwnd
1924	 * is fully utilized. If cwnd-limited then remember that fact for the
1925	 * current window. If not cwnd-limited then track the maximum number of
1926	 * outstanding packets in the current window. (If cwnd-limited then we
1927	 * chose to not update tp->max_packets_out to avoid an extra else
1928	 * clause with no functional impact.)
1929	 */
1930	if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1931	    is_cwnd_limited ||
1932	    (!tp->is_cwnd_limited &&
1933	     tp->packets_out > tp->max_packets_out)) {
1934		tp->is_cwnd_limited = is_cwnd_limited;
1935		tp->max_packets_out = tp->packets_out;
1936		tp->cwnd_usage_seq = tp->snd_nxt;
1937	}
1938
1939	if (tcp_is_cwnd_limited(sk)) {
1940		/* Network is feed fully. */
1941		tp->snd_cwnd_used = 0;
1942		tp->snd_cwnd_stamp = tcp_jiffies32;
1943	} else {
1944		/* Network starves. */
1945		if (tp->packets_out > tp->snd_cwnd_used)
1946			tp->snd_cwnd_used = tp->packets_out;
1947
1948		if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1949		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1950		    !ca_ops->cong_control)
1951			tcp_cwnd_application_limited(sk);
1952
1953		/* The following conditions together indicate the starvation
1954		 * is caused by insufficient sender buffer:
1955		 * 1) just sent some data (see tcp_write_xmit)
1956		 * 2) not cwnd limited (this else condition)
1957		 * 3) no more data to send (tcp_write_queue_empty())
1958		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1959		 */
1960		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1961		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1962		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1963			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1964	}
1965}
1966
1967/* Minshall's variant of the Nagle send check. */
1968static bool tcp_minshall_check(const struct tcp_sock *tp)
1969{
1970	return after(tp->snd_sml, tp->snd_una) &&
1971		!after(tp->snd_sml, tp->snd_nxt);
1972}
1973
1974/* Update snd_sml if this skb is under mss
1975 * Note that a TSO packet might end with a sub-mss segment
1976 * The test is really :
1977 * if ((skb->len % mss) != 0)
1978 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1979 * But we can avoid doing the divide again given we already have
1980 *  skb_pcount = skb->len / mss_now
1981 */
1982static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1983				const struct sk_buff *skb)
1984{
1985	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1986		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1987}
1988
1989/* Return false, if packet can be sent now without violation Nagle's rules:
1990 * 1. It is full sized. (provided by caller in %partial bool)
1991 * 2. Or it contains FIN. (already checked by caller)
1992 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1993 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1994 *    With Minshall's modification: all sent small packets are ACKed.
1995 */
1996static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1997			    int nonagle)
1998{
1999	return partial &&
2000		((nonagle & TCP_NAGLE_CORK) ||
2001		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
2002}
2003
2004/* Return how many segs we'd like on a TSO packet,
2005 * depending on current pacing rate, and how close the peer is.
2006 *
2007 * Rationale is:
2008 * - For close peers, we rather send bigger packets to reduce
2009 *   cpu costs, because occasional losses will be repaired fast.
2010 * - For long distance/rtt flows, we would like to get ACK clocking
2011 *   with 1 ACK per ms.
2012 *
2013 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2014 * in bigger TSO bursts. We we cut the RTT-based allowance in half
2015 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2016 * is below 1500 bytes after 6 * ~500 usec = 3ms.
2017 */
2018static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2019			    int min_tso_segs)
2020{
2021	unsigned long bytes;
2022	u32 r;
2023
2024	bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2025
2026	r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2027	if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2028		bytes += sk->sk_gso_max_size >> r;
2029
2030	bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2031
2032	return max_t(u32, bytes / mss_now, min_tso_segs);
2033}
2034
2035/* Return the number of segments we want in the skb we are transmitting.
2036 * See if congestion control module wants to decide; otherwise, autosize.
2037 */
2038static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2039{
2040	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2041	u32 min_tso, tso_segs;
2042
2043	min_tso = ca_ops->min_tso_segs ?
2044			ca_ops->min_tso_segs(sk) :
2045			READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2046
2047	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2048	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2049}
2050
2051/* Returns the portion of skb which can be sent right away */
2052static unsigned int tcp_mss_split_point(const struct sock *sk,
2053					const struct sk_buff *skb,
2054					unsigned int mss_now,
2055					unsigned int max_segs,
2056					int nonagle)
2057{
2058	const struct tcp_sock *tp = tcp_sk(sk);
2059	u32 partial, needed, window, max_len;
2060
2061	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2062	max_len = mss_now * max_segs;
2063
2064	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2065		return max_len;
2066
2067	needed = min(skb->len, window);
2068
2069	if (max_len <= needed)
2070		return max_len;
2071
2072	partial = needed % mss_now;
2073	/* If last segment is not a full MSS, check if Nagle rules allow us
2074	 * to include this last segment in this skb.
2075	 * Otherwise, we'll split the skb at last MSS boundary
2076	 */
2077	if (tcp_nagle_check(partial != 0, tp, nonagle))
2078		return needed - partial;
2079
2080	return needed;
2081}
2082
2083/* Can at least one segment of SKB be sent right now, according to the
2084 * congestion window rules?  If so, return how many segments are allowed.
2085 */
2086static u32 tcp_cwnd_test(const struct tcp_sock *tp)
 
2087{
2088	u32 in_flight, cwnd, halfcwnd;
 
 
 
 
2089
2090	in_flight = tcp_packets_in_flight(tp);
2091	cwnd = tcp_snd_cwnd(tp);
2092	if (in_flight >= cwnd)
2093		return 0;
2094
2095	/* For better scheduling, ensure we have at least
2096	 * 2 GSO packets in flight.
2097	 */
2098	halfcwnd = max(cwnd >> 1, 1U);
2099	return min(halfcwnd, cwnd - in_flight);
2100}
2101
2102/* Initialize TSO state of a skb.
2103 * This must be invoked the first time we consider transmitting
2104 * SKB onto the wire.
2105 */
2106static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
 
2107{
2108	int tso_segs = tcp_skb_pcount(skb);
2109
2110	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now))
2111		return tcp_set_skb_tso_segs(skb, mss_now);
2112
 
2113	return tso_segs;
2114}
2115
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2116
2117/* Return true if the Nagle test allows this packet to be
2118 * sent now.
2119 */
2120static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2121				  unsigned int cur_mss, int nonagle)
2122{
2123	/* Nagle rule does not apply to frames, which sit in the middle of the
2124	 * write_queue (they have no chances to get new data).
2125	 *
2126	 * This is implemented in the callers, where they modify the 'nonagle'
2127	 * argument based upon the location of SKB in the send queue.
2128	 */
2129	if (nonagle & TCP_NAGLE_PUSH)
2130		return true;
2131
2132	/* Don't use the nagle rule for urgent data (or for the final FIN). */
2133	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2134		return true;
 
 
 
2135
2136	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2137		return true;
2138
2139	return false;
2140}
2141
2142/* Does at least the first segment of SKB fit into the send window? */
2143static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2144			     const struct sk_buff *skb,
2145			     unsigned int cur_mss)
2146{
2147	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2148
2149	if (skb->len > cur_mss)
2150		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2151
2152	return !after(end_seq, tcp_wnd_end(tp));
2153}
2154
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2155/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2156 * which is put after SKB on the list.  It is very much like
2157 * tcp_fragment() except that it may make several kinds of assumptions
2158 * in order to speed up the splitting operation.  In particular, we
2159 * know that all the data is in scatter-gather pages, and that the
2160 * packet has never been sent out before (and thus is not cloned).
2161 */
2162static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2163			unsigned int mss_now, gfp_t gfp)
2164{
2165	int nlen = skb->len - len;
2166	struct sk_buff *buff;
 
2167	u8 flags;
2168
2169	/* All of a TSO frame must be composed of paged data.  */
2170	DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
 
2171
2172	buff = tcp_stream_alloc_skb(sk, gfp, true);
2173	if (unlikely(!buff))
2174		return -ENOMEM;
2175	skb_copy_decrypted(buff, skb);
2176	mptcp_skb_ext_copy(buff, skb);
2177
2178	sk_wmem_queued_add(sk, buff->truesize);
2179	sk_mem_charge(sk, buff->truesize);
2180	buff->truesize += nlen;
2181	skb->truesize -= nlen;
2182
2183	/* Correct the sequence numbers. */
2184	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2185	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2186	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2187
2188	/* PSH and FIN should only be set in the second packet. */
2189	flags = TCP_SKB_CB(skb)->tcp_flags;
2190	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2191	TCP_SKB_CB(buff)->tcp_flags = flags;
2192
2193	tcp_skb_fragment_eor(skb, buff);
 
2194
 
2195	skb_split(skb, buff, len);
2196	tcp_fragment_tstamp(skb, buff);
2197
2198	/* Fix up tso_factor for both original and new SKB.  */
2199	tcp_set_skb_tso_segs(skb, mss_now);
2200	tcp_set_skb_tso_segs(buff, mss_now);
2201
2202	/* Link BUFF into the send queue. */
2203	__skb_header_release(buff);
2204	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2205
2206	return 0;
2207}
2208
2209/* Try to defer sending, if possible, in order to minimize the amount
2210 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
2211 *
2212 * This algorithm is from John Heffner.
2213 */
2214static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2215				 bool *is_cwnd_limited,
2216				 bool *is_rwnd_limited,
2217				 u32 max_segs)
2218{
 
2219	const struct inet_connection_sock *icsk = inet_csk(sk);
2220	u32 send_win, cong_win, limit, in_flight;
2221	struct tcp_sock *tp = tcp_sk(sk);
2222	struct sk_buff *head;
2223	int win_divisor;
2224	s64 delta;
2225
2226	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2227		goto send_now;
2228
2229	/* Avoid bursty behavior by allowing defer
2230	 * only if the last write was recent (1 ms).
2231	 * Note that tp->tcp_wstamp_ns can be in the future if we have
2232	 * packets waiting in a qdisc or device for EDT delivery.
2233	 */
2234	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2235	if (delta > 0)
2236		goto send_now;
2237
2238	in_flight = tcp_packets_in_flight(tp);
2239
2240	BUG_ON(tcp_skb_pcount(skb) <= 1);
2241	BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2242
2243	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2244
2245	/* From in_flight test above, we know that cwnd > in_flight.  */
2246	cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2247
2248	limit = min(send_win, cong_win);
2249
2250	/* If a full-sized TSO skb can be sent, do it. */
2251	if (limit >= max_segs * tp->mss_cache)
2252		goto send_now;
2253
2254	/* Middle in queue won't get any more data, full sendable already? */
2255	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2256		goto send_now;
2257
2258	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2259	if (win_divisor) {
2260		u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2261
2262		/* If at least some fraction of a window is available,
2263		 * just use it.
2264		 */
2265		chunk /= win_divisor;
2266		if (limit >= chunk)
2267			goto send_now;
2268	} else {
2269		/* Different approach, try not to defer past a single
2270		 * ACK.  Receiver should ACK every other full sized
2271		 * frame, so if we have space for more than 3 frames
2272		 * then send now.
2273		 */
2274		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2275			goto send_now;
2276	}
2277
2278	/* TODO : use tsorted_sent_queue ? */
2279	head = tcp_rtx_queue_head(sk);
2280	if (!head)
2281		goto send_now;
2282	delta = tp->tcp_clock_cache - head->tstamp;
2283	/* If next ACK is likely to come too late (half srtt), do not defer */
2284	if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2285		goto send_now;
2286
2287	/* Ok, it looks like it is advisable to defer.
2288	 * Three cases are tracked :
2289	 * 1) We are cwnd-limited
2290	 * 2) We are rwnd-limited
2291	 * 3) We are application limited.
2292	 */
2293	if (cong_win < send_win) {
2294		if (cong_win <= skb->len) {
2295			*is_cwnd_limited = true;
2296			return true;
2297		}
2298	} else {
2299		if (send_win <= skb->len) {
2300			*is_rwnd_limited = true;
2301			return true;
2302		}
2303	}
2304
2305	/* If this packet won't get more data, do not wait. */
2306	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2307	    TCP_SKB_CB(skb)->eor)
2308		goto send_now;
2309
2310	return true;
2311
2312send_now:
2313	return false;
2314}
2315
2316static inline void tcp_mtu_check_reprobe(struct sock *sk)
2317{
2318	struct inet_connection_sock *icsk = inet_csk(sk);
2319	struct tcp_sock *tp = tcp_sk(sk);
2320	struct net *net = sock_net(sk);
2321	u32 interval;
2322	s32 delta;
2323
2324	interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2325	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2326	if (unlikely(delta >= interval * HZ)) {
2327		int mss = tcp_current_mss(sk);
2328
2329		/* Update current search range */
2330		icsk->icsk_mtup.probe_size = 0;
2331		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2332			sizeof(struct tcphdr) +
2333			icsk->icsk_af_ops->net_header_len;
2334		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2335
2336		/* Update probe time stamp */
2337		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2338	}
2339}
2340
2341static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2342{
2343	struct sk_buff *skb, *next;
2344
2345	skb = tcp_send_head(sk);
2346	tcp_for_write_queue_from_safe(skb, next, sk) {
2347		if (len <= skb->len)
2348			break;
2349
2350		if (tcp_has_tx_tstamp(skb) || !tcp_skb_can_collapse(skb, next))
2351			return false;
2352
2353		len -= skb->len;
2354	}
2355
2356	return true;
2357}
2358
2359static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2360			     int probe_size)
2361{
2362	skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2363	int i, todo, len = 0, nr_frags = 0;
2364	const struct sk_buff *skb;
2365
2366	if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2367		return -ENOMEM;
2368
2369	skb_queue_walk(&sk->sk_write_queue, skb) {
2370		const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2371
2372		if (skb_headlen(skb))
2373			return -EINVAL;
2374
2375		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2376			if (len >= probe_size)
2377				goto commit;
2378			todo = min_t(int, skb_frag_size(fragfrom),
2379				     probe_size - len);
2380			len += todo;
2381			if (lastfrag &&
2382			    skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2383			    skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2384						      skb_frag_size(lastfrag)) {
2385				skb_frag_size_add(lastfrag, todo);
2386				continue;
2387			}
2388			if (unlikely(nr_frags == MAX_SKB_FRAGS))
2389				return -E2BIG;
2390			skb_frag_page_copy(fragto, fragfrom);
2391			skb_frag_off_copy(fragto, fragfrom);
2392			skb_frag_size_set(fragto, todo);
2393			nr_frags++;
2394			lastfrag = fragto++;
2395		}
2396	}
2397commit:
2398	WARN_ON_ONCE(len != probe_size);
2399	for (i = 0; i < nr_frags; i++)
2400		skb_frag_ref(to, i);
2401
2402	skb_shinfo(to)->nr_frags = nr_frags;
2403	to->truesize += probe_size;
2404	to->len += probe_size;
2405	to->data_len += probe_size;
2406	__skb_header_release(to);
2407	return 0;
2408}
2409
2410/* tcp_mtu_probe() and tcp_grow_skb() can both eat an skb (src) if
2411 * all its payload was moved to another one (dst).
2412 * Make sure to transfer tcp_flags, eor, and tstamp.
2413 */
2414static void tcp_eat_one_skb(struct sock *sk,
2415			    struct sk_buff *dst,
2416			    struct sk_buff *src)
2417{
2418	TCP_SKB_CB(dst)->tcp_flags |= TCP_SKB_CB(src)->tcp_flags;
2419	TCP_SKB_CB(dst)->eor = TCP_SKB_CB(src)->eor;
2420	tcp_skb_collapse_tstamp(dst, src);
2421	tcp_unlink_write_queue(src, sk);
2422	tcp_wmem_free_skb(sk, src);
2423}
2424
2425/* Create a new MTU probe if we are ready.
2426 * MTU probe is regularly attempting to increase the path MTU by
2427 * deliberately sending larger packets.  This discovers routing
2428 * changes resulting in larger path MTUs.
2429 *
2430 * Returns 0 if we should wait to probe (no cwnd available),
2431 *         1 if a probe was sent,
2432 *         -1 otherwise
2433 */
2434static int tcp_mtu_probe(struct sock *sk)
2435{
2436	struct inet_connection_sock *icsk = inet_csk(sk);
2437	struct tcp_sock *tp = tcp_sk(sk);
 
2438	struct sk_buff *skb, *nskb, *next;
2439	struct net *net = sock_net(sk);
2440	int probe_size;
2441	int size_needed;
2442	int copy, len;
2443	int mss_now;
2444	int interval;
2445
2446	/* Not currently probing/verifying,
2447	 * not in recovery,
2448	 * have enough cwnd, and
2449	 * not SACKing (the variable headers throw things off)
2450	 */
2451	if (likely(!icsk->icsk_mtup.enabled ||
2452		   icsk->icsk_mtup.probe_size ||
2453		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2454		   tcp_snd_cwnd(tp) < 11 ||
2455		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2456		return -1;
2457
2458	/* Use binary search for probe_size between tcp_mss_base,
2459	 * and current mss_clamp. if (search_high - search_low)
2460	 * smaller than a threshold, backoff from probing.
2461	 */
2462	mss_now = tcp_current_mss(sk);
2463	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2464				    icsk->icsk_mtup.search_low) >> 1);
2465	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2466	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2467	/* When misfortune happens, we are reprobing actively,
2468	 * and then reprobe timer has expired. We stick with current
2469	 * probing process by not resetting search range to its orignal.
2470	 */
2471	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2472	    interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2473		/* Check whether enough time has elaplased for
2474		 * another round of probing.
2475		 */
2476		tcp_mtu_check_reprobe(sk);
2477		return -1;
2478	}
2479
2480	/* Have enough data in the send queue to probe? */
2481	if (tp->write_seq - tp->snd_nxt < size_needed)
2482		return -1;
2483
2484	if (tp->snd_wnd < size_needed)
2485		return -1;
2486	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2487		return 0;
2488
2489	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2490	if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2491		if (!tcp_packets_in_flight(tp))
2492			return -1;
2493		else
2494			return 0;
2495	}
2496
2497	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2498		return -1;
2499
2500	/* We're allowed to probe.  Build it now. */
2501	nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2502	if (!nskb)
2503		return -1;
2504
2505	/* build the payload, and be prepared to abort if this fails. */
2506	if (tcp_clone_payload(sk, nskb, probe_size)) {
2507		tcp_skb_tsorted_anchor_cleanup(nskb);
2508		consume_skb(nskb);
2509		return -1;
2510	}
2511	sk_wmem_queued_add(sk, nskb->truesize);
2512	sk_mem_charge(sk, nskb->truesize);
2513
2514	skb = tcp_send_head(sk);
2515	skb_copy_decrypted(nskb, skb);
2516	mptcp_skb_ext_copy(nskb, skb);
2517
2518	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2519	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2520	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
 
 
 
2521
2522	tcp_insert_write_queue_before(nskb, skb, sk);
2523	tcp_highest_sack_replace(sk, skb, nskb);
2524
2525	len = 0;
2526	tcp_for_write_queue_from_safe(skb, next, sk) {
2527		copy = min_t(int, skb->len, probe_size - len);
 
 
 
 
 
 
2528
2529		if (skb->len <= copy) {
2530			tcp_eat_one_skb(sk, nskb, skb);
 
 
 
 
2531		} else {
2532			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2533						   ~(TCPHDR_FIN|TCPHDR_PSH);
2534			__pskb_trim_head(skb, copy);
2535			tcp_set_skb_tso_segs(skb, mss_now);
 
 
 
 
 
 
 
2536			TCP_SKB_CB(skb)->seq += copy;
2537		}
2538
2539		len += copy;
2540
2541		if (len >= probe_size)
2542			break;
2543	}
2544	tcp_init_tso_segs(nskb, nskb->len);
2545
2546	/* We're ready to send.  If this fails, the probe will
2547	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2548	 */
2549	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2550		/* Decrement cwnd here because we are sending
2551		 * effectively two packets. */
2552		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2553		tcp_event_new_data_sent(sk, nskb);
2554
2555		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2556		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2557		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2558
2559		return 1;
2560	}
2561
2562	return -1;
2563}
2564
2565static bool tcp_pacing_check(struct sock *sk)
2566{
2567	struct tcp_sock *tp = tcp_sk(sk);
2568
2569	if (!tcp_needs_internal_pacing(sk))
2570		return false;
2571
2572	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2573		return false;
2574
2575	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2576		hrtimer_start(&tp->pacing_timer,
2577			      ns_to_ktime(tp->tcp_wstamp_ns),
2578			      HRTIMER_MODE_ABS_PINNED_SOFT);
2579		sock_hold(sk);
2580	}
2581	return true;
2582}
2583
2584static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2585{
2586	const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2587
2588	/* No skb in the rtx queue. */
2589	if (!node)
2590		return true;
2591
2592	/* Only one skb in rtx queue. */
2593	return !node->rb_left && !node->rb_right;
2594}
2595
2596/* TCP Small Queues :
2597 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2598 * (These limits are doubled for retransmits)
2599 * This allows for :
2600 *  - better RTT estimation and ACK scheduling
2601 *  - faster recovery
2602 *  - high rates
2603 * Alas, some drivers / subsystems require a fair amount
2604 * of queued bytes to ensure line rate.
2605 * One example is wifi aggregation (802.11 AMPDU)
2606 */
2607static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2608				  unsigned int factor)
2609{
2610	unsigned long limit;
2611
2612	limit = max_t(unsigned long,
2613		      2 * skb->truesize,
2614		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2615	if (sk->sk_pacing_status == SK_PACING_NONE)
2616		limit = min_t(unsigned long, limit,
2617			      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2618	limit <<= factor;
2619
2620	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2621	    tcp_sk(sk)->tcp_tx_delay) {
2622		u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2623				  tcp_sk(sk)->tcp_tx_delay;
2624
2625		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2626		 * approximate our needs assuming an ~100% skb->truesize overhead.
2627		 * USEC_PER_SEC is approximated by 2^20.
2628		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2629		 */
2630		extra_bytes >>= (20 - 1);
2631		limit += extra_bytes;
2632	}
2633	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2634		/* Always send skb if rtx queue is empty or has one skb.
2635		 * No need to wait for TX completion to call us back,
2636		 * after softirq/tasklet schedule.
2637		 * This helps when TX completions are delayed too much.
2638		 */
2639		if (tcp_rtx_queue_empty_or_single_skb(sk))
2640			return false;
2641
2642		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2643		/* It is possible TX completion already happened
2644		 * before we set TSQ_THROTTLED, so we must
2645		 * test again the condition.
2646		 */
2647		smp_mb__after_atomic();
2648		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2649			return true;
2650	}
2651	return false;
2652}
2653
2654static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2655{
2656	const u32 now = tcp_jiffies32;
2657	enum tcp_chrono old = tp->chrono_type;
2658
2659	if (old > TCP_CHRONO_UNSPEC)
2660		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2661	tp->chrono_start = now;
2662	tp->chrono_type = new;
2663}
2664
2665void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2666{
2667	struct tcp_sock *tp = tcp_sk(sk);
2668
2669	/* If there are multiple conditions worthy of tracking in a
2670	 * chronograph then the highest priority enum takes precedence
2671	 * over the other conditions. So that if something "more interesting"
2672	 * starts happening, stop the previous chrono and start a new one.
2673	 */
2674	if (type > tp->chrono_type)
2675		tcp_chrono_set(tp, type);
2676}
2677
2678void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2679{
2680	struct tcp_sock *tp = tcp_sk(sk);
2681
2682
2683	/* There are multiple conditions worthy of tracking in a
2684	 * chronograph, so that the highest priority enum takes
2685	 * precedence over the other conditions (see tcp_chrono_start).
2686	 * If a condition stops, we only stop chrono tracking if
2687	 * it's the "most interesting" or current chrono we are
2688	 * tracking and starts busy chrono if we have pending data.
2689	 */
2690	if (tcp_rtx_and_write_queues_empty(sk))
2691		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2692	else if (type == tp->chrono_type)
2693		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2694}
2695
2696/* First skb in the write queue is smaller than ideal packet size.
2697 * Check if we can move payload from the second skb in the queue.
2698 */
2699static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount)
2700{
2701	struct sk_buff *next_skb = skb->next;
2702	unsigned int nlen;
2703
2704	if (tcp_skb_is_last(sk, skb))
2705		return;
2706
2707	if (!tcp_skb_can_collapse(skb, next_skb))
2708		return;
2709
2710	nlen = min_t(u32, amount, next_skb->len);
2711	if (!nlen || !skb_shift(skb, next_skb, nlen))
2712		return;
2713
2714	TCP_SKB_CB(skb)->end_seq += nlen;
2715	TCP_SKB_CB(next_skb)->seq += nlen;
2716
2717	if (!next_skb->len) {
2718		/* In case FIN is set, we need to update end_seq */
2719		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2720
2721		tcp_eat_one_skb(sk, skb, next_skb);
2722	}
2723}
2724
2725/* This routine writes packets to the network.  It advances the
2726 * send_head.  This happens as incoming acks open up the remote
2727 * window for us.
2728 *
2729 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2730 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2731 * account rare use of URG, this is not a big flaw.
2732 *
2733 * Send at most one packet when push_one > 0. Temporarily ignore
2734 * cwnd limit to force at most one packet out when push_one == 2.
2735
2736 * Returns true, if no segments are in flight and we have queued segments,
2737 * but cannot send anything now because of SWS or another problem.
2738 */
2739static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2740			   int push_one, gfp_t gfp)
2741{
2742	struct tcp_sock *tp = tcp_sk(sk);
2743	struct sk_buff *skb;
2744	unsigned int tso_segs, sent_pkts;
2745	u32 cwnd_quota, max_segs;
2746	int result;
2747	bool is_cwnd_limited = false, is_rwnd_limited = false;
2748
2749	sent_pkts = 0;
2750
2751	tcp_mstamp_refresh(tp);
2752	if (!push_one) {
2753		/* Do MTU probing. */
2754		result = tcp_mtu_probe(sk);
2755		if (!result) {
2756			return false;
2757		} else if (result > 0) {
2758			sent_pkts = 1;
2759		}
2760	}
2761
2762	max_segs = tcp_tso_segs(sk, mss_now);
2763	while ((skb = tcp_send_head(sk))) {
2764		unsigned int limit;
2765		int missing_bytes;
2766
2767		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2768			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2769			tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2770			skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
2771			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2772			tcp_init_tso_segs(skb, mss_now);
2773			goto repair; /* Skip network transmission */
2774		}
2775
2776		if (tcp_pacing_check(sk))
 
2777			break;
2778
2779		cwnd_quota = tcp_cwnd_test(tp);
2780		if (!cwnd_quota) {
2781			if (push_one == 2)
2782				/* Force out a loss probe pkt. */
2783				cwnd_quota = 1;
2784			else
2785				break;
2786		}
2787		cwnd_quota = min(cwnd_quota, max_segs);
2788		missing_bytes = cwnd_quota * mss_now - skb->len;
2789		if (missing_bytes > 0)
2790			tcp_grow_skb(sk, skb, missing_bytes);
2791
2792		tso_segs = tcp_set_skb_tso_segs(skb, mss_now);
2793
2794		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2795			is_rwnd_limited = true;
2796			break;
2797		}
2798
2799		if (tso_segs == 1) {
2800			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2801						     (tcp_skb_is_last(sk, skb) ?
2802						      nonagle : TCP_NAGLE_PUSH))))
2803				break;
2804		} else {
2805			if (!push_one &&
2806			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2807						 &is_rwnd_limited, max_segs))
2808				break;
2809		}
2810
2811		limit = mss_now;
2812		if (tso_segs > 1 && !tcp_urg_mode(tp))
2813			limit = tcp_mss_split_point(sk, skb, mss_now,
2814						    cwnd_quota,
2815						    nonagle);
2816
2817		if (skb->len > limit &&
2818		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2819			break;
2820
2821		if (tcp_small_queue_check(sk, skb, 0))
2822			break;
2823
2824		/* Argh, we hit an empty skb(), presumably a thread
2825		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2826		 * We do not want to send a pure-ack packet and have
2827		 * a strange looking rtx queue with empty packet(s).
2828		 */
2829		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2830			break;
2831
2832		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2833			break;
2834
2835repair:
2836		/* Advance the send_head.  This one is sent out.
2837		 * This call will increment packets_out.
2838		 */
2839		tcp_event_new_data_sent(sk, skb);
2840
2841		tcp_minshall_update(tp, mss_now, skb);
2842		sent_pkts += tcp_skb_pcount(skb);
2843
2844		if (push_one)
2845			break;
2846	}
2847
2848	if (is_rwnd_limited)
2849		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2850	else
2851		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2852
2853	is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2854	if (likely(sent_pkts || is_cwnd_limited))
2855		tcp_cwnd_validate(sk, is_cwnd_limited);
2856
2857	if (likely(sent_pkts)) {
2858		if (tcp_in_cwnd_reduction(sk))
2859			tp->prr_out += sent_pkts;
2860
2861		/* Send one loss probe per tail loss episode. */
2862		if (push_one != 2)
2863			tcp_schedule_loss_probe(sk, false);
2864		return false;
2865	}
2866	return !tp->packets_out && !tcp_write_queue_empty(sk);
2867}
2868
2869bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2870{
2871	struct inet_connection_sock *icsk = inet_csk(sk);
2872	struct tcp_sock *tp = tcp_sk(sk);
2873	u32 timeout, timeout_us, rto_delta_us;
2874	int early_retrans;
2875
2876	/* Don't do any loss probe on a Fast Open connection before 3WHS
2877	 * finishes.
2878	 */
2879	if (rcu_access_pointer(tp->fastopen_rsk))
2880		return false;
2881
2882	early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2883	/* Schedule a loss probe in 2*RTT for SACK capable connections
2884	 * not in loss recovery, that are either limited by cwnd or application.
2885	 */
2886	if ((early_retrans != 3 && early_retrans != 4) ||
2887	    !tp->packets_out || !tcp_is_sack(tp) ||
2888	    (icsk->icsk_ca_state != TCP_CA_Open &&
2889	     icsk->icsk_ca_state != TCP_CA_CWR))
2890		return false;
2891
2892	/* Probe timeout is 2*rtt. Add minimum RTO to account
2893	 * for delayed ack when there's one outstanding packet. If no RTT
2894	 * sample is available then probe after TCP_TIMEOUT_INIT.
2895	 */
2896	if (tp->srtt_us) {
2897		timeout_us = tp->srtt_us >> 2;
2898		if (tp->packets_out == 1)
2899			timeout_us += tcp_rto_min_us(sk);
2900		else
2901			timeout_us += TCP_TIMEOUT_MIN_US;
2902		timeout = usecs_to_jiffies(timeout_us);
2903	} else {
2904		timeout = TCP_TIMEOUT_INIT;
2905	}
2906
2907	/* If the RTO formula yields an earlier time, then use that time. */
2908	rto_delta_us = advancing_rto ?
2909			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2910			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2911	if (rto_delta_us > 0)
2912		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2913
2914	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2915	return true;
2916}
2917
2918/* Thanks to skb fast clones, we can detect if a prior transmit of
2919 * a packet is still in a qdisc or driver queue.
2920 * In this case, there is very little point doing a retransmit !
2921 */
2922static bool skb_still_in_host_queue(struct sock *sk,
2923				    const struct sk_buff *skb)
2924{
2925	if (unlikely(skb_fclone_busy(sk, skb))) {
2926		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2927		smp_mb__after_atomic();
2928		if (skb_fclone_busy(sk, skb)) {
2929			NET_INC_STATS(sock_net(sk),
2930				      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2931			return true;
2932		}
2933	}
2934	return false;
2935}
2936
2937/* When probe timeout (PTO) fires, try send a new segment if possible, else
2938 * retransmit the last segment.
2939 */
2940void tcp_send_loss_probe(struct sock *sk)
2941{
2942	struct tcp_sock *tp = tcp_sk(sk);
2943	struct sk_buff *skb;
2944	int pcount;
2945	int mss = tcp_current_mss(sk);
2946
2947	/* At most one outstanding TLP */
2948	if (tp->tlp_high_seq)
2949		goto rearm_timer;
2950
2951	tp->tlp_retrans = 0;
2952	skb = tcp_send_head(sk);
2953	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2954		pcount = tp->packets_out;
2955		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2956		if (tp->packets_out > pcount)
2957			goto probe_sent;
2958		goto rearm_timer;
2959	}
2960	skb = skb_rb_last(&sk->tcp_rtx_queue);
2961	if (unlikely(!skb)) {
2962		tcp_warn_once(sk, tp->packets_out, "invalid inflight: ");
2963		smp_store_release(&inet_csk(sk)->icsk_pending, 0);
2964		return;
2965	}
2966
2967	if (skb_still_in_host_queue(sk, skb))
2968		goto rearm_timer;
2969
2970	pcount = tcp_skb_pcount(skb);
2971	if (WARN_ON(!pcount))
2972		goto rearm_timer;
2973
2974	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2975		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2976					  (pcount - 1) * mss, mss,
2977					  GFP_ATOMIC)))
2978			goto rearm_timer;
2979		skb = skb_rb_next(skb);
2980	}
2981
2982	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2983		goto rearm_timer;
2984
2985	if (__tcp_retransmit_skb(sk, skb, 1))
2986		goto rearm_timer;
2987
2988	tp->tlp_retrans = 1;
2989
2990probe_sent:
2991	/* Record snd_nxt for loss detection. */
2992	tp->tlp_high_seq = tp->snd_nxt;
2993
2994	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2995	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2996	smp_store_release(&inet_csk(sk)->icsk_pending, 0);
2997rearm_timer:
2998	tcp_rearm_rto(sk);
2999}
3000
3001/* Push out any pending frames which were held back due to
3002 * TCP_CORK or attempt at coalescing tiny packets.
3003 * The socket must be locked by the caller.
3004 */
3005void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
3006			       int nonagle)
3007{
3008	/* If we are closed, the bytes will have to remain here.
3009	 * In time closedown will finish, we empty the write queue and
3010	 * all will be happy.
3011	 */
3012	if (unlikely(sk->sk_state == TCP_CLOSE))
3013		return;
3014
3015	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
3016			   sk_gfp_mask(sk, GFP_ATOMIC)))
3017		tcp_check_probe_timer(sk);
3018}
3019
3020/* Send _single_ skb sitting at the send head. This function requires
3021 * true push pending frames to setup probe timer etc.
3022 */
3023void tcp_push_one(struct sock *sk, unsigned int mss_now)
3024{
3025	struct sk_buff *skb = tcp_send_head(sk);
3026
3027	BUG_ON(!skb || skb->len < mss_now);
3028
3029	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
3030}
3031
3032/* This function returns the amount that we can raise the
3033 * usable window based on the following constraints
3034 *
3035 * 1. The window can never be shrunk once it is offered (RFC 793)
3036 * 2. We limit memory per socket
3037 *
3038 * RFC 1122:
3039 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3040 *  RECV.NEXT + RCV.WIN fixed until:
3041 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3042 *
3043 * i.e. don't raise the right edge of the window until you can raise
3044 * it at least MSS bytes.
3045 *
3046 * Unfortunately, the recommended algorithm breaks header prediction,
3047 * since header prediction assumes th->window stays fixed.
3048 *
3049 * Strictly speaking, keeping th->window fixed violates the receiver
3050 * side SWS prevention criteria. The problem is that under this rule
3051 * a stream of single byte packets will cause the right side of the
3052 * window to always advance by a single byte.
3053 *
3054 * Of course, if the sender implements sender side SWS prevention
3055 * then this will not be a problem.
3056 *
3057 * BSD seems to make the following compromise:
3058 *
3059 *	If the free space is less than the 1/4 of the maximum
3060 *	space available and the free space is less than 1/2 mss,
3061 *	then set the window to 0.
3062 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3063 *	Otherwise, just prevent the window from shrinking
3064 *	and from being larger than the largest representable value.
3065 *
3066 * This prevents incremental opening of the window in the regime
3067 * where TCP is limited by the speed of the reader side taking
3068 * data out of the TCP receive queue. It does nothing about
3069 * those cases where the window is constrained on the sender side
3070 * because the pipeline is full.
3071 *
3072 * BSD also seems to "accidentally" limit itself to windows that are a
3073 * multiple of MSS, at least until the free space gets quite small.
3074 * This would appear to be a side effect of the mbuf implementation.
3075 * Combining these two algorithms results in the observed behavior
3076 * of having a fixed window size at almost all times.
3077 *
3078 * Below we obtain similar behavior by forcing the offered window to
3079 * a multiple of the mss when it is feasible to do so.
3080 *
3081 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3082 * Regular options like TIMESTAMP are taken into account.
3083 */
3084u32 __tcp_select_window(struct sock *sk)
3085{
3086	struct inet_connection_sock *icsk = inet_csk(sk);
3087	struct tcp_sock *tp = tcp_sk(sk);
3088	struct net *net = sock_net(sk);
3089	/* MSS for the peer's data.  Previous versions used mss_clamp
3090	 * here.  I don't know if the value based on our guesses
3091	 * of peer's MSS is better for the performance.  It's more correct
3092	 * but may be worse for the performance because of rcv_mss
3093	 * fluctuations.  --SAW  1998/11/1
3094	 */
3095	int mss = icsk->icsk_ack.rcv_mss;
3096	int free_space = tcp_space(sk);
3097	int allowed_space = tcp_full_space(sk);
3098	int full_space, window;
3099
3100	if (sk_is_mptcp(sk))
3101		mptcp_space(sk, &free_space, &allowed_space);
3102
3103	full_space = min_t(int, tp->window_clamp, allowed_space);
3104
3105	if (unlikely(mss > full_space)) {
3106		mss = full_space;
3107		if (mss <= 0)
3108			return 0;
3109	}
3110
3111	/* Only allow window shrink if the sysctl is enabled and we have
3112	 * a non-zero scaling factor in effect.
3113	 */
3114	if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3115		goto shrink_window_allowed;
3116
3117	/* do not allow window to shrink */
3118
3119	if (free_space < (full_space >> 1)) {
3120		icsk->icsk_ack.quick = 0;
3121
3122		if (tcp_under_memory_pressure(sk))
3123			tcp_adjust_rcv_ssthresh(sk);
3124
3125		/* free_space might become our new window, make sure we don't
3126		 * increase it due to wscale.
3127		 */
3128		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3129
3130		/* if free space is less than mss estimate, or is below 1/16th
3131		 * of the maximum allowed, try to move to zero-window, else
3132		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3133		 * new incoming data is dropped due to memory limits.
3134		 * With large window, mss test triggers way too late in order
3135		 * to announce zero window in time before rmem limit kicks in.
3136		 */
3137		if (free_space < (allowed_space >> 4) || free_space < mss)
3138			return 0;
3139	}
3140
3141	if (free_space > tp->rcv_ssthresh)
3142		free_space = tp->rcv_ssthresh;
3143
3144	/* Don't do rounding if we are using window scaling, since the
3145	 * scaled window will not line up with the MSS boundary anyway.
3146	 */
 
3147	if (tp->rx_opt.rcv_wscale) {
3148		window = free_space;
3149
3150		/* Advertise enough space so that it won't get scaled away.
3151		 * Import case: prevent zero window announcement if
3152		 * 1<<rcv_wscale > mss.
3153		 */
3154		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
 
 
3155	} else {
3156		window = tp->rcv_wnd;
3157		/* Get the largest window that is a nice multiple of mss.
3158		 * Window clamp already applied above.
3159		 * If our current window offering is within 1 mss of the
3160		 * free space we just keep it. This prevents the divide
3161		 * and multiply from happening most of the time.
3162		 * We also don't do any window rounding when the free space
3163		 * is too small.
3164		 */
3165		if (window <= free_space - mss || window > free_space)
3166			window = rounddown(free_space, mss);
3167		else if (mss == full_space &&
3168			 free_space > window + (full_space >> 1))
3169			window = free_space;
3170	}
3171
3172	return window;
3173
3174shrink_window_allowed:
3175	/* new window should always be an exact multiple of scaling factor */
3176	free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3177
3178	if (free_space < (full_space >> 1)) {
3179		icsk->icsk_ack.quick = 0;
3180
3181		if (tcp_under_memory_pressure(sk))
3182			tcp_adjust_rcv_ssthresh(sk);
3183
3184		/* if free space is too low, return a zero window */
3185		if (free_space < (allowed_space >> 4) || free_space < mss ||
3186			free_space < (1 << tp->rx_opt.rcv_wscale))
3187			return 0;
3188	}
3189
3190	if (free_space > tp->rcv_ssthresh) {
3191		free_space = tp->rcv_ssthresh;
3192		/* new window should always be an exact multiple of scaling factor
3193		 *
3194		 * For this case, we ALIGN "up" (increase free_space) because
3195		 * we know free_space is not zero here, it has been reduced from
3196		 * the memory-based limit, and rcv_ssthresh is not a hard limit
3197		 * (unlike sk_rcvbuf).
3198		 */
3199		free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3200	}
3201
3202	return free_space;
3203}
3204
3205void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3206			     const struct sk_buff *next_skb)
3207{
3208	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3209		const struct skb_shared_info *next_shinfo =
3210			skb_shinfo(next_skb);
3211		struct skb_shared_info *shinfo = skb_shinfo(skb);
3212
3213		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3214		shinfo->tskey = next_shinfo->tskey;
3215		TCP_SKB_CB(skb)->txstamp_ack |=
3216			TCP_SKB_CB(next_skb)->txstamp_ack;
3217	}
3218}
3219
3220/* Collapses two adjacent SKB's during retransmission. */
3221static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3222{
3223	struct tcp_sock *tp = tcp_sk(sk);
3224	struct sk_buff *next_skb = skb_rb_next(skb);
3225	int next_skb_size;
3226
 
3227	next_skb_size = next_skb->len;
3228
3229	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3230
3231	if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3232		return false;
 
3233
3234	tcp_highest_sack_replace(sk, next_skb, skb);
 
 
 
 
 
 
 
3235
3236	/* Update sequence range on original skb. */
3237	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3238
3239	/* Merge over control information. This moves PSH/FIN etc. over */
3240	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3241
3242	/* All done, get rid of second SKB and account for it so
3243	 * packet counting does not break.
3244	 */
3245	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3246	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3247
3248	/* changed transmit queue under us so clear hints */
3249	tcp_clear_retrans_hints_partial(tp);
3250	if (next_skb == tp->retransmit_skb_hint)
3251		tp->retransmit_skb_hint = skb;
3252
3253	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3254
3255	tcp_skb_collapse_tstamp(skb, next_skb);
3256
3257	tcp_rtx_queue_unlink_and_free(next_skb, sk);
3258	return true;
3259}
3260
3261/* Check if coalescing SKBs is legal. */
3262static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3263{
3264	if (tcp_skb_pcount(skb) > 1)
3265		return false;
 
 
 
3266	if (skb_cloned(skb))
3267		return false;
3268	if (!skb_frags_readable(skb))
3269		return false;
3270	/* Some heuristics for collapsing over SACK'd could be invented */
3271	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3272		return false;
3273
3274	return true;
3275}
3276
3277/* Collapse packets in the retransmit queue to make to create
3278 * less packets on the wire. This is only done on retransmission.
3279 */
3280static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3281				     int space)
3282{
3283	struct tcp_sock *tp = tcp_sk(sk);
3284	struct sk_buff *skb = to, *tmp;
3285	bool first = true;
3286
3287	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3288		return;
3289	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3290		return;
3291
3292	skb_rbtree_walk_from_safe(skb, tmp) {
3293		if (!tcp_can_collapse(sk, skb))
3294			break;
3295
3296		if (!tcp_skb_can_collapse(to, skb))
3297			break;
3298
3299		space -= skb->len;
3300
3301		if (first) {
3302			first = false;
3303			continue;
3304		}
3305
3306		if (space < 0)
3307			break;
 
 
 
 
 
3308
3309		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3310			break;
3311
3312		if (!tcp_collapse_retrans(sk, to))
3313			break;
3314	}
3315}
3316
3317/* This retransmits one SKB.  Policy decisions and retransmit queue
3318 * state updates are done by the caller.  Returns non-zero if an
3319 * error occurred which prevented the send.
3320 */
3321int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3322{
3323	struct inet_connection_sock *icsk = inet_csk(sk);
3324	struct tcp_sock *tp = tcp_sk(sk);
 
3325	unsigned int cur_mss;
3326	int diff, len, err;
3327	int avail_wnd;
3328
3329	/* Inconclusive MTU probe */
3330	if (icsk->icsk_mtup.probe_size)
3331		icsk->icsk_mtup.probe_size = 0;
 
3332
3333	if (skb_still_in_host_queue(sk, skb))
3334		return -EBUSY;
 
 
 
 
3335
3336start:
3337	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3338		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3339			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3340			TCP_SKB_CB(skb)->seq++;
3341			goto start;
3342		}
3343		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3344			WARN_ON_ONCE(1);
3345			return -EINVAL;
3346		}
3347		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3348			return -ENOMEM;
3349	}
3350
3351	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3352		return -EHOSTUNREACH; /* Routing failure or similar. */
3353
3354	cur_mss = tcp_current_mss(sk);
3355	avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3356
3357	/* If receiver has shrunk his window, and skb is out of
3358	 * new window, do not retransmit it. The exception is the
3359	 * case, when window is shrunk to zero. In this case
3360	 * our retransmit of one segment serves as a zero window probe.
3361	 */
3362	if (avail_wnd <= 0) {
3363		if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3364			return -EAGAIN;
3365		avail_wnd = cur_mss;
3366	}
3367
3368	len = cur_mss * segs;
3369	if (len > avail_wnd) {
3370		len = rounddown(avail_wnd, cur_mss);
3371		if (!len)
3372			len = avail_wnd;
3373	}
3374	if (skb->len > len) {
3375		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3376				 cur_mss, GFP_ATOMIC))
3377			return -ENOMEM; /* We'll try again later. */
3378	} else {
3379		if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3380			return -ENOMEM;
3381
3382		diff = tcp_skb_pcount(skb);
3383		tcp_set_skb_tso_segs(skb, cur_mss);
3384		diff -= tcp_skb_pcount(skb);
3385		if (diff)
3386			tcp_adjust_pcount(sk, skb, diff);
3387		avail_wnd = min_t(int, avail_wnd, cur_mss);
3388		if (skb->len < avail_wnd)
3389			tcp_retrans_try_collapse(sk, skb, avail_wnd);
3390	}
3391
3392	/* RFC3168, section 6.1.1.1. ECN fallback */
3393	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3394		tcp_ecn_clear_syn(sk, skb);
3395
3396	/* Update global and local TCP statistics. */
3397	segs = tcp_skb_pcount(skb);
3398	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3399	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3400		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3401	tp->total_retrans += segs;
3402	tp->bytes_retrans += skb->len;
3403
3404	/* make sure skb->data is aligned on arches that require it
3405	 * and check if ack-trimming & collapsing extended the headroom
3406	 * beyond what csum_start can cover.
3407	 */
3408	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3409		     skb_headroom(skb) >= 0xFFFF)) {
3410		struct sk_buff *nskb;
3411
3412		tcp_skb_tsorted_save(skb) {
3413			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3414			if (nskb) {
3415				nskb->dev = NULL;
3416				err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3417			} else {
3418				err = -ENOBUFS;
3419			}
3420		} tcp_skb_tsorted_restore(skb);
3421
3422		if (!err) {
3423			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3424			tcp_rate_skb_sent(sk, skb);
3425		}
3426	} else {
3427		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3428	}
3429
3430	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3431		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3432				  TCP_SKB_CB(skb)->seq, segs, err);
3433
3434	if (likely(!err)) {
3435		trace_tcp_retransmit_skb(sk, skb);
3436	} else if (err != -EBUSY) {
3437		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3438	}
3439
3440	/* To avoid taking spuriously low RTT samples based on a timestamp
3441	 * for a transmit that never happened, always mark EVER_RETRANS
 
3442	 */
3443	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
 
 
 
 
 
 
 
 
 
3444
3445	return err;
3446}
 
 
3447
3448int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3449{
3450	struct tcp_sock *tp = tcp_sk(sk);
3451	int err = __tcp_retransmit_skb(sk, skb, segs);
3452
3453	if (err == 0) {
 
 
 
 
 
3454#if FASTRETRANS_DEBUG > 0
3455		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3456			net_dbg_ratelimited("retrans_out leaked\n");
 
3457		}
3458#endif
 
 
3459		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3460		tp->retrans_out += tcp_skb_pcount(skb);
3461	}
3462
3463	/* Save stamp of the first (attempted) retransmit. */
3464	if (!tp->retrans_stamp)
3465		tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb);
3466
3467	if (tp->undo_retrans < 0)
3468		tp->undo_retrans = 0;
3469	tp->undo_retrans += tcp_skb_pcount(skb);
 
 
 
 
3470	return err;
3471}
3472
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3473/* This gets called after a retransmit timeout, and the initially
3474 * retransmitted data is acknowledged.  It tries to continue
3475 * resending the rest of the retransmit queue, until either
3476 * we've sent it all or the congestion window limit is reached.
 
 
 
3477 */
3478void tcp_xmit_retransmit_queue(struct sock *sk)
3479{
3480	const struct inet_connection_sock *icsk = inet_csk(sk);
3481	struct sk_buff *skb, *rtx_head, *hole = NULL;
3482	struct tcp_sock *tp = tcp_sk(sk);
3483	bool rearm_timer = false;
3484	u32 max_segs;
 
3485	int mib_idx;
 
3486
3487	if (!tp->packets_out)
3488		return;
3489
3490	rtx_head = tcp_rtx_queue_head(sk);
3491	skb = tp->retransmit_skb_hint ?: rtx_head;
3492	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3493	skb_rbtree_walk_from(skb) {
3494		__u8 sacked;
3495		int segs;
3496
3497		if (tcp_pacing_check(sk))
3498			break;
 
 
 
 
 
 
 
 
 
 
3499
 
 
3500		/* we could do better than to assign each time */
3501		if (!hole)
3502			tp->retransmit_skb_hint = skb;
3503
3504		segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3505		if (segs <= 0)
3506			break;
3507		sacked = TCP_SKB_CB(skb)->sacked;
3508		/* In case tcp_shift_skb_data() have aggregated large skbs,
3509		 * we need to make sure not sending too bigs TSO packets
3510		 */
3511		segs = min_t(int, segs, max_segs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3512
3513		if (tp->retrans_out >= tp->lost_out) {
3514			break;
3515		} else if (!(sacked & TCPCB_LOST)) {
3516			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3517				hole = skb;
3518			continue;
3519
3520		} else {
 
3521			if (icsk->icsk_ca_state != TCP_CA_Loss)
3522				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3523			else
3524				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3525		}
3526
3527		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3528			continue;
3529
3530		if (tcp_small_queue_check(sk, skb, 1))
3531			break;
3532
3533		if (tcp_retransmit_skb(sk, skb, segs))
3534			break;
3535
3536		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3537
3538		if (tcp_in_cwnd_reduction(sk))
3539			tp->prr_out += tcp_skb_pcount(skb);
3540
3541		if (skb == rtx_head &&
3542		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3543			rearm_timer = true;
3544
 
 
 
 
3545	}
3546	if (rearm_timer)
3547		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3548				     inet_csk(sk)->icsk_rto,
3549				     TCP_RTO_MAX);
3550}
3551
3552/* We allow to exceed memory limits for FIN packets to expedite
3553 * connection tear down and (memory) recovery.
3554 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3555 * or even be forced to close flow without any FIN.
3556 * In general, we want to allow one skb per socket to avoid hangs
3557 * with edge trigger epoll()
3558 */
3559void sk_forced_mem_schedule(struct sock *sk, int size)
3560{
3561	int delta, amt;
3562
3563	delta = size - sk->sk_forward_alloc;
3564	if (delta <= 0)
3565		return;
3566	amt = sk_mem_pages(delta);
3567	sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3568	sk_memory_allocated_add(sk, amt);
3569
3570	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3571		mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3572					gfp_memcg_charge() | __GFP_NOFAIL);
3573}
3574
3575/* Send a FIN. The caller locks the socket for us.
3576 * We should try to send a FIN packet really hard, but eventually give up.
3577 */
3578void tcp_send_fin(struct sock *sk)
3579{
3580	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3581	struct tcp_sock *tp = tcp_sk(sk);
 
 
3582
3583	/* Optimization, tack on the FIN if we have one skb in write queue and
3584	 * this skb was not yet sent, or we are under memory pressure.
3585	 * Note: in the latter case, FIN packet will be sent after a timeout,
3586	 * as TCP stack thinks it has already been transmitted.
3587	 */
3588	tskb = tail;
3589	if (!tskb && tcp_under_memory_pressure(sk))
3590		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3591
3592	if (tskb) {
3593		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3594		TCP_SKB_CB(tskb)->end_seq++;
3595		tp->write_seq++;
3596		if (!tail) {
3597			/* This means tskb was already sent.
3598			 * Pretend we included the FIN on previous transmit.
3599			 * We need to set tp->snd_nxt to the value it would have
3600			 * if FIN had been sent. This is because retransmit path
3601			 * does not change tp->snd_nxt.
3602			 */
3603			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3604			return;
3605		}
3606	} else {
3607		skb = alloc_skb_fclone(MAX_TCP_HEADER,
3608				       sk_gfp_mask(sk, GFP_ATOMIC |
3609						       __GFP_NOWARN));
3610		if (unlikely(!skb))
3611			return;
 
 
 
3612
3613		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3614		skb_reserve(skb, MAX_TCP_HEADER);
3615		sk_forced_mem_schedule(sk, skb->truesize);
3616		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3617		tcp_init_nondata_skb(skb, tp->write_seq,
3618				     TCPHDR_ACK | TCPHDR_FIN);
3619		tcp_queue_skb(sk, skb);
3620	}
3621	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3622}
3623
3624/* We get here when a process closes a file descriptor (either due to
3625 * an explicit close() or as a byproduct of exit()'ing) and there
3626 * was unread data in the receive queue.  This behavior is recommended
3627 * by RFC 2525, section 2.17.  -DaveM
3628 */
3629void tcp_send_active_reset(struct sock *sk, gfp_t priority,
3630			   enum sk_rst_reason reason)
3631{
3632	struct sk_buff *skb;
3633
3634	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3635
3636	/* NOTE: No TCP options attached and we never retransmit this. */
3637	skb = alloc_skb(MAX_TCP_HEADER, priority);
3638	if (!skb) {
3639		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3640		return;
3641	}
3642
3643	/* Reserve space for headers and prepare control bits. */
3644	skb_reserve(skb, MAX_TCP_HEADER);
3645	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3646			     TCPHDR_ACK | TCPHDR_RST);
3647	tcp_mstamp_refresh(tcp_sk(sk));
3648	/* Send it off. */
 
3649	if (tcp_transmit_skb(sk, skb, 0, priority))
3650		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3651
3652	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3653	 * skb here is different to the troublesome skb, so use NULL
3654	 */
3655	trace_tcp_send_reset(sk, NULL, reason);
3656}
3657
3658/* Send a crossed SYN-ACK during socket establishment.
3659 * WARNING: This routine must only be called when we have already sent
3660 * a SYN packet that crossed the incoming SYN that caused this routine
3661 * to get called. If this assumption fails then the initial rcv_wnd
3662 * and rcv_wscale values will not be correct.
3663 */
3664int tcp_send_synack(struct sock *sk)
3665{
3666	struct sk_buff *skb;
3667
3668	skb = tcp_rtx_queue_head(sk);
3669	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3670		pr_err("%s: wrong queue state\n", __func__);
3671		return -EFAULT;
3672	}
3673	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3674		if (skb_cloned(skb)) {
3675			struct sk_buff *nskb;
3676
3677			tcp_skb_tsorted_save(skb) {
3678				nskb = skb_copy(skb, GFP_ATOMIC);
3679			} tcp_skb_tsorted_restore(skb);
3680			if (!nskb)
3681				return -ENOMEM;
3682			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3683			tcp_highest_sack_replace(sk, skb, nskb);
3684			tcp_rtx_queue_unlink_and_free(skb, sk);
3685			__skb_header_release(nskb);
3686			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3687			sk_wmem_queued_add(sk, nskb->truesize);
3688			sk_mem_charge(sk, nskb->truesize);
3689			skb = nskb;
3690		}
3691
3692		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3693		tcp_ecn_send_synack(sk, skb);
3694	}
 
3695	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3696}
3697
3698/**
3699 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3700 * @sk: listener socket
3701 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3702 *       should not use it again.
3703 * @req: request_sock pointer
3704 * @foc: cookie for tcp fast open
3705 * @synack_type: Type of synack to prepare
3706 * @syn_skb: SYN packet just received.  It could be NULL for rtx case.
3707 */
3708struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3709				struct request_sock *req,
3710				struct tcp_fastopen_cookie *foc,
3711				enum tcp_synack_type synack_type,
3712				struct sk_buff *syn_skb)
3713{
3714	struct inet_request_sock *ireq = inet_rsk(req);
3715	const struct tcp_sock *tp = tcp_sk(sk);
3716	struct tcp_out_options opts;
3717	struct tcp_key key = {};
 
 
 
 
3718	struct sk_buff *skb;
 
3719	int tcp_header_size;
3720	struct tcphdr *th;
3721	int mss;
3722	u64 now;
3723
3724	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3725	if (unlikely(!skb)) {
3726		dst_release(dst);
 
3727		return NULL;
3728	}
3729	/* Reserve space for headers. */
3730	skb_reserve(skb, MAX_TCP_HEADER);
3731
3732	switch (synack_type) {
3733	case TCP_SYNACK_NORMAL:
3734		skb_set_owner_edemux(skb, req_to_sk(req));
3735		break;
3736	case TCP_SYNACK_COOKIE:
3737		/* Under synflood, we do not attach skb to a socket,
3738		 * to avoid false sharing.
3739		 */
3740		break;
3741	case TCP_SYNACK_FASTOPEN:
3742		/* sk is a const pointer, because we want to express multiple
3743		 * cpu might call us concurrently.
3744		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3745		 */
3746		skb_set_owner_w(skb, (struct sock *)sk);
3747		break;
3748	}
3749	skb_dst_set(skb, dst);
3750
3751	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3752
3753	memset(&opts, 0, sizeof(opts));
3754	now = tcp_clock_ns();
3755#ifdef CONFIG_SYN_COOKIES
3756	if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3757		skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3758				      SKB_CLOCK_MONOTONIC);
3759	else
3760#endif
3761	{
3762		skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
3763		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3764			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3765	}
3766
3767#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3768	rcu_read_lock();
3769#endif
3770	if (tcp_rsk_used_ao(req)) {
3771#ifdef CONFIG_TCP_AO
3772		struct tcp_ao_key *ao_key = NULL;
3773		u8 keyid = tcp_rsk(req)->ao_keyid;
3774		u8 rnext = tcp_rsk(req)->ao_rcv_next;
3775
3776		ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3777							    keyid, -1);
3778		/* If there is no matching key - avoid sending anything,
3779		 * especially usigned segments. It could try harder and lookup
3780		 * for another peer-matching key, but the peer has requested
3781		 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3782		 */
3783		if (unlikely(!ao_key)) {
3784			trace_tcp_ao_synack_no_key(sk, keyid, rnext);
3785			rcu_read_unlock();
3786			kfree_skb(skb);
3787			net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3788					     keyid);
3789			return NULL;
3790		}
3791		key.ao_key = ao_key;
3792		key.type = TCP_KEY_AO;
3793#endif
3794	} else {
3795#ifdef CONFIG_TCP_MD5SIG
3796		key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3797					req_to_sk(req));
3798		if (key.md5_key)
3799			key.type = TCP_KEY_MD5;
3800#endif
3801	}
3802	skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3803	/* bpf program will be interested in the tcp_flags */
3804	TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3805	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts,
3806					     &key, foc, synack_type, syn_skb)
3807					+ sizeof(*th);
3808
3809	skb_push(skb, tcp_header_size);
3810	skb_reset_transport_header(skb);
3811
3812	th = (struct tcphdr *)skb->data;
3813	memset(th, 0, sizeof(struct tcphdr));
3814	th->syn = 1;
3815	th->ack = 1;
3816	tcp_ecn_make_synack(req, th);
3817	th->source = htons(ireq->ir_num);
3818	th->dest = ireq->ir_rmt_port;
3819	skb->mark = ireq->ir_mark;
3820	skb->ip_summed = CHECKSUM_PARTIAL;
3821	th->seq = htonl(tcp_rsk(req)->snt_isn);
3822	/* XXX data is queued and acked as is. No buffer/window check */
3823	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3824
3825	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3826	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3827	tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key);
3828	th->doff = (tcp_header_size >> 2);
3829	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3830
3831	/* Okay, we have all we need - do the md5 hash if needed */
3832	if (tcp_key_is_md5(&key)) {
3833#ifdef CONFIG_TCP_MD5SIG
 
 
3834		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3835					key.md5_key, req_to_sk(req), skb);
3836#endif
3837	} else if (tcp_key_is_ao(&key)) {
3838#ifdef CONFIG_TCP_AO
3839		tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
3840					key.ao_key, req, skb,
3841					opts.hash_location - (u8 *)th, 0);
3842#endif
3843	}
3844#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3845	rcu_read_unlock();
3846#endif
3847
3848	bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3849				synack_type, &opts);
3850
3851	skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
3852	tcp_add_tx_delay(skb, tp);
3853
3854	return skb;
3855}
3856EXPORT_SYMBOL(tcp_make_synack);
3857
3858static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3859{
3860	struct inet_connection_sock *icsk = inet_csk(sk);
3861	const struct tcp_congestion_ops *ca;
3862	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3863
3864	if (ca_key == TCP_CA_UNSPEC)
3865		return;
3866
3867	rcu_read_lock();
3868	ca = tcp_ca_find_key(ca_key);
3869	if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3870		bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3871		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3872		icsk->icsk_ca_ops = ca;
3873	}
3874	rcu_read_unlock();
3875}
3876
3877/* Do all connect socket setups that can be done AF independent. */
3878static void tcp_connect_init(struct sock *sk)
3879{
3880	const struct dst_entry *dst = __sk_dst_get(sk);
3881	struct tcp_sock *tp = tcp_sk(sk);
3882	__u8 rcv_wscale;
3883	u32 rcv_wnd;
3884
3885	/* We'll fix this up when we get a response from the other end.
3886	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3887	 */
3888	tp->tcp_header_len = sizeof(struct tcphdr);
3889	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3890		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3891
3892	tcp_ao_connect_init(sk);
 
 
 
3893
3894	/* If user gave his TCP_MAXSEG, record it to clamp */
3895	if (tp->rx_opt.user_mss)
3896		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3897	tp->max_window = 0;
3898	tcp_mtup_init(sk);
3899	tcp_sync_mss(sk, dst_mtu(dst));
3900
3901	tcp_ca_dst_init(sk, dst);
3902
3903	if (!tp->window_clamp)
3904		WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW));
3905	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
 
 
3906
3907	tcp_initialize_rcv_mss(sk);
3908
3909	/* limit the window selection if the user enforce a smaller rx buffer */
3910	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3911	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3912		WRITE_ONCE(tp->window_clamp, tcp_full_space(sk));
3913
3914	rcv_wnd = tcp_rwnd_init_bpf(sk);
3915	if (rcv_wnd == 0)
3916		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3917
3918	tcp_select_initial_window(sk, tcp_full_space(sk),
3919				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3920				  &tp->rcv_wnd,
3921				  &tp->window_clamp,
3922				  READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3923				  &rcv_wscale,
3924				  rcv_wnd);
3925
3926	tp->rx_opt.rcv_wscale = rcv_wscale;
3927	tp->rcv_ssthresh = tp->rcv_wnd;
3928
3929	WRITE_ONCE(sk->sk_err, 0);
3930	sock_reset_flag(sk, SOCK_DONE);
3931	tp->snd_wnd = 0;
3932	tcp_init_wl(tp, 0);
3933	tcp_write_queue_purge(sk);
3934	tp->snd_una = tp->write_seq;
3935	tp->snd_sml = tp->write_seq;
3936	tp->snd_up = tp->write_seq;
3937	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3938
3939	if (likely(!tp->repair))
3940		tp->rcv_nxt = 0;
3941	else
3942		tp->rcv_tstamp = tcp_jiffies32;
3943	tp->rcv_wup = tp->rcv_nxt;
3944	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3945
3946	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3947	inet_csk(sk)->icsk_retransmits = 0;
3948	tcp_clear_retrans(tp);
3949}
3950
3951static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3952{
3953	struct tcp_sock *tp = tcp_sk(sk);
3954	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3955
3956	tcb->end_seq += skb->len;
3957	__skb_header_release(skb);
3958	sk_wmem_queued_add(sk, skb->truesize);
3959	sk_mem_charge(sk, skb->truesize);
3960	WRITE_ONCE(tp->write_seq, tcb->end_seq);
3961	tp->packets_out += tcp_skb_pcount(skb);
3962}
3963
3964/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3965 * queue a data-only packet after the regular SYN, such that regular SYNs
3966 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3967 * only the SYN sequence, the data are retransmitted in the first ACK.
3968 * If cookie is not cached or other error occurs, falls back to send a
3969 * regular SYN with Fast Open cookie request option.
3970 */
3971static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3972{
3973	struct inet_connection_sock *icsk = inet_csk(sk);
3974	struct tcp_sock *tp = tcp_sk(sk);
3975	struct tcp_fastopen_request *fo = tp->fastopen_req;
3976	struct page_frag *pfrag = sk_page_frag(sk);
3977	struct sk_buff *syn_data;
3978	int space, err = 0;
3979
3980	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3981	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3982		goto fallback;
3983
3984	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3985	 * user-MSS. Reserve maximum option space for middleboxes that add
3986	 * private TCP options. The cost is reduced data space in SYN :(
3987	 */
3988	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3989	/* Sync mss_cache after updating the mss_clamp */
3990	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3991
3992	space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3993		MAX_TCP_OPTION_SPACE;
3994
3995	space = min_t(size_t, space, fo->size);
3996
3997	if (space &&
3998	    !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
3999				  pfrag, sk->sk_allocation))
4000		goto fallback;
4001	syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
4002	if (!syn_data)
4003		goto fallback;
4004	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
4005	if (space) {
4006		space = min_t(size_t, space, pfrag->size - pfrag->offset);
4007		space = tcp_wmem_schedule(sk, space);
4008	}
4009	if (space) {
4010		space = copy_page_from_iter(pfrag->page, pfrag->offset,
4011					    space, &fo->data->msg_iter);
4012		if (unlikely(!space)) {
4013			tcp_skb_tsorted_anchor_cleanup(syn_data);
4014			kfree_skb(syn_data);
4015			goto fallback;
4016		}
4017		skb_fill_page_desc(syn_data, 0, pfrag->page,
4018				   pfrag->offset, space);
4019		page_ref_inc(pfrag->page);
4020		pfrag->offset += space;
4021		skb_len_add(syn_data, space);
4022		skb_zcopy_set(syn_data, fo->uarg, NULL);
4023	}
4024	/* No more data pending in inet_wait_for_connect() */
4025	if (space == fo->size)
4026		fo->data = NULL;
4027	fo->copied = space;
4028
4029	tcp_connect_queue_skb(sk, syn_data);
4030	if (syn_data->len)
4031		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
4032
4033	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
4034
4035	skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, SKB_CLOCK_MONOTONIC);
4036
4037	/* Now full SYN+DATA was cloned and sent (or not),
4038	 * remove the SYN from the original skb (syn_data)
4039	 * we keep in write queue in case of a retransmit, as we
4040	 * also have the SYN packet (with no data) in the same queue.
4041	 */
4042	TCP_SKB_CB(syn_data)->seq++;
4043	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4044	if (!err) {
4045		tp->syn_data = (fo->copied > 0);
4046		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
4047		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4048		goto done;
4049	}
4050
4051	/* data was not sent, put it in write_queue */
4052	__skb_queue_tail(&sk->sk_write_queue, syn_data);
4053	tp->packets_out -= tcp_skb_pcount(syn_data);
4054
4055fallback:
4056	/* Send a regular SYN with Fast Open cookie request option */
4057	if (fo->cookie.len > 0)
4058		fo->cookie.len = 0;
4059	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
4060	if (err)
4061		tp->syn_fastopen = 0;
4062done:
4063	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
4064	return err;
4065}
4066
4067/* Build a SYN and send it off. */
4068int tcp_connect(struct sock *sk)
4069{
4070	struct tcp_sock *tp = tcp_sk(sk);
4071	struct sk_buff *buff;
4072	int err;
4073
4074	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
4075
4076#if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4077	/* Has to be checked late, after setting daddr/saddr/ops.
4078	 * Return error if the peer has both a md5 and a tcp-ao key
4079	 * configured as this is ambiguous.
4080	 */
4081	if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4082					       lockdep_sock_is_held(sk)))) {
4083		bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4084		bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4085		struct tcp_ao_info *ao_info;
4086
4087		ao_info = rcu_dereference_check(tp->ao_info,
4088						lockdep_sock_is_held(sk));
4089		if (ao_info) {
4090			/* This is an extra check: tcp_ao_required() in
4091			 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4092			 * md5 keys on ao_required socket.
4093			 */
4094			needs_ao |= ao_info->ao_required;
4095			WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4096		}
4097		if (needs_md5 && needs_ao)
4098			return -EKEYREJECTED;
4099
4100		/* If we have a matching md5 key and no matching tcp-ao key
4101		 * then free up ao_info if allocated.
4102		 */
4103		if (needs_md5) {
4104			tcp_ao_destroy_sock(sk, false);
4105		} else if (needs_ao) {
4106			tcp_clear_md5_list(sk);
4107			kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4108						  lockdep_sock_is_held(sk)));
4109		}
4110	}
4111#endif
4112#ifdef CONFIG_TCP_AO
4113	if (unlikely(rcu_dereference_protected(tp->ao_info,
4114					       lockdep_sock_is_held(sk)))) {
4115		/* Don't allow connecting if ao is configured but no
4116		 * matching key is found.
4117		 */
4118		if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4119			return -EKEYREJECTED;
4120	}
4121#endif
4122
4123	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4124		return -EHOSTUNREACH; /* Routing failure or similar. */
4125
4126	tcp_connect_init(sk);
4127
4128	if (unlikely(tp->repair)) {
4129		tcp_finish_connect(sk, NULL);
4130		return 0;
4131	}
4132
4133	buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
4134	if (unlikely(!buff))
4135		return -ENOBUFS;
4136
4137	/* SYN eats a sequence byte, write_seq updated by
4138	 * tcp_connect_queue_skb().
4139	 */
4140	tcp_init_nondata_skb(buff, tp->write_seq, TCPHDR_SYN);
4141	tcp_mstamp_refresh(tp);
4142	tp->retrans_stamp = tcp_time_stamp_ts(tp);
4143	tcp_connect_queue_skb(sk, buff);
4144	tcp_ecn_send_syn(sk, buff);
4145	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
4146
4147	/* Send off SYN; include data in Fast Open. */
4148	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
4149	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
 
 
 
4150	if (err == -ECONNREFUSED)
4151		return err;
4152
4153	/* We change tp->snd_nxt after the tcp_transmit_skb() call
4154	 * in order to make this packet get counted in tcpOutSegs.
4155	 */
4156	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4157	tp->pushed_seq = tp->write_seq;
4158	buff = tcp_send_head(sk);
4159	if (unlikely(buff)) {
4160		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4161		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
4162	}
4163	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4164
4165	/* Timer for repeating the SYN until an answer. */
4166	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4167				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
4168	return 0;
4169}
4170EXPORT_SYMBOL(tcp_connect);
4171
4172u32 tcp_delack_max(const struct sock *sk)
4173{
4174	u32 delack_from_rto_min = max(tcp_rto_min(sk), 2) - 1;
4175
4176	return min(inet_csk(sk)->icsk_delack_max, delack_from_rto_min);
4177}
4178
4179/* Send out a delayed ack, the caller does the policy checking
4180 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
4181 * for details.
4182 */
4183void tcp_send_delayed_ack(struct sock *sk)
4184{
4185	struct inet_connection_sock *icsk = inet_csk(sk);
4186	int ato = icsk->icsk_ack.ato;
4187	unsigned long timeout;
4188
4189	if (ato > TCP_DELACK_MIN) {
4190		const struct tcp_sock *tp = tcp_sk(sk);
4191		int max_ato = HZ / 2;
4192
4193		if (inet_csk_in_pingpong_mode(sk) ||
4194		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4195			max_ato = TCP_DELACK_MAX;
4196
4197		/* Slow path, intersegment interval is "high". */
4198
4199		/* If some rtt estimate is known, use it to bound delayed ack.
4200		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4201		 * directly.
4202		 */
4203		if (tp->srtt_us) {
4204			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4205					TCP_DELACK_MIN);
4206
4207			if (rtt < max_ato)
4208				max_ato = rtt;
4209		}
4210
4211		ato = min(ato, max_ato);
4212	}
4213
4214	ato = min_t(u32, ato, tcp_delack_max(sk));
4215
4216	/* Stay within the limit we were given */
4217	timeout = jiffies + ato;
4218
4219	/* Use new timeout only if there wasn't a older one earlier. */
4220	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4221		/* If delack timer is about to expire, send ACK now. */
4222		if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
 
 
 
4223			tcp_send_ack(sk);
4224			return;
4225		}
4226
4227		if (!time_before(timeout, icsk->icsk_ack.timeout))
4228			timeout = icsk->icsk_ack.timeout;
4229	}
4230	smp_store_release(&icsk->icsk_ack.pending,
4231			  icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
4232	icsk->icsk_ack.timeout = timeout;
4233	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4234}
4235
4236/* This routine sends an ack and also updates the window. */
4237void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4238{
4239	struct sk_buff *buff;
4240
4241	/* If we have been reset, we may not send again. */
4242	if (sk->sk_state == TCP_CLOSE)
4243		return;
4244
4245	/* We are not putting this on the write queue, so
4246	 * tcp_transmit_skb() will set the ownership to this
4247	 * sock.
4248	 */
4249	buff = alloc_skb(MAX_TCP_HEADER,
4250			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4251	if (unlikely(!buff)) {
4252		struct inet_connection_sock *icsk = inet_csk(sk);
4253		unsigned long delay;
4254
4255		delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4256		if (delay < TCP_RTO_MAX)
4257			icsk->icsk_ack.retry++;
4258		inet_csk_schedule_ack(sk);
4259		icsk->icsk_ack.ato = TCP_ATO_MIN;
4260		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
 
4261		return;
4262	}
4263
4264	/* Reserve space for headers and prepare control bits. */
4265	skb_reserve(buff, MAX_TCP_HEADER);
4266	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4267
4268	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
4269	 * too much.
4270	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4271	 */
4272	skb_set_tcp_pure_ack(buff);
4273
4274	/* Send it off, this clears delayed acks for us. */
4275	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4276}
4277EXPORT_SYMBOL_GPL(__tcp_send_ack);
4278
4279void tcp_send_ack(struct sock *sk)
4280{
4281	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4282}
4283
4284/* This routine sends a packet with an out of date sequence
4285 * number. It assumes the other end will try to ack it.
4286 *
4287 * Question: what should we make while urgent mode?
4288 * 4.4BSD forces sending single byte of data. We cannot send
4289 * out of window data, because we have SND.NXT==SND.MAX...
4290 *
4291 * Current solution: to send TWO zero-length segments in urgent mode:
4292 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4293 * out-of-date with SND.UNA-1 to probe window.
4294 */
4295static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4296{
4297	struct tcp_sock *tp = tcp_sk(sk);
4298	struct sk_buff *skb;
4299
4300	/* We don't queue it, tcp_transmit_skb() sets ownership. */
4301	skb = alloc_skb(MAX_TCP_HEADER,
4302			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4303	if (!skb)
4304		return -1;
4305
4306	/* Reserve space for headers and set control bits. */
4307	skb_reserve(skb, MAX_TCP_HEADER);
4308	/* Use a previous sequence.  This should cause the other
4309	 * end to send an ack.  Don't queue or clone SKB, just
4310	 * send it.
4311	 */
4312	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4313	NET_INC_STATS(sock_net(sk), mib);
4314	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4315}
4316
4317/* Called from setsockopt( ... TCP_REPAIR ) */
4318void tcp_send_window_probe(struct sock *sk)
4319{
4320	if (sk->sk_state == TCP_ESTABLISHED) {
4321		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4322		tcp_mstamp_refresh(tcp_sk(sk));
4323		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4324	}
4325}
4326
4327/* Initiate keepalive or window probe from timer. */
4328int tcp_write_wakeup(struct sock *sk, int mib)
4329{
4330	struct tcp_sock *tp = tcp_sk(sk);
4331	struct sk_buff *skb;
4332
4333	if (sk->sk_state == TCP_CLOSE)
4334		return -1;
4335
4336	skb = tcp_send_head(sk);
4337	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4338		int err;
4339		unsigned int mss = tcp_current_mss(sk);
4340		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4341
4342		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4343			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4344
4345		/* We are probing the opening of a window
4346		 * but the window size is != 0
4347		 * must have been a result SWS avoidance ( sender )
4348		 */
4349		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4350		    skb->len > mss) {
4351			seg_size = min(seg_size, mss);
4352			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4353			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4354					 skb, seg_size, mss, GFP_ATOMIC))
4355				return -1;
4356		} else if (!tcp_skb_pcount(skb))
4357			tcp_set_skb_tso_segs(skb, mss);
4358
4359		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 
4360		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4361		if (!err)
4362			tcp_event_new_data_sent(sk, skb);
4363		return err;
4364	} else {
4365		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4366			tcp_xmit_probe_skb(sk, 1, mib);
4367		return tcp_xmit_probe_skb(sk, 0, mib);
4368	}
4369}
4370
4371/* A window probe timeout has occurred.  If window is not closed send
4372 * a partial packet else a zero probe.
4373 */
4374void tcp_send_probe0(struct sock *sk)
4375{
4376	struct inet_connection_sock *icsk = inet_csk(sk);
4377	struct tcp_sock *tp = tcp_sk(sk);
4378	struct net *net = sock_net(sk);
4379	unsigned long timeout;
4380	int err;
4381
4382	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4383
4384	if (tp->packets_out || tcp_write_queue_empty(sk)) {
4385		/* Cancel probe timer, if it is not required. */
4386		icsk->icsk_probes_out = 0;
4387		icsk->icsk_backoff = 0;
4388		icsk->icsk_probes_tstamp = 0;
4389		return;
4390	}
4391
4392	icsk->icsk_probes_out++;
4393	if (err <= 0) {
4394		if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4395			icsk->icsk_backoff++;
4396		timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
 
 
 
4397	} else {
4398		/* If packet was not sent due to local congestion,
4399		 * Let senders fight for local resources conservatively.
 
 
 
4400		 */
4401		timeout = TCP_RESOURCE_PROBE_INTERVAL;
4402	}
4403
4404	timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4405	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4406}
4407
4408int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4409{
4410	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4411	struct flowi fl;
4412	int res;
4413
4414	/* Paired with WRITE_ONCE() in sock_setsockopt() */
4415	if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4416		WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4417	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4418				  NULL);
4419	if (!res) {
4420		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4421		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4422		if (unlikely(tcp_passive_fastopen(sk))) {
4423			/* sk has const attribute because listeners are lockless.
4424			 * However in this case, we are dealing with a passive fastopen
4425			 * socket thus we can change total_retrans value.
4426			 */
4427			tcp_sk_rw(sk)->total_retrans++;
4428		}
4429		trace_tcp_retransmit_synack(sk, req);
4430	}
4431	return res;
4432}
4433EXPORT_SYMBOL(tcp_rtx_synack);
v3.1
 
   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#include <net/tcp.h>
 
 
  38
  39#include <linux/compiler.h>
  40#include <linux/gfp.h>
  41#include <linux/module.h>
 
 
  42
  43/* People can turn this off for buggy TCP's found in printers etc. */
  44int sysctl_tcp_retrans_collapse __read_mostly = 1;
  45
  46/* People can turn this on to work with those rare, broken TCPs that
  47 * interpret the window field as a signed quantity.
  48 */
  49int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
 
 
  50
  51/* This limits the percentage of the congestion window which we
  52 * will allow a single TSO frame to consume.  Building TSO frames
  53 * which are too large can cause TCP streams to be bursty.
  54 */
  55int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  56
  57int sysctl_tcp_mtu_probing __read_mostly = 0;
  58int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
  59
  60/* By default, RFC2861 behavior.  */
  61int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  62
  63int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  64EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  65
 
 
  66
  67/* Account for new data that has been sent to the network. */
  68static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  69{
 
  70	struct tcp_sock *tp = tcp_sk(sk);
  71	unsigned int prior_packets = tp->packets_out;
  72
  73	tcp_advance_send_head(sk, skb);
  74	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
 
 
  75
  76	/* Don't override Nagle indefinitely with F-RTO */
  77	if (tp->frto_counter == 2)
  78		tp->frto_counter = 3;
  79
  80	tp->packets_out += tcp_skb_pcount(skb);
  81	if (!prior_packets)
  82		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
  83					  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
 
 
 
  84}
  85
  86/* SND.NXT, if window was not shrunk.
 
  87 * If window has been shrunk, what should we make? It is not clear at all.
  88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  90 * invalid. OK, let's make this for now:
  91 */
  92static inline __u32 tcp_acceptable_seq(struct sock *sk)
  93{
  94	struct tcp_sock *tp = tcp_sk(sk);
  95
  96	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
 
 
  97		return tp->snd_nxt;
  98	else
  99		return tcp_wnd_end(tp);
 100}
 101
 102/* Calculate mss to advertise in SYN segment.
 103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 104 *
 105 * 1. It is independent of path mtu.
 106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 108 *    attached devices, because some buggy hosts are confused by
 109 *    large MSS.
 110 * 4. We do not make 3, we advertise MSS, calculated from first
 111 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 112 *    This may be overridden via information stored in routing table.
 113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 114 *    probably even Jumbo".
 115 */
 116static __u16 tcp_advertise_mss(struct sock *sk)
 117{
 118	struct tcp_sock *tp = tcp_sk(sk);
 119	struct dst_entry *dst = __sk_dst_get(sk);
 120	int mss = tp->advmss;
 121
 122	if (dst) {
 123		unsigned int metric = dst_metric_advmss(dst);
 124
 125		if (metric < mss) {
 126			mss = metric;
 127			tp->advmss = mss;
 128		}
 129	}
 130
 131	return (__u16)mss;
 132}
 133
 134/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 135 * This is the first part of cwnd validation mechanism. */
 136static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
 
 137{
 138	struct tcp_sock *tp = tcp_sk(sk);
 139	s32 delta = tcp_time_stamp - tp->lsndtime;
 140	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
 141	u32 cwnd = tp->snd_cwnd;
 142
 143	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 144
 145	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 146	restart_cwnd = min(restart_cwnd, cwnd);
 147
 148	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 149		cwnd >>= 1;
 150	tp->snd_cwnd = max(cwnd, restart_cwnd);
 151	tp->snd_cwnd_stamp = tcp_time_stamp;
 152	tp->snd_cwnd_used = 0;
 153}
 154
 155/* Congestion state accounting after a packet has been sent. */
 156static void tcp_event_data_sent(struct tcp_sock *tp,
 157				struct sk_buff *skb, struct sock *sk)
 158{
 159	struct inet_connection_sock *icsk = inet_csk(sk);
 160	const u32 now = tcp_time_stamp;
 161
 162	if (sysctl_tcp_slow_start_after_idle &&
 163	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
 164		tcp_cwnd_restart(sk, __sk_dst_get(sk));
 165
 166	tp->lsndtime = now;
 167
 168	/* If it is a reply for ato after last received
 169	 * packet, enter pingpong mode.
 170	 */
 171	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 172		icsk->icsk_ack.pingpong = 1;
 173}
 174
 175/* Account for an ACK we sent. */
 176static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
 177{
 178	tcp_dec_quickack_mode(sk, pkts);
 
 
 
 
 
 
 
 
 
 
 
 
 179	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 180}
 181
 182/* Determine a window scaling and initial window to offer.
 183 * Based on the assumption that the given amount of space
 184 * will be offered. Store the results in the tp structure.
 185 * NOTE: for smooth operation initial space offering should
 186 * be a multiple of mss if possible. We assume here that mss >= 1.
 187 * This MUST be enforced by all callers.
 188 */
 189void tcp_select_initial_window(int __space, __u32 mss,
 190			       __u32 *rcv_wnd, __u32 *window_clamp,
 191			       int wscale_ok, __u8 *rcv_wscale,
 192			       __u32 init_rcv_wnd)
 193{
 194	unsigned int space = (__space < 0 ? 0 : __space);
 
 195
 196	/* If no clamp set the clamp to the max possible scaled window */
 197	if (*window_clamp == 0)
 198		(*window_clamp) = (65535 << 14);
 199	space = min(*window_clamp, space);
 200
 201	/* Quantize space offering to a multiple of mss if possible. */
 202	if (space > mss)
 203		space = (space / mss) * mss;
 204
 205	/* NOTE: offering an initial window larger than 32767
 206	 * will break some buggy TCP stacks. If the admin tells us
 207	 * it is likely we could be speaking with such a buggy stack
 208	 * we will truncate our initial window offering to 32K-1
 209	 * unless the remote has sent us a window scaling option,
 210	 * which we interpret as a sign the remote TCP is not
 211	 * misinterpreting the window field as a signed quantity.
 212	 */
 213	if (sysctl_tcp_workaround_signed_windows)
 214		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 215	else
 216		(*rcv_wnd) = space;
 217
 218	(*rcv_wscale) = 0;
 
 
 
 219	if (wscale_ok) {
 220		/* Set window scaling on max possible window
 221		 * See RFC1323 for an explanation of the limit to 14
 222		 */
 223		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
 224		space = min_t(u32, space, *window_clamp);
 225		while (space > 65535 && (*rcv_wscale) < 14) {
 226			space >>= 1;
 227			(*rcv_wscale)++;
 228		}
 229	}
 230
 231	/* Set initial window to a value enough for senders starting with
 232	 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
 233	 * a limit on the initial window when mss is larger than 1460.
 234	 */
 235	if (mss > (1 << *rcv_wscale)) {
 236		int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
 237		if (mss > 1460)
 238			init_cwnd =
 239			max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
 240		/* when initializing use the value from init_rcv_wnd
 241		 * rather than the default from above
 242		 */
 243		if (init_rcv_wnd)
 244			*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 245		else
 246			*rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
 247	}
 248
 249	/* Set the clamp no higher than max representable value */
 250	(*window_clamp) = min(65535U << (*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 cur_win = tcp_receive_window(tp);
 263	u32 new_win = __tcp_select_window(sk);
 
 
 
 
 
 
 
 
 
 
 
 264
 265	/* Never shrink the offered window */
 
 266	if (new_win < cur_win) {
 267		/* Danger Will Robinson!
 268		 * Don't update rcv_wup/rcv_wnd here or else
 269		 * we will not be able to advertise a zero
 270		 * window in time.  --DaveM
 271		 *
 272		 * Relax Will Robinson.
 273		 */
 274		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 
 
 
 
 
 275	}
 
 276	tp->rcv_wnd = new_win;
 277	tp->rcv_wup = tp->rcv_nxt;
 278
 279	/* Make sure we do not exceed the maximum possible
 280	 * scaled window.
 281	 */
 282	if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
 
 283		new_win = min(new_win, MAX_TCP_WINDOW);
 284	else
 285		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 286
 287	/* RFC1323 scaling applied */
 288	new_win >>= tp->rx_opt.rcv_wscale;
 289
 290	/* If we advertise zero window, disable fast path. */
 291	if (new_win == 0)
 292		tp->pred_flags = 0;
 
 
 
 
 
 293
 294	return new_win;
 295}
 296
 297/* Packet ECN state for a SYN-ACK */
 298static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
 299{
 300	TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
 
 
 301	if (!(tp->ecn_flags & TCP_ECN_OK))
 302		TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
 
 
 
 303}
 304
 305/* Packet ECN state for a SYN.  */
 306static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
 307{
 308	struct tcp_sock *tp = tcp_sk(sk);
 
 
 
 
 
 
 
 
 
 
 309
 310	tp->ecn_flags = 0;
 311	if (sysctl_tcp_ecn == 1) {
 312		TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
 
 313		tp->ecn_flags = TCP_ECN_OK;
 
 
 314	}
 315}
 316
 317static __inline__ void
 318TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
 
 
 
 
 
 
 
 
 
 319{
 320	if (inet_rsk(req)->ecn_ok)
 321		th->ece = 1;
 322}
 323
 324/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 325 * be sent.
 326 */
 327static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
 328				int tcp_header_len)
 329{
 330	struct tcp_sock *tp = tcp_sk(sk);
 331
 332	if (tp->ecn_flags & TCP_ECN_OK) {
 333		/* Not-retransmitted data segment: set ECT and inject CWR. */
 334		if (skb->len != tcp_header_len &&
 335		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 336			INET_ECN_xmit(sk);
 337			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 338				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 339				tcp_hdr(skb)->cwr = 1;
 340				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 341			}
 342		} else {
 343			/* ACK or retransmitted segment: clear ECT|CE */
 344			INET_ECN_dontxmit(sk);
 345		}
 346		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 347			tcp_hdr(skb)->ece = 1;
 348	}
 349}
 350
 351/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 352 * auto increment end seqno.
 353 */
 354static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 355{
 356	skb->ip_summed = CHECKSUM_PARTIAL;
 357	skb->csum = 0;
 358
 359	TCP_SKB_CB(skb)->flags = flags;
 360	TCP_SKB_CB(skb)->sacked = 0;
 361
 362	skb_shinfo(skb)->gso_segs = 1;
 363	skb_shinfo(skb)->gso_size = 0;
 364	skb_shinfo(skb)->gso_type = 0;
 365
 366	TCP_SKB_CB(skb)->seq = seq;
 367	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 368		seq++;
 369	TCP_SKB_CB(skb)->end_seq = seq;
 370}
 371
 372static inline int tcp_urg_mode(const struct tcp_sock *tp)
 373{
 374	return tp->snd_una != tp->snd_up;
 375}
 376
 377#define OPTION_SACK_ADVERTISE	(1 << 0)
 378#define OPTION_TS		(1 << 1)
 379#define OPTION_MD5		(1 << 2)
 380#define OPTION_WSCALE		(1 << 3)
 381#define OPTION_COOKIE_EXTENSION	(1 << 4)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 382
 383struct tcp_out_options {
 384	u8 options;		/* bit field of OPTION_* */
 
 385	u8 ws;			/* window scale, 0 to disable */
 386	u8 num_sack_blocks;	/* number of SACK blocks to include */
 387	u8 hash_size;		/* bytes in hash_location */
 388	u16 mss;		/* 0 to disable */
 
 389	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 390	__u8 *hash_location;	/* temporary pointer, overloaded */
 
 391};
 392
 393/* The sysctl int routines are generic, so check consistency here.
 394 */
 395static u8 tcp_cookie_size_check(u8 desired)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 396{
 397	int cookie_size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 398
 399	if (desired > 0)
 400		/* previously specified */
 401		return desired;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 402
 403	cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
 404	if (cookie_size <= 0)
 405		/* no default specified */
 406		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 407
 408	if (cookie_size <= TCP_COOKIE_MIN)
 409		/* value too small, specify minimum */
 410		return TCP_COOKIE_MIN;
 411
 412	if (cookie_size >= TCP_COOKIE_MAX)
 413		/* value too large, specify maximum */
 414		return TCP_COOKIE_MAX;
 415
 416	if (cookie_size & 1)
 417		/* 8-bit multiple, illegal, fix it */
 418		cookie_size++;
 419
 420	return (u8)cookie_size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 421}
 422
 423/* Write previously computed TCP options to the packet.
 424 *
 425 * Beware: Something in the Internet is very sensitive to the ordering of
 426 * TCP options, we learned this through the hard way, so be careful here.
 427 * Luckily we can at least blame others for their non-compliance but from
 428 * inter-operatibility perspective it seems that we're somewhat stuck with
 429 * the ordering which we have been using if we want to keep working with
 430 * those broken things (not that it currently hurts anybody as there isn't
 431 * particular reason why the ordering would need to be changed).
 432 *
 433 * At least SACK_PERM as the first option is known to lead to a disaster
 434 * (but it may well be that other scenarios fail similarly).
 435 */
 436static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 437			      struct tcp_out_options *opts)
 438{
 439	u8 options = opts->options;	/* mungable copy */
 440
 441	/* Having both authentication and cookies for security is redundant,
 442	 * and there's certainly not enough room.  Instead, the cookie-less
 443	 * extension variant is proposed.
 444	 *
 445	 * Consider the pessimal case with authentication.  The options
 446	 * could look like:
 447	 *   COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
 448	 */
 449	if (unlikely(OPTION_MD5 & options)) {
 450		if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 451			*ptr++ = htonl((TCPOPT_COOKIE << 24) |
 452				       (TCPOLEN_COOKIE_BASE << 16) |
 453				       (TCPOPT_MD5SIG << 8) |
 454				       TCPOLEN_MD5SIG);
 455		} else {
 456			*ptr++ = htonl((TCPOPT_NOP << 24) |
 457				       (TCPOPT_NOP << 16) |
 458				       (TCPOPT_MD5SIG << 8) |
 459				       TCPOLEN_MD5SIG);
 460		}
 461		options &= ~OPTION_COOKIE_EXTENSION;
 462		/* overload cookie hash location */
 463		opts->hash_location = (__u8 *)ptr;
 464		ptr += 4;
 
 
 465	}
 466
 467	if (unlikely(opts->mss)) {
 468		*ptr++ = htonl((TCPOPT_MSS << 24) |
 469			       (TCPOLEN_MSS << 16) |
 470			       opts->mss);
 471	}
 472
 473	if (likely(OPTION_TS & options)) {
 474		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 475			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 476				       (TCPOLEN_SACK_PERM << 16) |
 477				       (TCPOPT_TIMESTAMP << 8) |
 478				       TCPOLEN_TIMESTAMP);
 479			options &= ~OPTION_SACK_ADVERTISE;
 480		} else {
 481			*ptr++ = htonl((TCPOPT_NOP << 24) |
 482				       (TCPOPT_NOP << 16) |
 483				       (TCPOPT_TIMESTAMP << 8) |
 484				       TCPOLEN_TIMESTAMP);
 485		}
 486		*ptr++ = htonl(opts->tsval);
 487		*ptr++ = htonl(opts->tsecr);
 488	}
 489
 490	/* Specification requires after timestamp, so do it now.
 491	 *
 492	 * Consider the pessimal case without authentication.  The options
 493	 * could look like:
 494	 *   MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
 495	 */
 496	if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 497		__u8 *cookie_copy = opts->hash_location;
 498		u8 cookie_size = opts->hash_size;
 499
 500		/* 8-bit multiple handled in tcp_cookie_size_check() above,
 501		 * and elsewhere.
 502		 */
 503		if (0x2 & cookie_size) {
 504			__u8 *p = (__u8 *)ptr;
 505
 506			/* 16-bit multiple */
 507			*p++ = TCPOPT_COOKIE;
 508			*p++ = TCPOLEN_COOKIE_BASE + cookie_size;
 509			*p++ = *cookie_copy++;
 510			*p++ = *cookie_copy++;
 511			ptr++;
 512			cookie_size -= 2;
 513		} else {
 514			/* 32-bit multiple */
 515			*ptr++ = htonl(((TCPOPT_NOP << 24) |
 516					(TCPOPT_NOP << 16) |
 517					(TCPOPT_COOKIE << 8) |
 518					TCPOLEN_COOKIE_BASE) +
 519				       cookie_size);
 520		}
 521
 522		if (cookie_size > 0) {
 523			memcpy(ptr, cookie_copy, cookie_size);
 524			ptr += (cookie_size / 4);
 525		}
 526	}
 527
 528	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 529		*ptr++ = htonl((TCPOPT_NOP << 24) |
 530			       (TCPOPT_NOP << 16) |
 531			       (TCPOPT_SACK_PERM << 8) |
 532			       TCPOLEN_SACK_PERM);
 533	}
 534
 535	if (unlikely(OPTION_WSCALE & options)) {
 536		*ptr++ = htonl((TCPOPT_NOP << 24) |
 537			       (TCPOPT_WINDOW << 16) |
 538			       (TCPOLEN_WINDOW << 8) |
 539			       opts->ws);
 540	}
 541
 542	if (unlikely(opts->num_sack_blocks)) {
 543		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 544			tp->duplicate_sack : tp->selective_acks;
 545		int this_sack;
 546
 547		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 548			       (TCPOPT_NOP  << 16) |
 549			       (TCPOPT_SACK <<  8) |
 550			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 551						     TCPOLEN_SACK_PERBLOCK)));
 552
 553		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 554		     ++this_sack) {
 555			*ptr++ = htonl(sp[this_sack].start_seq);
 556			*ptr++ = htonl(sp[this_sack].end_seq);
 557		}
 558
 559		tp->rx_opt.dsack = 0;
 560	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 561}
 562
 563/* Compute TCP options for SYN packets. This is not the final
 564 * network wire format yet.
 565 */
 566static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 567				struct tcp_out_options *opts,
 568				struct tcp_md5sig_key **md5) {
 
 569	struct tcp_sock *tp = tcp_sk(sk);
 570	struct tcp_cookie_values *cvp = tp->cookie_values;
 571	unsigned remaining = MAX_TCP_OPTION_SPACE;
 572	u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
 573			 tcp_cookie_size_check(cvp->cookie_desired) :
 574			 0;
 575
 576#ifdef CONFIG_TCP_MD5SIG
 577	*md5 = tp->af_specific->md5_lookup(sk, sk);
 578	if (*md5) {
 579		opts->options |= OPTION_MD5;
 580		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 
 
 
 
 
 
 581	}
 582#else
 583	*md5 = NULL;
 584#endif
 585
 586	/* We always get an MSS option.  The option bytes which will be seen in
 587	 * normal data packets should timestamps be used, must be in the MSS
 588	 * advertised.  But we subtract them from tp->mss_cache so that
 589	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 590	 * fact here if necessary.  If we don't do this correctly, as a
 591	 * receiver we won't recognize data packets as being full sized when we
 592	 * should, and thus we won't abide by the delayed ACK rules correctly.
 593	 * SACKs don't matter, we never delay an ACK when we have any of those
 594	 * going out.  */
 595	opts->mss = tcp_advertise_mss(sk);
 596	remaining -= TCPOLEN_MSS_ALIGNED;
 597
 598	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
 599		opts->options |= OPTION_TS;
 600		opts->tsval = TCP_SKB_CB(skb)->when;
 601		opts->tsecr = tp->rx_opt.ts_recent;
 602		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 603	}
 604	if (likely(sysctl_tcp_window_scaling)) {
 605		opts->ws = tp->rx_opt.rcv_wscale;
 606		opts->options |= OPTION_WSCALE;
 607		remaining -= TCPOLEN_WSCALE_ALIGNED;
 608	}
 609	if (likely(sysctl_tcp_sack)) {
 610		opts->options |= OPTION_SACK_ADVERTISE;
 611		if (unlikely(!(OPTION_TS & opts->options)))
 612			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 613	}
 614
 615	/* Note that timestamps are required by the specification.
 616	 *
 617	 * Odd numbers of bytes are prohibited by the specification, ensuring
 618	 * that the cookie is 16-bit aligned, and the resulting cookie pair is
 619	 * 32-bit aligned.
 620	 */
 621	if (*md5 == NULL &&
 622	    (OPTION_TS & opts->options) &&
 623	    cookie_size > 0) {
 624		int need = TCPOLEN_COOKIE_BASE + cookie_size;
 625
 626		if (0x2 & need) {
 627			/* 32-bit multiple */
 628			need += 2; /* NOPs */
 629
 630			if (need > remaining) {
 631				/* try shrinking cookie to fit */
 632				cookie_size -= 2;
 633				need -= 4;
 634			}
 635		}
 636		while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
 637			cookie_size -= 4;
 638			need -= 4;
 639		}
 640		if (TCP_COOKIE_MIN <= cookie_size) {
 641			opts->options |= OPTION_COOKIE_EXTENSION;
 642			opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
 643			opts->hash_size = cookie_size;
 644
 645			/* Remember for future incarnations. */
 646			cvp->cookie_desired = cookie_size;
 647
 648			if (cvp->cookie_desired != cvp->cookie_pair_size) {
 649				/* Currently use random bytes as a nonce,
 650				 * assuming these are completely unpredictable
 651				 * by hostile users of the same system.
 652				 */
 653				get_random_bytes(&cvp->cookie_pair[0],
 654						 cookie_size);
 655				cvp->cookie_pair_size = cookie_size;
 656			}
 657
 658			remaining -= need;
 659		}
 660	}
 
 
 
 661	return MAX_TCP_OPTION_SPACE - remaining;
 662}
 663
 664/* Set up TCP options for SYN-ACKs. */
 665static unsigned tcp_synack_options(struct sock *sk,
 666				   struct request_sock *req,
 667				   unsigned mss, struct sk_buff *skb,
 668				   struct tcp_out_options *opts,
 669				   struct tcp_md5sig_key **md5,
 670				   struct tcp_extend_values *xvp)
 
 
 671{
 672	struct inet_request_sock *ireq = inet_rsk(req);
 673	unsigned remaining = MAX_TCP_OPTION_SPACE;
 674	u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
 675			 xvp->cookie_plus :
 676			 0;
 677
 678#ifdef CONFIG_TCP_MD5SIG
 679	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
 680	if (*md5) {
 681		opts->options |= OPTION_MD5;
 682		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 683
 684		/* We can't fit any SACK blocks in a packet with MD5 + TS
 685		 * options. There was discussion about disabling SACK
 686		 * rather than TS in order to fit in better with old,
 687		 * buggy kernels, but that was deemed to be unnecessary.
 688		 */
 
 
 
 
 
 689		ireq->tstamp_ok &= !ireq->sack_ok;
 690	}
 691#else
 692	*md5 = NULL;
 693#endif
 694
 695	/* We always send an MSS option. */
 696	opts->mss = mss;
 697	remaining -= TCPOLEN_MSS_ALIGNED;
 698
 699	if (likely(ireq->wscale_ok)) {
 700		opts->ws = ireq->rcv_wscale;
 701		opts->options |= OPTION_WSCALE;
 702		remaining -= TCPOLEN_WSCALE_ALIGNED;
 703	}
 704	if (likely(ireq->tstamp_ok)) {
 705		opts->options |= OPTION_TS;
 706		opts->tsval = TCP_SKB_CB(skb)->when;
 707		opts->tsecr = req->ts_recent;
 
 708		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 709	}
 710	if (likely(ireq->sack_ok)) {
 711		opts->options |= OPTION_SACK_ADVERTISE;
 712		if (unlikely(!ireq->tstamp_ok))
 713			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 714	}
 
 
 715
 716	/* Similar rationale to tcp_syn_options() applies here, too.
 717	 * If the <SYN> options fit, the same options should fit now!
 718	 */
 719	if (*md5 == NULL &&
 720	    ireq->tstamp_ok &&
 721	    cookie_plus > TCPOLEN_COOKIE_BASE) {
 722		int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
 723
 724		if (0x2 & need) {
 725			/* 32-bit multiple */
 726			need += 2; /* NOPs */
 727		}
 728		if (need <= remaining) {
 729			opts->options |= OPTION_COOKIE_EXTENSION;
 730			opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
 731			remaining -= need;
 732		} else {
 733			/* There's no error return, so flag it. */
 734			xvp->cookie_out_never = 1; /* true */
 735			opts->hash_size = 0;
 736		}
 737	}
 
 
 
 
 
 
 
 
 738	return MAX_TCP_OPTION_SPACE - remaining;
 739}
 740
 741/* Compute TCP options for ESTABLISHED sockets. This is not the
 742 * final wire format yet.
 743 */
 744static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
 745					struct tcp_out_options *opts,
 746					struct tcp_md5sig_key **md5) {
 747	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
 748	struct tcp_sock *tp = tcp_sk(sk);
 749	unsigned size = 0;
 750	unsigned int eff_sacks;
 751
 752#ifdef CONFIG_TCP_MD5SIG
 753	*md5 = tp->af_specific->md5_lookup(sk, sk);
 754	if (unlikely(*md5)) {
 
 755		opts->options |= OPTION_MD5;
 756		size += TCPOLEN_MD5SIG_ALIGNED;
 
 
 
 757	}
 758#else
 759	*md5 = NULL;
 760#endif
 761
 762	if (likely(tp->rx_opt.tstamp_ok)) {
 763		opts->options |= OPTION_TS;
 764		opts->tsval = tcb ? tcb->when : 0;
 
 765		opts->tsecr = tp->rx_opt.ts_recent;
 766		size += TCPOLEN_TSTAMP_ALIGNED;
 767	}
 768
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 769	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 770	if (unlikely(eff_sacks)) {
 771		const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
 
 
 
 
 772		opts->num_sack_blocks =
 773			min_t(unsigned, eff_sacks,
 774			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 775			      TCPOLEN_SACK_PERBLOCK);
 
 776		size += TCPOLEN_SACK_BASE_ALIGNED +
 777			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 778	}
 779
 
 
 
 
 
 
 
 
 
 780	return size;
 781}
 782
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 783/* This routine actually transmits TCP packets queued in by
 784 * tcp_do_sendmsg().  This is used by both the initial
 785 * transmission and possible later retransmissions.
 786 * All SKB's seen here are completely headerless.  It is our
 787 * job to build the TCP header, and pass the packet down to
 788 * IP so it can do the same plus pass the packet off to the
 789 * device.
 790 *
 791 * We are working here with either a clone of the original
 792 * SKB, or a fresh unique copy made by the retransmit engine.
 793 */
 794static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
 795			    gfp_t gfp_mask)
 796{
 797	const struct inet_connection_sock *icsk = inet_csk(sk);
 798	struct inet_sock *inet;
 799	struct tcp_sock *tp;
 800	struct tcp_skb_cb *tcb;
 801	struct tcp_out_options opts;
 802	unsigned tcp_options_size, tcp_header_size;
 803	struct tcp_md5sig_key *md5;
 
 804	struct tcphdr *th;
 
 805	int err;
 806
 807	BUG_ON(!skb || !tcp_skb_pcount(skb));
 
 
 
 
 
 
 
 
 
 
 
 
 
 808
 809	/* If congestion control is doing timestamping, we must
 810	 * take such a timestamp before we potentially clone/copy.
 811	 */
 812	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
 813		__net_timestamp(skb);
 814
 815	if (likely(clone_it)) {
 816		if (unlikely(skb_cloned(skb)))
 817			skb = pskb_copy(skb, gfp_mask);
 818		else
 819			skb = skb_clone(skb, gfp_mask);
 820		if (unlikely(!skb))
 821			return -ENOBUFS;
 
 
 
 
 822	}
 823
 824	inet = inet_sk(sk);
 825	tp = tcp_sk(sk);
 826	tcb = TCP_SKB_CB(skb);
 827	memset(&opts, 0, sizeof(opts));
 828
 829	if (unlikely(tcb->flags & TCPHDR_SYN))
 830		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
 831	else
 832		tcp_options_size = tcp_established_options(sk, skb, &opts,
 833							   &md5);
 
 
 
 
 
 
 
 
 
 
 
 834	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
 835
 836	if (tcp_packets_in_flight(tp) == 0) {
 837		tcp_ca_event(sk, CA_EVENT_TX_START);
 838		skb->ooo_okay = 1;
 839	} else
 840		skb->ooo_okay = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 841
 842	skb_push(skb, tcp_header_size);
 843	skb_reset_transport_header(skb);
 844	skb_set_owner_w(skb, sk);
 
 
 
 
 
 
 845
 846	/* Build TCP header and checksum it. */
 847	th = tcp_hdr(skb);
 848	th->source		= inet->inet_sport;
 849	th->dest		= inet->inet_dport;
 850	th->seq			= htonl(tcb->seq);
 851	th->ack_seq		= htonl(tp->rcv_nxt);
 852	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
 853					tcb->flags);
 854
 855	if (unlikely(tcb->flags & TCPHDR_SYN)) {
 856		/* RFC1323: The window in SYN & SYN/ACK segments
 857		 * is never scaled.
 858		 */
 859		th->window	= htons(min(tp->rcv_wnd, 65535U));
 860	} else {
 861		th->window	= htons(tcp_select_window(sk));
 862	}
 863	th->check		= 0;
 864	th->urg_ptr		= 0;
 865
 866	/* The urg_mode check is necessary during a below snd_una win probe */
 867	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
 868		if (before(tp->snd_up, tcb->seq + 0x10000)) {
 869			th->urg_ptr = htons(tp->snd_up - tcb->seq);
 870			th->urg = 1;
 871		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
 872			th->urg_ptr = htons(0xFFFF);
 873			th->urg = 1;
 874		}
 875	}
 876
 877	tcp_options_write((__be32 *)(th + 1), tp, &opts);
 878	if (likely((tcb->flags & TCPHDR_SYN) == 0))
 879		TCP_ECN_send(sk, skb, tcp_header_size);
 
 
 
 
 
 
 
 
 
 880
 
 881#ifdef CONFIG_TCP_MD5SIG
 882	/* Calculate the MD5 hash, as we have all we need now */
 883	if (md5) {
 884		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
 885		tp->af_specific->calc_md5_hash(opts.hash_location,
 886					       md5, sk, NULL, skb);
 887	}
 888#endif
 
 
 889
 890	icsk->icsk_af_ops->send_check(sk, skb);
 
 
 
 
 
 
 891
 892	if (likely(tcb->flags & TCPHDR_ACK))
 893		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
 894
 895	if (skb->len != tcp_header_size)
 896		tcp_event_data_sent(tp, skb, sk);
 
 
 
 
 
 
 
 
 
 
 897
 898	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
 899		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
 900			      tcp_skb_pcount(skb));
 901
 902	err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
 903	if (likely(err <= 0))
 904		return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 905
 906	tcp_enter_cwr(sk, 1);
 907
 908	return net_xmit_eval(err);
 
 
 909}
 910
 911/* This routine just queues the buffer for sending.
 912 *
 913 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
 914 * otherwise socket can stall.
 915 */
 916static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
 917{
 918	struct tcp_sock *tp = tcp_sk(sk);
 919
 920	/* Advance write_seq and place onto the write_queue. */
 921	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
 922	skb_header_release(skb);
 923	tcp_add_write_queue_tail(sk, skb);
 924	sk->sk_wmem_queued += skb->truesize;
 925	sk_mem_charge(sk, skb->truesize);
 926}
 927
 928/* Initialize TSO segments for a packet. */
 929static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
 930				 unsigned int mss_now)
 931{
 932	if (skb->len <= mss_now || !sk_can_gso(sk) ||
 933	    skb->ip_summed == CHECKSUM_NONE) {
 
 934		/* Avoid the costly divide in the normal
 935		 * non-TSO case.
 936		 */
 937		skb_shinfo(skb)->gso_segs = 1;
 938		skb_shinfo(skb)->gso_size = 0;
 939		skb_shinfo(skb)->gso_type = 0;
 940	} else {
 941		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
 942		skb_shinfo(skb)->gso_size = mss_now;
 943		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
 944	}
 945}
 946
 947/* When a modification to fackets out becomes necessary, we need to check
 948 * skb is counted to fackets_out or not.
 949 */
 950static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
 951				   int decr)
 952{
 953	struct tcp_sock *tp = tcp_sk(sk);
 954
 955	if (!tp->sacked_out || tcp_is_reno(tp))
 956		return;
 957
 958	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
 959		tp->fackets_out -= decr;
 960}
 961
 962/* Pcount in the middle of the write queue got changed, we need to do various
 963 * tweaks to fix counters
 964 */
 965static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
 966{
 967	struct tcp_sock *tp = tcp_sk(sk);
 968
 969	tp->packets_out -= decr;
 970
 971	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
 972		tp->sacked_out -= decr;
 973	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
 974		tp->retrans_out -= decr;
 975	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
 976		tp->lost_out -= decr;
 977
 978	/* Reno case is special. Sigh... */
 979	if (tcp_is_reno(tp) && decr > 0)
 980		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
 981
 982	tcp_adjust_fackets_out(sk, skb, decr);
 983
 984	if (tp->lost_skb_hint &&
 985	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
 986	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
 987		tp->lost_cnt_hint -= decr;
 988
 989	tcp_verify_left_out(tp);
 990}
 991
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 992/* Function to create two new TCP segments.  Shrinks the given segment
 993 * to the specified size and appends a new segment with the rest of the
 994 * packet to the list.  This won't be called frequently, I hope.
 995 * Remember, these are still headerless SKBs at this point.
 996 */
 997int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
 998		 unsigned int mss_now)
 
 999{
1000	struct tcp_sock *tp = tcp_sk(sk);
1001	struct sk_buff *buff;
1002	int nsize, old_factor;
 
1003	int nlen;
1004	u8 flags;
1005
1006	if (WARN_ON(len > skb->len))
1007		return -EINVAL;
1008
1009	nsize = skb_headlen(skb) - len;
1010	if (nsize < 0)
1011		nsize = 0;
1012
1013	if (skb_cloned(skb) &&
1014	    skb_is_nonlinear(skb) &&
1015	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
 
 
 
 
 
 
 
 
 
 
1016		return -ENOMEM;
1017
1018	/* Get a new skb... force flag on. */
1019	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1020	if (buff == NULL)
1021		return -ENOMEM; /* We'll just try again later. */
 
 
1022
1023	sk->sk_wmem_queued += buff->truesize;
1024	sk_mem_charge(sk, buff->truesize);
1025	nlen = skb->len - len - nsize;
1026	buff->truesize += nlen;
1027	skb->truesize -= nlen;
1028
1029	/* Correct the sequence numbers. */
1030	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1031	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1032	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1033
1034	/* PSH and FIN should only be set in the second packet. */
1035	flags = TCP_SKB_CB(skb)->flags;
1036	TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1037	TCP_SKB_CB(buff)->flags = flags;
1038	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
 
1039
1040	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1041		/* Copy and checksum data tail into the new buffer. */
1042		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1043						       skb_put(buff, nsize),
1044						       nsize, 0);
1045
1046		skb_trim(skb, len);
1047
1048		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1049	} else {
1050		skb->ip_summed = CHECKSUM_PARTIAL;
1051		skb_split(skb, buff, len);
1052	}
1053
1054	buff->ip_summed = skb->ip_summed;
1055
1056	/* Looks stupid, but our code really uses when of
1057	 * skbs, which it never sent before. --ANK
1058	 */
1059	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1060	buff->tstamp = skb->tstamp;
1061
1062	old_factor = tcp_skb_pcount(skb);
1063
1064	/* Fix up tso_factor for both original and new SKB.  */
1065	tcp_set_skb_tso_segs(sk, skb, mss_now);
1066	tcp_set_skb_tso_segs(sk, buff, mss_now);
 
 
 
1067
1068	/* If this packet has been sent out already, we must
1069	 * adjust the various packet counters.
1070	 */
1071	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1072		int diff = old_factor - tcp_skb_pcount(skb) -
1073			tcp_skb_pcount(buff);
1074
1075		if (diff)
1076			tcp_adjust_pcount(sk, skb, diff);
1077	}
1078
1079	/* Link BUFF into the send queue. */
1080	skb_header_release(buff);
1081	tcp_insert_write_queue_after(skb, buff, sk);
 
 
1082
1083	return 0;
1084}
1085
1086/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1087 * eventually). The difference is that pulled data not copied, but
1088 * immediately discarded.
1089 */
1090static void __pskb_trim_head(struct sk_buff *skb, int len)
1091{
 
1092	int i, k, eat;
1093
 
1094	eat = len;
1095	k = 0;
1096	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1097		if (skb_shinfo(skb)->frags[i].size <= eat) {
1098			put_page(skb_shinfo(skb)->frags[i].page);
1099			eat -= skb_shinfo(skb)->frags[i].size;
 
 
 
1100		} else {
1101			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1102			if (eat) {
1103				skb_shinfo(skb)->frags[k].page_offset += eat;
1104				skb_shinfo(skb)->frags[k].size -= eat;
1105				eat = 0;
1106			}
1107			k++;
1108		}
1109	}
1110	skb_shinfo(skb)->nr_frags = k;
1111
1112	skb_reset_tail_pointer(skb);
1113	skb->data_len -= len;
1114	skb->len = skb->data_len;
 
1115}
1116
1117/* Remove acked data from a packet in the transmit queue. */
1118int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1119{
1120	if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
 
 
1121		return -ENOMEM;
1122
1123	/* If len == headlen, we avoid __skb_pull to preserve alignment. */
1124	if (unlikely(len < skb_headlen(skb)))
1125		__skb_pull(skb, len);
1126	else
1127		__pskb_trim_head(skb, len - skb_headlen(skb));
1128
1129	TCP_SKB_CB(skb)->seq += len;
1130	skb->ip_summed = CHECKSUM_PARTIAL;
1131
1132	skb->truesize	     -= len;
1133	sk->sk_wmem_queued   -= len;
1134	sk_mem_uncharge(sk, len);
1135	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1136
1137	/* Any change of skb->len requires recalculation of tso
1138	 * factor and mss.
1139	 */
1140	if (tcp_skb_pcount(skb) > 1)
1141		tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1142
1143	return 0;
1144}
1145
1146/* Calculate MSS. Not accounting for SACKs here.  */
1147int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1148{
1149	struct tcp_sock *tp = tcp_sk(sk);
1150	struct inet_connection_sock *icsk = inet_csk(sk);
1151	int mss_now;
1152
1153	/* Calculate base mss without TCP options:
1154	   It is MMS_S - sizeof(tcphdr) of rfc1122
1155	 */
1156	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1157
1158	/* Clamp it (mss_clamp does not include tcp options) */
1159	if (mss_now > tp->rx_opt.mss_clamp)
1160		mss_now = tp->rx_opt.mss_clamp;
1161
1162	/* Now subtract optional transport overhead */
1163	mss_now -= icsk->icsk_ext_hdr_len;
1164
1165	/* Then reserve room for full set of TCP options and 8 bytes of data */
1166	if (mss_now < 48)
1167		mss_now = 48;
 
 
1168
1169	/* Now subtract TCP options size, not including SACKs */
1170	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1171
1172	return mss_now;
 
 
1173}
 
1174
1175/* Inverse of above */
1176int tcp_mss_to_mtu(struct sock *sk, int mss)
1177{
1178	struct tcp_sock *tp = tcp_sk(sk);
1179	struct inet_connection_sock *icsk = inet_csk(sk);
1180	int mtu;
1181
1182	mtu = mss +
1183	      tp->tcp_header_len +
1184	      icsk->icsk_ext_hdr_len +
1185	      icsk->icsk_af_ops->net_header_len;
1186
1187	return mtu;
1188}
 
1189
1190/* MTU probing init per socket */
1191void tcp_mtup_init(struct sock *sk)
1192{
1193	struct tcp_sock *tp = tcp_sk(sk);
1194	struct inet_connection_sock *icsk = inet_csk(sk);
 
1195
1196	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1197	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1198			       icsk->icsk_af_ops->net_header_len;
1199	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1200	icsk->icsk_mtup.probe_size = 0;
 
 
1201}
1202EXPORT_SYMBOL(tcp_mtup_init);
1203
1204/* This function synchronize snd mss to current pmtu/exthdr set.
1205
1206   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1207   for TCP options, but includes only bare TCP header.
1208
1209   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1210   It is minimum of user_mss and mss received with SYN.
1211   It also does not include TCP options.
1212
1213   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1214
1215   tp->mss_cache is current effective sending mss, including
1216   all tcp options except for SACKs. It is evaluated,
1217   taking into account current pmtu, but never exceeds
1218   tp->rx_opt.mss_clamp.
1219
1220   NOTE1. rfc1122 clearly states that advertised MSS
1221   DOES NOT include either tcp or ip options.
1222
1223   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1224   are READ ONLY outside this function.		--ANK (980731)
1225 */
1226unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1227{
1228	struct tcp_sock *tp = tcp_sk(sk);
1229	struct inet_connection_sock *icsk = inet_csk(sk);
1230	int mss_now;
1231
1232	if (icsk->icsk_mtup.search_high > pmtu)
1233		icsk->icsk_mtup.search_high = pmtu;
1234
1235	mss_now = tcp_mtu_to_mss(sk, pmtu);
1236	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1237
1238	/* And store cached results */
1239	icsk->icsk_pmtu_cookie = pmtu;
1240	if (icsk->icsk_mtup.enabled)
1241		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1242	tp->mss_cache = mss_now;
1243
1244	return mss_now;
1245}
1246EXPORT_SYMBOL(tcp_sync_mss);
1247
1248/* Compute the current effective MSS, taking SACKs and IP options,
1249 * and even PMTU discovery events into account.
1250 */
1251unsigned int tcp_current_mss(struct sock *sk)
1252{
1253	struct tcp_sock *tp = tcp_sk(sk);
1254	struct dst_entry *dst = __sk_dst_get(sk);
1255	u32 mss_now;
1256	unsigned header_len;
1257	struct tcp_out_options opts;
1258	struct tcp_md5sig_key *md5;
1259
1260	mss_now = tp->mss_cache;
1261
1262	if (dst) {
1263		u32 mtu = dst_mtu(dst);
1264		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1265			mss_now = tcp_sync_mss(sk, mtu);
1266	}
1267
1268	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1269		     sizeof(struct tcphdr);
1270	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1271	 * some common options. If this is an odd packet (because we have SACK
1272	 * blocks etc) then our calculated header_len will be different, and
1273	 * we have to adjust mss_now correspondingly */
1274	if (header_len != tp->tcp_header_len) {
1275		int delta = (int) header_len - tp->tcp_header_len;
1276		mss_now -= delta;
1277	}
1278
1279	return mss_now;
1280}
1281
1282/* Congestion window validation. (RFC2861) */
1283static void tcp_cwnd_validate(struct sock *sk)
 
 
 
1284{
1285	struct tcp_sock *tp = tcp_sk(sk);
1286
1287	if (tp->packets_out >= tp->snd_cwnd) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1288		/* Network is feed fully. */
1289		tp->snd_cwnd_used = 0;
1290		tp->snd_cwnd_stamp = tcp_time_stamp;
1291	} else {
1292		/* Network starves. */
1293		if (tp->packets_out > tp->snd_cwnd_used)
1294			tp->snd_cwnd_used = tp->packets_out;
1295
1296		if (sysctl_tcp_slow_start_after_idle &&
1297		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
 
1298			tcp_cwnd_application_limited(sk);
 
 
 
 
 
 
 
 
 
 
 
 
1299	}
1300}
1301
1302/* Returns the portion of skb which can be sent right away without
1303 * introducing MSS oddities to segment boundaries. In rare cases where
1304 * mss_now != mss_cache, we will request caller to create a small skb
1305 * per input skb which could be mostly avoided here (if desired).
1306 *
1307 * We explicitly want to create a request for splitting write queue tail
1308 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1309 * thus all the complexity (cwnd_len is always MSS multiple which we
1310 * return whenever allowed by the other factors). Basically we need the
1311 * modulo only when the receiver window alone is the limiting factor or
1312 * when we would be allowed to send the split-due-to-Nagle skb fully.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1313 */
1314static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1315					unsigned int mss_now, unsigned int cwnd)
1316{
1317	struct tcp_sock *tp = tcp_sk(sk);
1318	u32 needed, window, cwnd_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1319
1320	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1321	cwnd_len = mss_now * cwnd;
1322
1323	if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1324		return cwnd_len;
1325
1326	needed = min(skb->len, window);
1327
1328	if (cwnd_len <= needed)
1329		return cwnd_len;
 
 
 
 
 
 
 
 
1330
1331	return needed - needed % mss_now;
1332}
1333
1334/* Can at least one segment of SKB be sent right now, according to the
1335 * congestion window rules?  If so, return how many segments are allowed.
1336 */
1337static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1338					 struct sk_buff *skb)
1339{
1340	u32 in_flight, cwnd;
1341
1342	/* Don't be strict about the congestion window for the final FIN.  */
1343	if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
1344		return 1;
1345
1346	in_flight = tcp_packets_in_flight(tp);
1347	cwnd = tp->snd_cwnd;
1348	if (in_flight < cwnd)
1349		return (cwnd - in_flight);
1350
1351	return 0;
 
 
 
 
1352}
1353
1354/* Initialize TSO state of a skb.
1355 * This must be invoked the first time we consider transmitting
1356 * SKB onto the wire.
1357 */
1358static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1359			     unsigned int mss_now)
1360{
1361	int tso_segs = tcp_skb_pcount(skb);
1362
1363	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1364		tcp_set_skb_tso_segs(sk, skb, mss_now);
1365		tso_segs = tcp_skb_pcount(skb);
1366	}
1367	return tso_segs;
1368}
1369
1370/* Minshall's variant of the Nagle send check. */
1371static inline int tcp_minshall_check(const struct tcp_sock *tp)
1372{
1373	return after(tp->snd_sml, tp->snd_una) &&
1374		!after(tp->snd_sml, tp->snd_nxt);
1375}
1376
1377/* Return 0, if packet can be sent now without violation Nagle's rules:
1378 * 1. It is full sized.
1379 * 2. Or it contains FIN. (already checked by caller)
1380 * 3. Or TCP_NODELAY was set.
1381 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1382 *    With Minshall's modification: all sent small packets are ACKed.
1383 */
1384static inline int tcp_nagle_check(const struct tcp_sock *tp,
1385				  const struct sk_buff *skb,
1386				  unsigned mss_now, int nonagle)
1387{
1388	return skb->len < mss_now &&
1389		((nonagle & TCP_NAGLE_CORK) ||
1390		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1391}
1392
1393/* Return non-zero if the Nagle test allows this packet to be
1394 * sent now.
1395 */
1396static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1397				 unsigned int cur_mss, int nonagle)
1398{
1399	/* Nagle rule does not apply to frames, which sit in the middle of the
1400	 * write_queue (they have no chances to get new data).
1401	 *
1402	 * This is implemented in the callers, where they modify the 'nonagle'
1403	 * argument based upon the location of SKB in the send queue.
1404	 */
1405	if (nonagle & TCP_NAGLE_PUSH)
1406		return 1;
1407
1408	/* Don't use the nagle rule for urgent data (or for the final FIN).
1409	 * Nagle can be ignored during F-RTO too (see RFC4138).
1410	 */
1411	if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1412	    (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
1413		return 1;
1414
1415	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1416		return 1;
1417
1418	return 0;
1419}
1420
1421/* Does at least the first segment of SKB fit into the send window? */
1422static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1423				   unsigned int cur_mss)
 
1424{
1425	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1426
1427	if (skb->len > cur_mss)
1428		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1429
1430	return !after(end_seq, tcp_wnd_end(tp));
1431}
1432
1433/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1434 * should be put on the wire right now.  If so, it returns the number of
1435 * packets allowed by the congestion window.
1436 */
1437static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1438				 unsigned int cur_mss, int nonagle)
1439{
1440	struct tcp_sock *tp = tcp_sk(sk);
1441	unsigned int cwnd_quota;
1442
1443	tcp_init_tso_segs(sk, skb, cur_mss);
1444
1445	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1446		return 0;
1447
1448	cwnd_quota = tcp_cwnd_test(tp, skb);
1449	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1450		cwnd_quota = 0;
1451
1452	return cwnd_quota;
1453}
1454
1455/* Test if sending is allowed right now. */
1456int tcp_may_send_now(struct sock *sk)
1457{
1458	struct tcp_sock *tp = tcp_sk(sk);
1459	struct sk_buff *skb = tcp_send_head(sk);
1460
1461	return skb &&
1462		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1463			     (tcp_skb_is_last(sk, skb) ?
1464			      tp->nonagle : TCP_NAGLE_PUSH));
1465}
1466
1467/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1468 * which is put after SKB on the list.  It is very much like
1469 * tcp_fragment() except that it may make several kinds of assumptions
1470 * in order to speed up the splitting operation.  In particular, we
1471 * know that all the data is in scatter-gather pages, and that the
1472 * packet has never been sent out before (and thus is not cloned).
1473 */
1474static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1475			unsigned int mss_now, gfp_t gfp)
1476{
 
1477	struct sk_buff *buff;
1478	int nlen = skb->len - len;
1479	u8 flags;
1480
1481	/* All of a TSO frame must be composed of paged data.  */
1482	if (skb->len != skb->data_len)
1483		return tcp_fragment(sk, skb, len, mss_now);
1484
1485	buff = sk_stream_alloc_skb(sk, 0, gfp);
1486	if (unlikely(buff == NULL))
1487		return -ENOMEM;
 
 
1488
1489	sk->sk_wmem_queued += buff->truesize;
1490	sk_mem_charge(sk, buff->truesize);
1491	buff->truesize += nlen;
1492	skb->truesize -= nlen;
1493
1494	/* Correct the sequence numbers. */
1495	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1496	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1497	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1498
1499	/* PSH and FIN should only be set in the second packet. */
1500	flags = TCP_SKB_CB(skb)->flags;
1501	TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1502	TCP_SKB_CB(buff)->flags = flags;
1503
1504	/* This packet was never sent out yet, so no SACK bits. */
1505	TCP_SKB_CB(buff)->sacked = 0;
1506
1507	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1508	skb_split(skb, buff, len);
 
1509
1510	/* Fix up tso_factor for both original and new SKB.  */
1511	tcp_set_skb_tso_segs(sk, skb, mss_now);
1512	tcp_set_skb_tso_segs(sk, buff, mss_now);
1513
1514	/* Link BUFF into the send queue. */
1515	skb_header_release(buff);
1516	tcp_insert_write_queue_after(skb, buff, sk);
1517
1518	return 0;
1519}
1520
1521/* Try to defer sending, if possible, in order to minimize the amount
1522 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1523 *
1524 * This algorithm is from John Heffner.
1525 */
1526static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
 
 
 
1527{
1528	struct tcp_sock *tp = tcp_sk(sk);
1529	const struct inet_connection_sock *icsk = inet_csk(sk);
1530	u32 send_win, cong_win, limit, in_flight;
 
 
1531	int win_divisor;
 
1532
1533	if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
1534		goto send_now;
1535
1536	if (icsk->icsk_ca_state != TCP_CA_Open)
1537		goto send_now;
1538
1539	/* Defer for less than two clock ticks. */
1540	if (tp->tso_deferred &&
1541	    (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
 
1542		goto send_now;
1543
1544	in_flight = tcp_packets_in_flight(tp);
1545
1546	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
 
1547
1548	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1549
1550	/* From in_flight test above, we know that cwnd > in_flight.  */
1551	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1552
1553	limit = min(send_win, cong_win);
1554
1555	/* If a full-sized TSO skb can be sent, do it. */
1556	if (limit >= sk->sk_gso_max_size)
1557		goto send_now;
1558
1559	/* Middle in queue won't get any more data, full sendable already? */
1560	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1561		goto send_now;
1562
1563	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1564	if (win_divisor) {
1565		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1566
1567		/* If at least some fraction of a window is available,
1568		 * just use it.
1569		 */
1570		chunk /= win_divisor;
1571		if (limit >= chunk)
1572			goto send_now;
1573	} else {
1574		/* Different approach, try not to defer past a single
1575		 * ACK.  Receiver should ACK every other full sized
1576		 * frame, so if we have space for more than 3 frames
1577		 * then send now.
1578		 */
1579		if (limit > tcp_max_burst(tp) * tp->mss_cache)
1580			goto send_now;
1581	}
1582
1583	/* Ok, it looks like it is advisable to defer.  */
1584	tp->tso_deferred = 1 | (jiffies << 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1585
1586	return 1;
1587
1588send_now:
1589	tp->tso_deferred = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1590	return 0;
1591}
1592
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1593/* Create a new MTU probe if we are ready.
1594 * MTU probe is regularly attempting to increase the path MTU by
1595 * deliberately sending larger packets.  This discovers routing
1596 * changes resulting in larger path MTUs.
1597 *
1598 * Returns 0 if we should wait to probe (no cwnd available),
1599 *         1 if a probe was sent,
1600 *         -1 otherwise
1601 */
1602static int tcp_mtu_probe(struct sock *sk)
1603{
 
1604	struct tcp_sock *tp = tcp_sk(sk);
1605	struct inet_connection_sock *icsk = inet_csk(sk);
1606	struct sk_buff *skb, *nskb, *next;
1607	int len;
1608	int probe_size;
1609	int size_needed;
1610	int copy;
1611	int mss_now;
 
1612
1613	/* Not currently probing/verifying,
1614	 * not in recovery,
1615	 * have enough cwnd, and
1616	 * not SACKing (the variable headers throw things off) */
1617	if (!icsk->icsk_mtup.enabled ||
1618	    icsk->icsk_mtup.probe_size ||
1619	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1620	    tp->snd_cwnd < 11 ||
1621	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
 
1622		return -1;
1623
1624	/* Very simple search strategy: just double the MSS. */
 
 
 
1625	mss_now = tcp_current_mss(sk);
1626	probe_size = 2 * tp->mss_cache;
 
1627	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1628	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1629		/* TODO: set timer for probe_converge_event */
 
 
 
 
 
 
 
 
 
1630		return -1;
1631	}
1632
1633	/* Have enough data in the send queue to probe? */
1634	if (tp->write_seq - tp->snd_nxt < size_needed)
1635		return -1;
1636
1637	if (tp->snd_wnd < size_needed)
1638		return -1;
1639	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1640		return 0;
1641
1642	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1643	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1644		if (!tcp_packets_in_flight(tp))
1645			return -1;
1646		else
1647			return 0;
1648	}
1649
 
 
 
1650	/* We're allowed to probe.  Build it now. */
1651	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
 
1652		return -1;
1653	sk->sk_wmem_queued += nskb->truesize;
 
 
 
 
 
 
 
1654	sk_mem_charge(sk, nskb->truesize);
1655
1656	skb = tcp_send_head(sk);
 
 
1657
1658	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1659	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1660	TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
1661	TCP_SKB_CB(nskb)->sacked = 0;
1662	nskb->csum = 0;
1663	nskb->ip_summed = skb->ip_summed;
1664
1665	tcp_insert_write_queue_before(nskb, skb, sk);
 
1666
1667	len = 0;
1668	tcp_for_write_queue_from_safe(skb, next, sk) {
1669		copy = min_t(int, skb->len, probe_size - len);
1670		if (nskb->ip_summed)
1671			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1672		else
1673			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1674							    skb_put(nskb, copy),
1675							    copy, nskb->csum);
1676
1677		if (skb->len <= copy) {
1678			/* We've eaten all the data from this skb.
1679			 * Throw it away. */
1680			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1681			tcp_unlink_write_queue(skb, sk);
1682			sk_wmem_free_skb(sk, skb);
1683		} else {
1684			TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1685						   ~(TCPHDR_FIN|TCPHDR_PSH);
1686			if (!skb_shinfo(skb)->nr_frags) {
1687				skb_pull(skb, copy);
1688				if (skb->ip_summed != CHECKSUM_PARTIAL)
1689					skb->csum = csum_partial(skb->data,
1690								 skb->len, 0);
1691			} else {
1692				__pskb_trim_head(skb, copy);
1693				tcp_set_skb_tso_segs(sk, skb, mss_now);
1694			}
1695			TCP_SKB_CB(skb)->seq += copy;
1696		}
1697
1698		len += copy;
1699
1700		if (len >= probe_size)
1701			break;
1702	}
1703	tcp_init_tso_segs(sk, nskb, nskb->len);
1704
1705	/* We're ready to send.  If this fails, the probe will
1706	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1707	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1708	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1709		/* Decrement cwnd here because we are sending
1710		 * effectively two packets. */
1711		tp->snd_cwnd--;
1712		tcp_event_new_data_sent(sk, nskb);
1713
1714		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1715		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1716		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1717
1718		return 1;
1719	}
1720
1721	return -1;
1722}
1723
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1724/* This routine writes packets to the network.  It advances the
1725 * send_head.  This happens as incoming acks open up the remote
1726 * window for us.
1727 *
1728 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1729 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1730 * account rare use of URG, this is not a big flaw.
1731 *
1732 * Returns 1, if no segments are in flight and we have queued segments, but
1733 * cannot send anything now because of SWS or another problem.
 
 
 
1734 */
1735static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1736			  int push_one, gfp_t gfp)
1737{
1738	struct tcp_sock *tp = tcp_sk(sk);
1739	struct sk_buff *skb;
1740	unsigned int tso_segs, sent_pkts;
1741	int cwnd_quota;
1742	int result;
 
1743
1744	sent_pkts = 0;
1745
 
1746	if (!push_one) {
1747		/* Do MTU probing. */
1748		result = tcp_mtu_probe(sk);
1749		if (!result) {
1750			return 0;
1751		} else if (result > 0) {
1752			sent_pkts = 1;
1753		}
1754	}
1755
 
1756	while ((skb = tcp_send_head(sk))) {
1757		unsigned int limit;
 
1758
1759		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1760		BUG_ON(!tso_segs);
 
 
 
 
 
 
1761
1762		cwnd_quota = tcp_cwnd_test(tp, skb);
1763		if (!cwnd_quota)
1764			break;
1765
1766		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1767			break;
 
1768
1769		if (tso_segs == 1) {
1770			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1771						     (tcp_skb_is_last(sk, skb) ?
1772						      nonagle : TCP_NAGLE_PUSH))))
1773				break;
1774		} else {
1775			if (!push_one && tcp_tso_should_defer(sk, skb))
 
 
1776				break;
1777		}
1778
1779		limit = mss_now;
1780		if (tso_segs > 1 && !tcp_urg_mode(tp))
1781			limit = tcp_mss_split_point(sk, skb, mss_now,
1782						    cwnd_quota);
 
1783
1784		if (skb->len > limit &&
1785		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1786			break;
1787
1788		TCP_SKB_CB(skb)->when = tcp_time_stamp;
 
 
 
 
 
 
 
 
 
1789
1790		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1791			break;
1792
 
1793		/* Advance the send_head.  This one is sent out.
1794		 * This call will increment packets_out.
1795		 */
1796		tcp_event_new_data_sent(sk, skb);
1797
1798		tcp_minshall_update(tp, mss_now, skb);
1799		sent_pkts++;
1800
1801		if (push_one)
1802			break;
1803	}
1804
 
 
 
 
 
 
 
 
 
1805	if (likely(sent_pkts)) {
1806		tcp_cwnd_validate(sk);
1807		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1808	}
1809	return !tp->packets_out && tcp_send_head(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1810}
1811
1812/* Push out any pending frames which were held back due to
1813 * TCP_CORK or attempt at coalescing tiny packets.
1814 * The socket must be locked by the caller.
1815 */
1816void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1817			       int nonagle)
1818{
1819	/* If we are closed, the bytes will have to remain here.
1820	 * In time closedown will finish, we empty the write queue and
1821	 * all will be happy.
1822	 */
1823	if (unlikely(sk->sk_state == TCP_CLOSE))
1824		return;
1825
1826	if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
 
1827		tcp_check_probe_timer(sk);
1828}
1829
1830/* Send _single_ skb sitting at the send head. This function requires
1831 * true push pending frames to setup probe timer etc.
1832 */
1833void tcp_push_one(struct sock *sk, unsigned int mss_now)
1834{
1835	struct sk_buff *skb = tcp_send_head(sk);
1836
1837	BUG_ON(!skb || skb->len < mss_now);
1838
1839	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1840}
1841
1842/* This function returns the amount that we can raise the
1843 * usable window based on the following constraints
1844 *
1845 * 1. The window can never be shrunk once it is offered (RFC 793)
1846 * 2. We limit memory per socket
1847 *
1848 * RFC 1122:
1849 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1850 *  RECV.NEXT + RCV.WIN fixed until:
1851 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1852 *
1853 * i.e. don't raise the right edge of the window until you can raise
1854 * it at least MSS bytes.
1855 *
1856 * Unfortunately, the recommended algorithm breaks header prediction,
1857 * since header prediction assumes th->window stays fixed.
1858 *
1859 * Strictly speaking, keeping th->window fixed violates the receiver
1860 * side SWS prevention criteria. The problem is that under this rule
1861 * a stream of single byte packets will cause the right side of the
1862 * window to always advance by a single byte.
1863 *
1864 * Of course, if the sender implements sender side SWS prevention
1865 * then this will not be a problem.
1866 *
1867 * BSD seems to make the following compromise:
1868 *
1869 *	If the free space is less than the 1/4 of the maximum
1870 *	space available and the free space is less than 1/2 mss,
1871 *	then set the window to 0.
1872 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1873 *	Otherwise, just prevent the window from shrinking
1874 *	and from being larger than the largest representable value.
1875 *
1876 * This prevents incremental opening of the window in the regime
1877 * where TCP is limited by the speed of the reader side taking
1878 * data out of the TCP receive queue. It does nothing about
1879 * those cases where the window is constrained on the sender side
1880 * because the pipeline is full.
1881 *
1882 * BSD also seems to "accidentally" limit itself to windows that are a
1883 * multiple of MSS, at least until the free space gets quite small.
1884 * This would appear to be a side effect of the mbuf implementation.
1885 * Combining these two algorithms results in the observed behavior
1886 * of having a fixed window size at almost all times.
1887 *
1888 * Below we obtain similar behavior by forcing the offered window to
1889 * a multiple of the mss when it is feasible to do so.
1890 *
1891 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1892 * Regular options like TIMESTAMP are taken into account.
1893 */
1894u32 __tcp_select_window(struct sock *sk)
1895{
1896	struct inet_connection_sock *icsk = inet_csk(sk);
1897	struct tcp_sock *tp = tcp_sk(sk);
 
1898	/* MSS for the peer's data.  Previous versions used mss_clamp
1899	 * here.  I don't know if the value based on our guesses
1900	 * of peer's MSS is better for the performance.  It's more correct
1901	 * but may be worse for the performance because of rcv_mss
1902	 * fluctuations.  --SAW  1998/11/1
1903	 */
1904	int mss = icsk->icsk_ack.rcv_mss;
1905	int free_space = tcp_space(sk);
1906	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1907	int window;
 
 
 
 
 
1908
1909	if (mss > full_space)
1910		mss = full_space;
 
 
 
 
 
 
 
 
 
 
 
1911
1912	if (free_space < (full_space >> 1)) {
1913		icsk->icsk_ack.quick = 0;
1914
1915		if (tcp_memory_pressure)
1916			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1917					       4U * tp->advmss);
 
 
 
 
1918
1919		if (free_space < mss)
 
 
 
 
 
 
 
1920			return 0;
1921	}
1922
1923	if (free_space > tp->rcv_ssthresh)
1924		free_space = tp->rcv_ssthresh;
1925
1926	/* Don't do rounding if we are using window scaling, since the
1927	 * scaled window will not line up with the MSS boundary anyway.
1928	 */
1929	window = tp->rcv_wnd;
1930	if (tp->rx_opt.rcv_wscale) {
1931		window = free_space;
1932
1933		/* Advertise enough space so that it won't get scaled away.
1934		 * Import case: prevent zero window announcement if
1935		 * 1<<rcv_wscale > mss.
1936		 */
1937		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1938			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1939				  << tp->rx_opt.rcv_wscale);
1940	} else {
 
1941		/* Get the largest window that is a nice multiple of mss.
1942		 * Window clamp already applied above.
1943		 * If our current window offering is within 1 mss of the
1944		 * free space we just keep it. This prevents the divide
1945		 * and multiply from happening most of the time.
1946		 * We also don't do any window rounding when the free space
1947		 * is too small.
1948		 */
1949		if (window <= free_space - mss || window > free_space)
1950			window = (free_space / mss) * mss;
1951		else if (mss == full_space &&
1952			 free_space > window + (full_space >> 1))
1953			window = free_space;
1954	}
1955
1956	return window;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1957}
1958
1959/* Collapses two adjacent SKB's during retransmission. */
1960static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1961{
1962	struct tcp_sock *tp = tcp_sk(sk);
1963	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1964	int skb_size, next_skb_size;
1965
1966	skb_size = skb->len;
1967	next_skb_size = next_skb->len;
1968
1969	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1970
1971	tcp_highest_sack_combine(sk, next_skb, skb);
1972
1973	tcp_unlink_write_queue(next_skb, sk);
1974
1975	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1976				  next_skb_size);
1977
1978	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1979		skb->ip_summed = CHECKSUM_PARTIAL;
1980
1981	if (skb->ip_summed != CHECKSUM_PARTIAL)
1982		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1983
1984	/* Update sequence range on original skb. */
1985	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1986
1987	/* Merge over control information. This moves PSH/FIN etc. over */
1988	TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1989
1990	/* All done, get rid of second SKB and account for it so
1991	 * packet counting does not break.
1992	 */
1993	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
 
1994
1995	/* changed transmit queue under us so clear hints */
1996	tcp_clear_retrans_hints_partial(tp);
1997	if (next_skb == tp->retransmit_skb_hint)
1998		tp->retransmit_skb_hint = skb;
1999
2000	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2001
2002	sk_wmem_free_skb(sk, next_skb);
 
 
 
2003}
2004
2005/* Check if coalescing SKBs is legal. */
2006static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2007{
2008	if (tcp_skb_pcount(skb) > 1)
2009		return 0;
2010	/* TODO: SACK collapsing could be used to remove this condition */
2011	if (skb_shinfo(skb)->nr_frags != 0)
2012		return 0;
2013	if (skb_cloned(skb))
2014		return 0;
2015	if (skb == tcp_send_head(sk))
2016		return 0;
2017	/* Some heurestics for collapsing over SACK'd could be invented */
2018	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2019		return 0;
2020
2021	return 1;
2022}
2023
2024/* Collapse packets in the retransmit queue to make to create
2025 * less packets on the wire. This is only done on retransmission.
2026 */
2027static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2028				     int space)
2029{
2030	struct tcp_sock *tp = tcp_sk(sk);
2031	struct sk_buff *skb = to, *tmp;
2032	int first = 1;
2033
2034	if (!sysctl_tcp_retrans_collapse)
2035		return;
2036	if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
2037		return;
2038
2039	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2040		if (!tcp_can_collapse(sk, skb))
2041			break;
2042
 
 
 
2043		space -= skb->len;
2044
2045		if (first) {
2046			first = 0;
2047			continue;
2048		}
2049
2050		if (space < 0)
2051			break;
2052		/* Punt if not enough space exists in the first SKB for
2053		 * the data in the second
2054		 */
2055		if (skb->len > skb_tailroom(to))
2056			break;
2057
2058		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2059			break;
2060
2061		tcp_collapse_retrans(sk, to);
 
2062	}
2063}
2064
2065/* This retransmits one SKB.  Policy decisions and retransmit queue
2066 * state updates are done by the caller.  Returns non-zero if an
2067 * error occurred which prevented the send.
2068 */
2069int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2070{
 
2071	struct tcp_sock *tp = tcp_sk(sk);
2072	struct inet_connection_sock *icsk = inet_csk(sk);
2073	unsigned int cur_mss;
2074	int err;
 
2075
2076	/* Inconslusive MTU probe */
2077	if (icsk->icsk_mtup.probe_size) {
2078		icsk->icsk_mtup.probe_size = 0;
2079	}
2080
2081	/* Do not sent more than we queued. 1/4 is reserved for possible
2082	 * copying overhead: fragmentation, tunneling, mangling etc.
2083	 */
2084	if (atomic_read(&sk->sk_wmem_alloc) >
2085	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2086		return -EAGAIN;
2087
 
2088	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2089		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2090			BUG();
 
 
 
 
 
 
 
2091		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2092			return -ENOMEM;
2093	}
2094
2095	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2096		return -EHOSTUNREACH; /* Routing failure or similar. */
2097
2098	cur_mss = tcp_current_mss(sk);
 
2099
2100	/* If receiver has shrunk his window, and skb is out of
2101	 * new window, do not retransmit it. The exception is the
2102	 * case, when window is shrunk to zero. In this case
2103	 * our retransmit serves as a zero window probe.
2104	 */
2105	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2106	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2107		return -EAGAIN;
 
 
2108
2109	if (skb->len > cur_mss) {
2110		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
 
 
 
 
 
 
 
2111			return -ENOMEM; /* We'll try again later. */
2112	} else {
2113		int oldpcount = tcp_skb_pcount(skb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2114
2115		if (unlikely(oldpcount > 1)) {
2116			tcp_init_tso_segs(sk, skb, cur_mss);
2117			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2118		}
 
 
2119	}
2120
2121	tcp_retrans_try_collapse(sk, skb, cur_mss);
 
 
 
 
 
 
 
 
2122
2123	/* Some Solaris stacks overoptimize and ignore the FIN on a
2124	 * retransmit when old data is attached.  So strip it off
2125	 * since it is cheap to do so and saves bytes on the network.
2126	 */
2127	if (skb->len > 0 &&
2128	    (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
2129	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2130		if (!pskb_trim(skb, 0)) {
2131			/* Reuse, even though it does some unnecessary work */
2132			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2133					     TCP_SKB_CB(skb)->flags);
2134			skb->ip_summed = CHECKSUM_NONE;
2135		}
2136	}
2137
2138	/* Make a copy, if the first transmission SKB clone we made
2139	 * is still in somebody's hands, else make a clone.
2140	 */
2141	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2142
2143	err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
 
 
 
2144
2145	if (err == 0) {
2146		/* Update global TCP statistics. */
2147		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2148
2149		tp->total_retrans++;
2150
2151#if FASTRETRANS_DEBUG > 0
2152		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2153			if (net_ratelimit())
2154				printk(KERN_DEBUG "retrans_out leaked.\n");
2155		}
2156#endif
2157		if (!tp->retrans_out)
2158			tp->lost_retrans_low = tp->snd_nxt;
2159		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2160		tp->retrans_out += tcp_skb_pcount(skb);
 
2161
2162		/* Save stamp of the first retransmit. */
2163		if (!tp->retrans_stamp)
2164			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2165
2166		tp->undo_retrans += tcp_skb_pcount(skb);
2167
2168		/* snd_nxt is stored to detect loss of retransmitted segment,
2169		 * see tcp_input.c tcp_sacktag_write_queue().
2170		 */
2171		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2172	}
2173	return err;
2174}
2175
2176/* Check if we forward retransmits are possible in the current
2177 * window/congestion state.
2178 */
2179static int tcp_can_forward_retransmit(struct sock *sk)
2180{
2181	const struct inet_connection_sock *icsk = inet_csk(sk);
2182	struct tcp_sock *tp = tcp_sk(sk);
2183
2184	/* Forward retransmissions are possible only during Recovery. */
2185	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2186		return 0;
2187
2188	/* No forward retransmissions in Reno are possible. */
2189	if (tcp_is_reno(tp))
2190		return 0;
2191
2192	/* Yeah, we have to make difficult choice between forward transmission
2193	 * and retransmission... Both ways have their merits...
2194	 *
2195	 * For now we do not retransmit anything, while we have some new
2196	 * segments to send. In the other cases, follow rule 3 for
2197	 * NextSeg() specified in RFC3517.
2198	 */
2199
2200	if (tcp_may_send_now(sk))
2201		return 0;
2202
2203	return 1;
2204}
2205
2206/* This gets called after a retransmit timeout, and the initially
2207 * retransmitted data is acknowledged.  It tries to continue
2208 * resending the rest of the retransmit queue, until either
2209 * we've sent it all or the congestion window limit is reached.
2210 * If doing SACK, the first ACK which comes back for a timeout
2211 * based retransmit packet might feed us FACK information again.
2212 * If so, we use it to avoid unnecessarily retransmissions.
2213 */
2214void tcp_xmit_retransmit_queue(struct sock *sk)
2215{
2216	const struct inet_connection_sock *icsk = inet_csk(sk);
 
2217	struct tcp_sock *tp = tcp_sk(sk);
2218	struct sk_buff *skb;
2219	struct sk_buff *hole = NULL;
2220	u32 last_lost;
2221	int mib_idx;
2222	int fwd_rexmitting = 0;
2223
2224	if (!tp->packets_out)
2225		return;
2226
2227	if (!tp->lost_out)
2228		tp->retransmit_high = tp->snd_una;
 
 
 
 
2229
2230	if (tp->retransmit_skb_hint) {
2231		skb = tp->retransmit_skb_hint;
2232		last_lost = TCP_SKB_CB(skb)->end_seq;
2233		if (after(last_lost, tp->retransmit_high))
2234			last_lost = tp->retransmit_high;
2235	} else {
2236		skb = tcp_write_queue_head(sk);
2237		last_lost = tp->snd_una;
2238	}
2239
2240	tcp_for_write_queue_from(skb, sk) {
2241		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2242
2243		if (skb == tcp_send_head(sk))
2244			break;
2245		/* we could do better than to assign each time */
2246		if (hole == NULL)
2247			tp->retransmit_skb_hint = skb;
2248
2249		/* Assume this retransmit will generate
2250		 * only one packet for congestion window
2251		 * calculation purposes.  This works because
2252		 * tcp_retransmit_skb() will chop up the
2253		 * packet to be MSS sized and all the
2254		 * packet counting works out.
2255		 */
2256		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2257			return;
2258
2259		if (fwd_rexmitting) {
2260begin_fwd:
2261			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2262				break;
2263			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2264
2265		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2266			tp->retransmit_high = last_lost;
2267			if (!tcp_can_forward_retransmit(sk))
2268				break;
2269			/* Backtrack if necessary to non-L'ed skb */
2270			if (hole != NULL) {
2271				skb = hole;
2272				hole = NULL;
2273			}
2274			fwd_rexmitting = 1;
2275			goto begin_fwd;
2276
 
 
2277		} else if (!(sacked & TCPCB_LOST)) {
2278			if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2279				hole = skb;
2280			continue;
2281
2282		} else {
2283			last_lost = TCP_SKB_CB(skb)->end_seq;
2284			if (icsk->icsk_ca_state != TCP_CA_Loss)
2285				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2286			else
2287				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2288		}
2289
2290		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2291			continue;
2292
2293		if (tcp_retransmit_skb(sk, skb))
2294			return;
2295		NET_INC_STATS_BH(sock_net(sk), mib_idx);
 
 
 
 
 
 
 
 
 
 
 
2296
2297		if (skb == tcp_write_queue_head(sk))
2298			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2299						  inet_csk(sk)->icsk_rto,
2300						  TCP_RTO_MAX);
2301	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2302}
2303
2304/* Send a fin.  The caller locks the socket for us.  This cannot be
2305 * allowed to fail queueing a FIN frame under any circumstances.
2306 */
2307void tcp_send_fin(struct sock *sk)
2308{
 
2309	struct tcp_sock *tp = tcp_sk(sk);
2310	struct sk_buff *skb = tcp_write_queue_tail(sk);
2311	int mss_now;
2312
2313	/* Optimization, tack on the FIN if we have a queue of
2314	 * unsent frames.  But be careful about outgoing SACKS
2315	 * and IP options.
 
2316	 */
2317	mss_now = tcp_current_mss(sk);
2318
2319	if (tcp_send_head(sk) != NULL) {
2320		TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
2321		TCP_SKB_CB(skb)->end_seq++;
 
 
2322		tp->write_seq++;
 
 
 
 
 
 
 
 
 
 
2323	} else {
2324		/* Socket is locked, keep trying until memory is available. */
2325		for (;;) {
2326			skb = alloc_skb_fclone(MAX_TCP_HEADER,
2327					       sk->sk_allocation);
2328			if (skb)
2329				break;
2330			yield();
2331		}
2332
2333		/* Reserve space for headers and prepare control bits. */
2334		skb_reserve(skb, MAX_TCP_HEADER);
 
2335		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2336		tcp_init_nondata_skb(skb, tp->write_seq,
2337				     TCPHDR_ACK | TCPHDR_FIN);
2338		tcp_queue_skb(sk, skb);
2339	}
2340	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2341}
2342
2343/* We get here when a process closes a file descriptor (either due to
2344 * an explicit close() or as a byproduct of exit()'ing) and there
2345 * was unread data in the receive queue.  This behavior is recommended
2346 * by RFC 2525, section 2.17.  -DaveM
2347 */
2348void tcp_send_active_reset(struct sock *sk, gfp_t priority)
 
2349{
2350	struct sk_buff *skb;
2351
 
 
2352	/* NOTE: No TCP options attached and we never retransmit this. */
2353	skb = alloc_skb(MAX_TCP_HEADER, priority);
2354	if (!skb) {
2355		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2356		return;
2357	}
2358
2359	/* Reserve space for headers and prepare control bits. */
2360	skb_reserve(skb, MAX_TCP_HEADER);
2361	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2362			     TCPHDR_ACK | TCPHDR_RST);
 
2363	/* Send it off. */
2364	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2365	if (tcp_transmit_skb(sk, skb, 0, priority))
2366		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2367
2368	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
 
 
 
2369}
2370
2371/* Send a crossed SYN-ACK during socket establishment.
2372 * WARNING: This routine must only be called when we have already sent
2373 * a SYN packet that crossed the incoming SYN that caused this routine
2374 * to get called. If this assumption fails then the initial rcv_wnd
2375 * and rcv_wscale values will not be correct.
2376 */
2377int tcp_send_synack(struct sock *sk)
2378{
2379	struct sk_buff *skb;
2380
2381	skb = tcp_write_queue_head(sk);
2382	if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
2383		printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2384		return -EFAULT;
2385	}
2386	if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
2387		if (skb_cloned(skb)) {
2388			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2389			if (nskb == NULL)
 
 
 
 
2390				return -ENOMEM;
2391			tcp_unlink_write_queue(skb, sk);
2392			skb_header_release(nskb);
2393			__tcp_add_write_queue_head(sk, nskb);
2394			sk_wmem_free_skb(sk, skb);
2395			sk->sk_wmem_queued += nskb->truesize;
 
2396			sk_mem_charge(sk, nskb->truesize);
2397			skb = nskb;
2398		}
2399
2400		TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
2401		TCP_ECN_send_synack(tcp_sk(sk), skb);
2402	}
2403	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2404	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2405}
2406
2407/* Prepare a SYN-ACK. */
2408struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
 
 
 
 
 
 
 
 
 
2409				struct request_sock *req,
2410				struct request_values *rvp)
 
 
2411{
 
 
2412	struct tcp_out_options opts;
2413	struct tcp_extend_values *xvp = tcp_xv(rvp);
2414	struct inet_request_sock *ireq = inet_rsk(req);
2415	struct tcp_sock *tp = tcp_sk(sk);
2416	const struct tcp_cookie_values *cvp = tp->cookie_values;
2417	struct tcphdr *th;
2418	struct sk_buff *skb;
2419	struct tcp_md5sig_key *md5;
2420	int tcp_header_size;
 
2421	int mss;
2422	int s_data_desired = 0;
2423
2424	if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2425		s_data_desired = cvp->s_data_desired;
2426	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2427	if (skb == NULL)
2428		return NULL;
2429
2430	/* Reserve space for headers. */
2431	skb_reserve(skb, MAX_TCP_HEADER);
2432
2433	skb_dst_set(skb, dst_clone(dst));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2434
2435	mss = dst_metric_advmss(dst);
2436	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2437		mss = tp->rx_opt.user_mss;
2438
2439	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2440		__u8 rcv_wscale;
2441		/* Set this up on the first call only */
2442		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2443
2444		/* limit the window selection if the user enforce a smaller rx buffer */
2445		if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2446		    (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2447			req->window_clamp = tcp_full_space(sk);
2448
2449		/* tcp_full_space because it is guaranteed to be the first packet */
2450		tcp_select_initial_window(tcp_full_space(sk),
2451			mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2452			&req->rcv_wnd,
2453			&req->window_clamp,
2454			ireq->wscale_ok,
2455			&rcv_wscale,
2456			dst_metric(dst, RTAX_INITRWND));
2457		ireq->rcv_wscale = rcv_wscale;
2458	}
2459
2460	memset(&opts, 0, sizeof(opts));
 
2461#ifdef CONFIG_SYN_COOKIES
2462	if (unlikely(req->cookie_ts))
2463		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
 
2464	else
2465#endif
2466	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2467	tcp_header_size = tcp_synack_options(sk, req, mss,
2468					     skb, &opts, &md5, xvp)
2469			+ sizeof(*th);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2470
2471	skb_push(skb, tcp_header_size);
2472	skb_reset_transport_header(skb);
2473
2474	th = tcp_hdr(skb);
2475	memset(th, 0, sizeof(struct tcphdr));
2476	th->syn = 1;
2477	th->ack = 1;
2478	TCP_ECN_make_synack(req, th);
2479	th->source = ireq->loc_port;
2480	th->dest = ireq->rmt_port;
2481	/* Setting of flags are superfluous here for callers (and ECE is
2482	 * not even correctly set)
2483	 */
2484	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2485			     TCPHDR_SYN | TCPHDR_ACK);
2486
2487	if (OPTION_COOKIE_EXTENSION & opts.options) {
2488		if (s_data_desired) {
2489			u8 *buf = skb_put(skb, s_data_desired);
2490
2491			/* copy data directly from the listening socket. */
2492			memcpy(buf, cvp->s_data_payload, s_data_desired);
2493			TCP_SKB_CB(skb)->end_seq += s_data_desired;
2494		}
2495
2496		if (opts.hash_size > 0) {
2497			__u32 workspace[SHA_WORKSPACE_WORDS];
2498			u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2499			u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2500
2501			/* Secret recipe depends on the Timestamp, (future)
2502			 * Sequence and Acknowledgment Numbers, Initiator
2503			 * Cookie, and others handled by IP variant caller.
2504			 */
2505			*tail-- ^= opts.tsval;
2506			*tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2507			*tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2508
2509			/* recommended */
2510			*tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2511			*tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2512
2513			sha_transform((__u32 *)&xvp->cookie_bakery[0],
2514				      (char *)mess,
2515				      &workspace[0]);
2516			opts.hash_location =
2517				(__u8 *)&xvp->cookie_bakery[0];
2518		}
2519	}
2520
2521	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2522	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2523
2524	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2525	th->window = htons(min(req->rcv_wnd, 65535U));
2526	tcp_options_write((__be32 *)(th + 1), tp, &opts);
2527	th->doff = (tcp_header_size >> 2);
2528	TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2529
 
 
2530#ifdef CONFIG_TCP_MD5SIG
2531	/* Okay, we have all we need - do the md5 hash if needed */
2532	if (md5) {
2533		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2534					       md5, NULL, req, skb);
 
 
 
 
 
 
 
2535	}
 
 
2536#endif
2537
 
 
 
 
 
 
2538	return skb;
2539}
2540EXPORT_SYMBOL(tcp_make_synack);
2541
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2542/* Do all connect socket setups that can be done AF independent. */
2543static void tcp_connect_init(struct sock *sk)
2544{
2545	struct dst_entry *dst = __sk_dst_get(sk);
2546	struct tcp_sock *tp = tcp_sk(sk);
2547	__u8 rcv_wscale;
 
2548
2549	/* We'll fix this up when we get a response from the other end.
2550	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2551	 */
2552	tp->tcp_header_len = sizeof(struct tcphdr) +
2553		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
 
2554
2555#ifdef CONFIG_TCP_MD5SIG
2556	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2557		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2558#endif
2559
2560	/* If user gave his TCP_MAXSEG, record it to clamp */
2561	if (tp->rx_opt.user_mss)
2562		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2563	tp->max_window = 0;
2564	tcp_mtup_init(sk);
2565	tcp_sync_mss(sk, dst_mtu(dst));
2566
 
 
2567	if (!tp->window_clamp)
2568		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2569	tp->advmss = dst_metric_advmss(dst);
2570	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2571		tp->advmss = tp->rx_opt.user_mss;
2572
2573	tcp_initialize_rcv_mss(sk);
2574
2575	/* limit the window selection if the user enforce a smaller rx buffer */
2576	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2577	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2578		tp->window_clamp = tcp_full_space(sk);
2579
2580	tcp_select_initial_window(tcp_full_space(sk),
 
 
 
 
2581				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2582				  &tp->rcv_wnd,
2583				  &tp->window_clamp,
2584				  sysctl_tcp_window_scaling,
2585				  &rcv_wscale,
2586				  dst_metric(dst, RTAX_INITRWND));
2587
2588	tp->rx_opt.rcv_wscale = rcv_wscale;
2589	tp->rcv_ssthresh = tp->rcv_wnd;
2590
2591	sk->sk_err = 0;
2592	sock_reset_flag(sk, SOCK_DONE);
2593	tp->snd_wnd = 0;
2594	tcp_init_wl(tp, 0);
 
2595	tp->snd_una = tp->write_seq;
2596	tp->snd_sml = tp->write_seq;
2597	tp->snd_up = tp->write_seq;
2598	tp->rcv_nxt = 0;
2599	tp->rcv_wup = 0;
2600	tp->copied_seq = 0;
 
 
 
 
 
2601
2602	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2603	inet_csk(sk)->icsk_retransmits = 0;
2604	tcp_clear_retrans(tp);
2605}
2606
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2607/* Build a SYN and send it off. */
2608int tcp_connect(struct sock *sk)
2609{
2610	struct tcp_sock *tp = tcp_sk(sk);
2611	struct sk_buff *buff;
2612	int err;
2613
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2614	tcp_connect_init(sk);
2615
2616	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2617	if (unlikely(buff == NULL))
 
 
 
 
 
2618		return -ENOBUFS;
2619
2620	/* Reserve space for headers. */
2621	skb_reserve(buff, MAX_TCP_HEADER);
2622
2623	tp->snd_nxt = tp->write_seq;
2624	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2625	TCP_ECN_send_syn(sk, buff);
2626
2627	/* Send it off. */
2628	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2629	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2630	skb_header_release(buff);
2631	__tcp_add_write_queue_tail(sk, buff);
2632	sk->sk_wmem_queued += buff->truesize;
2633	sk_mem_charge(sk, buff->truesize);
2634	tp->packets_out += tcp_skb_pcount(buff);
2635	err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2636	if (err == -ECONNREFUSED)
2637		return err;
2638
2639	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2640	 * in order to make this packet get counted in tcpOutSegs.
2641	 */
2642	tp->snd_nxt = tp->write_seq;
2643	tp->pushed_seq = tp->write_seq;
 
 
 
 
 
2644	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2645
2646	/* Timer for repeating the SYN until an answer. */
2647	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2648				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2649	return 0;
2650}
2651EXPORT_SYMBOL(tcp_connect);
2652
 
 
 
 
 
 
 
2653/* Send out a delayed ack, the caller does the policy checking
2654 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2655 * for details.
2656 */
2657void tcp_send_delayed_ack(struct sock *sk)
2658{
2659	struct inet_connection_sock *icsk = inet_csk(sk);
2660	int ato = icsk->icsk_ack.ato;
2661	unsigned long timeout;
2662
2663	if (ato > TCP_DELACK_MIN) {
2664		const struct tcp_sock *tp = tcp_sk(sk);
2665		int max_ato = HZ / 2;
2666
2667		if (icsk->icsk_ack.pingpong ||
2668		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2669			max_ato = TCP_DELACK_MAX;
2670
2671		/* Slow path, intersegment interval is "high". */
2672
2673		/* If some rtt estimate is known, use it to bound delayed ack.
2674		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2675		 * directly.
2676		 */
2677		if (tp->srtt) {
2678			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
 
2679
2680			if (rtt < max_ato)
2681				max_ato = rtt;
2682		}
2683
2684		ato = min(ato, max_ato);
2685	}
2686
 
 
2687	/* Stay within the limit we were given */
2688	timeout = jiffies + ato;
2689
2690	/* Use new timeout only if there wasn't a older one earlier. */
2691	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2692		/* If delack timer was blocked or is about to expire,
2693		 * send ACK now.
2694		 */
2695		if (icsk->icsk_ack.blocked ||
2696		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2697			tcp_send_ack(sk);
2698			return;
2699		}
2700
2701		if (!time_before(timeout, icsk->icsk_ack.timeout))
2702			timeout = icsk->icsk_ack.timeout;
2703	}
2704	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
 
2705	icsk->icsk_ack.timeout = timeout;
2706	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2707}
2708
2709/* This routine sends an ack and also updates the window. */
2710void tcp_send_ack(struct sock *sk)
2711{
2712	struct sk_buff *buff;
2713
2714	/* If we have been reset, we may not send again. */
2715	if (sk->sk_state == TCP_CLOSE)
2716		return;
2717
2718	/* We are not putting this on the write queue, so
2719	 * tcp_transmit_skb() will set the ownership to this
2720	 * sock.
2721	 */
2722	buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2723	if (buff == NULL) {
 
 
 
 
 
 
 
2724		inet_csk_schedule_ack(sk);
2725		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2726		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2727					  TCP_DELACK_MAX, TCP_RTO_MAX);
2728		return;
2729	}
2730
2731	/* Reserve space for headers and prepare control bits. */
2732	skb_reserve(buff, MAX_TCP_HEADER);
2733	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2734
 
 
 
 
 
 
2735	/* Send it off, this clears delayed acks for us. */
2736	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2737	tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
 
 
 
 
 
2738}
2739
2740/* This routine sends a packet with an out of date sequence
2741 * number. It assumes the other end will try to ack it.
2742 *
2743 * Question: what should we make while urgent mode?
2744 * 4.4BSD forces sending single byte of data. We cannot send
2745 * out of window data, because we have SND.NXT==SND.MAX...
2746 *
2747 * Current solution: to send TWO zero-length segments in urgent mode:
2748 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2749 * out-of-date with SND.UNA-1 to probe window.
2750 */
2751static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2752{
2753	struct tcp_sock *tp = tcp_sk(sk);
2754	struct sk_buff *skb;
2755
2756	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2757	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2758	if (skb == NULL)
 
2759		return -1;
2760
2761	/* Reserve space for headers and set control bits. */
2762	skb_reserve(skb, MAX_TCP_HEADER);
2763	/* Use a previous sequence.  This should cause the other
2764	 * end to send an ack.  Don't queue or clone SKB, just
2765	 * send it.
2766	 */
2767	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2768	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2769	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
 
 
 
 
 
 
 
 
 
 
2770}
2771
2772/* Initiate keepalive or window probe from timer. */
2773int tcp_write_wakeup(struct sock *sk)
2774{
2775	struct tcp_sock *tp = tcp_sk(sk);
2776	struct sk_buff *skb;
2777
2778	if (sk->sk_state == TCP_CLOSE)
2779		return -1;
2780
2781	if ((skb = tcp_send_head(sk)) != NULL &&
2782	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2783		int err;
2784		unsigned int mss = tcp_current_mss(sk);
2785		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2786
2787		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2788			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2789
2790		/* We are probing the opening of a window
2791		 * but the window size is != 0
2792		 * must have been a result SWS avoidance ( sender )
2793		 */
2794		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2795		    skb->len > mss) {
2796			seg_size = min(seg_size, mss);
2797			TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2798			if (tcp_fragment(sk, skb, seg_size, mss))
 
2799				return -1;
2800		} else if (!tcp_skb_pcount(skb))
2801			tcp_set_skb_tso_segs(sk, skb, mss);
2802
2803		TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2804		TCP_SKB_CB(skb)->when = tcp_time_stamp;
2805		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2806		if (!err)
2807			tcp_event_new_data_sent(sk, skb);
2808		return err;
2809	} else {
2810		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2811			tcp_xmit_probe_skb(sk, 1);
2812		return tcp_xmit_probe_skb(sk, 0);
2813	}
2814}
2815
2816/* A window probe timeout has occurred.  If window is not closed send
2817 * a partial packet else a zero probe.
2818 */
2819void tcp_send_probe0(struct sock *sk)
2820{
2821	struct inet_connection_sock *icsk = inet_csk(sk);
2822	struct tcp_sock *tp = tcp_sk(sk);
 
 
2823	int err;
2824
2825	err = tcp_write_wakeup(sk);
2826
2827	if (tp->packets_out || !tcp_send_head(sk)) {
2828		/* Cancel probe timer, if it is not required. */
2829		icsk->icsk_probes_out = 0;
2830		icsk->icsk_backoff = 0;
 
2831		return;
2832	}
2833
 
2834	if (err <= 0) {
2835		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2836			icsk->icsk_backoff++;
2837		icsk->icsk_probes_out++;
2838		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2839					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2840					  TCP_RTO_MAX);
2841	} else {
2842		/* If packet was not sent due to local congestion,
2843		 * do not backoff and do not remember icsk_probes_out.
2844		 * Let local senders to fight for local resources.
2845		 *
2846		 * Use accumulated backoff yet.
2847		 */
2848		if (!icsk->icsk_probes_out)
2849			icsk->icsk_probes_out = 1;
2850		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2851					  min(icsk->icsk_rto << icsk->icsk_backoff,
2852					      TCP_RESOURCE_PROBE_INTERVAL),
2853					  TCP_RTO_MAX);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2854	}
 
2855}