Linux Audio

Check our new training course

Loading...
v5.4
   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
  42#include <linux/compiler.h>
  43#include <linux/gfp.h>
  44#include <linux/module.h>
  45#include <linux/static_key.h>
  46
  47#include <trace/events/tcp.h>
 
  48
  49/* Refresh clocks of a TCP socket,
  50 * ensuring monotically increasing values.
  51 */
  52void tcp_mstamp_refresh(struct tcp_sock *tp)
  53{
  54	u64 val = tcp_clock_ns();
  55
  56	tp->tcp_clock_cache = val;
  57	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  58}
 
 
 
 
 
 
 
 
 
 
 
  59
  60static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  61			   int push_one, gfp_t gfp);
  62
  63/* Account for new data that has been sent to the network. */
  64static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  65{
  66	struct inet_connection_sock *icsk = inet_csk(sk);
  67	struct tcp_sock *tp = tcp_sk(sk);
  68	unsigned int prior_packets = tp->packets_out;
  69
  70	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
 
  71
  72	__skb_unlink(skb, &sk->sk_write_queue);
  73	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
 
  74
  75	tp->packets_out += tcp_skb_pcount(skb);
  76	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  77		tcp_rearm_rto(sk);
  78
  79	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  80		      tcp_skb_pcount(skb));
  81}
  82
  83/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  84 * window scaling factor due to loss of precision.
  85 * If window has been shrunk, what should we make? It is not clear at all.
  86 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  87 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  88 * invalid. OK, let's make this for now:
  89 */
  90static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  91{
  92	const struct tcp_sock *tp = tcp_sk(sk);
  93
  94	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
  95	    (tp->rx_opt.wscale_ok &&
  96	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
  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	const 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.
 136 */
 137void tcp_cwnd_restart(struct sock *sk, s32 delta)
 138{
 139	struct tcp_sock *tp = tcp_sk(sk);
 140	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 
 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_jiffies32;
 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 sock *sk)
 158{
 159	struct inet_connection_sock *icsk = inet_csk(sk);
 160	const u32 now = tcp_jiffies32;
 161
 162	if (tcp_packets_in_flight(tp) == 0)
 163		tcp_ca_event(sk, CA_EVENT_TX_START);
 164
 165	/* If this is the first data packet sent in response to the
 166	 * previous received data,
 167	 * and it is a reply for ato after last received packet,
 168	 * increase pingpong count.
 169	 */
 170	if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
 171	    (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 172		inet_csk_inc_pingpong_cnt(sk);
 173
 174	tp->lsndtime = now;
 
 
 
 
 
 
 175}
 176
 177/* Account for an ACK we sent. */
 178static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
 179				      u32 rcv_nxt)
 180{
 181	struct tcp_sock *tp = tcp_sk(sk);
 182
 183	if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
 184		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 185			      tp->compressed_ack - TCP_FASTRETRANS_THRESH);
 186		tp->compressed_ack = TCP_FASTRETRANS_THRESH;
 187		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 188			__sock_put(sk);
 189	}
 190
 191	if (unlikely(rcv_nxt != tp->rcv_nxt))
 192		return;  /* Special ACK sent by DCTCP to reflect ECN */
 193	tcp_dec_quickack_mode(sk, pkts);
 194	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 195}
 196
 197/* Determine a window scaling and initial window to offer.
 198 * Based on the assumption that the given amount of space
 199 * will be offered. Store the results in the tp structure.
 200 * NOTE: for smooth operation initial space offering should
 201 * be a multiple of mss if possible. We assume here that mss >= 1.
 202 * This MUST be enforced by all callers.
 203 */
 204void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 205			       __u32 *rcv_wnd, __u32 *window_clamp,
 206			       int wscale_ok, __u8 *rcv_wscale,
 207			       __u32 init_rcv_wnd)
 208{
 209	unsigned int space = (__space < 0 ? 0 : __space);
 210
 211	/* If no clamp set the clamp to the max possible scaled window */
 212	if (*window_clamp == 0)
 213		(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
 214	space = min(*window_clamp, space);
 215
 216	/* Quantize space offering to a multiple of mss if possible. */
 217	if (space > mss)
 218		space = rounddown(space, mss);
 219
 220	/* NOTE: offering an initial window larger than 32767
 221	 * will break some buggy TCP stacks. If the admin tells us
 222	 * it is likely we could be speaking with such a buggy stack
 223	 * we will truncate our initial window offering to 32K-1
 224	 * unless the remote has sent us a window scaling option,
 225	 * which we interpret as a sign the remote TCP is not
 226	 * misinterpreting the window field as a signed quantity.
 227	 */
 228	if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 229		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 230	else
 231		(*rcv_wnd) = min_t(u32, space, U16_MAX);
 232
 233	if (init_rcv_wnd)
 234		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 235
 236	*rcv_wscale = 0;
 237	if (wscale_ok) {
 238		/* Set window scaling on max possible window */
 239		space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
 240		space = max_t(u32, space, sysctl_rmem_max);
 
 241		space = min_t(u32, space, *window_clamp);
 242		*rcv_wscale = clamp_t(int, ilog2(space) - 15,
 243				      0, TCP_MAX_WSCALE);
 
 
 244	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 245	/* Set the clamp no higher than max representable value */
 246	(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
 247}
 248EXPORT_SYMBOL(tcp_select_initial_window);
 249
 250/* Chose a new window to advertise, update state in tcp_sock for the
 251 * socket, and return result with RFC1323 scaling applied.  The return
 252 * value can be stuffed directly into th->window for an outgoing
 253 * frame.
 254 */
 255static u16 tcp_select_window(struct sock *sk)
 256{
 257	struct tcp_sock *tp = tcp_sk(sk);
 258	u32 old_win = tp->rcv_wnd;
 259	u32 cur_win = tcp_receive_window(tp);
 260	u32 new_win = __tcp_select_window(sk);
 261
 262	/* Never shrink the offered window */
 263	if (new_win < cur_win) {
 264		/* Danger Will Robinson!
 265		 * Don't update rcv_wup/rcv_wnd here or else
 266		 * we will not be able to advertise a zero
 267		 * window in time.  --DaveM
 268		 *
 269		 * Relax Will Robinson.
 270		 */
 271		if (new_win == 0)
 272			NET_INC_STATS(sock_net(sk),
 273				      LINUX_MIB_TCPWANTZEROWINDOWADV);
 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 &&
 283	    sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
 284		new_win = min(new_win, MAX_TCP_WINDOW);
 285	else
 286		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 287
 288	/* RFC1323 scaling applied */
 289	new_win >>= tp->rx_opt.rcv_wscale;
 290
 291	/* If we advertise zero window, disable fast path. */
 292	if (new_win == 0) {
 293		tp->pred_flags = 0;
 294		if (old_win)
 295			NET_INC_STATS(sock_net(sk),
 296				      LINUX_MIB_TCPTOZEROWINDOWADV);
 297	} else if (old_win == 0) {
 298		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
 299	}
 300
 301	return new_win;
 302}
 303
 304/* Packet ECN state for a SYN-ACK */
 305static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 306{
 307	const struct tcp_sock *tp = tcp_sk(sk);
 308
 309	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 310	if (!(tp->ecn_flags & TCP_ECN_OK))
 311		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 312	else if (tcp_ca_needs_ecn(sk) ||
 313		 tcp_bpf_ca_needs_ecn(sk))
 314		INET_ECN_xmit(sk);
 315}
 316
 317/* Packet ECN state for a SYN.  */
 318static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 319{
 320	struct tcp_sock *tp = tcp_sk(sk);
 321	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 322	bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
 323		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 324
 325	if (!use_ecn) {
 326		const struct dst_entry *dst = __sk_dst_get(sk);
 327
 328		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 329			use_ecn = true;
 330	}
 331
 332	tp->ecn_flags = 0;
 333
 334	if (use_ecn) {
 335		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 336		tp->ecn_flags = TCP_ECN_OK;
 337		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 338			INET_ECN_xmit(sk);
 339	}
 340}
 341
 342static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 343{
 344	if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
 345		/* tp->ecn_flags are cleared at a later point in time when
 346		 * SYN ACK is ultimatively being received.
 347		 */
 348		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 349}
 350
 351static void
 352tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 353{
 354	if (inet_rsk(req)->ecn_ok)
 355		th->ece = 1;
 356}
 357
 358/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 359 * be sent.
 360 */
 361static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 362			 struct tcphdr *th, int tcp_header_len)
 363{
 364	struct tcp_sock *tp = tcp_sk(sk);
 365
 366	if (tp->ecn_flags & TCP_ECN_OK) {
 367		/* Not-retransmitted data segment: set ECT and inject CWR. */
 368		if (skb->len != tcp_header_len &&
 369		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 370			INET_ECN_xmit(sk);
 371			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 372				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 373				th->cwr = 1;
 374				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 375			}
 376		} else if (!tcp_ca_needs_ecn(sk)) {
 377			/* ACK or retransmitted segment: clear ECT|CE */
 378			INET_ECN_dontxmit(sk);
 379		}
 380		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 381			th->ece = 1;
 382	}
 383}
 384
 385/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 386 * auto increment end seqno.
 387 */
 388static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 389{
 390	skb->ip_summed = CHECKSUM_PARTIAL;
 
 391
 392	TCP_SKB_CB(skb)->tcp_flags = flags;
 393	TCP_SKB_CB(skb)->sacked = 0;
 394
 395	tcp_skb_pcount_set(skb, 1);
 
 
 396
 397	TCP_SKB_CB(skb)->seq = seq;
 398	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 399		seq++;
 400	TCP_SKB_CB(skb)->end_seq = seq;
 401}
 402
 403static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 404{
 405	return tp->snd_una != tp->snd_up;
 406}
 407
 408#define OPTION_SACK_ADVERTISE	(1 << 0)
 409#define OPTION_TS		(1 << 1)
 410#define OPTION_MD5		(1 << 2)
 411#define OPTION_WSCALE		(1 << 3)
 412#define OPTION_FAST_OPEN_COOKIE	(1 << 8)
 413#define OPTION_SMC		(1 << 9)
 414
 415static void smc_options_write(__be32 *ptr, u16 *options)
 416{
 417#if IS_ENABLED(CONFIG_SMC)
 418	if (static_branch_unlikely(&tcp_have_smc)) {
 419		if (unlikely(OPTION_SMC & *options)) {
 420			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 421				       (TCPOPT_NOP  << 16) |
 422				       (TCPOPT_EXP <<  8) |
 423				       (TCPOLEN_EXP_SMC_BASE));
 424			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 425		}
 426	}
 427#endif
 428}
 429
 430struct tcp_out_options {
 431	u16 options;		/* bit field of OPTION_* */
 432	u16 mss;		/* 0 to disable */
 433	u8 ws;			/* window scale, 0 to disable */
 434	u8 num_sack_blocks;	/* number of SACK blocks to include */
 435	u8 hash_size;		/* bytes in hash_location */
 436	__u8 *hash_location;	/* temporary pointer, overloaded */
 437	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 438	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 439};
 440
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 441/* Write previously computed TCP options to the packet.
 442 *
 443 * Beware: Something in the Internet is very sensitive to the ordering of
 444 * TCP options, we learned this through the hard way, so be careful here.
 445 * Luckily we can at least blame others for their non-compliance but from
 446 * inter-operability perspective it seems that we're somewhat stuck with
 447 * the ordering which we have been using if we want to keep working with
 448 * those broken things (not that it currently hurts anybody as there isn't
 449 * particular reason why the ordering would need to be changed).
 450 *
 451 * At least SACK_PERM as the first option is known to lead to a disaster
 452 * (but it may well be that other scenarios fail similarly).
 453 */
 454static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 455			      struct tcp_out_options *opts)
 456{
 457	u16 options = opts->options;	/* mungable copy */
 458
 
 
 
 
 
 
 
 
 459	if (unlikely(OPTION_MD5 & options)) {
 460		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 461			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 
 
 
 
 
 
 
 
 
 
 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	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 491		*ptr++ = htonl((TCPOPT_NOP << 24) |
 492			       (TCPOPT_NOP << 16) |
 493			       (TCPOPT_SACK_PERM << 8) |
 494			       TCPOLEN_SACK_PERM);
 495	}
 496
 497	if (unlikely(OPTION_WSCALE & options)) {
 498		*ptr++ = htonl((TCPOPT_NOP << 24) |
 499			       (TCPOPT_WINDOW << 16) |
 500			       (TCPOLEN_WINDOW << 8) |
 501			       opts->ws);
 502	}
 503
 504	if (unlikely(opts->num_sack_blocks)) {
 505		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 506			tp->duplicate_sack : tp->selective_acks;
 507		int this_sack;
 508
 509		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 510			       (TCPOPT_NOP  << 16) |
 511			       (TCPOPT_SACK <<  8) |
 512			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 513						     TCPOLEN_SACK_PERBLOCK)));
 514
 515		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 516		     ++this_sack) {
 517			*ptr++ = htonl(sp[this_sack].start_seq);
 518			*ptr++ = htonl(sp[this_sack].end_seq);
 519		}
 520
 521		tp->rx_opt.dsack = 0;
 522	}
 523
 524	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 525		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 526		u8 *p = (u8 *)ptr;
 527		u32 len; /* Fast Open option length */
 528
 529		if (foc->exp) {
 530			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 531			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 532				     TCPOPT_FASTOPEN_MAGIC);
 533			p += TCPOLEN_EXP_FASTOPEN_BASE;
 534		} else {
 535			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 536			*p++ = TCPOPT_FASTOPEN;
 537			*p++ = len;
 538		}
 539
 540		memcpy(p, foc->val, foc->len);
 541		if ((len & 3) == 2) {
 542			p[foc->len] = TCPOPT_NOP;
 543			p[foc->len + 1] = TCPOPT_NOP;
 544		}
 545		ptr += (len + 3) >> 2;
 546	}
 547
 548	smc_options_write(ptr, &options);
 549}
 550
 551static void smc_set_option(const struct tcp_sock *tp,
 552			   struct tcp_out_options *opts,
 553			   unsigned int *remaining)
 554{
 555#if IS_ENABLED(CONFIG_SMC)
 556	if (static_branch_unlikely(&tcp_have_smc)) {
 557		if (tp->syn_smc) {
 558			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 559				opts->options |= OPTION_SMC;
 560				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 561			}
 562		}
 563	}
 564#endif
 565}
 566
 567static void smc_set_option_cond(const struct tcp_sock *tp,
 568				const struct inet_request_sock *ireq,
 569				struct tcp_out_options *opts,
 570				unsigned int *remaining)
 571{
 572#if IS_ENABLED(CONFIG_SMC)
 573	if (static_branch_unlikely(&tcp_have_smc)) {
 574		if (tp->syn_smc && ireq->smc_ok) {
 575			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 576				opts->options |= OPTION_SMC;
 577				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 578			}
 579		}
 580	}
 581#endif
 582}
 583
 584/* Compute TCP options for SYN packets. This is not the final
 585 * network wire format yet.
 586 */
 587static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 588				struct tcp_out_options *opts,
 589				struct tcp_md5sig_key **md5)
 590{
 591	struct tcp_sock *tp = tcp_sk(sk);
 592	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 593	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 
 
 
 594
 595	*md5 = NULL;
 596#ifdef CONFIG_TCP_MD5SIG
 597	if (static_branch_unlikely(&tcp_md5_needed) &&
 598	    rcu_access_pointer(tp->md5sig_info)) {
 599		*md5 = tp->af_specific->md5_lookup(sk, sk);
 600		if (*md5) {
 601			opts->options |= OPTION_MD5;
 602			remaining -= TCPOLEN_MD5SIG_ALIGNED;
 603		}
 604	}
 
 
 605#endif
 606
 607	/* We always get an MSS option.  The option bytes which will be seen in
 608	 * normal data packets should timestamps be used, must be in the MSS
 609	 * advertised.  But we subtract them from tp->mss_cache so that
 610	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 611	 * fact here if necessary.  If we don't do this correctly, as a
 612	 * receiver we won't recognize data packets as being full sized when we
 613	 * should, and thus we won't abide by the delayed ACK rules correctly.
 614	 * SACKs don't matter, we never delay an ACK when we have any of those
 615	 * going out.  */
 616	opts->mss = tcp_advertise_mss(sk);
 617	remaining -= TCPOLEN_MSS_ALIGNED;
 618
 619	if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
 620		opts->options |= OPTION_TS;
 621		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
 622		opts->tsecr = tp->rx_opt.ts_recent;
 623		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 624	}
 625	if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
 626		opts->ws = tp->rx_opt.rcv_wscale;
 627		opts->options |= OPTION_WSCALE;
 628		remaining -= TCPOLEN_WSCALE_ALIGNED;
 629	}
 630	if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
 631		opts->options |= OPTION_SACK_ADVERTISE;
 632		if (unlikely(!(OPTION_TS & opts->options)))
 633			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 634	}
 635
 636	if (fastopen && fastopen->cookie.len >= 0) {
 637		u32 need = fastopen->cookie.len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 638
 639		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 640					       TCPOLEN_FASTOPEN_BASE;
 641		need = (need + 3) & ~3U;  /* Align to 32 bits */
 642		if (remaining >= need) {
 643			opts->options |= OPTION_FAST_OPEN_COOKIE;
 644			opts->fastopen_cookie = &fastopen->cookie;
 645			remaining -= need;
 646			tp->syn_fastopen = 1;
 647			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 648		}
 649	}
 650
 651	smc_set_option(tp, opts, &remaining);
 652
 653	return MAX_TCP_OPTION_SPACE - remaining;
 654}
 655
 656/* Set up TCP options for SYN-ACKs. */
 657static unsigned int tcp_synack_options(const struct sock *sk,
 658				       struct request_sock *req,
 659				       unsigned int mss, struct sk_buff *skb,
 660				       struct tcp_out_options *opts,
 661				       const struct tcp_md5sig_key *md5,
 662				       struct tcp_fastopen_cookie *foc)
 663{
 664	struct inet_request_sock *ireq = inet_rsk(req);
 665	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 
 
 
 666
 667#ifdef CONFIG_TCP_MD5SIG
 668	if (md5) {
 
 669		opts->options |= OPTION_MD5;
 670		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 671
 672		/* We can't fit any SACK blocks in a packet with MD5 + TS
 673		 * options. There was discussion about disabling SACK
 674		 * rather than TS in order to fit in better with old,
 675		 * buggy kernels, but that was deemed to be unnecessary.
 676		 */
 677		ireq->tstamp_ok &= !ireq->sack_ok;
 678	}
 
 
 679#endif
 680
 681	/* We always send an MSS option. */
 682	opts->mss = mss;
 683	remaining -= TCPOLEN_MSS_ALIGNED;
 684
 685	if (likely(ireq->wscale_ok)) {
 686		opts->ws = ireq->rcv_wscale;
 687		opts->options |= OPTION_WSCALE;
 688		remaining -= TCPOLEN_WSCALE_ALIGNED;
 689	}
 690	if (likely(ireq->tstamp_ok)) {
 691		opts->options |= OPTION_TS;
 692		opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
 693		opts->tsecr = req->ts_recent;
 694		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 695	}
 696	if (likely(ireq->sack_ok)) {
 697		opts->options |= OPTION_SACK_ADVERTISE;
 698		if (unlikely(!ireq->tstamp_ok))
 699			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 700	}
 701	if (foc != NULL && foc->len >= 0) {
 702		u32 need = foc->len;
 703
 704		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 705				   TCPOLEN_FASTOPEN_BASE;
 706		need = (need + 3) & ~3U;  /* Align to 32 bits */
 707		if (remaining >= need) {
 708			opts->options |= OPTION_FAST_OPEN_COOKIE;
 709			opts->fastopen_cookie = foc;
 
 
 
 
 
 
 
 
 
 710			remaining -= need;
 
 
 
 
 711		}
 712	}
 713
 714	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 715
 716	return MAX_TCP_OPTION_SPACE - remaining;
 717}
 718
 719/* Compute TCP options for ESTABLISHED sockets. This is not the
 720 * final wire format yet.
 721 */
 722static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 723					struct tcp_out_options *opts,
 724					struct tcp_md5sig_key **md5)
 725{
 726	struct tcp_sock *tp = tcp_sk(sk);
 727	unsigned int size = 0;
 728	unsigned int eff_sacks;
 729
 730	opts->options = 0;
 731
 732	*md5 = NULL;
 733#ifdef CONFIG_TCP_MD5SIG
 734	if (static_branch_unlikely(&tcp_md5_needed) &&
 735	    rcu_access_pointer(tp->md5sig_info)) {
 736		*md5 = tp->af_specific->md5_lookup(sk, sk);
 737		if (*md5) {
 738			opts->options |= OPTION_MD5;
 739			size += TCPOLEN_MD5SIG_ALIGNED;
 740		}
 741	}
 
 
 742#endif
 743
 744	if (likely(tp->rx_opt.tstamp_ok)) {
 745		opts->options |= OPTION_TS;
 746		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
 747		opts->tsecr = tp->rx_opt.ts_recent;
 748		size += TCPOLEN_TSTAMP_ALIGNED;
 749	}
 750
 751	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 752	if (unlikely(eff_sacks)) {
 753		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 754		opts->num_sack_blocks =
 755			min_t(unsigned int, eff_sacks,
 756			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 757			      TCPOLEN_SACK_PERBLOCK);
 758		size += TCPOLEN_SACK_BASE_ALIGNED +
 759			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 760	}
 761
 762	return size;
 763}
 764
 765
 766/* TCP SMALL QUEUES (TSQ)
 767 *
 768 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 769 * to reduce RTT and bufferbloat.
 770 * We do this using a special skb destructor (tcp_wfree).
 771 *
 772 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 773 * needs to be reallocated in a driver.
 774 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
 775 *
 776 * Since transmit from skb destructor is forbidden, we use a tasklet
 777 * to process all sockets that eventually need to send more skbs.
 778 * We use one tasklet per cpu, with its own queue of sockets.
 779 */
 780struct tsq_tasklet {
 781	struct tasklet_struct	tasklet;
 782	struct list_head	head; /* queue of tcp sockets */
 783};
 784static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
 785
 786static void tcp_tsq_write(struct sock *sk)
 787{
 788	if ((1 << sk->sk_state) &
 789	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
 790	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
 791		struct tcp_sock *tp = tcp_sk(sk);
 792
 793		if (tp->lost_out > tp->retrans_out &&
 794		    tp->snd_cwnd > tcp_packets_in_flight(tp)) {
 795			tcp_mstamp_refresh(tp);
 796			tcp_xmit_retransmit_queue(sk);
 797		}
 798
 799		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
 800			       0, GFP_ATOMIC);
 801	}
 802}
 803
 804static void tcp_tsq_handler(struct sock *sk)
 805{
 806	bh_lock_sock(sk);
 807	if (!sock_owned_by_user(sk))
 808		tcp_tsq_write(sk);
 809	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
 810		sock_hold(sk);
 811	bh_unlock_sock(sk);
 812}
 813/*
 814 * One tasklet per cpu tries to send more skbs.
 815 * We run in tasklet context but need to disable irqs when
 816 * transferring tsq->head because tcp_wfree() might
 817 * interrupt us (non NAPI drivers)
 818 */
 819static void tcp_tasklet_func(unsigned long data)
 820{
 821	struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
 822	LIST_HEAD(list);
 823	unsigned long flags;
 824	struct list_head *q, *n;
 825	struct tcp_sock *tp;
 826	struct sock *sk;
 827
 828	local_irq_save(flags);
 829	list_splice_init(&tsq->head, &list);
 830	local_irq_restore(flags);
 831
 832	list_for_each_safe(q, n, &list) {
 833		tp = list_entry(q, struct tcp_sock, tsq_node);
 834		list_del(&tp->tsq_node);
 835
 836		sk = (struct sock *)tp;
 837		smp_mb__before_atomic();
 838		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
 839
 840		tcp_tsq_handler(sk);
 841		sk_free(sk);
 842	}
 843}
 844
 845#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
 846			  TCPF_WRITE_TIMER_DEFERRED |	\
 847			  TCPF_DELACK_TIMER_DEFERRED |	\
 848			  TCPF_MTU_REDUCED_DEFERRED)
 849/**
 850 * tcp_release_cb - tcp release_sock() callback
 851 * @sk: socket
 852 *
 853 * called from release_sock() to perform protocol dependent
 854 * actions before socket release.
 855 */
 856void tcp_release_cb(struct sock *sk)
 857{
 858	unsigned long flags, nflags;
 859
 860	/* perform an atomic operation only if at least one flag is set */
 861	do {
 862		flags = sk->sk_tsq_flags;
 863		if (!(flags & TCP_DEFERRED_ALL))
 864			return;
 865		nflags = flags & ~TCP_DEFERRED_ALL;
 866	} while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
 867
 868	if (flags & TCPF_TSQ_DEFERRED) {
 869		tcp_tsq_write(sk);
 870		__sock_put(sk);
 871	}
 872	/* Here begins the tricky part :
 873	 * We are called from release_sock() with :
 874	 * 1) BH disabled
 875	 * 2) sk_lock.slock spinlock held
 876	 * 3) socket owned by us (sk->sk_lock.owned == 1)
 877	 *
 878	 * But following code is meant to be called from BH handlers,
 879	 * so we should keep BH disabled, but early release socket ownership
 880	 */
 881	sock_release_ownership(sk);
 882
 883	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
 884		tcp_write_timer_handler(sk);
 885		__sock_put(sk);
 886	}
 887	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
 888		tcp_delack_timer_handler(sk);
 889		__sock_put(sk);
 890	}
 891	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
 892		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
 893		__sock_put(sk);
 894	}
 895}
 896EXPORT_SYMBOL(tcp_release_cb);
 897
 898void __init tcp_tasklet_init(void)
 899{
 900	int i;
 901
 902	for_each_possible_cpu(i) {
 903		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
 904
 905		INIT_LIST_HEAD(&tsq->head);
 906		tasklet_init(&tsq->tasklet,
 907			     tcp_tasklet_func,
 908			     (unsigned long)tsq);
 909	}
 910}
 911
 912/*
 913 * Write buffer destructor automatically called from kfree_skb.
 914 * We can't xmit new skbs from this context, as we might already
 915 * hold qdisc lock.
 916 */
 917void tcp_wfree(struct sk_buff *skb)
 918{
 919	struct sock *sk = skb->sk;
 920	struct tcp_sock *tp = tcp_sk(sk);
 921	unsigned long flags, nval, oval;
 922
 923	/* Keep one reference on sk_wmem_alloc.
 924	 * Will be released by sk_free() from here or tcp_tasklet_func()
 925	 */
 926	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
 927
 928	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
 929	 * Wait until our queues (qdisc + devices) are drained.
 930	 * This gives :
 931	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
 932	 * - chance for incoming ACK (processed by another cpu maybe)
 933	 *   to migrate this flow (skb->ooo_okay will be eventually set)
 934	 */
 935	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
 936		goto out;
 937
 938	for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
 939		struct tsq_tasklet *tsq;
 940		bool empty;
 941
 942		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
 943			goto out;
 944
 945		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
 946		nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
 947		if (nval != oval)
 948			continue;
 949
 950		/* queue this socket to tasklet queue */
 951		local_irq_save(flags);
 952		tsq = this_cpu_ptr(&tsq_tasklet);
 953		empty = list_empty(&tsq->head);
 954		list_add(&tp->tsq_node, &tsq->head);
 955		if (empty)
 956			tasklet_schedule(&tsq->tasklet);
 957		local_irq_restore(flags);
 958		return;
 959	}
 960out:
 961	sk_free(sk);
 962}
 963
 964/* Note: Called under soft irq.
 965 * We can call TCP stack right away, unless socket is owned by user.
 966 */
 967enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
 968{
 969	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
 970	struct sock *sk = (struct sock *)tp;
 971
 972	tcp_tsq_handler(sk);
 973	sock_put(sk);
 974
 975	return HRTIMER_NORESTART;
 976}
 977
 978static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
 979				      u64 prior_wstamp)
 980{
 981	struct tcp_sock *tp = tcp_sk(sk);
 982
 983	if (sk->sk_pacing_status != SK_PACING_NONE) {
 984		unsigned long rate = sk->sk_pacing_rate;
 985
 986		/* Original sch_fq does not pace first 10 MSS
 987		 * Note that tp->data_segs_out overflows after 2^32 packets,
 988		 * this is a minor annoyance.
 989		 */
 990		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
 991			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
 992			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
 993
 994			/* take into account OS jitter */
 995			len_ns -= min_t(u64, len_ns / 2, credit);
 996			tp->tcp_wstamp_ns += len_ns;
 997		}
 998	}
 999	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1000}
1001
1002/* This routine actually transmits TCP packets queued in by
1003 * tcp_do_sendmsg().  This is used by both the initial
1004 * transmission and possible later retransmissions.
1005 * All SKB's seen here are completely headerless.  It is our
1006 * job to build the TCP header, and pass the packet down to
1007 * IP so it can do the same plus pass the packet off to the
1008 * device.
1009 *
1010 * We are working here with either a clone of the original
1011 * SKB, or a fresh unique copy made by the retransmit engine.
1012 */
1013static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1014			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1015{
1016	const struct inet_connection_sock *icsk = inet_csk(sk);
1017	struct inet_sock *inet;
1018	struct tcp_sock *tp;
1019	struct tcp_skb_cb *tcb;
1020	struct tcp_out_options opts;
1021	unsigned int tcp_options_size, tcp_header_size;
1022	struct sk_buff *oskb = NULL;
1023	struct tcp_md5sig_key *md5;
1024	struct tcphdr *th;
1025	u64 prior_wstamp;
1026	int err;
1027
1028	BUG_ON(!skb || !tcp_skb_pcount(skb));
1029	tp = tcp_sk(sk);
1030	prior_wstamp = tp->tcp_wstamp_ns;
1031	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1032	skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1033	if (clone_it) {
1034		TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1035			- tp->snd_una;
1036		oskb = skb;
1037
1038		tcp_skb_tsorted_save(oskb) {
1039			if (unlikely(skb_cloned(oskb)))
1040				skb = pskb_copy(oskb, gfp_mask);
1041			else
1042				skb = skb_clone(oskb, gfp_mask);
1043		} tcp_skb_tsorted_restore(oskb);
1044
 
 
 
 
 
 
 
 
 
 
 
1045		if (unlikely(!skb))
1046			return -ENOBUFS;
1047	}
1048
1049	inet = inet_sk(sk);
 
1050	tcb = TCP_SKB_CB(skb);
1051	memset(&opts, 0, sizeof(opts));
1052
1053	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1054		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1055	} else {
1056		tcp_options_size = tcp_established_options(sk, skb, &opts,
1057							   &md5);
1058		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1059		 * at receiver : This slightly improve GRO performance.
1060		 * Note that we do not force the PSH flag for non GSO packets,
1061		 * because they might be sent under high congestion events,
1062		 * and in this case it is better to delay the delivery of 1-MSS
1063		 * packets and thus the corresponding ACK packet that would
1064		 * release the following packet.
1065		 */
1066		if (tcp_skb_pcount(skb) > 1)
1067			tcb->tcp_flags |= TCPHDR_PSH;
1068	}
1069	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1070
1071	/* if no packet is in qdisc/device queue, then allow XPS to select
1072	 * another queue. We can be called from tcp_tsq_handler()
1073	 * which holds one reference to sk.
1074	 *
1075	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1076	 * One way to get this would be to set skb->truesize = 2 on them.
1077	 */
1078	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1079
1080	/* If we had to use memory reserve to allocate this skb,
1081	 * this might cause drops if packet is looped back :
1082	 * Other socket might not have SOCK_MEMALLOC.
1083	 * Packets not looped back do not care about pfmemalloc.
1084	 */
1085	skb->pfmemalloc = 0;
1086
1087	skb_push(skb, tcp_header_size);
1088	skb_reset_transport_header(skb);
1089
1090	skb_orphan(skb);
1091	skb->sk = sk;
1092	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1093	skb_set_hash_from_sk(skb, sk);
1094	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1095
1096	skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1097
1098	/* Build TCP header and checksum it. */
1099	th = (struct tcphdr *)skb->data;
1100	th->source		= inet->inet_sport;
1101	th->dest		= inet->inet_dport;
1102	th->seq			= htonl(tcb->seq);
1103	th->ack_seq		= htonl(rcv_nxt);
1104	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1105					tcb->tcp_flags);
1106
 
 
 
 
 
 
 
 
1107	th->check		= 0;
1108	th->urg_ptr		= 0;
1109
1110	/* The urg_mode check is necessary during a below snd_una win probe */
1111	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1112		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1113			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1114			th->urg = 1;
1115		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1116			th->urg_ptr = htons(0xFFFF);
1117			th->urg = 1;
1118		}
1119	}
1120
1121	tcp_options_write((__be32 *)(th + 1), tp, &opts);
1122	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1123	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1124		th->window      = htons(tcp_select_window(sk));
1125		tcp_ecn_send(sk, skb, th, tcp_header_size);
1126	} else {
1127		/* RFC1323: The window in SYN & SYN/ACK segments
1128		 * is never scaled.
1129		 */
1130		th->window	= htons(min(tp->rcv_wnd, 65535U));
1131	}
1132#ifdef CONFIG_TCP_MD5SIG
1133	/* Calculate the MD5 hash, as we have all we need now */
1134	if (md5) {
1135		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1136		tp->af_specific->calc_md5_hash(opts.hash_location,
1137					       md5, sk, skb);
1138	}
1139#endif
1140
1141	icsk->icsk_af_ops->send_check(sk, skb);
1142
1143	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1144		tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1145
1146	if (skb->len != tcp_header_size) {
1147		tcp_event_data_sent(tp, sk);
1148		tp->data_segs_out += tcp_skb_pcount(skb);
1149		tp->bytes_sent += skb->len - tcp_header_size;
1150	}
1151
1152	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1153		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1154			      tcp_skb_pcount(skb));
1155
1156	tp->segs_out += tcp_skb_pcount(skb);
1157	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1158	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1159	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1160
1161	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1162
1163	/* Cleanup our debris for IP stacks */
1164	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1165			       sizeof(struct inet6_skb_parm)));
1166
1167	tcp_add_tx_delay(skb, tp);
1168
1169	err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1170
1171	if (unlikely(err > 0)) {
1172		tcp_enter_cwr(sk);
1173		err = net_xmit_eval(err);
1174	}
1175	if (!err && oskb) {
1176		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1177		tcp_rate_skb_sent(sk, oskb);
1178	}
1179	return err;
1180}
1181
1182static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1183			    gfp_t gfp_mask)
1184{
1185	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1186				  tcp_sk(sk)->rcv_nxt);
1187}
1188
1189/* This routine just queues the buffer for sending.
1190 *
1191 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1192 * otherwise socket can stall.
1193 */
1194static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1195{
1196	struct tcp_sock *tp = tcp_sk(sk);
1197
1198	/* Advance write_seq and place onto the write_queue. */
1199	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1200	__skb_header_release(skb);
1201	tcp_add_write_queue_tail(sk, skb);
1202	sk_wmem_queued_add(sk, skb->truesize);
1203	sk_mem_charge(sk, skb->truesize);
1204}
1205
1206/* Initialize TSO segments for a packet. */
1207static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
 
1208{
1209	if (skb->len <= mss_now) {
 
1210		/* Avoid the costly divide in the normal
1211		 * non-TSO case.
1212		 */
1213		tcp_skb_pcount_set(skb, 1);
1214		TCP_SKB_CB(skb)->tcp_gso_size = 0;
 
1215	} else {
1216		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1217		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
 
1218	}
1219}
1220
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1221/* Pcount in the middle of the write queue got changed, we need to do various
1222 * tweaks to fix counters
1223 */
1224static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1225{
1226	struct tcp_sock *tp = tcp_sk(sk);
1227
1228	tp->packets_out -= decr;
1229
1230	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1231		tp->sacked_out -= decr;
1232	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1233		tp->retrans_out -= decr;
1234	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1235		tp->lost_out -= decr;
1236
1237	/* Reno case is special. Sigh... */
1238	if (tcp_is_reno(tp) && decr > 0)
1239		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1240
 
 
1241	if (tp->lost_skb_hint &&
1242	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1243	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1244		tp->lost_cnt_hint -= decr;
1245
1246	tcp_verify_left_out(tp);
1247}
1248
1249static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1250{
1251	return TCP_SKB_CB(skb)->txstamp_ack ||
1252		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1253}
1254
1255static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1256{
1257	struct skb_shared_info *shinfo = skb_shinfo(skb);
1258
1259	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1260	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1261		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1262		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1263
1264		shinfo->tx_flags &= ~tsflags;
1265		shinfo2->tx_flags |= tsflags;
1266		swap(shinfo->tskey, shinfo2->tskey);
1267		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1268		TCP_SKB_CB(skb)->txstamp_ack = 0;
1269	}
1270}
1271
1272static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1273{
1274	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1275	TCP_SKB_CB(skb)->eor = 0;
1276}
1277
1278/* Insert buff after skb on the write or rtx queue of sk.  */
1279static void tcp_insert_write_queue_after(struct sk_buff *skb,
1280					 struct sk_buff *buff,
1281					 struct sock *sk,
1282					 enum tcp_queue tcp_queue)
1283{
1284	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1285		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1286	else
1287		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1288}
1289
1290/* Function to create two new TCP segments.  Shrinks the given segment
1291 * to the specified size and appends a new segment with the rest of the
1292 * packet to the list.  This won't be called frequently, I hope.
1293 * Remember, these are still headerless SKBs at this point.
1294 */
1295int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1296		 struct sk_buff *skb, u32 len,
1297		 unsigned int mss_now, gfp_t gfp)
1298{
1299	struct tcp_sock *tp = tcp_sk(sk);
1300	struct sk_buff *buff;
1301	int nsize, old_factor;
1302	long limit;
1303	int nlen;
1304	u8 flags;
1305
1306	if (WARN_ON(len > skb->len))
1307		return -EINVAL;
1308
1309	nsize = skb_headlen(skb) - len;
1310	if (nsize < 0)
1311		nsize = 0;
1312
1313	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1314	 * We need some allowance to not penalize applications setting small
1315	 * SO_SNDBUF values.
1316	 * Also allow first and last skb in retransmit queue to be split.
1317	 */
1318	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1319	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1320		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1321		     skb != tcp_rtx_queue_head(sk) &&
1322		     skb != tcp_rtx_queue_tail(sk))) {
1323		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1324		return -ENOMEM;
1325	}
1326
1327	if (skb_unclone(skb, gfp))
1328		return -ENOMEM;
1329
1330	/* Get a new skb... force flag on. */
1331	buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1332	if (!buff)
1333		return -ENOMEM; /* We'll just try again later. */
1334	skb_copy_decrypted(buff, skb);
1335
1336	sk_wmem_queued_add(sk, buff->truesize);
1337	sk_mem_charge(sk, buff->truesize);
1338	nlen = skb->len - len - nsize;
1339	buff->truesize += nlen;
1340	skb->truesize -= nlen;
1341
1342	/* Correct the sequence numbers. */
1343	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1344	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1345	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1346
1347	/* PSH and FIN should only be set in the second packet. */
1348	flags = TCP_SKB_CB(skb)->tcp_flags;
1349	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1350	TCP_SKB_CB(buff)->tcp_flags = flags;
1351	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1352	tcp_skb_fragment_eor(skb, buff);
1353
1354	skb_split(skb, buff, len);
 
 
 
 
 
 
 
 
 
 
 
 
1355
1356	buff->ip_summed = CHECKSUM_PARTIAL;
1357
 
 
 
 
1358	buff->tstamp = skb->tstamp;
1359	tcp_fragment_tstamp(skb, buff);
1360
1361	old_factor = tcp_skb_pcount(skb);
1362
1363	/* Fix up tso_factor for both original and new SKB.  */
1364	tcp_set_skb_tso_segs(skb, mss_now);
1365	tcp_set_skb_tso_segs(buff, mss_now);
1366
1367	/* Update delivered info for the new segment */
1368	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1369
1370	/* If this packet has been sent out already, we must
1371	 * adjust the various packet counters.
1372	 */
1373	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1374		int diff = old_factor - tcp_skb_pcount(skb) -
1375			tcp_skb_pcount(buff);
1376
1377		if (diff)
1378			tcp_adjust_pcount(sk, skb, diff);
1379	}
1380
1381	/* Link BUFF into the send queue. */
1382	__skb_header_release(buff);
1383	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1384	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1385		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1386
1387	return 0;
1388}
1389
1390/* This is similar to __pskb_pull_tail(). The difference is that pulled
1391 * data is not copied, but immediately discarded.
 
1392 */
1393static int __pskb_trim_head(struct sk_buff *skb, int len)
1394{
1395	struct skb_shared_info *shinfo;
1396	int i, k, eat;
1397
1398	eat = min_t(int, len, skb_headlen(skb));
1399	if (eat) {
1400		__skb_pull(skb, eat);
1401		len -= eat;
1402		if (!len)
1403			return 0;
1404	}
1405	eat = len;
1406	k = 0;
1407	shinfo = skb_shinfo(skb);
1408	for (i = 0; i < shinfo->nr_frags; i++) {
1409		int size = skb_frag_size(&shinfo->frags[i]);
1410
1411		if (size <= eat) {
1412			skb_frag_unref(skb, i);
1413			eat -= size;
1414		} else {
1415			shinfo->frags[k] = shinfo->frags[i];
1416			if (eat) {
1417				skb_frag_off_add(&shinfo->frags[k], eat);
1418				skb_frag_size_sub(&shinfo->frags[k], eat);
1419				eat = 0;
1420			}
1421			k++;
1422		}
1423	}
1424	shinfo->nr_frags = k;
1425
 
1426	skb->data_len -= len;
1427	skb->len = skb->data_len;
1428	return len;
1429}
1430
1431/* Remove acked data from a packet in the transmit queue. */
1432int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1433{
1434	u32 delta_truesize;
1435
1436	if (skb_unclone(skb, GFP_ATOMIC))
1437		return -ENOMEM;
1438
1439	delta_truesize = __pskb_trim_head(skb, len);
 
 
 
 
1440
1441	TCP_SKB_CB(skb)->seq += len;
1442	skb->ip_summed = CHECKSUM_PARTIAL;
1443
1444	if (delta_truesize) {
1445		skb->truesize	   -= delta_truesize;
1446		sk_wmem_queued_add(sk, -delta_truesize);
1447		sk_mem_uncharge(sk, delta_truesize);
1448		sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1449	}
1450
1451	/* Any change of skb->len requires recalculation of tso factor. */
 
 
1452	if (tcp_skb_pcount(skb) > 1)
1453		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1454
1455	return 0;
1456}
1457
1458/* Calculate MSS not accounting any TCP options.  */
1459static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1460{
1461	const struct tcp_sock *tp = tcp_sk(sk);
1462	const struct inet_connection_sock *icsk = inet_csk(sk);
1463	int mss_now;
1464
1465	/* Calculate base mss without TCP options:
1466	   It is MMS_S - sizeof(tcphdr) of rfc1122
1467	 */
1468	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1469
1470	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1471	if (icsk->icsk_af_ops->net_frag_header_len) {
1472		const struct dst_entry *dst = __sk_dst_get(sk);
1473
1474		if (dst && dst_allfrag(dst))
1475			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1476	}
1477
1478	/* Clamp it (mss_clamp does not include tcp options) */
1479	if (mss_now > tp->rx_opt.mss_clamp)
1480		mss_now = tp->rx_opt.mss_clamp;
1481
1482	/* Now subtract optional transport overhead */
1483	mss_now -= icsk->icsk_ext_hdr_len;
1484
1485	/* Then reserve room for full set of TCP options and 8 bytes of data */
1486	mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1487	return mss_now;
1488}
1489
1490/* Calculate MSS. Not accounting for SACKs here.  */
1491int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1492{
1493	/* Subtract TCP options size, not including SACKs */
1494	return __tcp_mtu_to_mss(sk, pmtu) -
1495	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1496}
1497
1498/* Inverse of above */
1499int tcp_mss_to_mtu(struct sock *sk, int mss)
1500{
1501	const struct tcp_sock *tp = tcp_sk(sk);
1502	const struct inet_connection_sock *icsk = inet_csk(sk);
1503	int mtu;
1504
1505	mtu = mss +
1506	      tp->tcp_header_len +
1507	      icsk->icsk_ext_hdr_len +
1508	      icsk->icsk_af_ops->net_header_len;
1509
1510	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1511	if (icsk->icsk_af_ops->net_frag_header_len) {
1512		const struct dst_entry *dst = __sk_dst_get(sk);
1513
1514		if (dst && dst_allfrag(dst))
1515			mtu += icsk->icsk_af_ops->net_frag_header_len;
1516	}
1517	return mtu;
1518}
1519EXPORT_SYMBOL(tcp_mss_to_mtu);
1520
1521/* MTU probing init per socket */
1522void tcp_mtup_init(struct sock *sk)
1523{
1524	struct tcp_sock *tp = tcp_sk(sk);
1525	struct inet_connection_sock *icsk = inet_csk(sk);
1526	struct net *net = sock_net(sk);
1527
1528	icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1529	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1530			       icsk->icsk_af_ops->net_header_len;
1531	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1532	icsk->icsk_mtup.probe_size = 0;
1533	if (icsk->icsk_mtup.enabled)
1534		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1535}
1536EXPORT_SYMBOL(tcp_mtup_init);
1537
1538/* This function synchronize snd mss to current pmtu/exthdr set.
1539
1540   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1541   for TCP options, but includes only bare TCP header.
1542
1543   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1544   It is minimum of user_mss and mss received with SYN.
1545   It also does not include TCP options.
1546
1547   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1548
1549   tp->mss_cache is current effective sending mss, including
1550   all tcp options except for SACKs. It is evaluated,
1551   taking into account current pmtu, but never exceeds
1552   tp->rx_opt.mss_clamp.
1553
1554   NOTE1. rfc1122 clearly states that advertised MSS
1555   DOES NOT include either tcp or ip options.
1556
1557   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1558   are READ ONLY outside this function.		--ANK (980731)
1559 */
1560unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1561{
1562	struct tcp_sock *tp = tcp_sk(sk);
1563	struct inet_connection_sock *icsk = inet_csk(sk);
1564	int mss_now;
1565
1566	if (icsk->icsk_mtup.search_high > pmtu)
1567		icsk->icsk_mtup.search_high = pmtu;
1568
1569	mss_now = tcp_mtu_to_mss(sk, pmtu);
1570	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1571
1572	/* And store cached results */
1573	icsk->icsk_pmtu_cookie = pmtu;
1574	if (icsk->icsk_mtup.enabled)
1575		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1576	tp->mss_cache = mss_now;
1577
1578	return mss_now;
1579}
1580EXPORT_SYMBOL(tcp_sync_mss);
1581
1582/* Compute the current effective MSS, taking SACKs and IP options,
1583 * and even PMTU discovery events into account.
1584 */
1585unsigned int tcp_current_mss(struct sock *sk)
1586{
1587	const struct tcp_sock *tp = tcp_sk(sk);
1588	const struct dst_entry *dst = __sk_dst_get(sk);
1589	u32 mss_now;
1590	unsigned int header_len;
1591	struct tcp_out_options opts;
1592	struct tcp_md5sig_key *md5;
1593
1594	mss_now = tp->mss_cache;
1595
1596	if (dst) {
1597		u32 mtu = dst_mtu(dst);
1598		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1599			mss_now = tcp_sync_mss(sk, mtu);
1600	}
1601
1602	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1603		     sizeof(struct tcphdr);
1604	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1605	 * some common options. If this is an odd packet (because we have SACK
1606	 * blocks etc) then our calculated header_len will be different, and
1607	 * we have to adjust mss_now correspondingly */
1608	if (header_len != tp->tcp_header_len) {
1609		int delta = (int) header_len - tp->tcp_header_len;
1610		mss_now -= delta;
1611	}
1612
1613	return mss_now;
1614}
1615
1616/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1617 * As additional protections, we do not touch cwnd in retransmission phases,
1618 * and if application hit its sndbuf limit recently.
1619 */
1620static void tcp_cwnd_application_limited(struct sock *sk)
1621{
1622	struct tcp_sock *tp = tcp_sk(sk);
1623
1624	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1625	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1626		/* Limited by application or receiver window. */
1627		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1628		u32 win_used = max(tp->snd_cwnd_used, init_win);
1629		if (win_used < tp->snd_cwnd) {
1630			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1631			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1632		}
1633		tp->snd_cwnd_used = 0;
1634	}
1635	tp->snd_cwnd_stamp = tcp_jiffies32;
1636}
1637
1638static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1639{
1640	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1641	struct tcp_sock *tp = tcp_sk(sk);
1642
1643	/* Track the maximum number of outstanding packets in each
1644	 * window, and remember whether we were cwnd-limited then.
1645	 */
1646	if (!before(tp->snd_una, tp->max_packets_seq) ||
1647	    tp->packets_out > tp->max_packets_out) {
1648		tp->max_packets_out = tp->packets_out;
1649		tp->max_packets_seq = tp->snd_nxt;
1650		tp->is_cwnd_limited = is_cwnd_limited;
1651	}
1652
1653	if (tcp_is_cwnd_limited(sk)) {
1654		/* Network is feed fully. */
1655		tp->snd_cwnd_used = 0;
1656		tp->snd_cwnd_stamp = tcp_jiffies32;
1657	} else {
1658		/* Network starves. */
1659		if (tp->packets_out > tp->snd_cwnd_used)
1660			tp->snd_cwnd_used = tp->packets_out;
1661
1662		if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1663		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1664		    !ca_ops->cong_control)
1665			tcp_cwnd_application_limited(sk);
1666
1667		/* The following conditions together indicate the starvation
1668		 * is caused by insufficient sender buffer:
1669		 * 1) just sent some data (see tcp_write_xmit)
1670		 * 2) not cwnd limited (this else condition)
1671		 * 3) no more data to send (tcp_write_queue_empty())
1672		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1673		 */
1674		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1675		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1676		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1677			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1678	}
1679}
1680
1681/* Minshall's variant of the Nagle send check. */
1682static bool tcp_minshall_check(const struct tcp_sock *tp)
1683{
1684	return after(tp->snd_sml, tp->snd_una) &&
1685		!after(tp->snd_sml, tp->snd_nxt);
1686}
1687
1688/* Update snd_sml if this skb is under mss
1689 * Note that a TSO packet might end with a sub-mss segment
1690 * The test is really :
1691 * if ((skb->len % mss) != 0)
1692 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1693 * But we can avoid doing the divide again given we already have
1694 *  skb_pcount = skb->len / mss_now
1695 */
1696static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1697				const struct sk_buff *skb)
1698{
1699	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1700		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1701}
1702
1703/* Return false, if packet can be sent now without violation Nagle's rules:
1704 * 1. It is full sized. (provided by caller in %partial bool)
1705 * 2. Or it contains FIN. (already checked by caller)
1706 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1707 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1708 *    With Minshall's modification: all sent small packets are ACKed.
1709 */
1710static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1711			    int nonagle)
1712{
1713	return partial &&
1714		((nonagle & TCP_NAGLE_CORK) ||
1715		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1716}
1717
1718/* Return how many segs we'd like on a TSO packet,
1719 * to send one TSO packet per ms
1720 */
1721static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1722			    int min_tso_segs)
1723{
1724	u32 bytes, segs;
1725
1726	bytes = min_t(unsigned long,
1727		      sk->sk_pacing_rate >> sk->sk_pacing_shift,
1728		      sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1729
1730	/* Goal is to send at least one packet per ms,
1731	 * not one big TSO packet every 100 ms.
1732	 * This preserves ACK clocking and is consistent
1733	 * with tcp_tso_should_defer() heuristic.
1734	 */
1735	segs = max_t(u32, bytes / mss_now, min_tso_segs);
1736
1737	return segs;
1738}
1739
1740/* Return the number of segments we want in the skb we are transmitting.
1741 * See if congestion control module wants to decide; otherwise, autosize.
1742 */
1743static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
 
1744{
1745	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1746	u32 min_tso, tso_segs;
1747
1748	min_tso = ca_ops->min_tso_segs ?
1749			ca_ops->min_tso_segs(sk) :
1750			sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1751
1752	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1753	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1754}
1755
1756/* Returns the portion of skb which can be sent right away */
1757static unsigned int tcp_mss_split_point(const struct sock *sk,
1758					const struct sk_buff *skb,
1759					unsigned int mss_now,
1760					unsigned int max_segs,
1761					int nonagle)
1762{
1763	const struct tcp_sock *tp = tcp_sk(sk);
1764	u32 partial, needed, window, max_len;
1765
1766	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1767	max_len = mss_now * max_segs;
1768
1769	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1770		return max_len;
1771
1772	needed = min(skb->len, window);
1773
1774	if (max_len <= needed)
1775		return max_len;
1776
1777	partial = needed % mss_now;
1778	/* If last segment is not a full MSS, check if Nagle rules allow us
1779	 * to include this last segment in this skb.
1780	 * Otherwise, we'll split the skb at last MSS boundary
1781	 */
1782	if (tcp_nagle_check(partial != 0, tp, nonagle))
1783		return needed - partial;
1784
1785	return needed;
1786}
1787
1788/* Can at least one segment of SKB be sent right now, according to the
1789 * congestion window rules?  If so, return how many segments are allowed.
1790 */
1791static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1792					 const struct sk_buff *skb)
1793{
1794	u32 in_flight, cwnd, halfcwnd;
1795
1796	/* Don't be strict about the congestion window for the final FIN.  */
1797	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1798	    tcp_skb_pcount(skb) == 1)
1799		return 1;
1800
1801	in_flight = tcp_packets_in_flight(tp);
1802	cwnd = tp->snd_cwnd;
1803	if (in_flight >= cwnd)
1804		return 0;
1805
1806	/* For better scheduling, ensure we have at least
1807	 * 2 GSO packets in flight.
1808	 */
1809	halfcwnd = max(cwnd >> 1, 1U);
1810	return min(halfcwnd, cwnd - in_flight);
1811}
1812
1813/* Initialize TSO state of a skb.
1814 * This must be invoked the first time we consider transmitting
1815 * SKB onto the wire.
1816 */
1817static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
 
1818{
1819	int tso_segs = tcp_skb_pcount(skb);
1820
1821	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1822		tcp_set_skb_tso_segs(skb, mss_now);
1823		tso_segs = tcp_skb_pcount(skb);
1824	}
1825	return tso_segs;
1826}
1827
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1828
1829/* Return true if the Nagle test allows this packet to be
1830 * sent now.
1831 */
1832static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1833				  unsigned int cur_mss, int nonagle)
1834{
1835	/* Nagle rule does not apply to frames, which sit in the middle of the
1836	 * write_queue (they have no chances to get new data).
1837	 *
1838	 * This is implemented in the callers, where they modify the 'nonagle'
1839	 * argument based upon the location of SKB in the send queue.
1840	 */
1841	if (nonagle & TCP_NAGLE_PUSH)
1842		return true;
1843
1844	/* Don't use the nagle rule for urgent data (or for the final FIN). */
1845	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1846		return true;
 
 
 
1847
1848	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1849		return true;
1850
1851	return false;
1852}
1853
1854/* Does at least the first segment of SKB fit into the send window? */
1855static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1856			     const struct sk_buff *skb,
1857			     unsigned int cur_mss)
1858{
1859	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1860
1861	if (skb->len > cur_mss)
1862		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1863
1864	return !after(end_seq, tcp_wnd_end(tp));
1865}
1866
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1867/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1868 * which is put after SKB on the list.  It is very much like
1869 * tcp_fragment() except that it may make several kinds of assumptions
1870 * in order to speed up the splitting operation.  In particular, we
1871 * know that all the data is in scatter-gather pages, and that the
1872 * packet has never been sent out before (and thus is not cloned).
1873 */
1874static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1875			unsigned int mss_now, gfp_t gfp)
1876{
1877	int nlen = skb->len - len;
1878	struct sk_buff *buff;
 
1879	u8 flags;
1880
1881	/* All of a TSO frame must be composed of paged data.  */
1882	if (skb->len != skb->data_len)
1883		return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1884				    skb, len, mss_now, gfp);
1885
1886	buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1887	if (unlikely(!buff))
1888		return -ENOMEM;
1889	skb_copy_decrypted(buff, skb);
1890
1891	sk_wmem_queued_add(sk, buff->truesize);
1892	sk_mem_charge(sk, buff->truesize);
1893	buff->truesize += nlen;
1894	skb->truesize -= nlen;
1895
1896	/* Correct the sequence numbers. */
1897	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1898	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1899	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1900
1901	/* PSH and FIN should only be set in the second packet. */
1902	flags = TCP_SKB_CB(skb)->tcp_flags;
1903	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1904	TCP_SKB_CB(buff)->tcp_flags = flags;
1905
1906	/* This packet was never sent out yet, so no SACK bits. */
1907	TCP_SKB_CB(buff)->sacked = 0;
1908
1909	tcp_skb_fragment_eor(skb, buff);
1910
1911	buff->ip_summed = CHECKSUM_PARTIAL;
1912	skb_split(skb, buff, len);
1913	tcp_fragment_tstamp(skb, buff);
1914
1915	/* Fix up tso_factor for both original and new SKB.  */
1916	tcp_set_skb_tso_segs(skb, mss_now);
1917	tcp_set_skb_tso_segs(buff, mss_now);
1918
1919	/* Link BUFF into the send queue. */
1920	__skb_header_release(buff);
1921	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1922
1923	return 0;
1924}
1925
1926/* Try to defer sending, if possible, in order to minimize the amount
1927 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1928 *
1929 * This algorithm is from John Heffner.
1930 */
1931static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1932				 bool *is_cwnd_limited,
1933				 bool *is_rwnd_limited,
1934				 u32 max_segs)
1935{
 
1936	const struct inet_connection_sock *icsk = inet_csk(sk);
1937	u32 send_win, cong_win, limit, in_flight;
1938	struct tcp_sock *tp = tcp_sk(sk);
1939	struct sk_buff *head;
1940	int win_divisor;
1941	s64 delta;
1942
1943	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1944		goto send_now;
1945
1946	/* Avoid bursty behavior by allowing defer
1947	 * only if the last write was recent (1 ms).
1948	 * Note that tp->tcp_wstamp_ns can be in the future if we have
1949	 * packets waiting in a qdisc or device for EDT delivery.
1950	 */
1951	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1952	if (delta > 0)
1953		goto send_now;
1954
1955	in_flight = tcp_packets_in_flight(tp);
1956
1957	BUG_ON(tcp_skb_pcount(skb) <= 1);
1958	BUG_ON(tp->snd_cwnd <= in_flight);
1959
1960	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1961
1962	/* From in_flight test above, we know that cwnd > in_flight.  */
1963	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1964
1965	limit = min(send_win, cong_win);
1966
1967	/* If a full-sized TSO skb can be sent, do it. */
1968	if (limit >= max_segs * tp->mss_cache)
1969		goto send_now;
1970
1971	/* Middle in queue won't get any more data, full sendable already? */
1972	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1973		goto send_now;
1974
1975	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1976	if (win_divisor) {
1977		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1978
1979		/* If at least some fraction of a window is available,
1980		 * just use it.
1981		 */
1982		chunk /= win_divisor;
1983		if (limit >= chunk)
1984			goto send_now;
1985	} else {
1986		/* Different approach, try not to defer past a single
1987		 * ACK.  Receiver should ACK every other full sized
1988		 * frame, so if we have space for more than 3 frames
1989		 * then send now.
1990		 */
1991		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1992			goto send_now;
1993	}
1994
1995	/* TODO : use tsorted_sent_queue ? */
1996	head = tcp_rtx_queue_head(sk);
1997	if (!head)
1998		goto send_now;
1999	delta = tp->tcp_clock_cache - head->tstamp;
2000	/* If next ACK is likely to come too late (half srtt), do not defer */
2001	if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2002		goto send_now;
2003
2004	/* Ok, it looks like it is advisable to defer.
2005	 * Three cases are tracked :
2006	 * 1) We are cwnd-limited
2007	 * 2) We are rwnd-limited
2008	 * 3) We are application limited.
2009	 */
2010	if (cong_win < send_win) {
2011		if (cong_win <= skb->len) {
2012			*is_cwnd_limited = true;
2013			return true;
2014		}
2015	} else {
2016		if (send_win <= skb->len) {
2017			*is_rwnd_limited = true;
2018			return true;
2019		}
2020	}
2021
2022	/* If this packet won't get more data, do not wait. */
2023	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2024	    TCP_SKB_CB(skb)->eor)
2025		goto send_now;
2026
2027	return true;
2028
2029send_now:
2030	return false;
2031}
2032
2033static inline void tcp_mtu_check_reprobe(struct sock *sk)
2034{
2035	struct inet_connection_sock *icsk = inet_csk(sk);
2036	struct tcp_sock *tp = tcp_sk(sk);
2037	struct net *net = sock_net(sk);
2038	u32 interval;
2039	s32 delta;
2040
2041	interval = net->ipv4.sysctl_tcp_probe_interval;
2042	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2043	if (unlikely(delta >= interval * HZ)) {
2044		int mss = tcp_current_mss(sk);
2045
2046		/* Update current search range */
2047		icsk->icsk_mtup.probe_size = 0;
2048		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2049			sizeof(struct tcphdr) +
2050			icsk->icsk_af_ops->net_header_len;
2051		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2052
2053		/* Update probe time stamp */
2054		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2055	}
2056}
2057
2058static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2059{
2060	struct sk_buff *skb, *next;
2061
2062	skb = tcp_send_head(sk);
2063	tcp_for_write_queue_from_safe(skb, next, sk) {
2064		if (len <= skb->len)
2065			break;
2066
2067		if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2068			return false;
2069
2070		len -= skb->len;
2071	}
2072
2073	return true;
2074}
2075
2076/* Create a new MTU probe if we are ready.
2077 * MTU probe is regularly attempting to increase the path MTU by
2078 * deliberately sending larger packets.  This discovers routing
2079 * changes resulting in larger path MTUs.
2080 *
2081 * Returns 0 if we should wait to probe (no cwnd available),
2082 *         1 if a probe was sent,
2083 *         -1 otherwise
2084 */
2085static int tcp_mtu_probe(struct sock *sk)
2086{
2087	struct inet_connection_sock *icsk = inet_csk(sk);
2088	struct tcp_sock *tp = tcp_sk(sk);
 
2089	struct sk_buff *skb, *nskb, *next;
2090	struct net *net = sock_net(sk);
2091	int probe_size;
2092	int size_needed;
2093	int copy, len;
2094	int mss_now;
2095	int interval;
2096
2097	/* Not currently probing/verifying,
2098	 * not in recovery,
2099	 * have enough cwnd, and
2100	 * not SACKing (the variable headers throw things off)
2101	 */
2102	if (likely(!icsk->icsk_mtup.enabled ||
2103		   icsk->icsk_mtup.probe_size ||
2104		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2105		   tp->snd_cwnd < 11 ||
2106		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2107		return -1;
2108
2109	/* Use binary search for probe_size between tcp_mss_base,
2110	 * and current mss_clamp. if (search_high - search_low)
2111	 * smaller than a threshold, backoff from probing.
2112	 */
2113	mss_now = tcp_current_mss(sk);
2114	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2115				    icsk->icsk_mtup.search_low) >> 1);
2116	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2117	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2118	/* When misfortune happens, we are reprobing actively,
2119	 * and then reprobe timer has expired. We stick with current
2120	 * probing process by not resetting search range to its orignal.
2121	 */
2122	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2123		interval < net->ipv4.sysctl_tcp_probe_threshold) {
2124		/* Check whether enough time has elaplased for
2125		 * another round of probing.
2126		 */
2127		tcp_mtu_check_reprobe(sk);
2128		return -1;
2129	}
2130
2131	/* Have enough data in the send queue to probe? */
2132	if (tp->write_seq - tp->snd_nxt < size_needed)
2133		return -1;
2134
2135	if (tp->snd_wnd < size_needed)
2136		return -1;
2137	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2138		return 0;
2139
2140	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2141	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2142		if (!tcp_packets_in_flight(tp))
2143			return -1;
2144		else
2145			return 0;
2146	}
2147
2148	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2149		return -1;
2150
2151	/* We're allowed to probe.  Build it now. */
2152	nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2153	if (!nskb)
2154		return -1;
2155	sk_wmem_queued_add(sk, nskb->truesize);
2156	sk_mem_charge(sk, nskb->truesize);
2157
2158	skb = tcp_send_head(sk);
2159	skb_copy_decrypted(nskb, skb);
2160
2161	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2162	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2163	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2164	TCP_SKB_CB(nskb)->sacked = 0;
2165	nskb->csum = 0;
2166	nskb->ip_summed = CHECKSUM_PARTIAL;
2167
2168	tcp_insert_write_queue_before(nskb, skb, sk);
2169	tcp_highest_sack_replace(sk, skb, nskb);
2170
2171	len = 0;
2172	tcp_for_write_queue_from_safe(skb, next, sk) {
2173		copy = min_t(int, skb->len, probe_size - len);
2174		skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
 
 
 
 
 
2175
2176		if (skb->len <= copy) {
2177			/* We've eaten all the data from this skb.
2178			 * Throw it away. */
2179			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2180			/* If this is the last SKB we copy and eor is set
2181			 * we need to propagate it to the new skb.
2182			 */
2183			TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2184			tcp_skb_collapse_tstamp(nskb, skb);
2185			tcp_unlink_write_queue(skb, sk);
2186			sk_wmem_free_skb(sk, skb);
2187		} else {
2188			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2189						   ~(TCPHDR_FIN|TCPHDR_PSH);
2190			if (!skb_shinfo(skb)->nr_frags) {
2191				skb_pull(skb, copy);
 
 
 
2192			} else {
2193				__pskb_trim_head(skb, copy);
2194				tcp_set_skb_tso_segs(skb, mss_now);
2195			}
2196			TCP_SKB_CB(skb)->seq += copy;
2197		}
2198
2199		len += copy;
2200
2201		if (len >= probe_size)
2202			break;
2203	}
2204	tcp_init_tso_segs(nskb, nskb->len);
2205
2206	/* We're ready to send.  If this fails, the probe will
2207	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2208	 */
2209	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2210		/* Decrement cwnd here because we are sending
2211		 * effectively two packets. */
2212		tp->snd_cwnd--;
2213		tcp_event_new_data_sent(sk, nskb);
2214
2215		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2216		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2217		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2218
2219		return 1;
2220	}
2221
2222	return -1;
2223}
2224
2225static bool tcp_pacing_check(struct sock *sk)
2226{
2227	struct tcp_sock *tp = tcp_sk(sk);
2228
2229	if (!tcp_needs_internal_pacing(sk))
2230		return false;
2231
2232	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2233		return false;
2234
2235	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2236		hrtimer_start(&tp->pacing_timer,
2237			      ns_to_ktime(tp->tcp_wstamp_ns),
2238			      HRTIMER_MODE_ABS_PINNED_SOFT);
2239		sock_hold(sk);
2240	}
2241	return true;
2242}
2243
2244/* TCP Small Queues :
2245 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2246 * (These limits are doubled for retransmits)
2247 * This allows for :
2248 *  - better RTT estimation and ACK scheduling
2249 *  - faster recovery
2250 *  - high rates
2251 * Alas, some drivers / subsystems require a fair amount
2252 * of queued bytes to ensure line rate.
2253 * One example is wifi aggregation (802.11 AMPDU)
2254 */
2255static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2256				  unsigned int factor)
2257{
2258	unsigned long limit;
2259
2260	limit = max_t(unsigned long,
2261		      2 * skb->truesize,
2262		      sk->sk_pacing_rate >> sk->sk_pacing_shift);
2263	if (sk->sk_pacing_status == SK_PACING_NONE)
2264		limit = min_t(unsigned long, limit,
2265			      sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2266	limit <<= factor;
2267
2268	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2269	    tcp_sk(sk)->tcp_tx_delay) {
2270		u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2271
2272		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2273		 * approximate our needs assuming an ~100% skb->truesize overhead.
2274		 * USEC_PER_SEC is approximated by 2^20.
2275		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2276		 */
2277		extra_bytes >>= (20 - 1);
2278		limit += extra_bytes;
2279	}
2280	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2281		/* Always send skb if rtx queue is empty.
2282		 * No need to wait for TX completion to call us back,
2283		 * after softirq/tasklet schedule.
2284		 * This helps when TX completions are delayed too much.
2285		 */
2286		if (tcp_rtx_queue_empty(sk))
2287			return false;
2288
2289		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2290		/* It is possible TX completion already happened
2291		 * before we set TSQ_THROTTLED, so we must
2292		 * test again the condition.
2293		 */
2294		smp_mb__after_atomic();
2295		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2296			return true;
2297	}
2298	return false;
2299}
2300
2301static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2302{
2303	const u32 now = tcp_jiffies32;
2304	enum tcp_chrono old = tp->chrono_type;
2305
2306	if (old > TCP_CHRONO_UNSPEC)
2307		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2308	tp->chrono_start = now;
2309	tp->chrono_type = new;
2310}
2311
2312void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2313{
2314	struct tcp_sock *tp = tcp_sk(sk);
2315
2316	/* If there are multiple conditions worthy of tracking in a
2317	 * chronograph then the highest priority enum takes precedence
2318	 * over the other conditions. So that if something "more interesting"
2319	 * starts happening, stop the previous chrono and start a new one.
2320	 */
2321	if (type > tp->chrono_type)
2322		tcp_chrono_set(tp, type);
2323}
2324
2325void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2326{
2327	struct tcp_sock *tp = tcp_sk(sk);
2328
2329
2330	/* There are multiple conditions worthy of tracking in a
2331	 * chronograph, so that the highest priority enum takes
2332	 * precedence over the other conditions (see tcp_chrono_start).
2333	 * If a condition stops, we only stop chrono tracking if
2334	 * it's the "most interesting" or current chrono we are
2335	 * tracking and starts busy chrono if we have pending data.
2336	 */
2337	if (tcp_rtx_and_write_queues_empty(sk))
2338		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2339	else if (type == tp->chrono_type)
2340		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2341}
2342
2343/* This routine writes packets to the network.  It advances the
2344 * send_head.  This happens as incoming acks open up the remote
2345 * window for us.
2346 *
2347 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2348 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2349 * account rare use of URG, this is not a big flaw.
2350 *
2351 * Send at most one packet when push_one > 0. Temporarily ignore
2352 * cwnd limit to force at most one packet out when push_one == 2.
2353
2354 * Returns true, if no segments are in flight and we have queued segments,
2355 * but cannot send anything now because of SWS or another problem.
2356 */
2357static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2358			   int push_one, gfp_t gfp)
2359{
2360	struct tcp_sock *tp = tcp_sk(sk);
2361	struct sk_buff *skb;
2362	unsigned int tso_segs, sent_pkts;
2363	int cwnd_quota;
2364	int result;
2365	bool is_cwnd_limited = false, is_rwnd_limited = false;
2366	u32 max_segs;
2367
2368	sent_pkts = 0;
2369
2370	tcp_mstamp_refresh(tp);
2371	if (!push_one) {
2372		/* Do MTU probing. */
2373		result = tcp_mtu_probe(sk);
2374		if (!result) {
2375			return false;
2376		} else if (result > 0) {
2377			sent_pkts = 1;
2378		}
2379	}
2380
2381	max_segs = tcp_tso_segs(sk, mss_now);
2382	while ((skb = tcp_send_head(sk))) {
2383		unsigned int limit;
2384
2385		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2386			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2387			skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2388			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2389			tcp_init_tso_segs(skb, mss_now);
2390			goto repair; /* Skip network transmission */
2391		}
2392
2393		if (tcp_pacing_check(sk))
2394			break;
2395
2396		tso_segs = tcp_init_tso_segs(skb, mss_now);
2397		BUG_ON(!tso_segs);
2398
2399		cwnd_quota = tcp_cwnd_test(tp, skb);
2400		if (!cwnd_quota) {
2401			if (push_one == 2)
2402				/* Force out a loss probe pkt. */
2403				cwnd_quota = 1;
2404			else
2405				break;
2406		}
2407
2408		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2409			is_rwnd_limited = true;
2410			break;
2411		}
2412
2413		if (tso_segs == 1) {
2414			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2415						     (tcp_skb_is_last(sk, skb) ?
2416						      nonagle : TCP_NAGLE_PUSH))))
2417				break;
2418		} else {
2419			if (!push_one &&
2420			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2421						 &is_rwnd_limited, max_segs))
2422				break;
2423		}
2424
2425		limit = mss_now;
2426		if (tso_segs > 1 && !tcp_urg_mode(tp))
2427			limit = tcp_mss_split_point(sk, skb, mss_now,
2428						    min_t(unsigned int,
2429							  cwnd_quota,
2430							  max_segs),
2431						    nonagle);
2432
2433		if (skb->len > limit &&
2434		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2435			break;
2436
2437		if (tcp_small_queue_check(sk, skb, 0))
2438			break;
2439
2440		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2441			break;
2442
2443repair:
2444		/* Advance the send_head.  This one is sent out.
2445		 * This call will increment packets_out.
2446		 */
2447		tcp_event_new_data_sent(sk, skb);
2448
2449		tcp_minshall_update(tp, mss_now, skb);
2450		sent_pkts += tcp_skb_pcount(skb);
2451
2452		if (push_one)
2453			break;
2454	}
2455
2456	if (is_rwnd_limited)
2457		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2458	else
2459		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2460
2461	if (likely(sent_pkts)) {
2462		if (tcp_in_cwnd_reduction(sk))
2463			tp->prr_out += sent_pkts;
2464
2465		/* Send one loss probe per tail loss episode. */
2466		if (push_one != 2)
2467			tcp_schedule_loss_probe(sk, false);
2468		is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2469		tcp_cwnd_validate(sk, is_cwnd_limited);
2470		return false;
2471	}
2472	return !tp->packets_out && !tcp_write_queue_empty(sk);
2473}
2474
2475bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2476{
2477	struct inet_connection_sock *icsk = inet_csk(sk);
2478	struct tcp_sock *tp = tcp_sk(sk);
2479	u32 timeout, rto_delta_us;
2480	int early_retrans;
2481
2482	/* Don't do any loss probe on a Fast Open connection before 3WHS
2483	 * finishes.
2484	 */
2485	if (rcu_access_pointer(tp->fastopen_rsk))
2486		return false;
2487
2488	early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2489	/* Schedule a loss probe in 2*RTT for SACK capable connections
2490	 * not in loss recovery, that are either limited by cwnd or application.
2491	 */
2492	if ((early_retrans != 3 && early_retrans != 4) ||
2493	    !tp->packets_out || !tcp_is_sack(tp) ||
2494	    (icsk->icsk_ca_state != TCP_CA_Open &&
2495	     icsk->icsk_ca_state != TCP_CA_CWR))
2496		return false;
2497
2498	/* Probe timeout is 2*rtt. Add minimum RTO to account
2499	 * for delayed ack when there's one outstanding packet. If no RTT
2500	 * sample is available then probe after TCP_TIMEOUT_INIT.
2501	 */
2502	if (tp->srtt_us) {
2503		timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2504		if (tp->packets_out == 1)
2505			timeout += TCP_RTO_MIN;
2506		else
2507			timeout += TCP_TIMEOUT_MIN;
2508	} else {
2509		timeout = TCP_TIMEOUT_INIT;
2510	}
2511
2512	/* If the RTO formula yields an earlier time, then use that time. */
2513	rto_delta_us = advancing_rto ?
2514			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2515			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2516	if (rto_delta_us > 0)
2517		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2518
2519	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2520			     TCP_RTO_MAX, NULL);
2521	return true;
2522}
2523
2524/* Thanks to skb fast clones, we can detect if a prior transmit of
2525 * a packet is still in a qdisc or driver queue.
2526 * In this case, there is very little point doing a retransmit !
2527 */
2528static bool skb_still_in_host_queue(const struct sock *sk,
2529				    const struct sk_buff *skb)
2530{
2531	if (unlikely(skb_fclone_busy(sk, skb))) {
2532		NET_INC_STATS(sock_net(sk),
2533			      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2534		return true;
2535	}
2536	return false;
2537}
2538
2539/* When probe timeout (PTO) fires, try send a new segment if possible, else
2540 * retransmit the last segment.
2541 */
2542void tcp_send_loss_probe(struct sock *sk)
2543{
2544	struct tcp_sock *tp = tcp_sk(sk);
2545	struct sk_buff *skb;
2546	int pcount;
2547	int mss = tcp_current_mss(sk);
2548
2549	skb = tcp_send_head(sk);
2550	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2551		pcount = tp->packets_out;
2552		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2553		if (tp->packets_out > pcount)
2554			goto probe_sent;
2555		goto rearm_timer;
2556	}
2557	skb = skb_rb_last(&sk->tcp_rtx_queue);
2558	if (unlikely(!skb)) {
2559		WARN_ONCE(tp->packets_out,
2560			  "invalid inflight: %u state %u cwnd %u mss %d\n",
2561			  tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2562		inet_csk(sk)->icsk_pending = 0;
2563		return;
2564	}
2565
2566	/* At most one outstanding TLP retransmission. */
2567	if (tp->tlp_high_seq)
2568		goto rearm_timer;
2569
2570	if (skb_still_in_host_queue(sk, skb))
2571		goto rearm_timer;
2572
2573	pcount = tcp_skb_pcount(skb);
2574	if (WARN_ON(!pcount))
2575		goto rearm_timer;
2576
2577	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2578		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2579					  (pcount - 1) * mss, mss,
2580					  GFP_ATOMIC)))
2581			goto rearm_timer;
2582		skb = skb_rb_next(skb);
2583	}
2584
2585	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2586		goto rearm_timer;
2587
2588	if (__tcp_retransmit_skb(sk, skb, 1))
2589		goto rearm_timer;
2590
2591	/* Record snd_nxt for loss detection. */
2592	tp->tlp_high_seq = tp->snd_nxt;
2593
2594probe_sent:
2595	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2596	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2597	inet_csk(sk)->icsk_pending = 0;
2598rearm_timer:
2599	tcp_rearm_rto(sk);
2600}
2601
2602/* Push out any pending frames which were held back due to
2603 * TCP_CORK or attempt at coalescing tiny packets.
2604 * The socket must be locked by the caller.
2605 */
2606void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2607			       int nonagle)
2608{
2609	/* If we are closed, the bytes will have to remain here.
2610	 * In time closedown will finish, we empty the write queue and
2611	 * all will be happy.
2612	 */
2613	if (unlikely(sk->sk_state == TCP_CLOSE))
2614		return;
2615
2616	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2617			   sk_gfp_mask(sk, GFP_ATOMIC)))
2618		tcp_check_probe_timer(sk);
2619}
2620
2621/* Send _single_ skb sitting at the send head. This function requires
2622 * true push pending frames to setup probe timer etc.
2623 */
2624void tcp_push_one(struct sock *sk, unsigned int mss_now)
2625{
2626	struct sk_buff *skb = tcp_send_head(sk);
2627
2628	BUG_ON(!skb || skb->len < mss_now);
2629
2630	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2631}
2632
2633/* This function returns the amount that we can raise the
2634 * usable window based on the following constraints
2635 *
2636 * 1. The window can never be shrunk once it is offered (RFC 793)
2637 * 2. We limit memory per socket
2638 *
2639 * RFC 1122:
2640 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2641 *  RECV.NEXT + RCV.WIN fixed until:
2642 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2643 *
2644 * i.e. don't raise the right edge of the window until you can raise
2645 * it at least MSS bytes.
2646 *
2647 * Unfortunately, the recommended algorithm breaks header prediction,
2648 * since header prediction assumes th->window stays fixed.
2649 *
2650 * Strictly speaking, keeping th->window fixed violates the receiver
2651 * side SWS prevention criteria. The problem is that under this rule
2652 * a stream of single byte packets will cause the right side of the
2653 * window to always advance by a single byte.
2654 *
2655 * Of course, if the sender implements sender side SWS prevention
2656 * then this will not be a problem.
2657 *
2658 * BSD seems to make the following compromise:
2659 *
2660 *	If the free space is less than the 1/4 of the maximum
2661 *	space available and the free space is less than 1/2 mss,
2662 *	then set the window to 0.
2663 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2664 *	Otherwise, just prevent the window from shrinking
2665 *	and from being larger than the largest representable value.
2666 *
2667 * This prevents incremental opening of the window in the regime
2668 * where TCP is limited by the speed of the reader side taking
2669 * data out of the TCP receive queue. It does nothing about
2670 * those cases where the window is constrained on the sender side
2671 * because the pipeline is full.
2672 *
2673 * BSD also seems to "accidentally" limit itself to windows that are a
2674 * multiple of MSS, at least until the free space gets quite small.
2675 * This would appear to be a side effect of the mbuf implementation.
2676 * Combining these two algorithms results in the observed behavior
2677 * of having a fixed window size at almost all times.
2678 *
2679 * Below we obtain similar behavior by forcing the offered window to
2680 * a multiple of the mss when it is feasible to do so.
2681 *
2682 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2683 * Regular options like TIMESTAMP are taken into account.
2684 */
2685u32 __tcp_select_window(struct sock *sk)
2686{
2687	struct inet_connection_sock *icsk = inet_csk(sk);
2688	struct tcp_sock *tp = tcp_sk(sk);
2689	/* MSS for the peer's data.  Previous versions used mss_clamp
2690	 * here.  I don't know if the value based on our guesses
2691	 * of peer's MSS is better for the performance.  It's more correct
2692	 * but may be worse for the performance because of rcv_mss
2693	 * fluctuations.  --SAW  1998/11/1
2694	 */
2695	int mss = icsk->icsk_ack.rcv_mss;
2696	int free_space = tcp_space(sk);
2697	int allowed_space = tcp_full_space(sk);
2698	int full_space = min_t(int, tp->window_clamp, allowed_space);
2699	int window;
2700
2701	if (unlikely(mss > full_space)) {
2702		mss = full_space;
2703		if (mss <= 0)
2704			return 0;
2705	}
2706	if (free_space < (full_space >> 1)) {
2707		icsk->icsk_ack.quick = 0;
2708
2709		if (tcp_under_memory_pressure(sk))
2710			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2711					       4U * tp->advmss);
2712
2713		/* free_space might become our new window, make sure we don't
2714		 * increase it due to wscale.
2715		 */
2716		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2717
2718		/* if free space is less than mss estimate, or is below 1/16th
2719		 * of the maximum allowed, try to move to zero-window, else
2720		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2721		 * new incoming data is dropped due to memory limits.
2722		 * With large window, mss test triggers way too late in order
2723		 * to announce zero window in time before rmem limit kicks in.
2724		 */
2725		if (free_space < (allowed_space >> 4) || free_space < mss)
2726			return 0;
2727	}
2728
2729	if (free_space > tp->rcv_ssthresh)
2730		free_space = tp->rcv_ssthresh;
2731
2732	/* Don't do rounding if we are using window scaling, since the
2733	 * scaled window will not line up with the MSS boundary anyway.
2734	 */
 
2735	if (tp->rx_opt.rcv_wscale) {
2736		window = free_space;
2737
2738		/* Advertise enough space so that it won't get scaled away.
2739		 * Import case: prevent zero window announcement if
2740		 * 1<<rcv_wscale > mss.
2741		 */
2742		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
 
 
2743	} else {
2744		window = tp->rcv_wnd;
2745		/* Get the largest window that is a nice multiple of mss.
2746		 * Window clamp already applied above.
2747		 * If our current window offering is within 1 mss of the
2748		 * free space we just keep it. This prevents the divide
2749		 * and multiply from happening most of the time.
2750		 * We also don't do any window rounding when the free space
2751		 * is too small.
2752		 */
2753		if (window <= free_space - mss || window > free_space)
2754			window = rounddown(free_space, mss);
2755		else if (mss == full_space &&
2756			 free_space > window + (full_space >> 1))
2757			window = free_space;
2758	}
2759
2760	return window;
2761}
2762
2763void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2764			     const struct sk_buff *next_skb)
2765{
2766	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2767		const struct skb_shared_info *next_shinfo =
2768			skb_shinfo(next_skb);
2769		struct skb_shared_info *shinfo = skb_shinfo(skb);
2770
2771		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2772		shinfo->tskey = next_shinfo->tskey;
2773		TCP_SKB_CB(skb)->txstamp_ack |=
2774			TCP_SKB_CB(next_skb)->txstamp_ack;
2775	}
2776}
2777
2778/* Collapses two adjacent SKB's during retransmission. */
2779static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2780{
2781	struct tcp_sock *tp = tcp_sk(sk);
2782	struct sk_buff *next_skb = skb_rb_next(skb);
2783	int next_skb_size;
2784
 
2785	next_skb_size = next_skb->len;
2786
2787	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2788
2789	if (next_skb_size) {
2790		if (next_skb_size <= skb_availroom(skb))
2791			skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2792				      next_skb_size);
2793		else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2794			return false;
2795	}
2796	tcp_highest_sack_replace(sk, next_skb, skb);
 
 
 
 
2797
2798	/* Update sequence range on original skb. */
2799	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2800
2801	/* Merge over control information. This moves PSH/FIN etc. over */
2802	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2803
2804	/* All done, get rid of second SKB and account for it so
2805	 * packet counting does not break.
2806	 */
2807	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2808	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2809
2810	/* changed transmit queue under us so clear hints */
2811	tcp_clear_retrans_hints_partial(tp);
2812	if (next_skb == tp->retransmit_skb_hint)
2813		tp->retransmit_skb_hint = skb;
2814
2815	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2816
2817	tcp_skb_collapse_tstamp(skb, next_skb);
2818
2819	tcp_rtx_queue_unlink_and_free(next_skb, sk);
2820	return true;
2821}
2822
2823/* Check if coalescing SKBs is legal. */
2824static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2825{
2826	if (tcp_skb_pcount(skb) > 1)
2827		return false;
 
 
 
2828	if (skb_cloned(skb))
2829		return false;
2830	/* Some heuristics for collapsing over SACK'd could be invented */
 
 
2831	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2832		return false;
2833
2834	return true;
2835}
2836
2837/* Collapse packets in the retransmit queue to make to create
2838 * less packets on the wire. This is only done on retransmission.
2839 */
2840static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2841				     int space)
2842{
2843	struct tcp_sock *tp = tcp_sk(sk);
2844	struct sk_buff *skb = to, *tmp;
2845	bool first = true;
2846
2847	if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2848		return;
2849	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2850		return;
2851
2852	skb_rbtree_walk_from_safe(skb, tmp) {
2853		if (!tcp_can_collapse(sk, skb))
2854			break;
2855
2856		if (!tcp_skb_can_collapse_to(to))
2857			break;
2858
2859		space -= skb->len;
2860
2861		if (first) {
2862			first = false;
2863			continue;
2864		}
2865
2866		if (space < 0)
2867			break;
 
 
 
 
 
2868
2869		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2870			break;
2871
2872		if (!tcp_collapse_retrans(sk, to))
2873			break;
2874	}
2875}
2876
2877/* This retransmits one SKB.  Policy decisions and retransmit queue
2878 * state updates are done by the caller.  Returns non-zero if an
2879 * error occurred which prevented the send.
2880 */
2881int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2882{
2883	struct inet_connection_sock *icsk = inet_csk(sk);
2884	struct tcp_sock *tp = tcp_sk(sk);
 
2885	unsigned int cur_mss;
2886	int diff, len, err;
2887
2888
2889	/* Inconclusive MTU probe */
2890	if (icsk->icsk_mtup.probe_size)
2891		icsk->icsk_mtup.probe_size = 0;
 
2892
2893	/* Do not sent more than we queued. 1/4 is reserved for possible
2894	 * copying overhead: fragmentation, tunneling, mangling etc.
2895	 */
2896	if (refcount_read(&sk->sk_wmem_alloc) >
2897	    min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2898		  sk->sk_sndbuf))
2899		return -EAGAIN;
2900
2901	if (skb_still_in_host_queue(sk, skb))
2902		return -EBUSY;
2903
2904	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2905		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2906			WARN_ON_ONCE(1);
2907			return -EINVAL;
2908		}
2909		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2910			return -ENOMEM;
2911	}
2912
2913	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2914		return -EHOSTUNREACH; /* Routing failure or similar. */
2915
2916	cur_mss = tcp_current_mss(sk);
2917
2918	/* If receiver has shrunk his window, and skb is out of
2919	 * new window, do not retransmit it. The exception is the
2920	 * case, when window is shrunk to zero. In this case
2921	 * our retransmit serves as a zero window probe.
2922	 */
2923	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2924	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2925		return -EAGAIN;
2926
2927	len = cur_mss * segs;
2928	if (skb->len > len) {
2929		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2930				 cur_mss, GFP_ATOMIC))
2931			return -ENOMEM; /* We'll try again later. */
2932	} else {
2933		if (skb_unclone(skb, GFP_ATOMIC))
2934			return -ENOMEM;
2935
2936		diff = tcp_skb_pcount(skb);
2937		tcp_set_skb_tso_segs(skb, cur_mss);
2938		diff -= tcp_skb_pcount(skb);
2939		if (diff)
2940			tcp_adjust_pcount(sk, skb, diff);
2941		if (skb->len < cur_mss)
2942			tcp_retrans_try_collapse(sk, skb, cur_mss);
2943	}
2944
2945	/* RFC3168, section 6.1.1.1. ECN fallback */
2946	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2947		tcp_ecn_clear_syn(sk, skb);
2948
2949	/* Update global and local TCP statistics. */
2950	segs = tcp_skb_pcount(skb);
2951	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2952	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2953		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2954	tp->total_retrans += segs;
2955	tp->bytes_retrans += skb->len;
2956
2957	/* make sure skb->data is aligned on arches that require it
2958	 * and check if ack-trimming & collapsing extended the headroom
2959	 * beyond what csum_start can cover.
2960	 */
2961	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2962		     skb_headroom(skb) >= 0xFFFF)) {
2963		struct sk_buff *nskb;
2964
2965		tcp_skb_tsorted_save(skb) {
2966			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2967			err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2968				     -ENOBUFS;
2969		} tcp_skb_tsorted_restore(skb);
2970
2971		if (!err) {
2972			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
2973			tcp_rate_skb_sent(sk, skb);
2974		}
2975	} else {
2976		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2977	}
2978
2979	/* To avoid taking spuriously low RTT samples based on a timestamp
2980	 * for a transmit that never happened, always mark EVER_RETRANS
2981	 */
2982	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2983
2984	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2985		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2986				  TCP_SKB_CB(skb)->seq, segs, err);
2987
2988	if (likely(!err)) {
2989		trace_tcp_retransmit_skb(sk, skb);
2990	} else if (err != -EBUSY) {
2991		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2992	}
2993	return err;
2994}
2995
2996int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2997{
2998	struct tcp_sock *tp = tcp_sk(sk);
2999	int err = __tcp_retransmit_skb(sk, skb, segs);
3000
3001	if (err == 0) {
 
 
 
 
 
3002#if FASTRETRANS_DEBUG > 0
3003		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3004			net_dbg_ratelimited("retrans_out leaked\n");
 
3005		}
3006#endif
 
 
3007		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3008		tp->retrans_out += tcp_skb_pcount(skb);
3009	}
3010
3011	/* Save stamp of the first (attempted) retransmit. */
3012	if (!tp->retrans_stamp)
3013		tp->retrans_stamp = tcp_skb_timestamp(skb);
3014
3015	if (tp->undo_retrans < 0)
3016		tp->undo_retrans = 0;
3017	tp->undo_retrans += tcp_skb_pcount(skb);
 
 
 
 
3018	return err;
3019}
3020
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3021/* This gets called after a retransmit timeout, and the initially
3022 * retransmitted data is acknowledged.  It tries to continue
3023 * resending the rest of the retransmit queue, until either
3024 * we've sent it all or the congestion window limit is reached.
 
 
 
3025 */
3026void tcp_xmit_retransmit_queue(struct sock *sk)
3027{
3028	const struct inet_connection_sock *icsk = inet_csk(sk);
3029	struct sk_buff *skb, *rtx_head, *hole = NULL;
3030	struct tcp_sock *tp = tcp_sk(sk);
3031	u32 max_segs;
 
 
3032	int mib_idx;
 
3033
3034	if (!tp->packets_out)
3035		return;
3036
3037	rtx_head = tcp_rtx_queue_head(sk);
3038	skb = tp->retransmit_skb_hint ?: rtx_head;
3039	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3040	skb_rbtree_walk_from(skb) {
3041		__u8 sacked;
3042		int segs;
3043
3044		if (tcp_pacing_check(sk))
3045			break;
 
 
 
 
 
 
 
3046
 
 
 
 
 
3047		/* we could do better than to assign each time */
3048		if (!hole)
3049			tp->retransmit_skb_hint = skb;
3050
3051		segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3052		if (segs <= 0)
3053			return;
3054		sacked = TCP_SKB_CB(skb)->sacked;
3055		/* In case tcp_shift_skb_data() have aggregated large skbs,
3056		 * we need to make sure not sending too bigs TSO packets
3057		 */
3058		segs = min_t(int, segs, max_segs);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3059
3060		if (tp->retrans_out >= tp->lost_out) {
3061			break;
3062		} else if (!(sacked & TCPCB_LOST)) {
3063			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3064				hole = skb;
3065			continue;
3066
3067		} else {
 
3068			if (icsk->icsk_ca_state != TCP_CA_Loss)
3069				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3070			else
3071				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3072		}
3073
3074		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3075			continue;
3076
3077		if (tcp_small_queue_check(sk, skb, 1))
3078			return;
 
3079
3080		if (tcp_retransmit_skb(sk, skb, segs))
3081			return;
3082
3083		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3084
3085		if (tcp_in_cwnd_reduction(sk))
3086			tp->prr_out += tcp_skb_pcount(skb);
3087
3088		if (skb == rtx_head &&
3089		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3090			tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3091					     inet_csk(sk)->icsk_rto,
3092					     TCP_RTO_MAX,
3093					     skb);
3094	}
3095}
3096
3097/* We allow to exceed memory limits for FIN packets to expedite
3098 * connection tear down and (memory) recovery.
3099 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3100 * or even be forced to close flow without any FIN.
3101 * In general, we want to allow one skb per socket to avoid hangs
3102 * with edge trigger epoll()
3103 */
3104void sk_forced_mem_schedule(struct sock *sk, int size)
3105{
3106	int amt;
3107
3108	if (size <= sk->sk_forward_alloc)
3109		return;
3110	amt = sk_mem_pages(size);
3111	sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3112	sk_memory_allocated_add(sk, amt);
3113
3114	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3115		mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3116}
3117
3118/* Send a FIN. The caller locks the socket for us.
3119 * We should try to send a FIN packet really hard, but eventually give up.
3120 */
3121void tcp_send_fin(struct sock *sk)
3122{
3123	struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3124	struct tcp_sock *tp = tcp_sk(sk);
 
 
3125
3126	/* Optimization, tack on the FIN if we have one skb in write queue and
3127	 * this skb was not yet sent, or we are under memory pressure.
3128	 * Note: in the latter case, FIN packet will be sent after a timeout,
3129	 * as TCP stack thinks it has already been transmitted.
3130	 */
3131	if (!tskb && tcp_under_memory_pressure(sk))
3132		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3133
3134	if (tskb) {
3135		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3136		TCP_SKB_CB(tskb)->end_seq++;
3137		tp->write_seq++;
3138		if (tcp_write_queue_empty(sk)) {
3139			/* This means tskb was already sent.
3140			 * Pretend we included the FIN on previous transmit.
3141			 * We need to set tp->snd_nxt to the value it would have
3142			 * if FIN had been sent. This is because retransmit path
3143			 * does not change tp->snd_nxt.
3144			 */
3145			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3146			return;
3147		}
3148	} else {
3149		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3150		if (unlikely(!skb))
3151			return;
 
 
 
 
 
3152
3153		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3154		skb_reserve(skb, MAX_TCP_HEADER);
3155		sk_forced_mem_schedule(sk, skb->truesize);
3156		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3157		tcp_init_nondata_skb(skb, tp->write_seq,
3158				     TCPHDR_ACK | TCPHDR_FIN);
3159		tcp_queue_skb(sk, skb);
3160	}
3161	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3162}
3163
3164/* We get here when a process closes a file descriptor (either due to
3165 * an explicit close() or as a byproduct of exit()'ing) and there
3166 * was unread data in the receive queue.  This behavior is recommended
3167 * by RFC 2525, section 2.17.  -DaveM
3168 */
3169void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3170{
3171	struct sk_buff *skb;
3172
3173	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3174
3175	/* NOTE: No TCP options attached and we never retransmit this. */
3176	skb = alloc_skb(MAX_TCP_HEADER, priority);
3177	if (!skb) {
3178		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3179		return;
3180	}
3181
3182	/* Reserve space for headers and prepare control bits. */
3183	skb_reserve(skb, MAX_TCP_HEADER);
3184	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3185			     TCPHDR_ACK | TCPHDR_RST);
3186	tcp_mstamp_refresh(tcp_sk(sk));
3187	/* Send it off. */
 
3188	if (tcp_transmit_skb(sk, skb, 0, priority))
3189		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3190
3191	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3192	 * skb here is different to the troublesome skb, so use NULL
3193	 */
3194	trace_tcp_send_reset(sk, NULL);
3195}
3196
3197/* Send a crossed SYN-ACK during socket establishment.
3198 * WARNING: This routine must only be called when we have already sent
3199 * a SYN packet that crossed the incoming SYN that caused this routine
3200 * to get called. If this assumption fails then the initial rcv_wnd
3201 * and rcv_wscale values will not be correct.
3202 */
3203int tcp_send_synack(struct sock *sk)
3204{
3205	struct sk_buff *skb;
3206
3207	skb = tcp_rtx_queue_head(sk);
3208	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3209		pr_err("%s: wrong queue state\n", __func__);
3210		return -EFAULT;
3211	}
3212	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3213		if (skb_cloned(skb)) {
3214			struct sk_buff *nskb;
3215
3216			tcp_skb_tsorted_save(skb) {
3217				nskb = skb_copy(skb, GFP_ATOMIC);
3218			} tcp_skb_tsorted_restore(skb);
3219			if (!nskb)
3220				return -ENOMEM;
3221			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3222			tcp_rtx_queue_unlink_and_free(skb, sk);
3223			__skb_header_release(nskb);
3224			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3225			sk_wmem_queued_add(sk, nskb->truesize);
3226			sk_mem_charge(sk, nskb->truesize);
3227			skb = nskb;
3228		}
3229
3230		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3231		tcp_ecn_send_synack(sk, skb);
3232	}
 
3233	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3234}
3235
3236/**
3237 * tcp_make_synack - Prepare a SYN-ACK.
3238 * sk: listener socket
3239 * dst: dst entry attached to the SYNACK
3240 * req: request_sock pointer
3241 *
3242 * Allocate one skb and build a SYNACK packet.
3243 * @dst is consumed : Caller should not use it again.
3244 */
3245struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3246				struct request_sock *req,
3247				struct tcp_fastopen_cookie *foc,
3248				enum tcp_synack_type synack_type)
3249{
3250	struct inet_request_sock *ireq = inet_rsk(req);
3251	const struct tcp_sock *tp = tcp_sk(sk);
3252	struct tcp_md5sig_key *md5 = NULL;
3253	struct tcp_out_options opts;
 
 
 
 
 
3254	struct sk_buff *skb;
 
3255	int tcp_header_size;
3256	struct tcphdr *th;
3257	int mss;
3258	u64 now;
3259
3260	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3261	if (unlikely(!skb)) {
3262		dst_release(dst);
 
3263		return NULL;
3264	}
3265	/* Reserve space for headers. */
3266	skb_reserve(skb, MAX_TCP_HEADER);
3267
3268	switch (synack_type) {
3269	case TCP_SYNACK_NORMAL:
3270		skb_set_owner_w(skb, req_to_sk(req));
3271		break;
3272	case TCP_SYNACK_COOKIE:
3273		/* Under synflood, we do not attach skb to a socket,
3274		 * to avoid false sharing.
3275		 */
3276		break;
3277	case TCP_SYNACK_FASTOPEN:
3278		/* sk is a const pointer, because we want to express multiple
3279		 * cpu might call us concurrently.
3280		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3281		 */
3282		skb_set_owner_w(skb, (struct sock *)sk);
3283		break;
3284	}
3285	skb_dst_set(skb, dst);
3286
3287	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3288
3289	memset(&opts, 0, sizeof(opts));
3290	now = tcp_clock_ns();
3291#ifdef CONFIG_SYN_COOKIES
3292	if (unlikely(req->cookie_ts))
3293		skb->skb_mstamp_ns = cookie_init_timestamp(req);
3294	else
3295#endif
3296	{
3297		skb->skb_mstamp_ns = now;
3298		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3299			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3300	}
3301
3302#ifdef CONFIG_TCP_MD5SIG
3303	rcu_read_lock();
3304	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3305#endif
3306	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3307	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3308					     foc) + sizeof(*th);
3309
3310	skb_push(skb, tcp_header_size);
3311	skb_reset_transport_header(skb);
3312
3313	th = (struct tcphdr *)skb->data;
3314	memset(th, 0, sizeof(struct tcphdr));
3315	th->syn = 1;
3316	th->ack = 1;
3317	tcp_ecn_make_synack(req, th);
3318	th->source = htons(ireq->ir_num);
3319	th->dest = ireq->ir_rmt_port;
3320	skb->mark = ireq->ir_mark;
3321	skb->ip_summed = CHECKSUM_PARTIAL;
3322	th->seq = htonl(tcp_rsk(req)->snt_isn);
3323	/* XXX data is queued and acked as is. No buffer/window check */
3324	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3325
3326	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3327	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3328	tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3329	th->doff = (tcp_header_size >> 2);
3330	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3331
3332#ifdef CONFIG_TCP_MD5SIG
3333	/* Okay, we have all we need - do the md5 hash if needed */
3334	if (md5)
3335		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3336					       md5, req_to_sk(req), skb);
3337	rcu_read_unlock();
3338#endif
3339
3340	skb->skb_mstamp_ns = now;
3341	tcp_add_tx_delay(skb, tp);
3342
3343	return skb;
3344}
3345EXPORT_SYMBOL(tcp_make_synack);
3346
3347static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3348{
3349	struct inet_connection_sock *icsk = inet_csk(sk);
3350	const struct tcp_congestion_ops *ca;
3351	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3352
3353	if (ca_key == TCP_CA_UNSPEC)
3354		return;
3355
3356	rcu_read_lock();
3357	ca = tcp_ca_find_key(ca_key);
3358	if (likely(ca && try_module_get(ca->owner))) {
3359		module_put(icsk->icsk_ca_ops->owner);
3360		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3361		icsk->icsk_ca_ops = ca;
3362	}
3363	rcu_read_unlock();
3364}
3365
3366/* Do all connect socket setups that can be done AF independent. */
3367static void tcp_connect_init(struct sock *sk)
3368{
3369	const struct dst_entry *dst = __sk_dst_get(sk);
3370	struct tcp_sock *tp = tcp_sk(sk);
3371	__u8 rcv_wscale;
3372	u32 rcv_wnd;
3373
3374	/* We'll fix this up when we get a response from the other end.
3375	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3376	 */
3377	tp->tcp_header_len = sizeof(struct tcphdr);
3378	if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3379		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3380
3381#ifdef CONFIG_TCP_MD5SIG
3382	if (tp->af_specific->md5_lookup(sk, sk))
3383		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3384#endif
3385
3386	/* If user gave his TCP_MAXSEG, record it to clamp */
3387	if (tp->rx_opt.user_mss)
3388		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3389	tp->max_window = 0;
3390	tcp_mtup_init(sk);
3391	tcp_sync_mss(sk, dst_mtu(dst));
3392
3393	tcp_ca_dst_init(sk, dst);
3394
3395	if (!tp->window_clamp)
3396		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3397	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
 
 
3398
3399	tcp_initialize_rcv_mss(sk);
3400
3401	/* limit the window selection if the user enforce a smaller rx buffer */
3402	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3403	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3404		tp->window_clamp = tcp_full_space(sk);
3405
3406	rcv_wnd = tcp_rwnd_init_bpf(sk);
3407	if (rcv_wnd == 0)
3408		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3409
3410	tcp_select_initial_window(sk, tcp_full_space(sk),
3411				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3412				  &tp->rcv_wnd,
3413				  &tp->window_clamp,
3414				  sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3415				  &rcv_wscale,
3416				  rcv_wnd);
3417
3418	tp->rx_opt.rcv_wscale = rcv_wscale;
3419	tp->rcv_ssthresh = tp->rcv_wnd;
3420
3421	sk->sk_err = 0;
3422	sock_reset_flag(sk, SOCK_DONE);
3423	tp->snd_wnd = 0;
3424	tcp_init_wl(tp, 0);
3425	tcp_write_queue_purge(sk);
3426	tp->snd_una = tp->write_seq;
3427	tp->snd_sml = tp->write_seq;
3428	tp->snd_up = tp->write_seq;
3429	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3430
3431	if (likely(!tp->repair))
3432		tp->rcv_nxt = 0;
3433	else
3434		tp->rcv_tstamp = tcp_jiffies32;
3435	tp->rcv_wup = tp->rcv_nxt;
3436	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3437
3438	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3439	inet_csk(sk)->icsk_retransmits = 0;
3440	tcp_clear_retrans(tp);
3441}
3442
3443static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3444{
3445	struct tcp_sock *tp = tcp_sk(sk);
3446	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3447
3448	tcb->end_seq += skb->len;
3449	__skb_header_release(skb);
3450	sk_wmem_queued_add(sk, skb->truesize);
3451	sk_mem_charge(sk, skb->truesize);
3452	WRITE_ONCE(tp->write_seq, tcb->end_seq);
3453	tp->packets_out += tcp_skb_pcount(skb);
3454}
3455
3456/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3457 * queue a data-only packet after the regular SYN, such that regular SYNs
3458 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3459 * only the SYN sequence, the data are retransmitted in the first ACK.
3460 * If cookie is not cached or other error occurs, falls back to send a
3461 * regular SYN with Fast Open cookie request option.
3462 */
3463static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3464{
3465	struct tcp_sock *tp = tcp_sk(sk);
3466	struct tcp_fastopen_request *fo = tp->fastopen_req;
3467	int space, err = 0;
3468	struct sk_buff *syn_data;
3469
3470	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3471	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3472		goto fallback;
3473
3474	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3475	 * user-MSS. Reserve maximum option space for middleboxes that add
3476	 * private TCP options. The cost is reduced data space in SYN :(
3477	 */
3478	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3479
3480	space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3481		MAX_TCP_OPTION_SPACE;
3482
3483	space = min_t(size_t, space, fo->size);
3484
3485	/* limit to order-0 allocations */
3486	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3487
3488	syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3489	if (!syn_data)
3490		goto fallback;
3491	syn_data->ip_summed = CHECKSUM_PARTIAL;
3492	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3493	if (space) {
3494		int copied = copy_from_iter(skb_put(syn_data, space), space,
3495					    &fo->data->msg_iter);
3496		if (unlikely(!copied)) {
3497			tcp_skb_tsorted_anchor_cleanup(syn_data);
3498			kfree_skb(syn_data);
3499			goto fallback;
3500		}
3501		if (copied != space) {
3502			skb_trim(syn_data, copied);
3503			space = copied;
3504		}
3505		skb_zcopy_set(syn_data, fo->uarg, NULL);
3506	}
3507	/* No more data pending in inet_wait_for_connect() */
3508	if (space == fo->size)
3509		fo->data = NULL;
3510	fo->copied = space;
3511
3512	tcp_connect_queue_skb(sk, syn_data);
3513	if (syn_data->len)
3514		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3515
3516	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3517
3518	syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3519
3520	/* Now full SYN+DATA was cloned and sent (or not),
3521	 * remove the SYN from the original skb (syn_data)
3522	 * we keep in write queue in case of a retransmit, as we
3523	 * also have the SYN packet (with no data) in the same queue.
3524	 */
3525	TCP_SKB_CB(syn_data)->seq++;
3526	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3527	if (!err) {
3528		tp->syn_data = (fo->copied > 0);
3529		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3530		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3531		goto done;
3532	}
3533
3534	/* data was not sent, put it in write_queue */
3535	__skb_queue_tail(&sk->sk_write_queue, syn_data);
3536	tp->packets_out -= tcp_skb_pcount(syn_data);
3537
3538fallback:
3539	/* Send a regular SYN with Fast Open cookie request option */
3540	if (fo->cookie.len > 0)
3541		fo->cookie.len = 0;
3542	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3543	if (err)
3544		tp->syn_fastopen = 0;
3545done:
3546	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3547	return err;
3548}
3549
3550/* Build a SYN and send it off. */
3551int tcp_connect(struct sock *sk)
3552{
3553	struct tcp_sock *tp = tcp_sk(sk);
3554	struct sk_buff *buff;
3555	int err;
3556
3557	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3558
3559	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3560		return -EHOSTUNREACH; /* Routing failure or similar. */
3561
3562	tcp_connect_init(sk);
3563
3564	if (unlikely(tp->repair)) {
3565		tcp_finish_connect(sk, NULL);
3566		return 0;
3567	}
3568
3569	buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3570	if (unlikely(!buff))
3571		return -ENOBUFS;
3572
 
 
 
 
3573	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3574	tcp_mstamp_refresh(tp);
3575	tp->retrans_stamp = tcp_time_stamp(tp);
3576	tcp_connect_queue_skb(sk, buff);
3577	tcp_ecn_send_syn(sk, buff);
3578	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3579
3580	/* Send off SYN; include data in Fast Open. */
3581	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3582	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
 
 
3583	if (err == -ECONNREFUSED)
3584		return err;
3585
3586	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3587	 * in order to make this packet get counted in tcpOutSegs.
3588	 */
3589	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3590	tp->pushed_seq = tp->write_seq;
3591	buff = tcp_send_head(sk);
3592	if (unlikely(buff)) {
3593		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3594		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
3595	}
3596	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3597
3598	/* Timer for repeating the SYN until an answer. */
3599	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3600				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3601	return 0;
3602}
3603EXPORT_SYMBOL(tcp_connect);
3604
3605/* Send out a delayed ack, the caller does the policy checking
3606 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3607 * for details.
3608 */
3609void tcp_send_delayed_ack(struct sock *sk)
3610{
3611	struct inet_connection_sock *icsk = inet_csk(sk);
3612	int ato = icsk->icsk_ack.ato;
3613	unsigned long timeout;
3614
3615	if (ato > TCP_DELACK_MIN) {
3616		const struct tcp_sock *tp = tcp_sk(sk);
3617		int max_ato = HZ / 2;
3618
3619		if (inet_csk_in_pingpong_mode(sk) ||
3620		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3621			max_ato = TCP_DELACK_MAX;
3622
3623		/* Slow path, intersegment interval is "high". */
3624
3625		/* If some rtt estimate is known, use it to bound delayed ack.
3626		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3627		 * directly.
3628		 */
3629		if (tp->srtt_us) {
3630			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3631					TCP_DELACK_MIN);
3632
3633			if (rtt < max_ato)
3634				max_ato = rtt;
3635		}
3636
3637		ato = min(ato, max_ato);
3638	}
3639
3640	/* Stay within the limit we were given */
3641	timeout = jiffies + ato;
3642
3643	/* Use new timeout only if there wasn't a older one earlier. */
3644	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3645		/* If delack timer was blocked or is about to expire,
3646		 * send ACK now.
3647		 */
3648		if (icsk->icsk_ack.blocked ||
3649		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3650			tcp_send_ack(sk);
3651			return;
3652		}
3653
3654		if (!time_before(timeout, icsk->icsk_ack.timeout))
3655			timeout = icsk->icsk_ack.timeout;
3656	}
3657	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3658	icsk->icsk_ack.timeout = timeout;
3659	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3660}
3661
3662/* This routine sends an ack and also updates the window. */
3663void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3664{
3665	struct sk_buff *buff;
3666
3667	/* If we have been reset, we may not send again. */
3668	if (sk->sk_state == TCP_CLOSE)
3669		return;
3670
3671	/* We are not putting this on the write queue, so
3672	 * tcp_transmit_skb() will set the ownership to this
3673	 * sock.
3674	 */
3675	buff = alloc_skb(MAX_TCP_HEADER,
3676			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3677	if (unlikely(!buff)) {
3678		inet_csk_schedule_ack(sk);
3679		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3680		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3681					  TCP_DELACK_MAX, TCP_RTO_MAX);
3682		return;
3683	}
3684
3685	/* Reserve space for headers and prepare control bits. */
3686	skb_reserve(buff, MAX_TCP_HEADER);
3687	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3688
3689	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3690	 * too much.
3691	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3692	 */
3693	skb_set_tcp_pure_ack(buff);
3694
3695	/* Send it off, this clears delayed acks for us. */
3696	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3697}
3698EXPORT_SYMBOL_GPL(__tcp_send_ack);
3699
3700void tcp_send_ack(struct sock *sk)
3701{
3702	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3703}
3704
3705/* This routine sends a packet with an out of date sequence
3706 * number. It assumes the other end will try to ack it.
3707 *
3708 * Question: what should we make while urgent mode?
3709 * 4.4BSD forces sending single byte of data. We cannot send
3710 * out of window data, because we have SND.NXT==SND.MAX...
3711 *
3712 * Current solution: to send TWO zero-length segments in urgent mode:
3713 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3714 * out-of-date with SND.UNA-1 to probe window.
3715 */
3716static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3717{
3718	struct tcp_sock *tp = tcp_sk(sk);
3719	struct sk_buff *skb;
3720
3721	/* We don't queue it, tcp_transmit_skb() sets ownership. */
3722	skb = alloc_skb(MAX_TCP_HEADER,
3723			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3724	if (!skb)
3725		return -1;
3726
3727	/* Reserve space for headers and set control bits. */
3728	skb_reserve(skb, MAX_TCP_HEADER);
3729	/* Use a previous sequence.  This should cause the other
3730	 * end to send an ack.  Don't queue or clone SKB, just
3731	 * send it.
3732	 */
3733	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3734	NET_INC_STATS(sock_net(sk), mib);
3735	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3736}
3737
3738/* Called from setsockopt( ... TCP_REPAIR ) */
3739void tcp_send_window_probe(struct sock *sk)
3740{
3741	if (sk->sk_state == TCP_ESTABLISHED) {
3742		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3743		tcp_mstamp_refresh(tcp_sk(sk));
3744		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3745	}
3746}
3747
3748/* Initiate keepalive or window probe from timer. */
3749int tcp_write_wakeup(struct sock *sk, int mib)
3750{
3751	struct tcp_sock *tp = tcp_sk(sk);
3752	struct sk_buff *skb;
3753
3754	if (sk->sk_state == TCP_CLOSE)
3755		return -1;
3756
3757	skb = tcp_send_head(sk);
3758	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3759		int err;
3760		unsigned int mss = tcp_current_mss(sk);
3761		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3762
3763		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3764			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3765
3766		/* We are probing the opening of a window
3767		 * but the window size is != 0
3768		 * must have been a result SWS avoidance ( sender )
3769		 */
3770		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3771		    skb->len > mss) {
3772			seg_size = min(seg_size, mss);
3773			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3774			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3775					 skb, seg_size, mss, GFP_ATOMIC))
3776				return -1;
3777		} else if (!tcp_skb_pcount(skb))
3778			tcp_set_skb_tso_segs(skb, mss);
3779
3780		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 
3781		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3782		if (!err)
3783			tcp_event_new_data_sent(sk, skb);
3784		return err;
3785	} else {
3786		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3787			tcp_xmit_probe_skb(sk, 1, mib);
3788		return tcp_xmit_probe_skb(sk, 0, mib);
3789	}
3790}
3791
3792/* A window probe timeout has occurred.  If window is not closed send
3793 * a partial packet else a zero probe.
3794 */
3795void tcp_send_probe0(struct sock *sk)
3796{
3797	struct inet_connection_sock *icsk = inet_csk(sk);
3798	struct tcp_sock *tp = tcp_sk(sk);
3799	struct net *net = sock_net(sk);
3800	unsigned long timeout;
3801	int err;
3802
3803	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3804
3805	if (tp->packets_out || tcp_write_queue_empty(sk)) {
3806		/* Cancel probe timer, if it is not required. */
3807		icsk->icsk_probes_out = 0;
3808		icsk->icsk_backoff = 0;
3809		return;
3810	}
3811
3812	icsk->icsk_probes_out++;
3813	if (err <= 0) {
3814		if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3815			icsk->icsk_backoff++;
3816		timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
 
 
 
3817	} else {
3818		/* If packet was not sent due to local congestion,
3819		 * Let senders fight for local resources conservatively.
 
 
 
3820		 */
3821		timeout = TCP_RESOURCE_PROBE_INTERVAL;
3822	}
3823	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3824}
3825
3826int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3827{
3828	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3829	struct flowi fl;
3830	int res;
3831
3832	tcp_rsk(req)->txhash = net_tx_rndhash();
3833	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3834	if (!res) {
3835		__TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3836		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3837		if (unlikely(tcp_passive_fastopen(sk)))
3838			tcp_sk(sk)->total_retrans++;
3839		trace_tcp_retransmit_synack(sk, req);
3840	}
3841	return res;
3842}
3843EXPORT_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}