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:
  23 *		Pedro Roque	:	Fast Retransmit/Recovery.
  24 *					Two receive queues.
  25 *					Retransmit queue handled by TCP.
  26 *					Better retransmit timer handling.
  27 *					New congestion avoidance.
  28 *					Header prediction.
  29 *					Variable renaming.
  30 *
  31 *		Eric		:	Fast Retransmit.
  32 *		Randy Scott	:	MSS option defines.
  33 *		Eric Schenk	:	Fixes to slow start algorithm.
  34 *		Eric Schenk	:	Yet another double ACK bug.
  35 *		Eric Schenk	:	Delayed ACK bug fixes.
  36 *		Eric Schenk	:	Floyd style fast retrans war avoidance.
  37 *		David S. Miller	:	Don't allow zero congestion window.
  38 *		Eric Schenk	:	Fix retransmitter so that it sends
  39 *					next packet on ack of previous packet.
  40 *		Andi Kleen	:	Moved open_request checking here
  41 *					and process RSTs for open_requests.
  42 *		Andi Kleen	:	Better prune_queue, and other fixes.
  43 *		Andrey Savochkin:	Fix RTT measurements in the presence of
  44 *					timestamps.
  45 *		Andrey Savochkin:	Check sequence numbers correctly when
  46 *					removing SACKs due to in sequence incoming
  47 *					data segments.
  48 *		Andi Kleen:		Make sure we never ack data there is not
  49 *					enough room for. Also make this condition
  50 *					a fatal error if it might still happen.
  51 *		Andi Kleen:		Add tcp_measure_rcv_mss to make
  52 *					connections with MSS<min(MTU,ann. MSS)
  53 *					work without delayed acks.
  54 *		Andi Kleen:		Process packets with PSH set in the
  55 *					fast path.
  56 *		J Hadi Salim:		ECN support
  57 *	 	Andrei Gurtov,
  58 *		Pasi Sarolahti,
  59 *		Panu Kuhlberg:		Experimental audit of TCP (re)transmission
  60 *					engine. Lots of bugs are found.
  61 *		Pasi Sarolahti:		F-RTO for dealing with spurious RTOs
  62 */
  63
  64#define pr_fmt(fmt) "TCP: " fmt
  65
  66#include <linux/mm.h>
  67#include <linux/slab.h>
  68#include <linux/module.h>
  69#include <linux/sysctl.h>
  70#include <linux/kernel.h>
 
  71#include <net/dst.h>
  72#include <net/tcp.h>
 
  73#include <net/inet_common.h>
  74#include <linux/ipsec.h>
  75#include <asm/unaligned.h>
  76#include <net/netdma.h>
 
 
 
 
  77
  78int sysctl_tcp_timestamps __read_mostly = 1;
  79int sysctl_tcp_window_scaling __read_mostly = 1;
  80int sysctl_tcp_sack __read_mostly = 1;
  81int sysctl_tcp_fack __read_mostly = 1;
  82int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH;
  83EXPORT_SYMBOL(sysctl_tcp_reordering);
  84int sysctl_tcp_dsack __read_mostly = 1;
  85int sysctl_tcp_app_win __read_mostly = 31;
  86int sysctl_tcp_adv_win_scale __read_mostly = 1;
  87EXPORT_SYMBOL(sysctl_tcp_adv_win_scale);
  88
  89/* rfc5961 challenge ack rate limiting */
  90int sysctl_tcp_challenge_ack_limit = 100;
  91
  92int sysctl_tcp_stdurg __read_mostly;
  93int sysctl_tcp_rfc1337 __read_mostly;
  94int sysctl_tcp_max_orphans __read_mostly = NR_FILE;
  95int sysctl_tcp_frto __read_mostly = 2;
  96
  97int sysctl_tcp_thin_dupack __read_mostly;
  98
  99int sysctl_tcp_moderate_rcvbuf __read_mostly = 1;
 100int sysctl_tcp_early_retrans __read_mostly = 3;
 101
 102#define FLAG_DATA		0x01 /* Incoming frame contained data.		*/
 103#define FLAG_WIN_UPDATE		0x02 /* Incoming ACK was a window update.	*/
 104#define FLAG_DATA_ACKED		0x04 /* This ACK acknowledged new data.		*/
 105#define FLAG_RETRANS_DATA_ACKED	0x08 /* "" "" some of which was retransmitted.	*/
 106#define FLAG_SYN_ACKED		0x10 /* This ACK acknowledged SYN.		*/
 107#define FLAG_DATA_SACKED	0x20 /* New SACK.				*/
 108#define FLAG_ECE		0x40 /* ECE in this ACK				*/
 
 109#define FLAG_SLOWPATH		0x100 /* Do not skip RFC checks for window update.*/
 110#define FLAG_ORIG_SACK_ACKED	0x200 /* Never retransmitted data are (s)acked	*/
 111#define FLAG_SND_UNA_ADVANCED	0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */
 112#define FLAG_DSACKING_ACK	0x800 /* SACK blocks contained D-SACK info */
 
 113#define FLAG_SACK_RENEGING	0x2000 /* snd_una advanced to a sacked seq */
 114#define FLAG_UPDATE_TS_RECENT	0x4000 /* tcp_replace_ts_recent() */
 
 
 
 115
 116#define FLAG_ACKED		(FLAG_DATA_ACKED|FLAG_SYN_ACKED)
 117#define FLAG_NOT_DUP		(FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED)
 118#define FLAG_CA_ALERT		(FLAG_DATA_SACKED|FLAG_ECE)
 119#define FLAG_FORWARD_PROGRESS	(FLAG_ACKED|FLAG_DATA_SACKED)
 120
 121#define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH)
 122#define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH))
 123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 124/* Adapt the MSS value used to make delayed ack decision to the
 125 * real world.
 126 */
 127static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb)
 128{
 129	struct inet_connection_sock *icsk = inet_csk(sk);
 130	const unsigned int lss = icsk->icsk_ack.last_seg_size;
 131	unsigned int len;
 132
 133	icsk->icsk_ack.last_seg_size = 0;
 134
 135	/* skb->len may jitter because of SACKs, even if peer
 136	 * sends good full-sized frames.
 137	 */
 138	len = skb_shinfo(skb)->gso_size ? : skb->len;
 139	if (len >= icsk->icsk_ack.rcv_mss) {
 140		icsk->icsk_ack.rcv_mss = len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 141	} else {
 142		/* Otherwise, we make more careful check taking into account,
 143		 * that SACKs block is variable.
 144		 *
 145		 * "len" is invariant segment length, including TCP header.
 146		 */
 147		len += skb->data - skb_transport_header(skb);
 148		if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) ||
 149		    /* If PSH is not set, packet should be
 150		     * full sized, provided peer TCP is not badly broken.
 151		     * This observation (if it is correct 8)) allows
 152		     * to handle super-low mtu links fairly.
 153		     */
 154		    (len >= TCP_MIN_MSS + sizeof(struct tcphdr) &&
 155		     !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) {
 156			/* Subtract also invariant (if peer is RFC compliant),
 157			 * tcp header plus fixed timestamp option length.
 158			 * Resulting "len" is MSS free of SACK jitter.
 159			 */
 160			len -= tcp_sk(sk)->tcp_header_len;
 161			icsk->icsk_ack.last_seg_size = len;
 162			if (len == lss) {
 163				icsk->icsk_ack.rcv_mss = len;
 164				return;
 165			}
 166		}
 167		if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)
 168			icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2;
 169		icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
 170	}
 171}
 172
 173static void tcp_incr_quickack(struct sock *sk)
 174{
 175	struct inet_connection_sock *icsk = inet_csk(sk);
 176	unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss);
 177
 178	if (quickacks == 0)
 179		quickacks = 2;
 
 180	if (quickacks > icsk->icsk_ack.quick)
 181		icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS);
 182}
 183
 184static void tcp_enter_quickack_mode(struct sock *sk)
 185{
 186	struct inet_connection_sock *icsk = inet_csk(sk);
 187	tcp_incr_quickack(sk);
 188	icsk->icsk_ack.pingpong = 0;
 
 189	icsk->icsk_ack.ato = TCP_ATO_MIN;
 190}
 191
 192/* Send ACKs quickly, if "quick" count is not exhausted
 193 * and the session is not interactive.
 194 */
 195
 196static inline bool tcp_in_quickack_mode(const struct sock *sk)
 197{
 198	const struct inet_connection_sock *icsk = inet_csk(sk);
 
 199
 200	return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong;
 
 201}
 202
 203static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp)
 204{
 205	if (tp->ecn_flags & TCP_ECN_OK)
 206		tp->ecn_flags |= TCP_ECN_QUEUE_CWR;
 207}
 208
 209static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb)
 210{
 211	if (tcp_hdr(skb)->cwr)
 212		tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 
 
 
 
 
 
 
 
 213}
 214
 215static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp)
 216{
 217	tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR;
 218}
 219
 220static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb)
 221{
 222	if (!(tp->ecn_flags & TCP_ECN_OK))
 223		return;
 224
 225	switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) {
 226	case INET_ECN_NOT_ECT:
 227		/* Funny extension: if ECT is not set on a segment,
 228		 * and we already seen ECT on a previous segment,
 229		 * it is probably a retransmit.
 230		 */
 231		if (tp->ecn_flags & TCP_ECN_SEEN)
 232			tcp_enter_quickack_mode((struct sock *)tp);
 233		break;
 234	case INET_ECN_CE:
 
 
 
 235		if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) {
 236			/* Better not delay acks, sender can have a very low cwnd */
 237			tcp_enter_quickack_mode((struct sock *)tp);
 238			tp->ecn_flags |= TCP_ECN_DEMAND_CWR;
 239		}
 240		/* fallinto */
 
 241	default:
 
 
 242		tp->ecn_flags |= TCP_ECN_SEEN;
 
 243	}
 244}
 245
 246static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th)
 
 
 
 
 
 
 247{
 248	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr))
 249		tp->ecn_flags &= ~TCP_ECN_OK;
 250}
 251
 252static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th)
 253{
 254	if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr))
 255		tp->ecn_flags &= ~TCP_ECN_OK;
 256}
 257
 258static bool TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th)
 259{
 260	if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK))
 261		return true;
 262	return false;
 263}
 264
 265/* Buffer size and advertised window tuning.
 266 *
 267 * 1. Tuning sk->sk_sndbuf, when connection enters established state.
 268 */
 269
 270static void tcp_sndbuf_expand(struct sock *sk)
 271{
 272	const struct tcp_sock *tp = tcp_sk(sk);
 
 273	int sndmem, per_mss;
 274	u32 nr_segs;
 275
 276	/* Worst case is non GSO/TSO : each frame consumes one skb
 277	 * and skb->head is kmalloced using power of two area of memory
 278	 */
 279	per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) +
 280		  MAX_TCP_HEADER +
 281		  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
 282
 283	per_mss = roundup_pow_of_two(per_mss) +
 284		  SKB_DATA_ALIGN(sizeof(struct sk_buff));
 285
 286	nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd);
 287	nr_segs = max_t(u32, nr_segs, tp->reordering + 1);
 288
 289	/* Fast Recovery (RFC 5681 3.2) :
 290	 * Cubic needs 1.7 factor, rounded to 2 to include
 291	 * extra cushion (application might react slowly to POLLOUT)
 292	 */
 293	sndmem = 2 * nr_segs * per_mss;
 
 294
 295	if (sk->sk_sndbuf < sndmem)
 296		sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
 
 297}
 298
 299/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
 300 *
 301 * All tcp_full_space() is split to two parts: "network" buffer, allocated
 302 * forward and advertised in receiver window (tp->rcv_wnd) and
 303 * "application buffer", required to isolate scheduling/application
 304 * latencies from network.
 305 * window_clamp is maximal advertised window. It can be less than
 306 * tcp_full_space(), in this case tcp_full_space() - window_clamp
 307 * is reserved for "application" buffer. The less window_clamp is
 308 * the smoother our behaviour from viewpoint of network, but the lower
 309 * throughput and the higher sensitivity of the connection to losses. 8)
 310 *
 311 * rcv_ssthresh is more strict window_clamp used at "slow start"
 312 * phase to predict further behaviour of this connection.
 313 * It is used for two goals:
 314 * - to enforce header prediction at sender, even when application
 315 *   requires some significant "application buffer". It is check #1.
 316 * - to prevent pruning of receive queue because of misprediction
 317 *   of receiver window. Check #2.
 318 *
 319 * The scheme does not work when sender sends good segments opening
 320 * window and then starts to feed us spaghetti. But it should work
 321 * in common situations. Otherwise, we have to rely on queue collapsing.
 322 */
 323
 324/* Slow part of check#2. */
 325static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb)
 
 326{
 327	struct tcp_sock *tp = tcp_sk(sk);
 328	/* Optimize this! */
 329	int truesize = tcp_win_from_space(skb->truesize) >> 1;
 330	int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1;
 331
 332	while (tp->rcv_ssthresh <= window) {
 333		if (truesize <= skb->len)
 334			return 2 * inet_csk(sk)->icsk_ack.rcv_mss;
 335
 336		truesize >>= 1;
 337		window >>= 1;
 338	}
 339	return 0;
 340}
 341
 342static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 343{
 344	struct tcp_sock *tp = tcp_sk(sk);
 
 
 
 
 
 
 345
 346	/* Check #1 */
 347	if (tp->rcv_ssthresh < tp->window_clamp &&
 348	    (int)tp->rcv_ssthresh < tcp_space(sk) &&
 349	    !sk_under_memory_pressure(sk)) {
 350		int incr;
 351
 352		/* Check #2. Increase window, if skb with such overhead
 353		 * will fit to rcvbuf in future.
 354		 */
 355		if (tcp_win_from_space(skb->truesize) <= skb->len)
 356			incr = 2 * tp->advmss;
 357		else
 358			incr = __tcp_grow_window(sk, skb);
 359
 360		if (incr) {
 361			incr = max_t(int, incr, 2 * skb->len);
 362			tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr,
 363					       tp->window_clamp);
 364			inet_csk(sk)->icsk_ack.quick |= 1;
 365		}
 
 
 
 
 
 366	}
 367}
 368
 369/* 3. Tuning rcvbuf, when connection enters established state. */
 370static void tcp_fixup_rcvbuf(struct sock *sk)
 371{
 372	u32 mss = tcp_sk(sk)->advmss;
 373	int rcvmem;
 374
 375	rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) *
 376		 tcp_default_init_rwnd(mss);
 377
 378	/* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency
 379	 * Allow enough cushion so that sender is not limited by our window
 380	 */
 381	if (sysctl_tcp_moderate_rcvbuf)
 382		rcvmem <<= 2;
 383
 384	if (sk->sk_rcvbuf < rcvmem)
 385		sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]);
 386}
 387
 388/* 4. Try to fixup all. It is made immediately after connection enters
 389 *    established state.
 390 */
 391void tcp_init_buffer_space(struct sock *sk)
 392{
 
 393	struct tcp_sock *tp = tcp_sk(sk);
 394	int maxwin;
 395
 396	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 397		tcp_fixup_rcvbuf(sk);
 398	if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK))
 399		tcp_sndbuf_expand(sk);
 400
 401	tp->rcvq_space.space = tp->rcv_wnd;
 402	tp->rcvq_space.time = tcp_time_stamp;
 403	tp->rcvq_space.seq = tp->copied_seq;
 404
 405	maxwin = tcp_full_space(sk);
 406
 407	if (tp->window_clamp >= maxwin) {
 408		tp->window_clamp = maxwin;
 409
 410		if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss)
 411			tp->window_clamp = max(maxwin -
 412					       (maxwin >> sysctl_tcp_app_win),
 413					       4 * tp->advmss);
 414	}
 415
 416	/* Force reservation of one segment. */
 417	if (sysctl_tcp_app_win &&
 418	    tp->window_clamp > 2 * tp->advmss &&
 419	    tp->window_clamp + tp->advmss > maxwin)
 420		tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss);
 
 421
 422	tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp);
 423	tp->snd_cwnd_stamp = tcp_time_stamp;
 
 
 424}
 425
 426/* 5. Recalculate window clamp after socket hit its memory bounds. */
 427static void tcp_clamp_window(struct sock *sk)
 428{
 429	struct tcp_sock *tp = tcp_sk(sk);
 430	struct inet_connection_sock *icsk = inet_csk(sk);
 
 
 431
 432	icsk->icsk_ack.quick = 0;
 
 433
 434	if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] &&
 435	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) &&
 436	    !sk_under_memory_pressure(sk) &&
 437	    sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) {
 438		sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc),
 439				    sysctl_tcp_rmem[2]);
 440	}
 441	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf)
 442		tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss);
 443}
 444
 445/* Initialize RCV_MSS value.
 446 * RCV_MSS is an our guess about MSS used by the peer.
 447 * We haven't any direct information about the MSS.
 448 * It's better to underestimate the RCV_MSS rather than overestimate.
 449 * Overestimations make us ACKing less frequently than needed.
 450 * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss().
 451 */
 452void tcp_initialize_rcv_mss(struct sock *sk)
 453{
 454	const struct tcp_sock *tp = tcp_sk(sk);
 455	unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache);
 456
 457	hint = min(hint, tp->rcv_wnd / 2);
 458	hint = min(hint, TCP_MSS_DEFAULT);
 459	hint = max(hint, TCP_MIN_MSS);
 460
 461	inet_csk(sk)->icsk_ack.rcv_mss = hint;
 462}
 463EXPORT_SYMBOL(tcp_initialize_rcv_mss);
 464
 465/* Receiver "autotuning" code.
 466 *
 467 * The algorithm for RTT estimation w/o timestamps is based on
 468 * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL.
 469 * <http://public.lanl.gov/radiant/pubs.html#DRS>
 470 *
 471 * More detail on this code can be found at
 472 * <http://staff.psc.edu/jheffner/>,
 473 * though this reference is out of date.  A new paper
 474 * is pending.
 475 */
 476static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep)
 477{
 478	u32 new_sample = tp->rcv_rtt_est.rtt;
 479	long m = sample;
 480
 481	if (m == 0)
 482		m = 1;
 483
 484	if (new_sample != 0) {
 485		/* If we sample in larger samples in the non-timestamp
 486		 * case, we could grossly overestimate the RTT especially
 487		 * with chatty applications or bulk transfer apps which
 488		 * are stalled on filesystem I/O.
 489		 *
 490		 * Also, since we are only going for a minimum in the
 491		 * non-timestamp case, we do not smooth things out
 492		 * else with timestamps disabled convergence takes too
 493		 * long.
 494		 */
 495		if (!win_dep) {
 496			m -= (new_sample >> 3);
 497			new_sample += m;
 498		} else {
 499			m <<= 3;
 500			if (m < new_sample)
 501				new_sample = m;
 502		}
 503	} else {
 504		/* No previous measure. */
 505		new_sample = m << 3;
 506	}
 507
 508	if (tp->rcv_rtt_est.rtt != new_sample)
 509		tp->rcv_rtt_est.rtt = new_sample;
 510}
 511
 512static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp)
 513{
 
 
 514	if (tp->rcv_rtt_est.time == 0)
 515		goto new_measure;
 516	if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq))
 517		return;
 518	tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rcv_rtt_est.time, 1);
 
 
 
 519
 520new_measure:
 521	tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd;
 522	tp->rcv_rtt_est.time = tcp_time_stamp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 523}
 524
 525static inline void tcp_rcv_rtt_measure_ts(struct sock *sk,
 526					  const struct sk_buff *skb)
 527{
 528	struct tcp_sock *tp = tcp_sk(sk);
 529	if (tp->rx_opt.rcv_tsecr &&
 530	    (TCP_SKB_CB(skb)->end_seq -
 531	     TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss))
 532		tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0);
 
 
 
 
 
 
 
 
 533}
 534
 535/*
 536 * This function should be called every time data is copied to user space.
 537 * It calculates the appropriate TCP receive buffer space.
 538 */
 539void tcp_rcv_space_adjust(struct sock *sk)
 540{
 541	struct tcp_sock *tp = tcp_sk(sk);
 
 542	int time;
 543	int copied;
 544
 545	time = tcp_time_stamp - tp->rcvq_space.time;
 546	if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0)
 
 
 
 547		return;
 548
 549	/* Number of bytes copied to user in last RTT */
 550	copied = tp->copied_seq - tp->rcvq_space.seq;
 551	if (copied <= tp->rcvq_space.space)
 552		goto new_measure;
 553
 554	/* A bit of theory :
 555	 * copied = bytes received in previous RTT, our base window
 556	 * To cope with packet losses, we need a 2x factor
 557	 * To cope with slow start, and sender growing its cwin by 100 %
 558	 * every RTT, we need a 4x factor, because the ACK we are sending
 559	 * now is for the next RTT, not the current one :
 560	 * <prev RTT . ><current RTT .. ><next RTT .... >
 561	 */
 562
 563	if (sysctl_tcp_moderate_rcvbuf &&
 564	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 565		int rcvwin, rcvmem, rcvbuf;
 
 566
 567		/* minimal window to cope with packet losses, assuming
 568		 * steady state. Add some cushion because of small variations.
 569		 */
 570		rcvwin = (copied << 1) + 16 * tp->advmss;
 571
 572		/* If rate increased by 25%,
 573		 *	assume slow start, rcvwin = 3 * copied
 574		 * If rate increased by 50%,
 575		 *	assume sender can use 2x growth, rcvwin = 4 * copied
 576		 */
 577		if (copied >=
 578		    tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) {
 579			if (copied >=
 580			    tp->rcvq_space.space + (tp->rcvq_space.space >> 1))
 581				rcvwin <<= 1;
 582			else
 583				rcvwin += (rcvwin >> 1);
 584		}
 585
 586		rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER);
 587		while (tcp_win_from_space(rcvmem) < tp->advmss)
 588			rcvmem += 128;
 589
 590		rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]);
 
 591		if (rcvbuf > sk->sk_rcvbuf) {
 592			sk->sk_rcvbuf = rcvbuf;
 593
 594			/* Make the window clamp follow along.  */
 595			tp->window_clamp = rcvwin;
 
 596		}
 597	}
 598	tp->rcvq_space.space = copied;
 599
 600new_measure:
 601	tp->rcvq_space.seq = tp->copied_seq;
 602	tp->rcvq_space.time = tcp_time_stamp;
 
 
 
 
 
 
 
 
 
 
 603}
 604
 605/* There is something which you must keep in mind when you analyze the
 606 * behavior of the tp->ato delayed ack timeout interval.  When a
 607 * connection starts up, we want to ack as quickly as possible.  The
 608 * problem is that "good" TCP's do slow start at the beginning of data
 609 * transmission.  The means that until we send the first few ACK's the
 610 * sender will sit on his end and only queue most of his data, because
 611 * he can only send snd_cwnd unacked packets at any given time.  For
 612 * each ACK we send, he increments snd_cwnd and transmits more of his
 613 * queue.  -DaveM
 614 */
 615static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb)
 616{
 617	struct tcp_sock *tp = tcp_sk(sk);
 618	struct inet_connection_sock *icsk = inet_csk(sk);
 619	u32 now;
 620
 621	inet_csk_schedule_ack(sk);
 622
 623	tcp_measure_rcv_mss(sk, skb);
 624
 625	tcp_rcv_rtt_measure(tp);
 626
 627	now = tcp_time_stamp;
 628
 629	if (!icsk->icsk_ack.ato) {
 630		/* The _first_ data packet received, initialize
 631		 * delayed ACK engine.
 632		 */
 633		tcp_incr_quickack(sk);
 634		icsk->icsk_ack.ato = TCP_ATO_MIN;
 635	} else {
 636		int m = now - icsk->icsk_ack.lrcvtime;
 637
 638		if (m <= TCP_ATO_MIN / 2) {
 639			/* The fastest case is the first. */
 640			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2;
 641		} else if (m < icsk->icsk_ack.ato) {
 642			icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m;
 643			if (icsk->icsk_ack.ato > icsk->icsk_rto)
 644				icsk->icsk_ack.ato = icsk->icsk_rto;
 645		} else if (m > icsk->icsk_rto) {
 646			/* Too long gap. Apparently sender failed to
 647			 * restart window, so that we send ACKs quickly.
 648			 */
 649			tcp_incr_quickack(sk);
 650			sk_mem_reclaim(sk);
 651		}
 652	}
 653	icsk->icsk_ack.lrcvtime = now;
 
 654
 655	TCP_ECN_check_ce(tp, skb);
 656
 657	if (skb->len >= 128)
 658		tcp_grow_window(sk, skb);
 659}
 660
 661/* Called to compute a smoothed rtt estimate. The data fed to this
 662 * routine either comes from timestamps, or from segments that were
 663 * known _not_ to have been retransmitted [see Karn/Partridge
 664 * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88
 665 * piece by Van Jacobson.
 666 * NOTE: the next three routines used to be one big routine.
 667 * To save cycles in the RFC 1323 implementation it was better to break
 668 * it up into three procedures. -- erics
 669 */
 670static void tcp_rtt_estimator(struct sock *sk, long mrtt_us)
 671{
 672	struct tcp_sock *tp = tcp_sk(sk);
 673	long m = mrtt_us; /* RTT */
 674	u32 srtt = tp->srtt_us;
 675
 676	/*	The following amusing code comes from Jacobson's
 677	 *	article in SIGCOMM '88.  Note that rtt and mdev
 678	 *	are scaled versions of rtt and mean deviation.
 679	 *	This is designed to be as fast as possible
 680	 *	m stands for "measurement".
 681	 *
 682	 *	On a 1990 paper the rto value is changed to:
 683	 *	RTO = rtt + 4 * mdev
 684	 *
 685	 * Funny. This algorithm seems to be very broken.
 686	 * These formulae increase RTO, when it should be decreased, increase
 687	 * too slowly, when it should be increased quickly, decrease too quickly
 688	 * etc. I guess in BSD RTO takes ONE value, so that it is absolutely
 689	 * does not matter how to _calculate_ it. Seems, it was trap
 690	 * that VJ failed to avoid. 8)
 691	 */
 692	if (srtt != 0) {
 693		m -= (srtt >> 3);	/* m is now error in rtt est */
 694		srtt += m;		/* rtt = 7/8 rtt + 1/8 new */
 695		if (m < 0) {
 696			m = -m;		/* m is now abs(error) */
 697			m -= (tp->mdev_us >> 2);   /* similar update on mdev */
 698			/* This is similar to one of Eifel findings.
 699			 * Eifel blocks mdev updates when rtt decreases.
 700			 * This solution is a bit different: we use finer gain
 701			 * for mdev in this case (alpha*beta).
 702			 * Like Eifel it also prevents growth of rto,
 703			 * but also it limits too fast rto decreases,
 704			 * happening in pure Eifel.
 705			 */
 706			if (m > 0)
 707				m >>= 3;
 708		} else {
 709			m -= (tp->mdev_us >> 2);   /* similar update on mdev */
 710		}
 711		tp->mdev_us += m;		/* mdev = 3/4 mdev + 1/4 new */
 712		if (tp->mdev_us > tp->mdev_max_us) {
 713			tp->mdev_max_us = tp->mdev_us;
 714			if (tp->mdev_max_us > tp->rttvar_us)
 715				tp->rttvar_us = tp->mdev_max_us;
 716		}
 717		if (after(tp->snd_una, tp->rtt_seq)) {
 718			if (tp->mdev_max_us < tp->rttvar_us)
 719				tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2;
 720			tp->rtt_seq = tp->snd_nxt;
 721			tp->mdev_max_us = tcp_rto_min_us(sk);
 
 
 722		}
 723	} else {
 724		/* no previous measure. */
 725		srtt = m << 3;		/* take the measured time to be rtt */
 726		tp->mdev_us = m << 1;	/* make sure rto = 3*rtt */
 727		tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk));
 728		tp->mdev_max_us = tp->rttvar_us;
 729		tp->rtt_seq = tp->snd_nxt;
 
 
 730	}
 731	tp->srtt_us = max(1U, srtt);
 732}
 733
 734/* Set the sk_pacing_rate to allow proper sizing of TSO packets.
 735 * Note: TCP stack does not yet implement pacing.
 736 * FQ packet scheduler can be used to implement cheap but effective
 737 * TCP pacing, to smooth the burst on large writes when packets
 738 * in flight is significantly lower than cwnd (or rwin)
 739 */
 740static void tcp_update_pacing_rate(struct sock *sk)
 741{
 742	const struct tcp_sock *tp = tcp_sk(sk);
 743	u64 rate;
 744
 745	/* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */
 746	rate = (u64)tp->mss_cache * 2 * (USEC_PER_SEC << 3);
 747
 748	rate *= max(tp->snd_cwnd, tp->packets_out);
 
 
 
 
 
 
 
 
 
 
 
 
 
 749
 750	if (likely(tp->srtt_us))
 751		do_div(rate, tp->srtt_us);
 752
 753	/* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate
 754	 * without any lock. We want to make sure compiler wont store
 755	 * intermediate values in this location.
 756	 */
 757	ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate,
 758						sk->sk_max_pacing_rate);
 759}
 760
 761/* Calculate rto without backoff.  This is the second half of Van Jacobson's
 762 * routine referred to above.
 763 */
 764static void tcp_set_rto(struct sock *sk)
 765{
 766	const struct tcp_sock *tp = tcp_sk(sk);
 767	/* Old crap is replaced with new one. 8)
 768	 *
 769	 * More seriously:
 770	 * 1. If rtt variance happened to be less 50msec, it is hallucination.
 771	 *    It cannot be less due to utterly erratic ACK generation made
 772	 *    at least by solaris and freebsd. "Erratic ACKs" has _nothing_
 773	 *    to do with delayed acks, because at cwnd>2 true delack timeout
 774	 *    is invisible. Actually, Linux-2.4 also generates erratic
 775	 *    ACKs in some circumstances.
 776	 */
 777	inet_csk(sk)->icsk_rto = __tcp_set_rto(tp);
 778
 779	/* 2. Fixups made earlier cannot be right.
 780	 *    If we do not estimate RTO correctly without them,
 781	 *    all the algo is pure shit and should be replaced
 782	 *    with correct one. It is exactly, which we pretend to do.
 783	 */
 784
 785	/* NOTE: clamping at TCP_RTO_MIN is not required, current algo
 786	 * guarantees that rto is higher.
 787	 */
 788	tcp_bound_rto(sk);
 789}
 790
 791__u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst)
 792{
 793	__u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0);
 794
 795	if (!cwnd)
 796		cwnd = TCP_INIT_CWND;
 797	return min_t(__u32, cwnd, tp->snd_cwnd_clamp);
 798}
 799
 800/*
 801 * Packet counting of FACK is based on in-order assumptions, therefore TCP
 802 * disables it when reordering is detected
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 803 */
 804void tcp_disable_fack(struct tcp_sock *tp)
 
 805{
 806	/* RFC3517 uses different metric in lost marker => reset on change */
 807	if (tcp_is_fack(tp))
 808		tp->lost_skb_hint = NULL;
 809	tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED;
 810}
 
 
 
 
 
 
 
 
 
 
 
 
 
 811
 812/* Take a notice that peer is sending D-SACKs */
 813static void tcp_dsack_seen(struct tcp_sock *tp)
 814{
 815	tp->rx_opt.sack_ok |= TCP_DSACK_SEEN;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 816}
 817
 818static void tcp_update_reordering(struct sock *sk, const int metric,
 819				  const int ts)
 
 
 
 
 820{
 821	struct tcp_sock *tp = tcp_sk(sk);
 822	if (metric > tp->reordering) {
 823		int mib_idx;
 824
 825		tp->reordering = min(TCP_MAX_REORDERING, metric);
 826
 827		/* This exciting event is worth to be remembered. 8) */
 828		if (ts)
 829			mib_idx = LINUX_MIB_TCPTSREORDER;
 830		else if (tcp_is_reno(tp))
 831			mib_idx = LINUX_MIB_TCPRENOREORDER;
 832		else if (tcp_is_fack(tp))
 833			mib_idx = LINUX_MIB_TCPFACKREORDER;
 834		else
 835			mib_idx = LINUX_MIB_TCPSACKREORDER;
 836
 837		NET_INC_STATS_BH(sock_net(sk), mib_idx);
 
 838#if FASTRETRANS_DEBUG > 1
 839		pr_debug("Disorder%d %d %u f%u s%u rr%d\n",
 840			 tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state,
 841			 tp->reordering,
 842			 tp->fackets_out,
 843			 tp->sacked_out,
 844			 tp->undo_marker ? tp->undo_retrans : 0);
 845#endif
 846		tcp_disable_fack(tp);
 
 847	}
 848
 849	if (metric > 0)
 850		tcp_disable_early_retrans(tp);
 
 
 851}
 852
 853/* This must be called before lost_out is incremented */
 
 
 
 
 854static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb)
 855{
 856	if ((tp->retransmit_skb_hint == NULL) ||
 857	    before(TCP_SKB_CB(skb)->seq,
 858		   TCP_SKB_CB(tp->retransmit_skb_hint)->seq))
 
 859		tp->retransmit_skb_hint = skb;
 
 860
 861	if (!tp->lost_out ||
 862	    after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high))
 863		tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
 
 
 
 864}
 865
 866static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb)
 867{
 868	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 869		tcp_verify_retransmit_hint(tp, skb);
 
 
 
 870
 
 
 
 
 
 
 
 
 
 
 
 871		tp->lost_out += tcp_skb_pcount(skb);
 872		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 
 873	}
 874}
 875
 876static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp,
 877					    struct sk_buff *skb)
 
 878{
 879	tcp_verify_retransmit_hint(tp, skb);
 880
 881	if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) {
 882		tp->lost_out += tcp_skb_pcount(skb);
 883		TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
 884	}
 885}
 886
 887/* This procedure tags the retransmission queue when SACKs arrive.
 888 *
 889 * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L).
 890 * Packets in queue with these bits set are counted in variables
 891 * sacked_out, retrans_out and lost_out, correspondingly.
 892 *
 893 * Valid combinations are:
 894 * Tag  InFlight	Description
 895 * 0	1		- orig segment is in flight.
 896 * S	0		- nothing flies, orig reached receiver.
 897 * L	0		- nothing flies, orig lost by net.
 898 * R	2		- both orig and retransmit are in flight.
 899 * L|R	1		- orig is lost, retransmit is in flight.
 900 * S|R  1		- orig reached receiver, retrans is still in flight.
 901 * (L|S|R is logically valid, it could occur when L|R is sacked,
 902 *  but it is equivalent to plain S and code short-curcuits it to S.
 903 *  L|S is logically invalid, it would mean -1 packet in flight 8))
 904 *
 905 * These 6 states form finite state machine, controlled by the following events:
 906 * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue())
 907 * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue())
 908 * 3. Loss detection event of two flavors:
 909 *	A. Scoreboard estimator decided the packet is lost.
 910 *	   A'. Reno "three dupacks" marks head of queue lost.
 911 *	   A''. Its FACK modification, head until snd.fack is lost.
 912 *	B. SACK arrives sacking SND.NXT at the moment, when the
 913 *	   segment was retransmitted.
 914 * 4. D-SACK added new rule: D-SACK changes any tag to S.
 915 *
 916 * It is pleasant to note, that state diagram turns out to be commutative,
 917 * so that we are allowed not to be bothered by order of our actions,
 918 * when multiple events arrive simultaneously. (see the function below).
 919 *
 920 * Reordering detection.
 921 * --------------------
 922 * Reordering metric is maximal distance, which a packet can be displaced
 923 * in packet stream. With SACKs we can estimate it:
 924 *
 925 * 1. SACK fills old hole and the corresponding segment was not
 926 *    ever retransmitted -> reordering. Alas, we cannot use it
 927 *    when segment was retransmitted.
 928 * 2. The last flaw is solved with D-SACK. D-SACK arrives
 929 *    for retransmitted and already SACKed segment -> reordering..
 930 * Both of these heuristics are not used in Loss state, when we cannot
 931 * account for retransmits accurately.
 932 *
 933 * SACK block validation.
 934 * ----------------------
 935 *
 936 * SACK block range validation checks that the received SACK block fits to
 937 * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT.
 938 * Note that SND.UNA is not included to the range though being valid because
 939 * it means that the receiver is rather inconsistent with itself reporting
 940 * SACK reneging when it should advance SND.UNA. Such SACK block this is
 941 * perfectly valid, however, in light of RFC2018 which explicitly states
 942 * that "SACK block MUST reflect the newest segment.  Even if the newest
 943 * segment is going to be discarded ...", not that it looks very clever
 944 * in case of head skb. Due to potentional receiver driven attacks, we
 945 * choose to avoid immediate execution of a walk in write queue due to
 946 * reneging and defer head skb's loss recovery to standard loss recovery
 947 * procedure that will eventually trigger (nothing forbids us doing this).
 948 *
 949 * Implements also blockage to start_seq wrap-around. Problem lies in the
 950 * fact that though start_seq (s) is before end_seq (i.e., not reversed),
 951 * there's no guarantee that it will be before snd_nxt (n). The problem
 952 * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt
 953 * wrap (s_w):
 954 *
 955 *         <- outs wnd ->                          <- wrapzone ->
 956 *         u     e      n                         u_w   e_w  s n_w
 957 *         |     |      |                          |     |   |  |
 958 * |<------------+------+----- TCP seqno space --------------+---------->|
 959 * ...-- <2^31 ->|                                           |<--------...
 960 * ...---- >2^31 ------>|                                    |<--------...
 961 *
 962 * Current code wouldn't be vulnerable but it's better still to discard such
 963 * crazy SACK blocks. Doing this check for start_seq alone closes somewhat
 964 * similar case (end_seq after snd_nxt wrap) as earlier reversed check in
 965 * snd_nxt wrap -> snd_una region will then become "well defined", i.e.,
 966 * equal to the ideal case (infinite seqno space without wrap caused issues).
 967 *
 968 * With D-SACK the lower bound is extended to cover sequence space below
 969 * SND.UNA down to undo_marker, which is the last point of interest. Yet
 970 * again, D-SACK block must not to go across snd_una (for the same reason as
 971 * for the normal SACK blocks, explained above). But there all simplicity
 972 * ends, TCP might receive valid D-SACKs below that. As long as they reside
 973 * fully below undo_marker they do not affect behavior in anyway and can
 974 * therefore be safely ignored. In rare cases (which are more or less
 975 * theoretical ones), the D-SACK will nicely cross that boundary due to skb
 976 * fragmentation and packet reordering past skb's retransmission. To consider
 977 * them correctly, the acceptable range must be extended even more though
 978 * the exact amount is rather hard to quantify. However, tp->max_window can
 979 * be used as an exaggerated estimate.
 980 */
 981static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack,
 982				   u32 start_seq, u32 end_seq)
 983{
 984	/* Too far in future, or reversed (interpretation is ambiguous) */
 985	if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq))
 986		return false;
 987
 988	/* Nasty start_seq wrap-around check (see comments above) */
 989	if (!before(start_seq, tp->snd_nxt))
 990		return false;
 991
 992	/* In outstanding window? ...This is valid exit for D-SACKs too.
 993	 * start_seq == snd_una is non-sensical (see comments above)
 994	 */
 995	if (after(start_seq, tp->snd_una))
 996		return true;
 997
 998	if (!is_dsack || !tp->undo_marker)
 999		return false;
1000
1001	/* ...Then it's D-SACK, and must reside below snd_una completely */
1002	if (after(end_seq, tp->snd_una))
1003		return false;
1004
1005	if (!before(start_seq, tp->undo_marker))
1006		return true;
1007
1008	/* Too old */
1009	if (!after(end_seq, tp->undo_marker))
1010		return false;
1011
1012	/* Undo_marker boundary crossing (overestimates a lot). Known already:
1013	 *   start_seq < undo_marker and end_seq >= undo_marker.
1014	 */
1015	return !before(start_seq, end_seq - tp->max_window);
1016}
1017
1018/* Check for lost retransmit. This superb idea is borrowed from "ratehalving".
1019 * Event "B". Later note: FACK people cheated me again 8), we have to account
1020 * for reordering! Ugly, but should help.
1021 *
1022 * Search retransmitted skbs from write_queue that were sent when snd_nxt was
1023 * less than what is now known to be received by the other end (derived from
1024 * highest SACK block). Also calculate the lowest snd_nxt among the remaining
1025 * retransmitted skbs to avoid some costly processing per ACKs.
1026 */
1027static void tcp_mark_lost_retrans(struct sock *sk)
1028{
1029	const struct inet_connection_sock *icsk = inet_csk(sk);
1030	struct tcp_sock *tp = tcp_sk(sk);
1031	struct sk_buff *skb;
1032	int cnt = 0;
1033	u32 new_low_seq = tp->snd_nxt;
1034	u32 received_upto = tcp_highest_sack_seq(tp);
1035
1036	if (!tcp_is_fack(tp) || !tp->retrans_out ||
1037	    !after(received_upto, tp->lost_retrans_low) ||
1038	    icsk->icsk_ca_state != TCP_CA_Recovery)
1039		return;
1040
1041	tcp_for_write_queue(skb, sk) {
1042		u32 ack_seq = TCP_SKB_CB(skb)->ack_seq;
1043
1044		if (skb == tcp_send_head(sk))
1045			break;
1046		if (cnt == tp->retrans_out)
1047			break;
1048		if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1049			continue;
1050
1051		if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS))
1052			continue;
1053
1054		/* TODO: We would like to get rid of tcp_is_fack(tp) only
1055		 * constraint here (see above) but figuring out that at
1056		 * least tp->reordering SACK blocks reside between ack_seq
1057		 * and received_upto is not easy task to do cheaply with
1058		 * the available datastructures.
1059		 *
1060		 * Whether FACK should check here for tp->reordering segs
1061		 * in-between one could argue for either way (it would be
1062		 * rather simple to implement as we could count fack_count
1063		 * during the walk and do tp->fackets_out - fack_count).
1064		 */
1065		if (after(received_upto, ack_seq)) {
1066			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
1067			tp->retrans_out -= tcp_skb_pcount(skb);
1068
1069			tcp_skb_mark_lost_uncond_verify(tp, skb);
1070			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT);
1071		} else {
1072			if (before(ack_seq, new_low_seq))
1073				new_low_seq = ack_seq;
1074			cnt += tcp_skb_pcount(skb);
1075		}
1076	}
1077
1078	if (tp->retrans_out)
1079		tp->lost_retrans_low = new_low_seq;
1080}
1081
1082static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb,
1083			    struct tcp_sack_block_wire *sp, int num_sacks,
1084			    u32 prior_snd_una)
1085{
1086	struct tcp_sock *tp = tcp_sk(sk);
1087	u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq);
1088	u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq);
1089	bool dup_sack = false;
1090
1091	if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) {
1092		dup_sack = true;
1093		tcp_dsack_seen(tp);
1094		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV);
1095	} else if (num_sacks > 1) {
1096		u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq);
1097		u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq);
1098
1099		if (!after(end_seq_0, end_seq_1) &&
1100		    !before(start_seq_0, start_seq_1)) {
1101			dup_sack = true;
1102			tcp_dsack_seen(tp);
1103			NET_INC_STATS_BH(sock_net(sk),
1104					LINUX_MIB_TCPDSACKOFORECV);
1105		}
 
 
 
 
1106	}
1107
 
 
1108	/* D-SACK for already forgotten data... Do dumb counting. */
1109	if (dup_sack && tp->undo_marker && tp->undo_retrans &&
1110	    !after(end_seq_0, prior_snd_una) &&
1111	    after(end_seq_0, tp->undo_marker))
1112		tp->undo_retrans--;
1113
1114	return dup_sack;
1115}
1116
1117struct tcp_sacktag_state {
1118	int	reord;
1119	int	fack_count;
1120	long	rtt_us; /* RTT measured by SACKing never-retransmitted data */
1121	int	flag;
1122};
1123
1124/* Check if skb is fully within the SACK block. In presence of GSO skbs,
1125 * the incoming SACK may not exactly match but we can find smaller MSS
1126 * aligned portion of it that matches. Therefore we might need to fragment
1127 * which may fail and creates some hassle (caller must handle error case
1128 * returns).
1129 *
1130 * FIXME: this could be merged to shift decision code
1131 */
1132static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb,
1133				  u32 start_seq, u32 end_seq)
1134{
1135	int err;
1136	bool in_sack;
1137	unsigned int pkt_len;
1138	unsigned int mss;
1139
1140	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1141		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1142
1143	if (tcp_skb_pcount(skb) > 1 && !in_sack &&
1144	    after(TCP_SKB_CB(skb)->end_seq, start_seq)) {
1145		mss = tcp_skb_mss(skb);
1146		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1147
1148		if (!in_sack) {
1149			pkt_len = start_seq - TCP_SKB_CB(skb)->seq;
1150			if (pkt_len < mss)
1151				pkt_len = mss;
1152		} else {
1153			pkt_len = end_seq - TCP_SKB_CB(skb)->seq;
1154			if (pkt_len < mss)
1155				return -EINVAL;
1156		}
1157
1158		/* Round if necessary so that SACKs cover only full MSSes
1159		 * and/or the remaining small portion (if present)
1160		 */
1161		if (pkt_len > mss) {
1162			unsigned int new_len = (pkt_len / mss) * mss;
1163			if (!in_sack && new_len < pkt_len) {
1164				new_len += mss;
1165				if (new_len > skb->len)
1166					return 0;
1167			}
1168			pkt_len = new_len;
1169		}
1170		err = tcp_fragment(sk, skb, pkt_len, mss);
 
 
 
 
 
1171		if (err < 0)
1172			return err;
1173	}
1174
1175	return in_sack;
1176}
1177
1178/* Mark the given newly-SACKed range as such, adjusting counters and hints. */
1179static u8 tcp_sacktag_one(struct sock *sk,
1180			  struct tcp_sacktag_state *state, u8 sacked,
1181			  u32 start_seq, u32 end_seq,
1182			  int dup_sack, int pcount,
1183			  const struct skb_mstamp *xmit_time)
1184{
1185	struct tcp_sock *tp = tcp_sk(sk);
1186	int fack_count = state->fack_count;
1187
1188	/* Account D-SACK for retransmitted packet. */
1189	if (dup_sack && (sacked & TCPCB_RETRANS)) {
1190		if (tp->undo_marker && tp->undo_retrans &&
1191		    after(end_seq, tp->undo_marker))
1192			tp->undo_retrans--;
1193		if (sacked & TCPCB_SACKED_ACKED)
1194			state->reord = min(fack_count, state->reord);
 
1195	}
1196
1197	/* Nothing to do; acked frame is about to be dropped (was ACKed). */
1198	if (!after(end_seq, tp->snd_una))
1199		return sacked;
1200
1201	if (!(sacked & TCPCB_SACKED_ACKED)) {
 
 
1202		if (sacked & TCPCB_SACKED_RETRANS) {
1203			/* If the segment is not tagged as lost,
1204			 * we do not clear RETRANS, believing
1205			 * that retransmission is still in flight.
1206			 */
1207			if (sacked & TCPCB_LOST) {
1208				sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS);
1209				tp->lost_out -= pcount;
1210				tp->retrans_out -= pcount;
1211			}
1212		} else {
1213			if (!(sacked & TCPCB_RETRANS)) {
1214				/* New sack for not retransmitted frame,
1215				 * which was in hole. It is reordering.
1216				 */
1217				if (before(start_seq,
1218					   tcp_highest_sack_seq(tp)))
1219					state->reord = min(fack_count,
1220							   state->reord);
 
1221				if (!after(end_seq, tp->high_seq))
1222					state->flag |= FLAG_ORIG_SACK_ACKED;
1223				/* Pick the earliest sequence sacked for RTT */
1224				if (state->rtt_us < 0) {
1225					struct skb_mstamp now;
1226
1227					skb_mstamp_get(&now);
1228					state->rtt_us = skb_mstamp_us_delta(&now,
1229								xmit_time);
1230				}
1231			}
1232
1233			if (sacked & TCPCB_LOST) {
1234				sacked &= ~TCPCB_LOST;
1235				tp->lost_out -= pcount;
1236			}
1237		}
1238
1239		sacked |= TCPCB_SACKED_ACKED;
1240		state->flag |= FLAG_DATA_SACKED;
1241		tp->sacked_out += pcount;
1242
1243		fack_count += pcount;
1244
1245		/* Lost marker hint past SACKed? Tweak RFC3517 cnt */
1246		if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) &&
1247		    before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq))
1248			tp->lost_cnt_hint += pcount;
1249
1250		if (fack_count > tp->fackets_out)
1251			tp->fackets_out = fack_count;
1252	}
1253
1254	/* D-SACK. We can detect redundant retransmission in S|R and plain R
1255	 * frames and clear it. undo_retrans is decreased above, L|R frames
1256	 * are accounted above as well.
1257	 */
1258	if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) {
1259		sacked &= ~TCPCB_SACKED_RETRANS;
1260		tp->retrans_out -= pcount;
1261	}
1262
1263	return sacked;
1264}
1265
1266/* Shift newly-SACKed bytes from this skb to the immediately previous
1267 * already-SACKed sk_buff. Mark the newly-SACKed bytes as such.
1268 */
1269static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb,
 
1270			    struct tcp_sacktag_state *state,
1271			    unsigned int pcount, int shifted, int mss,
1272			    bool dup_sack)
1273{
1274	struct tcp_sock *tp = tcp_sk(sk);
1275	struct sk_buff *prev = tcp_write_queue_prev(sk, skb);
1276	u32 start_seq = TCP_SKB_CB(skb)->seq;	/* start of newly-SACKed */
1277	u32 end_seq = start_seq + shifted;	/* end of newly-SACKed */
1278
1279	BUG_ON(!pcount);
1280
1281	/* Adjust counters and hints for the newly sacked sequence
1282	 * range but discard the return value since prev is already
1283	 * marked. We must tag the range first because the seq
1284	 * advancement below implicitly advances
1285	 * tcp_highest_sack_seq() when skb is highest_sack.
1286	 */
1287	tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked,
1288			start_seq, end_seq, dup_sack, pcount,
1289			&skb->skb_mstamp);
 
1290
1291	if (skb == tp->lost_skb_hint)
1292		tp->lost_cnt_hint += pcount;
1293
1294	TCP_SKB_CB(prev)->end_seq += shifted;
1295	TCP_SKB_CB(skb)->seq += shifted;
1296
1297	skb_shinfo(prev)->gso_segs += pcount;
1298	BUG_ON(skb_shinfo(skb)->gso_segs < pcount);
1299	skb_shinfo(skb)->gso_segs -= pcount;
1300
1301	/* When we're adding to gso_segs == 1, gso_size will be zero,
1302	 * in theory this shouldn't be necessary but as long as DSACK
1303	 * code can come after this skb later on it's better to keep
1304	 * setting gso_size to something.
1305	 */
1306	if (!skb_shinfo(prev)->gso_size) {
1307		skb_shinfo(prev)->gso_size = mss;
1308		skb_shinfo(prev)->gso_type = sk->sk_gso_type;
1309	}
1310
1311	/* CHECKME: To clear or not to clear? Mimics normal skb currently */
1312	if (skb_shinfo(skb)->gso_segs <= 1) {
1313		skb_shinfo(skb)->gso_size = 0;
1314		skb_shinfo(skb)->gso_type = 0;
1315	}
1316
1317	/* Difference in this won't matter, both ACKed by the same cumul. ACK */
1318	TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS);
1319
1320	if (skb->len > 0) {
1321		BUG_ON(!tcp_skb_pcount(skb));
1322		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED);
1323		return false;
1324	}
1325
1326	/* Whole SKB was eaten :-) */
1327
1328	if (skb == tp->retransmit_skb_hint)
1329		tp->retransmit_skb_hint = prev;
1330	if (skb == tp->lost_skb_hint) {
1331		tp->lost_skb_hint = prev;
1332		tp->lost_cnt_hint -= tcp_skb_pcount(prev);
1333	}
1334
1335	TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
 
1336	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1337		TCP_SKB_CB(prev)->end_seq++;
1338
1339	if (skb == tcp_highest_sack(sk))
1340		tcp_advance_highest_sack(sk, skb);
1341
1342	tcp_unlink_write_queue(skb, sk);
1343	sk_wmem_free_skb(sk, skb);
 
1344
1345	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED);
 
 
1346
1347	return true;
1348}
1349
1350/* I wish gso_size would have a bit more sane initialization than
1351 * something-or-zero which complicates things
1352 */
1353static int tcp_skb_seglen(const struct sk_buff *skb)
1354{
1355	return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb);
1356}
1357
1358/* Shifting pages past head area doesn't work */
1359static int skb_can_shift(const struct sk_buff *skb)
1360{
1361	return !skb_headlen(skb) && skb_is_nonlinear(skb);
1362}
1363
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1364/* Try collapsing SACK blocks spanning across multiple skbs to a single
1365 * skb.
1366 */
1367static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb,
1368					  struct tcp_sacktag_state *state,
1369					  u32 start_seq, u32 end_seq,
1370					  bool dup_sack)
1371{
1372	struct tcp_sock *tp = tcp_sk(sk);
1373	struct sk_buff *prev;
1374	int mss;
1375	int pcount = 0;
1376	int len;
1377	int in_sack;
1378
1379	if (!sk_can_gso(sk))
1380		goto fallback;
1381
1382	/* Normally R but no L won't result in plain S */
1383	if (!dup_sack &&
1384	    (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS)
1385		goto fallback;
1386	if (!skb_can_shift(skb))
1387		goto fallback;
1388	/* This frame is about to be dropped (was ACKed). */
1389	if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
1390		goto fallback;
1391
1392	/* Can only happen with delayed DSACK + discard craziness */
1393	if (unlikely(skb == tcp_write_queue_head(sk)))
 
1394		goto fallback;
1395	prev = tcp_write_queue_prev(sk, skb);
1396
1397	if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED)
1398		goto fallback;
1399
 
 
 
1400	in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) &&
1401		  !before(end_seq, TCP_SKB_CB(skb)->end_seq);
1402
1403	if (in_sack) {
1404		len = skb->len;
1405		pcount = tcp_skb_pcount(skb);
1406		mss = tcp_skb_seglen(skb);
1407
1408		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1409		 * drop this restriction as unnecessary
1410		 */
1411		if (mss != tcp_skb_seglen(prev))
1412			goto fallback;
1413	} else {
1414		if (!after(TCP_SKB_CB(skb)->end_seq, start_seq))
1415			goto noop;
1416		/* CHECKME: This is non-MSS split case only?, this will
1417		 * cause skipped skbs due to advancing loop btw, original
1418		 * has that feature too
1419		 */
1420		if (tcp_skb_pcount(skb) <= 1)
1421			goto noop;
1422
1423		in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq);
1424		if (!in_sack) {
1425			/* TODO: head merge to next could be attempted here
1426			 * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)),
1427			 * though it might not be worth of the additional hassle
1428			 *
1429			 * ...we can probably just fallback to what was done
1430			 * previously. We could try merging non-SACKed ones
1431			 * as well but it probably isn't going to buy off
1432			 * because later SACKs might again split them, and
1433			 * it would make skb timestamp tracking considerably
1434			 * harder problem.
1435			 */
1436			goto fallback;
1437		}
1438
1439		len = end_seq - TCP_SKB_CB(skb)->seq;
1440		BUG_ON(len < 0);
1441		BUG_ON(len > skb->len);
1442
1443		/* MSS boundaries should be honoured or else pcount will
1444		 * severely break even though it makes things bit trickier.
1445		 * Optimize common case to avoid most of the divides
1446		 */
1447		mss = tcp_skb_mss(skb);
1448
1449		/* TODO: Fix DSACKs to not fragment already SACKed and we can
1450		 * drop this restriction as unnecessary
1451		 */
1452		if (mss != tcp_skb_seglen(prev))
1453			goto fallback;
1454
1455		if (len == mss) {
1456			pcount = 1;
1457		} else if (len < mss) {
1458			goto noop;
1459		} else {
1460			pcount = len / mss;
1461			len = pcount * mss;
1462		}
1463	}
1464
1465	/* tcp_sacktag_one() won't SACK-tag ranges below snd_una */
1466	if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una))
1467		goto fallback;
1468
1469	if (!skb_shift(prev, skb, len))
1470		goto fallback;
1471	if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack))
1472		goto out;
1473
1474	/* Hole filled allows collapsing with the next as well, this is very
1475	 * useful when hole on every nth skb pattern happens
1476	 */
1477	if (prev == tcp_write_queue_tail(sk))
 
1478		goto out;
1479	skb = tcp_write_queue_next(sk, prev);
1480
1481	if (!skb_can_shift(skb) ||
1482	    (skb == tcp_send_head(sk)) ||
1483	    ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) ||
1484	    (mss != tcp_skb_seglen(skb)))
1485		goto out;
1486
 
 
1487	len = skb->len;
1488	if (skb_shift(prev, skb, len)) {
1489		pcount += tcp_skb_pcount(skb);
1490		tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0);
1491	}
1492
1493out:
1494	state->fack_count += pcount;
1495	return prev;
1496
1497noop:
1498	return skb;
1499
1500fallback:
1501	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK);
1502	return NULL;
1503}
1504
1505static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk,
1506					struct tcp_sack_block *next_dup,
1507					struct tcp_sacktag_state *state,
1508					u32 start_seq, u32 end_seq,
1509					bool dup_sack_in)
1510{
1511	struct tcp_sock *tp = tcp_sk(sk);
1512	struct sk_buff *tmp;
1513
1514	tcp_for_write_queue_from(skb, sk) {
1515		int in_sack = 0;
1516		bool dup_sack = dup_sack_in;
1517
1518		if (skb == tcp_send_head(sk))
1519			break;
1520
1521		/* queue is in-order => we can short-circuit the walk early */
1522		if (!before(TCP_SKB_CB(skb)->seq, end_seq))
1523			break;
1524
1525		if ((next_dup != NULL) &&
1526		    before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) {
1527			in_sack = tcp_match_skb_to_sack(sk, skb,
1528							next_dup->start_seq,
1529							next_dup->end_seq);
1530			if (in_sack > 0)
1531				dup_sack = true;
1532		}
1533
1534		/* skb reference here is a bit tricky to get right, since
1535		 * shifting can eat and free both this skb and the next,
1536		 * so not even _safe variant of the loop is enough.
1537		 */
1538		if (in_sack <= 0) {
1539			tmp = tcp_shift_skb_data(sk, skb, state,
1540						 start_seq, end_seq, dup_sack);
1541			if (tmp != NULL) {
1542				if (tmp != skb) {
1543					skb = tmp;
1544					continue;
1545				}
1546
1547				in_sack = 0;
1548			} else {
1549				in_sack = tcp_match_skb_to_sack(sk, skb,
1550								start_seq,
1551								end_seq);
1552			}
1553		}
1554
1555		if (unlikely(in_sack < 0))
1556			break;
1557
1558		if (in_sack) {
1559			TCP_SKB_CB(skb)->sacked =
1560				tcp_sacktag_one(sk,
1561						state,
1562						TCP_SKB_CB(skb)->sacked,
1563						TCP_SKB_CB(skb)->seq,
1564						TCP_SKB_CB(skb)->end_seq,
1565						dup_sack,
1566						tcp_skb_pcount(skb),
1567						&skb->skb_mstamp);
 
 
 
1568
1569			if (!before(TCP_SKB_CB(skb)->seq,
1570				    tcp_highest_sack_seq(tp)))
1571				tcp_advance_highest_sack(sk, skb);
1572		}
1573
1574		state->fack_count += tcp_skb_pcount(skb);
1575	}
1576	return skb;
1577}
1578
1579/* Avoid all extra work that is being done by sacktag while walking in
1580 * a normal way
1581 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1582static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk,
1583					struct tcp_sacktag_state *state,
1584					u32 skip_to_seq)
1585{
1586	tcp_for_write_queue_from(skb, sk) {
1587		if (skb == tcp_send_head(sk))
1588			break;
1589
1590		if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq))
1591			break;
1592
1593		state->fack_count += tcp_skb_pcount(skb);
1594	}
1595	return skb;
1596}
1597
1598static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb,
1599						struct sock *sk,
1600						struct tcp_sack_block *next_dup,
1601						struct tcp_sacktag_state *state,
1602						u32 skip_to_seq)
1603{
1604	if (next_dup == NULL)
1605		return skb;
1606
1607	if (before(next_dup->start_seq, skip_to_seq)) {
1608		skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq);
1609		skb = tcp_sacktag_walk(skb, sk, NULL, state,
1610				       next_dup->start_seq, next_dup->end_seq,
1611				       1);
1612	}
1613
1614	return skb;
1615}
1616
1617static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache)
1618{
1619	return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1620}
1621
1622static int
1623tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb,
1624			u32 prior_snd_una, long *sack_rtt_us)
1625{
1626	struct tcp_sock *tp = tcp_sk(sk);
1627	const unsigned char *ptr = (skb_transport_header(ack_skb) +
1628				    TCP_SKB_CB(ack_skb)->sacked);
1629	struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2);
1630	struct tcp_sack_block sp[TCP_NUM_SACKS];
1631	struct tcp_sack_block *cache;
1632	struct tcp_sacktag_state state;
1633	struct sk_buff *skb;
1634	int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3);
1635	int used_sacks;
1636	bool found_dup_sack = false;
1637	int i, j;
1638	int first_sack_index;
1639
1640	state.flag = 0;
1641	state.reord = tp->packets_out;
1642	state.rtt_us = -1L;
1643
1644	if (!tp->sacked_out) {
1645		if (WARN_ON(tp->fackets_out))
1646			tp->fackets_out = 0;
1647		tcp_highest_sack_reset(sk);
1648	}
1649
1650	found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire,
1651					 num_sacks, prior_snd_una);
1652	if (found_dup_sack)
1653		state.flag |= FLAG_DSACKING_ACK;
1654
1655	/* Eliminate too old ACKs, but take into
1656	 * account more or less fresh ones, they can
1657	 * contain valid SACK info.
1658	 */
1659	if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window))
1660		return 0;
1661
1662	if (!tp->packets_out)
1663		goto out;
1664
1665	used_sacks = 0;
1666	first_sack_index = 0;
1667	for (i = 0; i < num_sacks; i++) {
1668		bool dup_sack = !i && found_dup_sack;
1669
1670		sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq);
1671		sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq);
1672
1673		if (!tcp_is_sackblock_valid(tp, dup_sack,
1674					    sp[used_sacks].start_seq,
1675					    sp[used_sacks].end_seq)) {
1676			int mib_idx;
1677
1678			if (dup_sack) {
1679				if (!tp->undo_marker)
1680					mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO;
1681				else
1682					mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD;
1683			} else {
1684				/* Don't count olds caused by ACK reordering */
1685				if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) &&
1686				    !after(sp[used_sacks].end_seq, tp->snd_una))
1687					continue;
1688				mib_idx = LINUX_MIB_TCPSACKDISCARD;
1689			}
1690
1691			NET_INC_STATS_BH(sock_net(sk), mib_idx);
1692			if (i == 0)
1693				first_sack_index = -1;
1694			continue;
1695		}
1696
1697		/* Ignore very old stuff early */
1698		if (!after(sp[used_sacks].end_seq, prior_snd_una))
 
 
1699			continue;
 
1700
1701		used_sacks++;
1702	}
1703
1704	/* order SACK blocks to allow in order walk of the retrans queue */
1705	for (i = used_sacks - 1; i > 0; i--) {
1706		for (j = 0; j < i; j++) {
1707			if (after(sp[j].start_seq, sp[j + 1].start_seq)) {
1708				swap(sp[j], sp[j + 1]);
1709
1710				/* Track where the first SACK block goes to */
1711				if (j == first_sack_index)
1712					first_sack_index = j + 1;
1713			}
1714		}
1715	}
1716
1717	skb = tcp_write_queue_head(sk);
1718	state.fack_count = 0;
1719	i = 0;
1720
1721	if (!tp->sacked_out) {
1722		/* It's already past, so skip checking against it */
1723		cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache);
1724	} else {
1725		cache = tp->recv_sack_cache;
1726		/* Skip empty blocks in at head of the cache */
1727		while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq &&
1728		       !cache->end_seq)
1729			cache++;
1730	}
1731
1732	while (i < used_sacks) {
1733		u32 start_seq = sp[i].start_seq;
1734		u32 end_seq = sp[i].end_seq;
1735		bool dup_sack = (found_dup_sack && (i == first_sack_index));
1736		struct tcp_sack_block *next_dup = NULL;
1737
1738		if (found_dup_sack && ((i + 1) == first_sack_index))
1739			next_dup = &sp[i + 1];
1740
1741		/* Skip too early cached blocks */
1742		while (tcp_sack_cache_ok(tp, cache) &&
1743		       !before(start_seq, cache->end_seq))
1744			cache++;
1745
1746		/* Can skip some work by looking recv_sack_cache? */
1747		if (tcp_sack_cache_ok(tp, cache) && !dup_sack &&
1748		    after(end_seq, cache->start_seq)) {
1749
1750			/* Head todo? */
1751			if (before(start_seq, cache->start_seq)) {
1752				skb = tcp_sacktag_skip(skb, sk, &state,
1753						       start_seq);
1754				skb = tcp_sacktag_walk(skb, sk, next_dup,
1755						       &state,
1756						       start_seq,
1757						       cache->start_seq,
1758						       dup_sack);
1759			}
1760
1761			/* Rest of the block already fully processed? */
1762			if (!after(end_seq, cache->end_seq))
1763				goto advance_sp;
1764
1765			skb = tcp_maybe_skipping_dsack(skb, sk, next_dup,
1766						       &state,
1767						       cache->end_seq);
1768
1769			/* ...tail remains todo... */
1770			if (tcp_highest_sack_seq(tp) == cache->end_seq) {
1771				/* ...but better entrypoint exists! */
1772				skb = tcp_highest_sack(sk);
1773				if (skb == NULL)
1774					break;
1775				state.fack_count = tp->fackets_out;
1776				cache++;
1777				goto walk;
1778			}
1779
1780			skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq);
1781			/* Check overlap against next cached too (past this one already) */
1782			cache++;
1783			continue;
1784		}
1785
1786		if (!before(start_seq, tcp_highest_sack_seq(tp))) {
1787			skb = tcp_highest_sack(sk);
1788			if (skb == NULL)
1789				break;
1790			state.fack_count = tp->fackets_out;
1791		}
1792		skb = tcp_sacktag_skip(skb, sk, &state, start_seq);
1793
1794walk:
1795		skb = tcp_sacktag_walk(skb, sk, next_dup, &state,
1796				       start_seq, end_seq, dup_sack);
1797
1798advance_sp:
1799		i++;
1800	}
1801
1802	/* Clear the head of the cache sack blocks so we can skip it next time */
1803	for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) {
1804		tp->recv_sack_cache[i].start_seq = 0;
1805		tp->recv_sack_cache[i].end_seq = 0;
1806	}
1807	for (j = 0; j < used_sacks; j++)
1808		tp->recv_sack_cache[i++] = sp[j];
1809
1810	tcp_mark_lost_retrans(sk);
 
1811
1812	tcp_verify_left_out(tp);
1813
1814	if ((state.reord < tp->fackets_out) &&
1815	    ((inet_csk(sk)->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker))
1816		tcp_update_reordering(sk, tp->fackets_out - state.reord, 0);
1817
1818out:
1819
1820#if FASTRETRANS_DEBUG > 0
1821	WARN_ON((int)tp->sacked_out < 0);
1822	WARN_ON((int)tp->lost_out < 0);
1823	WARN_ON((int)tp->retrans_out < 0);
1824	WARN_ON((int)tcp_packets_in_flight(tp) < 0);
1825#endif
1826	*sack_rtt_us = state.rtt_us;
1827	return state.flag;
1828}
1829
1830/* Limits sacked_out so that sum with lost_out isn't ever larger than
1831 * packets_out. Returns false if sacked_out adjustement wasn't necessary.
1832 */
1833static bool tcp_limit_reno_sacked(struct tcp_sock *tp)
1834{
1835	u32 holes;
1836
1837	holes = max(tp->lost_out, 1U);
1838	holes = min(holes, tp->packets_out);
1839
1840	if ((tp->sacked_out + holes) > tp->packets_out) {
1841		tp->sacked_out = tp->packets_out - holes;
1842		return true;
1843	}
1844	return false;
1845}
1846
1847/* If we receive more dupacks than we expected counting segments
1848 * in assumption of absent reordering, interpret this as reordering.
1849 * The only another reason could be bug in receiver TCP.
1850 */
1851static void tcp_check_reno_reordering(struct sock *sk, const int addend)
1852{
1853	struct tcp_sock *tp = tcp_sk(sk);
1854	if (tcp_limit_reno_sacked(tp))
1855		tcp_update_reordering(sk, tp->packets_out + addend, 0);
 
 
 
 
 
 
1856}
1857
1858/* Emulate SACKs for SACKless connection: account for a new dupack. */
1859
1860static void tcp_add_reno_sack(struct sock *sk)
1861{
1862	struct tcp_sock *tp = tcp_sk(sk);
1863	tp->sacked_out++;
1864	tcp_check_reno_reordering(sk, 0);
1865	tcp_verify_left_out(tp);
 
 
 
 
 
 
 
 
1866}
1867
1868/* Account for ACK, ACKing some data in Reno Recovery phase. */
1869
1870static void tcp_remove_reno_sacks(struct sock *sk, int acked)
1871{
1872	struct tcp_sock *tp = tcp_sk(sk);
1873
1874	if (acked > 0) {
1875		/* One ACK acked hole. The rest eat duplicate ACKs. */
 
 
1876		if (acked - 1 >= tp->sacked_out)
1877			tp->sacked_out = 0;
1878		else
1879			tp->sacked_out -= acked - 1;
1880	}
1881	tcp_check_reno_reordering(sk, acked);
1882	tcp_verify_left_out(tp);
1883}
1884
1885static inline void tcp_reset_reno_sack(struct tcp_sock *tp)
1886{
1887	tp->sacked_out = 0;
1888}
1889
1890static void tcp_clear_retrans_partial(struct tcp_sock *tp)
1891{
1892	tp->retrans_out = 0;
1893	tp->lost_out = 0;
1894
1895	tp->undo_marker = 0;
1896	tp->undo_retrans = 0;
 
 
 
 
 
1897}
1898
1899void tcp_clear_retrans(struct tcp_sock *tp)
1900{
1901	tcp_clear_retrans_partial(tp);
1902
1903	tp->fackets_out = 0;
1904	tp->sacked_out = 0;
 
 
 
 
 
 
 
1905}
1906
1907/* Enter Loss state. If "how" is not zero, forget all SACK information
 
 
 
 
 
 
1908 * and reset tags completely, otherwise preserve SACKs. If receiver
1909 * dropped its ofo queue, we will know this due to reneging detection.
1910 */
1911void tcp_enter_loss(struct sock *sk, int how)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1912{
1913	const struct inet_connection_sock *icsk = inet_csk(sk);
1914	struct tcp_sock *tp = tcp_sk(sk);
1915	struct sk_buff *skb;
1916	bool new_recovery = false;
 
 
 
1917
1918	/* Reduce ssthresh if it has not yet been made inside this window. */
1919	if (icsk->icsk_ca_state <= TCP_CA_Disorder ||
1920	    !after(tp->high_seq, tp->snd_una) ||
1921	    (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) {
1922		new_recovery = true;
1923		tp->prior_ssthresh = tcp_current_ssthresh(sk);
 
1924		tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk);
1925		tcp_ca_event(sk, CA_EVENT_LOSS);
 
1926	}
1927	tp->snd_cwnd	   = 1;
1928	tp->snd_cwnd_cnt   = 0;
1929	tp->snd_cwnd_stamp = tcp_time_stamp;
1930
1931	tcp_clear_retrans_partial(tp);
1932
1933	if (tcp_is_reno(tp))
1934		tcp_reset_reno_sack(tp);
1935
1936	tp->undo_marker = tp->snd_una;
1937	if (how) {
1938		tp->sacked_out = 0;
1939		tp->fackets_out = 0;
1940	}
1941	tcp_clear_all_retrans_hints(tp);
1942
1943	tcp_for_write_queue(skb, sk) {
1944		if (skb == tcp_send_head(sk))
1945			break;
1946
1947		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1948			tp->undo_marker = 0;
1949
1950		TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED;
1951		if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) {
1952			TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED;
1953			TCP_SKB_CB(skb)->sacked |= TCPCB_LOST;
1954			tp->lost_out += tcp_skb_pcount(skb);
1955			tp->retransmit_high = TCP_SKB_CB(skb)->end_seq;
1956		}
1957	}
1958	tcp_verify_left_out(tp);
1959
1960	/* Timeout in disordered state after receiving substantial DUPACKs
1961	 * suggests that the degree of reordering is over-estimated.
1962	 */
 
1963	if (icsk->icsk_ca_state <= TCP_CA_Disorder &&
1964	    tp->sacked_out >= sysctl_tcp_reordering)
1965		tp->reordering = min_t(unsigned int, tp->reordering,
1966				       sysctl_tcp_reordering);
 
1967	tcp_set_ca_state(sk, TCP_CA_Loss);
1968	tp->high_seq = tp->snd_nxt;
1969	TCP_ECN_queue_cwr(tp);
 
1970
1971	/* F-RTO RFC5682 sec 3.1 step 1: retransmit SND.UNA if no previous
1972	 * loss recovery is underway except recurring timeout(s) on
1973	 * the same SND.UNA (sec 3.2). Disable F-RTO on path MTU probing
1974	 */
1975	tp->frto = sysctl_tcp_frto &&
1976		   (new_recovery || icsk->icsk_retransmits) &&
1977		   !inet_csk(sk)->icsk_mtup.probe_size;
1978}
1979
1980/* If ACK arrived pointing to a remembered SACK, it means that our
1981 * remembered SACKs do not reflect real state of receiver i.e.
1982 * receiver _host_ is heavily congested (or buggy).
1983 *
1984 * Do processing similar to RTO timeout.
1985 */
1986static bool tcp_check_sack_reneging(struct sock *sk, int flag)
1987{
1988	if (flag & FLAG_SACK_RENEGING) {
1989		struct inet_connection_sock *icsk = inet_csk(sk);
1990		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING);
 
 
 
 
 
 
1991
1992		tcp_enter_loss(sk, 1);
1993		icsk->icsk_retransmits++;
1994		tcp_retransmit_skb(sk, tcp_write_queue_head(sk));
1995		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
1996					  icsk->icsk_rto, TCP_RTO_MAX);
 
1997		return true;
1998	}
1999	return false;
2000}
2001
2002static inline int tcp_fackets_out(const struct tcp_sock *tp)
2003{
2004	return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out;
2005}
2006
2007/* Heurestics to calculate number of duplicate ACKs. There's no dupACKs
2008 * counter when SACK is enabled (without SACK, sacked_out is used for
2009 * that purpose).
2010 *
2011 * Instead, with FACK TCP uses fackets_out that includes both SACKed
2012 * segments up to the highest received SACK block so far and holes in
2013 * between them.
2014 *
2015 * With reordering, holes may still be in flight, so RFC3517 recovery
2016 * uses pure sacked_out (total number of SACKed segments) even though
2017 * it violates the RFC that uses duplicate ACKs, often these are equal
2018 * but when e.g. out-of-window ACKs or packet duplication occurs,
2019 * they differ. Since neither occurs due to loss, TCP should really
2020 * ignore them.
2021 */
2022static inline int tcp_dupack_heuristics(const struct tcp_sock *tp)
2023{
2024	return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1;
2025}
2026
2027static bool tcp_pause_early_retransmit(struct sock *sk, int flag)
2028{
2029	struct tcp_sock *tp = tcp_sk(sk);
2030	unsigned long delay;
2031
2032	/* Delay early retransmit and entering fast recovery for
2033	 * max(RTT/4, 2msec) unless ack has ECE mark, no RTT samples
2034	 * available, or RTO is scheduled to fire first.
2035	 */
2036	if (sysctl_tcp_early_retrans < 2 || sysctl_tcp_early_retrans > 3 ||
2037	    (flag & FLAG_ECE) || !tp->srtt_us)
2038		return false;
2039
2040	delay = max(usecs_to_jiffies(tp->srtt_us >> 5),
2041		    msecs_to_jiffies(2));
2042
2043	if (!time_after(inet_csk(sk)->icsk_timeout, (jiffies + delay)))
2044		return false;
2045
2046	inet_csk_reset_xmit_timer(sk, ICSK_TIME_EARLY_RETRANS, delay,
2047				  TCP_RTO_MAX);
2048	return true;
2049}
2050
2051/* Linux NewReno/SACK/FACK/ECN state machine.
2052 * --------------------------------------
2053 *
2054 * "Open"	Normal state, no dubious events, fast path.
2055 * "Disorder"   In all the respects it is "Open",
2056 *		but requires a bit more attention. It is entered when
2057 *		we see some SACKs or dupacks. It is split of "Open"
2058 *		mainly to move some processing from fast path to slow one.
2059 * "CWR"	CWND was reduced due to some Congestion Notification event.
2060 *		It can be ECN, ICMP source quench, local device congestion.
2061 * "Recovery"	CWND was reduced, we are fast-retransmitting.
2062 * "Loss"	CWND was reduced due to RTO timeout or SACK reneging.
2063 *
2064 * tcp_fastretrans_alert() is entered:
2065 * - each incoming ACK, if state is not "Open"
2066 * - when arrived ACK is unusual, namely:
2067 *	* SACK
2068 *	* Duplicate ACK.
2069 *	* ECN ECE.
2070 *
2071 * Counting packets in flight is pretty simple.
2072 *
2073 *	in_flight = packets_out - left_out + retrans_out
2074 *
2075 *	packets_out is SND.NXT-SND.UNA counted in packets.
2076 *
2077 *	retrans_out is number of retransmitted segments.
2078 *
2079 *	left_out is number of segments left network, but not ACKed yet.
2080 *
2081 *		left_out = sacked_out + lost_out
2082 *
2083 *     sacked_out: Packets, which arrived to receiver out of order
2084 *		   and hence not ACKed. With SACKs this number is simply
2085 *		   amount of SACKed data. Even without SACKs
2086 *		   it is easy to give pretty reliable estimate of this number,
2087 *		   counting duplicate ACKs.
2088 *
2089 *       lost_out: Packets lost by network. TCP has no explicit
2090 *		   "loss notification" feedback from network (for now).
2091 *		   It means that this number can be only _guessed_.
2092 *		   Actually, it is the heuristics to predict lossage that
2093 *		   distinguishes different algorithms.
2094 *
2095 *	F.e. after RTO, when all the queue is considered as lost,
2096 *	lost_out = packets_out and in_flight = retrans_out.
2097 *
2098 *		Essentially, we have now two algorithms counting
2099 *		lost packets.
2100 *
2101 *		FACK: It is the simplest heuristics. As soon as we decided
2102 *		that something is lost, we decide that _all_ not SACKed
2103 *		packets until the most forward SACK are lost. I.e.
2104 *		lost_out = fackets_out - sacked_out and left_out = fackets_out.
2105 *		It is absolutely correct estimate, if network does not reorder
2106 *		packets. And it loses any connection to reality when reordering
2107 *		takes place. We use FACK by default until reordering
2108 *		is suspected on the path to this destination.
2109 *
2110 *		NewReno: when Recovery is entered, we assume that one segment
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2111 *		is lost (classic Reno). While we are in Recovery and
2112 *		a partial ACK arrives, we assume that one more packet
2113 *		is lost (NewReno). This heuristics are the same in NewReno
2114 *		and SACK.
2115 *
2116 *  Imagine, that's all! Forget about all this shamanism about CWND inflation
2117 *  deflation etc. CWND is real congestion window, never inflated, changes
2118 *  only according to classic VJ rules.
2119 *
2120 * Really tricky (and requiring careful tuning) part of algorithm
2121 * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue().
2122 * The first determines the moment _when_ we should reduce CWND and,
2123 * hence, slow down forward transmission. In fact, it determines the moment
2124 * when we decide that hole is caused by loss, rather than by a reorder.
2125 *
2126 * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill
2127 * holes, caused by lost packets.
2128 *
2129 * And the most logically complicated part of algorithm is undo
2130 * heuristics. We detect false retransmits due to both too early
2131 * fast retransmit (reordering) and underestimated RTO, analyzing
2132 * timestamps and D-SACKs. When we detect that some segments were
2133 * retransmitted by mistake and CWND reduction was wrong, we undo
2134 * window reduction and abort recovery phase. This logic is hidden
2135 * inside several functions named tcp_try_undo_<something>.
2136 */
2137
2138/* This function decides, when we should leave Disordered state
2139 * and enter Recovery phase, reducing congestion window.
2140 *
2141 * Main question: may we further continue forward transmission
2142 * with the same cwnd?
2143 */
2144static bool tcp_time_to_recover(struct sock *sk, int flag)
2145{
2146	struct tcp_sock *tp = tcp_sk(sk);
2147	__u32 packets_out;
2148
2149	/* Trick#1: The loss is proven. */
2150	if (tp->lost_out)
2151		return true;
2152
2153	/* Not-A-Trick#2 : Classic rule... */
2154	if (tcp_dupack_heuristics(tp) > tp->reordering)
2155		return true;
2156
2157	/* Trick#4: It is still not OK... But will it be useful to delay
2158	 * recovery more?
2159	 */
2160	packets_out = tp->packets_out;
2161	if (packets_out <= tp->reordering &&
2162	    tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) &&
2163	    !tcp_may_send_now(sk)) {
2164		/* We have nothing to send. This connection is limited
2165		 * either by receiver window or by application.
2166		 */
2167		return true;
2168	}
2169
2170	/* If a thin stream is detected, retransmit after first
2171	 * received dupack. Employ only if SACK is supported in order
2172	 * to avoid possible corner-case series of spurious retransmissions
2173	 * Use only if there are no unsent data.
2174	 */
2175	if ((tp->thin_dupack || sysctl_tcp_thin_dupack) &&
2176	    tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 &&
2177	    tcp_is_sack(tp) && !tcp_send_head(sk))
2178		return true;
2179
2180	/* Trick#6: TCP early retransmit, per RFC5827.  To avoid spurious
2181	 * retransmissions due to small network reorderings, we implement
2182	 * Mitigation A.3 in the RFC and delay the retransmission for a short
2183	 * interval if appropriate.
2184	 */
2185	if (tp->do_early_retrans && !tp->retrans_out && tp->sacked_out &&
2186	    (tp->packets_out >= (tp->sacked_out + 1) && tp->packets_out < 4) &&
2187	    !tcp_may_send_now(sk))
2188		return !tcp_pause_early_retransmit(sk, flag);
2189
2190	return false;
2191}
2192
2193/* Detect loss in event "A" above by marking head of queue up as lost.
2194 * For FACK or non-SACK(Reno) senders, the first "packets" number of segments
2195 * are considered lost. For RFC3517 SACK, a segment is considered lost if it
2196 * has at least tp->reordering SACKed seqments above it; "packets" refers to
2197 * the maximum SACKed segments to pass before reaching this limit.
2198 */
2199static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head)
2200{
2201	struct tcp_sock *tp = tcp_sk(sk);
2202	struct sk_buff *skb;
2203	int cnt, oldcnt;
2204	int err;
2205	unsigned int mss;
2206	/* Use SACK to deduce losses of new sequences sent during recovery */
2207	const u32 loss_high = tcp_is_sack(tp) ?  tp->snd_nxt : tp->high_seq;
2208
2209	WARN_ON(packets > tp->packets_out);
2210	if (tp->lost_skb_hint) {
2211		skb = tp->lost_skb_hint;
2212		cnt = tp->lost_cnt_hint;
2213		/* Head already handled? */
2214		if (mark_head && skb != tcp_write_queue_head(sk))
2215			return;
 
2216	} else {
2217		skb = tcp_write_queue_head(sk);
2218		cnt = 0;
2219	}
2220
2221	tcp_for_write_queue_from(skb, sk) {
2222		if (skb == tcp_send_head(sk))
2223			break;
2224		/* TODO: do this better */
2225		/* this is not the most efficient way to do this... */
2226		tp->lost_skb_hint = skb;
2227		tp->lost_cnt_hint = cnt;
2228
2229		if (after(TCP_SKB_CB(skb)->end_seq, loss_high))
2230			break;
2231
2232		oldcnt = cnt;
2233		if (tcp_is_fack(tp) || tcp_is_reno(tp) ||
2234		    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
2235			cnt += tcp_skb_pcount(skb);
2236
2237		if (cnt > packets) {
2238			if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) ||
2239			    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) ||
2240			    (oldcnt >= packets))
2241				break;
2242
2243			mss = skb_shinfo(skb)->gso_size;
2244			err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss);
2245			if (err < 0)
2246				break;
2247			cnt = packets;
2248		}
2249
2250		tcp_skb_mark_lost(tp, skb);
 
2251
2252		if (mark_head)
2253			break;
2254	}
2255	tcp_verify_left_out(tp);
2256}
2257
2258/* Account newly detected lost packet(s) */
2259
2260static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit)
2261{
2262	struct tcp_sock *tp = tcp_sk(sk);
2263
2264	if (tcp_is_reno(tp)) {
2265		tcp_mark_head_lost(sk, 1, 1);
2266	} else if (tcp_is_fack(tp)) {
2267		int lost = tp->fackets_out - tp->reordering;
2268		if (lost <= 0)
2269			lost = 1;
2270		tcp_mark_head_lost(sk, lost, 0);
2271	} else {
2272		int sacked_upto = tp->sacked_out - tp->reordering;
2273		if (sacked_upto >= 0)
2274			tcp_mark_head_lost(sk, sacked_upto, 0);
2275		else if (fast_rexmit)
2276			tcp_mark_head_lost(sk, 1, 1);
2277	}
2278}
2279
2280/* CWND moderation, preventing bursts due to too big ACKs
2281 * in dubious situations.
 
 
 
 
 
 
2282 */
2283static inline void tcp_moderate_cwnd(struct tcp_sock *tp)
 
2284{
2285	tp->snd_cwnd = min(tp->snd_cwnd,
2286			   tcp_packets_in_flight(tp) + tcp_max_burst(tp));
2287	tp->snd_cwnd_stamp = tcp_time_stamp;
2288}
2289
2290/* Nothing was retransmitted or returned timestamp is less
2291 * than timestamp of the first retransmission.
2292 */
2293static inline bool tcp_packet_delayed(const struct tcp_sock *tp)
2294{
2295	return !tp->retrans_stamp ||
2296		(tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2297		 before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp));
 
 
 
 
 
 
 
 
 
 
 
 
 
2298}
2299
2300/* Undo procedures. */
2301
2302#if FASTRETRANS_DEBUG > 1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2303static void DBGUNDO(struct sock *sk, const char *msg)
2304{
 
2305	struct tcp_sock *tp = tcp_sk(sk);
2306	struct inet_sock *inet = inet_sk(sk);
2307
2308	if (sk->sk_family == AF_INET) {
2309		pr_debug("Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n",
2310			 msg,
2311			 &inet->inet_daddr, ntohs(inet->inet_dport),
2312			 tp->snd_cwnd, tcp_left_out(tp),
2313			 tp->snd_ssthresh, tp->prior_ssthresh,
2314			 tp->packets_out);
2315	}
2316#if IS_ENABLED(CONFIG_IPV6)
2317	else if (sk->sk_family == AF_INET6) {
2318		struct ipv6_pinfo *np = inet6_sk(sk);
2319		pr_debug("Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n",
2320			 msg,
2321			 &np->daddr, ntohs(inet->inet_dport),
2322			 tp->snd_cwnd, tcp_left_out(tp),
2323			 tp->snd_ssthresh, tp->prior_ssthresh,
2324			 tp->packets_out);
2325	}
2326#endif
2327}
2328#else
2329#define DBGUNDO(x...) do { } while (0)
2330#endif
 
2331
2332static void tcp_undo_cwnd_reduction(struct sock *sk, bool unmark_loss)
2333{
2334	struct tcp_sock *tp = tcp_sk(sk);
2335
2336	if (unmark_loss) {
2337		struct sk_buff *skb;
2338
2339		tcp_for_write_queue(skb, sk) {
2340			if (skb == tcp_send_head(sk))
2341				break;
2342			TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST;
2343		}
2344		tp->lost_out = 0;
2345		tcp_clear_all_retrans_hints(tp);
2346	}
2347
2348	if (tp->prior_ssthresh) {
2349		const struct inet_connection_sock *icsk = inet_csk(sk);
2350
2351		if (icsk->icsk_ca_ops->undo_cwnd)
2352			tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk);
2353		else
2354			tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1);
2355
2356		if (tp->prior_ssthresh > tp->snd_ssthresh) {
2357			tp->snd_ssthresh = tp->prior_ssthresh;
2358			TCP_ECN_withdraw_cwr(tp);
2359		}
2360	} else {
2361		tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh);
2362	}
2363	tp->snd_cwnd_stamp = tcp_time_stamp;
2364	tp->undo_marker = 0;
 
2365}
2366
2367static inline bool tcp_may_undo(const struct tcp_sock *tp)
2368{
2369	return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp));
2370}
2371
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2372/* People celebrate: "We love our President!" */
2373static bool tcp_try_undo_recovery(struct sock *sk)
2374{
2375	struct tcp_sock *tp = tcp_sk(sk);
2376
2377	if (tcp_may_undo(tp)) {
2378		int mib_idx;
2379
2380		/* Happy end! We did not retransmit anything
2381		 * or our original transmission succeeded.
2382		 */
2383		DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans");
2384		tcp_undo_cwnd_reduction(sk, false);
2385		if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss)
2386			mib_idx = LINUX_MIB_TCPLOSSUNDO;
2387		else
2388			mib_idx = LINUX_MIB_TCPFULLUNDO;
2389
2390		NET_INC_STATS_BH(sock_net(sk), mib_idx);
 
 
2391	}
2392	if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) {
2393		/* Hold old state until something *above* high_seq
2394		 * is ACKed. For Reno it is MUST to prevent false
2395		 * fast retransmits (RFC2582). SACK TCP is safe. */
2396		tcp_moderate_cwnd(tp);
2397		return true;
2398	}
2399	tcp_set_ca_state(sk, TCP_CA_Open);
 
2400	return false;
2401}
2402
2403/* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */
2404static bool tcp_try_undo_dsack(struct sock *sk)
2405{
2406	struct tcp_sock *tp = tcp_sk(sk);
2407
2408	if (tp->undo_marker && !tp->undo_retrans) {
 
 
2409		DBGUNDO(sk, "D-SACK");
2410		tcp_undo_cwnd_reduction(sk, false);
2411		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO);
2412		return true;
2413	}
2414	return false;
2415}
2416
2417/* We can clear retrans_stamp when there are no retransmissions in the
2418 * window. It would seem that it is trivially available for us in
2419 * tp->retrans_out, however, that kind of assumptions doesn't consider
2420 * what will happen if errors occur when sending retransmission for the
2421 * second time. ...It could the that such segment has only
2422 * TCPCB_EVER_RETRANS set at the present time. It seems that checking
2423 * the head skb is enough except for some reneging corner cases that
2424 * are not worth the effort.
2425 *
2426 * Main reason for all this complexity is the fact that connection dying
2427 * time now depends on the validity of the retrans_stamp, in particular,
2428 * that successive retransmissions of a segment must not advance
2429 * retrans_stamp under any conditions.
2430 */
2431static bool tcp_any_retrans_done(const struct sock *sk)
2432{
2433	const struct tcp_sock *tp = tcp_sk(sk);
2434	struct sk_buff *skb;
2435
2436	if (tp->retrans_out)
2437		return true;
2438
2439	skb = tcp_write_queue_head(sk);
2440	if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS))
2441		return true;
2442
2443	return false;
2444}
2445
2446/* Undo during loss recovery after partial ACK or using F-RTO. */
2447static bool tcp_try_undo_loss(struct sock *sk, bool frto_undo)
2448{
2449	struct tcp_sock *tp = tcp_sk(sk);
2450
2451	if (frto_undo || tcp_may_undo(tp)) {
2452		tcp_undo_cwnd_reduction(sk, true);
2453
2454		DBGUNDO(sk, "partial loss");
2455		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO);
2456		if (frto_undo)
2457			NET_INC_STATS_BH(sock_net(sk),
2458					 LINUX_MIB_TCPSPURIOUSRTOS);
2459		inet_csk(sk)->icsk_retransmits = 0;
2460		if (frto_undo || tcp_is_sack(tp))
 
 
2461			tcp_set_ca_state(sk, TCP_CA_Open);
 
 
2462		return true;
2463	}
2464	return false;
2465}
2466
2467/* The cwnd reduction in CWR and Recovery use the PRR algorithm
2468 * https://datatracker.ietf.org/doc/draft-ietf-tcpm-proportional-rate-reduction/
2469 * It computes the number of packets to send (sndcnt) based on packets newly
2470 * delivered:
2471 *   1) If the packets in flight is larger than ssthresh, PRR spreads the
2472 *	cwnd reductions across a full RTT.
2473 *   2) If packets in flight is lower than ssthresh (such as due to excess
2474 *	losses and/or application stalls), do not perform any further cwnd
2475 *	reductions, but instead slow start up to ssthresh.
2476 */
2477static void tcp_init_cwnd_reduction(struct sock *sk, const bool set_ssthresh)
2478{
2479	struct tcp_sock *tp = tcp_sk(sk);
2480
2481	tp->high_seq = tp->snd_nxt;
2482	tp->tlp_high_seq = 0;
2483	tp->snd_cwnd_cnt = 0;
2484	tp->prior_cwnd = tp->snd_cwnd;
2485	tp->prr_delivered = 0;
2486	tp->prr_out = 0;
2487	if (set_ssthresh)
2488		tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk);
2489	TCP_ECN_queue_cwr(tp);
2490}
2491
2492static void tcp_cwnd_reduction(struct sock *sk, const int prior_unsacked,
2493			       int fast_rexmit)
2494{
2495	struct tcp_sock *tp = tcp_sk(sk);
2496	int sndcnt = 0;
2497	int delta = tp->snd_ssthresh - tcp_packets_in_flight(tp);
2498	int newly_acked_sacked = prior_unsacked -
2499				 (tp->packets_out - tp->sacked_out);
 
2500
2501	tp->prr_delivered += newly_acked_sacked;
2502	if (tcp_packets_in_flight(tp) > tp->snd_ssthresh) {
2503		u64 dividend = (u64)tp->snd_ssthresh * tp->prr_delivered +
2504			       tp->prior_cwnd - 1;
2505		sndcnt = div_u64(dividend, tp->prior_cwnd) - tp->prr_out;
2506	} else {
2507		sndcnt = min_t(int, delta,
2508			       max_t(int, tp->prr_delivered - tp->prr_out,
2509				     newly_acked_sacked) + 1);
2510	}
2511
2512	sndcnt = max(sndcnt, (fast_rexmit ? 1 : 0));
2513	tp->snd_cwnd = tcp_packets_in_flight(tp) + sndcnt;
 
 
2514}
2515
2516static inline void tcp_end_cwnd_reduction(struct sock *sk)
2517{
2518	struct tcp_sock *tp = tcp_sk(sk);
2519
 
 
 
2520	/* Reset cwnd to ssthresh in CWR or Recovery (unless it's undone) */
2521	if (inet_csk(sk)->icsk_ca_state == TCP_CA_CWR ||
2522	    (tp->undo_marker && tp->snd_ssthresh < TCP_INFINITE_SSTHRESH)) {
2523		tp->snd_cwnd = tp->snd_ssthresh;
2524		tp->snd_cwnd_stamp = tcp_time_stamp;
2525	}
2526	tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR);
2527}
2528
2529/* Enter CWR state. Disable cwnd undo since congestion is proven with ECN */
2530void tcp_enter_cwr(struct sock *sk, const int set_ssthresh)
2531{
2532	struct tcp_sock *tp = tcp_sk(sk);
2533
2534	tp->prior_ssthresh = 0;
2535	if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2536		tp->undo_marker = 0;
2537		tcp_init_cwnd_reduction(sk, set_ssthresh);
2538		tcp_set_ca_state(sk, TCP_CA_CWR);
2539	}
2540}
 
2541
2542static void tcp_try_keep_open(struct sock *sk)
2543{
2544	struct tcp_sock *tp = tcp_sk(sk);
2545	int state = TCP_CA_Open;
2546
2547	if (tcp_left_out(tp) || tcp_any_retrans_done(sk))
2548		state = TCP_CA_Disorder;
2549
2550	if (inet_csk(sk)->icsk_ca_state != state) {
2551		tcp_set_ca_state(sk, state);
2552		tp->high_seq = tp->snd_nxt;
2553	}
2554}
2555
2556static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked)
2557{
2558	struct tcp_sock *tp = tcp_sk(sk);
2559
2560	tcp_verify_left_out(tp);
2561
2562	if (!tcp_any_retrans_done(sk))
2563		tp->retrans_stamp = 0;
2564
2565	if (flag & FLAG_ECE)
2566		tcp_enter_cwr(sk, 1);
2567
2568	if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) {
2569		tcp_try_keep_open(sk);
2570	} else {
2571		tcp_cwnd_reduction(sk, prior_unsacked, 0);
2572	}
2573}
2574
2575static void tcp_mtup_probe_failed(struct sock *sk)
2576{
2577	struct inet_connection_sock *icsk = inet_csk(sk);
2578
2579	icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1;
2580	icsk->icsk_mtup.probe_size = 0;
 
2581}
2582
2583static void tcp_mtup_probe_success(struct sock *sk)
2584{
2585	struct tcp_sock *tp = tcp_sk(sk);
2586	struct inet_connection_sock *icsk = inet_csk(sk);
 
2587
2588	/* FIXME: breaks with very large cwnd */
2589	tp->prior_ssthresh = tcp_current_ssthresh(sk);
2590	tp->snd_cwnd = tp->snd_cwnd *
2591		       tcp_mss_to_mtu(sk, tp->mss_cache) /
2592		       icsk->icsk_mtup.probe_size;
 
 
 
2593	tp->snd_cwnd_cnt = 0;
2594	tp->snd_cwnd_stamp = tcp_time_stamp;
2595	tp->snd_ssthresh = tcp_current_ssthresh(sk);
2596
2597	icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size;
2598	icsk->icsk_mtup.probe_size = 0;
2599	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2600}
2601
2602/* Do a simple retransmit without using the backoff mechanisms in
2603 * tcp_timer. This is used for path mtu discovery.
2604 * The socket is already locked here.
2605 */
2606void tcp_simple_retransmit(struct sock *sk)
2607{
2608	const struct inet_connection_sock *icsk = inet_csk(sk);
2609	struct tcp_sock *tp = tcp_sk(sk);
2610	struct sk_buff *skb;
2611	unsigned int mss = tcp_current_mss(sk);
2612	u32 prior_lost = tp->lost_out;
2613
2614	tcp_for_write_queue(skb, sk) {
2615		if (skb == tcp_send_head(sk))
2616			break;
2617		if (tcp_skb_seglen(skb) > mss &&
2618		    !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) {
2619			if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2620				TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS;
2621				tp->retrans_out -= tcp_skb_pcount(skb);
2622			}
2623			tcp_skb_mark_lost_uncond_verify(tp, skb);
2624		}
 
 
 
 
 
 
 
2625	}
2626
2627	tcp_clear_retrans_hints_partial(tp);
2628
2629	if (prior_lost == tp->lost_out)
2630		return;
2631
2632	if (tcp_is_reno(tp))
2633		tcp_limit_reno_sacked(tp);
2634
2635	tcp_verify_left_out(tp);
2636
2637	/* Don't muck with the congestion window here.
2638	 * Reason is that we do not increase amount of _data_
2639	 * in network, but units changed and effective
2640	 * cwnd/ssthresh really reduced now.
2641	 */
2642	if (icsk->icsk_ca_state != TCP_CA_Loss) {
2643		tp->high_seq = tp->snd_nxt;
2644		tp->snd_ssthresh = tcp_current_ssthresh(sk);
2645		tp->prior_ssthresh = 0;
2646		tp->undo_marker = 0;
2647		tcp_set_ca_state(sk, TCP_CA_Loss);
2648	}
2649	tcp_xmit_retransmit_queue(sk);
2650}
2651EXPORT_SYMBOL(tcp_simple_retransmit);
2652
2653static void tcp_enter_recovery(struct sock *sk, bool ece_ack)
2654{
2655	struct tcp_sock *tp = tcp_sk(sk);
2656	int mib_idx;
2657
 
 
 
2658	if (tcp_is_reno(tp))
2659		mib_idx = LINUX_MIB_TCPRENORECOVERY;
2660	else
2661		mib_idx = LINUX_MIB_TCPSACKRECOVERY;
2662
2663	NET_INC_STATS_BH(sock_net(sk), mib_idx);
2664
2665	tp->prior_ssthresh = 0;
2666	tp->undo_marker = tp->snd_una;
2667	tp->undo_retrans = tp->retrans_out;
2668
2669	if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) {
2670		if (!ece_ack)
2671			tp->prior_ssthresh = tcp_current_ssthresh(sk);
2672		tcp_init_cwnd_reduction(sk, true);
2673	}
2674	tcp_set_ca_state(sk, TCP_CA_Recovery);
2675}
2676
 
 
 
 
 
 
 
 
2677/* Process an ACK in CA_Loss state. Move to CA_Open if lost data are
2678 * recovered or spurious. Otherwise retransmits more on partial ACKs.
2679 */
2680static void tcp_process_loss(struct sock *sk, int flag, bool is_dupack)
 
2681{
2682	struct inet_connection_sock *icsk = inet_csk(sk);
2683	struct tcp_sock *tp = tcp_sk(sk);
2684	bool recovered = !before(tp->snd_una, tp->high_seq);
2685
 
 
 
 
2686	if (tp->frto) { /* F-RTO RFC5682 sec 3.1 (sack enhanced version). */
2687		/* Step 3.b. A timeout is spurious if not all data are
2688		 * lost, i.e., never-retransmitted data are (s)acked.
2689		 */
2690		if (tcp_try_undo_loss(sk, flag & FLAG_ORIG_SACK_ACKED))
 
2691			return;
2692
2693		if (after(tp->snd_nxt, tp->high_seq) &&
2694		    (flag & FLAG_DATA_SACKED || is_dupack)) {
2695			tp->frto = 0; /* Loss was real: 2nd part of step 3.a */
2696		} else if (flag & FLAG_SND_UNA_ADVANCED && !recovered) {
2697			tp->high_seq = tp->snd_nxt;
2698			__tcp_push_pending_frames(sk, tcp_current_mss(sk),
2699						  TCP_NAGLE_OFF);
2700			if (after(tp->snd_nxt, tp->high_seq))
2701				return; /* Step 2.b */
 
 
 
 
 
2702			tp->frto = 0;
2703		}
2704	}
2705
2706	if (recovered) {
2707		/* F-RTO RFC5682 sec 3.1 step 2.a and 1st part of step 3.a */
2708		icsk->icsk_retransmits = 0;
2709		tcp_try_undo_recovery(sk);
2710		return;
2711	}
2712	if (flag & FLAG_DATA_ACKED)
2713		icsk->icsk_retransmits = 0;
2714	if (tcp_is_reno(tp)) {
2715		/* A Reno DUPACK means new data in F-RTO step 2.b above are
2716		 * delivered. Lower inflight to clock out (re)tranmissions.
2717		 */
2718		if (after(tp->snd_nxt, tp->high_seq) && is_dupack)
2719			tcp_add_reno_sack(sk);
2720		else if (flag & FLAG_SND_UNA_ADVANCED)
2721			tcp_reset_reno_sack(tp);
2722	}
2723	if (tcp_try_undo_loss(sk, false))
2724		return;
2725	tcp_xmit_retransmit_queue(sk);
 
 
 
 
 
 
2726}
2727
2728/* Undo during fast recovery after partial ACK. */
2729static bool tcp_try_undo_partial(struct sock *sk, const int acked,
2730				 const int prior_unsacked)
2731{
2732	struct tcp_sock *tp = tcp_sk(sk);
2733
2734	if (tp->undo_marker && tcp_packet_delayed(tp)) {
2735		/* Plain luck! Hole if filled with delayed
2736		 * packet, rather than with a retransmit.
2737		 */
2738		tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1);
2739
2740		/* We are getting evidence that the reordering degree is higher
2741		 * than we realized. If there are no retransmits out then we
2742		 * can undo. Otherwise we clock out new packets but do not
2743		 * mark more packets lost or retransmit more.
2744		 */
2745		if (tp->retrans_out) {
2746			tcp_cwnd_reduction(sk, prior_unsacked, 0);
2747			return true;
2748		}
2749
2750		if (!tcp_any_retrans_done(sk))
2751			tp->retrans_stamp = 0;
2752
2753		DBGUNDO(sk, "partial recovery");
2754		tcp_undo_cwnd_reduction(sk, true);
2755		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO);
2756		tcp_try_keep_open(sk);
2757		return true;
 
 
2758	}
2759	return false;
2760}
2761
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2762/* Process an event, which can update packets-in-flight not trivially.
2763 * Main goal of this function is to calculate new estimate for left_out,
2764 * taking into account both packets sitting in receiver's buffer and
2765 * packets lost by network.
2766 *
2767 * Besides that it does CWND reduction, when packet loss is detected
2768 * and changes state of machine.
 
2769 *
2770 * It does _not_ decide what to send, it is made in function
2771 * tcp_xmit_retransmit_queue().
2772 */
2773static void tcp_fastretrans_alert(struct sock *sk, const int acked,
2774				  const int prior_unsacked,
2775				  bool is_dupack, int flag)
2776{
2777	struct inet_connection_sock *icsk = inet_csk(sk);
2778	struct tcp_sock *tp = tcp_sk(sk);
2779	bool do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) &&
2780				    (tcp_fackets_out(tp) > tp->reordering));
2781	int fast_rexmit = 0;
 
2782
2783	if (WARN_ON(!tp->packets_out && tp->sacked_out))
2784		tp->sacked_out = 0;
2785	if (WARN_ON(!tp->sacked_out && tp->fackets_out))
2786		tp->fackets_out = 0;
2787
2788	/* Now state machine starts.
2789	 * A. ECE, hence prohibit cwnd undoing, the reduction is required. */
2790	if (flag & FLAG_ECE)
2791		tp->prior_ssthresh = 0;
2792
2793	/* B. In all the states check for reneging SACKs. */
2794	if (tcp_check_sack_reneging(sk, flag))
2795		return;
2796
2797	/* C. Check consistency of the current state. */
2798	tcp_verify_left_out(tp);
2799
2800	/* D. Check state exit conditions. State can be terminated
2801	 *    when high_seq is ACKed. */
2802	if (icsk->icsk_ca_state == TCP_CA_Open) {
2803		WARN_ON(tp->retrans_out != 0);
2804		tp->retrans_stamp = 0;
2805	} else if (!before(tp->snd_una, tp->high_seq)) {
2806		switch (icsk->icsk_ca_state) {
2807		case TCP_CA_CWR:
2808			/* CWR is to be held something *above* high_seq
2809			 * is ACKed for CWR bit to reach receiver. */
2810			if (tp->snd_una != tp->high_seq) {
2811				tcp_end_cwnd_reduction(sk);
2812				tcp_set_ca_state(sk, TCP_CA_Open);
2813			}
2814			break;
2815
2816		case TCP_CA_Recovery:
2817			if (tcp_is_reno(tp))
2818				tcp_reset_reno_sack(tp);
2819			if (tcp_try_undo_recovery(sk))
2820				return;
2821			tcp_end_cwnd_reduction(sk);
2822			break;
2823		}
2824	}
2825
2826	/* E. Process state. */
2827	switch (icsk->icsk_ca_state) {
2828	case TCP_CA_Recovery:
2829		if (!(flag & FLAG_SND_UNA_ADVANCED)) {
2830			if (tcp_is_reno(tp) && is_dupack)
2831				tcp_add_reno_sack(sk);
2832		} else {
2833			if (tcp_try_undo_partial(sk, acked, prior_unsacked))
2834				return;
2835			/* Partial ACK arrived. Force fast retransmit. */
2836			do_lost = tcp_is_reno(tp) ||
2837				  tcp_fackets_out(tp) > tp->reordering;
2838		}
2839		if (tcp_try_undo_dsack(sk)) {
2840			tcp_try_keep_open(sk);
2841			return;
 
 
 
 
 
 
 
 
 
 
 
 
2842		}
2843		break;
2844	case TCP_CA_Loss:
2845		tcp_process_loss(sk, flag, is_dupack);
2846		if (icsk->icsk_ca_state != TCP_CA_Open)
 
 
 
 
2847			return;
2848		/* Fall through to processing in Open state. */
 
2849	default:
2850		if (tcp_is_reno(tp)) {
2851			if (flag & FLAG_SND_UNA_ADVANCED)
2852				tcp_reset_reno_sack(tp);
2853			if (is_dupack)
2854				tcp_add_reno_sack(sk);
2855		}
2856
2857		if (icsk->icsk_ca_state <= TCP_CA_Disorder)
2858			tcp_try_undo_dsack(sk);
2859
 
2860		if (!tcp_time_to_recover(sk, flag)) {
2861			tcp_try_to_open(sk, flag, prior_unsacked);
2862			return;
2863		}
2864
2865		/* MTU probe failure: don't reduce cwnd */
2866		if (icsk->icsk_ca_state < TCP_CA_CWR &&
2867		    icsk->icsk_mtup.probe_size &&
2868		    tp->snd_una == tp->mtu_probe.probe_seq_start) {
2869			tcp_mtup_probe_failed(sk);
2870			/* Restores the reduction we did in tcp_mtup_probe() */
2871			tp->snd_cwnd++;
2872			tcp_simple_retransmit(sk);
2873			return;
2874		}
2875
2876		/* Otherwise enter Recovery state */
2877		tcp_enter_recovery(sk, (flag & FLAG_ECE));
2878		fast_rexmit = 1;
2879	}
2880
2881	if (do_lost)
2882		tcp_update_scoreboard(sk, fast_rexmit);
2883	tcp_cwnd_reduction(sk, prior_unsacked, fast_rexmit);
2884	tcp_xmit_retransmit_queue(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2885}
2886
2887static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag,
2888				      long seq_rtt_us, long sack_rtt_us)
 
2889{
2890	const struct tcp_sock *tp = tcp_sk(sk);
2891
2892	/* Prefer RTT measured from ACK's timing to TS-ECR. This is because
2893	 * broken middle-boxes or peers may corrupt TS-ECR fields. But
2894	 * Karn's algorithm forbids taking RTT if some retransmitted data
2895	 * is acked (RFC6298).
2896	 */
2897	if (flag & FLAG_RETRANS_DATA_ACKED)
2898		seq_rtt_us = -1L;
2899
2900	if (seq_rtt_us < 0)
2901		seq_rtt_us = sack_rtt_us;
2902
2903	/* RTTM Rule: A TSecr value received in a segment is used to
2904	 * update the averaged RTT measurement only if the segment
2905	 * acknowledges some new data, i.e., only if it advances the
2906	 * left edge of the send window.
2907	 * See draft-ietf-tcplw-high-performance-00, section 3.3.
2908	 */
2909	if (seq_rtt_us < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
2910	    flag & FLAG_ACKED)
2911		seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - tp->rx_opt.rcv_tsecr);
2912
 
2913	if (seq_rtt_us < 0)
2914		return false;
2915
 
 
 
 
 
2916	tcp_rtt_estimator(sk, seq_rtt_us);
2917	tcp_set_rto(sk);
2918
2919	/* RFC6298: only reset backoff on valid RTT measurement. */
2920	inet_csk(sk)->icsk_backoff = 0;
2921	return true;
2922}
2923
2924/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */
2925static void tcp_synack_rtt_meas(struct sock *sk, const u32 synack_stamp)
2926{
2927	struct tcp_sock *tp = tcp_sk(sk);
2928	long seq_rtt_us = -1L;
2929
2930	if (synack_stamp && !tp->total_retrans)
2931		seq_rtt_us = jiffies_to_usecs(tcp_time_stamp - synack_stamp);
2932
2933	/* If the ACK acks both the SYNACK and the (Fast Open'd) data packets
2934	 * sent in SYN_RECV, SYNACK RTT is the smooth RTT computed in tcp_ack()
2935	 */
2936	if (!tp->srtt_us)
2937		tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt_us, -1L);
2938}
2939
2940static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 acked, u32 in_flight)
 
2941{
2942	const struct inet_connection_sock *icsk = inet_csk(sk);
2943	icsk->icsk_ca_ops->cong_avoid(sk, ack, acked, in_flight);
2944	tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp;
 
2945}
2946
2947/* Restart timer after forward progress on connection.
2948 * RFC2988 recommends to restart timer to now+rto.
2949 */
2950void tcp_rearm_rto(struct sock *sk)
2951{
2952	const struct inet_connection_sock *icsk = inet_csk(sk);
2953	struct tcp_sock *tp = tcp_sk(sk);
2954
2955	/* If the retrans timer is currently being used by Fast Open
2956	 * for SYN-ACK retrans purpose, stay put.
2957	 */
2958	if (tp->fastopen_rsk)
2959		return;
2960
2961	if (!tp->packets_out) {
2962		inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS);
2963	} else {
2964		u32 rto = inet_csk(sk)->icsk_rto;
2965		/* Offset the time elapsed after installing regular RTO */
2966		if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
2967		    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
2968			struct sk_buff *skb = tcp_write_queue_head(sk);
2969			const u32 rto_time_stamp = TCP_SKB_CB(skb)->when + rto;
2970			s32 delta = (s32)(rto_time_stamp - tcp_time_stamp);
2971			/* delta may not be positive if the socket is locked
2972			 * when the retrans timer fires and is rescheduled.
2973			 */
2974			if (delta > 0)
2975				rto = delta;
2976		}
2977		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, rto,
2978					  TCP_RTO_MAX);
2979	}
2980}
2981
2982/* This function is called when the delayed ER timer fires. TCP enters
2983 * fast recovery and performs fast-retransmit.
2984 */
2985void tcp_resume_early_retransmit(struct sock *sk)
2986{
2987	struct tcp_sock *tp = tcp_sk(sk);
2988
2989	tcp_rearm_rto(sk);
2990
2991	/* Stop if ER is disabled after the delayed ER timer is scheduled */
2992	if (!tp->do_early_retrans)
2993		return;
2994
2995	tcp_enter_recovery(sk, false);
2996	tcp_update_scoreboard(sk, 1);
2997	tcp_xmit_retransmit_queue(sk);
2998}
2999
3000/* If we get here, the whole TSO packet has not been acked. */
3001static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb)
3002{
3003	struct tcp_sock *tp = tcp_sk(sk);
3004	u32 packets_acked;
3005
3006	BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una));
3007
3008	packets_acked = tcp_skb_pcount(skb);
3009	if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3010		return 0;
3011	packets_acked -= tcp_skb_pcount(skb);
3012
3013	if (packets_acked) {
3014		BUG_ON(tcp_skb_pcount(skb) == 0);
3015		BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq));
3016	}
3017
3018	return packets_acked;
3019}
3020
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3021/* Remove acknowledged frames from the retransmission queue. If our packet
3022 * is before the ack sequence we can discard it as it's confirmed to have
3023 * arrived at the other end.
3024 */
3025static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets,
3026			       u32 prior_snd_una, long sack_rtt_us)
 
3027{
3028	const struct inet_connection_sock *icsk = inet_csk(sk);
3029	struct skb_mstamp first_ackt, last_ackt, now;
3030	struct tcp_sock *tp = tcp_sk(sk);
3031	u32 prior_sacked = tp->sacked_out;
3032	u32 reord = tp->packets_out;
 
3033	bool fully_acked = true;
3034	long ca_seq_rtt_us = -1L;
3035	long seq_rtt_us = -1L;
3036	struct sk_buff *skb;
3037	u32 pkts_acked = 0;
3038	bool rtt_update;
3039	int flag = 0;
3040
3041	first_ackt.v64 = 0;
3042
3043	while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) {
3044		struct tcp_skb_cb *scb = TCP_SKB_CB(skb);
 
3045		u8 sacked = scb->sacked;
3046		u32 acked_pcount;
3047
3048		/* Determine how many packets and what bytes were acked, tso and else */
3049		if (after(scb->end_seq, tp->snd_una)) {
3050			if (tcp_skb_pcount(skb) == 1 ||
3051			    !after(tp->snd_una, scb->seq))
3052				break;
3053
3054			acked_pcount = tcp_tso_acked(sk, skb);
3055			if (!acked_pcount)
3056				break;
3057
3058			fully_acked = false;
3059		} else {
3060			acked_pcount = tcp_skb_pcount(skb);
3061		}
3062
3063		if (sacked & TCPCB_RETRANS) {
3064			if (sacked & TCPCB_SACKED_RETRANS)
3065				tp->retrans_out -= acked_pcount;
3066			flag |= FLAG_RETRANS_DATA_ACKED;
3067		} else {
3068			last_ackt = skb->skb_mstamp;
3069			WARN_ON_ONCE(last_ackt.v64 == 0);
3070			if (!first_ackt.v64)
3071				first_ackt = last_ackt;
3072
3073			if (!(sacked & TCPCB_SACKED_ACKED))
3074				reord = min(pkts_acked, reord);
3075			if (!after(scb->end_seq, tp->high_seq))
3076				flag |= FLAG_ORIG_SACK_ACKED;
3077		}
3078
3079		if (sacked & TCPCB_SACKED_ACKED)
3080			tp->sacked_out -= acked_pcount;
 
 
 
 
 
 
3081		if (sacked & TCPCB_LOST)
3082			tp->lost_out -= acked_pcount;
3083
3084		tp->packets_out -= acked_pcount;
3085		pkts_acked += acked_pcount;
 
3086
3087		/* Initial outgoing SYN's get put onto the write_queue
3088		 * just like anything else we transmit.  It is not
3089		 * true data, and if we misinform our callers that
3090		 * this ACK acks real data, we will erroneously exit
3091		 * connection startup slow start one packet too
3092		 * quickly.  This is severely frowned upon behavior.
3093		 */
3094		if (!(scb->tcp_flags & TCPHDR_SYN)) {
3095			flag |= FLAG_DATA_ACKED;
3096		} else {
3097			flag |= FLAG_SYN_ACKED;
3098			tp->retrans_stamp = 0;
3099		}
3100
3101		if (!fully_acked)
3102			break;
3103
3104		tcp_unlink_write_queue(skb, sk);
3105		sk_wmem_free_skb(sk, skb);
3106		if (skb == tp->retransmit_skb_hint)
 
3107			tp->retransmit_skb_hint = NULL;
3108		if (skb == tp->lost_skb_hint)
3109			tp->lost_skb_hint = NULL;
 
 
3110	}
3111
 
 
 
3112	if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una)))
3113		tp->snd_up = tp->snd_una;
3114
3115	if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
3116		flag |= FLAG_SACK_RENEGING;
3117
3118	skb_mstamp_get(&now);
3119	if (first_ackt.v64) {
3120		seq_rtt_us = skb_mstamp_us_delta(&now, &first_ackt);
3121		ca_seq_rtt_us = skb_mstamp_us_delta(&now, &last_ackt);
 
 
 
 
 
 
 
 
 
 
 
 
 
3122	}
3123
3124	rtt_update = tcp_ack_update_rtt(sk, flag, seq_rtt_us, sack_rtt_us);
 
 
 
 
3125
3126	if (flag & FLAG_ACKED) {
3127		const struct tcp_congestion_ops *ca_ops
3128			= inet_csk(sk)->icsk_ca_ops;
3129
3130		tcp_rearm_rto(sk);
3131		if (unlikely(icsk->icsk_mtup.probe_size &&
3132			     !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) {
3133			tcp_mtup_probe_success(sk);
3134		}
3135
3136		if (tcp_is_reno(tp)) {
3137			tcp_remove_reno_sacks(sk, pkts_acked);
 
 
 
 
 
 
 
 
 
3138		} else {
3139			int delta;
3140
3141			/* Non-retransmitted hole got filled? That's reordering */
3142			if (reord < prior_fackets)
3143				tcp_update_reordering(sk, tp->fackets_out - reord, 0);
3144
3145			delta = tcp_is_fack(tp) ? pkts_acked :
3146						  prior_sacked - tp->sacked_out;
3147			tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta);
3148		}
3149
3150		tp->fackets_out -= min(pkts_acked, tp->fackets_out);
3151
3152		if (ca_ops->pkts_acked)
3153			ca_ops->pkts_acked(sk, pkts_acked, ca_seq_rtt_us);
3154
3155	} else if (skb && rtt_update && sack_rtt_us >= 0 &&
3156		   sack_rtt_us > skb_mstamp_us_delta(&now, &skb->skb_mstamp)) {
 
3157		/* Do not re-arm RTO if the sack RTT is measured from data sent
3158		 * after when the head was last (re)transmitted. Otherwise the
3159		 * timeout may continue to extend in loss recovery.
3160		 */
3161		tcp_rearm_rto(sk);
 
 
 
 
 
 
 
 
 
3162	}
3163
3164#if FASTRETRANS_DEBUG > 0
3165	WARN_ON((int)tp->sacked_out < 0);
3166	WARN_ON((int)tp->lost_out < 0);
3167	WARN_ON((int)tp->retrans_out < 0);
3168	if (!tp->packets_out && tcp_is_sack(tp)) {
3169		icsk = inet_csk(sk);
3170		if (tp->lost_out) {
3171			pr_debug("Leak l=%u %d\n",
3172				 tp->lost_out, icsk->icsk_ca_state);
3173			tp->lost_out = 0;
3174		}
3175		if (tp->sacked_out) {
3176			pr_debug("Leak s=%u %d\n",
3177				 tp->sacked_out, icsk->icsk_ca_state);
3178			tp->sacked_out = 0;
3179		}
3180		if (tp->retrans_out) {
3181			pr_debug("Leak r=%u %d\n",
3182				 tp->retrans_out, icsk->icsk_ca_state);
3183			tp->retrans_out = 0;
3184		}
3185	}
3186#endif
3187	return flag;
3188}
3189
3190static void tcp_ack_probe(struct sock *sk)
3191{
3192	const struct tcp_sock *tp = tcp_sk(sk);
3193	struct inet_connection_sock *icsk = inet_csk(sk);
 
 
3194
3195	/* Was it a usable window open? */
3196
3197	if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) {
 
3198		icsk->icsk_backoff = 0;
 
3199		inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0);
3200		/* Socket must be waked up by subsequent tcp_data_snd_check().
3201		 * This function is not for random using!
3202		 */
3203	} else {
3204		inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3205					  min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3206					  TCP_RTO_MAX);
 
3207	}
3208}
3209
3210static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag)
3211{
3212	return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) ||
3213		inet_csk(sk)->icsk_ca_state != TCP_CA_Open;
3214}
3215
3216/* Decide wheather to run the increase function of congestion control. */
3217static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag)
3218{
3219	if (tcp_in_cwnd_reduction(sk))
3220		return false;
3221
3222	/* If reordering is high then always grow cwnd whenever data is
3223	 * delivered regardless of its ordering. Otherwise stay conservative
3224	 * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/
3225	 * new SACK or ECE mark may first advance cwnd here and later reduce
3226	 * cwnd in tcp_fastretrans_alert() based on more states.
3227	 */
3228	if (tcp_sk(sk)->reordering > sysctl_tcp_reordering)
 
3229		return flag & FLAG_FORWARD_PROGRESS;
3230
3231	return flag & FLAG_DATA_ACKED;
3232}
3233
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3234/* Check that window update is acceptable.
3235 * The function assumes that snd_una<=ack<=snd_next.
3236 */
3237static inline bool tcp_may_update_window(const struct tcp_sock *tp,
3238					const u32 ack, const u32 ack_seq,
3239					const u32 nwin)
3240{
3241	return	after(ack, tp->snd_una) ||
3242		after(ack_seq, tp->snd_wl1) ||
3243		(ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3244}
3245
3246/* Update our send window.
3247 *
3248 * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2
3249 * and in FreeBSD. NetBSD's one is even worse.) is wrong.
3250 */
3251static int tcp_ack_update_window(struct sock *sk, const struct sk_buff *skb, u32 ack,
3252				 u32 ack_seq)
3253{
3254	struct tcp_sock *tp = tcp_sk(sk);
3255	int flag = 0;
3256	u32 nwin = ntohs(tcp_hdr(skb)->window);
3257
3258	if (likely(!tcp_hdr(skb)->syn))
3259		nwin <<= tp->rx_opt.snd_wscale;
3260
3261	if (tcp_may_update_window(tp, ack, ack_seq, nwin)) {
3262		flag |= FLAG_WIN_UPDATE;
3263		tcp_update_wl(tp, ack_seq);
3264
3265		if (tp->snd_wnd != nwin) {
3266			tp->snd_wnd = nwin;
3267
3268			/* Note, it is the only place, where
3269			 * fast path is recovered for sending TCP.
3270			 */
3271			tp->pred_flags = 0;
3272			tcp_fast_path_check(sk);
3273
 
 
 
3274			if (nwin > tp->max_window) {
3275				tp->max_window = nwin;
3276				tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie);
3277			}
3278		}
3279	}
3280
3281	tp->snd_una = ack;
3282
3283	return flag;
3284}
3285
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3286/* RFC 5961 7 [ACK Throttling] */
3287static void tcp_send_challenge_ack(struct sock *sk)
3288{
3289	/* unprotected vars, we dont care of overwrites */
3290	static u32 challenge_timestamp;
3291	static unsigned int challenge_count;
3292	u32 now = jiffies / HZ;
3293
3294	if (now != challenge_timestamp) {
3295		challenge_timestamp = now;
3296		challenge_count = 0;
3297	}
3298	if (++challenge_count <= sysctl_tcp_challenge_ack_limit) {
3299		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPCHALLENGEACK);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3300		tcp_send_ack(sk);
3301	}
3302}
3303
3304static void tcp_store_ts_recent(struct tcp_sock *tp)
3305{
3306	tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval;
3307	tp->rx_opt.ts_recent_stamp = get_seconds();
3308}
3309
3310static void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq)
3311{
3312	if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) {
3313		/* PAWS bug workaround wrt. ACK frames, the PAWS discard
3314		 * extra check below makes sure this can only happen
3315		 * for pure ACK frames.  -DaveM
3316		 *
3317		 * Not only, also it occurs for expired timestamps.
3318		 */
3319
3320		if (tcp_paws_check(&tp->rx_opt, 0))
3321			tcp_store_ts_recent(tp);
3322	}
3323}
3324
3325/* This routine deals with acks during a TLP episode.
3326 * Ref: loss detection algorithm in draft-dukkipati-tcpm-tcp-loss-probe.
3327 */
3328static void tcp_process_tlp_ack(struct sock *sk, u32 ack, int flag)
3329{
3330	struct tcp_sock *tp = tcp_sk(sk);
3331	bool is_tlp_dupack = (ack == tp->tlp_high_seq) &&
3332			     !(flag & (FLAG_SND_UNA_ADVANCED |
3333				       FLAG_NOT_DUP | FLAG_DATA_SACKED));
3334
3335	/* Mark the end of TLP episode on receiving TLP dupack or when
3336	 * ack is after tlp_high_seq.
3337	 */
3338	if (is_tlp_dupack) {
3339		tp->tlp_high_seq = 0;
3340		return;
3341	}
3342
3343	if (after(ack, tp->tlp_high_seq)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3344		tp->tlp_high_seq = 0;
3345		/* Don't reduce cwnd if DSACK arrives for TLP retrans. */
3346		if (!(flag & FLAG_DSACKING_ACK)) {
3347			tcp_init_cwnd_reduction(sk, true);
3348			tcp_set_ca_state(sk, TCP_CA_CWR);
3349			tcp_end_cwnd_reduction(sk);
3350			tcp_try_keep_open(sk);
3351			NET_INC_STATS_BH(sock_net(sk),
3352					 LINUX_MIB_TCPLOSSPROBERECOVERY);
3353		}
3354	}
3355}
3356
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3357/* This routine deals with incoming acks, but not outgoing ones. */
3358static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag)
3359{
3360	struct inet_connection_sock *icsk = inet_csk(sk);
3361	struct tcp_sock *tp = tcp_sk(sk);
 
 
3362	u32 prior_snd_una = tp->snd_una;
 
3363	u32 ack_seq = TCP_SKB_CB(skb)->seq;
3364	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3365	bool is_dupack = false;
3366	u32 prior_in_flight;
3367	u32 prior_fackets;
3368	int prior_packets = tp->packets_out;
3369	const int prior_unsacked = tp->packets_out - tp->sacked_out;
3370	int acked = 0; /* Number of packets newly acked */
3371	long sack_rtt_us = -1L;
 
 
 
 
 
 
 
 
3372
3373	/* If the ack is older than previous acks
3374	 * then we can probably ignore it.
3375	 */
3376	if (before(ack, prior_snd_una)) {
 
 
 
 
3377		/* RFC 5961 5.2 [Blind Data Injection Attack].[Mitigation] */
3378		if (before(ack, prior_snd_una - tp->max_window)) {
3379			tcp_send_challenge_ack(sk);
3380			return -1;
 
3381		}
3382		goto old_ack;
3383	}
3384
3385	/* If the ack includes data we haven't sent yet, discard
3386	 * this segment (RFC793 Section 3.9).
3387	 */
3388	if (after(ack, tp->snd_nxt))
3389		goto invalid_ack;
3390
3391	if (icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
3392	    icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
3393		tcp_rearm_rto(sk);
3394
3395	if (after(ack, prior_snd_una))
3396		flag |= FLAG_SND_UNA_ADVANCED;
 
 
 
 
 
 
 
 
3397
3398	prior_fackets = tp->fackets_out;
3399	prior_in_flight = tcp_packets_in_flight(tp);
3400
3401	/* ts_recent update must be made after we are sure that the packet
3402	 * is in window.
3403	 */
3404	if (flag & FLAG_UPDATE_TS_RECENT)
3405		tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq);
3406
3407	if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) {
 
3408		/* Window is constant, pure forward advance.
3409		 * No more checks are required.
3410		 * Note, we use the fact that SND.UNA>=SND.WL2.
3411		 */
3412		tcp_update_wl(tp, ack_seq);
3413		tp->snd_una = ack;
3414		flag |= FLAG_WIN_UPDATE;
3415
3416		tcp_ca_event(sk, CA_EVENT_FAST_ACK);
3417
3418		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS);
3419	} else {
 
 
3420		if (ack_seq != TCP_SKB_CB(skb)->end_seq)
3421			flag |= FLAG_DATA;
3422		else
3423			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS);
3424
3425		flag |= tcp_ack_update_window(sk, skb, ack, ack_seq);
3426
3427		if (TCP_SKB_CB(skb)->sacked)
3428			flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3429							&sack_rtt_us);
3430
3431		if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb)))
3432			flag |= FLAG_ECE;
 
 
 
 
 
 
3433
3434		tcp_ca_event(sk, CA_EVENT_SLOW_ACK);
 
 
 
3435	}
3436
 
 
 
 
 
 
 
 
 
3437	/* We passed data and got it acked, remove any soft error
3438	 * log. Something worked...
3439	 */
3440	sk->sk_err_soft = 0;
3441	icsk->icsk_probes_out = 0;
3442	tp->rcv_tstamp = tcp_time_stamp;
3443	if (!prior_packets)
3444		goto no_queue;
3445
3446	/* See if we can take anything off of the retransmit queue. */
3447	acked = tp->packets_out;
3448	flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una,
3449				    sack_rtt_us);
3450	acked -= tp->packets_out;
3451
3452	/* Advance cwnd if state allows */
3453	if (tcp_may_raise_cwnd(sk, flag))
3454		tcp_cong_avoid(sk, ack, acked, prior_in_flight);
3455
3456	if (tcp_ack_is_dubious(sk, flag)) {
3457		is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP));
3458		tcp_fastretrans_alert(sk, acked, prior_unsacked,
3459				      is_dupack, flag);
3460	}
3461	if (tp->tlp_high_seq)
3462		tcp_process_tlp_ack(sk, ack, flag);
3463
3464	if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) {
3465		struct dst_entry *dst = __sk_dst_get(sk);
3466		if (dst)
3467			dst_confirm(dst);
3468	}
3469
3470	if (icsk->icsk_pending == ICSK_TIME_RETRANS)
3471		tcp_schedule_loss_probe(sk);
3472	tcp_update_pacing_rate(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3473	return 1;
3474
3475no_queue:
3476	/* If data was DSACKed, see if we can undo a cwnd reduction. */
3477	if (flag & FLAG_DSACKING_ACK)
3478		tcp_fastretrans_alert(sk, acked, prior_unsacked,
3479				      is_dupack, flag);
 
 
3480	/* If this ack opens up a zero window, clear backoff.  It was
3481	 * being used to time the probes, and is probably far higher than
3482	 * it needs to be for normal retransmission.
3483	 */
3484	if (tcp_send_head(sk))
3485		tcp_ack_probe(sk);
3486
3487	if (tp->tlp_high_seq)
3488		tcp_process_tlp_ack(sk, ack, flag);
3489	return 1;
3490
3491invalid_ack:
3492	SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3493	return -1;
3494
3495old_ack:
3496	/* If data was SACKed, tag it and see if we should send more data.
3497	 * If data was DSACKed, see if we can undo a cwnd reduction.
3498	 */
3499	if (TCP_SKB_CB(skb)->sacked) {
3500		flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una,
3501						&sack_rtt_us);
3502		tcp_fastretrans_alert(sk, acked, prior_unsacked,
3503				      is_dupack, flag);
 
 
3504	}
3505
3506	SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt);
3507	return 0;
3508}
3509
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3510/* Look for tcp options. Normally only called on SYN and SYNACK packets.
3511 * But, this can also be called on packets in the established flow when
3512 * the fast version below fails.
3513 */
3514void tcp_parse_options(const struct sk_buff *skb,
 
3515		       struct tcp_options_received *opt_rx, int estab,
3516		       struct tcp_fastopen_cookie *foc)
3517{
3518	const unsigned char *ptr;
3519	const struct tcphdr *th = tcp_hdr(skb);
3520	int length = (th->doff * 4) - sizeof(struct tcphdr);
3521
3522	ptr = (const unsigned char *)(th + 1);
3523	opt_rx->saw_tstamp = 0;
 
3524
3525	while (length > 0) {
3526		int opcode = *ptr++;
3527		int opsize;
3528
3529		switch (opcode) {
3530		case TCPOPT_EOL:
3531			return;
3532		case TCPOPT_NOP:	/* Ref: RFC 793 section 3.1 */
3533			length--;
3534			continue;
3535		default:
 
 
3536			opsize = *ptr++;
3537			if (opsize < 2) /* "silly options" */
3538				return;
3539			if (opsize > length)
3540				return;	/* don't parse partial options */
3541			switch (opcode) {
3542			case TCPOPT_MSS:
3543				if (opsize == TCPOLEN_MSS && th->syn && !estab) {
3544					u16 in_mss = get_unaligned_be16(ptr);
3545					if (in_mss) {
3546						if (opt_rx->user_mss &&
3547						    opt_rx->user_mss < in_mss)
3548							in_mss = opt_rx->user_mss;
3549						opt_rx->mss_clamp = in_mss;
3550					}
3551				}
3552				break;
3553			case TCPOPT_WINDOW:
3554				if (opsize == TCPOLEN_WINDOW && th->syn &&
3555				    !estab && sysctl_tcp_window_scaling) {
3556					__u8 snd_wscale = *(__u8 *)ptr;
3557					opt_rx->wscale_ok = 1;
3558					if (snd_wscale > 14) {
3559						net_info_ratelimited("%s: Illegal window scaling value %d >14 received\n",
3560								     __func__,
3561								     snd_wscale);
3562						snd_wscale = 14;
 
3563					}
3564					opt_rx->snd_wscale = snd_wscale;
3565				}
3566				break;
3567			case TCPOPT_TIMESTAMP:
3568				if ((opsize == TCPOLEN_TIMESTAMP) &&
3569				    ((estab && opt_rx->tstamp_ok) ||
3570				     (!estab && sysctl_tcp_timestamps))) {
3571					opt_rx->saw_tstamp = 1;
3572					opt_rx->rcv_tsval = get_unaligned_be32(ptr);
3573					opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4);
3574				}
3575				break;
3576			case TCPOPT_SACK_PERM:
3577				if (opsize == TCPOLEN_SACK_PERM && th->syn &&
3578				    !estab && sysctl_tcp_sack) {
3579					opt_rx->sack_ok = TCP_SACK_SEEN;
3580					tcp_sack_reset(opt_rx);
3581				}
3582				break;
3583
3584			case TCPOPT_SACK:
3585				if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) &&
3586				   !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) &&
3587				   opt_rx->sack_ok) {
3588					TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th;
3589				}
3590				break;
3591#ifdef CONFIG_TCP_MD5SIG
3592			case TCPOPT_MD5SIG:
3593				/*
3594				 * The MD5 Hash has already been
3595				 * checked (see tcp_v{4,6}_do_rcv()).
3596				 */
3597				break;
3598#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
3599			case TCPOPT_EXP:
3600				/* Fast Open option shares code 254 using a
3601				 * 16 bits magic number. It's valid only in
3602				 * SYN or SYN-ACK with an even size.
3603				 */
3604				if (opsize < TCPOLEN_EXP_FASTOPEN_BASE ||
3605				    get_unaligned_be16(ptr) != TCPOPT_FASTOPEN_MAGIC ||
3606				    foc == NULL || !th->syn || (opsize & 1))
 
 
 
 
 
 
 
3607					break;
3608				foc->len = opsize - TCPOLEN_EXP_FASTOPEN_BASE;
3609				if (foc->len >= TCP_FASTOPEN_COOKIE_MIN &&
3610				    foc->len <= TCP_FASTOPEN_COOKIE_MAX)
3611					memcpy(foc->val, ptr + 2, foc->len);
3612				else if (foc->len != 0)
3613					foc->len = -1;
3614				break;
3615
 
 
3616			}
3617			ptr += opsize-2;
3618			length -= opsize;
3619		}
3620	}
3621}
3622EXPORT_SYMBOL(tcp_parse_options);
3623
3624static bool tcp_parse_aligned_timestamp(struct tcp_sock *tp, const struct tcphdr *th)
3625{
3626	const __be32 *ptr = (const __be32 *)(th + 1);
3627
3628	if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16)
3629			  | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) {
3630		tp->rx_opt.saw_tstamp = 1;
3631		++ptr;
3632		tp->rx_opt.rcv_tsval = ntohl(*ptr);
3633		++ptr;
3634		if (*ptr)
3635			tp->rx_opt.rcv_tsecr = ntohl(*ptr) - tp->tsoffset;
3636		else
3637			tp->rx_opt.rcv_tsecr = 0;
3638		return true;
3639	}
3640	return false;
3641}
3642
3643/* Fast parse options. This hopes to only see timestamps.
3644 * If it is wrong it falls back on tcp_parse_options().
3645 */
3646static bool tcp_fast_parse_options(const struct sk_buff *skb,
 
3647				   const struct tcphdr *th, struct tcp_sock *tp)
3648{
3649	/* In the spirit of fast parsing, compare doff directly to constant
3650	 * values.  Because equality is used, short doff can be ignored here.
3651	 */
3652	if (th->doff == (sizeof(*th) / 4)) {
3653		tp->rx_opt.saw_tstamp = 0;
3654		return false;
3655	} else if (tp->rx_opt.tstamp_ok &&
3656		   th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) {
3657		if (tcp_parse_aligned_timestamp(tp, th))
3658			return true;
3659	}
3660
3661	tcp_parse_options(skb, &tp->rx_opt, 1, NULL);
3662	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
3663		tp->rx_opt.rcv_tsecr -= tp->tsoffset;
3664
3665	return true;
3666}
3667
3668#ifdef CONFIG_TCP_MD5SIG
3669/*
3670 * Parse MD5 Signature option
3671 */
3672const u8 *tcp_parse_md5sig_option(const struct tcphdr *th)
 
3673{
3674	int length = (th->doff << 2) - sizeof(*th);
3675	const u8 *ptr = (const u8 *)(th + 1);
 
3676
3677	/* If the TCP option is too short, we can short cut */
3678	if (length < TCPOLEN_MD5SIG)
3679		return NULL;
3680
3681	while (length > 0) {
 
 
 
 
3682		int opcode = *ptr++;
3683		int opsize;
3684
3685		switch (opcode) {
3686		case TCPOPT_EOL:
3687			return NULL;
3688		case TCPOPT_NOP:
3689			length--;
3690			continue;
3691		default:
3692			opsize = *ptr++;
3693			if (opsize < 2 || opsize > length)
3694				return NULL;
3695			if (opcode == TCPOPT_MD5SIG)
3696				return opsize == TCPOLEN_MD5SIG ? ptr : NULL;
 
 
 
 
 
 
 
 
 
 
 
3697		}
3698		ptr += opsize - 2;
3699		length -= opsize;
3700	}
3701	return NULL;
3702}
3703EXPORT_SYMBOL(tcp_parse_md5sig_option);
3704#endif
3705
3706/* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM
3707 *
3708 * It is not fatal. If this ACK does _not_ change critical state (seqs, window)
3709 * it can pass through stack. So, the following predicate verifies that
3710 * this segment is not used for anything but congestion avoidance or
3711 * fast retransmit. Moreover, we even are able to eliminate most of such
3712 * second order effects, if we apply some small "replay" window (~RTO)
3713 * to timestamp space.
3714 *
3715 * All these measures still do not guarantee that we reject wrapped ACKs
3716 * on networks with high bandwidth, when sequence space is recycled fastly,
3717 * but it guarantees that such events will be very rare and do not affect
3718 * connection seriously. This doesn't look nice, but alas, PAWS is really
3719 * buggy extension.
3720 *
3721 * [ Later note. Even worse! It is buggy for segments _with_ data. RFC
3722 * states that events when retransmit arrives after original data are rare.
3723 * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is
3724 * the biggest problem on large power networks even with minor reordering.
3725 * OK, let's give it small replay window. If peer clock is even 1hz, it is safe
3726 * up to bandwidth of 18Gigabit/sec. 8) ]
3727 */
3728
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3729static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb)
3730{
3731	const struct tcp_sock *tp = tcp_sk(sk);
3732	const struct tcphdr *th = tcp_hdr(skb);
3733	u32 seq = TCP_SKB_CB(skb)->seq;
3734	u32 ack = TCP_SKB_CB(skb)->ack_seq;
3735
3736	return (/* 1. Pure ACK with correct sequence number. */
3737		(th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) &&
3738
3739		/* 2. ... and duplicate ACK. */
3740		ack == tp->snd_una &&
3741
3742		/* 3. ... and does not update window. */
3743		!tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) &&
3744
3745		/* 4. ... and sits in replay window. */
3746		(s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ);
 
3747}
3748
3749static inline bool tcp_paws_discard(const struct sock *sk,
3750				   const struct sk_buff *skb)
3751{
3752	const struct tcp_sock *tp = tcp_sk(sk);
3753
3754	return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) &&
3755	       !tcp_disordered_ack(sk, skb);
3756}
3757
3758/* Check segment sequence number for validity.
3759 *
3760 * Segment controls are considered valid, if the segment
3761 * fits to the window after truncation to the window. Acceptability
3762 * of data (and SYN, FIN, of course) is checked separately.
3763 * See tcp_data_queue(), for example.
3764 *
3765 * Also, controls (RST is main one) are accepted using RCV.WUP instead
3766 * of RCV.NXT. Peer still did not advance his SND.UNA when we
3767 * delayed ACK, so that hisSND.UNA<=ourRCV.WUP.
3768 * (borrowed from freebsd)
3769 */
3770
3771static inline bool tcp_sequence(const struct tcp_sock *tp, u32 seq, u32 end_seq)
 
3772{
3773	return	!before(end_seq, tp->rcv_wup) &&
3774		!after(seq, tp->rcv_nxt + tcp_receive_window(tp));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3775}
 
3776
3777/* When we get a reset we do this. */
3778void tcp_reset(struct sock *sk)
3779{
 
 
 
 
 
 
 
 
 
 
3780	/* We want the right error as BSD sees it (and indeed as we do). */
3781	switch (sk->sk_state) {
3782	case TCP_SYN_SENT:
3783		sk->sk_err = ECONNREFUSED;
3784		break;
3785	case TCP_CLOSE_WAIT:
3786		sk->sk_err = EPIPE;
3787		break;
3788	case TCP_CLOSE:
3789		return;
3790	default:
3791		sk->sk_err = ECONNRESET;
3792	}
3793	/* This barrier is coupled with smp_rmb() in tcp_poll() */
3794	smp_wmb();
3795
3796	if (!sock_flag(sk, SOCK_DEAD))
3797		sk->sk_error_report(sk);
3798
3799	tcp_done(sk);
3800}
3801
3802/*
3803 * 	Process the FIN bit. This now behaves as it is supposed to work
3804 *	and the FIN takes effect when it is validly part of sequence
3805 *	space. Not before when we get holes.
3806 *
3807 *	If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT
3808 *	(and thence onto LAST-ACK and finally, CLOSE, we never enter
3809 *	TIME-WAIT)
3810 *
3811 *	If we are in FINWAIT-1, a received FIN indicates simultaneous
3812 *	close and we go into CLOSING (and later onto TIME-WAIT)
3813 *
3814 *	If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT.
3815 */
3816static void tcp_fin(struct sock *sk)
3817{
3818	struct tcp_sock *tp = tcp_sk(sk);
3819	const struct dst_entry *dst;
3820
3821	inet_csk_schedule_ack(sk);
3822
3823	sk->sk_shutdown |= RCV_SHUTDOWN;
3824	sock_set_flag(sk, SOCK_DONE);
3825
3826	switch (sk->sk_state) {
3827	case TCP_SYN_RECV:
3828	case TCP_ESTABLISHED:
3829		/* Move to CLOSE_WAIT */
3830		tcp_set_state(sk, TCP_CLOSE_WAIT);
3831		dst = __sk_dst_get(sk);
3832		if (!dst || !dst_metric(dst, RTAX_QUICKACK))
3833			inet_csk(sk)->icsk_ack.pingpong = 1;
3834		break;
3835
3836	case TCP_CLOSE_WAIT:
3837	case TCP_CLOSING:
3838		/* Received a retransmission of the FIN, do
3839		 * nothing.
3840		 */
3841		break;
3842	case TCP_LAST_ACK:
3843		/* RFC793: Remain in the LAST-ACK state. */
3844		break;
3845
3846	case TCP_FIN_WAIT1:
3847		/* This case occurs when a simultaneous close
3848		 * happens, we must ack the received FIN and
3849		 * enter the CLOSING state.
3850		 */
3851		tcp_send_ack(sk);
3852		tcp_set_state(sk, TCP_CLOSING);
3853		break;
3854	case TCP_FIN_WAIT2:
3855		/* Received a FIN -- send ACK and enter TIME_WAIT. */
3856		tcp_send_ack(sk);
3857		tcp_time_wait(sk, TCP_TIME_WAIT, 0);
3858		break;
3859	default:
3860		/* Only TCP_LISTEN and TCP_CLOSE are left, in these
3861		 * cases we should never reach this piece of code.
3862		 */
3863		pr_err("%s: Impossible, sk->sk_state=%d\n",
3864		       __func__, sk->sk_state);
3865		break;
3866	}
3867
3868	/* It _is_ possible, that we have something out-of-order _after_ FIN.
3869	 * Probably, we should reset in this case. For now drop them.
3870	 */
3871	__skb_queue_purge(&tp->out_of_order_queue);
3872	if (tcp_is_sack(tp))
3873		tcp_sack_reset(&tp->rx_opt);
3874	sk_mem_reclaim(sk);
3875
3876	if (!sock_flag(sk, SOCK_DEAD)) {
3877		sk->sk_state_change(sk);
3878
3879		/* Do not send POLL_HUP for half duplex close. */
3880		if (sk->sk_shutdown == SHUTDOWN_MASK ||
3881		    sk->sk_state == TCP_CLOSE)
3882			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
3883		else
3884			sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
3885	}
3886}
3887
3888static inline bool tcp_sack_extend(struct tcp_sack_block *sp, u32 seq,
3889				  u32 end_seq)
3890{
3891	if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) {
3892		if (before(seq, sp->start_seq))
3893			sp->start_seq = seq;
3894		if (after(end_seq, sp->end_seq))
3895			sp->end_seq = end_seq;
3896		return true;
3897	}
3898	return false;
3899}
3900
3901static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq)
3902{
3903	struct tcp_sock *tp = tcp_sk(sk);
3904
3905	if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3906		int mib_idx;
3907
3908		if (before(seq, tp->rcv_nxt))
3909			mib_idx = LINUX_MIB_TCPDSACKOLDSENT;
3910		else
3911			mib_idx = LINUX_MIB_TCPDSACKOFOSENT;
3912
3913		NET_INC_STATS_BH(sock_net(sk), mib_idx);
3914
3915		tp->rx_opt.dsack = 1;
3916		tp->duplicate_sack[0].start_seq = seq;
3917		tp->duplicate_sack[0].end_seq = end_seq;
3918	}
3919}
3920
3921static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq)
3922{
3923	struct tcp_sock *tp = tcp_sk(sk);
3924
3925	if (!tp->rx_opt.dsack)
3926		tcp_dsack_set(sk, seq, end_seq);
3927	else
3928		tcp_sack_extend(tp->duplicate_sack, seq, end_seq);
3929}
3930
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3931static void tcp_send_dupack(struct sock *sk, const struct sk_buff *skb)
3932{
3933	struct tcp_sock *tp = tcp_sk(sk);
3934
3935	if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
3936	    before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
3937		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
3938		tcp_enter_quickack_mode(sk);
3939
3940		if (tcp_is_sack(tp) && sysctl_tcp_dsack) {
3941			u32 end_seq = TCP_SKB_CB(skb)->end_seq;
3942
 
3943			if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt))
3944				end_seq = tp->rcv_nxt;
3945			tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq);
3946		}
3947	}
3948
3949	tcp_send_ack(sk);
3950}
3951
3952/* These routines update the SACK block as out-of-order packets arrive or
3953 * in-order packets close up the sequence space.
3954 */
3955static void tcp_sack_maybe_coalesce(struct tcp_sock *tp)
3956{
3957	int this_sack;
3958	struct tcp_sack_block *sp = &tp->selective_acks[0];
3959	struct tcp_sack_block *swalk = sp + 1;
3960
3961	/* See if the recent change to the first SACK eats into
3962	 * or hits the sequence space of other SACK blocks, if so coalesce.
3963	 */
3964	for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) {
3965		if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) {
3966			int i;
3967
3968			/* Zap SWALK, by moving every further SACK up by one slot.
3969			 * Decrease num_sacks.
3970			 */
3971			tp->rx_opt.num_sacks--;
3972			for (i = this_sack; i < tp->rx_opt.num_sacks; i++)
3973				sp[i] = sp[i + 1];
3974			continue;
3975		}
3976		this_sack++, swalk++;
 
3977	}
3978}
3979
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3980static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq)
3981{
3982	struct tcp_sock *tp = tcp_sk(sk);
3983	struct tcp_sack_block *sp = &tp->selective_acks[0];
3984	int cur_sacks = tp->rx_opt.num_sacks;
3985	int this_sack;
3986
3987	if (!cur_sacks)
3988		goto new_sack;
3989
3990	for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) {
3991		if (tcp_sack_extend(sp, seq, end_seq)) {
 
 
3992			/* Rotate this_sack to the first one. */
3993			for (; this_sack > 0; this_sack--, sp--)
3994				swap(*sp, *(sp - 1));
3995			if (cur_sacks > 1)
3996				tcp_sack_maybe_coalesce(tp);
3997			return;
3998		}
3999	}
4000
 
 
 
4001	/* Could not find an adjacent existing SACK, build a new one,
4002	 * put it at the front, and shift everyone else down.  We
4003	 * always know there is at least one SACK present already here.
4004	 *
4005	 * If the sack array is full, forget about the last one.
4006	 */
4007	if (this_sack >= TCP_NUM_SACKS) {
4008		this_sack--;
4009		tp->rx_opt.num_sacks--;
4010		sp--;
4011	}
4012	for (; this_sack > 0; this_sack--, sp--)
4013		*sp = *(sp - 1);
4014
4015new_sack:
4016	/* Build the new head SACK, and we're done. */
4017	sp->start_seq = seq;
4018	sp->end_seq = end_seq;
4019	tp->rx_opt.num_sacks++;
4020}
4021
4022/* RCV.NXT advances, some SACKs should be eaten. */
4023
4024static void tcp_sack_remove(struct tcp_sock *tp)
4025{
4026	struct tcp_sack_block *sp = &tp->selective_acks[0];
4027	int num_sacks = tp->rx_opt.num_sacks;
4028	int this_sack;
4029
4030	/* Empty ofo queue, hence, all the SACKs are eaten. Clear. */
4031	if (skb_queue_empty(&tp->out_of_order_queue)) {
4032		tp->rx_opt.num_sacks = 0;
4033		return;
4034	}
4035
4036	for (this_sack = 0; this_sack < num_sacks;) {
4037		/* Check if the start of the sack is covered by RCV.NXT. */
4038		if (!before(tp->rcv_nxt, sp->start_seq)) {
4039			int i;
4040
4041			/* RCV.NXT must cover all the block! */
4042			WARN_ON(before(tp->rcv_nxt, sp->end_seq));
4043
4044			/* Zap this SACK, by moving forward any other SACKS. */
4045			for (i = this_sack+1; i < num_sacks; i++)
4046				tp->selective_acks[i-1] = tp->selective_acks[i];
4047			num_sacks--;
4048			continue;
4049		}
4050		this_sack++;
4051		sp++;
4052	}
4053	tp->rx_opt.num_sacks = num_sacks;
4054}
4055
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4056/* This one checks to see if we can put data from the
4057 * out_of_order queue into the receive_queue.
4058 */
4059static void tcp_ofo_queue(struct sock *sk)
4060{
4061	struct tcp_sock *tp = tcp_sk(sk);
4062	__u32 dsack_high = tp->rcv_nxt;
4063	struct sk_buff *skb;
4064
4065	while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) {
 
 
 
 
4066		if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
4067			break;
4068
4069		if (before(TCP_SKB_CB(skb)->seq, dsack_high)) {
4070			__u32 dsack = dsack_high;
4071			if (before(TCP_SKB_CB(skb)->end_seq, dsack_high))
4072				dsack_high = TCP_SKB_CB(skb)->end_seq;
4073			tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack);
4074		}
 
 
4075
4076		if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
4077			SOCK_DEBUG(sk, "ofo packet was already received\n");
4078			__skb_unlink(skb, &tp->out_of_order_queue);
4079			__kfree_skb(skb);
4080			continue;
4081		}
4082		SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n",
4083			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4084			   TCP_SKB_CB(skb)->end_seq);
4085
4086		__skb_unlink(skb, &tp->out_of_order_queue);
4087		__skb_queue_tail(&sk->sk_receive_queue, skb);
4088		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4089		if (tcp_hdr(skb)->fin)
 
 
 
4090			tcp_fin(sk);
 
 
 
 
 
4091	}
4092}
4093
4094static bool tcp_prune_ofo_queue(struct sock *sk);
4095static int tcp_prune_queue(struct sock *sk);
4096
4097static int tcp_try_rmem_schedule(struct sock *sk, struct sk_buff *skb,
4098				 unsigned int size)
4099{
4100	if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf ||
4101	    !sk_rmem_schedule(sk, skb, size)) {
4102
4103		if (tcp_prune_queue(sk) < 0)
4104			return -1;
4105
4106		if (!sk_rmem_schedule(sk, skb, size)) {
4107			if (!tcp_prune_ofo_queue(sk))
4108				return -1;
4109
4110			if (!sk_rmem_schedule(sk, skb, size))
4111				return -1;
4112		}
4113	}
4114	return 0;
4115}
4116
4117/**
4118 * tcp_try_coalesce - try to merge skb to prior one
4119 * @sk: socket
4120 * @to: prior buffer
4121 * @from: buffer to add in queue
4122 * @fragstolen: pointer to boolean
4123 *
4124 * Before queueing skb @from after @to, try to merge them
4125 * to reduce overall memory use and queue lengths, if cost is small.
4126 * Packets in ofo or receive queues can stay a long time.
4127 * Better try to coalesce them right now to avoid future collapses.
4128 * Returns true if caller should free @from instead of queueing it
4129 */
4130static bool tcp_try_coalesce(struct sock *sk,
4131			     struct sk_buff *to,
4132			     struct sk_buff *from,
4133			     bool *fragstolen)
4134{
4135	int delta;
4136
4137	*fragstolen = false;
4138
4139	if (tcp_hdr(from)->fin)
4140		return false;
4141
4142	/* Its possible this segment overlaps with prior segment in queue */
4143	if (TCP_SKB_CB(from)->seq != TCP_SKB_CB(to)->end_seq)
4144		return false;
4145
4146	if (!skb_try_coalesce(to, from, fragstolen, &delta))
4147		return false;
4148
4149	atomic_add(delta, &sk->sk_rmem_alloc);
4150	sk_mem_charge(sk, delta);
4151	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOALESCE);
4152	TCP_SKB_CB(to)->end_seq = TCP_SKB_CB(from)->end_seq;
4153	TCP_SKB_CB(to)->ack_seq = TCP_SKB_CB(from)->ack_seq;
4154	return true;
4155}
4156
4157static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb)
4158{
4159	struct tcp_sock *tp = tcp_sk(sk);
 
4160	struct sk_buff *skb1;
4161	u32 seq, end_seq;
 
4162
4163	TCP_ECN_check_ce(tp, skb);
 
4164
4165	if (unlikely(tcp_try_rmem_schedule(sk, skb, skb->truesize))) {
4166		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFODROP);
4167		__kfree_skb(skb);
 
4168		return;
4169	}
4170
4171	/* Disable header prediction. */
4172	tp->pred_flags = 0;
4173	inet_csk_schedule_ack(sk);
4174
4175	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOQUEUE);
4176	SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n",
4177		   tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
 
4178
4179	skb1 = skb_peek_tail(&tp->out_of_order_queue);
4180	if (!skb1) {
4181		/* Initial out of order segment, build 1 SACK. */
4182		if (tcp_is_sack(tp)) {
4183			tp->rx_opt.num_sacks = 1;
4184			tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq;
4185			tp->selective_acks[0].end_seq =
4186						TCP_SKB_CB(skb)->end_seq;
4187		}
4188		__skb_queue_head(&tp->out_of_order_queue, skb);
 
 
4189		goto end;
4190	}
4191
4192	seq = TCP_SKB_CB(skb)->seq;
4193	end_seq = TCP_SKB_CB(skb)->end_seq;
4194
4195	if (seq == TCP_SKB_CB(skb1)->end_seq) {
4196		bool fragstolen;
4197
4198		if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) {
4199			__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4200		} else {
4201			tcp_grow_window(sk, skb);
4202			kfree_skb_partial(skb, fragstolen);
4203			skb = NULL;
4204		}
4205
4206		if (!tp->rx_opt.num_sacks ||
4207		    tp->selective_acks[0].end_seq != seq)
4208			goto add_sack;
4209
4210		/* Common case: data arrive in order after hole. */
4211		tp->selective_acks[0].end_seq = end_seq;
4212		goto end;
4213	}
4214
4215	/* Find place to insert this segment. */
4216	while (1) {
4217		if (!after(TCP_SKB_CB(skb1)->seq, seq))
4218			break;
4219		if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) {
4220			skb1 = NULL;
4221			break;
4222		}
4223		skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1);
4224	}
4225
4226	/* Do skb overlap to previous one? */
4227	if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) {
4228		if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4229			/* All the bits are present. Drop. */
4230			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4231			__kfree_skb(skb);
4232			skb = NULL;
4233			tcp_dsack_set(sk, seq, end_seq);
4234			goto add_sack;
4235		}
4236		if (after(seq, TCP_SKB_CB(skb1)->seq)) {
4237			/* Partial overlap. */
4238			tcp_dsack_set(sk, seq,
4239				      TCP_SKB_CB(skb1)->end_seq);
4240		} else {
4241			if (skb_queue_is_first(&tp->out_of_order_queue,
4242					       skb1))
4243				skb1 = NULL;
4244			else
4245				skb1 = skb_queue_prev(
4246					&tp->out_of_order_queue,
4247					skb1);
4248		}
4249	}
4250	if (!skb1)
4251		__skb_queue_head(&tp->out_of_order_queue, skb);
4252	else
4253		__skb_queue_after(&tp->out_of_order_queue, skb1, skb);
4254
4255	/* And clean segments covered by new one as whole. */
4256	while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) {
4257		skb1 = skb_queue_next(&tp->out_of_order_queue, skb);
4258
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4259		if (!after(end_seq, TCP_SKB_CB(skb1)->seq))
4260			break;
4261		if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) {
4262			tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4263					 end_seq);
4264			break;
4265		}
4266		__skb_unlink(skb1, &tp->out_of_order_queue);
4267		tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq,
4268				 TCP_SKB_CB(skb1)->end_seq);
4269		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPOFOMERGE);
4270		__kfree_skb(skb1);
4271	}
 
 
 
4272
4273add_sack:
4274	if (tcp_is_sack(tp))
4275		tcp_sack_new_ofo_skb(sk, seq, end_seq);
4276end:
4277	if (skb) {
4278		tcp_grow_window(sk, skb);
 
 
 
 
 
4279		skb_set_owner_r(skb, sk);
4280	}
4281}
4282
4283static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen,
4284		  bool *fragstolen)
4285{
4286	int eaten;
4287	struct sk_buff *tail = skb_peek_tail(&sk->sk_receive_queue);
4288
4289	__skb_pull(skb, hdrlen);
4290	eaten = (tail &&
4291		 tcp_try_coalesce(sk, tail, skb, fragstolen)) ? 1 : 0;
4292	tcp_sk(sk)->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
 
4293	if (!eaten) {
4294		__skb_queue_tail(&sk->sk_receive_queue, skb);
4295		skb_set_owner_r(skb, sk);
4296	}
4297	return eaten;
4298}
4299
4300int tcp_send_rcvq(struct sock *sk, struct msghdr *msg, size_t size)
4301{
4302	struct sk_buff *skb = NULL;
4303	struct tcphdr *th;
 
4304	bool fragstolen;
4305
4306	if (size == 0)
4307		return 0;
4308
4309	skb = alloc_skb(size + sizeof(*th), sk->sk_allocation);
 
 
 
 
 
 
 
 
4310	if (!skb)
4311		goto err;
4312
4313	if (tcp_try_rmem_schedule(sk, skb, size + sizeof(*th)))
4314		goto err_free;
 
4315
4316	th = (struct tcphdr *)skb_put(skb, sizeof(*th));
4317	skb_reset_transport_header(skb);
4318	memset(th, 0, sizeof(*th));
 
4319
4320	if (memcpy_fromiovec(skb_put(skb, size), msg->msg_iov, size))
 
4321		goto err_free;
4322
4323	TCP_SKB_CB(skb)->seq = tcp_sk(sk)->rcv_nxt;
4324	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + size;
4325	TCP_SKB_CB(skb)->ack_seq = tcp_sk(sk)->snd_una - 1;
4326
4327	if (tcp_queue_rcv(sk, skb, sizeof(*th), &fragstolen)) {
4328		WARN_ON_ONCE(fragstolen); /* should not happen */
4329		__kfree_skb(skb);
4330	}
4331	return size;
4332
4333err_free:
4334	kfree_skb(skb);
4335err:
4336	return -ENOMEM;
 
 
 
 
 
 
 
4337}
4338
4339static void tcp_data_queue(struct sock *sk, struct sk_buff *skb)
4340{
4341	const struct tcphdr *th = tcp_hdr(skb);
4342	struct tcp_sock *tp = tcp_sk(sk);
4343	int eaten = -1;
4344	bool fragstolen = false;
4345
4346	if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq)
4347		goto drop;
4348
4349	skb_dst_drop(skb);
4350	__skb_pull(skb, th->doff * 4);
 
 
 
 
 
4351
4352	TCP_ECN_accept_cwr(tp, skb);
 
 
 
 
 
4353
 
4354	tp->rx_opt.dsack = 0;
4355
4356	/*  Queue data for delivery to the user.
4357	 *  Packets in sequence go to the receive queue.
4358	 *  Out of sequence packets to the out_of_order_queue.
4359	 */
4360	if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) {
4361		if (tcp_receive_window(tp) == 0)
 
 
 
 
 
 
 
 
 
 
 
 
4362			goto out_of_window;
 
4363
4364		/* Ok. In sequence. In window. */
4365		if (tp->ucopy.task == current &&
4366		    tp->copied_seq == tp->rcv_nxt && tp->ucopy.len &&
4367		    sock_owned_by_user(sk) && !tp->urg_data) {
4368			int chunk = min_t(unsigned int, skb->len,
4369					  tp->ucopy.len);
4370
4371			__set_current_state(TASK_RUNNING);
4372
4373			local_bh_enable();
4374			if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) {
4375				tp->ucopy.len -= chunk;
4376				tp->copied_seq += chunk;
4377				eaten = (chunk == skb->len);
4378				tcp_rcv_space_adjust(sk);
4379			}
4380			local_bh_disable();
4381		}
4382
4383		if (eaten <= 0) {
4384queue_and_out:
4385			if (eaten < 0 &&
4386			    tcp_try_rmem_schedule(sk, skb, skb->truesize))
4387				goto drop;
4388
4389			eaten = tcp_queue_rcv(sk, skb, 0, &fragstolen);
4390		}
4391		tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
4392		if (skb->len)
4393			tcp_event_data_recv(sk, skb);
4394		if (th->fin)
4395			tcp_fin(sk);
4396
4397		if (!skb_queue_empty(&tp->out_of_order_queue)) {
4398			tcp_ofo_queue(sk);
4399
4400			/* RFC2581. 4.2. SHOULD send immediate ACK, when
4401			 * gap in queue is filled.
4402			 */
4403			if (skb_queue_empty(&tp->out_of_order_queue))
4404				inet_csk(sk)->icsk_ack.pingpong = 0;
4405		}
4406
4407		if (tp->rx_opt.num_sacks)
4408			tcp_sack_remove(tp);
4409
4410		tcp_fast_path_check(sk);
4411
4412		if (eaten > 0)
4413			kfree_skb_partial(skb, fragstolen);
4414		if (!sock_flag(sk, SOCK_DEAD))
4415			sk->sk_data_ready(sk);
4416		return;
4417	}
4418
4419	if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) {
 
4420		/* A retransmit, 2nd most common case.  Force an immediate ack. */
4421		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST);
 
4422		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq);
4423
4424out_of_window:
4425		tcp_enter_quickack_mode(sk);
4426		inet_csk_schedule_ack(sk);
4427drop:
4428		__kfree_skb(skb);
4429		return;
4430	}
4431
4432	/* Out of window. F.e. zero window probe. */
4433	if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp)))
 
 
4434		goto out_of_window;
4435
4436	tcp_enter_quickack_mode(sk);
4437
4438	if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) {
4439		/* Partial packet, seq < rcv_next < end_seq */
4440		SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n",
4441			   tp->rcv_nxt, TCP_SKB_CB(skb)->seq,
4442			   TCP_SKB_CB(skb)->end_seq);
4443
4444		tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt);
4445
4446		/* If window is closed, drop tail of packet. But after
4447		 * remembering D-SACK for its head made in previous line.
4448		 */
4449		if (!tcp_receive_window(tp))
 
 
4450			goto out_of_window;
 
4451		goto queue_and_out;
4452	}
4453
4454	tcp_data_queue_ofo(sk, skb);
4455}
4456
 
 
 
 
 
 
 
 
4457static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb,
4458					struct sk_buff_head *list)
 
4459{
4460	struct sk_buff *next = NULL;
4461
4462	if (!skb_queue_is_last(list, skb))
4463		next = skb_queue_next(list, skb);
 
 
4464
4465	__skb_unlink(skb, list);
4466	__kfree_skb(skb);
4467	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED);
4468
4469	return next;
4470}
4471
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4472/* Collapse contiguous sequence of skbs head..tail with
4473 * sequence numbers start..end.
4474 *
4475 * If tail is NULL, this means until the end of the list.
4476 *
4477 * Segments with FIN/SYN are not collapsed (only because this
4478 * simplifies code)
4479 */
4480static void
4481tcp_collapse(struct sock *sk, struct sk_buff_head *list,
4482	     struct sk_buff *head, struct sk_buff *tail,
4483	     u32 start, u32 end)
4484{
4485	struct sk_buff *skb, *n;
 
4486	bool end_of_skbs;
4487
4488	/* First, check that queue is collapsible and find
4489	 * the point where collapsing can be useful. */
4490	skb = head;
4491restart:
4492	end_of_skbs = true;
4493	skb_queue_walk_from_safe(list, skb, n) {
4494		if (skb == tail)
4495			break;
 
 
4496		/* No new bits? It is possible on ofo queue. */
4497		if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4498			skb = tcp_collapse_one(sk, skb, list);
4499			if (!skb)
4500				break;
4501			goto restart;
4502		}
4503
4504		/* The first skb to collapse is:
4505		 * - not SYN/FIN and
4506		 * - bloated or contains data before "start" or
4507		 *   overlaps to the next one.
4508		 */
4509		if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin &&
4510		    (tcp_win_from_space(skb->truesize) > skb->len ||
4511		     before(TCP_SKB_CB(skb)->seq, start))) {
4512			end_of_skbs = false;
4513			break;
4514		}
4515
4516		if (!skb_queue_is_last(list, skb)) {
4517			struct sk_buff *next = skb_queue_next(list, skb);
4518			if (next != tail &&
4519			    TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) {
4520				end_of_skbs = false;
4521				break;
4522			}
4523		}
4524
 
4525		/* Decided to skip this, advance start seq. */
4526		start = TCP_SKB_CB(skb)->end_seq;
4527	}
4528	if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin)
 
 
4529		return;
4530
 
 
4531	while (before(start, end)) {
 
4532		struct sk_buff *nskb;
4533		unsigned int header = skb_headroom(skb);
4534		int copy = SKB_MAX_ORDER(header, 0);
4535
4536		/* Too big header? This can happen with IPv6. */
4537		if (copy < 0)
4538			return;
4539		if (end - start < copy)
4540			copy = end - start;
4541		nskb = alloc_skb(copy + header, GFP_ATOMIC);
4542		if (!nskb)
4543			return;
4544
4545		skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head);
4546		skb_set_network_header(nskb, (skb_network_header(skb) -
4547					      skb->head));
4548		skb_set_transport_header(nskb, (skb_transport_header(skb) -
4549						skb->head));
4550		skb_reserve(nskb, header);
4551		memcpy(nskb->head, skb->head, header);
4552		memcpy(nskb->cb, skb->cb, sizeof(skb->cb));
 
4553		TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start;
4554		__skb_queue_before(list, skb, nskb);
 
 
 
4555		skb_set_owner_r(nskb, sk);
 
4556
4557		/* Copy data, releasing collapsed skbs. */
4558		while (copy > 0) {
4559			int offset = start - TCP_SKB_CB(skb)->seq;
4560			int size = TCP_SKB_CB(skb)->end_seq - start;
4561
4562			BUG_ON(offset < 0);
4563			if (size > 0) {
4564				size = min(copy, size);
4565				if (skb_copy_bits(skb, offset, skb_put(nskb, size), size))
4566					BUG();
4567				TCP_SKB_CB(nskb)->end_seq += size;
4568				copy -= size;
4569				start += size;
4570			}
4571			if (!before(start, TCP_SKB_CB(skb)->end_seq)) {
4572				skb = tcp_collapse_one(sk, skb, list);
4573				if (!skb ||
4574				    skb == tail ||
4575				    tcp_hdr(skb)->syn ||
4576				    tcp_hdr(skb)->fin)
4577					return;
 
4578			}
4579		}
4580	}
 
 
 
4581}
4582
4583/* Collapse ofo queue. Algorithm: select contiguous sequence of skbs
4584 * and tcp_collapse() them until all the queue is collapsed.
4585 */
4586static void tcp_collapse_ofo_queue(struct sock *sk)
4587{
4588	struct tcp_sock *tp = tcp_sk(sk);
4589	struct sk_buff *skb = skb_peek(&tp->out_of_order_queue);
4590	struct sk_buff *head;
4591	u32 start, end;
4592
4593	if (skb == NULL)
 
 
 
4594		return;
4595
4596	start = TCP_SKB_CB(skb)->seq;
4597	end = TCP_SKB_CB(skb)->end_seq;
4598	head = skb;
4599
4600	for (;;) {
4601		struct sk_buff *next = NULL;
4602
4603		if (!skb_queue_is_last(&tp->out_of_order_queue, skb))
4604			next = skb_queue_next(&tp->out_of_order_queue, skb);
4605		skb = next;
4606
4607		/* Segment is terminated when we see gap or when
4608		 * we are at the end of all the queue. */
 
4609		if (!skb ||
4610		    after(TCP_SKB_CB(skb)->seq, end) ||
4611		    before(TCP_SKB_CB(skb)->end_seq, start)) {
4612			tcp_collapse(sk, &tp->out_of_order_queue,
4613				     head, skb, start, end);
4614			head = skb;
4615			if (!skb)
4616				break;
4617			/* Start new segment */
 
 
 
 
 
 
 
 
 
4618			start = TCP_SKB_CB(skb)->seq;
 
4619			end = TCP_SKB_CB(skb)->end_seq;
4620		} else {
4621			if (before(TCP_SKB_CB(skb)->seq, start))
4622				start = TCP_SKB_CB(skb)->seq;
4623			if (after(TCP_SKB_CB(skb)->end_seq, end))
4624				end = TCP_SKB_CB(skb)->end_seq;
4625		}
4626	}
4627}
4628
4629/*
4630 * Purge the out-of-order queue.
4631 * Return true if queue was pruned.
4632 */
4633static bool tcp_prune_ofo_queue(struct sock *sk)
4634{
4635	struct tcp_sock *tp = tcp_sk(sk);
4636	bool res = false;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4637
4638	if (!skb_queue_empty(&tp->out_of_order_queue)) {
4639		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED);
4640		__skb_queue_purge(&tp->out_of_order_queue);
4641
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4642		/* Reset SACK state.  A conforming SACK implementation will
4643		 * do the same at a timeout based retransmit.  When a connection
4644		 * is in a sad state like this, we care only about integrity
4645		 * of the connection not performance.
4646		 */
4647		if (tp->rx_opt.sack_ok)
4648			tcp_sack_reset(&tp->rx_opt);
4649		sk_mem_reclaim(sk);
4650		res = true;
4651	}
4652	return res;
4653}
4654
4655/* Reduce allocated memory if we can, trying to get
4656 * the socket within its memory limits again.
4657 *
4658 * Return less than zero if we should start dropping frames
4659 * until the socket owning process reads some of the data
4660 * to stabilize the situation.
4661 */
4662static int tcp_prune_queue(struct sock *sk)
4663{
4664	struct tcp_sock *tp = tcp_sk(sk);
4665
4666	SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq);
4667
4668	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED);
4669
4670	if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf)
4671		tcp_clamp_window(sk);
4672	else if (sk_under_memory_pressure(sk))
4673		tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss);
 
 
 
4674
4675	tcp_collapse_ofo_queue(sk);
4676	if (!skb_queue_empty(&sk->sk_receive_queue))
4677		tcp_collapse(sk, &sk->sk_receive_queue,
4678			     skb_peek(&sk->sk_receive_queue),
4679			     NULL,
4680			     tp->copied_seq, tp->rcv_nxt);
4681	sk_mem_reclaim(sk);
4682
4683	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4684		return 0;
4685
4686	/* Collapsing did not help, destructive actions follow.
4687	 * This must not ever occur. */
4688
4689	tcp_prune_ofo_queue(sk);
4690
4691	if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf)
4692		return 0;
4693
4694	/* If we are really being abused, tell the caller to silently
4695	 * drop receive data on the floor.  It will get retransmitted
4696	 * and hopefully then we'll have sufficient space.
4697	 */
4698	NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED);
4699
4700	/* Massive buffer overcommit. */
4701	tp->pred_flags = 0;
4702	return -1;
4703}
4704
4705/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
4706 * As additional protections, we do not touch cwnd in retransmission phases,
4707 * and if application hit its sndbuf limit recently.
4708 */
4709void tcp_cwnd_application_limited(struct sock *sk)
4710{
4711	struct tcp_sock *tp = tcp_sk(sk);
4712
4713	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
4714	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
4715		/* Limited by application or receiver window. */
4716		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
4717		u32 win_used = max(tp->snd_cwnd_used, init_win);
4718		if (win_used < tp->snd_cwnd) {
4719			tp->snd_ssthresh = tcp_current_ssthresh(sk);
4720			tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
4721		}
4722		tp->snd_cwnd_used = 0;
4723	}
4724	tp->snd_cwnd_stamp = tcp_time_stamp;
4725}
4726
4727static bool tcp_should_expand_sndbuf(const struct sock *sk)
4728{
4729	const struct tcp_sock *tp = tcp_sk(sk);
4730
4731	/* If the user specified a specific send buffer setting, do
4732	 * not modify it.
4733	 */
4734	if (sk->sk_userlocks & SOCK_SNDBUF_LOCK)
4735		return false;
4736
4737	/* If we are under global TCP memory pressure, do not expand.  */
4738	if (sk_under_memory_pressure(sk))
 
 
 
 
 
 
 
 
 
4739		return false;
 
4740
4741	/* If we are under soft global TCP memory pressure, do not expand.  */
4742	if (sk_memory_allocated(sk) >= sk_prot_mem_limits(sk, 0))
4743		return false;
4744
4745	/* If we filled the congestion window, do not expand.  */
4746	if (tp->packets_out >= tp->snd_cwnd)
4747		return false;
4748
4749	return true;
4750}
4751
4752/* When incoming ACK allowed to free some skb from write_queue,
4753 * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket
4754 * on the exit from tcp input handler.
4755 *
4756 * PROBLEM: sndbuf expansion does not work well with largesend.
4757 */
4758static void tcp_new_space(struct sock *sk)
4759{
4760	struct tcp_sock *tp = tcp_sk(sk);
4761
4762	if (tcp_should_expand_sndbuf(sk)) {
4763		tcp_sndbuf_expand(sk);
4764		tp->snd_cwnd_stamp = tcp_time_stamp;
4765	}
4766
4767	sk->sk_write_space(sk);
4768}
4769
4770static void tcp_check_space(struct sock *sk)
 
 
 
 
 
 
 
 
 
 
4771{
4772	if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) {
4773		sock_reset_flag(sk, SOCK_QUEUE_SHRUNK);
4774		if (sk->sk_socket &&
4775		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags))
4776			tcp_new_space(sk);
 
 
4777	}
4778}
4779
4780static inline void tcp_data_snd_check(struct sock *sk)
4781{
4782	tcp_push_pending_frames(sk);
4783	tcp_check_space(sk);
4784}
4785
4786/*
4787 * Check if sending an ack is needed.
4788 */
4789static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible)
4790{
4791	struct tcp_sock *tp = tcp_sk(sk);
 
4792
4793	    /* More than one full frame received... */
4794	if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss &&
4795	     /* ... and right edge of window advances far enough.
4796	      * (tcp_recvmsg() will send ACK otherwise). Or...
 
 
4797	      */
4798	     __tcp_select_window(sk) >= tp->rcv_wnd) ||
 
4799	    /* We ACK each frame or... */
4800	    tcp_in_quickack_mode(sk) ||
4801	    /* We have out of order data. */
4802	    (ofo_possible && skb_peek(&tp->out_of_order_queue))) {
4803		/* Then ack it now */
 
 
 
 
 
 
 
 
4804		tcp_send_ack(sk);
4805	} else {
4806		/* Else, send delayed ack. */
 
 
4807		tcp_send_delayed_ack(sk);
 
4808	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4809}
4810
4811static inline void tcp_ack_snd_check(struct sock *sk)
4812{
4813	if (!inet_csk_ack_scheduled(sk)) {
4814		/* We sent a data segment already. */
4815		return;
4816	}
4817	__tcp_ack_snd_check(sk, 1);
4818}
4819
4820/*
4821 *	This routine is only called when we have urgent data
4822 *	signaled. Its the 'slow' part of tcp_urg. It could be
4823 *	moved inline now as tcp_urg is only called from one
4824 *	place. We handle URGent data wrong. We have to - as
4825 *	BSD still doesn't use the correction from RFC961.
4826 *	For 1003.1g we should support a new option TCP_STDURG to permit
4827 *	either form (or just set the sysctl tcp_stdurg).
4828 */
4829
4830static void tcp_check_urg(struct sock *sk, const struct tcphdr *th)
4831{
4832	struct tcp_sock *tp = tcp_sk(sk);
4833	u32 ptr = ntohs(th->urg_ptr);
4834
4835	if (ptr && !sysctl_tcp_stdurg)
4836		ptr--;
4837	ptr += ntohl(th->seq);
4838
4839	/* Ignore urgent data that we've already seen and read. */
4840	if (after(tp->copied_seq, ptr))
4841		return;
4842
4843	/* Do not replay urg ptr.
4844	 *
4845	 * NOTE: interesting situation not covered by specs.
4846	 * Misbehaving sender may send urg ptr, pointing to segment,
4847	 * which we already have in ofo queue. We are not able to fetch
4848	 * such data and will stay in TCP_URG_NOTYET until will be eaten
4849	 * by recvmsg(). Seems, we are not obliged to handle such wicked
4850	 * situations. But it is worth to think about possibility of some
4851	 * DoSes using some hypothetical application level deadlock.
4852	 */
4853	if (before(ptr, tp->rcv_nxt))
4854		return;
4855
4856	/* Do we already have a newer (or duplicate) urgent pointer? */
4857	if (tp->urg_data && !after(ptr, tp->urg_seq))
4858		return;
4859
4860	/* Tell the world about our new urgent pointer. */
4861	sk_send_sigurg(sk);
4862
4863	/* We may be adding urgent data when the last byte read was
4864	 * urgent. To do this requires some care. We cannot just ignore
4865	 * tp->copied_seq since we would read the last urgent byte again
4866	 * as data, nor can we alter copied_seq until this data arrives
4867	 * or we break the semantics of SIOCATMARK (and thus sockatmark())
4868	 *
4869	 * NOTE. Double Dutch. Rendering to plain English: author of comment
4870	 * above did something sort of 	send("A", MSG_OOB); send("B", MSG_OOB);
4871	 * and expect that both A and B disappear from stream. This is _wrong_.
4872	 * Though this happens in BSD with high probability, this is occasional.
4873	 * Any application relying on this is buggy. Note also, that fix "works"
4874	 * only in this artificial test. Insert some normal data between A and B and we will
4875	 * decline of BSD again. Verdict: it is better to remove to trap
4876	 * buggy users.
4877	 */
4878	if (tp->urg_seq == tp->copied_seq && tp->urg_data &&
4879	    !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) {
4880		struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
4881		tp->copied_seq++;
4882		if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) {
4883			__skb_unlink(skb, &sk->sk_receive_queue);
4884			__kfree_skb(skb);
4885		}
4886	}
4887
4888	tp->urg_data = TCP_URG_NOTYET;
4889	tp->urg_seq = ptr;
4890
4891	/* Disable header prediction. */
4892	tp->pred_flags = 0;
4893}
4894
4895/* This is the 'fast' part of urgent handling. */
4896static void tcp_urg(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th)
4897{
4898	struct tcp_sock *tp = tcp_sk(sk);
4899
4900	/* Check if we get a new urgent pointer - normally not. */
4901	if (th->urg)
4902		tcp_check_urg(sk, th);
4903
4904	/* Do we wait for any urgent data? - normally not... */
4905	if (tp->urg_data == TCP_URG_NOTYET) {
4906		u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) -
4907			  th->syn;
4908
4909		/* Is the urgent pointer pointing into this packet? */
4910		if (ptr < skb->len) {
4911			u8 tmp;
4912			if (skb_copy_bits(skb, ptr, &tmp, 1))
4913				BUG();
4914			tp->urg_data = TCP_URG_VALID | tmp;
4915			if (!sock_flag(sk, SOCK_DEAD))
4916				sk->sk_data_ready(sk);
4917		}
4918	}
4919}
4920
4921static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen)
4922{
4923	struct tcp_sock *tp = tcp_sk(sk);
4924	int chunk = skb->len - hlen;
4925	int err;
4926
4927	local_bh_enable();
4928	if (skb_csum_unnecessary(skb))
4929		err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk);
4930	else
4931		err = skb_copy_and_csum_datagram_iovec(skb, hlen,
4932						       tp->ucopy.iov);
4933
4934	if (!err) {
4935		tp->ucopy.len -= chunk;
4936		tp->copied_seq += chunk;
4937		tcp_rcv_space_adjust(sk);
4938	}
4939
4940	local_bh_disable();
4941	return err;
4942}
4943
4944static __sum16 __tcp_checksum_complete_user(struct sock *sk,
4945					    struct sk_buff *skb)
4946{
4947	__sum16 result;
4948
4949	if (sock_owned_by_user(sk)) {
4950		local_bh_enable();
4951		result = __tcp_checksum_complete(skb);
4952		local_bh_disable();
4953	} else {
4954		result = __tcp_checksum_complete(skb);
4955	}
4956	return result;
4957}
4958
4959static inline bool tcp_checksum_complete_user(struct sock *sk,
4960					     struct sk_buff *skb)
4961{
4962	return !skb_csum_unnecessary(skb) &&
4963	       __tcp_checksum_complete_user(sk, skb);
4964}
4965
4966#ifdef CONFIG_NET_DMA
4967static bool tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb,
4968				  int hlen)
4969{
4970	struct tcp_sock *tp = tcp_sk(sk);
4971	int chunk = skb->len - hlen;
4972	int dma_cookie;
4973	bool copied_early = false;
4974
4975	if (tp->ucopy.wakeup)
4976		return false;
4977
4978	if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list)
4979		tp->ucopy.dma_chan = net_dma_find_channel();
4980
4981	if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) {
4982
4983		dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan,
4984							 skb, hlen,
4985							 tp->ucopy.iov, chunk,
4986							 tp->ucopy.pinned_list);
4987
4988		if (dma_cookie < 0)
4989			goto out;
4990
4991		tp->ucopy.dma_cookie = dma_cookie;
4992		copied_early = true;
4993
4994		tp->ucopy.len -= chunk;
4995		tp->copied_seq += chunk;
4996		tcp_rcv_space_adjust(sk);
4997
4998		if ((tp->ucopy.len == 0) ||
4999		    (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) ||
5000		    (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) {
5001			tp->ucopy.wakeup = 1;
5002			sk->sk_data_ready(sk);
5003		}
5004	} else if (chunk > 0) {
5005		tp->ucopy.wakeup = 1;
5006		sk->sk_data_ready(sk);
5007	}
5008out:
5009	return copied_early;
5010}
5011#endif /* CONFIG_NET_DMA */
5012
5013/* Does PAWS and seqno based validation of an incoming segment, flags will
5014 * play significant role here.
5015 */
5016static bool tcp_validate_incoming(struct sock *sk, struct sk_buff *skb,
5017				  const struct tcphdr *th, int syn_inerr)
5018{
5019	struct tcp_sock *tp = tcp_sk(sk);
 
5020
5021	/* RFC1323: H1. Apply PAWS check first. */
5022	if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp &&
 
5023	    tcp_paws_discard(sk, skb)) {
5024		if (!th->rst) {
5025			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED);
5026			tcp_send_dupack(sk, skb);
 
 
 
 
 
 
5027			goto discard;
5028		}
5029		/* Reset is accepted even if it did not pass PAWS. */
5030	}
5031
5032	/* Step 1: check sequence number */
5033	if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) {
 
5034		/* RFC793, page 37: "In all states except SYN-SENT, all reset
5035		 * (RST) segments are validated by checking their SEQ-fields."
5036		 * And page 69: "If an incoming segment is not acceptable,
5037		 * an acknowledgment should be sent in reply (unless the RST
5038		 * bit is set, if so drop the segment and return)".
5039		 */
5040		if (!th->rst) {
5041			if (th->syn)
5042				goto syn_challenge;
5043			tcp_send_dupack(sk, skb);
 
 
 
 
 
5044		}
5045		goto discard;
5046	}
5047
5048	/* Step 2: check RST bit */
5049	if (th->rst) {
5050		/* RFC 5961 3.2 :
5051		 * If sequence number exactly matches RCV.NXT, then
 
 
 
5052		 *     RESET the connection
5053		 * else
5054		 *     Send a challenge ACK
5055		 */
5056		if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt)
5057			tcp_reset(sk);
5058		else
5059			tcp_send_challenge_ack(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5060		goto discard;
5061	}
5062
5063	/* step 3: check security and precedence [ignored] */
5064
5065	/* step 4: Check for a SYN
5066	 * RFC 5691 4.2 : Send a challenge ack
5067	 */
5068	if (th->syn) {
 
 
 
 
 
5069syn_challenge:
5070		if (syn_inerr)
5071			TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
5072		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNCHALLENGE);
5073		tcp_send_challenge_ack(sk);
 
5074		goto discard;
5075	}
5076
 
 
 
5077	return true;
5078
5079discard:
 
 
 
 
 
5080	__kfree_skb(skb);
5081	return false;
5082}
5083
5084/*
5085 *	TCP receive function for the ESTABLISHED state.
5086 *
5087 *	It is split into a fast path and a slow path. The fast path is
5088 * 	disabled when:
5089 *	- A zero window was announced from us - zero window probing
5090 *        is only handled properly in the slow path.
5091 *	- Out of order segments arrived.
5092 *	- Urgent data is expected.
5093 *	- There is no buffer space left
5094 *	- Unexpected TCP flags/window values/header lengths are received
5095 *	  (detected by checking the TCP header against pred_flags)
5096 *	- Data is sent in both directions. Fast path only supports pure senders
5097 *	  or pure receivers (this means either the sequence number or the ack
5098 *	  value must stay constant)
5099 *	- Unexpected TCP option.
5100 *
5101 *	When these conditions are not satisfied it drops into a standard
5102 *	receive procedure patterned after RFC793 to handle all cases.
5103 *	The first three cases are guaranteed by proper pred_flags setting,
5104 *	the rest is checked inline. Fast processing is turned on in
5105 *	tcp_data_queue when everything is OK.
5106 */
5107void tcp_rcv_established(struct sock *sk, struct sk_buff *skb,
5108			 const struct tcphdr *th, unsigned int len)
5109{
 
 
5110	struct tcp_sock *tp = tcp_sk(sk);
 
5111
5112	if (unlikely(sk->sk_rx_dst == NULL))
 
 
 
 
5113		inet_csk(sk)->icsk_af_ops->sk_rx_dst_set(sk, skb);
5114	/*
5115	 *	Header prediction.
5116	 *	The code loosely follows the one in the famous
5117	 *	"30 instruction TCP receive" Van Jacobson mail.
5118	 *
5119	 *	Van's trick is to deposit buffers into socket queue
5120	 *	on a device interrupt, to call tcp_recv function
5121	 *	on the receive process context and checksum and copy
5122	 *	the buffer to user space. smart...
5123	 *
5124	 *	Our current scheme is not silly either but we take the
5125	 *	extra cost of the net_bh soft interrupt processing...
5126	 *	We do checksum and copy also but from device to kernel.
5127	 */
5128
5129	tp->rx_opt.saw_tstamp = 0;
5130
5131	/*	pred_flags is 0xS?10 << 16 + snd_wnd
5132	 *	if header_prediction is to be made
5133	 *	'S' will always be tp->tcp_header_len >> 2
5134	 *	'?' will be 0 for the fast path, otherwise pred_flags is 0 to
5135	 *  turn it off	(when there are holes in the receive
5136	 *	 space for instance)
5137	 *	PSH flag is ignored.
5138	 */
5139
5140	if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags &&
5141	    TCP_SKB_CB(skb)->seq == tp->rcv_nxt &&
5142	    !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
5143		int tcp_header_len = tp->tcp_header_len;
5144
5145		/* Timestamp header prediction: tcp_header_len
5146		 * is automatically equal to th->doff*4 due to pred_flags
5147		 * match.
5148		 */
5149
5150		/* Check timestamp */
5151		if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) {
5152			/* No? Slow path! */
5153			if (!tcp_parse_aligned_timestamp(tp, th))
5154				goto slow_path;
5155
5156			/* If PAWS failed, check it more carefully in slow path */
5157			if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0)
5158				goto slow_path;
5159
5160			/* DO NOT update ts_recent here, if checksum fails
5161			 * and timestamp was corrupted part, it will result
5162			 * in a hung connection since we will drop all
5163			 * future packets due to the PAWS test.
5164			 */
5165		}
5166
5167		if (len <= tcp_header_len) {
5168			/* Bulk data transfer: sender */
5169			if (len == tcp_header_len) {
5170				/* Predicted packet is in window by definition.
5171				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5172				 * Hence, check seq<=rcv_wup reduces to:
5173				 */
5174				if (tcp_header_len ==
5175				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5176				    tp->rcv_nxt == tp->rcv_wup)
5177					tcp_store_ts_recent(tp);
5178
5179				/* We know that such packets are checksummed
5180				 * on entry.
5181				 */
5182				tcp_ack(sk, skb, 0);
5183				__kfree_skb(skb);
5184				tcp_data_snd_check(sk);
 
 
 
 
 
5185				return;
5186			} else { /* Header too small */
5187				TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
 
5188				goto discard;
5189			}
5190		} else {
5191			int eaten = 0;
5192			int copied_early = 0;
5193			bool fragstolen = false;
5194
5195			if (tp->copied_seq == tp->rcv_nxt &&
5196			    len - tcp_header_len <= tp->ucopy.len) {
5197#ifdef CONFIG_NET_DMA
5198				if (tp->ucopy.task == current &&
5199				    sock_owned_by_user(sk) &&
5200				    tcp_dma_try_early_copy(sk, skb, tcp_header_len)) {
5201					copied_early = 1;
5202					eaten = 1;
5203				}
5204#endif
5205				if (tp->ucopy.task == current &&
5206				    sock_owned_by_user(sk) && !copied_early) {
5207					__set_current_state(TASK_RUNNING);
5208
5209					if (!tcp_copy_to_iovec(sk, skb, tcp_header_len))
5210						eaten = 1;
5211				}
5212				if (eaten) {
5213					/* Predicted packet is in window by definition.
5214					 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5215					 * Hence, check seq<=rcv_wup reduces to:
5216					 */
5217					if (tcp_header_len ==
5218					    (sizeof(struct tcphdr) +
5219					     TCPOLEN_TSTAMP_ALIGNED) &&
5220					    tp->rcv_nxt == tp->rcv_wup)
5221						tcp_store_ts_recent(tp);
5222
5223					tcp_rcv_rtt_measure_ts(sk, skb);
5224
5225					__skb_pull(skb, tcp_header_len);
5226					tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
5227					NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER);
5228				}
5229				if (copied_early)
5230					tcp_cleanup_rbuf(sk, skb->len);
5231			}
5232			if (!eaten) {
5233				if (tcp_checksum_complete_user(sk, skb))
5234					goto csum_error;
5235
5236				if ((int)skb->truesize > sk->sk_forward_alloc)
5237					goto step5;
5238
5239				/* Predicted packet is in window by definition.
5240				 * seq == rcv_nxt and rcv_wup <= rcv_nxt.
5241				 * Hence, check seq<=rcv_wup reduces to:
5242				 */
5243				if (tcp_header_len ==
5244				    (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) &&
5245				    tp->rcv_nxt == tp->rcv_wup)
5246					tcp_store_ts_recent(tp);
5247
5248				tcp_rcv_rtt_measure_ts(sk, skb);
5249
5250				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS);
 
5251
5252				/* Bulk data transfer: receiver */
5253				eaten = tcp_queue_rcv(sk, skb, tcp_header_len,
5254						      &fragstolen);
5255			}
 
 
 
 
 
 
 
 
 
 
 
 
 
5256
5257			tcp_event_data_recv(sk, skb);
5258
5259			if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) {
5260				/* Well, only one small jumplet in fast path... */
5261				tcp_ack(sk, skb, FLAG_DATA);
5262				tcp_data_snd_check(sk);
5263				if (!inet_csk_ack_scheduled(sk))
5264					goto no_ack;
 
 
5265			}
5266
5267			if (!copied_early || tp->rcv_nxt != tp->rcv_wup)
5268				__tcp_ack_snd_check(sk, 0);
5269no_ack:
5270#ifdef CONFIG_NET_DMA
5271			if (copied_early)
5272				__skb_queue_tail(&sk->sk_async_wait_queue, skb);
5273			else
5274#endif
5275			if (eaten)
5276				kfree_skb_partial(skb, fragstolen);
5277			sk->sk_data_ready(sk);
5278			return;
5279		}
5280	}
5281
5282slow_path:
5283	if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb))
5284		goto csum_error;
5285
5286	if (!th->ack && !th->rst)
 
5287		goto discard;
 
5288
5289	/*
5290	 *	Standard slow path.
5291	 */
5292
5293	if (!tcp_validate_incoming(sk, skb, th, 1))
5294		return;
5295
5296step5:
5297	if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0)
 
 
5298		goto discard;
5299
5300	tcp_rcv_rtt_measure_ts(sk, skb);
5301
5302	/* Process urgent data. */
5303	tcp_urg(sk, skb, th);
5304
5305	/* step 7: process the segment text */
5306	tcp_data_queue(sk, skb);
5307
5308	tcp_data_snd_check(sk);
5309	tcp_ack_snd_check(sk);
5310	return;
5311
5312csum_error:
5313	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS);
5314	TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS);
 
 
5315
5316discard:
5317	__kfree_skb(skb);
5318}
5319EXPORT_SYMBOL(tcp_rcv_established);
5320
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5321void tcp_finish_connect(struct sock *sk, struct sk_buff *skb)
5322{
5323	struct tcp_sock *tp = tcp_sk(sk);
5324	struct inet_connection_sock *icsk = inet_csk(sk);
5325
 
5326	tcp_set_state(sk, TCP_ESTABLISHED);
 
5327
5328	if (skb != NULL) {
5329		icsk->icsk_af_ops->sk_rx_dst_set(sk, skb);
5330		security_inet_conn_established(sk, skb);
 
5331	}
5332
5333	/* Make sure socket is routed, for correct metrics.  */
5334	icsk->icsk_af_ops->rebuild_header(sk);
5335
5336	tcp_init_metrics(sk);
5337
5338	tcp_init_congestion_control(sk);
5339
5340	/* Prevent spurious tcp_cwnd_restart() on first data
5341	 * packet.
5342	 */
5343	tp->lsndtime = tcp_time_stamp;
5344
5345	tcp_init_buffer_space(sk);
5346
5347	if (sock_flag(sk, SOCK_KEEPOPEN))
5348		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp));
5349
5350	if (!tp->rx_opt.snd_wscale)
5351		__tcp_fast_path_on(tp, tp->snd_wnd);
5352	else
5353		tp->pred_flags = 0;
5354
5355	if (!sock_flag(sk, SOCK_DEAD)) {
5356		sk->sk_state_change(sk);
5357		sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5358	}
5359}
5360
5361static bool tcp_rcv_fastopen_synack(struct sock *sk, struct sk_buff *synack,
5362				    struct tcp_fastopen_cookie *cookie)
5363{
5364	struct tcp_sock *tp = tcp_sk(sk);
5365	struct sk_buff *data = tp->syn_data ? tcp_write_queue_head(sk) : NULL;
5366	u16 mss = tp->rx_opt.mss_clamp;
5367	bool syn_drop;
5368
5369	if (mss == tp->rx_opt.user_mss) {
5370		struct tcp_options_received opt;
5371
5372		/* Get original SYNACK MSS value if user MSS sets mss_clamp */
5373		tcp_clear_options(&opt);
5374		opt.user_mss = opt.mss_clamp = 0;
5375		tcp_parse_options(synack, &opt, 0, NULL);
5376		mss = opt.mss_clamp;
5377	}
5378
5379	if (!tp->syn_fastopen)  /* Ignore an unsolicited cookie */
 
5380		cookie->len = -1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5381
5382	/* The SYN-ACK neither has cookie nor acknowledges the data. Presumably
5383	 * the remote receives only the retransmitted (regular) SYNs: either
5384	 * the original SYN-data or the corresponding SYN-ACK is lost.
5385	 */
5386	syn_drop = (cookie->len <= 0 && data && tp->total_retrans);
5387
5388	tcp_fastopen_cache_set(sk, mss, cookie, syn_drop);
5389
5390	if (data) { /* Retransmit unacked data in SYN */
5391		tcp_for_write_queue_from(data, sk) {
5392			if (data == tcp_send_head(sk) ||
5393			    __tcp_retransmit_skb(sk, data))
5394				break;
5395		}
5396		tcp_rearm_rto(sk);
5397		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
 
 
5398		return true;
5399	}
5400	tp->syn_data_acked = tp->syn_data;
5401	if (tp->syn_data_acked)
5402		NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
 
 
 
 
 
 
 
5403	return false;
5404}
5405
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5406static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
5407					 const struct tcphdr *th, unsigned int len)
5408{
5409	struct inet_connection_sock *icsk = inet_csk(sk);
5410	struct tcp_sock *tp = tcp_sk(sk);
5411	struct tcp_fastopen_cookie foc = { .len = -1 };
5412	int saved_clamp = tp->rx_opt.mss_clamp;
 
 
5413
5414	tcp_parse_options(skb, &tp->rx_opt, 0, &foc);
5415	if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr)
5416		tp->rx_opt.rcv_tsecr -= tp->tsoffset;
5417
5418	if (th->ack) {
5419		/* rfc793:
5420		 * "If the state is SYN-SENT then
5421		 *    first check the ACK bit
5422		 *      If the ACK bit is set
5423		 *	  If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send
5424		 *        a reset (unless the RST bit is set, if so drop
5425		 *        the segment and return)"
5426		 */
5427		if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
5428		    after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
 
 
 
 
 
 
5429			goto reset_and_undo;
 
5430
5431		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
5432		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,
5433			     tcp_time_stamp)) {
5434			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED);
 
 
5435			goto reset_and_undo;
5436		}
5437
5438		/* Now ACK is acceptable.
5439		 *
5440		 * "If the RST bit is set
5441		 *    If the ACK was acceptable then signal the user "error:
5442		 *    connection reset", drop the segment, enter CLOSED state,
5443		 *    delete TCB, and return."
5444		 */
5445
5446		if (th->rst) {
5447			tcp_reset(sk);
5448			goto discard;
 
 
5449		}
5450
5451		/* rfc793:
5452		 *   "fifth, if neither of the SYN or RST bits is set then
5453		 *    drop the segment and return."
5454		 *
5455		 *    See note below!
5456		 *                                        --ANK(990513)
5457		 */
5458		if (!th->syn)
 
5459			goto discard_and_undo;
5460
5461		/* rfc793:
5462		 *   "If the SYN bit is on ...
5463		 *    are acceptable then ...
5464		 *    (our SYN has been ACKed), change the connection
5465		 *    state to ESTABLISHED..."
5466		 */
5467
5468		TCP_ECN_rcv_synack(tp, th);
5469
5470		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
 
5471		tcp_ack(sk, skb, FLAG_SLOWPATH);
5472
5473		/* Ok.. it's good. Set up sequence numbers and
5474		 * move to established.
5475		 */
5476		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
5477		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5478
5479		/* RFC1323: The window in SYN & SYN/ACK segments is
5480		 * never scaled.
5481		 */
5482		tp->snd_wnd = ntohs(th->window);
5483
5484		if (!tp->rx_opt.wscale_ok) {
5485			tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0;
5486			tp->window_clamp = min(tp->window_clamp, 65535U);
 
5487		}
5488
5489		if (tp->rx_opt.saw_tstamp) {
5490			tp->rx_opt.tstamp_ok	   = 1;
5491			tp->tcp_header_len =
5492				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5493			tp->advmss	    -= TCPOLEN_TSTAMP_ALIGNED;
5494			tcp_store_ts_recent(tp);
5495		} else {
5496			tp->tcp_header_len = sizeof(struct tcphdr);
5497		}
5498
5499		if (tcp_is_sack(tp) && sysctl_tcp_fack)
5500			tcp_enable_fack(tp);
5501
5502		tcp_mtup_init(sk);
5503		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5504		tcp_initialize_rcv_mss(sk);
5505
5506		/* Remember, tcp_poll() does not lock socket!
5507		 * Change state from SYN-SENT only after copied_seq
5508		 * is initialized. */
5509		tp->copied_seq = tp->rcv_nxt;
 
 
5510
5511		smp_mb();
5512
5513		tcp_finish_connect(sk, skb);
5514
5515		if ((tp->syn_fastopen || tp->syn_data) &&
5516		    tcp_rcv_fastopen_synack(sk, skb, &foc))
5517			return -1;
5518
 
 
 
 
 
 
5519		if (sk->sk_write_pending ||
5520		    icsk->icsk_accept_queue.rskq_defer_accept ||
5521		    icsk->icsk_ack.pingpong) {
5522			/* Save one ACK. Data will be ready after
5523			 * several ticks, if write_pending is set.
5524			 *
5525			 * It may be deleted, but with this feature tcpdumps
5526			 * look so _wonderfully_ clever, that I was not able
5527			 * to stand against the temptation 8)     --ANK
5528			 */
5529			inet_csk_schedule_ack(sk);
5530			icsk->icsk_ack.lrcvtime = tcp_time_stamp;
5531			tcp_enter_quickack_mode(sk);
5532			inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
5533						  TCP_DELACK_MAX, TCP_RTO_MAX);
5534
5535discard:
5536			__kfree_skb(skb);
5537			return 0;
5538		} else {
5539			tcp_send_ack(sk);
5540		}
 
5541		return -1;
5542	}
5543
5544	/* No ACK in the segment */
5545
5546	if (th->rst) {
5547		/* rfc793:
5548		 * "If the RST bit is set
5549		 *
5550		 *      Otherwise (no ACK) drop the segment and return."
5551		 */
5552
5553		goto discard_and_undo;
5554	}
5555
5556	/* PAWS check. */
5557	if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp &&
5558	    tcp_paws_reject(&tp->rx_opt, 0))
 
5559		goto discard_and_undo;
5560
5561	if (th->syn) {
5562		/* We see SYN without ACK. It is attempt of
5563		 * simultaneous connect with crossed SYNs.
5564		 * Particularly, it can be connect to self.
5565		 */
 
 
 
 
 
 
 
 
 
 
5566		tcp_set_state(sk, TCP_SYN_RECV);
5567
5568		if (tp->rx_opt.saw_tstamp) {
5569			tp->rx_opt.tstamp_ok = 1;
5570			tcp_store_ts_recent(tp);
5571			tp->tcp_header_len =
5572				sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED;
5573		} else {
5574			tp->tcp_header_len = sizeof(struct tcphdr);
5575		}
5576
5577		tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
 
5578		tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1;
5579
5580		/* RFC1323: The window in SYN & SYN/ACK segments is
5581		 * never scaled.
5582		 */
5583		tp->snd_wnd    = ntohs(th->window);
5584		tp->snd_wl1    = TCP_SKB_CB(skb)->seq;
5585		tp->max_window = tp->snd_wnd;
5586
5587		TCP_ECN_rcv_syn(tp, th);
5588
5589		tcp_mtup_init(sk);
5590		tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
5591		tcp_initialize_rcv_mss(sk);
5592
5593		tcp_send_synack(sk);
5594#if 0
5595		/* Note, we could accept data and URG from this segment.
5596		 * There are no obstacles to make this (except that we must
5597		 * either change tcp_recvmsg() to prevent it from returning data
5598		 * before 3WHS completes per RFC793, or employ TCP Fast Open).
5599		 *
5600		 * However, if we ignore data in ACKless segments sometimes,
5601		 * we have no reasons to accept it sometimes.
5602		 * Also, seems the code doing it in step6 of tcp_rcv_state_process
5603		 * is not flawless. So, discard packet for sanity.
5604		 * Uncomment this return to process the data.
5605		 */
5606		return -1;
5607#else
5608		goto discard;
5609#endif
5610	}
5611	/* "fifth, if neither of the SYN or RST bits is set then
5612	 * drop the segment and return."
5613	 */
5614
5615discard_and_undo:
5616	tcp_clear_options(&tp->rx_opt);
5617	tp->rx_opt.mss_clamp = saved_clamp;
5618	goto discard;
 
5619
5620reset_and_undo:
5621	tcp_clear_options(&tp->rx_opt);
5622	tp->rx_opt.mss_clamp = saved_clamp;
5623	return 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5624}
5625
5626/*
5627 *	This function implements the receiving procedure of RFC 793 for
5628 *	all states except ESTABLISHED and TIME_WAIT.
5629 *	It's called from both tcp_v4_rcv and tcp_v6_rcv and should be
5630 *	address independent.
5631 */
5632
5633int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb,
5634			  const struct tcphdr *th, unsigned int len)
5635{
5636	struct tcp_sock *tp = tcp_sk(sk);
5637	struct inet_connection_sock *icsk = inet_csk(sk);
 
5638	struct request_sock *req;
5639	int queued = 0;
5640	bool acceptable;
5641	u32 synack_stamp;
5642
5643	tp->rx_opt.saw_tstamp = 0;
5644
5645	switch (sk->sk_state) {
5646	case TCP_CLOSE:
 
5647		goto discard;
5648
5649	case TCP_LISTEN:
5650		if (th->ack)
5651			return 1;
5652
5653		if (th->rst)
 
5654			goto discard;
5655
5656		if (th->syn) {
5657			if (th->fin)
 
5658				goto discard;
5659			if (icsk->icsk_af_ops->conn_request(sk, skb) < 0)
5660				return 1;
5661
5662			/* Now we have several options: In theory there is
5663			 * nothing else in the frame. KA9Q has an option to
5664			 * send data with the syn, BSD accepts data with the
5665			 * syn up to the [to be] advertised window and
5666			 * Solaris 2.1 gives you a protocol error. For now
5667			 * we just ignore it, that fits the spec precisely
5668			 * and avoids incompatibilities. It would be nice in
5669			 * future to drop through and process the data.
5670			 *
5671			 * Now that TTCP is starting to be used we ought to
5672			 * queue this data.
5673			 * But, this leaves one open to an easy denial of
5674			 * service attack, and SYN cookies can't defend
5675			 * against this problem. So, we drop the data
5676			 * in the interest of security over speed unless
5677			 * it's still in use.
5678			 */
5679			kfree_skb(skb);
 
 
 
 
 
 
5680			return 0;
5681		}
 
5682		goto discard;
5683
5684	case TCP_SYN_SENT:
5685		queued = tcp_rcv_synsent_state_process(sk, skb, th, len);
 
 
5686		if (queued >= 0)
5687			return queued;
5688
5689		/* Do step6 onward by hand. */
5690		tcp_urg(sk, skb, th);
5691		__kfree_skb(skb);
5692		tcp_data_snd_check(sk);
5693		return 0;
5694	}
5695
5696	req = tp->fastopen_rsk;
5697	if (req != NULL) {
 
 
 
 
 
5698		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
5699		    sk->sk_state != TCP_FIN_WAIT1);
5700
5701		if (tcp_check_req(sk, skb, req, NULL, true) == NULL)
 
5702			goto discard;
 
5703	}
5704
5705	if (!th->ack && !th->rst)
 
5706		goto discard;
5707
5708	if (!tcp_validate_incoming(sk, skb, th, 0))
5709		return 0;
5710
5711	/* step 5: check the ACK field */
5712	acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH |
5713				      FLAG_UPDATE_TS_RECENT) > 0;
5714
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5715	switch (sk->sk_state) {
5716	case TCP_SYN_RECV:
5717		if (!acceptable)
5718			return 1;
 
5719
5720		/* Once we leave TCP_SYN_RECV, we no longer need req
5721		 * so release it.
5722		 */
5723		if (req) {
5724			synack_stamp = tcp_rsk(req)->snt_synack;
5725			tp->total_retrans = req->num_retrans;
5726			reqsk_fastopen_remove(sk, req, false);
5727		} else {
5728			synack_stamp = tp->lsndtime;
5729			/* Make sure socket is routed, for correct metrics. */
5730			icsk->icsk_af_ops->rebuild_header(sk);
5731			tcp_init_congestion_control(sk);
5732
5733			tcp_mtup_init(sk);
5734			tp->copied_seq = tp->rcv_nxt;
5735			tcp_init_buffer_space(sk);
5736		}
 
5737		smp_mb();
5738		tcp_set_state(sk, TCP_ESTABLISHED);
5739		sk->sk_state_change(sk);
5740
5741		/* Note, that this wakeup is only for marginal crossed SYN case.
5742		 * Passively open sockets are not waked up, because
5743		 * sk->sk_sleep == NULL and sk->sk_socket == NULL.
5744		 */
5745		if (sk->sk_socket)
5746			sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT);
5747
5748		tp->snd_una = TCP_SKB_CB(skb)->ack_seq;
5749		tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale;
5750		tcp_init_wl(tp, TCP_SKB_CB(skb)->seq);
5751		tcp_synack_rtt_meas(sk, synack_stamp);
5752
5753		if (tp->rx_opt.tstamp_ok)
5754			tp->advmss -= TCPOLEN_TSTAMP_ALIGNED;
5755
5756		if (req) {
5757			/* Re-arm the timer because data may have been sent out.
5758			 * This is similar to the regular data transmission case
5759			 * when new data has just been ack'ed.
5760			 *
5761			 * (TFO) - we could try to be more aggressive and
5762			 * retransmitting any data sooner based on when they
5763			 * are sent out.
5764			 */
5765			tcp_rearm_rto(sk);
5766		} else
5767			tcp_init_metrics(sk);
5768
5769		tcp_update_pacing_rate(sk);
5770
5771		/* Prevent spurious tcp_cwnd_restart() on first data packet */
5772		tp->lsndtime = tcp_time_stamp;
5773
5774		tcp_initialize_rcv_mss(sk);
5775		tcp_fast_path_on(tp);
 
 
5776		break;
5777
5778	case TCP_FIN_WAIT1: {
5779		struct dst_entry *dst;
5780		int tmo;
5781
5782		/* If we enter the TCP_FIN_WAIT1 state and we are a
5783		 * Fast Open socket and this is the first acceptable
5784		 * ACK we have received, this would have acknowledged
5785		 * our SYNACK so stop the SYNACK timer.
5786		 */
5787		if (req != NULL) {
5788			/* Return RST if ack_seq is invalid.
5789			 * Note that RFC793 only says to generate a
5790			 * DUPACK for it but for TCP Fast Open it seems
5791			 * better to treat this case like TCP_SYN_RECV
5792			 * above.
5793			 */
5794			if (!acceptable)
5795				return 1;
5796			/* We no longer need the request sock. */
5797			reqsk_fastopen_remove(sk, req, false);
5798			tcp_rearm_rto(sk);
5799		}
5800		if (tp->snd_una != tp->write_seq)
5801			break;
5802
5803		tcp_set_state(sk, TCP_FIN_WAIT2);
5804		sk->sk_shutdown |= SEND_SHUTDOWN;
5805
5806		dst = __sk_dst_get(sk);
5807		if (dst)
5808			dst_confirm(dst);
5809
5810		if (!sock_flag(sk, SOCK_DEAD)) {
5811			/* Wake up lingering close() */
5812			sk->sk_state_change(sk);
5813			break;
5814		}
5815
5816		if (tp->linger2 < 0 ||
5817		    (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5818		     after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) {
 
 
 
 
 
 
 
5819			tcp_done(sk);
5820			NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5821			return 1;
5822		}
5823
5824		tmo = tcp_fin_time(sk);
5825		if (tmo > TCP_TIMEWAIT_LEN) {
5826			inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN);
5827		} else if (th->fin || sock_owned_by_user(sk)) {
5828			/* Bad case. We could lose such FIN otherwise.
5829			 * It is not a big problem, but it looks confusing
5830			 * and not so rare event. We still can lose it now,
5831			 * if it spins in bh_lock_sock(), but it is really
5832			 * marginal case.
5833			 */
5834			inet_csk_reset_keepalive_timer(sk, tmo);
5835		} else {
5836			tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
5837			goto discard;
5838		}
5839		break;
5840	}
5841
5842	case TCP_CLOSING:
5843		if (tp->snd_una == tp->write_seq) {
5844			tcp_time_wait(sk, TCP_TIME_WAIT, 0);
5845			goto discard;
5846		}
5847		break;
5848
5849	case TCP_LAST_ACK:
5850		if (tp->snd_una == tp->write_seq) {
5851			tcp_update_metrics(sk);
5852			tcp_done(sk);
5853			goto discard;
5854		}
5855		break;
5856	}
5857
5858	/* step 6: check the URG bit */
5859	tcp_urg(sk, skb, th);
5860
5861	/* step 7: process the segment text */
5862	switch (sk->sk_state) {
5863	case TCP_CLOSE_WAIT:
5864	case TCP_CLOSING:
5865	case TCP_LAST_ACK:
5866		if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt))
 
 
 
 
 
5867			break;
 
 
5868	case TCP_FIN_WAIT1:
5869	case TCP_FIN_WAIT2:
5870		/* RFC 793 says to queue data in these states,
5871		 * RFC 1122 says we MUST send a reset.
5872		 * BSD 4.4 also does reset.
5873		 */
5874		if (sk->sk_shutdown & RCV_SHUTDOWN) {
5875			if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq &&
5876			    after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) {
5877				NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
5878				tcp_reset(sk);
5879				return 1;
5880			}
5881		}
5882		/* Fall through */
5883	case TCP_ESTABLISHED:
5884		tcp_data_queue(sk, skb);
5885		queued = 1;
5886		break;
5887	}
5888
5889	/* tcp_data could move socket to TIME-WAIT */
5890	if (sk->sk_state != TCP_CLOSE) {
5891		tcp_data_snd_check(sk);
5892		tcp_ack_snd_check(sk);
5893	}
5894
5895	if (!queued) {
5896discard:
5897		__kfree_skb(skb);
5898	}
5899	return 0;
 
 
 
 
5900}
5901EXPORT_SYMBOL(tcp_rcv_state_process);