Linux Audio

Check our new training course

Loading...
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}
v3.5.6
   1/*
   2 * INET		An implementation of the TCP/IP protocol suite for the LINUX
   3 *		operating system.  INET is implemented using the  BSD Socket
   4 *		interface as the means of communication with the user level.
   5 *
   6 *		Implementation of the Transmission Control Protocol(TCP).
   7 *
   8 * Authors:	Ross Biro
   9 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  10 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
  11 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
  12 *		Florian La Roche, <flla@stud.uni-sb.de>
  13 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  14 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
  15 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
  16 *		Matthew Dillon, <dillon@apollo.west.oic.com>
  17 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  18 *		Jorge Cwik, <jorge@laser.satlink.net>
  19 */
  20
  21/*
  22 * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
  23 *				:	Fragmentation on mtu decrease
  24 *				:	Segment collapse on retransmit
  25 *				:	AF independence
  26 *
  27 *		Linus Torvalds	:	send_delayed_ack
  28 *		David S. Miller	:	Charge memory using the right skb
  29 *					during syn/ack processing.
  30 *		David S. Miller :	Output engine completely rewritten.
  31 *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
  32 *		Cacophonix Gaul :	draft-minshall-nagle-01
  33 *		J Hadi Salim	:	ECN support
  34 *
  35 */
  36
  37#define pr_fmt(fmt) "TCP: " fmt
  38
  39#include <net/tcp.h>
  40
  41#include <linux/compiler.h>
  42#include <linux/gfp.h>
  43#include <linux/module.h>
  44
  45/* People can turn this off for buggy TCP's found in printers etc. */
  46int sysctl_tcp_retrans_collapse __read_mostly = 1;
  47
  48/* People can turn this on to work with those rare, broken TCPs that
  49 * interpret the window field as a signed quantity.
  50 */
  51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
  52
  53/* This limits the percentage of the congestion window which we
  54 * will allow a single TSO frame to consume.  Building TSO frames
  55 * which are too large can cause TCP streams to be bursty.
  56 */
  57int sysctl_tcp_tso_win_divisor __read_mostly = 3;
  58
  59int sysctl_tcp_mtu_probing __read_mostly = 0;
  60int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
  61
  62/* By default, RFC2861 behavior.  */
  63int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
  64
  65int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
  66EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
  67
  68
  69/* Account for new data that has been sent to the network. */
  70static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
  71{
  72	struct tcp_sock *tp = tcp_sk(sk);
  73	unsigned int prior_packets = tp->packets_out;
  74
  75	tcp_advance_send_head(sk, skb);
  76	tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
  77
  78	/* Don't override Nagle indefinitely with F-RTO */
  79	if (tp->frto_counter == 2)
  80		tp->frto_counter = 3;
  81
  82	tp->packets_out += tcp_skb_pcount(skb);
  83	if (!prior_packets || tp->early_retrans_delayed)
  84		tcp_rearm_rto(sk);
 
  85}
  86
  87/* SND.NXT, if window was not shrunk.
  88 * If window has been shrunk, what should we make? It is not clear at all.
  89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  91 * invalid. OK, let's make this for now:
  92 */
  93static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  94{
  95	const struct tcp_sock *tp = tcp_sk(sk);
  96
  97	if (!before(tcp_wnd_end(tp), tp->snd_nxt))
  98		return tp->snd_nxt;
  99	else
 100		return tcp_wnd_end(tp);
 101}
 102
 103/* Calculate mss to advertise in SYN segment.
 104 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 105 *
 106 * 1. It is independent of path mtu.
 107 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 108 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 109 *    attached devices, because some buggy hosts are confused by
 110 *    large MSS.
 111 * 4. We do not make 3, we advertise MSS, calculated from first
 112 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 113 *    This may be overridden via information stored in routing table.
 114 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 115 *    probably even Jumbo".
 116 */
 117static __u16 tcp_advertise_mss(struct sock *sk)
 118{
 119	struct tcp_sock *tp = tcp_sk(sk);
 120	const struct dst_entry *dst = __sk_dst_get(sk);
 121	int mss = tp->advmss;
 122
 123	if (dst) {
 124		unsigned int metric = dst_metric_advmss(dst);
 125
 126		if (metric < mss) {
 127			mss = metric;
 128			tp->advmss = mss;
 129		}
 130	}
 131
 132	return (__u16)mss;
 133}
 134
 135/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 136 * This is the first part of cwnd validation mechanism. */
 137static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
 138{
 139	struct tcp_sock *tp = tcp_sk(sk);
 140	s32 delta = tcp_time_stamp - tp->lsndtime;
 141	u32 restart_cwnd = tcp_init_cwnd(tp, dst);
 142	u32 cwnd = tp->snd_cwnd;
 143
 144	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 145
 146	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 147	restart_cwnd = min(restart_cwnd, cwnd);
 148
 149	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 150		cwnd >>= 1;
 151	tp->snd_cwnd = max(cwnd, restart_cwnd);
 152	tp->snd_cwnd_stamp = tcp_time_stamp;
 153	tp->snd_cwnd_used = 0;
 154}
 155
 156/* Congestion state accounting after a packet has been sent. */
 157static void tcp_event_data_sent(struct tcp_sock *tp,
 158				struct sock *sk)
 159{
 160	struct inet_connection_sock *icsk = inet_csk(sk);
 161	const u32 now = tcp_time_stamp;
 162
 163	if (sysctl_tcp_slow_start_after_idle &&
 164	    (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
 165		tcp_cwnd_restart(sk, __sk_dst_get(sk));
 166
 167	tp->lsndtime = now;
 168
 169	/* If it is a reply for ato after last received
 170	 * packet, enter pingpong mode.
 171	 */
 172	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 173		icsk->icsk_ack.pingpong = 1;
 174}
 175
 176/* Account for an ACK we sent. */
 177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
 178{
 179	tcp_dec_quickack_mode(sk, pkts);
 180	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 181}
 182
 183/* Determine a window scaling and initial window to offer.
 184 * Based on the assumption that the given amount of space
 185 * will be offered. Store the results in the tp structure.
 186 * NOTE: for smooth operation initial space offering should
 187 * be a multiple of mss if possible. We assume here that mss >= 1.
 188 * This MUST be enforced by all callers.
 189 */
 190void tcp_select_initial_window(int __space, __u32 mss,
 191			       __u32 *rcv_wnd, __u32 *window_clamp,
 192			       int wscale_ok, __u8 *rcv_wscale,
 193			       __u32 init_rcv_wnd)
 194{
 195	unsigned int space = (__space < 0 ? 0 : __space);
 196
 197	/* If no clamp set the clamp to the max possible scaled window */
 198	if (*window_clamp == 0)
 199		(*window_clamp) = (65535 << 14);
 200	space = min(*window_clamp, space);
 201
 202	/* Quantize space offering to a multiple of mss if possible. */
 203	if (space > mss)
 204		space = (space / mss) * mss;
 205
 206	/* NOTE: offering an initial window larger than 32767
 207	 * will break some buggy TCP stacks. If the admin tells us
 208	 * it is likely we could be speaking with such a buggy stack
 209	 * we will truncate our initial window offering to 32K-1
 210	 * unless the remote has sent us a window scaling option,
 211	 * which we interpret as a sign the remote TCP is not
 212	 * misinterpreting the window field as a signed quantity.
 213	 */
 214	if (sysctl_tcp_workaround_signed_windows)
 215		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 216	else
 217		(*rcv_wnd) = space;
 218
 219	(*rcv_wscale) = 0;
 220	if (wscale_ok) {
 221		/* Set window scaling on max possible window
 222		 * See RFC1323 for an explanation of the limit to 14
 223		 */
 224		space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
 225		space = min_t(u32, space, *window_clamp);
 226		while (space > 65535 && (*rcv_wscale) < 14) {
 227			space >>= 1;
 228			(*rcv_wscale)++;
 229		}
 230	}
 231
 232	/* Set initial window to a value enough for senders starting with
 233	 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
 234	 * a limit on the initial window when mss is larger than 1460.
 235	 */
 236	if (mss > (1 << *rcv_wscale)) {
 237		int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
 238		if (mss > 1460)
 239			init_cwnd =
 240			max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
 241		/* when initializing use the value from init_rcv_wnd
 242		 * rather than the default from above
 243		 */
 244		if (init_rcv_wnd)
 245			*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 246		else
 247			*rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
 248	}
 249
 250	/* Set the clamp no higher than max representable value */
 251	(*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
 252}
 253EXPORT_SYMBOL(tcp_select_initial_window);
 254
 255/* Chose a new window to advertise, update state in tcp_sock for the
 256 * socket, and return result with RFC1323 scaling applied.  The return
 257 * value can be stuffed directly into th->window for an outgoing
 258 * frame.
 259 */
 260static u16 tcp_select_window(struct sock *sk)
 261{
 262	struct tcp_sock *tp = tcp_sk(sk);
 263	u32 cur_win = tcp_receive_window(tp);
 264	u32 new_win = __tcp_select_window(sk);
 265
 266	/* Never shrink the offered window */
 267	if (new_win < cur_win) {
 268		/* Danger Will Robinson!
 269		 * Don't update rcv_wup/rcv_wnd here or else
 270		 * we will not be able to advertise a zero
 271		 * window in time.  --DaveM
 272		 *
 273		 * Relax Will Robinson.
 274		 */
 275		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 276	}
 277	tp->rcv_wnd = new_win;
 278	tp->rcv_wup = tp->rcv_nxt;
 279
 280	/* Make sure we do not exceed the maximum possible
 281	 * scaled window.
 282	 */
 283	if (!tp->rx_opt.rcv_wscale && 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
 295	return new_win;
 296}
 297
 298/* Packet ECN state for a SYN-ACK */
 299static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
 300{
 301	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 302	if (!(tp->ecn_flags & TCP_ECN_OK))
 303		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 304}
 305
 306/* Packet ECN state for a SYN.  */
 307static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
 308{
 309	struct tcp_sock *tp = tcp_sk(sk);
 310
 311	tp->ecn_flags = 0;
 312	if (sysctl_tcp_ecn == 1) {
 313		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 314		tp->ecn_flags = TCP_ECN_OK;
 315	}
 316}
 317
 318static __inline__ void
 319TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
 320{
 321	if (inet_rsk(req)->ecn_ok)
 322		th->ece = 1;
 323}
 324
 325/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 326 * be sent.
 327 */
 328static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
 329				int tcp_header_len)
 330{
 331	struct tcp_sock *tp = tcp_sk(sk);
 332
 333	if (tp->ecn_flags & TCP_ECN_OK) {
 334		/* Not-retransmitted data segment: set ECT and inject CWR. */
 335		if (skb->len != tcp_header_len &&
 336		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 337			INET_ECN_xmit(sk);
 338			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 339				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 340				tcp_hdr(skb)->cwr = 1;
 341				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 342			}
 343		} else {
 344			/* ACK or retransmitted segment: clear ECT|CE */
 345			INET_ECN_dontxmit(sk);
 346		}
 347		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 348			tcp_hdr(skb)->ece = 1;
 349	}
 350}
 351
 352/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 353 * auto increment end seqno.
 354 */
 355static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 356{
 357	skb->ip_summed = CHECKSUM_PARTIAL;
 358	skb->csum = 0;
 359
 360	TCP_SKB_CB(skb)->tcp_flags = flags;
 361	TCP_SKB_CB(skb)->sacked = 0;
 362
 363	skb_shinfo(skb)->gso_segs = 1;
 364	skb_shinfo(skb)->gso_size = 0;
 365	skb_shinfo(skb)->gso_type = 0;
 366
 367	TCP_SKB_CB(skb)->seq = seq;
 368	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 369		seq++;
 370	TCP_SKB_CB(skb)->end_seq = seq;
 371}
 372
 373static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 374{
 375	return tp->snd_una != tp->snd_up;
 376}
 377
 378#define OPTION_SACK_ADVERTISE	(1 << 0)
 379#define OPTION_TS		(1 << 1)
 380#define OPTION_MD5		(1 << 2)
 381#define OPTION_WSCALE		(1 << 3)
 382#define OPTION_COOKIE_EXTENSION	(1 << 4)
 383
 384struct tcp_out_options {
 385	u8 options;		/* bit field of OPTION_* */
 386	u8 ws;			/* window scale, 0 to disable */
 387	u8 num_sack_blocks;	/* number of SACK blocks to include */
 388	u8 hash_size;		/* bytes in hash_location */
 389	u16 mss;		/* 0 to disable */
 390	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 391	__u8 *hash_location;	/* temporary pointer, overloaded */
 392};
 393
 394/* The sysctl int routines are generic, so check consistency here.
 395 */
 396static u8 tcp_cookie_size_check(u8 desired)
 397{
 398	int cookie_size;
 399
 400	if (desired > 0)
 401		/* previously specified */
 402		return desired;
 403
 404	cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
 405	if (cookie_size <= 0)
 406		/* no default specified */
 407		return 0;
 408
 409	if (cookie_size <= TCP_COOKIE_MIN)
 410		/* value too small, specify minimum */
 411		return TCP_COOKIE_MIN;
 412
 413	if (cookie_size >= TCP_COOKIE_MAX)
 414		/* value too large, specify maximum */
 415		return TCP_COOKIE_MAX;
 416
 417	if (cookie_size & 1)
 418		/* 8-bit multiple, illegal, fix it */
 419		cookie_size++;
 420
 421	return (u8)cookie_size;
 422}
 423
 424/* Write previously computed TCP options to the packet.
 425 *
 426 * Beware: Something in the Internet is very sensitive to the ordering of
 427 * TCP options, we learned this through the hard way, so be careful here.
 428 * Luckily we can at least blame others for their non-compliance but from
 429 * inter-operatibility perspective it seems that we're somewhat stuck with
 430 * the ordering which we have been using if we want to keep working with
 431 * those broken things (not that it currently hurts anybody as there isn't
 432 * particular reason why the ordering would need to be changed).
 433 *
 434 * At least SACK_PERM as the first option is known to lead to a disaster
 435 * (but it may well be that other scenarios fail similarly).
 436 */
 437static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
 438			      struct tcp_out_options *opts)
 439{
 440	u8 options = opts->options;	/* mungable copy */
 441
 442	/* Having both authentication and cookies for security is redundant,
 443	 * and there's certainly not enough room.  Instead, the cookie-less
 444	 * extension variant is proposed.
 445	 *
 446	 * Consider the pessimal case with authentication.  The options
 447	 * could look like:
 448	 *   COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
 449	 */
 450	if (unlikely(OPTION_MD5 & options)) {
 451		if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 452			*ptr++ = htonl((TCPOPT_COOKIE << 24) |
 453				       (TCPOLEN_COOKIE_BASE << 16) |
 454				       (TCPOPT_MD5SIG << 8) |
 455				       TCPOLEN_MD5SIG);
 456		} else {
 457			*ptr++ = htonl((TCPOPT_NOP << 24) |
 458				       (TCPOPT_NOP << 16) |
 459				       (TCPOPT_MD5SIG << 8) |
 460				       TCPOLEN_MD5SIG);
 461		}
 462		options &= ~OPTION_COOKIE_EXTENSION;
 463		/* overload cookie hash location */
 464		opts->hash_location = (__u8 *)ptr;
 465		ptr += 4;
 466	}
 467
 468	if (unlikely(opts->mss)) {
 469		*ptr++ = htonl((TCPOPT_MSS << 24) |
 470			       (TCPOLEN_MSS << 16) |
 471			       opts->mss);
 472	}
 473
 474	if (likely(OPTION_TS & options)) {
 475		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 476			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 477				       (TCPOLEN_SACK_PERM << 16) |
 478				       (TCPOPT_TIMESTAMP << 8) |
 479				       TCPOLEN_TIMESTAMP);
 480			options &= ~OPTION_SACK_ADVERTISE;
 481		} else {
 482			*ptr++ = htonl((TCPOPT_NOP << 24) |
 483				       (TCPOPT_NOP << 16) |
 484				       (TCPOPT_TIMESTAMP << 8) |
 485				       TCPOLEN_TIMESTAMP);
 486		}
 487		*ptr++ = htonl(opts->tsval);
 488		*ptr++ = htonl(opts->tsecr);
 489	}
 490
 491	/* Specification requires after timestamp, so do it now.
 492	 *
 493	 * Consider the pessimal case without authentication.  The options
 494	 * could look like:
 495	 *   MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
 496	 */
 497	if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
 498		__u8 *cookie_copy = opts->hash_location;
 499		u8 cookie_size = opts->hash_size;
 500
 501		/* 8-bit multiple handled in tcp_cookie_size_check() above,
 502		 * and elsewhere.
 503		 */
 504		if (0x2 & cookie_size) {
 505			__u8 *p = (__u8 *)ptr;
 506
 507			/* 16-bit multiple */
 508			*p++ = TCPOPT_COOKIE;
 509			*p++ = TCPOLEN_COOKIE_BASE + cookie_size;
 510			*p++ = *cookie_copy++;
 511			*p++ = *cookie_copy++;
 512			ptr++;
 513			cookie_size -= 2;
 514		} else {
 515			/* 32-bit multiple */
 516			*ptr++ = htonl(((TCPOPT_NOP << 24) |
 517					(TCPOPT_NOP << 16) |
 518					(TCPOPT_COOKIE << 8) |
 519					TCPOLEN_COOKIE_BASE) +
 520				       cookie_size);
 521		}
 522
 523		if (cookie_size > 0) {
 524			memcpy(ptr, cookie_copy, cookie_size);
 525			ptr += (cookie_size / 4);
 526		}
 527	}
 528
 529	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 530		*ptr++ = htonl((TCPOPT_NOP << 24) |
 531			       (TCPOPT_NOP << 16) |
 532			       (TCPOPT_SACK_PERM << 8) |
 533			       TCPOLEN_SACK_PERM);
 534	}
 535
 536	if (unlikely(OPTION_WSCALE & options)) {
 537		*ptr++ = htonl((TCPOPT_NOP << 24) |
 538			       (TCPOPT_WINDOW << 16) |
 539			       (TCPOLEN_WINDOW << 8) |
 540			       opts->ws);
 541	}
 542
 543	if (unlikely(opts->num_sack_blocks)) {
 544		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 545			tp->duplicate_sack : tp->selective_acks;
 546		int this_sack;
 547
 548		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 549			       (TCPOPT_NOP  << 16) |
 550			       (TCPOPT_SACK <<  8) |
 551			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 552						     TCPOLEN_SACK_PERBLOCK)));
 553
 554		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 555		     ++this_sack) {
 556			*ptr++ = htonl(sp[this_sack].start_seq);
 557			*ptr++ = htonl(sp[this_sack].end_seq);
 558		}
 559
 560		tp->rx_opt.dsack = 0;
 561	}
 562}
 563
 564/* Compute TCP options for SYN packets. This is not the final
 565 * network wire format yet.
 566 */
 567static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 568				struct tcp_out_options *opts,
 569				struct tcp_md5sig_key **md5)
 570{
 571	struct tcp_sock *tp = tcp_sk(sk);
 572	struct tcp_cookie_values *cvp = tp->cookie_values;
 573	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 574	u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
 575			 tcp_cookie_size_check(cvp->cookie_desired) :
 576			 0;
 577
 578#ifdef CONFIG_TCP_MD5SIG
 579	*md5 = tp->af_specific->md5_lookup(sk, sk);
 580	if (*md5) {
 581		opts->options |= OPTION_MD5;
 582		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 583	}
 584#else
 585	*md5 = NULL;
 586#endif
 587
 588	/* We always get an MSS option.  The option bytes which will be seen in
 589	 * normal data packets should timestamps be used, must be in the MSS
 590	 * advertised.  But we subtract them from tp->mss_cache so that
 591	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 592	 * fact here if necessary.  If we don't do this correctly, as a
 593	 * receiver we won't recognize data packets as being full sized when we
 594	 * should, and thus we won't abide by the delayed ACK rules correctly.
 595	 * SACKs don't matter, we never delay an ACK when we have any of those
 596	 * going out.  */
 597	opts->mss = tcp_advertise_mss(sk);
 598	remaining -= TCPOLEN_MSS_ALIGNED;
 599
 600	if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
 601		opts->options |= OPTION_TS;
 602		opts->tsval = TCP_SKB_CB(skb)->when;
 603		opts->tsecr = tp->rx_opt.ts_recent;
 604		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 605	}
 606	if (likely(sysctl_tcp_window_scaling)) {
 607		opts->ws = tp->rx_opt.rcv_wscale;
 608		opts->options |= OPTION_WSCALE;
 609		remaining -= TCPOLEN_WSCALE_ALIGNED;
 610	}
 611	if (likely(sysctl_tcp_sack)) {
 612		opts->options |= OPTION_SACK_ADVERTISE;
 613		if (unlikely(!(OPTION_TS & opts->options)))
 614			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 615	}
 616
 617	/* Note that timestamps are required by the specification.
 618	 *
 619	 * Odd numbers of bytes are prohibited by the specification, ensuring
 620	 * that the cookie is 16-bit aligned, and the resulting cookie pair is
 621	 * 32-bit aligned.
 622	 */
 623	if (*md5 == NULL &&
 624	    (OPTION_TS & opts->options) &&
 625	    cookie_size > 0) {
 626		int need = TCPOLEN_COOKIE_BASE + cookie_size;
 627
 628		if (0x2 & need) {
 629			/* 32-bit multiple */
 630			need += 2; /* NOPs */
 631
 632			if (need > remaining) {
 633				/* try shrinking cookie to fit */
 634				cookie_size -= 2;
 635				need -= 4;
 636			}
 637		}
 638		while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
 639			cookie_size -= 4;
 640			need -= 4;
 641		}
 642		if (TCP_COOKIE_MIN <= cookie_size) {
 643			opts->options |= OPTION_COOKIE_EXTENSION;
 644			opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
 645			opts->hash_size = cookie_size;
 646
 647			/* Remember for future incarnations. */
 648			cvp->cookie_desired = cookie_size;
 649
 650			if (cvp->cookie_desired != cvp->cookie_pair_size) {
 651				/* Currently use random bytes as a nonce,
 652				 * assuming these are completely unpredictable
 653				 * by hostile users of the same system.
 654				 */
 655				get_random_bytes(&cvp->cookie_pair[0],
 656						 cookie_size);
 657				cvp->cookie_pair_size = cookie_size;
 658			}
 659
 660			remaining -= need;
 661		}
 662	}
 663	return MAX_TCP_OPTION_SPACE - remaining;
 664}
 665
 666/* Set up TCP options for SYN-ACKs. */
 667static unsigned int tcp_synack_options(struct sock *sk,
 668				   struct request_sock *req,
 669				   unsigned int mss, struct sk_buff *skb,
 670				   struct tcp_out_options *opts,
 671				   struct tcp_md5sig_key **md5,
 672				   struct tcp_extend_values *xvp)
 673{
 674	struct inet_request_sock *ireq = inet_rsk(req);
 675	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 676	u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
 677			 xvp->cookie_plus :
 678			 0;
 679
 680#ifdef CONFIG_TCP_MD5SIG
 681	*md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
 682	if (*md5) {
 683		opts->options |= OPTION_MD5;
 684		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 685
 686		/* We can't fit any SACK blocks in a packet with MD5 + TS
 687		 * options. There was discussion about disabling SACK
 688		 * rather than TS in order to fit in better with old,
 689		 * buggy kernels, but that was deemed to be unnecessary.
 690		 */
 691		ireq->tstamp_ok &= !ireq->sack_ok;
 692	}
 693#else
 694	*md5 = NULL;
 695#endif
 696
 697	/* We always send an MSS option. */
 698	opts->mss = mss;
 699	remaining -= TCPOLEN_MSS_ALIGNED;
 700
 701	if (likely(ireq->wscale_ok)) {
 702		opts->ws = ireq->rcv_wscale;
 703		opts->options |= OPTION_WSCALE;
 704		remaining -= TCPOLEN_WSCALE_ALIGNED;
 705	}
 706	if (likely(ireq->tstamp_ok)) {
 707		opts->options |= OPTION_TS;
 708		opts->tsval = TCP_SKB_CB(skb)->when;
 709		opts->tsecr = req->ts_recent;
 710		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 711	}
 712	if (likely(ireq->sack_ok)) {
 713		opts->options |= OPTION_SACK_ADVERTISE;
 714		if (unlikely(!ireq->tstamp_ok))
 715			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 716	}
 717
 718	/* Similar rationale to tcp_syn_options() applies here, too.
 719	 * If the <SYN> options fit, the same options should fit now!
 720	 */
 721	if (*md5 == NULL &&
 722	    ireq->tstamp_ok &&
 723	    cookie_plus > TCPOLEN_COOKIE_BASE) {
 724		int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
 725
 726		if (0x2 & need) {
 727			/* 32-bit multiple */
 728			need += 2; /* NOPs */
 729		}
 730		if (need <= remaining) {
 731			opts->options |= OPTION_COOKIE_EXTENSION;
 732			opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
 733			remaining -= need;
 734		} else {
 735			/* There's no error return, so flag it. */
 736			xvp->cookie_out_never = 1; /* true */
 737			opts->hash_size = 0;
 738		}
 739	}
 740	return MAX_TCP_OPTION_SPACE - remaining;
 741}
 742
 743/* Compute TCP options for ESTABLISHED sockets. This is not the
 744 * final wire format yet.
 745 */
 746static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 747					struct tcp_out_options *opts,
 748					struct tcp_md5sig_key **md5)
 749{
 750	struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
 751	struct tcp_sock *tp = tcp_sk(sk);
 752	unsigned int size = 0;
 753	unsigned int eff_sacks;
 754
 755#ifdef CONFIG_TCP_MD5SIG
 756	*md5 = tp->af_specific->md5_lookup(sk, sk);
 757	if (unlikely(*md5)) {
 758		opts->options |= OPTION_MD5;
 759		size += TCPOLEN_MD5SIG_ALIGNED;
 760	}
 761#else
 762	*md5 = NULL;
 763#endif
 764
 765	if (likely(tp->rx_opt.tstamp_ok)) {
 766		opts->options |= OPTION_TS;
 767		opts->tsval = tcb ? tcb->when : 0;
 768		opts->tsecr = tp->rx_opt.ts_recent;
 769		size += TCPOLEN_TSTAMP_ALIGNED;
 770	}
 771
 772	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 773	if (unlikely(eff_sacks)) {
 774		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 775		opts->num_sack_blocks =
 776			min_t(unsigned int, eff_sacks,
 777			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 778			      TCPOLEN_SACK_PERBLOCK);
 779		size += TCPOLEN_SACK_BASE_ALIGNED +
 780			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 781	}
 782
 783	return size;
 784}
 785
 786/* This routine actually transmits TCP packets queued in by
 787 * tcp_do_sendmsg().  This is used by both the initial
 788 * transmission and possible later retransmissions.
 789 * All SKB's seen here are completely headerless.  It is our
 790 * job to build the TCP header, and pass the packet down to
 791 * IP so it can do the same plus pass the packet off to the
 792 * device.
 793 *
 794 * We are working here with either a clone of the original
 795 * SKB, or a fresh unique copy made by the retransmit engine.
 796 */
 797static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
 798			    gfp_t gfp_mask)
 799{
 800	const struct inet_connection_sock *icsk = inet_csk(sk);
 801	struct inet_sock *inet;
 802	struct tcp_sock *tp;
 803	struct tcp_skb_cb *tcb;
 804	struct tcp_out_options opts;
 805	unsigned int tcp_options_size, tcp_header_size;
 806	struct tcp_md5sig_key *md5;
 807	struct tcphdr *th;
 808	int err;
 809
 810	BUG_ON(!skb || !tcp_skb_pcount(skb));
 811
 812	/* If congestion control is doing timestamping, we must
 813	 * take such a timestamp before we potentially clone/copy.
 814	 */
 815	if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
 816		__net_timestamp(skb);
 817
 818	if (likely(clone_it)) {
 819		if (unlikely(skb_cloned(skb)))
 820			skb = pskb_copy(skb, gfp_mask);
 821		else
 822			skb = skb_clone(skb, gfp_mask);
 823		if (unlikely(!skb))
 824			return -ENOBUFS;
 825	}
 826
 827	inet = inet_sk(sk);
 828	tp = tcp_sk(sk);
 829	tcb = TCP_SKB_CB(skb);
 830	memset(&opts, 0, sizeof(opts));
 831
 832	if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
 833		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
 834	else
 835		tcp_options_size = tcp_established_options(sk, skb, &opts,
 836							   &md5);
 837	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
 838
 839	if (tcp_packets_in_flight(tp) == 0) {
 840		tcp_ca_event(sk, CA_EVENT_TX_START);
 841		skb->ooo_okay = 1;
 842	} else
 843		skb->ooo_okay = 0;
 844
 845	skb_push(skb, tcp_header_size);
 846	skb_reset_transport_header(skb);
 847	skb_set_owner_w(skb, sk);
 848
 849	/* Build TCP header and checksum it. */
 850	th = tcp_hdr(skb);
 851	th->source		= inet->inet_sport;
 852	th->dest		= inet->inet_dport;
 853	th->seq			= htonl(tcb->seq);
 854	th->ack_seq		= htonl(tp->rcv_nxt);
 855	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
 856					tcb->tcp_flags);
 857
 858	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
 859		/* RFC1323: The window in SYN & SYN/ACK segments
 860		 * is never scaled.
 861		 */
 862		th->window	= htons(min(tp->rcv_wnd, 65535U));
 863	} else {
 864		th->window	= htons(tcp_select_window(sk));
 865	}
 866	th->check		= 0;
 867	th->urg_ptr		= 0;
 868
 869	/* The urg_mode check is necessary during a below snd_una win probe */
 870	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
 871		if (before(tp->snd_up, tcb->seq + 0x10000)) {
 872			th->urg_ptr = htons(tp->snd_up - tcb->seq);
 873			th->urg = 1;
 874		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
 875			th->urg_ptr = htons(0xFFFF);
 876			th->urg = 1;
 877		}
 878	}
 879
 880	tcp_options_write((__be32 *)(th + 1), tp, &opts);
 881	if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
 882		TCP_ECN_send(sk, skb, tcp_header_size);
 883
 884#ifdef CONFIG_TCP_MD5SIG
 885	/* Calculate the MD5 hash, as we have all we need now */
 886	if (md5) {
 887		sk_nocaps_add(sk, NETIF_F_GSO_MASK);
 888		tp->af_specific->calc_md5_hash(opts.hash_location,
 889					       md5, sk, NULL, skb);
 890	}
 891#endif
 892
 893	icsk->icsk_af_ops->send_check(sk, skb);
 894
 895	if (likely(tcb->tcp_flags & TCPHDR_ACK))
 896		tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
 897
 898	if (skb->len != tcp_header_size)
 899		tcp_event_data_sent(tp, sk);
 900
 901	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
 902		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
 903			      tcp_skb_pcount(skb));
 904
 905	err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
 906	if (likely(err <= 0))
 907		return err;
 908
 909	tcp_enter_cwr(sk, 1);
 910
 911	return net_xmit_eval(err);
 912}
 913
 914/* This routine just queues the buffer for sending.
 915 *
 916 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
 917 * otherwise socket can stall.
 918 */
 919static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
 920{
 921	struct tcp_sock *tp = tcp_sk(sk);
 922
 923	/* Advance write_seq and place onto the write_queue. */
 924	tp->write_seq = TCP_SKB_CB(skb)->end_seq;
 925	skb_header_release(skb);
 926	tcp_add_write_queue_tail(sk, skb);
 927	sk->sk_wmem_queued += skb->truesize;
 928	sk_mem_charge(sk, skb->truesize);
 929}
 930
 931/* Initialize TSO segments for a packet. */
 932static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
 933				 unsigned int mss_now)
 934{
 935	if (skb->len <= mss_now || !sk_can_gso(sk) ||
 936	    skb->ip_summed == CHECKSUM_NONE) {
 937		/* Avoid the costly divide in the normal
 938		 * non-TSO case.
 939		 */
 940		skb_shinfo(skb)->gso_segs = 1;
 941		skb_shinfo(skb)->gso_size = 0;
 942		skb_shinfo(skb)->gso_type = 0;
 943	} else {
 944		skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
 945		skb_shinfo(skb)->gso_size = mss_now;
 946		skb_shinfo(skb)->gso_type = sk->sk_gso_type;
 947	}
 948}
 949
 950/* When a modification to fackets out becomes necessary, we need to check
 951 * skb is counted to fackets_out or not.
 952 */
 953static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
 954				   int decr)
 955{
 956	struct tcp_sock *tp = tcp_sk(sk);
 957
 958	if (!tp->sacked_out || tcp_is_reno(tp))
 959		return;
 960
 961	if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
 962		tp->fackets_out -= decr;
 963}
 964
 965/* Pcount in the middle of the write queue got changed, we need to do various
 966 * tweaks to fix counters
 967 */
 968static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
 969{
 970	struct tcp_sock *tp = tcp_sk(sk);
 971
 972	tp->packets_out -= decr;
 973
 974	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
 975		tp->sacked_out -= decr;
 976	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
 977		tp->retrans_out -= decr;
 978	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
 979		tp->lost_out -= decr;
 980
 981	/* Reno case is special. Sigh... */
 982	if (tcp_is_reno(tp) && decr > 0)
 983		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
 984
 985	tcp_adjust_fackets_out(sk, skb, decr);
 986
 987	if (tp->lost_skb_hint &&
 988	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
 989	    (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
 990		tp->lost_cnt_hint -= decr;
 991
 992	tcp_verify_left_out(tp);
 993}
 994
 995/* Function to create two new TCP segments.  Shrinks the given segment
 996 * to the specified size and appends a new segment with the rest of the
 997 * packet to the list.  This won't be called frequently, I hope.
 998 * Remember, these are still headerless SKBs at this point.
 999 */
1000int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1001		 unsigned int mss_now)
1002{
1003	struct tcp_sock *tp = tcp_sk(sk);
1004	struct sk_buff *buff;
1005	int nsize, old_factor;
1006	int nlen;
1007	u8 flags;
1008
1009	if (WARN_ON(len > skb->len))
1010		return -EINVAL;
1011
1012	nsize = skb_headlen(skb) - len;
1013	if (nsize < 0)
1014		nsize = 0;
1015
1016	if (skb_cloned(skb) &&
1017	    skb_is_nonlinear(skb) &&
1018	    pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1019		return -ENOMEM;
1020
1021	/* Get a new skb... force flag on. */
1022	buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1023	if (buff == NULL)
1024		return -ENOMEM; /* We'll just try again later. */
1025
1026	sk->sk_wmem_queued += buff->truesize;
1027	sk_mem_charge(sk, buff->truesize);
1028	nlen = skb->len - len - nsize;
1029	buff->truesize += nlen;
1030	skb->truesize -= nlen;
1031
1032	/* Correct the sequence numbers. */
1033	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1034	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1035	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1036
1037	/* PSH and FIN should only be set in the second packet. */
1038	flags = TCP_SKB_CB(skb)->tcp_flags;
1039	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1040	TCP_SKB_CB(buff)->tcp_flags = flags;
1041	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1042
1043	if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1044		/* Copy and checksum data tail into the new buffer. */
1045		buff->csum = csum_partial_copy_nocheck(skb->data + len,
1046						       skb_put(buff, nsize),
1047						       nsize, 0);
1048
1049		skb_trim(skb, len);
1050
1051		skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1052	} else {
1053		skb->ip_summed = CHECKSUM_PARTIAL;
1054		skb_split(skb, buff, len);
1055	}
1056
1057	buff->ip_summed = skb->ip_summed;
1058
1059	/* Looks stupid, but our code really uses when of
1060	 * skbs, which it never sent before. --ANK
1061	 */
1062	TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1063	buff->tstamp = skb->tstamp;
1064
1065	old_factor = tcp_skb_pcount(skb);
1066
1067	/* Fix up tso_factor for both original and new SKB.  */
1068	tcp_set_skb_tso_segs(sk, skb, mss_now);
1069	tcp_set_skb_tso_segs(sk, buff, mss_now);
1070
1071	/* If this packet has been sent out already, we must
1072	 * adjust the various packet counters.
1073	 */
1074	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1075		int diff = old_factor - tcp_skb_pcount(skb) -
1076			tcp_skb_pcount(buff);
1077
1078		if (diff)
1079			tcp_adjust_pcount(sk, skb, diff);
1080	}
1081
1082	/* Link BUFF into the send queue. */
1083	skb_header_release(buff);
1084	tcp_insert_write_queue_after(skb, buff, sk);
1085
1086	return 0;
1087}
1088
1089/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1090 * eventually). The difference is that pulled data not copied, but
1091 * immediately discarded.
1092 */
1093static void __pskb_trim_head(struct sk_buff *skb, int len)
1094{
1095	int i, k, eat;
1096
1097	eat = min_t(int, len, skb_headlen(skb));
1098	if (eat) {
1099		__skb_pull(skb, eat);
1100		skb->avail_size -= eat;
1101		len -= eat;
1102		if (!len)
1103			return;
1104	}
1105	eat = len;
1106	k = 0;
1107	for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1108		int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1109
1110		if (size <= eat) {
1111			skb_frag_unref(skb, i);
1112			eat -= size;
1113		} else {
1114			skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1115			if (eat) {
1116				skb_shinfo(skb)->frags[k].page_offset += eat;
1117				skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1118				eat = 0;
1119			}
1120			k++;
1121		}
1122	}
1123	skb_shinfo(skb)->nr_frags = k;
1124
1125	skb_reset_tail_pointer(skb);
1126	skb->data_len -= len;
1127	skb->len = skb->data_len;
1128}
1129
1130/* Remove acked data from a packet in the transmit queue. */
1131int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1132{
1133	if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1134		return -ENOMEM;
1135
1136	__pskb_trim_head(skb, len);
 
 
 
 
1137
1138	TCP_SKB_CB(skb)->seq += len;
1139	skb->ip_summed = CHECKSUM_PARTIAL;
1140
1141	skb->truesize	     -= len;
1142	sk->sk_wmem_queued   -= len;
1143	sk_mem_uncharge(sk, len);
1144	sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1145
1146	/* Any change of skb->len requires recalculation of tso factor. */
 
 
1147	if (tcp_skb_pcount(skb) > 1)
1148		tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1149
1150	return 0;
1151}
1152
1153/* Calculate MSS. Not accounting for SACKs here.  */
1154int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1155{
1156	const struct tcp_sock *tp = tcp_sk(sk);
1157	const struct inet_connection_sock *icsk = inet_csk(sk);
1158	int mss_now;
1159
1160	/* Calculate base mss without TCP options:
1161	   It is MMS_S - sizeof(tcphdr) of rfc1122
1162	 */
1163	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1164
1165	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1166	if (icsk->icsk_af_ops->net_frag_header_len) {
1167		const struct dst_entry *dst = __sk_dst_get(sk);
1168
1169		if (dst && dst_allfrag(dst))
1170			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1171	}
1172
1173	/* Clamp it (mss_clamp does not include tcp options) */
1174	if (mss_now > tp->rx_opt.mss_clamp)
1175		mss_now = tp->rx_opt.mss_clamp;
1176
1177	/* Now subtract optional transport overhead */
1178	mss_now -= icsk->icsk_ext_hdr_len;
1179
1180	/* Then reserve room for full set of TCP options and 8 bytes of data */
1181	if (mss_now < 48)
1182		mss_now = 48;
1183
1184	/* Now subtract TCP options size, not including SACKs */
1185	mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1186
1187	return mss_now;
1188}
1189
1190/* Inverse of above */
1191int tcp_mss_to_mtu(struct sock *sk, int mss)
1192{
1193	const struct tcp_sock *tp = tcp_sk(sk);
1194	const struct inet_connection_sock *icsk = inet_csk(sk);
1195	int mtu;
1196
1197	mtu = mss +
1198	      tp->tcp_header_len +
1199	      icsk->icsk_ext_hdr_len +
1200	      icsk->icsk_af_ops->net_header_len;
1201
1202	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1203	if (icsk->icsk_af_ops->net_frag_header_len) {
1204		const struct dst_entry *dst = __sk_dst_get(sk);
1205
1206		if (dst && dst_allfrag(dst))
1207			mtu += icsk->icsk_af_ops->net_frag_header_len;
1208	}
1209	return mtu;
1210}
1211
1212/* MTU probing init per socket */
1213void tcp_mtup_init(struct sock *sk)
1214{
1215	struct tcp_sock *tp = tcp_sk(sk);
1216	struct inet_connection_sock *icsk = inet_csk(sk);
1217
1218	icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1219	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1220			       icsk->icsk_af_ops->net_header_len;
1221	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1222	icsk->icsk_mtup.probe_size = 0;
1223}
1224EXPORT_SYMBOL(tcp_mtup_init);
1225
1226/* This function synchronize snd mss to current pmtu/exthdr set.
1227
1228   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1229   for TCP options, but includes only bare TCP header.
1230
1231   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1232   It is minimum of user_mss and mss received with SYN.
1233   It also does not include TCP options.
1234
1235   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1236
1237   tp->mss_cache is current effective sending mss, including
1238   all tcp options except for SACKs. It is evaluated,
1239   taking into account current pmtu, but never exceeds
1240   tp->rx_opt.mss_clamp.
1241
1242   NOTE1. rfc1122 clearly states that advertised MSS
1243   DOES NOT include either tcp or ip options.
1244
1245   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1246   are READ ONLY outside this function.		--ANK (980731)
1247 */
1248unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1249{
1250	struct tcp_sock *tp = tcp_sk(sk);
1251	struct inet_connection_sock *icsk = inet_csk(sk);
1252	int mss_now;
1253
1254	if (icsk->icsk_mtup.search_high > pmtu)
1255		icsk->icsk_mtup.search_high = pmtu;
1256
1257	mss_now = tcp_mtu_to_mss(sk, pmtu);
1258	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1259
1260	/* And store cached results */
1261	icsk->icsk_pmtu_cookie = pmtu;
1262	if (icsk->icsk_mtup.enabled)
1263		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1264	tp->mss_cache = mss_now;
1265
1266	return mss_now;
1267}
1268EXPORT_SYMBOL(tcp_sync_mss);
1269
1270/* Compute the current effective MSS, taking SACKs and IP options,
1271 * and even PMTU discovery events into account.
1272 */
1273unsigned int tcp_current_mss(struct sock *sk)
1274{
1275	const struct tcp_sock *tp = tcp_sk(sk);
1276	const struct dst_entry *dst = __sk_dst_get(sk);
1277	u32 mss_now;
1278	unsigned int header_len;
1279	struct tcp_out_options opts;
1280	struct tcp_md5sig_key *md5;
1281
1282	mss_now = tp->mss_cache;
1283
1284	if (dst) {
1285		u32 mtu = dst_mtu(dst);
1286		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1287			mss_now = tcp_sync_mss(sk, mtu);
1288	}
1289
1290	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1291		     sizeof(struct tcphdr);
1292	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1293	 * some common options. If this is an odd packet (because we have SACK
1294	 * blocks etc) then our calculated header_len will be different, and
1295	 * we have to adjust mss_now correspondingly */
1296	if (header_len != tp->tcp_header_len) {
1297		int delta = (int) header_len - tp->tcp_header_len;
1298		mss_now -= delta;
1299	}
1300
1301	return mss_now;
1302}
1303
1304/* Congestion window validation. (RFC2861) */
1305static void tcp_cwnd_validate(struct sock *sk)
1306{
1307	struct tcp_sock *tp = tcp_sk(sk);
1308
1309	if (tp->packets_out >= tp->snd_cwnd) {
1310		/* Network is feed fully. */
1311		tp->snd_cwnd_used = 0;
1312		tp->snd_cwnd_stamp = tcp_time_stamp;
1313	} else {
1314		/* Network starves. */
1315		if (tp->packets_out > tp->snd_cwnd_used)
1316			tp->snd_cwnd_used = tp->packets_out;
1317
1318		if (sysctl_tcp_slow_start_after_idle &&
1319		    (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1320			tcp_cwnd_application_limited(sk);
1321	}
1322}
1323
1324/* Returns the portion of skb which can be sent right away without
1325 * introducing MSS oddities to segment boundaries. In rare cases where
1326 * mss_now != mss_cache, we will request caller to create a small skb
1327 * per input skb which could be mostly avoided here (if desired).
1328 *
1329 * We explicitly want to create a request for splitting write queue tail
1330 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1331 * thus all the complexity (cwnd_len is always MSS multiple which we
1332 * return whenever allowed by the other factors). Basically we need the
1333 * modulo only when the receiver window alone is the limiting factor or
1334 * when we would be allowed to send the split-due-to-Nagle skb fully.
1335 */
1336static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1337					unsigned int mss_now, unsigned int max_segs)
1338{
1339	const struct tcp_sock *tp = tcp_sk(sk);
1340	u32 needed, window, max_len;
1341
1342	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1343	max_len = mss_now * max_segs;
1344
1345	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1346		return max_len;
1347
1348	needed = min(skb->len, window);
1349
1350	if (max_len <= needed)
1351		return max_len;
1352
1353	return needed - needed % mss_now;
1354}
1355
1356/* Can at least one segment of SKB be sent right now, according to the
1357 * congestion window rules?  If so, return how many segments are allowed.
1358 */
1359static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1360					 const struct sk_buff *skb)
1361{
1362	u32 in_flight, cwnd;
1363
1364	/* Don't be strict about the congestion window for the final FIN.  */
1365	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1366	    tcp_skb_pcount(skb) == 1)
1367		return 1;
1368
1369	in_flight = tcp_packets_in_flight(tp);
1370	cwnd = tp->snd_cwnd;
1371	if (in_flight < cwnd)
1372		return (cwnd - in_flight);
1373
1374	return 0;
1375}
1376
1377/* Initialize TSO state of a skb.
1378 * This must be invoked the first time we consider transmitting
1379 * SKB onto the wire.
1380 */
1381static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1382			     unsigned int mss_now)
1383{
1384	int tso_segs = tcp_skb_pcount(skb);
1385
1386	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1387		tcp_set_skb_tso_segs(sk, skb, mss_now);
1388		tso_segs = tcp_skb_pcount(skb);
1389	}
1390	return tso_segs;
1391}
1392
1393/* Minshall's variant of the Nagle send check. */
1394static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1395{
1396	return after(tp->snd_sml, tp->snd_una) &&
1397		!after(tp->snd_sml, tp->snd_nxt);
1398}
1399
1400/* Return false, if packet can be sent now without violation Nagle's rules:
1401 * 1. It is full sized.
1402 * 2. Or it contains FIN. (already checked by caller)
1403 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1404 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1405 *    With Minshall's modification: all sent small packets are ACKed.
1406 */
1407static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1408				  const struct sk_buff *skb,
1409				  unsigned int mss_now, int nonagle)
1410{
1411	return skb->len < mss_now &&
1412		((nonagle & TCP_NAGLE_CORK) ||
1413		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1414}
1415
1416/* Return true if the Nagle test allows this packet to be
1417 * sent now.
1418 */
1419static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1420				  unsigned int cur_mss, int nonagle)
1421{
1422	/* Nagle rule does not apply to frames, which sit in the middle of the
1423	 * write_queue (they have no chances to get new data).
1424	 *
1425	 * This is implemented in the callers, where they modify the 'nonagle'
1426	 * argument based upon the location of SKB in the send queue.
1427	 */
1428	if (nonagle & TCP_NAGLE_PUSH)
1429		return true;
1430
1431	/* Don't use the nagle rule for urgent data (or for the final FIN).
1432	 * Nagle can be ignored during F-RTO too (see RFC4138).
1433	 */
1434	if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1435	    (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1436		return true;
1437
1438	if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1439		return true;
1440
1441	return false;
1442}
1443
1444/* Does at least the first segment of SKB fit into the send window? */
1445static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1446			     const struct sk_buff *skb,
1447			     unsigned int cur_mss)
1448{
1449	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1450
1451	if (skb->len > cur_mss)
1452		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1453
1454	return !after(end_seq, tcp_wnd_end(tp));
1455}
1456
1457/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1458 * should be put on the wire right now.  If so, it returns the number of
1459 * packets allowed by the congestion window.
1460 */
1461static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1462				 unsigned int cur_mss, int nonagle)
1463{
1464	const struct tcp_sock *tp = tcp_sk(sk);
1465	unsigned int cwnd_quota;
1466
1467	tcp_init_tso_segs(sk, skb, cur_mss);
1468
1469	if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1470		return 0;
1471
1472	cwnd_quota = tcp_cwnd_test(tp, skb);
1473	if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1474		cwnd_quota = 0;
1475
1476	return cwnd_quota;
1477}
1478
1479/* Test if sending is allowed right now. */
1480bool tcp_may_send_now(struct sock *sk)
1481{
1482	const struct tcp_sock *tp = tcp_sk(sk);
1483	struct sk_buff *skb = tcp_send_head(sk);
1484
1485	return skb &&
1486		tcp_snd_test(sk, skb, tcp_current_mss(sk),
1487			     (tcp_skb_is_last(sk, skb) ?
1488			      tp->nonagle : TCP_NAGLE_PUSH));
1489}
1490
1491/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1492 * which is put after SKB on the list.  It is very much like
1493 * tcp_fragment() except that it may make several kinds of assumptions
1494 * in order to speed up the splitting operation.  In particular, we
1495 * know that all the data is in scatter-gather pages, and that the
1496 * packet has never been sent out before (and thus is not cloned).
1497 */
1498static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1499			unsigned int mss_now, gfp_t gfp)
1500{
1501	struct sk_buff *buff;
1502	int nlen = skb->len - len;
1503	u8 flags;
1504
1505	/* All of a TSO frame must be composed of paged data.  */
1506	if (skb->len != skb->data_len)
1507		return tcp_fragment(sk, skb, len, mss_now);
1508
1509	buff = sk_stream_alloc_skb(sk, 0, gfp);
1510	if (unlikely(buff == NULL))
1511		return -ENOMEM;
1512
1513	sk->sk_wmem_queued += buff->truesize;
1514	sk_mem_charge(sk, buff->truesize);
1515	buff->truesize += nlen;
1516	skb->truesize -= nlen;
1517
1518	/* Correct the sequence numbers. */
1519	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1520	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1521	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1522
1523	/* PSH and FIN should only be set in the second packet. */
1524	flags = TCP_SKB_CB(skb)->tcp_flags;
1525	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1526	TCP_SKB_CB(buff)->tcp_flags = flags;
1527
1528	/* This packet was never sent out yet, so no SACK bits. */
1529	TCP_SKB_CB(buff)->sacked = 0;
1530
1531	buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1532	skb_split(skb, buff, len);
1533
1534	/* Fix up tso_factor for both original and new SKB.  */
1535	tcp_set_skb_tso_segs(sk, skb, mss_now);
1536	tcp_set_skb_tso_segs(sk, buff, mss_now);
1537
1538	/* Link BUFF into the send queue. */
1539	skb_header_release(buff);
1540	tcp_insert_write_queue_after(skb, buff, sk);
1541
1542	return 0;
1543}
1544
1545/* Try to defer sending, if possible, in order to minimize the amount
1546 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
1547 *
1548 * This algorithm is from John Heffner.
1549 */
1550static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1551{
1552	struct tcp_sock *tp = tcp_sk(sk);
1553	const struct inet_connection_sock *icsk = inet_csk(sk);
1554	u32 send_win, cong_win, limit, in_flight;
1555	int win_divisor;
1556
1557	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1558		goto send_now;
1559
1560	if (icsk->icsk_ca_state != TCP_CA_Open)
1561		goto send_now;
1562
1563	/* Defer for less than two clock ticks. */
1564	if (tp->tso_deferred &&
1565	    (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1566		goto send_now;
1567
1568	in_flight = tcp_packets_in_flight(tp);
1569
1570	BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1571
1572	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1573
1574	/* From in_flight test above, we know that cwnd > in_flight.  */
1575	cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1576
1577	limit = min(send_win, cong_win);
1578
1579	/* If a full-sized TSO skb can be sent, do it. */
1580	if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1581			   sk->sk_gso_max_segs * tp->mss_cache))
1582		goto send_now;
1583
1584	/* Middle in queue won't get any more data, full sendable already? */
1585	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1586		goto send_now;
1587
1588	win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1589	if (win_divisor) {
1590		u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1591
1592		/* If at least some fraction of a window is available,
1593		 * just use it.
1594		 */
1595		chunk /= win_divisor;
1596		if (limit >= chunk)
1597			goto send_now;
1598	} else {
1599		/* Different approach, try not to defer past a single
1600		 * ACK.  Receiver should ACK every other full sized
1601		 * frame, so if we have space for more than 3 frames
1602		 * then send now.
1603		 */
1604		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1605			goto send_now;
1606	}
1607
1608	/* Ok, it looks like it is advisable to defer.  */
1609	tp->tso_deferred = 1 | (jiffies << 1);
1610
1611	return true;
1612
1613send_now:
1614	tp->tso_deferred = 0;
1615	return false;
1616}
1617
1618/* Create a new MTU probe if we are ready.
1619 * MTU probe is regularly attempting to increase the path MTU by
1620 * deliberately sending larger packets.  This discovers routing
1621 * changes resulting in larger path MTUs.
1622 *
1623 * Returns 0 if we should wait to probe (no cwnd available),
1624 *         1 if a probe was sent,
1625 *         -1 otherwise
1626 */
1627static int tcp_mtu_probe(struct sock *sk)
1628{
1629	struct tcp_sock *tp = tcp_sk(sk);
1630	struct inet_connection_sock *icsk = inet_csk(sk);
1631	struct sk_buff *skb, *nskb, *next;
1632	int len;
1633	int probe_size;
1634	int size_needed;
1635	int copy;
1636	int mss_now;
1637
1638	/* Not currently probing/verifying,
1639	 * not in recovery,
1640	 * have enough cwnd, and
1641	 * not SACKing (the variable headers throw things off) */
1642	if (!icsk->icsk_mtup.enabled ||
1643	    icsk->icsk_mtup.probe_size ||
1644	    inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1645	    tp->snd_cwnd < 11 ||
1646	    tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1647		return -1;
1648
1649	/* Very simple search strategy: just double the MSS. */
1650	mss_now = tcp_current_mss(sk);
1651	probe_size = 2 * tp->mss_cache;
1652	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1653	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1654		/* TODO: set timer for probe_converge_event */
1655		return -1;
1656	}
1657
1658	/* Have enough data in the send queue to probe? */
1659	if (tp->write_seq - tp->snd_nxt < size_needed)
1660		return -1;
1661
1662	if (tp->snd_wnd < size_needed)
1663		return -1;
1664	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1665		return 0;
1666
1667	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
1668	if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1669		if (!tcp_packets_in_flight(tp))
1670			return -1;
1671		else
1672			return 0;
1673	}
1674
1675	/* We're allowed to probe.  Build it now. */
1676	if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1677		return -1;
1678	sk->sk_wmem_queued += nskb->truesize;
1679	sk_mem_charge(sk, nskb->truesize);
1680
1681	skb = tcp_send_head(sk);
1682
1683	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1684	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1685	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1686	TCP_SKB_CB(nskb)->sacked = 0;
1687	nskb->csum = 0;
1688	nskb->ip_summed = skb->ip_summed;
1689
1690	tcp_insert_write_queue_before(nskb, skb, sk);
1691
1692	len = 0;
1693	tcp_for_write_queue_from_safe(skb, next, sk) {
1694		copy = min_t(int, skb->len, probe_size - len);
1695		if (nskb->ip_summed)
1696			skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1697		else
1698			nskb->csum = skb_copy_and_csum_bits(skb, 0,
1699							    skb_put(nskb, copy),
1700							    copy, nskb->csum);
1701
1702		if (skb->len <= copy) {
1703			/* We've eaten all the data from this skb.
1704			 * Throw it away. */
1705			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1706			tcp_unlink_write_queue(skb, sk);
1707			sk_wmem_free_skb(sk, skb);
1708		} else {
1709			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1710						   ~(TCPHDR_FIN|TCPHDR_PSH);
1711			if (!skb_shinfo(skb)->nr_frags) {
1712				skb_pull(skb, copy);
1713				if (skb->ip_summed != CHECKSUM_PARTIAL)
1714					skb->csum = csum_partial(skb->data,
1715								 skb->len, 0);
1716			} else {
1717				__pskb_trim_head(skb, copy);
1718				tcp_set_skb_tso_segs(sk, skb, mss_now);
1719			}
1720			TCP_SKB_CB(skb)->seq += copy;
1721		}
1722
1723		len += copy;
1724
1725		if (len >= probe_size)
1726			break;
1727	}
1728	tcp_init_tso_segs(sk, nskb, nskb->len);
1729
1730	/* We're ready to send.  If this fails, the probe will
1731	 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1732	TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1733	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1734		/* Decrement cwnd here because we are sending
1735		 * effectively two packets. */
1736		tp->snd_cwnd--;
1737		tcp_event_new_data_sent(sk, nskb);
1738
1739		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1740		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1741		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1742
1743		return 1;
1744	}
1745
1746	return -1;
1747}
1748
1749/* This routine writes packets to the network.  It advances the
1750 * send_head.  This happens as incoming acks open up the remote
1751 * window for us.
1752 *
1753 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1754 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1755 * account rare use of URG, this is not a big flaw.
1756 *
1757 * Returns true, if no segments are in flight and we have queued segments,
1758 * but cannot send anything now because of SWS or another problem.
1759 */
1760static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1761			   int push_one, gfp_t gfp)
1762{
1763	struct tcp_sock *tp = tcp_sk(sk);
1764	struct sk_buff *skb;
1765	unsigned int tso_segs, sent_pkts;
1766	int cwnd_quota;
1767	int result;
1768
1769	sent_pkts = 0;
1770
1771	if (!push_one) {
1772		/* Do MTU probing. */
1773		result = tcp_mtu_probe(sk);
1774		if (!result) {
1775			return false;
1776		} else if (result > 0) {
1777			sent_pkts = 1;
1778		}
1779	}
1780
1781	while ((skb = tcp_send_head(sk))) {
1782		unsigned int limit;
1783
1784		tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1785		BUG_ON(!tso_segs);
1786
1787		cwnd_quota = tcp_cwnd_test(tp, skb);
1788		if (!cwnd_quota)
1789			break;
1790
1791		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1792			break;
1793
1794		if (tso_segs == 1) {
1795			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1796						     (tcp_skb_is_last(sk, skb) ?
1797						      nonagle : TCP_NAGLE_PUSH))))
1798				break;
1799		} else {
1800			if (!push_one && tcp_tso_should_defer(sk, skb))
1801				break;
1802		}
1803
1804		limit = mss_now;
1805		if (tso_segs > 1 && !tcp_urg_mode(tp))
1806			limit = tcp_mss_split_point(sk, skb, mss_now,
1807						    min_t(unsigned int,
1808							  cwnd_quota,
1809							  sk->sk_gso_max_segs));
1810
1811		if (skb->len > limit &&
1812		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1813			break;
1814
1815		TCP_SKB_CB(skb)->when = tcp_time_stamp;
1816
1817		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1818			break;
1819
1820		/* Advance the send_head.  This one is sent out.
1821		 * This call will increment packets_out.
1822		 */
1823		tcp_event_new_data_sent(sk, skb);
1824
1825		tcp_minshall_update(tp, mss_now, skb);
1826		sent_pkts += tcp_skb_pcount(skb);
1827
1828		if (push_one)
1829			break;
1830	}
1831	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
1832		tp->prr_out += sent_pkts;
1833
1834	if (likely(sent_pkts)) {
1835		tcp_cwnd_validate(sk);
1836		return false;
1837	}
1838	return !tp->packets_out && tcp_send_head(sk);
1839}
1840
1841/* Push out any pending frames which were held back due to
1842 * TCP_CORK or attempt at coalescing tiny packets.
1843 * The socket must be locked by the caller.
1844 */
1845void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1846			       int nonagle)
1847{
1848	/* If we are closed, the bytes will have to remain here.
1849	 * In time closedown will finish, we empty the write queue and
1850	 * all will be happy.
1851	 */
1852	if (unlikely(sk->sk_state == TCP_CLOSE))
1853		return;
1854
1855	if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1856		tcp_check_probe_timer(sk);
1857}
1858
1859/* Send _single_ skb sitting at the send head. This function requires
1860 * true push pending frames to setup probe timer etc.
1861 */
1862void tcp_push_one(struct sock *sk, unsigned int mss_now)
1863{
1864	struct sk_buff *skb = tcp_send_head(sk);
1865
1866	BUG_ON(!skb || skb->len < mss_now);
1867
1868	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1869}
1870
1871/* This function returns the amount that we can raise the
1872 * usable window based on the following constraints
1873 *
1874 * 1. The window can never be shrunk once it is offered (RFC 793)
1875 * 2. We limit memory per socket
1876 *
1877 * RFC 1122:
1878 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1879 *  RECV.NEXT + RCV.WIN fixed until:
1880 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1881 *
1882 * i.e. don't raise the right edge of the window until you can raise
1883 * it at least MSS bytes.
1884 *
1885 * Unfortunately, the recommended algorithm breaks header prediction,
1886 * since header prediction assumes th->window stays fixed.
1887 *
1888 * Strictly speaking, keeping th->window fixed violates the receiver
1889 * side SWS prevention criteria. The problem is that under this rule
1890 * a stream of single byte packets will cause the right side of the
1891 * window to always advance by a single byte.
1892 *
1893 * Of course, if the sender implements sender side SWS prevention
1894 * then this will not be a problem.
1895 *
1896 * BSD seems to make the following compromise:
1897 *
1898 *	If the free space is less than the 1/4 of the maximum
1899 *	space available and the free space is less than 1/2 mss,
1900 *	then set the window to 0.
1901 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1902 *	Otherwise, just prevent the window from shrinking
1903 *	and from being larger than the largest representable value.
1904 *
1905 * This prevents incremental opening of the window in the regime
1906 * where TCP is limited by the speed of the reader side taking
1907 * data out of the TCP receive queue. It does nothing about
1908 * those cases where the window is constrained on the sender side
1909 * because the pipeline is full.
1910 *
1911 * BSD also seems to "accidentally" limit itself to windows that are a
1912 * multiple of MSS, at least until the free space gets quite small.
1913 * This would appear to be a side effect of the mbuf implementation.
1914 * Combining these two algorithms results in the observed behavior
1915 * of having a fixed window size at almost all times.
1916 *
1917 * Below we obtain similar behavior by forcing the offered window to
1918 * a multiple of the mss when it is feasible to do so.
1919 *
1920 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1921 * Regular options like TIMESTAMP are taken into account.
1922 */
1923u32 __tcp_select_window(struct sock *sk)
1924{
1925	struct inet_connection_sock *icsk = inet_csk(sk);
1926	struct tcp_sock *tp = tcp_sk(sk);
1927	/* MSS for the peer's data.  Previous versions used mss_clamp
1928	 * here.  I don't know if the value based on our guesses
1929	 * of peer's MSS is better for the performance.  It's more correct
1930	 * but may be worse for the performance because of rcv_mss
1931	 * fluctuations.  --SAW  1998/11/1
1932	 */
1933	int mss = icsk->icsk_ack.rcv_mss;
1934	int free_space = tcp_space(sk);
1935	int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1936	int window;
1937
1938	if (mss > full_space)
1939		mss = full_space;
1940
1941	if (free_space < (full_space >> 1)) {
1942		icsk->icsk_ack.quick = 0;
1943
1944		if (sk_under_memory_pressure(sk))
1945			tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1946					       4U * tp->advmss);
1947
1948		if (free_space < mss)
1949			return 0;
1950	}
1951
1952	if (free_space > tp->rcv_ssthresh)
1953		free_space = tp->rcv_ssthresh;
1954
1955	/* Don't do rounding if we are using window scaling, since the
1956	 * scaled window will not line up with the MSS boundary anyway.
1957	 */
1958	window = tp->rcv_wnd;
1959	if (tp->rx_opt.rcv_wscale) {
1960		window = free_space;
1961
1962		/* Advertise enough space so that it won't get scaled away.
1963		 * Import case: prevent zero window announcement if
1964		 * 1<<rcv_wscale > mss.
1965		 */
1966		if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1967			window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1968				  << tp->rx_opt.rcv_wscale);
1969	} else {
1970		/* Get the largest window that is a nice multiple of mss.
1971		 * Window clamp already applied above.
1972		 * If our current window offering is within 1 mss of the
1973		 * free space we just keep it. This prevents the divide
1974		 * and multiply from happening most of the time.
1975		 * We also don't do any window rounding when the free space
1976		 * is too small.
1977		 */
1978		if (window <= free_space - mss || window > free_space)
1979			window = (free_space / mss) * mss;
1980		else if (mss == full_space &&
1981			 free_space > window + (full_space >> 1))
1982			window = free_space;
1983	}
1984
1985	return window;
1986}
1987
1988/* Collapses two adjacent SKB's during retransmission. */
1989static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1990{
1991	struct tcp_sock *tp = tcp_sk(sk);
1992	struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1993	int skb_size, next_skb_size;
1994
1995	skb_size = skb->len;
1996	next_skb_size = next_skb->len;
1997
1998	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1999
2000	tcp_highest_sack_combine(sk, next_skb, skb);
2001
2002	tcp_unlink_write_queue(next_skb, sk);
2003
2004	skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2005				  next_skb_size);
2006
2007	if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2008		skb->ip_summed = CHECKSUM_PARTIAL;
2009
2010	if (skb->ip_summed != CHECKSUM_PARTIAL)
2011		skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2012
2013	/* Update sequence range on original skb. */
2014	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2015
2016	/* Merge over control information. This moves PSH/FIN etc. over */
2017	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2018
2019	/* All done, get rid of second SKB and account for it so
2020	 * packet counting does not break.
2021	 */
2022	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2023
2024	/* changed transmit queue under us so clear hints */
2025	tcp_clear_retrans_hints_partial(tp);
2026	if (next_skb == tp->retransmit_skb_hint)
2027		tp->retransmit_skb_hint = skb;
2028
2029	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2030
2031	sk_wmem_free_skb(sk, next_skb);
2032}
2033
2034/* Check if coalescing SKBs is legal. */
2035static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2036{
2037	if (tcp_skb_pcount(skb) > 1)
2038		return false;
2039	/* TODO: SACK collapsing could be used to remove this condition */
2040	if (skb_shinfo(skb)->nr_frags != 0)
2041		return false;
2042	if (skb_cloned(skb))
2043		return false;
2044	if (skb == tcp_send_head(sk))
2045		return false;
2046	/* Some heurestics for collapsing over SACK'd could be invented */
2047	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2048		return false;
2049
2050	return true;
2051}
2052
2053/* Collapse packets in the retransmit queue to make to create
2054 * less packets on the wire. This is only done on retransmission.
2055 */
2056static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2057				     int space)
2058{
2059	struct tcp_sock *tp = tcp_sk(sk);
2060	struct sk_buff *skb = to, *tmp;
2061	bool first = true;
2062
2063	if (!sysctl_tcp_retrans_collapse)
2064		return;
2065	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2066		return;
2067
2068	tcp_for_write_queue_from_safe(skb, tmp, sk) {
2069		if (!tcp_can_collapse(sk, skb))
2070			break;
2071
2072		space -= skb->len;
2073
2074		if (first) {
2075			first = false;
2076			continue;
2077		}
2078
2079		if (space < 0)
2080			break;
2081		/* Punt if not enough space exists in the first SKB for
2082		 * the data in the second
2083		 */
2084		if (skb->len > skb_availroom(to))
2085			break;
2086
2087		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2088			break;
2089
2090		tcp_collapse_retrans(sk, to);
2091	}
2092}
2093
2094/* This retransmits one SKB.  Policy decisions and retransmit queue
2095 * state updates are done by the caller.  Returns non-zero if an
2096 * error occurred which prevented the send.
2097 */
2098int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2099{
2100	struct tcp_sock *tp = tcp_sk(sk);
2101	struct inet_connection_sock *icsk = inet_csk(sk);
2102	unsigned int cur_mss;
2103	int err;
2104
2105	/* Inconslusive MTU probe */
2106	if (icsk->icsk_mtup.probe_size) {
2107		icsk->icsk_mtup.probe_size = 0;
2108	}
2109
2110	/* Do not sent more than we queued. 1/4 is reserved for possible
2111	 * copying overhead: fragmentation, tunneling, mangling etc.
2112	 */
2113	if (atomic_read(&sk->sk_wmem_alloc) >
2114	    min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2115		return -EAGAIN;
2116
2117	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2118		if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2119			BUG();
2120		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2121			return -ENOMEM;
2122	}
2123
2124	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2125		return -EHOSTUNREACH; /* Routing failure or similar. */
2126
2127	cur_mss = tcp_current_mss(sk);
2128
2129	/* If receiver has shrunk his window, and skb is out of
2130	 * new window, do not retransmit it. The exception is the
2131	 * case, when window is shrunk to zero. In this case
2132	 * our retransmit serves as a zero window probe.
2133	 */
2134	if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2135	    TCP_SKB_CB(skb)->seq != tp->snd_una)
2136		return -EAGAIN;
2137
2138	if (skb->len > cur_mss) {
2139		if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2140			return -ENOMEM; /* We'll try again later. */
2141	} else {
2142		int oldpcount = tcp_skb_pcount(skb);
2143
2144		if (unlikely(oldpcount > 1)) {
2145			tcp_init_tso_segs(sk, skb, cur_mss);
2146			tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2147		}
2148	}
2149
2150	tcp_retrans_try_collapse(sk, skb, cur_mss);
2151
2152	/* Some Solaris stacks overoptimize and ignore the FIN on a
2153	 * retransmit when old data is attached.  So strip it off
2154	 * since it is cheap to do so and saves bytes on the network.
2155	 */
2156	if (skb->len > 0 &&
2157	    (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2158	    tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2159		if (!pskb_trim(skb, 0)) {
2160			/* Reuse, even though it does some unnecessary work */
2161			tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2162					     TCP_SKB_CB(skb)->tcp_flags);
2163			skb->ip_summed = CHECKSUM_NONE;
2164		}
2165	}
2166
2167	/* Make a copy, if the first transmission SKB clone we made
2168	 * is still in somebody's hands, else make a clone.
2169	 */
2170	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2171
2172	/* make sure skb->data is aligned on arches that require it */
2173	if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2174		struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2175						   GFP_ATOMIC);
2176		err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2177			     -ENOBUFS;
2178	} else {
2179		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2180	}
2181
2182	if (err == 0) {
2183		/* Update global TCP statistics. */
2184		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2185
2186		tp->total_retrans++;
2187
2188#if FASTRETRANS_DEBUG > 0
2189		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2190			net_dbg_ratelimited("retrans_out leaked\n");
 
2191		}
2192#endif
2193		if (!tp->retrans_out)
2194			tp->lost_retrans_low = tp->snd_nxt;
2195		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2196		tp->retrans_out += tcp_skb_pcount(skb);
2197
2198		/* Save stamp of the first retransmit. */
2199		if (!tp->retrans_stamp)
2200			tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2201
2202		tp->undo_retrans += tcp_skb_pcount(skb);
2203
2204		/* snd_nxt is stored to detect loss of retransmitted segment,
2205		 * see tcp_input.c tcp_sacktag_write_queue().
2206		 */
2207		TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2208	}
2209	return err;
2210}
2211
2212/* Check if we forward retransmits are possible in the current
2213 * window/congestion state.
2214 */
2215static bool tcp_can_forward_retransmit(struct sock *sk)
2216{
2217	const struct inet_connection_sock *icsk = inet_csk(sk);
2218	const struct tcp_sock *tp = tcp_sk(sk);
2219
2220	/* Forward retransmissions are possible only during Recovery. */
2221	if (icsk->icsk_ca_state != TCP_CA_Recovery)
2222		return false;
2223
2224	/* No forward retransmissions in Reno are possible. */
2225	if (tcp_is_reno(tp))
2226		return false;
2227
2228	/* Yeah, we have to make difficult choice between forward transmission
2229	 * and retransmission... Both ways have their merits...
2230	 *
2231	 * For now we do not retransmit anything, while we have some new
2232	 * segments to send. In the other cases, follow rule 3 for
2233	 * NextSeg() specified in RFC3517.
2234	 */
2235
2236	if (tcp_may_send_now(sk))
2237		return false;
2238
2239	return true;
2240}
2241
2242/* This gets called after a retransmit timeout, and the initially
2243 * retransmitted data is acknowledged.  It tries to continue
2244 * resending the rest of the retransmit queue, until either
2245 * we've sent it all or the congestion window limit is reached.
2246 * If doing SACK, the first ACK which comes back for a timeout
2247 * based retransmit packet might feed us FACK information again.
2248 * If so, we use it to avoid unnecessarily retransmissions.
2249 */
2250void tcp_xmit_retransmit_queue(struct sock *sk)
2251{
2252	const struct inet_connection_sock *icsk = inet_csk(sk);
2253	struct tcp_sock *tp = tcp_sk(sk);
2254	struct sk_buff *skb;
2255	struct sk_buff *hole = NULL;
2256	u32 last_lost;
2257	int mib_idx;
2258	int fwd_rexmitting = 0;
2259
2260	if (!tp->packets_out)
2261		return;
2262
2263	if (!tp->lost_out)
2264		tp->retransmit_high = tp->snd_una;
2265
2266	if (tp->retransmit_skb_hint) {
2267		skb = tp->retransmit_skb_hint;
2268		last_lost = TCP_SKB_CB(skb)->end_seq;
2269		if (after(last_lost, tp->retransmit_high))
2270			last_lost = tp->retransmit_high;
2271	} else {
2272		skb = tcp_write_queue_head(sk);
2273		last_lost = tp->snd_una;
2274	}
2275
2276	tcp_for_write_queue_from(skb, sk) {
2277		__u8 sacked = TCP_SKB_CB(skb)->sacked;
2278
2279		if (skb == tcp_send_head(sk))
2280			break;
2281		/* we could do better than to assign each time */
2282		if (hole == NULL)
2283			tp->retransmit_skb_hint = skb;
2284
2285		/* Assume this retransmit will generate
2286		 * only one packet for congestion window
2287		 * calculation purposes.  This works because
2288		 * tcp_retransmit_skb() will chop up the
2289		 * packet to be MSS sized and all the
2290		 * packet counting works out.
2291		 */
2292		if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2293			return;
2294
2295		if (fwd_rexmitting) {
2296begin_fwd:
2297			if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2298				break;
2299			mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2300
2301		} else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2302			tp->retransmit_high = last_lost;
2303			if (!tcp_can_forward_retransmit(sk))
2304				break;
2305			/* Backtrack if necessary to non-L'ed skb */
2306			if (hole != NULL) {
2307				skb = hole;
2308				hole = NULL;
2309			}
2310			fwd_rexmitting = 1;
2311			goto begin_fwd;
2312
2313		} else if (!(sacked & TCPCB_LOST)) {
2314			if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2315				hole = skb;
2316			continue;
2317
2318		} else {
2319			last_lost = TCP_SKB_CB(skb)->end_seq;
2320			if (icsk->icsk_ca_state != TCP_CA_Loss)
2321				mib_idx = LINUX_MIB_TCPFASTRETRANS;
2322			else
2323				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2324		}
2325
2326		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2327			continue;
2328
2329		if (tcp_retransmit_skb(sk, skb)) {
2330			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2331			return;
2332		}
2333		NET_INC_STATS_BH(sock_net(sk), mib_idx);
2334
2335		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2336			tp->prr_out += tcp_skb_pcount(skb);
2337
2338		if (skb == tcp_write_queue_head(sk))
2339			inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2340						  inet_csk(sk)->icsk_rto,
2341						  TCP_RTO_MAX);
2342	}
2343}
2344
2345/* Send a fin.  The caller locks the socket for us.  This cannot be
2346 * allowed to fail queueing a FIN frame under any circumstances.
2347 */
2348void tcp_send_fin(struct sock *sk)
2349{
2350	struct tcp_sock *tp = tcp_sk(sk);
2351	struct sk_buff *skb = tcp_write_queue_tail(sk);
2352	int mss_now;
2353
2354	/* Optimization, tack on the FIN if we have a queue of
2355	 * unsent frames.  But be careful about outgoing SACKS
2356	 * and IP options.
2357	 */
2358	mss_now = tcp_current_mss(sk);
2359
2360	if (tcp_send_head(sk) != NULL) {
2361		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2362		TCP_SKB_CB(skb)->end_seq++;
2363		tp->write_seq++;
2364	} else {
2365		/* Socket is locked, keep trying until memory is available. */
2366		for (;;) {
2367			skb = alloc_skb_fclone(MAX_TCP_HEADER,
2368					       sk->sk_allocation);
2369			if (skb)
2370				break;
2371			yield();
2372		}
2373
2374		/* Reserve space for headers and prepare control bits. */
2375		skb_reserve(skb, MAX_TCP_HEADER);
2376		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2377		tcp_init_nondata_skb(skb, tp->write_seq,
2378				     TCPHDR_ACK | TCPHDR_FIN);
2379		tcp_queue_skb(sk, skb);
2380	}
2381	__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2382}
2383
2384/* We get here when a process closes a file descriptor (either due to
2385 * an explicit close() or as a byproduct of exit()'ing) and there
2386 * was unread data in the receive queue.  This behavior is recommended
2387 * by RFC 2525, section 2.17.  -DaveM
2388 */
2389void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2390{
2391	struct sk_buff *skb;
2392
2393	/* NOTE: No TCP options attached and we never retransmit this. */
2394	skb = alloc_skb(MAX_TCP_HEADER, priority);
2395	if (!skb) {
2396		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2397		return;
2398	}
2399
2400	/* Reserve space for headers and prepare control bits. */
2401	skb_reserve(skb, MAX_TCP_HEADER);
2402	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2403			     TCPHDR_ACK | TCPHDR_RST);
2404	/* Send it off. */
2405	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2406	if (tcp_transmit_skb(sk, skb, 0, priority))
2407		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2408
2409	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2410}
2411
2412/* Send a crossed SYN-ACK during socket establishment.
2413 * WARNING: This routine must only be called when we have already sent
2414 * a SYN packet that crossed the incoming SYN that caused this routine
2415 * to get called. If this assumption fails then the initial rcv_wnd
2416 * and rcv_wscale values will not be correct.
2417 */
2418int tcp_send_synack(struct sock *sk)
2419{
2420	struct sk_buff *skb;
2421
2422	skb = tcp_write_queue_head(sk);
2423	if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2424		pr_debug("%s: wrong queue state\n", __func__);
2425		return -EFAULT;
2426	}
2427	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2428		if (skb_cloned(skb)) {
2429			struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2430			if (nskb == NULL)
2431				return -ENOMEM;
2432			tcp_unlink_write_queue(skb, sk);
2433			skb_header_release(nskb);
2434			__tcp_add_write_queue_head(sk, nskb);
2435			sk_wmem_free_skb(sk, skb);
2436			sk->sk_wmem_queued += nskb->truesize;
2437			sk_mem_charge(sk, nskb->truesize);
2438			skb = nskb;
2439		}
2440
2441		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2442		TCP_ECN_send_synack(tcp_sk(sk), skb);
2443	}
2444	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2445	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2446}
2447
2448/* Prepare a SYN-ACK. */
2449struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2450				struct request_sock *req,
2451				struct request_values *rvp)
2452{
2453	struct tcp_out_options opts;
2454	struct tcp_extend_values *xvp = tcp_xv(rvp);
2455	struct inet_request_sock *ireq = inet_rsk(req);
2456	struct tcp_sock *tp = tcp_sk(sk);
2457	const struct tcp_cookie_values *cvp = tp->cookie_values;
2458	struct tcphdr *th;
2459	struct sk_buff *skb;
2460	struct tcp_md5sig_key *md5;
2461	int tcp_header_size;
2462	int mss;
2463	int s_data_desired = 0;
2464
2465	if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2466		s_data_desired = cvp->s_data_desired;
2467	skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2468	if (skb == NULL)
2469		return NULL;
2470
2471	/* Reserve space for headers. */
2472	skb_reserve(skb, MAX_TCP_HEADER);
2473
2474	skb_dst_set(skb, dst_clone(dst));
2475
2476	mss = dst_metric_advmss(dst);
2477	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2478		mss = tp->rx_opt.user_mss;
2479
2480	if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2481		__u8 rcv_wscale;
2482		/* Set this up on the first call only */
2483		req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2484
2485		/* limit the window selection if the user enforce a smaller rx buffer */
2486		if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2487		    (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2488			req->window_clamp = tcp_full_space(sk);
2489
2490		/* tcp_full_space because it is guaranteed to be the first packet */
2491		tcp_select_initial_window(tcp_full_space(sk),
2492			mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2493			&req->rcv_wnd,
2494			&req->window_clamp,
2495			ireq->wscale_ok,
2496			&rcv_wscale,
2497			dst_metric(dst, RTAX_INITRWND));
2498		ireq->rcv_wscale = rcv_wscale;
2499	}
2500
2501	memset(&opts, 0, sizeof(opts));
2502#ifdef CONFIG_SYN_COOKIES
2503	if (unlikely(req->cookie_ts))
2504		TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2505	else
2506#endif
2507	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2508	tcp_header_size = tcp_synack_options(sk, req, mss,
2509					     skb, &opts, &md5, xvp)
2510			+ sizeof(*th);
2511
2512	skb_push(skb, tcp_header_size);
2513	skb_reset_transport_header(skb);
2514
2515	th = tcp_hdr(skb);
2516	memset(th, 0, sizeof(struct tcphdr));
2517	th->syn = 1;
2518	th->ack = 1;
2519	TCP_ECN_make_synack(req, th);
2520	th->source = ireq->loc_port;
2521	th->dest = ireq->rmt_port;
2522	/* Setting of flags are superfluous here for callers (and ECE is
2523	 * not even correctly set)
2524	 */
2525	tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2526			     TCPHDR_SYN | TCPHDR_ACK);
2527
2528	if (OPTION_COOKIE_EXTENSION & opts.options) {
2529		if (s_data_desired) {
2530			u8 *buf = skb_put(skb, s_data_desired);
2531
2532			/* copy data directly from the listening socket. */
2533			memcpy(buf, cvp->s_data_payload, s_data_desired);
2534			TCP_SKB_CB(skb)->end_seq += s_data_desired;
2535		}
2536
2537		if (opts.hash_size > 0) {
2538			__u32 workspace[SHA_WORKSPACE_WORDS];
2539			u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2540			u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2541
2542			/* Secret recipe depends on the Timestamp, (future)
2543			 * Sequence and Acknowledgment Numbers, Initiator
2544			 * Cookie, and others handled by IP variant caller.
2545			 */
2546			*tail-- ^= opts.tsval;
2547			*tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2548			*tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2549
2550			/* recommended */
2551			*tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2552			*tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2553
2554			sha_transform((__u32 *)&xvp->cookie_bakery[0],
2555				      (char *)mess,
2556				      &workspace[0]);
2557			opts.hash_location =
2558				(__u8 *)&xvp->cookie_bakery[0];
2559		}
2560	}
2561
2562	th->seq = htonl(TCP_SKB_CB(skb)->seq);
2563	th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2564
2565	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2566	th->window = htons(min(req->rcv_wnd, 65535U));
2567	tcp_options_write((__be32 *)(th + 1), tp, &opts);
2568	th->doff = (tcp_header_size >> 2);
2569	TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2570
2571#ifdef CONFIG_TCP_MD5SIG
2572	/* Okay, we have all we need - do the md5 hash if needed */
2573	if (md5) {
2574		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2575					       md5, NULL, req, skb);
2576	}
2577#endif
2578
2579	return skb;
2580}
2581EXPORT_SYMBOL(tcp_make_synack);
2582
2583/* Do all connect socket setups that can be done AF independent. */
2584void tcp_connect_init(struct sock *sk)
2585{
2586	const struct dst_entry *dst = __sk_dst_get(sk);
2587	struct tcp_sock *tp = tcp_sk(sk);
2588	__u8 rcv_wscale;
2589
2590	/* We'll fix this up when we get a response from the other end.
2591	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2592	 */
2593	tp->tcp_header_len = sizeof(struct tcphdr) +
2594		(sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2595
2596#ifdef CONFIG_TCP_MD5SIG
2597	if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2598		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2599#endif
2600
2601	/* If user gave his TCP_MAXSEG, record it to clamp */
2602	if (tp->rx_opt.user_mss)
2603		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2604	tp->max_window = 0;
2605	tcp_mtup_init(sk);
2606	tcp_sync_mss(sk, dst_mtu(dst));
2607
2608	if (!tp->window_clamp)
2609		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2610	tp->advmss = dst_metric_advmss(dst);
2611	if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2612		tp->advmss = tp->rx_opt.user_mss;
2613
2614	tcp_initialize_rcv_mss(sk);
2615
2616	/* limit the window selection if the user enforce a smaller rx buffer */
2617	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2618	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2619		tp->window_clamp = tcp_full_space(sk);
2620
2621	tcp_select_initial_window(tcp_full_space(sk),
2622				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2623				  &tp->rcv_wnd,
2624				  &tp->window_clamp,
2625				  sysctl_tcp_window_scaling,
2626				  &rcv_wscale,
2627				  dst_metric(dst, RTAX_INITRWND));
2628
2629	tp->rx_opt.rcv_wscale = rcv_wscale;
2630	tp->rcv_ssthresh = tp->rcv_wnd;
2631
2632	sk->sk_err = 0;
2633	sock_reset_flag(sk, SOCK_DONE);
2634	tp->snd_wnd = 0;
2635	tcp_init_wl(tp, 0);
2636	tp->snd_una = tp->write_seq;
2637	tp->snd_sml = tp->write_seq;
2638	tp->snd_up = tp->write_seq;
2639	tp->snd_nxt = tp->write_seq;
2640
2641	if (likely(!tp->repair))
2642		tp->rcv_nxt = 0;
2643	tp->rcv_wup = tp->rcv_nxt;
2644	tp->copied_seq = tp->rcv_nxt;
2645
2646	inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2647	inet_csk(sk)->icsk_retransmits = 0;
2648	tcp_clear_retrans(tp);
2649}
2650
2651/* Build a SYN and send it off. */
2652int tcp_connect(struct sock *sk)
2653{
2654	struct tcp_sock *tp = tcp_sk(sk);
2655	struct sk_buff *buff;
2656	int err;
2657
2658	tcp_connect_init(sk);
2659
2660	buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2661	if (unlikely(buff == NULL))
2662		return -ENOBUFS;
2663
2664	/* Reserve space for headers. */
2665	skb_reserve(buff, MAX_TCP_HEADER);
2666
 
2667	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2668	TCP_ECN_send_syn(sk, buff);
2669
2670	/* Send it off. */
2671	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2672	tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2673	skb_header_release(buff);
2674	__tcp_add_write_queue_tail(sk, buff);
2675	sk->sk_wmem_queued += buff->truesize;
2676	sk_mem_charge(sk, buff->truesize);
2677	tp->packets_out += tcp_skb_pcount(buff);
2678	err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2679	if (err == -ECONNREFUSED)
2680		return err;
2681
2682	/* We change tp->snd_nxt after the tcp_transmit_skb() call
2683	 * in order to make this packet get counted in tcpOutSegs.
2684	 */
2685	tp->snd_nxt = tp->write_seq;
2686	tp->pushed_seq = tp->write_seq;
2687	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2688
2689	/* Timer for repeating the SYN until an answer. */
2690	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2691				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2692	return 0;
2693}
2694EXPORT_SYMBOL(tcp_connect);
2695
2696/* Send out a delayed ack, the caller does the policy checking
2697 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
2698 * for details.
2699 */
2700void tcp_send_delayed_ack(struct sock *sk)
2701{
2702	struct inet_connection_sock *icsk = inet_csk(sk);
2703	int ato = icsk->icsk_ack.ato;
2704	unsigned long timeout;
2705
2706	if (ato > TCP_DELACK_MIN) {
2707		const struct tcp_sock *tp = tcp_sk(sk);
2708		int max_ato = HZ / 2;
2709
2710		if (icsk->icsk_ack.pingpong ||
2711		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2712			max_ato = TCP_DELACK_MAX;
2713
2714		/* Slow path, intersegment interval is "high". */
2715
2716		/* If some rtt estimate is known, use it to bound delayed ack.
2717		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2718		 * directly.
2719		 */
2720		if (tp->srtt) {
2721			int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2722
2723			if (rtt < max_ato)
2724				max_ato = rtt;
2725		}
2726
2727		ato = min(ato, max_ato);
2728	}
2729
2730	/* Stay within the limit we were given */
2731	timeout = jiffies + ato;
2732
2733	/* Use new timeout only if there wasn't a older one earlier. */
2734	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2735		/* If delack timer was blocked or is about to expire,
2736		 * send ACK now.
2737		 */
2738		if (icsk->icsk_ack.blocked ||
2739		    time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2740			tcp_send_ack(sk);
2741			return;
2742		}
2743
2744		if (!time_before(timeout, icsk->icsk_ack.timeout))
2745			timeout = icsk->icsk_ack.timeout;
2746	}
2747	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2748	icsk->icsk_ack.timeout = timeout;
2749	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2750}
2751
2752/* This routine sends an ack and also updates the window. */
2753void tcp_send_ack(struct sock *sk)
2754{
2755	struct sk_buff *buff;
2756
2757	/* If we have been reset, we may not send again. */
2758	if (sk->sk_state == TCP_CLOSE)
2759		return;
2760
2761	/* We are not putting this on the write queue, so
2762	 * tcp_transmit_skb() will set the ownership to this
2763	 * sock.
2764	 */
2765	buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2766	if (buff == NULL) {
2767		inet_csk_schedule_ack(sk);
2768		inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2769		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2770					  TCP_DELACK_MAX, TCP_RTO_MAX);
2771		return;
2772	}
2773
2774	/* Reserve space for headers and prepare control bits. */
2775	skb_reserve(buff, MAX_TCP_HEADER);
2776	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2777
2778	/* Send it off, this clears delayed acks for us. */
2779	TCP_SKB_CB(buff)->when = tcp_time_stamp;
2780	tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2781}
2782
2783/* This routine sends a packet with an out of date sequence
2784 * number. It assumes the other end will try to ack it.
2785 *
2786 * Question: what should we make while urgent mode?
2787 * 4.4BSD forces sending single byte of data. We cannot send
2788 * out of window data, because we have SND.NXT==SND.MAX...
2789 *
2790 * Current solution: to send TWO zero-length segments in urgent mode:
2791 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2792 * out-of-date with SND.UNA-1 to probe window.
2793 */
2794static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2795{
2796	struct tcp_sock *tp = tcp_sk(sk);
2797	struct sk_buff *skb;
2798
2799	/* We don't queue it, tcp_transmit_skb() sets ownership. */
2800	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2801	if (skb == NULL)
2802		return -1;
2803
2804	/* Reserve space for headers and set control bits. */
2805	skb_reserve(skb, MAX_TCP_HEADER);
2806	/* Use a previous sequence.  This should cause the other
2807	 * end to send an ack.  Don't queue or clone SKB, just
2808	 * send it.
2809	 */
2810	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2811	TCP_SKB_CB(skb)->when = tcp_time_stamp;
2812	return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2813}
2814
2815void tcp_send_window_probe(struct sock *sk)
2816{
2817	if (sk->sk_state == TCP_ESTABLISHED) {
2818		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
2819		tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
2820		tcp_xmit_probe_skb(sk, 0);
2821	}
2822}
2823
2824/* Initiate keepalive or window probe from timer. */
2825int tcp_write_wakeup(struct sock *sk)
2826{
2827	struct tcp_sock *tp = tcp_sk(sk);
2828	struct sk_buff *skb;
2829
2830	if (sk->sk_state == TCP_CLOSE)
2831		return -1;
2832
2833	if ((skb = tcp_send_head(sk)) != NULL &&
2834	    before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2835		int err;
2836		unsigned int mss = tcp_current_mss(sk);
2837		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2838
2839		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2840			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2841
2842		/* We are probing the opening of a window
2843		 * but the window size is != 0
2844		 * must have been a result SWS avoidance ( sender )
2845		 */
2846		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2847		    skb->len > mss) {
2848			seg_size = min(seg_size, mss);
2849			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2850			if (tcp_fragment(sk, skb, seg_size, mss))
2851				return -1;
2852		} else if (!tcp_skb_pcount(skb))
2853			tcp_set_skb_tso_segs(sk, skb, mss);
2854
2855		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2856		TCP_SKB_CB(skb)->when = tcp_time_stamp;
2857		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2858		if (!err)
2859			tcp_event_new_data_sent(sk, skb);
2860		return err;
2861	} else {
2862		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2863			tcp_xmit_probe_skb(sk, 1);
2864		return tcp_xmit_probe_skb(sk, 0);
2865	}
2866}
2867
2868/* A window probe timeout has occurred.  If window is not closed send
2869 * a partial packet else a zero probe.
2870 */
2871void tcp_send_probe0(struct sock *sk)
2872{
2873	struct inet_connection_sock *icsk = inet_csk(sk);
2874	struct tcp_sock *tp = tcp_sk(sk);
2875	int err;
2876
2877	err = tcp_write_wakeup(sk);
2878
2879	if (tp->packets_out || !tcp_send_head(sk)) {
2880		/* Cancel probe timer, if it is not required. */
2881		icsk->icsk_probes_out = 0;
2882		icsk->icsk_backoff = 0;
2883		return;
2884	}
2885
2886	if (err <= 0) {
2887		if (icsk->icsk_backoff < sysctl_tcp_retries2)
2888			icsk->icsk_backoff++;
2889		icsk->icsk_probes_out++;
2890		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2891					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2892					  TCP_RTO_MAX);
2893	} else {
2894		/* If packet was not sent due to local congestion,
2895		 * do not backoff and do not remember icsk_probes_out.
2896		 * Let local senders to fight for local resources.
2897		 *
2898		 * Use accumulated backoff yet.
2899		 */
2900		if (!icsk->icsk_probes_out)
2901			icsk->icsk_probes_out = 1;
2902		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2903					  min(icsk->icsk_rto << icsk->icsk_backoff,
2904					      TCP_RESOURCE_PROBE_INTERVAL),
2905					  TCP_RTO_MAX);
2906	}
2907}