Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * INET		An implementation of the TCP/IP protocol suite for the LINUX
   4 *		operating system.  INET is implemented using the  BSD Socket
   5 *		interface as the means of communication with the user level.
   6 *
   7 *		Implementation of the Transmission Control Protocol(TCP).
   8 *
   9 * Authors:	Ross Biro
  10 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  11 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
  12 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
  13 *		Florian La Roche, <flla@stud.uni-sb.de>
  14 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  15 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
  16 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
  17 *		Matthew Dillon, <dillon@apollo.west.oic.com>
  18 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  19 *		Jorge Cwik, <jorge@laser.satlink.net>
  20 */
  21
  22/*
  23 * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
  24 *				:	Fragmentation on mtu decrease
  25 *				:	Segment collapse on retransmit
  26 *				:	AF independence
  27 *
  28 *		Linus Torvalds	:	send_delayed_ack
  29 *		David S. Miller	:	Charge memory using the right skb
  30 *					during syn/ack processing.
  31 *		David S. Miller :	Output engine completely rewritten.
  32 *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
  33 *		Cacophonix Gaul :	draft-minshall-nagle-01
  34 *		J Hadi Salim	:	ECN support
  35 *
  36 */
  37
  38#define pr_fmt(fmt) "TCP: " fmt
  39
  40#include <net/tcp.h>
  41#include <net/mptcp.h>
 
  42
  43#include <linux/compiler.h>
  44#include <linux/gfp.h>
  45#include <linux/module.h>
  46#include <linux/static_key.h>
 
  47
  48#include <trace/events/tcp.h>
  49
  50/* Refresh clocks of a TCP socket,
  51 * ensuring monotically increasing values.
  52 */
  53void tcp_mstamp_refresh(struct tcp_sock *tp)
  54{
  55	u64 val = tcp_clock_ns();
  56
  57	tp->tcp_clock_cache = val;
  58	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  59}
  60
  61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  62			   int push_one, gfp_t gfp);
  63
  64/* Account for new data that has been sent to the network. */
  65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  66{
  67	struct inet_connection_sock *icsk = inet_csk(sk);
  68	struct tcp_sock *tp = tcp_sk(sk);
  69	unsigned int prior_packets = tp->packets_out;
  70
  71	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
  72
  73	__skb_unlink(skb, &sk->sk_write_queue);
  74	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  75
  76	if (tp->highest_sack == NULL)
  77		tp->highest_sack = skb;
  78
  79	tp->packets_out += tcp_skb_pcount(skb);
  80	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  81		tcp_rearm_rto(sk);
  82
  83	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  84		      tcp_skb_pcount(skb));
  85	tcp_check_space(sk);
  86}
  87
  88/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  89 * window scaling factor due to loss of precision.
  90 * If window has been shrunk, what should we make? It is not clear at all.
  91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  93 * invalid. OK, let's make this for now:
  94 */
  95static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  96{
  97	const struct tcp_sock *tp = tcp_sk(sk);
  98
  99	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
 100	    (tp->rx_opt.wscale_ok &&
 101	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
 102		return tp->snd_nxt;
 103	else
 104		return tcp_wnd_end(tp);
 105}
 106
 107/* Calculate mss to advertise in SYN segment.
 108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 109 *
 110 * 1. It is independent of path mtu.
 111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 113 *    attached devices, because some buggy hosts are confused by
 114 *    large MSS.
 115 * 4. We do not make 3, we advertise MSS, calculated from first
 116 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 117 *    This may be overridden via information stored in routing table.
 118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 119 *    probably even Jumbo".
 120 */
 121static __u16 tcp_advertise_mss(struct sock *sk)
 122{
 123	struct tcp_sock *tp = tcp_sk(sk);
 124	const struct dst_entry *dst = __sk_dst_get(sk);
 125	int mss = tp->advmss;
 126
 127	if (dst) {
 128		unsigned int metric = dst_metric_advmss(dst);
 129
 130		if (metric < mss) {
 131			mss = metric;
 132			tp->advmss = mss;
 133		}
 134	}
 135
 136	return (__u16)mss;
 137}
 138
 139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 140 * This is the first part of cwnd validation mechanism.
 141 */
 142void tcp_cwnd_restart(struct sock *sk, s32 delta)
 143{
 144	struct tcp_sock *tp = tcp_sk(sk);
 145	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 146	u32 cwnd = tcp_snd_cwnd(tp);
 147
 148	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 149
 150	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 151	restart_cwnd = min(restart_cwnd, cwnd);
 152
 153	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 154		cwnd >>= 1;
 155	tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
 156	tp->snd_cwnd_stamp = tcp_jiffies32;
 157	tp->snd_cwnd_used = 0;
 158}
 159
 160/* Congestion state accounting after a packet has been sent. */
 161static void tcp_event_data_sent(struct tcp_sock *tp,
 162				struct sock *sk)
 163{
 164	struct inet_connection_sock *icsk = inet_csk(sk);
 165	const u32 now = tcp_jiffies32;
 166
 167	if (tcp_packets_in_flight(tp) == 0)
 168		tcp_ca_event(sk, CA_EVENT_TX_START);
 169
 170	tp->lsndtime = now;
 171
 172	/* If it is a reply for ato after last received
 173	 * packet, enter pingpong mode.
 174	 */
 175	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 176		inet_csk_enter_pingpong_mode(sk);
 177}
 178
 179/* Account for an ACK we sent. */
 180static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
 181				      u32 rcv_nxt)
 182{
 183	struct tcp_sock *tp = tcp_sk(sk);
 184
 185	if (unlikely(tp->compressed_ack)) {
 186		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 187			      tp->compressed_ack);
 188		tp->compressed_ack = 0;
 189		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 190			__sock_put(sk);
 191	}
 192
 193	if (unlikely(rcv_nxt != tp->rcv_nxt))
 194		return;  /* Special ACK sent by DCTCP to reflect ECN */
 195	tcp_dec_quickack_mode(sk, pkts);
 196	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 197}
 198
 199/* Determine a window scaling and initial window to offer.
 200 * Based on the assumption that the given amount of space
 201 * will be offered. Store the results in the tp structure.
 202 * NOTE: for smooth operation initial space offering should
 203 * be a multiple of mss if possible. We assume here that mss >= 1.
 204 * This MUST be enforced by all callers.
 205 */
 206void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 207			       __u32 *rcv_wnd, __u32 *window_clamp,
 208			       int wscale_ok, __u8 *rcv_wscale,
 209			       __u32 init_rcv_wnd)
 210{
 211	unsigned int space = (__space < 0 ? 0 : __space);
 
 212
 213	/* If no clamp set the clamp to the max possible scaled window */
 214	if (*window_clamp == 0)
 215		(*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
 216	space = min(*window_clamp, space);
 217
 218	/* Quantize space offering to a multiple of mss if possible. */
 219	if (space > mss)
 220		space = rounddown(space, mss);
 221
 222	/* NOTE: offering an initial window larger than 32767
 223	 * will break some buggy TCP stacks. If the admin tells us
 224	 * it is likely we could be speaking with such a buggy stack
 225	 * we will truncate our initial window offering to 32K-1
 226	 * unless the remote has sent us a window scaling option,
 227	 * which we interpret as a sign the remote TCP is not
 228	 * misinterpreting the window field as a signed quantity.
 229	 */
 230	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
 231		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 232	else
 233		(*rcv_wnd) = min_t(u32, space, U16_MAX);
 234
 235	if (init_rcv_wnd)
 236		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 237
 238	*rcv_wscale = 0;
 239	if (wscale_ok) {
 240		/* Set window scaling on max possible window */
 241		space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
 242		space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
 243		space = min_t(u32, space, *window_clamp);
 244		*rcv_wscale = clamp_t(int, ilog2(space) - 15,
 245				      0, TCP_MAX_WSCALE);
 246	}
 247	/* Set the clamp no higher than max representable value */
 248	(*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
 
 249}
 250EXPORT_SYMBOL(tcp_select_initial_window);
 251
 252/* Chose a new window to advertise, update state in tcp_sock for the
 253 * socket, and return result with RFC1323 scaling applied.  The return
 254 * value can be stuffed directly into th->window for an outgoing
 255 * frame.
 256 */
 257static u16 tcp_select_window(struct sock *sk)
 258{
 259	struct tcp_sock *tp = tcp_sk(sk);
 
 260	u32 old_win = tp->rcv_wnd;
 261	u32 cur_win = tcp_receive_window(tp);
 262	u32 new_win = __tcp_select_window(sk);
 
 
 
 
 
 
 
 
 
 263
 264	/* Never shrink the offered window */
 
 265	if (new_win < cur_win) {
 266		/* Danger Will Robinson!
 267		 * Don't update rcv_wup/rcv_wnd here or else
 268		 * we will not be able to advertise a zero
 269		 * window in time.  --DaveM
 270		 *
 271		 * Relax Will Robinson.
 272		 */
 273		if (new_win == 0)
 274			NET_INC_STATS(sock_net(sk),
 275				      LINUX_MIB_TCPWANTZEROWINDOWADV);
 276		new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 
 
 277	}
 
 278	tp->rcv_wnd = new_win;
 279	tp->rcv_wup = tp->rcv_nxt;
 280
 281	/* Make sure we do not exceed the maximum possible
 282	 * scaled window.
 283	 */
 284	if (!tp->rx_opt.rcv_wscale &&
 285	    READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
 286		new_win = min(new_win, MAX_TCP_WINDOW);
 287	else
 288		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 289
 290	/* RFC1323 scaling applied */
 291	new_win >>= tp->rx_opt.rcv_wscale;
 292
 293	/* If we advertise zero window, disable fast path. */
 294	if (new_win == 0) {
 295		tp->pred_flags = 0;
 296		if (old_win)
 297			NET_INC_STATS(sock_net(sk),
 298				      LINUX_MIB_TCPTOZEROWINDOWADV);
 299	} else if (old_win == 0) {
 300		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
 301	}
 302
 303	return new_win;
 304}
 305
 306/* Packet ECN state for a SYN-ACK */
 307static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 308{
 309	const struct tcp_sock *tp = tcp_sk(sk);
 310
 311	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 312	if (!(tp->ecn_flags & TCP_ECN_OK))
 313		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 314	else if (tcp_ca_needs_ecn(sk) ||
 315		 tcp_bpf_ca_needs_ecn(sk))
 316		INET_ECN_xmit(sk);
 317}
 318
 319/* Packet ECN state for a SYN.  */
 320static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 321{
 322	struct tcp_sock *tp = tcp_sk(sk);
 323	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 324	bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
 325		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 326
 327	if (!use_ecn) {
 328		const struct dst_entry *dst = __sk_dst_get(sk);
 329
 330		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 331			use_ecn = true;
 332	}
 333
 334	tp->ecn_flags = 0;
 335
 336	if (use_ecn) {
 337		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 338		tp->ecn_flags = TCP_ECN_OK;
 339		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 340			INET_ECN_xmit(sk);
 341	}
 342}
 343
 344static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 345{
 346	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
 347		/* tp->ecn_flags are cleared at a later point in time when
 348		 * SYN ACK is ultimatively being received.
 349		 */
 350		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 351}
 352
 353static void
 354tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 355{
 356	if (inet_rsk(req)->ecn_ok)
 357		th->ece = 1;
 358}
 359
 360/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 361 * be sent.
 362 */
 363static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 364			 struct tcphdr *th, int tcp_header_len)
 365{
 366	struct tcp_sock *tp = tcp_sk(sk);
 367
 368	if (tp->ecn_flags & TCP_ECN_OK) {
 369		/* Not-retransmitted data segment: set ECT and inject CWR. */
 370		if (skb->len != tcp_header_len &&
 371		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 372			INET_ECN_xmit(sk);
 373			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 374				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 375				th->cwr = 1;
 376				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 377			}
 378		} else if (!tcp_ca_needs_ecn(sk)) {
 379			/* ACK or retransmitted segment: clear ECT|CE */
 380			INET_ECN_dontxmit(sk);
 381		}
 382		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 383			th->ece = 1;
 384	}
 385}
 386
 387/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 388 * auto increment end seqno.
 389 */
 390static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 391{
 392	skb->ip_summed = CHECKSUM_PARTIAL;
 393
 394	TCP_SKB_CB(skb)->tcp_flags = flags;
 395
 396	tcp_skb_pcount_set(skb, 1);
 397
 398	TCP_SKB_CB(skb)->seq = seq;
 399	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 400		seq++;
 401	TCP_SKB_CB(skb)->end_seq = seq;
 402}
 403
 404static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 405{
 406	return tp->snd_una != tp->snd_up;
 407}
 408
 409#define OPTION_SACK_ADVERTISE	BIT(0)
 410#define OPTION_TS		BIT(1)
 411#define OPTION_MD5		BIT(2)
 412#define OPTION_WSCALE		BIT(3)
 413#define OPTION_FAST_OPEN_COOKIE	BIT(8)
 414#define OPTION_SMC		BIT(9)
 415#define OPTION_MPTCP		BIT(10)
 
 416
 417static void smc_options_write(__be32 *ptr, u16 *options)
 418{
 419#if IS_ENABLED(CONFIG_SMC)
 420	if (static_branch_unlikely(&tcp_have_smc)) {
 421		if (unlikely(OPTION_SMC & *options)) {
 422			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 423				       (TCPOPT_NOP  << 16) |
 424				       (TCPOPT_EXP <<  8) |
 425				       (TCPOLEN_EXP_SMC_BASE));
 426			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 427		}
 428	}
 429#endif
 430}
 431
 432struct tcp_out_options {
 433	u16 options;		/* bit field of OPTION_* */
 434	u16 mss;		/* 0 to disable */
 435	u8 ws;			/* window scale, 0 to disable */
 436	u8 num_sack_blocks;	/* number of SACK blocks to include */
 437	u8 hash_size;		/* bytes in hash_location */
 438	u8 bpf_opt_len;		/* length of BPF hdr option */
 439	__u8 *hash_location;	/* temporary pointer, overloaded */
 440	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 441	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 442	struct mptcp_out_options mptcp;
 443};
 444
 445static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
 446				struct tcp_sock *tp,
 447				struct tcp_out_options *opts)
 448{
 449#if IS_ENABLED(CONFIG_MPTCP)
 450	if (unlikely(OPTION_MPTCP & opts->options))
 451		mptcp_write_options(th, ptr, tp, &opts->mptcp);
 452#endif
 453}
 454
 455#ifdef CONFIG_CGROUP_BPF
 456static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
 457					enum tcp_synack_type synack_type)
 458{
 459	if (unlikely(!skb))
 460		return BPF_WRITE_HDR_TCP_CURRENT_MSS;
 461
 462	if (unlikely(synack_type == TCP_SYNACK_COOKIE))
 463		return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
 464
 465	return 0;
 466}
 467
 468/* req, syn_skb and synack_type are used when writing synack */
 469static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 470				  struct request_sock *req,
 471				  struct sk_buff *syn_skb,
 472				  enum tcp_synack_type synack_type,
 473				  struct tcp_out_options *opts,
 474				  unsigned int *remaining)
 475{
 476	struct bpf_sock_ops_kern sock_ops;
 477	int err;
 478
 479	if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
 480					   BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
 481	    !*remaining)
 482		return;
 483
 484	/* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
 485
 486	/* init sock_ops */
 487	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 488
 489	sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
 490
 491	if (req) {
 492		/* The listen "sk" cannot be passed here because
 493		 * it is not locked.  It would not make too much
 494		 * sense to do bpf_setsockopt(listen_sk) based
 495		 * on individual connection request also.
 496		 *
 497		 * Thus, "req" is passed here and the cgroup-bpf-progs
 498		 * of the listen "sk" will be run.
 499		 *
 500		 * "req" is also used here for fastopen even the "sk" here is
 501		 * a fullsock "child" sk.  It is to keep the behavior
 502		 * consistent between fastopen and non-fastopen on
 503		 * the bpf programming side.
 504		 */
 505		sock_ops.sk = (struct sock *)req;
 506		sock_ops.syn_skb = syn_skb;
 507	} else {
 508		sock_owned_by_me(sk);
 509
 510		sock_ops.is_fullsock = 1;
 511		sock_ops.sk = sk;
 512	}
 513
 514	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 515	sock_ops.remaining_opt_len = *remaining;
 516	/* tcp_current_mss() does not pass a skb */
 517	if (skb)
 518		bpf_skops_init_skb(&sock_ops, skb, 0);
 519
 520	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 521
 522	if (err || sock_ops.remaining_opt_len == *remaining)
 523		return;
 524
 525	opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
 526	/* round up to 4 bytes */
 527	opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
 528
 529	*remaining -= opts->bpf_opt_len;
 530}
 531
 532static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 533				    struct request_sock *req,
 534				    struct sk_buff *syn_skb,
 535				    enum tcp_synack_type synack_type,
 536				    struct tcp_out_options *opts)
 537{
 538	u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
 539	struct bpf_sock_ops_kern sock_ops;
 540	int err;
 541
 542	if (likely(!max_opt_len))
 543		return;
 544
 545	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 546
 547	sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
 548
 549	if (req) {
 550		sock_ops.sk = (struct sock *)req;
 551		sock_ops.syn_skb = syn_skb;
 552	} else {
 553		sock_owned_by_me(sk);
 554
 555		sock_ops.is_fullsock = 1;
 556		sock_ops.sk = sk;
 557	}
 558
 559	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 560	sock_ops.remaining_opt_len = max_opt_len;
 561	first_opt_off = tcp_hdrlen(skb) - max_opt_len;
 562	bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
 563
 564	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 565
 566	if (err)
 567		nr_written = 0;
 568	else
 569		nr_written = max_opt_len - sock_ops.remaining_opt_len;
 570
 571	if (nr_written < max_opt_len)
 572		memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
 573		       max_opt_len - nr_written);
 574}
 575#else
 576static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 577				  struct request_sock *req,
 578				  struct sk_buff *syn_skb,
 579				  enum tcp_synack_type synack_type,
 580				  struct tcp_out_options *opts,
 581				  unsigned int *remaining)
 582{
 583}
 584
 585static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 586				    struct request_sock *req,
 587				    struct sk_buff *syn_skb,
 588				    enum tcp_synack_type synack_type,
 589				    struct tcp_out_options *opts)
 590{
 591}
 592#endif
 593
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 594/* Write previously computed TCP options to the packet.
 595 *
 596 * Beware: Something in the Internet is very sensitive to the ordering of
 597 * TCP options, we learned this through the hard way, so be careful here.
 598 * Luckily we can at least blame others for their non-compliance but from
 599 * inter-operability perspective it seems that we're somewhat stuck with
 600 * the ordering which we have been using if we want to keep working with
 601 * those broken things (not that it currently hurts anybody as there isn't
 602 * particular reason why the ordering would need to be changed).
 603 *
 604 * At least SACK_PERM as the first option is known to lead to a disaster
 605 * (but it may well be that other scenarios fail similarly).
 606 */
 607static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
 608			      struct tcp_out_options *opts)
 
 
 609{
 610	__be32 *ptr = (__be32 *)(th + 1);
 611	u16 options = opts->options;	/* mungable copy */
 612
 613	if (unlikely(OPTION_MD5 & options)) {
 614		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 615			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 616		/* overload cookie hash location */
 617		opts->hash_location = (__u8 *)ptr;
 618		ptr += 4;
 
 
 619	}
 620
 621	if (unlikely(opts->mss)) {
 622		*ptr++ = htonl((TCPOPT_MSS << 24) |
 623			       (TCPOLEN_MSS << 16) |
 624			       opts->mss);
 625	}
 626
 627	if (likely(OPTION_TS & options)) {
 628		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 629			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 630				       (TCPOLEN_SACK_PERM << 16) |
 631				       (TCPOPT_TIMESTAMP << 8) |
 632				       TCPOLEN_TIMESTAMP);
 633			options &= ~OPTION_SACK_ADVERTISE;
 634		} else {
 635			*ptr++ = htonl((TCPOPT_NOP << 24) |
 636				       (TCPOPT_NOP << 16) |
 637				       (TCPOPT_TIMESTAMP << 8) |
 638				       TCPOLEN_TIMESTAMP);
 639		}
 640		*ptr++ = htonl(opts->tsval);
 641		*ptr++ = htonl(opts->tsecr);
 642	}
 643
 644	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 645		*ptr++ = htonl((TCPOPT_NOP << 24) |
 646			       (TCPOPT_NOP << 16) |
 647			       (TCPOPT_SACK_PERM << 8) |
 648			       TCPOLEN_SACK_PERM);
 649	}
 650
 651	if (unlikely(OPTION_WSCALE & options)) {
 652		*ptr++ = htonl((TCPOPT_NOP << 24) |
 653			       (TCPOPT_WINDOW << 16) |
 654			       (TCPOLEN_WINDOW << 8) |
 655			       opts->ws);
 656	}
 657
 658	if (unlikely(opts->num_sack_blocks)) {
 659		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 660			tp->duplicate_sack : tp->selective_acks;
 661		int this_sack;
 662
 663		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 664			       (TCPOPT_NOP  << 16) |
 665			       (TCPOPT_SACK <<  8) |
 666			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 667						     TCPOLEN_SACK_PERBLOCK)));
 668
 669		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 670		     ++this_sack) {
 671			*ptr++ = htonl(sp[this_sack].start_seq);
 672			*ptr++ = htonl(sp[this_sack].end_seq);
 673		}
 674
 675		tp->rx_opt.dsack = 0;
 676	}
 677
 678	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 679		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 680		u8 *p = (u8 *)ptr;
 681		u32 len; /* Fast Open option length */
 682
 683		if (foc->exp) {
 684			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 685			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 686				     TCPOPT_FASTOPEN_MAGIC);
 687			p += TCPOLEN_EXP_FASTOPEN_BASE;
 688		} else {
 689			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 690			*p++ = TCPOPT_FASTOPEN;
 691			*p++ = len;
 692		}
 693
 694		memcpy(p, foc->val, foc->len);
 695		if ((len & 3) == 2) {
 696			p[foc->len] = TCPOPT_NOP;
 697			p[foc->len + 1] = TCPOPT_NOP;
 698		}
 699		ptr += (len + 3) >> 2;
 700	}
 701
 702	smc_options_write(ptr, &options);
 703
 704	mptcp_options_write(th, ptr, tp, opts);
 705}
 706
 707static void smc_set_option(const struct tcp_sock *tp,
 708			   struct tcp_out_options *opts,
 709			   unsigned int *remaining)
 710{
 711#if IS_ENABLED(CONFIG_SMC)
 712	if (static_branch_unlikely(&tcp_have_smc)) {
 713		if (tp->syn_smc) {
 714			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 715				opts->options |= OPTION_SMC;
 716				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 717			}
 718		}
 719	}
 720#endif
 721}
 722
 723static void smc_set_option_cond(const struct tcp_sock *tp,
 724				const struct inet_request_sock *ireq,
 725				struct tcp_out_options *opts,
 726				unsigned int *remaining)
 727{
 728#if IS_ENABLED(CONFIG_SMC)
 729	if (static_branch_unlikely(&tcp_have_smc)) {
 730		if (tp->syn_smc && ireq->smc_ok) {
 731			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 732				opts->options |= OPTION_SMC;
 733				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 734			}
 735		}
 736	}
 737#endif
 738}
 739
 740static void mptcp_set_option_cond(const struct request_sock *req,
 741				  struct tcp_out_options *opts,
 742				  unsigned int *remaining)
 743{
 744	if (rsk_is_mptcp(req)) {
 745		unsigned int size;
 746
 747		if (mptcp_synack_options(req, &size, &opts->mptcp)) {
 748			if (*remaining >= size) {
 749				opts->options |= OPTION_MPTCP;
 750				*remaining -= size;
 751			}
 752		}
 753	}
 754}
 755
 756/* Compute TCP options for SYN packets. This is not the final
 757 * network wire format yet.
 758 */
 759static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 760				struct tcp_out_options *opts,
 761				struct tcp_md5sig_key **md5)
 762{
 763	struct tcp_sock *tp = tcp_sk(sk);
 764	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 765	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 
 766
 767	*md5 = NULL;
 768#ifdef CONFIG_TCP_MD5SIG
 769	if (static_branch_unlikely(&tcp_md5_needed.key) &&
 770	    rcu_access_pointer(tp->md5sig_info)) {
 771		*md5 = tp->af_specific->md5_lookup(sk, sk);
 772		if (*md5) {
 773			opts->options |= OPTION_MD5;
 774			remaining -= TCPOLEN_MD5SIG_ALIGNED;
 
 
 775		}
 776	}
 777#endif
 778
 779	/* We always get an MSS option.  The option bytes which will be seen in
 780	 * normal data packets should timestamps be used, must be in the MSS
 781	 * advertised.  But we subtract them from tp->mss_cache so that
 782	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 783	 * fact here if necessary.  If we don't do this correctly, as a
 784	 * receiver we won't recognize data packets as being full sized when we
 785	 * should, and thus we won't abide by the delayed ACK rules correctly.
 786	 * SACKs don't matter, we never delay an ACK when we have any of those
 787	 * going out.  */
 788	opts->mss = tcp_advertise_mss(sk);
 789	remaining -= TCPOLEN_MSS_ALIGNED;
 790
 791	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
 792		opts->options |= OPTION_TS;
 793		opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
 794		opts->tsecr = tp->rx_opt.ts_recent;
 795		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 796	}
 797	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
 798		opts->ws = tp->rx_opt.rcv_wscale;
 799		opts->options |= OPTION_WSCALE;
 800		remaining -= TCPOLEN_WSCALE_ALIGNED;
 801	}
 802	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
 803		opts->options |= OPTION_SACK_ADVERTISE;
 804		if (unlikely(!(OPTION_TS & opts->options)))
 805			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 806	}
 807
 808	if (fastopen && fastopen->cookie.len >= 0) {
 809		u32 need = fastopen->cookie.len;
 810
 811		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 812					       TCPOLEN_FASTOPEN_BASE;
 813		need = (need + 3) & ~3U;  /* Align to 32 bits */
 814		if (remaining >= need) {
 815			opts->options |= OPTION_FAST_OPEN_COOKIE;
 816			opts->fastopen_cookie = &fastopen->cookie;
 817			remaining -= need;
 818			tp->syn_fastopen = 1;
 819			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 820		}
 821	}
 822
 823	smc_set_option(tp, opts, &remaining);
 824
 825	if (sk_is_mptcp(sk)) {
 826		unsigned int size;
 827
 828		if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
 829			opts->options |= OPTION_MPTCP;
 830			remaining -= size;
 
 
 831		}
 832	}
 833
 834	bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
 835
 836	return MAX_TCP_OPTION_SPACE - remaining;
 837}
 838
 839/* Set up TCP options for SYN-ACKs. */
 840static unsigned int tcp_synack_options(const struct sock *sk,
 841				       struct request_sock *req,
 842				       unsigned int mss, struct sk_buff *skb,
 843				       struct tcp_out_options *opts,
 844				       const struct tcp_md5sig_key *md5,
 845				       struct tcp_fastopen_cookie *foc,
 846				       enum tcp_synack_type synack_type,
 847				       struct sk_buff *syn_skb)
 848{
 849	struct inet_request_sock *ireq = inet_rsk(req);
 850	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 851
 852#ifdef CONFIG_TCP_MD5SIG
 853	if (md5) {
 854		opts->options |= OPTION_MD5;
 855		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 856
 857		/* We can't fit any SACK blocks in a packet with MD5 + TS
 858		 * options. There was discussion about disabling SACK
 859		 * rather than TS in order to fit in better with old,
 860		 * buggy kernels, but that was deemed to be unnecessary.
 861		 */
 862		if (synack_type != TCP_SYNACK_COOKIE)
 863			ireq->tstamp_ok &= !ireq->sack_ok;
 
 
 
 
 864	}
 865#endif
 866
 867	/* We always send an MSS option. */
 868	opts->mss = mss;
 869	remaining -= TCPOLEN_MSS_ALIGNED;
 870
 871	if (likely(ireq->wscale_ok)) {
 872		opts->ws = ireq->rcv_wscale;
 873		opts->options |= OPTION_WSCALE;
 874		remaining -= TCPOLEN_WSCALE_ALIGNED;
 875	}
 876	if (likely(ireq->tstamp_ok)) {
 877		opts->options |= OPTION_TS;
 878		opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
 879		opts->tsecr = req->ts_recent;
 
 880		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 881	}
 882	if (likely(ireq->sack_ok)) {
 883		opts->options |= OPTION_SACK_ADVERTISE;
 884		if (unlikely(!ireq->tstamp_ok))
 885			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 886	}
 887	if (foc != NULL && foc->len >= 0) {
 888		u32 need = foc->len;
 889
 890		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 891				   TCPOLEN_FASTOPEN_BASE;
 892		need = (need + 3) & ~3U;  /* Align to 32 bits */
 893		if (remaining >= need) {
 894			opts->options |= OPTION_FAST_OPEN_COOKIE;
 895			opts->fastopen_cookie = foc;
 896			remaining -= need;
 897		}
 898	}
 899
 900	mptcp_set_option_cond(req, opts, &remaining);
 901
 902	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 903
 904	bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
 905			      synack_type, opts, &remaining);
 906
 907	return MAX_TCP_OPTION_SPACE - remaining;
 908}
 909
 910/* Compute TCP options for ESTABLISHED sockets. This is not the
 911 * final wire format yet.
 912 */
 913static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 914					struct tcp_out_options *opts,
 915					struct tcp_md5sig_key **md5)
 916{
 917	struct tcp_sock *tp = tcp_sk(sk);
 918	unsigned int size = 0;
 919	unsigned int eff_sacks;
 920
 921	opts->options = 0;
 922
 923	*md5 = NULL;
 924#ifdef CONFIG_TCP_MD5SIG
 925	if (static_branch_unlikely(&tcp_md5_needed.key) &&
 926	    rcu_access_pointer(tp->md5sig_info)) {
 927		*md5 = tp->af_specific->md5_lookup(sk, sk);
 928		if (*md5) {
 929			opts->options |= OPTION_MD5;
 930			size += TCPOLEN_MD5SIG_ALIGNED;
 931		}
 932	}
 933#endif
 934
 935	if (likely(tp->rx_opt.tstamp_ok)) {
 936		opts->options |= OPTION_TS;
 937		opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
 
 938		opts->tsecr = tp->rx_opt.ts_recent;
 939		size += TCPOLEN_TSTAMP_ALIGNED;
 940	}
 941
 942	/* MPTCP options have precedence over SACK for the limited TCP
 943	 * option space because a MPTCP connection would be forced to
 944	 * fall back to regular TCP if a required multipath option is
 945	 * missing. SACK still gets a chance to use whatever space is
 946	 * left.
 947	 */
 948	if (sk_is_mptcp(sk)) {
 949		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 950		unsigned int opt_size = 0;
 951
 952		if (mptcp_established_options(sk, skb, &opt_size, remaining,
 953					      &opts->mptcp)) {
 954			opts->options |= OPTION_MPTCP;
 955			size += opt_size;
 956		}
 957	}
 958
 959	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
 960	if (unlikely(eff_sacks)) {
 961		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 962		if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
 963					 TCPOLEN_SACK_PERBLOCK))
 964			return size;
 965
 966		opts->num_sack_blocks =
 967			min_t(unsigned int, eff_sacks,
 968			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
 969			      TCPOLEN_SACK_PERBLOCK);
 970
 971		size += TCPOLEN_SACK_BASE_ALIGNED +
 972			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
 973	}
 974
 975	if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
 976					    BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
 977		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
 978
 979		bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
 980
 981		size = MAX_TCP_OPTION_SPACE - remaining;
 982	}
 983
 984	return size;
 985}
 986
 987
 988/* TCP SMALL QUEUES (TSQ)
 989 *
 990 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
 991 * to reduce RTT and bufferbloat.
 992 * We do this using a special skb destructor (tcp_wfree).
 993 *
 994 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
 995 * needs to be reallocated in a driver.
 996 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
 997 *
 998 * Since transmit from skb destructor is forbidden, we use a tasklet
 999 * to process all sockets that eventually need to send more skbs.
1000 * We use one tasklet per cpu, with its own queue of sockets.
1001 */
1002struct tsq_tasklet {
1003	struct tasklet_struct	tasklet;
1004	struct list_head	head; /* queue of tcp sockets */
1005};
1006static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1007
1008static void tcp_tsq_write(struct sock *sk)
1009{
1010	if ((1 << sk->sk_state) &
1011	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1012	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
1013		struct tcp_sock *tp = tcp_sk(sk);
1014
1015		if (tp->lost_out > tp->retrans_out &&
1016		    tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1017			tcp_mstamp_refresh(tp);
1018			tcp_xmit_retransmit_queue(sk);
1019		}
1020
1021		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1022			       0, GFP_ATOMIC);
1023	}
1024}
1025
1026static void tcp_tsq_handler(struct sock *sk)
1027{
1028	bh_lock_sock(sk);
1029	if (!sock_owned_by_user(sk))
1030		tcp_tsq_write(sk);
1031	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1032		sock_hold(sk);
1033	bh_unlock_sock(sk);
1034}
1035/*
1036 * One tasklet per cpu tries to send more skbs.
1037 * We run in tasklet context but need to disable irqs when
1038 * transferring tsq->head because tcp_wfree() might
1039 * interrupt us (non NAPI drivers)
1040 */
1041static void tcp_tasklet_func(struct tasklet_struct *t)
1042{
1043	struct tsq_tasklet *tsq = from_tasklet(tsq,  t, tasklet);
1044	LIST_HEAD(list);
1045	unsigned long flags;
1046	struct list_head *q, *n;
1047	struct tcp_sock *tp;
1048	struct sock *sk;
1049
1050	local_irq_save(flags);
1051	list_splice_init(&tsq->head, &list);
1052	local_irq_restore(flags);
1053
1054	list_for_each_safe(q, n, &list) {
1055		tp = list_entry(q, struct tcp_sock, tsq_node);
1056		list_del(&tp->tsq_node);
1057
1058		sk = (struct sock *)tp;
1059		smp_mb__before_atomic();
1060		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1061
1062		tcp_tsq_handler(sk);
1063		sk_free(sk);
1064	}
1065}
1066
1067#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
1068			  TCPF_WRITE_TIMER_DEFERRED |	\
1069			  TCPF_DELACK_TIMER_DEFERRED |	\
1070			  TCPF_MTU_REDUCED_DEFERRED)
 
1071/**
1072 * tcp_release_cb - tcp release_sock() callback
1073 * @sk: socket
1074 *
1075 * called from release_sock() to perform protocol dependent
1076 * actions before socket release.
1077 */
1078void tcp_release_cb(struct sock *sk)
1079{
1080	unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1081	unsigned long nflags;
1082
1083	/* perform an atomic operation only if at least one flag is set */
1084	do {
1085		if (!(flags & TCP_DEFERRED_ALL))
1086			return;
1087		nflags = flags & ~TCP_DEFERRED_ALL;
1088	} while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1089
1090	if (flags & TCPF_TSQ_DEFERRED) {
1091		tcp_tsq_write(sk);
1092		__sock_put(sk);
1093	}
1094	/* Here begins the tricky part :
1095	 * We are called from release_sock() with :
1096	 * 1) BH disabled
1097	 * 2) sk_lock.slock spinlock held
1098	 * 3) socket owned by us (sk->sk_lock.owned == 1)
1099	 *
1100	 * But following code is meant to be called from BH handlers,
1101	 * so we should keep BH disabled, but early release socket ownership
1102	 */
1103	sock_release_ownership(sk);
1104
1105	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1106		tcp_write_timer_handler(sk);
1107		__sock_put(sk);
1108	}
1109	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1110		tcp_delack_timer_handler(sk);
1111		__sock_put(sk);
1112	}
1113	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1114		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1115		__sock_put(sk);
1116	}
 
 
1117}
1118EXPORT_SYMBOL(tcp_release_cb);
1119
1120void __init tcp_tasklet_init(void)
1121{
1122	int i;
1123
1124	for_each_possible_cpu(i) {
1125		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1126
1127		INIT_LIST_HEAD(&tsq->head);
1128		tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1129	}
1130}
1131
1132/*
1133 * Write buffer destructor automatically called from kfree_skb.
1134 * We can't xmit new skbs from this context, as we might already
1135 * hold qdisc lock.
1136 */
1137void tcp_wfree(struct sk_buff *skb)
1138{
1139	struct sock *sk = skb->sk;
1140	struct tcp_sock *tp = tcp_sk(sk);
1141	unsigned long flags, nval, oval;
1142	struct tsq_tasklet *tsq;
1143	bool empty;
1144
1145	/* Keep one reference on sk_wmem_alloc.
1146	 * Will be released by sk_free() from here or tcp_tasklet_func()
1147	 */
1148	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1149
1150	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
1151	 * Wait until our queues (qdisc + devices) are drained.
1152	 * This gives :
1153	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1154	 * - chance for incoming ACK (processed by another cpu maybe)
1155	 *   to migrate this flow (skb->ooo_okay will be eventually set)
1156	 */
1157	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1158		goto out;
1159
1160	oval = smp_load_acquire(&sk->sk_tsq_flags);
1161	do {
1162		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1163			goto out;
1164
1165		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1166	} while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1167
1168	/* queue this socket to tasklet queue */
1169	local_irq_save(flags);
1170	tsq = this_cpu_ptr(&tsq_tasklet);
1171	empty = list_empty(&tsq->head);
1172	list_add(&tp->tsq_node, &tsq->head);
1173	if (empty)
1174		tasklet_schedule(&tsq->tasklet);
1175	local_irq_restore(flags);
1176	return;
1177out:
1178	sk_free(sk);
1179}
1180
1181/* Note: Called under soft irq.
1182 * We can call TCP stack right away, unless socket is owned by user.
1183 */
1184enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1185{
1186	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1187	struct sock *sk = (struct sock *)tp;
1188
1189	tcp_tsq_handler(sk);
1190	sock_put(sk);
1191
1192	return HRTIMER_NORESTART;
1193}
1194
1195static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1196				      u64 prior_wstamp)
1197{
1198	struct tcp_sock *tp = tcp_sk(sk);
1199
1200	if (sk->sk_pacing_status != SK_PACING_NONE) {
1201		unsigned long rate = sk->sk_pacing_rate;
1202
1203		/* Original sch_fq does not pace first 10 MSS
1204		 * Note that tp->data_segs_out overflows after 2^32 packets,
1205		 * this is a minor annoyance.
1206		 */
1207		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1208			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1209			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1210
1211			/* take into account OS jitter */
1212			len_ns -= min_t(u64, len_ns / 2, credit);
1213			tp->tcp_wstamp_ns += len_ns;
1214		}
1215	}
1216	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1217}
1218
1219INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1220INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1221INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1222
1223/* This routine actually transmits TCP packets queued in by
1224 * tcp_do_sendmsg().  This is used by both the initial
1225 * transmission and possible later retransmissions.
1226 * All SKB's seen here are completely headerless.  It is our
1227 * job to build the TCP header, and pass the packet down to
1228 * IP so it can do the same plus pass the packet off to the
1229 * device.
1230 *
1231 * We are working here with either a clone of the original
1232 * SKB, or a fresh unique copy made by the retransmit engine.
1233 */
1234static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1235			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1236{
1237	const struct inet_connection_sock *icsk = inet_csk(sk);
1238	struct inet_sock *inet;
1239	struct tcp_sock *tp;
1240	struct tcp_skb_cb *tcb;
1241	struct tcp_out_options opts;
1242	unsigned int tcp_options_size, tcp_header_size;
1243	struct sk_buff *oskb = NULL;
1244	struct tcp_md5sig_key *md5;
1245	struct tcphdr *th;
1246	u64 prior_wstamp;
1247	int err;
1248
1249	BUG_ON(!skb || !tcp_skb_pcount(skb));
1250	tp = tcp_sk(sk);
1251	prior_wstamp = tp->tcp_wstamp_ns;
1252	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1253	skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1254	if (clone_it) {
1255		oskb = skb;
1256
1257		tcp_skb_tsorted_save(oskb) {
1258			if (unlikely(skb_cloned(oskb)))
1259				skb = pskb_copy(oskb, gfp_mask);
1260			else
1261				skb = skb_clone(oskb, gfp_mask);
1262		} tcp_skb_tsorted_restore(oskb);
1263
1264		if (unlikely(!skb))
1265			return -ENOBUFS;
1266		/* retransmit skbs might have a non zero value in skb->dev
1267		 * because skb->dev is aliased with skb->rbnode.rb_left
1268		 */
1269		skb->dev = NULL;
1270	}
1271
1272	inet = inet_sk(sk);
1273	tcb = TCP_SKB_CB(skb);
1274	memset(&opts, 0, sizeof(opts));
1275
 
1276	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1277		tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1278	} else {
1279		tcp_options_size = tcp_established_options(sk, skb, &opts,
1280							   &md5);
1281		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1282		 * at receiver : This slightly improve GRO performance.
1283		 * Note that we do not force the PSH flag for non GSO packets,
1284		 * because they might be sent under high congestion events,
1285		 * and in this case it is better to delay the delivery of 1-MSS
1286		 * packets and thus the corresponding ACK packet that would
1287		 * release the following packet.
1288		 */
1289		if (tcp_skb_pcount(skb) > 1)
1290			tcb->tcp_flags |= TCPHDR_PSH;
1291	}
1292	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1293
1294	/* if no packet is in qdisc/device queue, then allow XPS to select
1295	 * another queue. We can be called from tcp_tsq_handler()
1296	 * which holds one reference to sk.
1297	 *
1298	 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1299	 * One way to get this would be to set skb->truesize = 2 on them.
 
 
 
 
 
 
1300	 */
1301	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
 
1302
1303	/* If we had to use memory reserve to allocate this skb,
1304	 * this might cause drops if packet is looped back :
1305	 * Other socket might not have SOCK_MEMALLOC.
1306	 * Packets not looped back do not care about pfmemalloc.
1307	 */
1308	skb->pfmemalloc = 0;
1309
1310	skb_push(skb, tcp_header_size);
1311	skb_reset_transport_header(skb);
1312
1313	skb_orphan(skb);
1314	skb->sk = sk;
1315	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1316	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1317
1318	skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1319
1320	/* Build TCP header and checksum it. */
1321	th = (struct tcphdr *)skb->data;
1322	th->source		= inet->inet_sport;
1323	th->dest		= inet->inet_dport;
1324	th->seq			= htonl(tcb->seq);
1325	th->ack_seq		= htonl(rcv_nxt);
1326	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1327					tcb->tcp_flags);
1328
1329	th->check		= 0;
1330	th->urg_ptr		= 0;
1331
1332	/* The urg_mode check is necessary during a below snd_una win probe */
1333	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1334		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1335			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1336			th->urg = 1;
1337		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1338			th->urg_ptr = htons(0xFFFF);
1339			th->urg = 1;
1340		}
1341	}
1342
1343	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1344	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1345		th->window      = htons(tcp_select_window(sk));
1346		tcp_ecn_send(sk, skb, th, tcp_header_size);
1347	} else {
1348		/* RFC1323: The window in SYN & SYN/ACK segments
1349		 * is never scaled.
1350		 */
1351		th->window	= htons(min(tp->rcv_wnd, 65535U));
1352	}
1353
1354	tcp_options_write(th, tp, &opts);
1355
 
1356#ifdef CONFIG_TCP_MD5SIG
1357	/* Calculate the MD5 hash, as we have all we need now */
1358	if (md5) {
1359		sk_gso_disable(sk);
1360		tp->af_specific->calc_md5_hash(opts.hash_location,
1361					       md5, sk, skb);
1362	}
1363#endif
 
 
 
 
 
 
 
 
 
 
1364
1365	/* BPF prog is the last one writing header option */
1366	bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1367
1368	INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1369			   tcp_v6_send_check, tcp_v4_send_check,
1370			   sk, skb);
1371
1372	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1373		tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1374
1375	if (skb->len != tcp_header_size) {
1376		tcp_event_data_sent(tp, sk);
1377		tp->data_segs_out += tcp_skb_pcount(skb);
1378		tp->bytes_sent += skb->len - tcp_header_size;
1379	}
1380
1381	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1382		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1383			      tcp_skb_pcount(skb));
1384
1385	tp->segs_out += tcp_skb_pcount(skb);
1386	skb_set_hash_from_sk(skb, sk);
1387	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1388	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1389	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1390
1391	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1392
1393	/* Cleanup our debris for IP stacks */
1394	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1395			       sizeof(struct inet6_skb_parm)));
1396
1397	tcp_add_tx_delay(skb, tp);
1398
1399	err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1400				 inet6_csk_xmit, ip_queue_xmit,
1401				 sk, skb, &inet->cork.fl);
1402
1403	if (unlikely(err > 0)) {
1404		tcp_enter_cwr(sk);
1405		err = net_xmit_eval(err);
1406	}
1407	if (!err && oskb) {
1408		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1409		tcp_rate_skb_sent(sk, oskb);
1410	}
1411	return err;
1412}
1413
1414static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1415			    gfp_t gfp_mask)
1416{
1417	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1418				  tcp_sk(sk)->rcv_nxt);
1419}
1420
1421/* This routine just queues the buffer for sending.
1422 *
1423 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1424 * otherwise socket can stall.
1425 */
1426static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1427{
1428	struct tcp_sock *tp = tcp_sk(sk);
1429
1430	/* Advance write_seq and place onto the write_queue. */
1431	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1432	__skb_header_release(skb);
1433	tcp_add_write_queue_tail(sk, skb);
1434	sk_wmem_queued_add(sk, skb->truesize);
1435	sk_mem_charge(sk, skb->truesize);
1436}
1437
1438/* Initialize TSO segments for a packet. */
1439static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1440{
 
 
1441	if (skb->len <= mss_now) {
1442		/* Avoid the costly divide in the normal
1443		 * non-TSO case.
1444		 */
1445		tcp_skb_pcount_set(skb, 1);
1446		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1447	} else {
1448		tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1449		TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1450	}
 
 
 
 
1451}
1452
1453/* Pcount in the middle of the write queue got changed, we need to do various
1454 * tweaks to fix counters
1455 */
1456static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1457{
1458	struct tcp_sock *tp = tcp_sk(sk);
1459
1460	tp->packets_out -= decr;
1461
1462	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1463		tp->sacked_out -= decr;
1464	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1465		tp->retrans_out -= decr;
1466	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1467		tp->lost_out -= decr;
1468
1469	/* Reno case is special. Sigh... */
1470	if (tcp_is_reno(tp) && decr > 0)
1471		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1472
1473	if (tp->lost_skb_hint &&
1474	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1475	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1476		tp->lost_cnt_hint -= decr;
1477
1478	tcp_verify_left_out(tp);
1479}
1480
1481static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1482{
1483	return TCP_SKB_CB(skb)->txstamp_ack ||
1484		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1485}
1486
1487static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1488{
1489	struct skb_shared_info *shinfo = skb_shinfo(skb);
1490
1491	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1492	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1493		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1494		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1495
1496		shinfo->tx_flags &= ~tsflags;
1497		shinfo2->tx_flags |= tsflags;
1498		swap(shinfo->tskey, shinfo2->tskey);
1499		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1500		TCP_SKB_CB(skb)->txstamp_ack = 0;
1501	}
1502}
1503
1504static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1505{
1506	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1507	TCP_SKB_CB(skb)->eor = 0;
1508}
1509
1510/* Insert buff after skb on the write or rtx queue of sk.  */
1511static void tcp_insert_write_queue_after(struct sk_buff *skb,
1512					 struct sk_buff *buff,
1513					 struct sock *sk,
1514					 enum tcp_queue tcp_queue)
1515{
1516	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1517		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1518	else
1519		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1520}
1521
1522/* Function to create two new TCP segments.  Shrinks the given segment
1523 * to the specified size and appends a new segment with the rest of the
1524 * packet to the list.  This won't be called frequently, I hope.
1525 * Remember, these are still headerless SKBs at this point.
1526 */
1527int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1528		 struct sk_buff *skb, u32 len,
1529		 unsigned int mss_now, gfp_t gfp)
1530{
1531	struct tcp_sock *tp = tcp_sk(sk);
1532	struct sk_buff *buff;
1533	int nsize, old_factor;
1534	long limit;
1535	int nlen;
1536	u8 flags;
1537
1538	if (WARN_ON(len > skb->len))
1539		return -EINVAL;
1540
1541	nsize = skb_headlen(skb) - len;
1542	if (nsize < 0)
1543		nsize = 0;
1544
1545	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1546	 * We need some allowance to not penalize applications setting small
1547	 * SO_SNDBUF values.
1548	 * Also allow first and last skb in retransmit queue to be split.
1549	 */
1550	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1551	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1552		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1553		     skb != tcp_rtx_queue_head(sk) &&
1554		     skb != tcp_rtx_queue_tail(sk))) {
1555		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1556		return -ENOMEM;
1557	}
1558
1559	if (skb_unclone_keeptruesize(skb, gfp))
1560		return -ENOMEM;
1561
1562	/* Get a new skb... force flag on. */
1563	buff = tcp_stream_alloc_skb(sk, nsize, gfp, true);
1564	if (!buff)
1565		return -ENOMEM; /* We'll just try again later. */
1566	skb_copy_decrypted(buff, skb);
1567	mptcp_skb_ext_copy(buff, skb);
1568
1569	sk_wmem_queued_add(sk, buff->truesize);
1570	sk_mem_charge(sk, buff->truesize);
1571	nlen = skb->len - len - nsize;
1572	buff->truesize += nlen;
1573	skb->truesize -= nlen;
1574
1575	/* Correct the sequence numbers. */
1576	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1577	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1578	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1579
1580	/* PSH and FIN should only be set in the second packet. */
1581	flags = TCP_SKB_CB(skb)->tcp_flags;
1582	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1583	TCP_SKB_CB(buff)->tcp_flags = flags;
1584	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1585	tcp_skb_fragment_eor(skb, buff);
1586
1587	skb_split(skb, buff, len);
1588
1589	skb_set_delivery_time(buff, skb->tstamp, true);
1590	tcp_fragment_tstamp(skb, buff);
1591
1592	old_factor = tcp_skb_pcount(skb);
1593
1594	/* Fix up tso_factor for both original and new SKB.  */
1595	tcp_set_skb_tso_segs(skb, mss_now);
1596	tcp_set_skb_tso_segs(buff, mss_now);
1597
1598	/* Update delivered info for the new segment */
1599	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1600
1601	/* If this packet has been sent out already, we must
1602	 * adjust the various packet counters.
1603	 */
1604	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1605		int diff = old_factor - tcp_skb_pcount(skb) -
1606			tcp_skb_pcount(buff);
1607
1608		if (diff)
1609			tcp_adjust_pcount(sk, skb, diff);
1610	}
1611
1612	/* Link BUFF into the send queue. */
1613	__skb_header_release(buff);
1614	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1615	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1616		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1617
1618	return 0;
1619}
1620
1621/* This is similar to __pskb_pull_tail(). The difference is that pulled
1622 * data is not copied, but immediately discarded.
1623 */
1624static int __pskb_trim_head(struct sk_buff *skb, int len)
1625{
1626	struct skb_shared_info *shinfo;
1627	int i, k, eat;
1628
1629	eat = min_t(int, len, skb_headlen(skb));
1630	if (eat) {
1631		__skb_pull(skb, eat);
1632		len -= eat;
1633		if (!len)
1634			return 0;
1635	}
1636	eat = len;
1637	k = 0;
1638	shinfo = skb_shinfo(skb);
1639	for (i = 0; i < shinfo->nr_frags; i++) {
1640		int size = skb_frag_size(&shinfo->frags[i]);
1641
1642		if (size <= eat) {
1643			skb_frag_unref(skb, i);
1644			eat -= size;
1645		} else {
1646			shinfo->frags[k] = shinfo->frags[i];
1647			if (eat) {
1648				skb_frag_off_add(&shinfo->frags[k], eat);
1649				skb_frag_size_sub(&shinfo->frags[k], eat);
1650				eat = 0;
1651			}
1652			k++;
1653		}
1654	}
1655	shinfo->nr_frags = k;
1656
1657	skb->data_len -= len;
1658	skb->len = skb->data_len;
1659	return len;
1660}
1661
1662/* Remove acked data from a packet in the transmit queue. */
1663int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1664{
1665	u32 delta_truesize;
1666
1667	if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1668		return -ENOMEM;
1669
1670	delta_truesize = __pskb_trim_head(skb, len);
1671
1672	TCP_SKB_CB(skb)->seq += len;
1673
1674	if (delta_truesize) {
1675		skb->truesize	   -= delta_truesize;
1676		sk_wmem_queued_add(sk, -delta_truesize);
1677		if (!skb_zcopy_pure(skb))
1678			sk_mem_uncharge(sk, delta_truesize);
1679	}
1680
1681	/* Any change of skb->len requires recalculation of tso factor. */
1682	if (tcp_skb_pcount(skb) > 1)
1683		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1684
1685	return 0;
1686}
1687
1688/* Calculate MSS not accounting any TCP options.  */
1689static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1690{
1691	const struct tcp_sock *tp = tcp_sk(sk);
1692	const struct inet_connection_sock *icsk = inet_csk(sk);
1693	int mss_now;
1694
1695	/* Calculate base mss without TCP options:
1696	   It is MMS_S - sizeof(tcphdr) of rfc1122
1697	 */
1698	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1699
1700	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1701	if (icsk->icsk_af_ops->net_frag_header_len) {
1702		const struct dst_entry *dst = __sk_dst_get(sk);
1703
1704		if (dst && dst_allfrag(dst))
1705			mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1706	}
1707
1708	/* Clamp it (mss_clamp does not include tcp options) */
1709	if (mss_now > tp->rx_opt.mss_clamp)
1710		mss_now = tp->rx_opt.mss_clamp;
1711
1712	/* Now subtract optional transport overhead */
1713	mss_now -= icsk->icsk_ext_hdr_len;
1714
1715	/* Then reserve room for full set of TCP options and 8 bytes of data */
1716	mss_now = max(mss_now,
1717		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1718	return mss_now;
1719}
1720
1721/* Calculate MSS. Not accounting for SACKs here.  */
1722int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1723{
1724	/* Subtract TCP options size, not including SACKs */
1725	return __tcp_mtu_to_mss(sk, pmtu) -
1726	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1727}
1728EXPORT_SYMBOL(tcp_mtu_to_mss);
1729
1730/* Inverse of above */
1731int tcp_mss_to_mtu(struct sock *sk, int mss)
1732{
1733	const struct tcp_sock *tp = tcp_sk(sk);
1734	const struct inet_connection_sock *icsk = inet_csk(sk);
1735	int mtu;
1736
1737	mtu = mss +
1738	      tp->tcp_header_len +
1739	      icsk->icsk_ext_hdr_len +
1740	      icsk->icsk_af_ops->net_header_len;
1741
1742	/* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1743	if (icsk->icsk_af_ops->net_frag_header_len) {
1744		const struct dst_entry *dst = __sk_dst_get(sk);
1745
1746		if (dst && dst_allfrag(dst))
1747			mtu += icsk->icsk_af_ops->net_frag_header_len;
1748	}
1749	return mtu;
1750}
1751EXPORT_SYMBOL(tcp_mss_to_mtu);
1752
1753/* MTU probing init per socket */
1754void tcp_mtup_init(struct sock *sk)
1755{
1756	struct tcp_sock *tp = tcp_sk(sk);
1757	struct inet_connection_sock *icsk = inet_csk(sk);
1758	struct net *net = sock_net(sk);
1759
1760	icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1761	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1762			       icsk->icsk_af_ops->net_header_len;
1763	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1764	icsk->icsk_mtup.probe_size = 0;
1765	if (icsk->icsk_mtup.enabled)
1766		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1767}
1768EXPORT_SYMBOL(tcp_mtup_init);
1769
1770/* This function synchronize snd mss to current pmtu/exthdr set.
1771
1772   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1773   for TCP options, but includes only bare TCP header.
1774
1775   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1776   It is minimum of user_mss and mss received with SYN.
1777   It also does not include TCP options.
1778
1779   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1780
1781   tp->mss_cache is current effective sending mss, including
1782   all tcp options except for SACKs. It is evaluated,
1783   taking into account current pmtu, but never exceeds
1784   tp->rx_opt.mss_clamp.
1785
1786   NOTE1. rfc1122 clearly states that advertised MSS
1787   DOES NOT include either tcp or ip options.
1788
1789   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1790   are READ ONLY outside this function.		--ANK (980731)
1791 */
1792unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1793{
1794	struct tcp_sock *tp = tcp_sk(sk);
1795	struct inet_connection_sock *icsk = inet_csk(sk);
1796	int mss_now;
1797
1798	if (icsk->icsk_mtup.search_high > pmtu)
1799		icsk->icsk_mtup.search_high = pmtu;
1800
1801	mss_now = tcp_mtu_to_mss(sk, pmtu);
1802	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1803
1804	/* And store cached results */
1805	icsk->icsk_pmtu_cookie = pmtu;
1806	if (icsk->icsk_mtup.enabled)
1807		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1808	tp->mss_cache = mss_now;
1809
1810	return mss_now;
1811}
1812EXPORT_SYMBOL(tcp_sync_mss);
1813
1814/* Compute the current effective MSS, taking SACKs and IP options,
1815 * and even PMTU discovery events into account.
1816 */
1817unsigned int tcp_current_mss(struct sock *sk)
1818{
1819	const struct tcp_sock *tp = tcp_sk(sk);
1820	const struct dst_entry *dst = __sk_dst_get(sk);
1821	u32 mss_now;
1822	unsigned int header_len;
1823	struct tcp_out_options opts;
1824	struct tcp_md5sig_key *md5;
1825
1826	mss_now = tp->mss_cache;
1827
1828	if (dst) {
1829		u32 mtu = dst_mtu(dst);
1830		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1831			mss_now = tcp_sync_mss(sk, mtu);
1832	}
1833
1834	header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1835		     sizeof(struct tcphdr);
1836	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1837	 * some common options. If this is an odd packet (because we have SACK
1838	 * blocks etc) then our calculated header_len will be different, and
1839	 * we have to adjust mss_now correspondingly */
1840	if (header_len != tp->tcp_header_len) {
1841		int delta = (int) header_len - tp->tcp_header_len;
1842		mss_now -= delta;
1843	}
1844
1845	return mss_now;
1846}
1847
1848/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1849 * As additional protections, we do not touch cwnd in retransmission phases,
1850 * and if application hit its sndbuf limit recently.
1851 */
1852static void tcp_cwnd_application_limited(struct sock *sk)
1853{
1854	struct tcp_sock *tp = tcp_sk(sk);
1855
1856	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1857	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1858		/* Limited by application or receiver window. */
1859		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1860		u32 win_used = max(tp->snd_cwnd_used, init_win);
1861		if (win_used < tcp_snd_cwnd(tp)) {
1862			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1863			tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1864		}
1865		tp->snd_cwnd_used = 0;
1866	}
1867	tp->snd_cwnd_stamp = tcp_jiffies32;
1868}
1869
1870static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1871{
1872	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1873	struct tcp_sock *tp = tcp_sk(sk);
1874
1875	/* Track the strongest available signal of the degree to which the cwnd
1876	 * is fully utilized. If cwnd-limited then remember that fact for the
1877	 * current window. If not cwnd-limited then track the maximum number of
1878	 * outstanding packets in the current window. (If cwnd-limited then we
1879	 * chose to not update tp->max_packets_out to avoid an extra else
1880	 * clause with no functional impact.)
1881	 */
1882	if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1883	    is_cwnd_limited ||
1884	    (!tp->is_cwnd_limited &&
1885	     tp->packets_out > tp->max_packets_out)) {
1886		tp->is_cwnd_limited = is_cwnd_limited;
1887		tp->max_packets_out = tp->packets_out;
1888		tp->cwnd_usage_seq = tp->snd_nxt;
1889	}
1890
1891	if (tcp_is_cwnd_limited(sk)) {
1892		/* Network is feed fully. */
1893		tp->snd_cwnd_used = 0;
1894		tp->snd_cwnd_stamp = tcp_jiffies32;
1895	} else {
1896		/* Network starves. */
1897		if (tp->packets_out > tp->snd_cwnd_used)
1898			tp->snd_cwnd_used = tp->packets_out;
1899
1900		if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1901		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1902		    !ca_ops->cong_control)
1903			tcp_cwnd_application_limited(sk);
1904
1905		/* The following conditions together indicate the starvation
1906		 * is caused by insufficient sender buffer:
1907		 * 1) just sent some data (see tcp_write_xmit)
1908		 * 2) not cwnd limited (this else condition)
1909		 * 3) no more data to send (tcp_write_queue_empty())
1910		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1911		 */
1912		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1913		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1914		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1915			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1916	}
1917}
1918
1919/* Minshall's variant of the Nagle send check. */
1920static bool tcp_minshall_check(const struct tcp_sock *tp)
1921{
1922	return after(tp->snd_sml, tp->snd_una) &&
1923		!after(tp->snd_sml, tp->snd_nxt);
1924}
1925
1926/* Update snd_sml if this skb is under mss
1927 * Note that a TSO packet might end with a sub-mss segment
1928 * The test is really :
1929 * if ((skb->len % mss) != 0)
1930 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1931 * But we can avoid doing the divide again given we already have
1932 *  skb_pcount = skb->len / mss_now
1933 */
1934static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1935				const struct sk_buff *skb)
1936{
1937	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1938		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1939}
1940
1941/* Return false, if packet can be sent now without violation Nagle's rules:
1942 * 1. It is full sized. (provided by caller in %partial bool)
1943 * 2. Or it contains FIN. (already checked by caller)
1944 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1945 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1946 *    With Minshall's modification: all sent small packets are ACKed.
1947 */
1948static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1949			    int nonagle)
1950{
1951	return partial &&
1952		((nonagle & TCP_NAGLE_CORK) ||
1953		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1954}
1955
1956/* Return how many segs we'd like on a TSO packet,
1957 * depending on current pacing rate, and how close the peer is.
1958 *
1959 * Rationale is:
1960 * - For close peers, we rather send bigger packets to reduce
1961 *   cpu costs, because occasional losses will be repaired fast.
1962 * - For long distance/rtt flows, we would like to get ACK clocking
1963 *   with 1 ACK per ms.
1964 *
1965 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1966 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1967 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1968 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1969 */
1970static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1971			    int min_tso_segs)
1972{
1973	unsigned long bytes;
1974	u32 r;
1975
1976	bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1977
1978	r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1979	if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1980		bytes += sk->sk_gso_max_size >> r;
1981
1982	bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1983
1984	return max_t(u32, bytes / mss_now, min_tso_segs);
1985}
1986
1987/* Return the number of segments we want in the skb we are transmitting.
1988 * See if congestion control module wants to decide; otherwise, autosize.
1989 */
1990static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1991{
1992	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1993	u32 min_tso, tso_segs;
1994
1995	min_tso = ca_ops->min_tso_segs ?
1996			ca_ops->min_tso_segs(sk) :
1997			READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1998
1999	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2000	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2001}
2002
2003/* Returns the portion of skb which can be sent right away */
2004static unsigned int tcp_mss_split_point(const struct sock *sk,
2005					const struct sk_buff *skb,
2006					unsigned int mss_now,
2007					unsigned int max_segs,
2008					int nonagle)
2009{
2010	const struct tcp_sock *tp = tcp_sk(sk);
2011	u32 partial, needed, window, max_len;
2012
2013	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2014	max_len = mss_now * max_segs;
2015
2016	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2017		return max_len;
2018
2019	needed = min(skb->len, window);
2020
2021	if (max_len <= needed)
2022		return max_len;
2023
2024	partial = needed % mss_now;
2025	/* If last segment is not a full MSS, check if Nagle rules allow us
2026	 * to include this last segment in this skb.
2027	 * Otherwise, we'll split the skb at last MSS boundary
2028	 */
2029	if (tcp_nagle_check(partial != 0, tp, nonagle))
2030		return needed - partial;
2031
2032	return needed;
2033}
2034
2035/* Can at least one segment of SKB be sent right now, according to the
2036 * congestion window rules?  If so, return how many segments are allowed.
2037 */
2038static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2039					 const struct sk_buff *skb)
2040{
2041	u32 in_flight, cwnd, halfcwnd;
2042
2043	/* Don't be strict about the congestion window for the final FIN.  */
2044	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2045	    tcp_skb_pcount(skb) == 1)
2046		return 1;
2047
2048	in_flight = tcp_packets_in_flight(tp);
2049	cwnd = tcp_snd_cwnd(tp);
2050	if (in_flight >= cwnd)
2051		return 0;
2052
2053	/* For better scheduling, ensure we have at least
2054	 * 2 GSO packets in flight.
2055	 */
2056	halfcwnd = max(cwnd >> 1, 1U);
2057	return min(halfcwnd, cwnd - in_flight);
2058}
2059
2060/* Initialize TSO state of a skb.
2061 * This must be invoked the first time we consider transmitting
2062 * SKB onto the wire.
2063 */
2064static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2065{
2066	int tso_segs = tcp_skb_pcount(skb);
2067
2068	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2069		tcp_set_skb_tso_segs(skb, mss_now);
2070		tso_segs = tcp_skb_pcount(skb);
2071	}
2072	return tso_segs;
2073}
2074
2075
2076/* Return true if the Nagle test allows this packet to be
2077 * sent now.
2078 */
2079static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2080				  unsigned int cur_mss, int nonagle)
2081{
2082	/* Nagle rule does not apply to frames, which sit in the middle of the
2083	 * write_queue (they have no chances to get new data).
2084	 *
2085	 * This is implemented in the callers, where they modify the 'nonagle'
2086	 * argument based upon the location of SKB in the send queue.
2087	 */
2088	if (nonagle & TCP_NAGLE_PUSH)
2089		return true;
2090
2091	/* Don't use the nagle rule for urgent data (or for the final FIN). */
2092	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2093		return true;
2094
2095	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2096		return true;
2097
2098	return false;
2099}
2100
2101/* Does at least the first segment of SKB fit into the send window? */
2102static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2103			     const struct sk_buff *skb,
2104			     unsigned int cur_mss)
2105{
2106	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2107
2108	if (skb->len > cur_mss)
2109		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2110
2111	return !after(end_seq, tcp_wnd_end(tp));
2112}
2113
2114/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2115 * which is put after SKB on the list.  It is very much like
2116 * tcp_fragment() except that it may make several kinds of assumptions
2117 * in order to speed up the splitting operation.  In particular, we
2118 * know that all the data is in scatter-gather pages, and that the
2119 * packet has never been sent out before (and thus is not cloned).
2120 */
2121static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2122			unsigned int mss_now, gfp_t gfp)
2123{
2124	int nlen = skb->len - len;
2125	struct sk_buff *buff;
2126	u8 flags;
2127
2128	/* All of a TSO frame must be composed of paged data.  */
2129	if (skb->len != skb->data_len)
2130		return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2131				    skb, len, mss_now, gfp);
2132
2133	buff = tcp_stream_alloc_skb(sk, 0, gfp, true);
2134	if (unlikely(!buff))
2135		return -ENOMEM;
2136	skb_copy_decrypted(buff, skb);
2137	mptcp_skb_ext_copy(buff, skb);
2138
2139	sk_wmem_queued_add(sk, buff->truesize);
2140	sk_mem_charge(sk, buff->truesize);
2141	buff->truesize += nlen;
2142	skb->truesize -= nlen;
2143
2144	/* Correct the sequence numbers. */
2145	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2146	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2147	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2148
2149	/* PSH and FIN should only be set in the second packet. */
2150	flags = TCP_SKB_CB(skb)->tcp_flags;
2151	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2152	TCP_SKB_CB(buff)->tcp_flags = flags;
2153
2154	tcp_skb_fragment_eor(skb, buff);
2155
2156	skb_split(skb, buff, len);
2157	tcp_fragment_tstamp(skb, buff);
2158
2159	/* Fix up tso_factor for both original and new SKB.  */
2160	tcp_set_skb_tso_segs(skb, mss_now);
2161	tcp_set_skb_tso_segs(buff, mss_now);
2162
2163	/* Link BUFF into the send queue. */
2164	__skb_header_release(buff);
2165	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2166
2167	return 0;
2168}
2169
2170/* Try to defer sending, if possible, in order to minimize the amount
2171 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
2172 *
2173 * This algorithm is from John Heffner.
2174 */
2175static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2176				 bool *is_cwnd_limited,
2177				 bool *is_rwnd_limited,
2178				 u32 max_segs)
2179{
2180	const struct inet_connection_sock *icsk = inet_csk(sk);
2181	u32 send_win, cong_win, limit, in_flight;
2182	struct tcp_sock *tp = tcp_sk(sk);
2183	struct sk_buff *head;
2184	int win_divisor;
2185	s64 delta;
2186
2187	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2188		goto send_now;
2189
2190	/* Avoid bursty behavior by allowing defer
2191	 * only if the last write was recent (1 ms).
2192	 * Note that tp->tcp_wstamp_ns can be in the future if we have
2193	 * packets waiting in a qdisc or device for EDT delivery.
2194	 */
2195	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2196	if (delta > 0)
2197		goto send_now;
2198
2199	in_flight = tcp_packets_in_flight(tp);
2200
2201	BUG_ON(tcp_skb_pcount(skb) <= 1);
2202	BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2203
2204	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2205
2206	/* From in_flight test above, we know that cwnd > in_flight.  */
2207	cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2208
2209	limit = min(send_win, cong_win);
2210
2211	/* If a full-sized TSO skb can be sent, do it. */
2212	if (limit >= max_segs * tp->mss_cache)
2213		goto send_now;
2214
2215	/* Middle in queue won't get any more data, full sendable already? */
2216	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2217		goto send_now;
2218
2219	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2220	if (win_divisor) {
2221		u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2222
2223		/* If at least some fraction of a window is available,
2224		 * just use it.
2225		 */
2226		chunk /= win_divisor;
2227		if (limit >= chunk)
2228			goto send_now;
2229	} else {
2230		/* Different approach, try not to defer past a single
2231		 * ACK.  Receiver should ACK every other full sized
2232		 * frame, so if we have space for more than 3 frames
2233		 * then send now.
2234		 */
2235		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2236			goto send_now;
2237	}
2238
2239	/* TODO : use tsorted_sent_queue ? */
2240	head = tcp_rtx_queue_head(sk);
2241	if (!head)
2242		goto send_now;
2243	delta = tp->tcp_clock_cache - head->tstamp;
2244	/* If next ACK is likely to come too late (half srtt), do not defer */
2245	if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2246		goto send_now;
2247
2248	/* Ok, it looks like it is advisable to defer.
2249	 * Three cases are tracked :
2250	 * 1) We are cwnd-limited
2251	 * 2) We are rwnd-limited
2252	 * 3) We are application limited.
2253	 */
2254	if (cong_win < send_win) {
2255		if (cong_win <= skb->len) {
2256			*is_cwnd_limited = true;
2257			return true;
2258		}
2259	} else {
2260		if (send_win <= skb->len) {
2261			*is_rwnd_limited = true;
2262			return true;
2263		}
2264	}
2265
2266	/* If this packet won't get more data, do not wait. */
2267	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2268	    TCP_SKB_CB(skb)->eor)
2269		goto send_now;
2270
2271	return true;
2272
2273send_now:
2274	return false;
2275}
2276
2277static inline void tcp_mtu_check_reprobe(struct sock *sk)
2278{
2279	struct inet_connection_sock *icsk = inet_csk(sk);
2280	struct tcp_sock *tp = tcp_sk(sk);
2281	struct net *net = sock_net(sk);
2282	u32 interval;
2283	s32 delta;
2284
2285	interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2286	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2287	if (unlikely(delta >= interval * HZ)) {
2288		int mss = tcp_current_mss(sk);
2289
2290		/* Update current search range */
2291		icsk->icsk_mtup.probe_size = 0;
2292		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2293			sizeof(struct tcphdr) +
2294			icsk->icsk_af_ops->net_header_len;
2295		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2296
2297		/* Update probe time stamp */
2298		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2299	}
2300}
2301
2302static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2303{
2304	struct sk_buff *skb, *next;
2305
2306	skb = tcp_send_head(sk);
2307	tcp_for_write_queue_from_safe(skb, next, sk) {
2308		if (len <= skb->len)
2309			break;
2310
2311		if (unlikely(TCP_SKB_CB(skb)->eor) ||
2312		    tcp_has_tx_tstamp(skb) ||
2313		    !skb_pure_zcopy_same(skb, next))
2314			return false;
2315
2316		len -= skb->len;
2317	}
2318
2319	return true;
2320}
2321
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2322/* Create a new MTU probe if we are ready.
2323 * MTU probe is regularly attempting to increase the path MTU by
2324 * deliberately sending larger packets.  This discovers routing
2325 * changes resulting in larger path MTUs.
2326 *
2327 * Returns 0 if we should wait to probe (no cwnd available),
2328 *         1 if a probe was sent,
2329 *         -1 otherwise
2330 */
2331static int tcp_mtu_probe(struct sock *sk)
2332{
2333	struct inet_connection_sock *icsk = inet_csk(sk);
2334	struct tcp_sock *tp = tcp_sk(sk);
2335	struct sk_buff *skb, *nskb, *next;
2336	struct net *net = sock_net(sk);
2337	int probe_size;
2338	int size_needed;
2339	int copy, len;
2340	int mss_now;
2341	int interval;
2342
2343	/* Not currently probing/verifying,
2344	 * not in recovery,
2345	 * have enough cwnd, and
2346	 * not SACKing (the variable headers throw things off)
2347	 */
2348	if (likely(!icsk->icsk_mtup.enabled ||
2349		   icsk->icsk_mtup.probe_size ||
2350		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2351		   tcp_snd_cwnd(tp) < 11 ||
2352		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2353		return -1;
2354
2355	/* Use binary search for probe_size between tcp_mss_base,
2356	 * and current mss_clamp. if (search_high - search_low)
2357	 * smaller than a threshold, backoff from probing.
2358	 */
2359	mss_now = tcp_current_mss(sk);
2360	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2361				    icsk->icsk_mtup.search_low) >> 1);
2362	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2363	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2364	/* When misfortune happens, we are reprobing actively,
2365	 * and then reprobe timer has expired. We stick with current
2366	 * probing process by not resetting search range to its orignal.
2367	 */
2368	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2369	    interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2370		/* Check whether enough time has elaplased for
2371		 * another round of probing.
2372		 */
2373		tcp_mtu_check_reprobe(sk);
2374		return -1;
2375	}
2376
2377	/* Have enough data in the send queue to probe? */
2378	if (tp->write_seq - tp->snd_nxt < size_needed)
2379		return -1;
2380
2381	if (tp->snd_wnd < size_needed)
2382		return -1;
2383	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2384		return 0;
2385
2386	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2387	if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2388		if (!tcp_packets_in_flight(tp))
2389			return -1;
2390		else
2391			return 0;
2392	}
2393
2394	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2395		return -1;
2396
2397	/* We're allowed to probe.  Build it now. */
2398	nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2399	if (!nskb)
2400		return -1;
 
 
 
 
 
 
 
2401	sk_wmem_queued_add(sk, nskb->truesize);
2402	sk_mem_charge(sk, nskb->truesize);
2403
2404	skb = tcp_send_head(sk);
2405	skb_copy_decrypted(nskb, skb);
2406	mptcp_skb_ext_copy(nskb, skb);
2407
2408	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2409	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2410	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2411
2412	tcp_insert_write_queue_before(nskb, skb, sk);
2413	tcp_highest_sack_replace(sk, skb, nskb);
2414
2415	len = 0;
2416	tcp_for_write_queue_from_safe(skb, next, sk) {
2417		copy = min_t(int, skb->len, probe_size - len);
2418		skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2419
2420		if (skb->len <= copy) {
2421			/* We've eaten all the data from this skb.
2422			 * Throw it away. */
2423			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2424			/* If this is the last SKB we copy and eor is set
2425			 * we need to propagate it to the new skb.
2426			 */
2427			TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2428			tcp_skb_collapse_tstamp(nskb, skb);
2429			tcp_unlink_write_queue(skb, sk);
2430			tcp_wmem_free_skb(sk, skb);
2431		} else {
2432			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2433						   ~(TCPHDR_FIN|TCPHDR_PSH);
2434			if (!skb_shinfo(skb)->nr_frags) {
2435				skb_pull(skb, copy);
2436			} else {
2437				__pskb_trim_head(skb, copy);
2438				tcp_set_skb_tso_segs(skb, mss_now);
2439			}
2440			TCP_SKB_CB(skb)->seq += copy;
2441		}
2442
2443		len += copy;
2444
2445		if (len >= probe_size)
2446			break;
2447	}
2448	tcp_init_tso_segs(nskb, nskb->len);
2449
2450	/* We're ready to send.  If this fails, the probe will
2451	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2452	 */
2453	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2454		/* Decrement cwnd here because we are sending
2455		 * effectively two packets. */
2456		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2457		tcp_event_new_data_sent(sk, nskb);
2458
2459		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2460		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2461		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2462
2463		return 1;
2464	}
2465
2466	return -1;
2467}
2468
2469static bool tcp_pacing_check(struct sock *sk)
2470{
2471	struct tcp_sock *tp = tcp_sk(sk);
2472
2473	if (!tcp_needs_internal_pacing(sk))
2474		return false;
2475
2476	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2477		return false;
2478
2479	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2480		hrtimer_start(&tp->pacing_timer,
2481			      ns_to_ktime(tp->tcp_wstamp_ns),
2482			      HRTIMER_MODE_ABS_PINNED_SOFT);
2483		sock_hold(sk);
2484	}
2485	return true;
2486}
2487
 
 
 
 
 
 
 
 
 
 
 
 
2488/* TCP Small Queues :
2489 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2490 * (These limits are doubled for retransmits)
2491 * This allows for :
2492 *  - better RTT estimation and ACK scheduling
2493 *  - faster recovery
2494 *  - high rates
2495 * Alas, some drivers / subsystems require a fair amount
2496 * of queued bytes to ensure line rate.
2497 * One example is wifi aggregation (802.11 AMPDU)
2498 */
2499static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2500				  unsigned int factor)
2501{
2502	unsigned long limit;
2503
2504	limit = max_t(unsigned long,
2505		      2 * skb->truesize,
2506		      sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2507	if (sk->sk_pacing_status == SK_PACING_NONE)
2508		limit = min_t(unsigned long, limit,
2509			      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2510	limit <<= factor;
2511
2512	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2513	    tcp_sk(sk)->tcp_tx_delay) {
2514		u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
 
2515
2516		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2517		 * approximate our needs assuming an ~100% skb->truesize overhead.
2518		 * USEC_PER_SEC is approximated by 2^20.
2519		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2520		 */
2521		extra_bytes >>= (20 - 1);
2522		limit += extra_bytes;
2523	}
2524	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2525		/* Always send skb if rtx queue is empty.
2526		 * No need to wait for TX completion to call us back,
2527		 * after softirq/tasklet schedule.
2528		 * This helps when TX completions are delayed too much.
2529		 */
2530		if (tcp_rtx_queue_empty(sk))
2531			return false;
2532
2533		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2534		/* It is possible TX completion already happened
2535		 * before we set TSQ_THROTTLED, so we must
2536		 * test again the condition.
2537		 */
2538		smp_mb__after_atomic();
2539		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2540			return true;
2541	}
2542	return false;
2543}
2544
2545static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2546{
2547	const u32 now = tcp_jiffies32;
2548	enum tcp_chrono old = tp->chrono_type;
2549
2550	if (old > TCP_CHRONO_UNSPEC)
2551		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2552	tp->chrono_start = now;
2553	tp->chrono_type = new;
2554}
2555
2556void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2557{
2558	struct tcp_sock *tp = tcp_sk(sk);
2559
2560	/* If there are multiple conditions worthy of tracking in a
2561	 * chronograph then the highest priority enum takes precedence
2562	 * over the other conditions. So that if something "more interesting"
2563	 * starts happening, stop the previous chrono and start a new one.
2564	 */
2565	if (type > tp->chrono_type)
2566		tcp_chrono_set(tp, type);
2567}
2568
2569void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2570{
2571	struct tcp_sock *tp = tcp_sk(sk);
2572
2573
2574	/* There are multiple conditions worthy of tracking in a
2575	 * chronograph, so that the highest priority enum takes
2576	 * precedence over the other conditions (see tcp_chrono_start).
2577	 * If a condition stops, we only stop chrono tracking if
2578	 * it's the "most interesting" or current chrono we are
2579	 * tracking and starts busy chrono if we have pending data.
2580	 */
2581	if (tcp_rtx_and_write_queues_empty(sk))
2582		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2583	else if (type == tp->chrono_type)
2584		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2585}
2586
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2587/* This routine writes packets to the network.  It advances the
2588 * send_head.  This happens as incoming acks open up the remote
2589 * window for us.
2590 *
2591 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2592 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2593 * account rare use of URG, this is not a big flaw.
2594 *
2595 * Send at most one packet when push_one > 0. Temporarily ignore
2596 * cwnd limit to force at most one packet out when push_one == 2.
2597
2598 * Returns true, if no segments are in flight and we have queued segments,
2599 * but cannot send anything now because of SWS or another problem.
2600 */
2601static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2602			   int push_one, gfp_t gfp)
2603{
2604	struct tcp_sock *tp = tcp_sk(sk);
2605	struct sk_buff *skb;
2606	unsigned int tso_segs, sent_pkts;
2607	int cwnd_quota;
2608	int result;
2609	bool is_cwnd_limited = false, is_rwnd_limited = false;
2610	u32 max_segs;
2611
2612	sent_pkts = 0;
2613
2614	tcp_mstamp_refresh(tp);
2615	if (!push_one) {
2616		/* Do MTU probing. */
2617		result = tcp_mtu_probe(sk);
2618		if (!result) {
2619			return false;
2620		} else if (result > 0) {
2621			sent_pkts = 1;
2622		}
2623	}
2624
2625	max_segs = tcp_tso_segs(sk, mss_now);
2626	while ((skb = tcp_send_head(sk))) {
2627		unsigned int limit;
 
2628
2629		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2630			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2631			tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2632			skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2633			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2634			tcp_init_tso_segs(skb, mss_now);
2635			goto repair; /* Skip network transmission */
2636		}
2637
2638		if (tcp_pacing_check(sk))
2639			break;
2640
2641		tso_segs = tcp_init_tso_segs(skb, mss_now);
2642		BUG_ON(!tso_segs);
2643
2644		cwnd_quota = tcp_cwnd_test(tp, skb);
2645		if (!cwnd_quota) {
2646			if (push_one == 2)
2647				/* Force out a loss probe pkt. */
2648				cwnd_quota = 1;
2649			else
2650				break;
2651		}
 
 
 
 
 
 
2652
2653		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2654			is_rwnd_limited = true;
2655			break;
2656		}
2657
2658		if (tso_segs == 1) {
2659			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2660						     (tcp_skb_is_last(sk, skb) ?
2661						      nonagle : TCP_NAGLE_PUSH))))
2662				break;
2663		} else {
2664			if (!push_one &&
2665			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2666						 &is_rwnd_limited, max_segs))
2667				break;
2668		}
2669
2670		limit = mss_now;
2671		if (tso_segs > 1 && !tcp_urg_mode(tp))
2672			limit = tcp_mss_split_point(sk, skb, mss_now,
2673						    min_t(unsigned int,
2674							  cwnd_quota,
2675							  max_segs),
2676						    nonagle);
2677
2678		if (skb->len > limit &&
2679		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2680			break;
2681
2682		if (tcp_small_queue_check(sk, skb, 0))
2683			break;
2684
2685		/* Argh, we hit an empty skb(), presumably a thread
2686		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2687		 * We do not want to send a pure-ack packet and have
2688		 * a strange looking rtx queue with empty packet(s).
2689		 */
2690		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2691			break;
2692
2693		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2694			break;
2695
2696repair:
2697		/* Advance the send_head.  This one is sent out.
2698		 * This call will increment packets_out.
2699		 */
2700		tcp_event_new_data_sent(sk, skb);
2701
2702		tcp_minshall_update(tp, mss_now, skb);
2703		sent_pkts += tcp_skb_pcount(skb);
2704
2705		if (push_one)
2706			break;
2707	}
2708
2709	if (is_rwnd_limited)
2710		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2711	else
2712		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2713
2714	is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2715	if (likely(sent_pkts || is_cwnd_limited))
2716		tcp_cwnd_validate(sk, is_cwnd_limited);
2717
2718	if (likely(sent_pkts)) {
2719		if (tcp_in_cwnd_reduction(sk))
2720			tp->prr_out += sent_pkts;
2721
2722		/* Send one loss probe per tail loss episode. */
2723		if (push_one != 2)
2724			tcp_schedule_loss_probe(sk, false);
2725		return false;
2726	}
2727	return !tp->packets_out && !tcp_write_queue_empty(sk);
2728}
2729
2730bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2731{
2732	struct inet_connection_sock *icsk = inet_csk(sk);
2733	struct tcp_sock *tp = tcp_sk(sk);
2734	u32 timeout, rto_delta_us;
2735	int early_retrans;
2736
2737	/* Don't do any loss probe on a Fast Open connection before 3WHS
2738	 * finishes.
2739	 */
2740	if (rcu_access_pointer(tp->fastopen_rsk))
2741		return false;
2742
2743	early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2744	/* Schedule a loss probe in 2*RTT for SACK capable connections
2745	 * not in loss recovery, that are either limited by cwnd or application.
2746	 */
2747	if ((early_retrans != 3 && early_retrans != 4) ||
2748	    !tp->packets_out || !tcp_is_sack(tp) ||
2749	    (icsk->icsk_ca_state != TCP_CA_Open &&
2750	     icsk->icsk_ca_state != TCP_CA_CWR))
2751		return false;
2752
2753	/* Probe timeout is 2*rtt. Add minimum RTO to account
2754	 * for delayed ack when there's one outstanding packet. If no RTT
2755	 * sample is available then probe after TCP_TIMEOUT_INIT.
2756	 */
2757	if (tp->srtt_us) {
2758		timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2759		if (tp->packets_out == 1)
2760			timeout += TCP_RTO_MIN;
2761		else
2762			timeout += TCP_TIMEOUT_MIN;
 
2763	} else {
2764		timeout = TCP_TIMEOUT_INIT;
2765	}
2766
2767	/* If the RTO formula yields an earlier time, then use that time. */
2768	rto_delta_us = advancing_rto ?
2769			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2770			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2771	if (rto_delta_us > 0)
2772		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2773
2774	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2775	return true;
2776}
2777
2778/* Thanks to skb fast clones, we can detect if a prior transmit of
2779 * a packet is still in a qdisc or driver queue.
2780 * In this case, there is very little point doing a retransmit !
2781 */
2782static bool skb_still_in_host_queue(struct sock *sk,
2783				    const struct sk_buff *skb)
2784{
2785	if (unlikely(skb_fclone_busy(sk, skb))) {
2786		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2787		smp_mb__after_atomic();
2788		if (skb_fclone_busy(sk, skb)) {
2789			NET_INC_STATS(sock_net(sk),
2790				      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2791			return true;
2792		}
2793	}
2794	return false;
2795}
2796
2797/* When probe timeout (PTO) fires, try send a new segment if possible, else
2798 * retransmit the last segment.
2799 */
2800void tcp_send_loss_probe(struct sock *sk)
2801{
2802	struct tcp_sock *tp = tcp_sk(sk);
2803	struct sk_buff *skb;
2804	int pcount;
2805	int mss = tcp_current_mss(sk);
2806
2807	/* At most one outstanding TLP */
2808	if (tp->tlp_high_seq)
2809		goto rearm_timer;
2810
2811	tp->tlp_retrans = 0;
2812	skb = tcp_send_head(sk);
2813	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2814		pcount = tp->packets_out;
2815		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2816		if (tp->packets_out > pcount)
2817			goto probe_sent;
2818		goto rearm_timer;
2819	}
2820	skb = skb_rb_last(&sk->tcp_rtx_queue);
2821	if (unlikely(!skb)) {
2822		WARN_ONCE(tp->packets_out,
2823			  "invalid inflight: %u state %u cwnd %u mss %d\n",
2824			  tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2825		inet_csk(sk)->icsk_pending = 0;
2826		return;
2827	}
2828
2829	if (skb_still_in_host_queue(sk, skb))
2830		goto rearm_timer;
2831
2832	pcount = tcp_skb_pcount(skb);
2833	if (WARN_ON(!pcount))
2834		goto rearm_timer;
2835
2836	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2837		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2838					  (pcount - 1) * mss, mss,
2839					  GFP_ATOMIC)))
2840			goto rearm_timer;
2841		skb = skb_rb_next(skb);
2842	}
2843
2844	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2845		goto rearm_timer;
2846
2847	if (__tcp_retransmit_skb(sk, skb, 1))
2848		goto rearm_timer;
2849
2850	tp->tlp_retrans = 1;
2851
2852probe_sent:
2853	/* Record snd_nxt for loss detection. */
2854	tp->tlp_high_seq = tp->snd_nxt;
2855
2856	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2857	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2858	inet_csk(sk)->icsk_pending = 0;
2859rearm_timer:
2860	tcp_rearm_rto(sk);
2861}
2862
2863/* Push out any pending frames which were held back due to
2864 * TCP_CORK or attempt at coalescing tiny packets.
2865 * The socket must be locked by the caller.
2866 */
2867void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2868			       int nonagle)
2869{
2870	/* If we are closed, the bytes will have to remain here.
2871	 * In time closedown will finish, we empty the write queue and
2872	 * all will be happy.
2873	 */
2874	if (unlikely(sk->sk_state == TCP_CLOSE))
2875		return;
2876
2877	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2878			   sk_gfp_mask(sk, GFP_ATOMIC)))
2879		tcp_check_probe_timer(sk);
2880}
2881
2882/* Send _single_ skb sitting at the send head. This function requires
2883 * true push pending frames to setup probe timer etc.
2884 */
2885void tcp_push_one(struct sock *sk, unsigned int mss_now)
2886{
2887	struct sk_buff *skb = tcp_send_head(sk);
2888
2889	BUG_ON(!skb || skb->len < mss_now);
2890
2891	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2892}
2893
2894/* This function returns the amount that we can raise the
2895 * usable window based on the following constraints
2896 *
2897 * 1. The window can never be shrunk once it is offered (RFC 793)
2898 * 2. We limit memory per socket
2899 *
2900 * RFC 1122:
2901 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2902 *  RECV.NEXT + RCV.WIN fixed until:
2903 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2904 *
2905 * i.e. don't raise the right edge of the window until you can raise
2906 * it at least MSS bytes.
2907 *
2908 * Unfortunately, the recommended algorithm breaks header prediction,
2909 * since header prediction assumes th->window stays fixed.
2910 *
2911 * Strictly speaking, keeping th->window fixed violates the receiver
2912 * side SWS prevention criteria. The problem is that under this rule
2913 * a stream of single byte packets will cause the right side of the
2914 * window to always advance by a single byte.
2915 *
2916 * Of course, if the sender implements sender side SWS prevention
2917 * then this will not be a problem.
2918 *
2919 * BSD seems to make the following compromise:
2920 *
2921 *	If the free space is less than the 1/4 of the maximum
2922 *	space available and the free space is less than 1/2 mss,
2923 *	then set the window to 0.
2924 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2925 *	Otherwise, just prevent the window from shrinking
2926 *	and from being larger than the largest representable value.
2927 *
2928 * This prevents incremental opening of the window in the regime
2929 * where TCP is limited by the speed of the reader side taking
2930 * data out of the TCP receive queue. It does nothing about
2931 * those cases where the window is constrained on the sender side
2932 * because the pipeline is full.
2933 *
2934 * BSD also seems to "accidentally" limit itself to windows that are a
2935 * multiple of MSS, at least until the free space gets quite small.
2936 * This would appear to be a side effect of the mbuf implementation.
2937 * Combining these two algorithms results in the observed behavior
2938 * of having a fixed window size at almost all times.
2939 *
2940 * Below we obtain similar behavior by forcing the offered window to
2941 * a multiple of the mss when it is feasible to do so.
2942 *
2943 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2944 * Regular options like TIMESTAMP are taken into account.
2945 */
2946u32 __tcp_select_window(struct sock *sk)
2947{
2948	struct inet_connection_sock *icsk = inet_csk(sk);
2949	struct tcp_sock *tp = tcp_sk(sk);
 
2950	/* MSS for the peer's data.  Previous versions used mss_clamp
2951	 * here.  I don't know if the value based on our guesses
2952	 * of peer's MSS is better for the performance.  It's more correct
2953	 * but may be worse for the performance because of rcv_mss
2954	 * fluctuations.  --SAW  1998/11/1
2955	 */
2956	int mss = icsk->icsk_ack.rcv_mss;
2957	int free_space = tcp_space(sk);
2958	int allowed_space = tcp_full_space(sk);
2959	int full_space, window;
2960
2961	if (sk_is_mptcp(sk))
2962		mptcp_space(sk, &free_space, &allowed_space);
2963
2964	full_space = min_t(int, tp->window_clamp, allowed_space);
2965
2966	if (unlikely(mss > full_space)) {
2967		mss = full_space;
2968		if (mss <= 0)
2969			return 0;
2970	}
 
 
 
 
 
 
 
 
 
2971	if (free_space < (full_space >> 1)) {
2972		icsk->icsk_ack.quick = 0;
2973
2974		if (tcp_under_memory_pressure(sk))
2975			tcp_adjust_rcv_ssthresh(sk);
2976
2977		/* free_space might become our new window, make sure we don't
2978		 * increase it due to wscale.
2979		 */
2980		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2981
2982		/* if free space is less than mss estimate, or is below 1/16th
2983		 * of the maximum allowed, try to move to zero-window, else
2984		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2985		 * new incoming data is dropped due to memory limits.
2986		 * With large window, mss test triggers way too late in order
2987		 * to announce zero window in time before rmem limit kicks in.
2988		 */
2989		if (free_space < (allowed_space >> 4) || free_space < mss)
2990			return 0;
2991	}
2992
2993	if (free_space > tp->rcv_ssthresh)
2994		free_space = tp->rcv_ssthresh;
2995
2996	/* Don't do rounding if we are using window scaling, since the
2997	 * scaled window will not line up with the MSS boundary anyway.
2998	 */
2999	if (tp->rx_opt.rcv_wscale) {
3000		window = free_space;
3001
3002		/* Advertise enough space so that it won't get scaled away.
3003		 * Import case: prevent zero window announcement if
3004		 * 1<<rcv_wscale > mss.
3005		 */
3006		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3007	} else {
3008		window = tp->rcv_wnd;
3009		/* Get the largest window that is a nice multiple of mss.
3010		 * Window clamp already applied above.
3011		 * If our current window offering is within 1 mss of the
3012		 * free space we just keep it. This prevents the divide
3013		 * and multiply from happening most of the time.
3014		 * We also don't do any window rounding when the free space
3015		 * is too small.
3016		 */
3017		if (window <= free_space - mss || window > free_space)
3018			window = rounddown(free_space, mss);
3019		else if (mss == full_space &&
3020			 free_space > window + (full_space >> 1))
3021			window = free_space;
3022	}
3023
3024	return window;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3025}
3026
3027void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3028			     const struct sk_buff *next_skb)
3029{
3030	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3031		const struct skb_shared_info *next_shinfo =
3032			skb_shinfo(next_skb);
3033		struct skb_shared_info *shinfo = skb_shinfo(skb);
3034
3035		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3036		shinfo->tskey = next_shinfo->tskey;
3037		TCP_SKB_CB(skb)->txstamp_ack |=
3038			TCP_SKB_CB(next_skb)->txstamp_ack;
3039	}
3040}
3041
3042/* Collapses two adjacent SKB's during retransmission. */
3043static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3044{
3045	struct tcp_sock *tp = tcp_sk(sk);
3046	struct sk_buff *next_skb = skb_rb_next(skb);
3047	int next_skb_size;
3048
3049	next_skb_size = next_skb->len;
3050
3051	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3052
3053	if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3054		return false;
3055
3056	tcp_highest_sack_replace(sk, next_skb, skb);
3057
3058	/* Update sequence range on original skb. */
3059	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3060
3061	/* Merge over control information. This moves PSH/FIN etc. over */
3062	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3063
3064	/* All done, get rid of second SKB and account for it so
3065	 * packet counting does not break.
3066	 */
3067	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3068	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3069
3070	/* changed transmit queue under us so clear hints */
3071	tcp_clear_retrans_hints_partial(tp);
3072	if (next_skb == tp->retransmit_skb_hint)
3073		tp->retransmit_skb_hint = skb;
3074
3075	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3076
3077	tcp_skb_collapse_tstamp(skb, next_skb);
3078
3079	tcp_rtx_queue_unlink_and_free(next_skb, sk);
3080	return true;
3081}
3082
3083/* Check if coalescing SKBs is legal. */
3084static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3085{
3086	if (tcp_skb_pcount(skb) > 1)
3087		return false;
3088	if (skb_cloned(skb))
3089		return false;
 
 
3090	/* Some heuristics for collapsing over SACK'd could be invented */
3091	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3092		return false;
3093
3094	return true;
3095}
3096
3097/* Collapse packets in the retransmit queue to make to create
3098 * less packets on the wire. This is only done on retransmission.
3099 */
3100static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3101				     int space)
3102{
3103	struct tcp_sock *tp = tcp_sk(sk);
3104	struct sk_buff *skb = to, *tmp;
3105	bool first = true;
3106
3107	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3108		return;
3109	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3110		return;
3111
3112	skb_rbtree_walk_from_safe(skb, tmp) {
3113		if (!tcp_can_collapse(sk, skb))
3114			break;
3115
3116		if (!tcp_skb_can_collapse(to, skb))
3117			break;
3118
3119		space -= skb->len;
3120
3121		if (first) {
3122			first = false;
3123			continue;
3124		}
3125
3126		if (space < 0)
3127			break;
3128
3129		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3130			break;
3131
3132		if (!tcp_collapse_retrans(sk, to))
3133			break;
3134	}
3135}
3136
3137/* This retransmits one SKB.  Policy decisions and retransmit queue
3138 * state updates are done by the caller.  Returns non-zero if an
3139 * error occurred which prevented the send.
3140 */
3141int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3142{
3143	struct inet_connection_sock *icsk = inet_csk(sk);
3144	struct tcp_sock *tp = tcp_sk(sk);
3145	unsigned int cur_mss;
3146	int diff, len, err;
3147	int avail_wnd;
3148
3149	/* Inconclusive MTU probe */
3150	if (icsk->icsk_mtup.probe_size)
3151		icsk->icsk_mtup.probe_size = 0;
3152
3153	if (skb_still_in_host_queue(sk, skb))
3154		return -EBUSY;
3155
 
3156	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
 
 
 
 
 
3157		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3158			WARN_ON_ONCE(1);
3159			return -EINVAL;
3160		}
3161		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3162			return -ENOMEM;
3163	}
3164
3165	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3166		return -EHOSTUNREACH; /* Routing failure or similar. */
3167
3168	cur_mss = tcp_current_mss(sk);
3169	avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3170
3171	/* If receiver has shrunk his window, and skb is out of
3172	 * new window, do not retransmit it. The exception is the
3173	 * case, when window is shrunk to zero. In this case
3174	 * our retransmit of one segment serves as a zero window probe.
3175	 */
3176	if (avail_wnd <= 0) {
3177		if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3178			return -EAGAIN;
3179		avail_wnd = cur_mss;
3180	}
3181
3182	len = cur_mss * segs;
3183	if (len > avail_wnd) {
3184		len = rounddown(avail_wnd, cur_mss);
3185		if (!len)
3186			len = avail_wnd;
3187	}
3188	if (skb->len > len) {
3189		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3190				 cur_mss, GFP_ATOMIC))
3191			return -ENOMEM; /* We'll try again later. */
3192	} else {
3193		if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3194			return -ENOMEM;
3195
3196		diff = tcp_skb_pcount(skb);
3197		tcp_set_skb_tso_segs(skb, cur_mss);
3198		diff -= tcp_skb_pcount(skb);
3199		if (diff)
3200			tcp_adjust_pcount(sk, skb, diff);
3201		avail_wnd = min_t(int, avail_wnd, cur_mss);
3202		if (skb->len < avail_wnd)
3203			tcp_retrans_try_collapse(sk, skb, avail_wnd);
3204	}
3205
3206	/* RFC3168, section 6.1.1.1. ECN fallback */
3207	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3208		tcp_ecn_clear_syn(sk, skb);
3209
3210	/* Update global and local TCP statistics. */
3211	segs = tcp_skb_pcount(skb);
3212	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3213	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3214		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3215	tp->total_retrans += segs;
3216	tp->bytes_retrans += skb->len;
3217
3218	/* make sure skb->data is aligned on arches that require it
3219	 * and check if ack-trimming & collapsing extended the headroom
3220	 * beyond what csum_start can cover.
3221	 */
3222	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3223		     skb_headroom(skb) >= 0xFFFF)) {
3224		struct sk_buff *nskb;
3225
3226		tcp_skb_tsorted_save(skb) {
3227			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3228			if (nskb) {
3229				nskb->dev = NULL;
3230				err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3231			} else {
3232				err = -ENOBUFS;
3233			}
3234		} tcp_skb_tsorted_restore(skb);
3235
3236		if (!err) {
3237			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3238			tcp_rate_skb_sent(sk, skb);
3239		}
3240	} else {
3241		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3242	}
3243
3244	/* To avoid taking spuriously low RTT samples based on a timestamp
3245	 * for a transmit that never happened, always mark EVER_RETRANS
3246	 */
3247	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3248
3249	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3250		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3251				  TCP_SKB_CB(skb)->seq, segs, err);
3252
3253	if (likely(!err)) {
3254		trace_tcp_retransmit_skb(sk, skb);
3255	} else if (err != -EBUSY) {
3256		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3257	}
 
 
 
 
 
 
3258	return err;
3259}
3260
3261int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3262{
3263	struct tcp_sock *tp = tcp_sk(sk);
3264	int err = __tcp_retransmit_skb(sk, skb, segs);
3265
3266	if (err == 0) {
3267#if FASTRETRANS_DEBUG > 0
3268		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3269			net_dbg_ratelimited("retrans_out leaked\n");
3270		}
3271#endif
3272		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3273		tp->retrans_out += tcp_skb_pcount(skb);
3274	}
3275
3276	/* Save stamp of the first (attempted) retransmit. */
3277	if (!tp->retrans_stamp)
3278		tp->retrans_stamp = tcp_skb_timestamp(skb);
3279
3280	if (tp->undo_retrans < 0)
3281		tp->undo_retrans = 0;
3282	tp->undo_retrans += tcp_skb_pcount(skb);
3283	return err;
3284}
3285
3286/* This gets called after a retransmit timeout, and the initially
3287 * retransmitted data is acknowledged.  It tries to continue
3288 * resending the rest of the retransmit queue, until either
3289 * we've sent it all or the congestion window limit is reached.
3290 */
3291void tcp_xmit_retransmit_queue(struct sock *sk)
3292{
3293	const struct inet_connection_sock *icsk = inet_csk(sk);
3294	struct sk_buff *skb, *rtx_head, *hole = NULL;
3295	struct tcp_sock *tp = tcp_sk(sk);
3296	bool rearm_timer = false;
3297	u32 max_segs;
3298	int mib_idx;
3299
3300	if (!tp->packets_out)
3301		return;
3302
3303	rtx_head = tcp_rtx_queue_head(sk);
3304	skb = tp->retransmit_skb_hint ?: rtx_head;
3305	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3306	skb_rbtree_walk_from(skb) {
3307		__u8 sacked;
3308		int segs;
3309
3310		if (tcp_pacing_check(sk))
3311			break;
3312
3313		/* we could do better than to assign each time */
3314		if (!hole)
3315			tp->retransmit_skb_hint = skb;
3316
3317		segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3318		if (segs <= 0)
3319			break;
3320		sacked = TCP_SKB_CB(skb)->sacked;
3321		/* In case tcp_shift_skb_data() have aggregated large skbs,
3322		 * we need to make sure not sending too bigs TSO packets
3323		 */
3324		segs = min_t(int, segs, max_segs);
3325
3326		if (tp->retrans_out >= tp->lost_out) {
3327			break;
3328		} else if (!(sacked & TCPCB_LOST)) {
3329			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3330				hole = skb;
3331			continue;
3332
3333		} else {
3334			if (icsk->icsk_ca_state != TCP_CA_Loss)
3335				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3336			else
3337				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3338		}
3339
3340		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3341			continue;
3342
3343		if (tcp_small_queue_check(sk, skb, 1))
3344			break;
3345
3346		if (tcp_retransmit_skb(sk, skb, segs))
3347			break;
3348
3349		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3350
3351		if (tcp_in_cwnd_reduction(sk))
3352			tp->prr_out += tcp_skb_pcount(skb);
3353
3354		if (skb == rtx_head &&
3355		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3356			rearm_timer = true;
3357
3358	}
3359	if (rearm_timer)
3360		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3361				     inet_csk(sk)->icsk_rto,
3362				     TCP_RTO_MAX);
3363}
3364
3365/* We allow to exceed memory limits for FIN packets to expedite
3366 * connection tear down and (memory) recovery.
3367 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3368 * or even be forced to close flow without any FIN.
3369 * In general, we want to allow one skb per socket to avoid hangs
3370 * with edge trigger epoll()
3371 */
3372void sk_forced_mem_schedule(struct sock *sk, int size)
3373{
3374	int delta, amt;
3375
3376	delta = size - sk->sk_forward_alloc;
3377	if (delta <= 0)
3378		return;
3379	amt = sk_mem_pages(delta);
3380	sk->sk_forward_alloc += amt << PAGE_SHIFT;
3381	sk_memory_allocated_add(sk, amt);
3382
3383	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3384		mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3385					gfp_memcg_charge() | __GFP_NOFAIL);
3386}
3387
3388/* Send a FIN. The caller locks the socket for us.
3389 * We should try to send a FIN packet really hard, but eventually give up.
3390 */
3391void tcp_send_fin(struct sock *sk)
3392{
3393	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3394	struct tcp_sock *tp = tcp_sk(sk);
3395
3396	/* Optimization, tack on the FIN if we have one skb in write queue and
3397	 * this skb was not yet sent, or we are under memory pressure.
3398	 * Note: in the latter case, FIN packet will be sent after a timeout,
3399	 * as TCP stack thinks it has already been transmitted.
3400	 */
3401	tskb = tail;
3402	if (!tskb && tcp_under_memory_pressure(sk))
3403		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3404
3405	if (tskb) {
3406		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3407		TCP_SKB_CB(tskb)->end_seq++;
3408		tp->write_seq++;
3409		if (!tail) {
3410			/* This means tskb was already sent.
3411			 * Pretend we included the FIN on previous transmit.
3412			 * We need to set tp->snd_nxt to the value it would have
3413			 * if FIN had been sent. This is because retransmit path
3414			 * does not change tp->snd_nxt.
3415			 */
3416			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3417			return;
3418		}
3419	} else {
3420		skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
 
 
3421		if (unlikely(!skb))
3422			return;
3423
3424		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3425		skb_reserve(skb, MAX_TCP_HEADER);
3426		sk_forced_mem_schedule(sk, skb->truesize);
3427		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3428		tcp_init_nondata_skb(skb, tp->write_seq,
3429				     TCPHDR_ACK | TCPHDR_FIN);
3430		tcp_queue_skb(sk, skb);
3431	}
3432	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3433}
3434
3435/* We get here when a process closes a file descriptor (either due to
3436 * an explicit close() or as a byproduct of exit()'ing) and there
3437 * was unread data in the receive queue.  This behavior is recommended
3438 * by RFC 2525, section 2.17.  -DaveM
3439 */
3440void tcp_send_active_reset(struct sock *sk, gfp_t priority)
 
3441{
3442	struct sk_buff *skb;
3443
3444	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3445
3446	/* NOTE: No TCP options attached and we never retransmit this. */
3447	skb = alloc_skb(MAX_TCP_HEADER, priority);
3448	if (!skb) {
3449		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3450		return;
3451	}
3452
3453	/* Reserve space for headers and prepare control bits. */
3454	skb_reserve(skb, MAX_TCP_HEADER);
3455	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3456			     TCPHDR_ACK | TCPHDR_RST);
3457	tcp_mstamp_refresh(tcp_sk(sk));
3458	/* Send it off. */
3459	if (tcp_transmit_skb(sk, skb, 0, priority))
3460		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3461
3462	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3463	 * skb here is different to the troublesome skb, so use NULL
3464	 */
3465	trace_tcp_send_reset(sk, NULL);
3466}
3467
3468/* Send a crossed SYN-ACK during socket establishment.
3469 * WARNING: This routine must only be called when we have already sent
3470 * a SYN packet that crossed the incoming SYN that caused this routine
3471 * to get called. If this assumption fails then the initial rcv_wnd
3472 * and rcv_wscale values will not be correct.
3473 */
3474int tcp_send_synack(struct sock *sk)
3475{
3476	struct sk_buff *skb;
3477
3478	skb = tcp_rtx_queue_head(sk);
3479	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3480		pr_err("%s: wrong queue state\n", __func__);
3481		return -EFAULT;
3482	}
3483	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3484		if (skb_cloned(skb)) {
3485			struct sk_buff *nskb;
3486
3487			tcp_skb_tsorted_save(skb) {
3488				nskb = skb_copy(skb, GFP_ATOMIC);
3489			} tcp_skb_tsorted_restore(skb);
3490			if (!nskb)
3491				return -ENOMEM;
3492			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3493			tcp_highest_sack_replace(sk, skb, nskb);
3494			tcp_rtx_queue_unlink_and_free(skb, sk);
3495			__skb_header_release(nskb);
3496			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3497			sk_wmem_queued_add(sk, nskb->truesize);
3498			sk_mem_charge(sk, nskb->truesize);
3499			skb = nskb;
3500		}
3501
3502		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3503		tcp_ecn_send_synack(sk, skb);
3504	}
3505	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3506}
3507
3508/**
3509 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3510 * @sk: listener socket
3511 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3512 *       should not use it again.
3513 * @req: request_sock pointer
3514 * @foc: cookie for tcp fast open
3515 * @synack_type: Type of synack to prepare
3516 * @syn_skb: SYN packet just received.  It could be NULL for rtx case.
3517 */
3518struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3519				struct request_sock *req,
3520				struct tcp_fastopen_cookie *foc,
3521				enum tcp_synack_type synack_type,
3522				struct sk_buff *syn_skb)
3523{
3524	struct inet_request_sock *ireq = inet_rsk(req);
3525	const struct tcp_sock *tp = tcp_sk(sk);
3526	struct tcp_md5sig_key *md5 = NULL;
3527	struct tcp_out_options opts;
 
3528	struct sk_buff *skb;
3529	int tcp_header_size;
3530	struct tcphdr *th;
3531	int mss;
3532	u64 now;
3533
3534	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3535	if (unlikely(!skb)) {
3536		dst_release(dst);
3537		return NULL;
3538	}
3539	/* Reserve space for headers. */
3540	skb_reserve(skb, MAX_TCP_HEADER);
3541
3542	switch (synack_type) {
3543	case TCP_SYNACK_NORMAL:
3544		skb_set_owner_w(skb, req_to_sk(req));
3545		break;
3546	case TCP_SYNACK_COOKIE:
3547		/* Under synflood, we do not attach skb to a socket,
3548		 * to avoid false sharing.
3549		 */
3550		break;
3551	case TCP_SYNACK_FASTOPEN:
3552		/* sk is a const pointer, because we want to express multiple
3553		 * cpu might call us concurrently.
3554		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3555		 */
3556		skb_set_owner_w(skb, (struct sock *)sk);
3557		break;
3558	}
3559	skb_dst_set(skb, dst);
3560
3561	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3562
3563	memset(&opts, 0, sizeof(opts));
3564	now = tcp_clock_ns();
3565#ifdef CONFIG_SYN_COOKIES
3566	if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3567		skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3568				      true);
3569	else
3570#endif
3571	{
3572		skb_set_delivery_time(skb, now, true);
3573		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3574			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3575	}
3576
3577#ifdef CONFIG_TCP_MD5SIG
3578	rcu_read_lock();
3579	md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3580#endif
3581	skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3582	/* bpf program will be interested in the tcp_flags */
3583	TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3584	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3585					     foc, synack_type,
3586					     syn_skb) + sizeof(*th);
3587
3588	skb_push(skb, tcp_header_size);
3589	skb_reset_transport_header(skb);
3590
3591	th = (struct tcphdr *)skb->data;
3592	memset(th, 0, sizeof(struct tcphdr));
3593	th->syn = 1;
3594	th->ack = 1;
3595	tcp_ecn_make_synack(req, th);
3596	th->source = htons(ireq->ir_num);
3597	th->dest = ireq->ir_rmt_port;
3598	skb->mark = ireq->ir_mark;
3599	skb->ip_summed = CHECKSUM_PARTIAL;
3600	th->seq = htonl(tcp_rsk(req)->snt_isn);
3601	/* XXX data is queued and acked as is. No buffer/window check */
3602	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3603
3604	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3605	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3606	tcp_options_write(th, NULL, &opts);
3607	th->doff = (tcp_header_size >> 2);
3608	__TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3609
3610#ifdef CONFIG_TCP_MD5SIG
3611	/* Okay, we have all we need - do the md5 hash if needed */
3612	if (md5)
 
3613		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3614					       md5, req_to_sk(req), skb);
 
 
 
 
 
 
 
 
 
3615	rcu_read_unlock();
3616#endif
3617
3618	bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3619				synack_type, &opts);
3620
3621	skb_set_delivery_time(skb, now, true);
3622	tcp_add_tx_delay(skb, tp);
3623
3624	return skb;
3625}
3626EXPORT_SYMBOL(tcp_make_synack);
3627
3628static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3629{
3630	struct inet_connection_sock *icsk = inet_csk(sk);
3631	const struct tcp_congestion_ops *ca;
3632	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3633
3634	if (ca_key == TCP_CA_UNSPEC)
3635		return;
3636
3637	rcu_read_lock();
3638	ca = tcp_ca_find_key(ca_key);
3639	if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3640		bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3641		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3642		icsk->icsk_ca_ops = ca;
3643	}
3644	rcu_read_unlock();
3645}
3646
3647/* Do all connect socket setups that can be done AF independent. */
3648static void tcp_connect_init(struct sock *sk)
3649{
3650	const struct dst_entry *dst = __sk_dst_get(sk);
3651	struct tcp_sock *tp = tcp_sk(sk);
3652	__u8 rcv_wscale;
3653	u32 rcv_wnd;
3654
3655	/* We'll fix this up when we get a response from the other end.
3656	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3657	 */
3658	tp->tcp_header_len = sizeof(struct tcphdr);
3659	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3660		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3661
3662#ifdef CONFIG_TCP_MD5SIG
3663	if (tp->af_specific->md5_lookup(sk, sk))
3664		tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3665#endif
3666
3667	/* If user gave his TCP_MAXSEG, record it to clamp */
3668	if (tp->rx_opt.user_mss)
3669		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3670	tp->max_window = 0;
3671	tcp_mtup_init(sk);
3672	tcp_sync_mss(sk, dst_mtu(dst));
3673
3674	tcp_ca_dst_init(sk, dst);
3675
3676	if (!tp->window_clamp)
3677		tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3678	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3679
3680	tcp_initialize_rcv_mss(sk);
3681
3682	/* limit the window selection if the user enforce a smaller rx buffer */
3683	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3684	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3685		tp->window_clamp = tcp_full_space(sk);
3686
3687	rcv_wnd = tcp_rwnd_init_bpf(sk);
3688	if (rcv_wnd == 0)
3689		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3690
3691	tcp_select_initial_window(sk, tcp_full_space(sk),
3692				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3693				  &tp->rcv_wnd,
3694				  &tp->window_clamp,
3695				  READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3696				  &rcv_wscale,
3697				  rcv_wnd);
3698
3699	tp->rx_opt.rcv_wscale = rcv_wscale;
3700	tp->rcv_ssthresh = tp->rcv_wnd;
3701
3702	sk->sk_err = 0;
3703	sock_reset_flag(sk, SOCK_DONE);
3704	tp->snd_wnd = 0;
3705	tcp_init_wl(tp, 0);
3706	tcp_write_queue_purge(sk);
3707	tp->snd_una = tp->write_seq;
3708	tp->snd_sml = tp->write_seq;
3709	tp->snd_up = tp->write_seq;
3710	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3711
3712	if (likely(!tp->repair))
3713		tp->rcv_nxt = 0;
3714	else
3715		tp->rcv_tstamp = tcp_jiffies32;
3716	tp->rcv_wup = tp->rcv_nxt;
3717	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3718
3719	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3720	inet_csk(sk)->icsk_retransmits = 0;
3721	tcp_clear_retrans(tp);
3722}
3723
3724static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3725{
3726	struct tcp_sock *tp = tcp_sk(sk);
3727	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3728
3729	tcb->end_seq += skb->len;
3730	__skb_header_release(skb);
3731	sk_wmem_queued_add(sk, skb->truesize);
3732	sk_mem_charge(sk, skb->truesize);
3733	WRITE_ONCE(tp->write_seq, tcb->end_seq);
3734	tp->packets_out += tcp_skb_pcount(skb);
3735}
3736
3737/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3738 * queue a data-only packet after the regular SYN, such that regular SYNs
3739 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3740 * only the SYN sequence, the data are retransmitted in the first ACK.
3741 * If cookie is not cached or other error occurs, falls back to send a
3742 * regular SYN with Fast Open cookie request option.
3743 */
3744static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3745{
3746	struct inet_connection_sock *icsk = inet_csk(sk);
3747	struct tcp_sock *tp = tcp_sk(sk);
3748	struct tcp_fastopen_request *fo = tp->fastopen_req;
3749	int space, err = 0;
3750	struct sk_buff *syn_data;
 
3751
3752	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3753	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3754		goto fallback;
3755
3756	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3757	 * user-MSS. Reserve maximum option space for middleboxes that add
3758	 * private TCP options. The cost is reduced data space in SYN :(
3759	 */
3760	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3761	/* Sync mss_cache after updating the mss_clamp */
3762	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3763
3764	space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3765		MAX_TCP_OPTION_SPACE;
3766
3767	space = min_t(size_t, space, fo->size);
3768
3769	/* limit to order-0 allocations */
3770	space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3771
3772	syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false);
 
3773	if (!syn_data)
3774		goto fallback;
3775	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3776	if (space) {
3777		int copied = copy_from_iter(skb_put(syn_data, space), space,
3778					    &fo->data->msg_iter);
3779		if (unlikely(!copied)) {
 
 
 
 
3780			tcp_skb_tsorted_anchor_cleanup(syn_data);
3781			kfree_skb(syn_data);
3782			goto fallback;
3783		}
3784		if (copied != space) {
3785			skb_trim(syn_data, copied);
3786			space = copied;
3787		}
 
3788		skb_zcopy_set(syn_data, fo->uarg, NULL);
3789	}
3790	/* No more data pending in inet_wait_for_connect() */
3791	if (space == fo->size)
3792		fo->data = NULL;
3793	fo->copied = space;
3794
3795	tcp_connect_queue_skb(sk, syn_data);
3796	if (syn_data->len)
3797		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3798
3799	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3800
3801	skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3802
3803	/* Now full SYN+DATA was cloned and sent (or not),
3804	 * remove the SYN from the original skb (syn_data)
3805	 * we keep in write queue in case of a retransmit, as we
3806	 * also have the SYN packet (with no data) in the same queue.
3807	 */
3808	TCP_SKB_CB(syn_data)->seq++;
3809	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3810	if (!err) {
3811		tp->syn_data = (fo->copied > 0);
3812		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3813		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3814		goto done;
3815	}
3816
3817	/* data was not sent, put it in write_queue */
3818	__skb_queue_tail(&sk->sk_write_queue, syn_data);
3819	tp->packets_out -= tcp_skb_pcount(syn_data);
3820
3821fallback:
3822	/* Send a regular SYN with Fast Open cookie request option */
3823	if (fo->cookie.len > 0)
3824		fo->cookie.len = 0;
3825	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3826	if (err)
3827		tp->syn_fastopen = 0;
3828done:
3829	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
3830	return err;
3831}
3832
3833/* Build a SYN and send it off. */
3834int tcp_connect(struct sock *sk)
3835{
3836	struct tcp_sock *tp = tcp_sk(sk);
3837	struct sk_buff *buff;
3838	int err;
3839
3840	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3841
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3842	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3843		return -EHOSTUNREACH; /* Routing failure or similar. */
3844
3845	tcp_connect_init(sk);
3846
3847	if (unlikely(tp->repair)) {
3848		tcp_finish_connect(sk, NULL);
3849		return 0;
3850	}
3851
3852	buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3853	if (unlikely(!buff))
3854		return -ENOBUFS;
3855
3856	tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
 
 
 
3857	tcp_mstamp_refresh(tp);
3858	tp->retrans_stamp = tcp_time_stamp(tp);
3859	tcp_connect_queue_skb(sk, buff);
3860	tcp_ecn_send_syn(sk, buff);
3861	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3862
3863	/* Send off SYN; include data in Fast Open. */
3864	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3865	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3866	if (err == -ECONNREFUSED)
3867		return err;
3868
3869	/* We change tp->snd_nxt after the tcp_transmit_skb() call
3870	 * in order to make this packet get counted in tcpOutSegs.
3871	 */
3872	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3873	tp->pushed_seq = tp->write_seq;
3874	buff = tcp_send_head(sk);
3875	if (unlikely(buff)) {
3876		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3877		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
3878	}
3879	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3880
3881	/* Timer for repeating the SYN until an answer. */
3882	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3883				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3884	return 0;
3885}
3886EXPORT_SYMBOL(tcp_connect);
3887
 
 
 
 
 
 
 
3888/* Send out a delayed ack, the caller does the policy checking
3889 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
3890 * for details.
3891 */
3892void tcp_send_delayed_ack(struct sock *sk)
3893{
3894	struct inet_connection_sock *icsk = inet_csk(sk);
3895	int ato = icsk->icsk_ack.ato;
3896	unsigned long timeout;
3897
3898	if (ato > TCP_DELACK_MIN) {
3899		const struct tcp_sock *tp = tcp_sk(sk);
3900		int max_ato = HZ / 2;
3901
3902		if (inet_csk_in_pingpong_mode(sk) ||
3903		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3904			max_ato = TCP_DELACK_MAX;
3905
3906		/* Slow path, intersegment interval is "high". */
3907
3908		/* If some rtt estimate is known, use it to bound delayed ack.
3909		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3910		 * directly.
3911		 */
3912		if (tp->srtt_us) {
3913			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3914					TCP_DELACK_MIN);
3915
3916			if (rtt < max_ato)
3917				max_ato = rtt;
3918		}
3919
3920		ato = min(ato, max_ato);
3921	}
3922
3923	ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3924
3925	/* Stay within the limit we were given */
3926	timeout = jiffies + ato;
3927
3928	/* Use new timeout only if there wasn't a older one earlier. */
3929	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3930		/* If delack timer is about to expire, send ACK now. */
3931		if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3932			tcp_send_ack(sk);
3933			return;
3934		}
3935
3936		if (!time_before(timeout, icsk->icsk_ack.timeout))
3937			timeout = icsk->icsk_ack.timeout;
3938	}
3939	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
 
3940	icsk->icsk_ack.timeout = timeout;
3941	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3942}
3943
3944/* This routine sends an ack and also updates the window. */
3945void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3946{
3947	struct sk_buff *buff;
3948
3949	/* If we have been reset, we may not send again. */
3950	if (sk->sk_state == TCP_CLOSE)
3951		return;
3952
3953	/* We are not putting this on the write queue, so
3954	 * tcp_transmit_skb() will set the ownership to this
3955	 * sock.
3956	 */
3957	buff = alloc_skb(MAX_TCP_HEADER,
3958			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3959	if (unlikely(!buff)) {
3960		struct inet_connection_sock *icsk = inet_csk(sk);
3961		unsigned long delay;
3962
3963		delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3964		if (delay < TCP_RTO_MAX)
3965			icsk->icsk_ack.retry++;
3966		inet_csk_schedule_ack(sk);
3967		icsk->icsk_ack.ato = TCP_ATO_MIN;
3968		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3969		return;
3970	}
3971
3972	/* Reserve space for headers and prepare control bits. */
3973	skb_reserve(buff, MAX_TCP_HEADER);
3974	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3975
3976	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
3977	 * too much.
3978	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3979	 */
3980	skb_set_tcp_pure_ack(buff);
3981
3982	/* Send it off, this clears delayed acks for us. */
3983	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3984}
3985EXPORT_SYMBOL_GPL(__tcp_send_ack);
3986
3987void tcp_send_ack(struct sock *sk)
3988{
3989	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3990}
3991
3992/* This routine sends a packet with an out of date sequence
3993 * number. It assumes the other end will try to ack it.
3994 *
3995 * Question: what should we make while urgent mode?
3996 * 4.4BSD forces sending single byte of data. We cannot send
3997 * out of window data, because we have SND.NXT==SND.MAX...
3998 *
3999 * Current solution: to send TWO zero-length segments in urgent mode:
4000 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4001 * out-of-date with SND.UNA-1 to probe window.
4002 */
4003static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4004{
4005	struct tcp_sock *tp = tcp_sk(sk);
4006	struct sk_buff *skb;
4007
4008	/* We don't queue it, tcp_transmit_skb() sets ownership. */
4009	skb = alloc_skb(MAX_TCP_HEADER,
4010			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4011	if (!skb)
4012		return -1;
4013
4014	/* Reserve space for headers and set control bits. */
4015	skb_reserve(skb, MAX_TCP_HEADER);
4016	/* Use a previous sequence.  This should cause the other
4017	 * end to send an ack.  Don't queue or clone SKB, just
4018	 * send it.
4019	 */
4020	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4021	NET_INC_STATS(sock_net(sk), mib);
4022	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4023}
4024
4025/* Called from setsockopt( ... TCP_REPAIR ) */
4026void tcp_send_window_probe(struct sock *sk)
4027{
4028	if (sk->sk_state == TCP_ESTABLISHED) {
4029		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4030		tcp_mstamp_refresh(tcp_sk(sk));
4031		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4032	}
4033}
4034
4035/* Initiate keepalive or window probe from timer. */
4036int tcp_write_wakeup(struct sock *sk, int mib)
4037{
4038	struct tcp_sock *tp = tcp_sk(sk);
4039	struct sk_buff *skb;
4040
4041	if (sk->sk_state == TCP_CLOSE)
4042		return -1;
4043
4044	skb = tcp_send_head(sk);
4045	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4046		int err;
4047		unsigned int mss = tcp_current_mss(sk);
4048		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4049
4050		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4051			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4052
4053		/* We are probing the opening of a window
4054		 * but the window size is != 0
4055		 * must have been a result SWS avoidance ( sender )
4056		 */
4057		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4058		    skb->len > mss) {
4059			seg_size = min(seg_size, mss);
4060			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4061			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4062					 skb, seg_size, mss, GFP_ATOMIC))
4063				return -1;
4064		} else if (!tcp_skb_pcount(skb))
4065			tcp_set_skb_tso_segs(skb, mss);
4066
4067		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4068		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4069		if (!err)
4070			tcp_event_new_data_sent(sk, skb);
4071		return err;
4072	} else {
4073		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4074			tcp_xmit_probe_skb(sk, 1, mib);
4075		return tcp_xmit_probe_skb(sk, 0, mib);
4076	}
4077}
4078
4079/* A window probe timeout has occurred.  If window is not closed send
4080 * a partial packet else a zero probe.
4081 */
4082void tcp_send_probe0(struct sock *sk)
4083{
4084	struct inet_connection_sock *icsk = inet_csk(sk);
4085	struct tcp_sock *tp = tcp_sk(sk);
4086	struct net *net = sock_net(sk);
4087	unsigned long timeout;
4088	int err;
4089
4090	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4091
4092	if (tp->packets_out || tcp_write_queue_empty(sk)) {
4093		/* Cancel probe timer, if it is not required. */
4094		icsk->icsk_probes_out = 0;
4095		icsk->icsk_backoff = 0;
4096		icsk->icsk_probes_tstamp = 0;
4097		return;
4098	}
4099
4100	icsk->icsk_probes_out++;
4101	if (err <= 0) {
4102		if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4103			icsk->icsk_backoff++;
4104		timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4105	} else {
4106		/* If packet was not sent due to local congestion,
4107		 * Let senders fight for local resources conservatively.
4108		 */
4109		timeout = TCP_RESOURCE_PROBE_INTERVAL;
4110	}
4111
4112	timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4113	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4114}
4115
4116int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4117{
4118	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4119	struct flowi fl;
4120	int res;
4121
4122	/* Paired with WRITE_ONCE() in sock_setsockopt() */
4123	if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4124		tcp_rsk(req)->txhash = net_tx_rndhash();
4125	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4126				  NULL);
4127	if (!res) {
4128		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4129		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4130		if (unlikely(tcp_passive_fastopen(sk)))
4131			tcp_sk(sk)->total_retrans++;
 
 
 
 
 
4132		trace_tcp_retransmit_synack(sk, req);
4133	}
4134	return res;
4135}
4136EXPORT_SYMBOL(tcp_rtx_synack);
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * INET		An implementation of the TCP/IP protocol suite for the LINUX
   4 *		operating system.  INET is implemented using the  BSD Socket
   5 *		interface as the means of communication with the user level.
   6 *
   7 *		Implementation of the Transmission Control Protocol(TCP).
   8 *
   9 * Authors:	Ross Biro
  10 *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
  11 *		Mark Evans, <evansmp@uhura.aston.ac.uk>
  12 *		Corey Minyard <wf-rch!minyard@relay.EU.net>
  13 *		Florian La Roche, <flla@stud.uni-sb.de>
  14 *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
  15 *		Linus Torvalds, <torvalds@cs.helsinki.fi>
  16 *		Alan Cox, <gw4pts@gw4pts.ampr.org>
  17 *		Matthew Dillon, <dillon@apollo.west.oic.com>
  18 *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
  19 *		Jorge Cwik, <jorge@laser.satlink.net>
  20 */
  21
  22/*
  23 * Changes:	Pedro Roque	:	Retransmit queue handled by TCP.
  24 *				:	Fragmentation on mtu decrease
  25 *				:	Segment collapse on retransmit
  26 *				:	AF independence
  27 *
  28 *		Linus Torvalds	:	send_delayed_ack
  29 *		David S. Miller	:	Charge memory using the right skb
  30 *					during syn/ack processing.
  31 *		David S. Miller :	Output engine completely rewritten.
  32 *		Andrea Arcangeli:	SYNACK carry ts_recent in tsecr.
  33 *		Cacophonix Gaul :	draft-minshall-nagle-01
  34 *		J Hadi Salim	:	ECN support
  35 *
  36 */
  37
  38#define pr_fmt(fmt) "TCP: " fmt
  39
  40#include <net/tcp.h>
  41#include <net/mptcp.h>
  42#include <net/proto_memory.h>
  43
  44#include <linux/compiler.h>
  45#include <linux/gfp.h>
  46#include <linux/module.h>
  47#include <linux/static_key.h>
  48#include <linux/skbuff_ref.h>
  49
  50#include <trace/events/tcp.h>
  51
  52/* Refresh clocks of a TCP socket,
  53 * ensuring monotically increasing values.
  54 */
  55void tcp_mstamp_refresh(struct tcp_sock *tp)
  56{
  57	u64 val = tcp_clock_ns();
  58
  59	tp->tcp_clock_cache = val;
  60	tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
  61}
  62
  63static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
  64			   int push_one, gfp_t gfp);
  65
  66/* Account for new data that has been sent to the network. */
  67static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
  68{
  69	struct inet_connection_sock *icsk = inet_csk(sk);
  70	struct tcp_sock *tp = tcp_sk(sk);
  71	unsigned int prior_packets = tp->packets_out;
  72
  73	WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
  74
  75	__skb_unlink(skb, &sk->sk_write_queue);
  76	tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
  77
  78	if (tp->highest_sack == NULL)
  79		tp->highest_sack = skb;
  80
  81	tp->packets_out += tcp_skb_pcount(skb);
  82	if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
  83		tcp_rearm_rto(sk);
  84
  85	NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
  86		      tcp_skb_pcount(skb));
  87	tcp_check_space(sk);
  88}
  89
  90/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
  91 * window scaling factor due to loss of precision.
  92 * If window has been shrunk, what should we make? It is not clear at all.
  93 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
  94 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
  95 * invalid. OK, let's make this for now:
  96 */
  97static inline __u32 tcp_acceptable_seq(const struct sock *sk)
  98{
  99	const struct tcp_sock *tp = tcp_sk(sk);
 100
 101	if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
 102	    (tp->rx_opt.wscale_ok &&
 103	     ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
 104		return tp->snd_nxt;
 105	else
 106		return tcp_wnd_end(tp);
 107}
 108
 109/* Calculate mss to advertise in SYN segment.
 110 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
 111 *
 112 * 1. It is independent of path mtu.
 113 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
 114 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
 115 *    attached devices, because some buggy hosts are confused by
 116 *    large MSS.
 117 * 4. We do not make 3, we advertise MSS, calculated from first
 118 *    hop device mtu, but allow to raise it to ip_rt_min_advmss.
 119 *    This may be overridden via information stored in routing table.
 120 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
 121 *    probably even Jumbo".
 122 */
 123static __u16 tcp_advertise_mss(struct sock *sk)
 124{
 125	struct tcp_sock *tp = tcp_sk(sk);
 126	const struct dst_entry *dst = __sk_dst_get(sk);
 127	int mss = tp->advmss;
 128
 129	if (dst) {
 130		unsigned int metric = dst_metric_advmss(dst);
 131
 132		if (metric < mss) {
 133			mss = metric;
 134			tp->advmss = mss;
 135		}
 136	}
 137
 138	return (__u16)mss;
 139}
 140
 141/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
 142 * This is the first part of cwnd validation mechanism.
 143 */
 144void tcp_cwnd_restart(struct sock *sk, s32 delta)
 145{
 146	struct tcp_sock *tp = tcp_sk(sk);
 147	u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
 148	u32 cwnd = tcp_snd_cwnd(tp);
 149
 150	tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
 151
 152	tp->snd_ssthresh = tcp_current_ssthresh(sk);
 153	restart_cwnd = min(restart_cwnd, cwnd);
 154
 155	while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
 156		cwnd >>= 1;
 157	tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
 158	tp->snd_cwnd_stamp = tcp_jiffies32;
 159	tp->snd_cwnd_used = 0;
 160}
 161
 162/* Congestion state accounting after a packet has been sent. */
 163static void tcp_event_data_sent(struct tcp_sock *tp,
 164				struct sock *sk)
 165{
 166	struct inet_connection_sock *icsk = inet_csk(sk);
 167	const u32 now = tcp_jiffies32;
 168
 169	if (tcp_packets_in_flight(tp) == 0)
 170		tcp_ca_event(sk, CA_EVENT_TX_START);
 171
 172	tp->lsndtime = now;
 173
 174	/* If it is a reply for ato after last received
 175	 * packet, increase pingpong count.
 176	 */
 177	if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
 178		inet_csk_inc_pingpong_cnt(sk);
 179}
 180
 181/* Account for an ACK we sent. */
 182static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
 
 183{
 184	struct tcp_sock *tp = tcp_sk(sk);
 185
 186	if (unlikely(tp->compressed_ack)) {
 187		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
 188			      tp->compressed_ack);
 189		tp->compressed_ack = 0;
 190		if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
 191			__sock_put(sk);
 192	}
 193
 194	if (unlikely(rcv_nxt != tp->rcv_nxt))
 195		return;  /* Special ACK sent by DCTCP to reflect ECN */
 196	tcp_dec_quickack_mode(sk);
 197	inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
 198}
 199
 200/* Determine a window scaling and initial window to offer.
 201 * Based on the assumption that the given amount of space
 202 * will be offered. Store the results in the tp structure.
 203 * NOTE: for smooth operation initial space offering should
 204 * be a multiple of mss if possible. We assume here that mss >= 1.
 205 * This MUST be enforced by all callers.
 206 */
 207void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
 208			       __u32 *rcv_wnd, __u32 *__window_clamp,
 209			       int wscale_ok, __u8 *rcv_wscale,
 210			       __u32 init_rcv_wnd)
 211{
 212	unsigned int space = (__space < 0 ? 0 : __space);
 213	u32 window_clamp = READ_ONCE(*__window_clamp);
 214
 215	/* If no clamp set the clamp to the max possible scaled window */
 216	if (window_clamp == 0)
 217		window_clamp = (U16_MAX << TCP_MAX_WSCALE);
 218	space = min(window_clamp, space);
 219
 220	/* Quantize space offering to a multiple of mss if possible. */
 221	if (space > mss)
 222		space = rounddown(space, mss);
 223
 224	/* NOTE: offering an initial window larger than 32767
 225	 * will break some buggy TCP stacks. If the admin tells us
 226	 * it is likely we could be speaking with such a buggy stack
 227	 * we will truncate our initial window offering to 32K-1
 228	 * unless the remote has sent us a window scaling option,
 229	 * which we interpret as a sign the remote TCP is not
 230	 * misinterpreting the window field as a signed quantity.
 231	 */
 232	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
 233		(*rcv_wnd) = min(space, MAX_TCP_WINDOW);
 234	else
 235		(*rcv_wnd) = space;
 236
 237	if (init_rcv_wnd)
 238		*rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
 239
 240	*rcv_wscale = 0;
 241	if (wscale_ok) {
 242		/* Set window scaling on max possible window */
 243		space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
 244		space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
 245		space = min_t(u32, space, window_clamp);
 246		*rcv_wscale = clamp_t(int, ilog2(space) - 15,
 247				      0, TCP_MAX_WSCALE);
 248	}
 249	/* Set the clamp no higher than max representable value */
 250	WRITE_ONCE(*__window_clamp,
 251		   min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp));
 252}
 253EXPORT_SYMBOL(tcp_select_initial_window);
 254
 255/* Chose a new window to advertise, update state in tcp_sock for the
 256 * socket, and return result with RFC1323 scaling applied.  The return
 257 * value can be stuffed directly into th->window for an outgoing
 258 * frame.
 259 */
 260static u16 tcp_select_window(struct sock *sk)
 261{
 262	struct tcp_sock *tp = tcp_sk(sk);
 263	struct net *net = sock_net(sk);
 264	u32 old_win = tp->rcv_wnd;
 265	u32 cur_win, new_win;
 266
 267	/* Make the window 0 if we failed to queue the data because we
 268	 * are out of memory.
 269	 */
 270	if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) {
 271		tp->pred_flags = 0;
 272		tp->rcv_wnd = 0;
 273		tp->rcv_wup = tp->rcv_nxt;
 274		return 0;
 275	}
 276
 277	cur_win = tcp_receive_window(tp);
 278	new_win = __tcp_select_window(sk);
 279	if (new_win < cur_win) {
 280		/* Danger Will Robinson!
 281		 * Don't update rcv_wup/rcv_wnd here or else
 282		 * we will not be able to advertise a zero
 283		 * window in time.  --DaveM
 284		 *
 285		 * Relax Will Robinson.
 286		 */
 287		if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
 288			/* Never shrink the offered window */
 289			if (new_win == 0)
 290				NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
 291			new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
 292		}
 293	}
 294
 295	tp->rcv_wnd = new_win;
 296	tp->rcv_wup = tp->rcv_nxt;
 297
 298	/* Make sure we do not exceed the maximum possible
 299	 * scaled window.
 300	 */
 301	if (!tp->rx_opt.rcv_wscale &&
 302	    READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
 303		new_win = min(new_win, MAX_TCP_WINDOW);
 304	else
 305		new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
 306
 307	/* RFC1323 scaling applied */
 308	new_win >>= tp->rx_opt.rcv_wscale;
 309
 310	/* If we advertise zero window, disable fast path. */
 311	if (new_win == 0) {
 312		tp->pred_flags = 0;
 313		if (old_win)
 314			NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
 
 315	} else if (old_win == 0) {
 316		NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
 317	}
 318
 319	return new_win;
 320}
 321
 322/* Packet ECN state for a SYN-ACK */
 323static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
 324{
 325	const struct tcp_sock *tp = tcp_sk(sk);
 326
 327	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
 328	if (!(tp->ecn_flags & TCP_ECN_OK))
 329		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
 330	else if (tcp_ca_needs_ecn(sk) ||
 331		 tcp_bpf_ca_needs_ecn(sk))
 332		INET_ECN_xmit(sk);
 333}
 334
 335/* Packet ECN state for a SYN.  */
 336static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
 337{
 338	struct tcp_sock *tp = tcp_sk(sk);
 339	bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
 340	bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
 341		tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
 342
 343	if (!use_ecn) {
 344		const struct dst_entry *dst = __sk_dst_get(sk);
 345
 346		if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
 347			use_ecn = true;
 348	}
 349
 350	tp->ecn_flags = 0;
 351
 352	if (use_ecn) {
 353		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
 354		tp->ecn_flags = TCP_ECN_OK;
 355		if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
 356			INET_ECN_xmit(sk);
 357	}
 358}
 359
 360static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
 361{
 362	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
 363		/* tp->ecn_flags are cleared at a later point in time when
 364		 * SYN ACK is ultimatively being received.
 365		 */
 366		TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
 367}
 368
 369static void
 370tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
 371{
 372	if (inet_rsk(req)->ecn_ok)
 373		th->ece = 1;
 374}
 375
 376/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
 377 * be sent.
 378 */
 379static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
 380			 struct tcphdr *th, int tcp_header_len)
 381{
 382	struct tcp_sock *tp = tcp_sk(sk);
 383
 384	if (tp->ecn_flags & TCP_ECN_OK) {
 385		/* Not-retransmitted data segment: set ECT and inject CWR. */
 386		if (skb->len != tcp_header_len &&
 387		    !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
 388			INET_ECN_xmit(sk);
 389			if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
 390				tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
 391				th->cwr = 1;
 392				skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
 393			}
 394		} else if (!tcp_ca_needs_ecn(sk)) {
 395			/* ACK or retransmitted segment: clear ECT|CE */
 396			INET_ECN_dontxmit(sk);
 397		}
 398		if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
 399			th->ece = 1;
 400	}
 401}
 402
 403/* Constructs common control bits of non-data skb. If SYN/FIN is present,
 404 * auto increment end seqno.
 405 */
 406static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
 407{
 408	skb->ip_summed = CHECKSUM_PARTIAL;
 409
 410	TCP_SKB_CB(skb)->tcp_flags = flags;
 411
 412	tcp_skb_pcount_set(skb, 1);
 413
 414	TCP_SKB_CB(skb)->seq = seq;
 415	if (flags & (TCPHDR_SYN | TCPHDR_FIN))
 416		seq++;
 417	TCP_SKB_CB(skb)->end_seq = seq;
 418}
 419
 420static inline bool tcp_urg_mode(const struct tcp_sock *tp)
 421{
 422	return tp->snd_una != tp->snd_up;
 423}
 424
 425#define OPTION_SACK_ADVERTISE	BIT(0)
 426#define OPTION_TS		BIT(1)
 427#define OPTION_MD5		BIT(2)
 428#define OPTION_WSCALE		BIT(3)
 429#define OPTION_FAST_OPEN_COOKIE	BIT(8)
 430#define OPTION_SMC		BIT(9)
 431#define OPTION_MPTCP		BIT(10)
 432#define OPTION_AO		BIT(11)
 433
 434static void smc_options_write(__be32 *ptr, u16 *options)
 435{
 436#if IS_ENABLED(CONFIG_SMC)
 437	if (static_branch_unlikely(&tcp_have_smc)) {
 438		if (unlikely(OPTION_SMC & *options)) {
 439			*ptr++ = htonl((TCPOPT_NOP  << 24) |
 440				       (TCPOPT_NOP  << 16) |
 441				       (TCPOPT_EXP <<  8) |
 442				       (TCPOLEN_EXP_SMC_BASE));
 443			*ptr++ = htonl(TCPOPT_SMC_MAGIC);
 444		}
 445	}
 446#endif
 447}
 448
 449struct tcp_out_options {
 450	u16 options;		/* bit field of OPTION_* */
 451	u16 mss;		/* 0 to disable */
 452	u8 ws;			/* window scale, 0 to disable */
 453	u8 num_sack_blocks;	/* number of SACK blocks to include */
 454	u8 hash_size;		/* bytes in hash_location */
 455	u8 bpf_opt_len;		/* length of BPF hdr option */
 456	__u8 *hash_location;	/* temporary pointer, overloaded */
 457	__u32 tsval, tsecr;	/* need to include OPTION_TS */
 458	struct tcp_fastopen_cookie *fastopen_cookie;	/* Fast open cookie */
 459	struct mptcp_out_options mptcp;
 460};
 461
 462static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
 463				struct tcp_sock *tp,
 464				struct tcp_out_options *opts)
 465{
 466#if IS_ENABLED(CONFIG_MPTCP)
 467	if (unlikely(OPTION_MPTCP & opts->options))
 468		mptcp_write_options(th, ptr, tp, &opts->mptcp);
 469#endif
 470}
 471
 472#ifdef CONFIG_CGROUP_BPF
 473static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
 474					enum tcp_synack_type synack_type)
 475{
 476	if (unlikely(!skb))
 477		return BPF_WRITE_HDR_TCP_CURRENT_MSS;
 478
 479	if (unlikely(synack_type == TCP_SYNACK_COOKIE))
 480		return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
 481
 482	return 0;
 483}
 484
 485/* req, syn_skb and synack_type are used when writing synack */
 486static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 487				  struct request_sock *req,
 488				  struct sk_buff *syn_skb,
 489				  enum tcp_synack_type synack_type,
 490				  struct tcp_out_options *opts,
 491				  unsigned int *remaining)
 492{
 493	struct bpf_sock_ops_kern sock_ops;
 494	int err;
 495
 496	if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
 497					   BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
 498	    !*remaining)
 499		return;
 500
 501	/* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
 502
 503	/* init sock_ops */
 504	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 505
 506	sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
 507
 508	if (req) {
 509		/* The listen "sk" cannot be passed here because
 510		 * it is not locked.  It would not make too much
 511		 * sense to do bpf_setsockopt(listen_sk) based
 512		 * on individual connection request also.
 513		 *
 514		 * Thus, "req" is passed here and the cgroup-bpf-progs
 515		 * of the listen "sk" will be run.
 516		 *
 517		 * "req" is also used here for fastopen even the "sk" here is
 518		 * a fullsock "child" sk.  It is to keep the behavior
 519		 * consistent between fastopen and non-fastopen on
 520		 * the bpf programming side.
 521		 */
 522		sock_ops.sk = (struct sock *)req;
 523		sock_ops.syn_skb = syn_skb;
 524	} else {
 525		sock_owned_by_me(sk);
 526
 527		sock_ops.is_fullsock = 1;
 528		sock_ops.sk = sk;
 529	}
 530
 531	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 532	sock_ops.remaining_opt_len = *remaining;
 533	/* tcp_current_mss() does not pass a skb */
 534	if (skb)
 535		bpf_skops_init_skb(&sock_ops, skb, 0);
 536
 537	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 538
 539	if (err || sock_ops.remaining_opt_len == *remaining)
 540		return;
 541
 542	opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
 543	/* round up to 4 bytes */
 544	opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
 545
 546	*remaining -= opts->bpf_opt_len;
 547}
 548
 549static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 550				    struct request_sock *req,
 551				    struct sk_buff *syn_skb,
 552				    enum tcp_synack_type synack_type,
 553				    struct tcp_out_options *opts)
 554{
 555	u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
 556	struct bpf_sock_ops_kern sock_ops;
 557	int err;
 558
 559	if (likely(!max_opt_len))
 560		return;
 561
 562	memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
 563
 564	sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
 565
 566	if (req) {
 567		sock_ops.sk = (struct sock *)req;
 568		sock_ops.syn_skb = syn_skb;
 569	} else {
 570		sock_owned_by_me(sk);
 571
 572		sock_ops.is_fullsock = 1;
 573		sock_ops.sk = sk;
 574	}
 575
 576	sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
 577	sock_ops.remaining_opt_len = max_opt_len;
 578	first_opt_off = tcp_hdrlen(skb) - max_opt_len;
 579	bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
 580
 581	err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
 582
 583	if (err)
 584		nr_written = 0;
 585	else
 586		nr_written = max_opt_len - sock_ops.remaining_opt_len;
 587
 588	if (nr_written < max_opt_len)
 589		memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
 590		       max_opt_len - nr_written);
 591}
 592#else
 593static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
 594				  struct request_sock *req,
 595				  struct sk_buff *syn_skb,
 596				  enum tcp_synack_type synack_type,
 597				  struct tcp_out_options *opts,
 598				  unsigned int *remaining)
 599{
 600}
 601
 602static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
 603				    struct request_sock *req,
 604				    struct sk_buff *syn_skb,
 605				    enum tcp_synack_type synack_type,
 606				    struct tcp_out_options *opts)
 607{
 608}
 609#endif
 610
 611static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
 612				      const struct tcp_request_sock *tcprsk,
 613				      struct tcp_out_options *opts,
 614				      struct tcp_key *key, __be32 *ptr)
 615{
 616#ifdef CONFIG_TCP_AO
 617	u8 maclen = tcp_ao_maclen(key->ao_key);
 618
 619	if (tcprsk) {
 620		u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
 621
 622		*ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
 623			       (tcprsk->ao_keyid << 8) |
 624			       (tcprsk->ao_rcv_next));
 625	} else {
 626		struct tcp_ao_key *rnext_key;
 627		struct tcp_ao_info *ao_info;
 628
 629		ao_info = rcu_dereference_check(tp->ao_info,
 630			lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
 631		rnext_key = READ_ONCE(ao_info->rnext_key);
 632		if (WARN_ON_ONCE(!rnext_key))
 633			return ptr;
 634		*ptr++ = htonl((TCPOPT_AO << 24) |
 635			       (tcp_ao_len(key->ao_key) << 16) |
 636			       (key->ao_key->sndid << 8) |
 637			       (rnext_key->rcvid));
 638	}
 639	opts->hash_location = (__u8 *)ptr;
 640	ptr += maclen / sizeof(*ptr);
 641	if (unlikely(maclen % sizeof(*ptr))) {
 642		memset(ptr, TCPOPT_NOP, sizeof(*ptr));
 643		ptr++;
 644	}
 645#endif
 646	return ptr;
 647}
 648
 649/* Write previously computed TCP options to the packet.
 650 *
 651 * Beware: Something in the Internet is very sensitive to the ordering of
 652 * TCP options, we learned this through the hard way, so be careful here.
 653 * Luckily we can at least blame others for their non-compliance but from
 654 * inter-operability perspective it seems that we're somewhat stuck with
 655 * the ordering which we have been using if we want to keep working with
 656 * those broken things (not that it currently hurts anybody as there isn't
 657 * particular reason why the ordering would need to be changed).
 658 *
 659 * At least SACK_PERM as the first option is known to lead to a disaster
 660 * (but it may well be that other scenarios fail similarly).
 661 */
 662static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
 663			      const struct tcp_request_sock *tcprsk,
 664			      struct tcp_out_options *opts,
 665			      struct tcp_key *key)
 666{
 667	__be32 *ptr = (__be32 *)(th + 1);
 668	u16 options = opts->options;	/* mungable copy */
 669
 670	if (tcp_key_is_md5(key)) {
 671		*ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
 672			       (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
 673		/* overload cookie hash location */
 674		opts->hash_location = (__u8 *)ptr;
 675		ptr += 4;
 676	} else if (tcp_key_is_ao(key)) {
 677		ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
 678	}
 
 679	if (unlikely(opts->mss)) {
 680		*ptr++ = htonl((TCPOPT_MSS << 24) |
 681			       (TCPOLEN_MSS << 16) |
 682			       opts->mss);
 683	}
 684
 685	if (likely(OPTION_TS & options)) {
 686		if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 687			*ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
 688				       (TCPOLEN_SACK_PERM << 16) |
 689				       (TCPOPT_TIMESTAMP << 8) |
 690				       TCPOLEN_TIMESTAMP);
 691			options &= ~OPTION_SACK_ADVERTISE;
 692		} else {
 693			*ptr++ = htonl((TCPOPT_NOP << 24) |
 694				       (TCPOPT_NOP << 16) |
 695				       (TCPOPT_TIMESTAMP << 8) |
 696				       TCPOLEN_TIMESTAMP);
 697		}
 698		*ptr++ = htonl(opts->tsval);
 699		*ptr++ = htonl(opts->tsecr);
 700	}
 701
 702	if (unlikely(OPTION_SACK_ADVERTISE & options)) {
 703		*ptr++ = htonl((TCPOPT_NOP << 24) |
 704			       (TCPOPT_NOP << 16) |
 705			       (TCPOPT_SACK_PERM << 8) |
 706			       TCPOLEN_SACK_PERM);
 707	}
 708
 709	if (unlikely(OPTION_WSCALE & options)) {
 710		*ptr++ = htonl((TCPOPT_NOP << 24) |
 711			       (TCPOPT_WINDOW << 16) |
 712			       (TCPOLEN_WINDOW << 8) |
 713			       opts->ws);
 714	}
 715
 716	if (unlikely(opts->num_sack_blocks)) {
 717		struct tcp_sack_block *sp = tp->rx_opt.dsack ?
 718			tp->duplicate_sack : tp->selective_acks;
 719		int this_sack;
 720
 721		*ptr++ = htonl((TCPOPT_NOP  << 24) |
 722			       (TCPOPT_NOP  << 16) |
 723			       (TCPOPT_SACK <<  8) |
 724			       (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
 725						     TCPOLEN_SACK_PERBLOCK)));
 726
 727		for (this_sack = 0; this_sack < opts->num_sack_blocks;
 728		     ++this_sack) {
 729			*ptr++ = htonl(sp[this_sack].start_seq);
 730			*ptr++ = htonl(sp[this_sack].end_seq);
 731		}
 732
 733		tp->rx_opt.dsack = 0;
 734	}
 735
 736	if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
 737		struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
 738		u8 *p = (u8 *)ptr;
 739		u32 len; /* Fast Open option length */
 740
 741		if (foc->exp) {
 742			len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
 743			*ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
 744				     TCPOPT_FASTOPEN_MAGIC);
 745			p += TCPOLEN_EXP_FASTOPEN_BASE;
 746		} else {
 747			len = TCPOLEN_FASTOPEN_BASE + foc->len;
 748			*p++ = TCPOPT_FASTOPEN;
 749			*p++ = len;
 750		}
 751
 752		memcpy(p, foc->val, foc->len);
 753		if ((len & 3) == 2) {
 754			p[foc->len] = TCPOPT_NOP;
 755			p[foc->len + 1] = TCPOPT_NOP;
 756		}
 757		ptr += (len + 3) >> 2;
 758	}
 759
 760	smc_options_write(ptr, &options);
 761
 762	mptcp_options_write(th, ptr, tp, opts);
 763}
 764
 765static void smc_set_option(const struct tcp_sock *tp,
 766			   struct tcp_out_options *opts,
 767			   unsigned int *remaining)
 768{
 769#if IS_ENABLED(CONFIG_SMC)
 770	if (static_branch_unlikely(&tcp_have_smc)) {
 771		if (tp->syn_smc) {
 772			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 773				opts->options |= OPTION_SMC;
 774				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 775			}
 776		}
 777	}
 778#endif
 779}
 780
 781static void smc_set_option_cond(const struct tcp_sock *tp,
 782				const struct inet_request_sock *ireq,
 783				struct tcp_out_options *opts,
 784				unsigned int *remaining)
 785{
 786#if IS_ENABLED(CONFIG_SMC)
 787	if (static_branch_unlikely(&tcp_have_smc)) {
 788		if (tp->syn_smc && ireq->smc_ok) {
 789			if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
 790				opts->options |= OPTION_SMC;
 791				*remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
 792			}
 793		}
 794	}
 795#endif
 796}
 797
 798static void mptcp_set_option_cond(const struct request_sock *req,
 799				  struct tcp_out_options *opts,
 800				  unsigned int *remaining)
 801{
 802	if (rsk_is_mptcp(req)) {
 803		unsigned int size;
 804
 805		if (mptcp_synack_options(req, &size, &opts->mptcp)) {
 806			if (*remaining >= size) {
 807				opts->options |= OPTION_MPTCP;
 808				*remaining -= size;
 809			}
 810		}
 811	}
 812}
 813
 814/* Compute TCP options for SYN packets. This is not the final
 815 * network wire format yet.
 816 */
 817static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
 818				struct tcp_out_options *opts,
 819				struct tcp_key *key)
 820{
 821	struct tcp_sock *tp = tcp_sk(sk);
 822	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 823	struct tcp_fastopen_request *fastopen = tp->fastopen_req;
 824	bool timestamps;
 825
 826	/* Better than switch (key.type) as it has static branches */
 827	if (tcp_key_is_md5(key)) {
 828		timestamps = false;
 829		opts->options |= OPTION_MD5;
 830		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 831	} else {
 832		timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
 833		if (tcp_key_is_ao(key)) {
 834			opts->options |= OPTION_AO;
 835			remaining -= tcp_ao_len_aligned(key->ao_key);
 836		}
 837	}
 
 838
 839	/* We always get an MSS option.  The option bytes which will be seen in
 840	 * normal data packets should timestamps be used, must be in the MSS
 841	 * advertised.  But we subtract them from tp->mss_cache so that
 842	 * calculations in tcp_sendmsg are simpler etc.  So account for this
 843	 * fact here if necessary.  If we don't do this correctly, as a
 844	 * receiver we won't recognize data packets as being full sized when we
 845	 * should, and thus we won't abide by the delayed ACK rules correctly.
 846	 * SACKs don't matter, we never delay an ACK when we have any of those
 847	 * going out.  */
 848	opts->mss = tcp_advertise_mss(sk);
 849	remaining -= TCPOLEN_MSS_ALIGNED;
 850
 851	if (likely(timestamps)) {
 852		opts->options |= OPTION_TS;
 853		opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset;
 854		opts->tsecr = tp->rx_opt.ts_recent;
 855		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 856	}
 857	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
 858		opts->ws = tp->rx_opt.rcv_wscale;
 859		opts->options |= OPTION_WSCALE;
 860		remaining -= TCPOLEN_WSCALE_ALIGNED;
 861	}
 862	if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
 863		opts->options |= OPTION_SACK_ADVERTISE;
 864		if (unlikely(!(OPTION_TS & opts->options)))
 865			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 866	}
 867
 868	if (fastopen && fastopen->cookie.len >= 0) {
 869		u32 need = fastopen->cookie.len;
 870
 871		need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 872					       TCPOLEN_FASTOPEN_BASE;
 873		need = (need + 3) & ~3U;  /* Align to 32 bits */
 874		if (remaining >= need) {
 875			opts->options |= OPTION_FAST_OPEN_COOKIE;
 876			opts->fastopen_cookie = &fastopen->cookie;
 877			remaining -= need;
 878			tp->syn_fastopen = 1;
 879			tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
 880		}
 881	}
 882
 883	smc_set_option(tp, opts, &remaining);
 884
 885	if (sk_is_mptcp(sk)) {
 886		unsigned int size;
 887
 888		if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
 889			if (remaining >= size) {
 890				opts->options |= OPTION_MPTCP;
 891				remaining -= size;
 892			}
 893		}
 894	}
 895
 896	bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
 897
 898	return MAX_TCP_OPTION_SPACE - remaining;
 899}
 900
 901/* Set up TCP options for SYN-ACKs. */
 902static unsigned int tcp_synack_options(const struct sock *sk,
 903				       struct request_sock *req,
 904				       unsigned int mss, struct sk_buff *skb,
 905				       struct tcp_out_options *opts,
 906				       const struct tcp_key *key,
 907				       struct tcp_fastopen_cookie *foc,
 908				       enum tcp_synack_type synack_type,
 909				       struct sk_buff *syn_skb)
 910{
 911	struct inet_request_sock *ireq = inet_rsk(req);
 912	unsigned int remaining = MAX_TCP_OPTION_SPACE;
 913
 914	if (tcp_key_is_md5(key)) {
 
 915		opts->options |= OPTION_MD5;
 916		remaining -= TCPOLEN_MD5SIG_ALIGNED;
 917
 918		/* We can't fit any SACK blocks in a packet with MD5 + TS
 919		 * options. There was discussion about disabling SACK
 920		 * rather than TS in order to fit in better with old,
 921		 * buggy kernels, but that was deemed to be unnecessary.
 922		 */
 923		if (synack_type != TCP_SYNACK_COOKIE)
 924			ireq->tstamp_ok &= !ireq->sack_ok;
 925	} else if (tcp_key_is_ao(key)) {
 926		opts->options |= OPTION_AO;
 927		remaining -= tcp_ao_len_aligned(key->ao_key);
 928		ireq->tstamp_ok &= !ireq->sack_ok;
 929	}
 
 930
 931	/* We always send an MSS option. */
 932	opts->mss = mss;
 933	remaining -= TCPOLEN_MSS_ALIGNED;
 934
 935	if (likely(ireq->wscale_ok)) {
 936		opts->ws = ireq->rcv_wscale;
 937		opts->options |= OPTION_WSCALE;
 938		remaining -= TCPOLEN_WSCALE_ALIGNED;
 939	}
 940	if (likely(ireq->tstamp_ok)) {
 941		opts->options |= OPTION_TS;
 942		opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) +
 943			      tcp_rsk(req)->ts_off;
 944		opts->tsecr = READ_ONCE(req->ts_recent);
 945		remaining -= TCPOLEN_TSTAMP_ALIGNED;
 946	}
 947	if (likely(ireq->sack_ok)) {
 948		opts->options |= OPTION_SACK_ADVERTISE;
 949		if (unlikely(!ireq->tstamp_ok))
 950			remaining -= TCPOLEN_SACKPERM_ALIGNED;
 951	}
 952	if (foc != NULL && foc->len >= 0) {
 953		u32 need = foc->len;
 954
 955		need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
 956				   TCPOLEN_FASTOPEN_BASE;
 957		need = (need + 3) & ~3U;  /* Align to 32 bits */
 958		if (remaining >= need) {
 959			opts->options |= OPTION_FAST_OPEN_COOKIE;
 960			opts->fastopen_cookie = foc;
 961			remaining -= need;
 962		}
 963	}
 964
 965	mptcp_set_option_cond(req, opts, &remaining);
 966
 967	smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
 968
 969	bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
 970			      synack_type, opts, &remaining);
 971
 972	return MAX_TCP_OPTION_SPACE - remaining;
 973}
 974
 975/* Compute TCP options for ESTABLISHED sockets. This is not the
 976 * final wire format yet.
 977 */
 978static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
 979					struct tcp_out_options *opts,
 980					struct tcp_key *key)
 981{
 982	struct tcp_sock *tp = tcp_sk(sk);
 983	unsigned int size = 0;
 984	unsigned int eff_sacks;
 985
 986	opts->options = 0;
 987
 988	/* Better than switch (key.type) as it has static branches */
 989	if (tcp_key_is_md5(key)) {
 990		opts->options |= OPTION_MD5;
 991		size += TCPOLEN_MD5SIG_ALIGNED;
 992	} else if (tcp_key_is_ao(key)) {
 993		opts->options |= OPTION_AO;
 994		size += tcp_ao_len_aligned(key->ao_key);
 
 
 995	}
 
 996
 997	if (likely(tp->rx_opt.tstamp_ok)) {
 998		opts->options |= OPTION_TS;
 999		opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) +
1000				tp->tsoffset : 0;
1001		opts->tsecr = tp->rx_opt.ts_recent;
1002		size += TCPOLEN_TSTAMP_ALIGNED;
1003	}
1004
1005	/* MPTCP options have precedence over SACK for the limited TCP
1006	 * option space because a MPTCP connection would be forced to
1007	 * fall back to regular TCP if a required multipath option is
1008	 * missing. SACK still gets a chance to use whatever space is
1009	 * left.
1010	 */
1011	if (sk_is_mptcp(sk)) {
1012		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1013		unsigned int opt_size = 0;
1014
1015		if (mptcp_established_options(sk, skb, &opt_size, remaining,
1016					      &opts->mptcp)) {
1017			opts->options |= OPTION_MPTCP;
1018			size += opt_size;
1019		}
1020	}
1021
1022	eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1023	if (unlikely(eff_sacks)) {
1024		const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1025		if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
1026					 TCPOLEN_SACK_PERBLOCK))
1027			return size;
1028
1029		opts->num_sack_blocks =
1030			min_t(unsigned int, eff_sacks,
1031			      (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1032			      TCPOLEN_SACK_PERBLOCK);
1033
1034		size += TCPOLEN_SACK_BASE_ALIGNED +
1035			opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1036	}
1037
1038	if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1039					    BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1040		unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1041
1042		bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1043
1044		size = MAX_TCP_OPTION_SPACE - remaining;
1045	}
1046
1047	return size;
1048}
1049
1050
1051/* TCP SMALL QUEUES (TSQ)
1052 *
1053 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1054 * to reduce RTT and bufferbloat.
1055 * We do this using a special skb destructor (tcp_wfree).
1056 *
1057 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1058 * needs to be reallocated in a driver.
1059 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1060 *
1061 * Since transmit from skb destructor is forbidden, we use a tasklet
1062 * to process all sockets that eventually need to send more skbs.
1063 * We use one tasklet per cpu, with its own queue of sockets.
1064 */
1065struct tsq_tasklet {
1066	struct tasklet_struct	tasklet;
1067	struct list_head	head; /* queue of tcp sockets */
1068};
1069static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1070
1071static void tcp_tsq_write(struct sock *sk)
1072{
1073	if ((1 << sk->sk_state) &
1074	    (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1075	     TCPF_CLOSE_WAIT  | TCPF_LAST_ACK)) {
1076		struct tcp_sock *tp = tcp_sk(sk);
1077
1078		if (tp->lost_out > tp->retrans_out &&
1079		    tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1080			tcp_mstamp_refresh(tp);
1081			tcp_xmit_retransmit_queue(sk);
1082		}
1083
1084		tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1085			       0, GFP_ATOMIC);
1086	}
1087}
1088
1089static void tcp_tsq_handler(struct sock *sk)
1090{
1091	bh_lock_sock(sk);
1092	if (!sock_owned_by_user(sk))
1093		tcp_tsq_write(sk);
1094	else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1095		sock_hold(sk);
1096	bh_unlock_sock(sk);
1097}
1098/*
1099 * One tasklet per cpu tries to send more skbs.
1100 * We run in tasklet context but need to disable irqs when
1101 * transferring tsq->head because tcp_wfree() might
1102 * interrupt us (non NAPI drivers)
1103 */
1104static void tcp_tasklet_func(struct tasklet_struct *t)
1105{
1106	struct tsq_tasklet *tsq = from_tasklet(tsq,  t, tasklet);
1107	LIST_HEAD(list);
1108	unsigned long flags;
1109	struct list_head *q, *n;
1110	struct tcp_sock *tp;
1111	struct sock *sk;
1112
1113	local_irq_save(flags);
1114	list_splice_init(&tsq->head, &list);
1115	local_irq_restore(flags);
1116
1117	list_for_each_safe(q, n, &list) {
1118		tp = list_entry(q, struct tcp_sock, tsq_node);
1119		list_del(&tp->tsq_node);
1120
1121		sk = (struct sock *)tp;
1122		smp_mb__before_atomic();
1123		clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1124
1125		tcp_tsq_handler(sk);
1126		sk_free(sk);
1127	}
1128}
1129
1130#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED |		\
1131			  TCPF_WRITE_TIMER_DEFERRED |	\
1132			  TCPF_DELACK_TIMER_DEFERRED |	\
1133			  TCPF_MTU_REDUCED_DEFERRED |	\
1134			  TCPF_ACK_DEFERRED)
1135/**
1136 * tcp_release_cb - tcp release_sock() callback
1137 * @sk: socket
1138 *
1139 * called from release_sock() to perform protocol dependent
1140 * actions before socket release.
1141 */
1142void tcp_release_cb(struct sock *sk)
1143{
1144	unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1145	unsigned long nflags;
1146
1147	/* perform an atomic operation only if at least one flag is set */
1148	do {
1149		if (!(flags & TCP_DEFERRED_ALL))
1150			return;
1151		nflags = flags & ~TCP_DEFERRED_ALL;
1152	} while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1153
1154	if (flags & TCPF_TSQ_DEFERRED) {
1155		tcp_tsq_write(sk);
1156		__sock_put(sk);
1157	}
 
 
 
 
 
 
 
 
 
 
1158
1159	if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1160		tcp_write_timer_handler(sk);
1161		__sock_put(sk);
1162	}
1163	if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1164		tcp_delack_timer_handler(sk);
1165		__sock_put(sk);
1166	}
1167	if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1168		inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1169		__sock_put(sk);
1170	}
1171	if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1172		tcp_send_ack(sk);
1173}
1174EXPORT_SYMBOL(tcp_release_cb);
1175
1176void __init tcp_tasklet_init(void)
1177{
1178	int i;
1179
1180	for_each_possible_cpu(i) {
1181		struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1182
1183		INIT_LIST_HEAD(&tsq->head);
1184		tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1185	}
1186}
1187
1188/*
1189 * Write buffer destructor automatically called from kfree_skb.
1190 * We can't xmit new skbs from this context, as we might already
1191 * hold qdisc lock.
1192 */
1193void tcp_wfree(struct sk_buff *skb)
1194{
1195	struct sock *sk = skb->sk;
1196	struct tcp_sock *tp = tcp_sk(sk);
1197	unsigned long flags, nval, oval;
1198	struct tsq_tasklet *tsq;
1199	bool empty;
1200
1201	/* Keep one reference on sk_wmem_alloc.
1202	 * Will be released by sk_free() from here or tcp_tasklet_func()
1203	 */
1204	WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1205
1206	/* If this softirq is serviced by ksoftirqd, we are likely under stress.
1207	 * Wait until our queues (qdisc + devices) are drained.
1208	 * This gives :
1209	 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1210	 * - chance for incoming ACK (processed by another cpu maybe)
1211	 *   to migrate this flow (skb->ooo_okay will be eventually set)
1212	 */
1213	if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1214		goto out;
1215
1216	oval = smp_load_acquire(&sk->sk_tsq_flags);
1217	do {
1218		if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1219			goto out;
1220
1221		nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1222	} while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1223
1224	/* queue this socket to tasklet queue */
1225	local_irq_save(flags);
1226	tsq = this_cpu_ptr(&tsq_tasklet);
1227	empty = list_empty(&tsq->head);
1228	list_add(&tp->tsq_node, &tsq->head);
1229	if (empty)
1230		tasklet_schedule(&tsq->tasklet);
1231	local_irq_restore(flags);
1232	return;
1233out:
1234	sk_free(sk);
1235}
1236
1237/* Note: Called under soft irq.
1238 * We can call TCP stack right away, unless socket is owned by user.
1239 */
1240enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1241{
1242	struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1243	struct sock *sk = (struct sock *)tp;
1244
1245	tcp_tsq_handler(sk);
1246	sock_put(sk);
1247
1248	return HRTIMER_NORESTART;
1249}
1250
1251static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1252				      u64 prior_wstamp)
1253{
1254	struct tcp_sock *tp = tcp_sk(sk);
1255
1256	if (sk->sk_pacing_status != SK_PACING_NONE) {
1257		unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1258
1259		/* Original sch_fq does not pace first 10 MSS
1260		 * Note that tp->data_segs_out overflows after 2^32 packets,
1261		 * this is a minor annoyance.
1262		 */
1263		if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1264			u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1265			u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1266
1267			/* take into account OS jitter */
1268			len_ns -= min_t(u64, len_ns / 2, credit);
1269			tp->tcp_wstamp_ns += len_ns;
1270		}
1271	}
1272	list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1273}
1274
1275INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1276INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1277INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1278
1279/* This routine actually transmits TCP packets queued in by
1280 * tcp_do_sendmsg().  This is used by both the initial
1281 * transmission and possible later retransmissions.
1282 * All SKB's seen here are completely headerless.  It is our
1283 * job to build the TCP header, and pass the packet down to
1284 * IP so it can do the same plus pass the packet off to the
1285 * device.
1286 *
1287 * We are working here with either a clone of the original
1288 * SKB, or a fresh unique copy made by the retransmit engine.
1289 */
1290static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1291			      int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1292{
1293	const struct inet_connection_sock *icsk = inet_csk(sk);
1294	struct inet_sock *inet;
1295	struct tcp_sock *tp;
1296	struct tcp_skb_cb *tcb;
1297	struct tcp_out_options opts;
1298	unsigned int tcp_options_size, tcp_header_size;
1299	struct sk_buff *oskb = NULL;
1300	struct tcp_key key;
1301	struct tcphdr *th;
1302	u64 prior_wstamp;
1303	int err;
1304
1305	BUG_ON(!skb || !tcp_skb_pcount(skb));
1306	tp = tcp_sk(sk);
1307	prior_wstamp = tp->tcp_wstamp_ns;
1308	tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1309	skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
1310	if (clone_it) {
1311		oskb = skb;
1312
1313		tcp_skb_tsorted_save(oskb) {
1314			if (unlikely(skb_cloned(oskb)))
1315				skb = pskb_copy(oskb, gfp_mask);
1316			else
1317				skb = skb_clone(oskb, gfp_mask);
1318		} tcp_skb_tsorted_restore(oskb);
1319
1320		if (unlikely(!skb))
1321			return -ENOBUFS;
1322		/* retransmit skbs might have a non zero value in skb->dev
1323		 * because skb->dev is aliased with skb->rbnode.rb_left
1324		 */
1325		skb->dev = NULL;
1326	}
1327
1328	inet = inet_sk(sk);
1329	tcb = TCP_SKB_CB(skb);
1330	memset(&opts, 0, sizeof(opts));
1331
1332	tcp_get_current_key(sk, &key);
1333	if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1334		tcp_options_size = tcp_syn_options(sk, skb, &opts, &key);
1335	} else {
1336		tcp_options_size = tcp_established_options(sk, skb, &opts, &key);
 
1337		/* Force a PSH flag on all (GSO) packets to expedite GRO flush
1338		 * at receiver : This slightly improve GRO performance.
1339		 * Note that we do not force the PSH flag for non GSO packets,
1340		 * because they might be sent under high congestion events,
1341		 * and in this case it is better to delay the delivery of 1-MSS
1342		 * packets and thus the corresponding ACK packet that would
1343		 * release the following packet.
1344		 */
1345		if (tcp_skb_pcount(skb) > 1)
1346			tcb->tcp_flags |= TCPHDR_PSH;
1347	}
1348	tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1349
1350	/* We set skb->ooo_okay to one if this packet can select
1351	 * a different TX queue than prior packets of this flow,
1352	 * to avoid self inflicted reorders.
1353	 * The 'other' queue decision is based on current cpu number
1354	 * if XPS is enabled, or sk->sk_txhash otherwise.
1355	 * We can switch to another (and better) queue if:
1356	 * 1) No packet with payload is in qdisc/device queues.
1357	 *    Delays in TX completion can defeat the test
1358	 *    even if packets were already sent.
1359	 * 2) Or rtx queue is empty.
1360	 *    This mitigates above case if ACK packets for
1361	 *    all prior packets were already processed.
1362	 */
1363	skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1364			tcp_rtx_queue_empty(sk);
1365
1366	/* If we had to use memory reserve to allocate this skb,
1367	 * this might cause drops if packet is looped back :
1368	 * Other socket might not have SOCK_MEMALLOC.
1369	 * Packets not looped back do not care about pfmemalloc.
1370	 */
1371	skb->pfmemalloc = 0;
1372
1373	skb_push(skb, tcp_header_size);
1374	skb_reset_transport_header(skb);
1375
1376	skb_orphan(skb);
1377	skb->sk = sk;
1378	skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1379	refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1380
1381	skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1382
1383	/* Build TCP header and checksum it. */
1384	th = (struct tcphdr *)skb->data;
1385	th->source		= inet->inet_sport;
1386	th->dest		= inet->inet_dport;
1387	th->seq			= htonl(tcb->seq);
1388	th->ack_seq		= htonl(rcv_nxt);
1389	*(((__be16 *)th) + 6)	= htons(((tcp_header_size >> 2) << 12) |
1390					tcb->tcp_flags);
1391
1392	th->check		= 0;
1393	th->urg_ptr		= 0;
1394
1395	/* The urg_mode check is necessary during a below snd_una win probe */
1396	if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1397		if (before(tp->snd_up, tcb->seq + 0x10000)) {
1398			th->urg_ptr = htons(tp->snd_up - tcb->seq);
1399			th->urg = 1;
1400		} else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1401			th->urg_ptr = htons(0xFFFF);
1402			th->urg = 1;
1403		}
1404	}
1405
1406	skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1407	if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1408		th->window      = htons(tcp_select_window(sk));
1409		tcp_ecn_send(sk, skb, th, tcp_header_size);
1410	} else {
1411		/* RFC1323: The window in SYN & SYN/ACK segments
1412		 * is never scaled.
1413		 */
1414		th->window	= htons(min(tp->rcv_wnd, 65535U));
1415	}
1416
1417	tcp_options_write(th, tp, NULL, &opts, &key);
1418
1419	if (tcp_key_is_md5(&key)) {
1420#ifdef CONFIG_TCP_MD5SIG
1421		/* Calculate the MD5 hash, as we have all we need now */
 
1422		sk_gso_disable(sk);
1423		tp->af_specific->calc_md5_hash(opts.hash_location,
1424					       key.md5_key, sk, skb);
 
1425#endif
1426	} else if (tcp_key_is_ao(&key)) {
1427		int err;
1428
1429		err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th,
1430					  opts.hash_location);
1431		if (err) {
1432			kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED);
1433			return -ENOMEM;
1434		}
1435	}
1436
1437	/* BPF prog is the last one writing header option */
1438	bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1439
1440	INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1441			   tcp_v6_send_check, tcp_v4_send_check,
1442			   sk, skb);
1443
1444	if (likely(tcb->tcp_flags & TCPHDR_ACK))
1445		tcp_event_ack_sent(sk, rcv_nxt);
1446
1447	if (skb->len != tcp_header_size) {
1448		tcp_event_data_sent(tp, sk);
1449		tp->data_segs_out += tcp_skb_pcount(skb);
1450		tp->bytes_sent += skb->len - tcp_header_size;
1451	}
1452
1453	if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1454		TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1455			      tcp_skb_pcount(skb));
1456
1457	tp->segs_out += tcp_skb_pcount(skb);
1458	skb_set_hash_from_sk(skb, sk);
1459	/* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1460	skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1461	skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1462
1463	/* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1464
1465	/* Cleanup our debris for IP stacks */
1466	memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1467			       sizeof(struct inet6_skb_parm)));
1468
1469	tcp_add_tx_delay(skb, tp);
1470
1471	err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1472				 inet6_csk_xmit, ip_queue_xmit,
1473				 sk, skb, &inet->cork.fl);
1474
1475	if (unlikely(err > 0)) {
1476		tcp_enter_cwr(sk);
1477		err = net_xmit_eval(err);
1478	}
1479	if (!err && oskb) {
1480		tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1481		tcp_rate_skb_sent(sk, oskb);
1482	}
1483	return err;
1484}
1485
1486static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1487			    gfp_t gfp_mask)
1488{
1489	return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1490				  tcp_sk(sk)->rcv_nxt);
1491}
1492
1493/* This routine just queues the buffer for sending.
1494 *
1495 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1496 * otherwise socket can stall.
1497 */
1498static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1499{
1500	struct tcp_sock *tp = tcp_sk(sk);
1501
1502	/* Advance write_seq and place onto the write_queue. */
1503	WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1504	__skb_header_release(skb);
1505	tcp_add_write_queue_tail(sk, skb);
1506	sk_wmem_queued_add(sk, skb->truesize);
1507	sk_mem_charge(sk, skb->truesize);
1508}
1509
1510/* Initialize TSO segments for a packet. */
1511static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1512{
1513	int tso_segs;
1514
1515	if (skb->len <= mss_now) {
1516		/* Avoid the costly divide in the normal
1517		 * non-TSO case.
1518		 */
 
1519		TCP_SKB_CB(skb)->tcp_gso_size = 0;
1520		tcp_skb_pcount_set(skb, 1);
1521		return 1;
 
1522	}
1523	TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1524	tso_segs = DIV_ROUND_UP(skb->len, mss_now);
1525	tcp_skb_pcount_set(skb, tso_segs);
1526	return tso_segs;
1527}
1528
1529/* Pcount in the middle of the write queue got changed, we need to do various
1530 * tweaks to fix counters
1531 */
1532static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1533{
1534	struct tcp_sock *tp = tcp_sk(sk);
1535
1536	tp->packets_out -= decr;
1537
1538	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1539		tp->sacked_out -= decr;
1540	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1541		tp->retrans_out -= decr;
1542	if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1543		tp->lost_out -= decr;
1544
1545	/* Reno case is special. Sigh... */
1546	if (tcp_is_reno(tp) && decr > 0)
1547		tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1548
1549	if (tp->lost_skb_hint &&
1550	    before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1551	    (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1552		tp->lost_cnt_hint -= decr;
1553
1554	tcp_verify_left_out(tp);
1555}
1556
1557static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1558{
1559	return TCP_SKB_CB(skb)->txstamp_ack ||
1560		(skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1561}
1562
1563static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1564{
1565	struct skb_shared_info *shinfo = skb_shinfo(skb);
1566
1567	if (unlikely(tcp_has_tx_tstamp(skb)) &&
1568	    !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1569		struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1570		u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1571
1572		shinfo->tx_flags &= ~tsflags;
1573		shinfo2->tx_flags |= tsflags;
1574		swap(shinfo->tskey, shinfo2->tskey);
1575		TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1576		TCP_SKB_CB(skb)->txstamp_ack = 0;
1577	}
1578}
1579
1580static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1581{
1582	TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1583	TCP_SKB_CB(skb)->eor = 0;
1584}
1585
1586/* Insert buff after skb on the write or rtx queue of sk.  */
1587static void tcp_insert_write_queue_after(struct sk_buff *skb,
1588					 struct sk_buff *buff,
1589					 struct sock *sk,
1590					 enum tcp_queue tcp_queue)
1591{
1592	if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1593		__skb_queue_after(&sk->sk_write_queue, skb, buff);
1594	else
1595		tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1596}
1597
1598/* Function to create two new TCP segments.  Shrinks the given segment
1599 * to the specified size and appends a new segment with the rest of the
1600 * packet to the list.  This won't be called frequently, I hope.
1601 * Remember, these are still headerless SKBs at this point.
1602 */
1603int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1604		 struct sk_buff *skb, u32 len,
1605		 unsigned int mss_now, gfp_t gfp)
1606{
1607	struct tcp_sock *tp = tcp_sk(sk);
1608	struct sk_buff *buff;
1609	int old_factor;
1610	long limit;
1611	int nlen;
1612	u8 flags;
1613
1614	if (WARN_ON(len > skb->len))
1615		return -EINVAL;
1616
1617	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
 
 
1618
1619	/* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1620	 * We need some allowance to not penalize applications setting small
1621	 * SO_SNDBUF values.
1622	 * Also allow first and last skb in retransmit queue to be split.
1623	 */
1624	limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1625	if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1626		     tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1627		     skb != tcp_rtx_queue_head(sk) &&
1628		     skb != tcp_rtx_queue_tail(sk))) {
1629		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1630		return -ENOMEM;
1631	}
1632
1633	if (skb_unclone_keeptruesize(skb, gfp))
1634		return -ENOMEM;
1635
1636	/* Get a new skb... force flag on. */
1637	buff = tcp_stream_alloc_skb(sk, gfp, true);
1638	if (!buff)
1639		return -ENOMEM; /* We'll just try again later. */
1640	skb_copy_decrypted(buff, skb);
1641	mptcp_skb_ext_copy(buff, skb);
1642
1643	sk_wmem_queued_add(sk, buff->truesize);
1644	sk_mem_charge(sk, buff->truesize);
1645	nlen = skb->len - len;
1646	buff->truesize += nlen;
1647	skb->truesize -= nlen;
1648
1649	/* Correct the sequence numbers. */
1650	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1651	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1652	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1653
1654	/* PSH and FIN should only be set in the second packet. */
1655	flags = TCP_SKB_CB(skb)->tcp_flags;
1656	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1657	TCP_SKB_CB(buff)->tcp_flags = flags;
1658	TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1659	tcp_skb_fragment_eor(skb, buff);
1660
1661	skb_split(skb, buff, len);
1662
1663	skb_set_delivery_time(buff, skb->tstamp, SKB_CLOCK_MONOTONIC);
1664	tcp_fragment_tstamp(skb, buff);
1665
1666	old_factor = tcp_skb_pcount(skb);
1667
1668	/* Fix up tso_factor for both original and new SKB.  */
1669	tcp_set_skb_tso_segs(skb, mss_now);
1670	tcp_set_skb_tso_segs(buff, mss_now);
1671
1672	/* Update delivered info for the new segment */
1673	TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1674
1675	/* If this packet has been sent out already, we must
1676	 * adjust the various packet counters.
1677	 */
1678	if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1679		int diff = old_factor - tcp_skb_pcount(skb) -
1680			tcp_skb_pcount(buff);
1681
1682		if (diff)
1683			tcp_adjust_pcount(sk, skb, diff);
1684	}
1685
1686	/* Link BUFF into the send queue. */
1687	__skb_header_release(buff);
1688	tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1689	if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1690		list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1691
1692	return 0;
1693}
1694
1695/* This is similar to __pskb_pull_tail(). The difference is that pulled
1696 * data is not copied, but immediately discarded.
1697 */
1698static int __pskb_trim_head(struct sk_buff *skb, int len)
1699{
1700	struct skb_shared_info *shinfo;
1701	int i, k, eat;
1702
1703	DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
 
 
 
 
 
 
1704	eat = len;
1705	k = 0;
1706	shinfo = skb_shinfo(skb);
1707	for (i = 0; i < shinfo->nr_frags; i++) {
1708		int size = skb_frag_size(&shinfo->frags[i]);
1709
1710		if (size <= eat) {
1711			skb_frag_unref(skb, i);
1712			eat -= size;
1713		} else {
1714			shinfo->frags[k] = shinfo->frags[i];
1715			if (eat) {
1716				skb_frag_off_add(&shinfo->frags[k], eat);
1717				skb_frag_size_sub(&shinfo->frags[k], eat);
1718				eat = 0;
1719			}
1720			k++;
1721		}
1722	}
1723	shinfo->nr_frags = k;
1724
1725	skb->data_len -= len;
1726	skb->len = skb->data_len;
1727	return len;
1728}
1729
1730/* Remove acked data from a packet in the transmit queue. */
1731int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1732{
1733	u32 delta_truesize;
1734
1735	if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1736		return -ENOMEM;
1737
1738	delta_truesize = __pskb_trim_head(skb, len);
1739
1740	TCP_SKB_CB(skb)->seq += len;
1741
1742	skb->truesize	   -= delta_truesize;
1743	sk_wmem_queued_add(sk, -delta_truesize);
1744	if (!skb_zcopy_pure(skb))
1745		sk_mem_uncharge(sk, delta_truesize);
 
 
1746
1747	/* Any change of skb->len requires recalculation of tso factor. */
1748	if (tcp_skb_pcount(skb) > 1)
1749		tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1750
1751	return 0;
1752}
1753
1754/* Calculate MSS not accounting any TCP options.  */
1755static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1756{
1757	const struct tcp_sock *tp = tcp_sk(sk);
1758	const struct inet_connection_sock *icsk = inet_csk(sk);
1759	int mss_now;
1760
1761	/* Calculate base mss without TCP options:
1762	   It is MMS_S - sizeof(tcphdr) of rfc1122
1763	 */
1764	mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1765
 
 
 
 
 
 
 
 
1766	/* Clamp it (mss_clamp does not include tcp options) */
1767	if (mss_now > tp->rx_opt.mss_clamp)
1768		mss_now = tp->rx_opt.mss_clamp;
1769
1770	/* Now subtract optional transport overhead */
1771	mss_now -= icsk->icsk_ext_hdr_len;
1772
1773	/* Then reserve room for full set of TCP options and 8 bytes of data */
1774	mss_now = max(mss_now,
1775		      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1776	return mss_now;
1777}
1778
1779/* Calculate MSS. Not accounting for SACKs here.  */
1780int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1781{
1782	/* Subtract TCP options size, not including SACKs */
1783	return __tcp_mtu_to_mss(sk, pmtu) -
1784	       (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1785}
1786EXPORT_SYMBOL(tcp_mtu_to_mss);
1787
1788/* Inverse of above */
1789int tcp_mss_to_mtu(struct sock *sk, int mss)
1790{
1791	const struct tcp_sock *tp = tcp_sk(sk);
1792	const struct inet_connection_sock *icsk = inet_csk(sk);
 
1793
1794	return mss +
1795	      tp->tcp_header_len +
1796	      icsk->icsk_ext_hdr_len +
1797	      icsk->icsk_af_ops->net_header_len;
 
 
 
 
 
 
 
 
 
1798}
1799EXPORT_SYMBOL(tcp_mss_to_mtu);
1800
1801/* MTU probing init per socket */
1802void tcp_mtup_init(struct sock *sk)
1803{
1804	struct tcp_sock *tp = tcp_sk(sk);
1805	struct inet_connection_sock *icsk = inet_csk(sk);
1806	struct net *net = sock_net(sk);
1807
1808	icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1809	icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1810			       icsk->icsk_af_ops->net_header_len;
1811	icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1812	icsk->icsk_mtup.probe_size = 0;
1813	if (icsk->icsk_mtup.enabled)
1814		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1815}
1816EXPORT_SYMBOL(tcp_mtup_init);
1817
1818/* This function synchronize snd mss to current pmtu/exthdr set.
1819
1820   tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1821   for TCP options, but includes only bare TCP header.
1822
1823   tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1824   It is minimum of user_mss and mss received with SYN.
1825   It also does not include TCP options.
1826
1827   inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1828
1829   tp->mss_cache is current effective sending mss, including
1830   all tcp options except for SACKs. It is evaluated,
1831   taking into account current pmtu, but never exceeds
1832   tp->rx_opt.mss_clamp.
1833
1834   NOTE1. rfc1122 clearly states that advertised MSS
1835   DOES NOT include either tcp or ip options.
1836
1837   NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1838   are READ ONLY outside this function.		--ANK (980731)
1839 */
1840unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1841{
1842	struct tcp_sock *tp = tcp_sk(sk);
1843	struct inet_connection_sock *icsk = inet_csk(sk);
1844	int mss_now;
1845
1846	if (icsk->icsk_mtup.search_high > pmtu)
1847		icsk->icsk_mtup.search_high = pmtu;
1848
1849	mss_now = tcp_mtu_to_mss(sk, pmtu);
1850	mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1851
1852	/* And store cached results */
1853	icsk->icsk_pmtu_cookie = pmtu;
1854	if (icsk->icsk_mtup.enabled)
1855		mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1856	tp->mss_cache = mss_now;
1857
1858	return mss_now;
1859}
1860EXPORT_SYMBOL(tcp_sync_mss);
1861
1862/* Compute the current effective MSS, taking SACKs and IP options,
1863 * and even PMTU discovery events into account.
1864 */
1865unsigned int tcp_current_mss(struct sock *sk)
1866{
1867	const struct tcp_sock *tp = tcp_sk(sk);
1868	const struct dst_entry *dst = __sk_dst_get(sk);
1869	u32 mss_now;
1870	unsigned int header_len;
1871	struct tcp_out_options opts;
1872	struct tcp_key key;
1873
1874	mss_now = tp->mss_cache;
1875
1876	if (dst) {
1877		u32 mtu = dst_mtu(dst);
1878		if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1879			mss_now = tcp_sync_mss(sk, mtu);
1880	}
1881	tcp_get_current_key(sk, &key);
1882	header_len = tcp_established_options(sk, NULL, &opts, &key) +
1883		     sizeof(struct tcphdr);
1884	/* The mss_cache is sized based on tp->tcp_header_len, which assumes
1885	 * some common options. If this is an odd packet (because we have SACK
1886	 * blocks etc) then our calculated header_len will be different, and
1887	 * we have to adjust mss_now correspondingly */
1888	if (header_len != tp->tcp_header_len) {
1889		int delta = (int) header_len - tp->tcp_header_len;
1890		mss_now -= delta;
1891	}
1892
1893	return mss_now;
1894}
1895
1896/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1897 * As additional protections, we do not touch cwnd in retransmission phases,
1898 * and if application hit its sndbuf limit recently.
1899 */
1900static void tcp_cwnd_application_limited(struct sock *sk)
1901{
1902	struct tcp_sock *tp = tcp_sk(sk);
1903
1904	if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1905	    sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1906		/* Limited by application or receiver window. */
1907		u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1908		u32 win_used = max(tp->snd_cwnd_used, init_win);
1909		if (win_used < tcp_snd_cwnd(tp)) {
1910			tp->snd_ssthresh = tcp_current_ssthresh(sk);
1911			tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1912		}
1913		tp->snd_cwnd_used = 0;
1914	}
1915	tp->snd_cwnd_stamp = tcp_jiffies32;
1916}
1917
1918static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1919{
1920	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1921	struct tcp_sock *tp = tcp_sk(sk);
1922
1923	/* Track the strongest available signal of the degree to which the cwnd
1924	 * is fully utilized. If cwnd-limited then remember that fact for the
1925	 * current window. If not cwnd-limited then track the maximum number of
1926	 * outstanding packets in the current window. (If cwnd-limited then we
1927	 * chose to not update tp->max_packets_out to avoid an extra else
1928	 * clause with no functional impact.)
1929	 */
1930	if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1931	    is_cwnd_limited ||
1932	    (!tp->is_cwnd_limited &&
1933	     tp->packets_out > tp->max_packets_out)) {
1934		tp->is_cwnd_limited = is_cwnd_limited;
1935		tp->max_packets_out = tp->packets_out;
1936		tp->cwnd_usage_seq = tp->snd_nxt;
1937	}
1938
1939	if (tcp_is_cwnd_limited(sk)) {
1940		/* Network is feed fully. */
1941		tp->snd_cwnd_used = 0;
1942		tp->snd_cwnd_stamp = tcp_jiffies32;
1943	} else {
1944		/* Network starves. */
1945		if (tp->packets_out > tp->snd_cwnd_used)
1946			tp->snd_cwnd_used = tp->packets_out;
1947
1948		if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1949		    (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1950		    !ca_ops->cong_control)
1951			tcp_cwnd_application_limited(sk);
1952
1953		/* The following conditions together indicate the starvation
1954		 * is caused by insufficient sender buffer:
1955		 * 1) just sent some data (see tcp_write_xmit)
1956		 * 2) not cwnd limited (this else condition)
1957		 * 3) no more data to send (tcp_write_queue_empty())
1958		 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1959		 */
1960		if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1961		    test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1962		    (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1963			tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1964	}
1965}
1966
1967/* Minshall's variant of the Nagle send check. */
1968static bool tcp_minshall_check(const struct tcp_sock *tp)
1969{
1970	return after(tp->snd_sml, tp->snd_una) &&
1971		!after(tp->snd_sml, tp->snd_nxt);
1972}
1973
1974/* Update snd_sml if this skb is under mss
1975 * Note that a TSO packet might end with a sub-mss segment
1976 * The test is really :
1977 * if ((skb->len % mss) != 0)
1978 *        tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1979 * But we can avoid doing the divide again given we already have
1980 *  skb_pcount = skb->len / mss_now
1981 */
1982static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1983				const struct sk_buff *skb)
1984{
1985	if (skb->len < tcp_skb_pcount(skb) * mss_now)
1986		tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1987}
1988
1989/* Return false, if packet can be sent now without violation Nagle's rules:
1990 * 1. It is full sized. (provided by caller in %partial bool)
1991 * 2. Or it contains FIN. (already checked by caller)
1992 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1993 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1994 *    With Minshall's modification: all sent small packets are ACKed.
1995 */
1996static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1997			    int nonagle)
1998{
1999	return partial &&
2000		((nonagle & TCP_NAGLE_CORK) ||
2001		 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
2002}
2003
2004/* Return how many segs we'd like on a TSO packet,
2005 * depending on current pacing rate, and how close the peer is.
2006 *
2007 * Rationale is:
2008 * - For close peers, we rather send bigger packets to reduce
2009 *   cpu costs, because occasional losses will be repaired fast.
2010 * - For long distance/rtt flows, we would like to get ACK clocking
2011 *   with 1 ACK per ms.
2012 *
2013 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2014 * in bigger TSO bursts. We we cut the RTT-based allowance in half
2015 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2016 * is below 1500 bytes after 6 * ~500 usec = 3ms.
2017 */
2018static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2019			    int min_tso_segs)
2020{
2021	unsigned long bytes;
2022	u32 r;
2023
2024	bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2025
2026	r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2027	if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2028		bytes += sk->sk_gso_max_size >> r;
2029
2030	bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2031
2032	return max_t(u32, bytes / mss_now, min_tso_segs);
2033}
2034
2035/* Return the number of segments we want in the skb we are transmitting.
2036 * See if congestion control module wants to decide; otherwise, autosize.
2037 */
2038static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2039{
2040	const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2041	u32 min_tso, tso_segs;
2042
2043	min_tso = ca_ops->min_tso_segs ?
2044			ca_ops->min_tso_segs(sk) :
2045			READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2046
2047	tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2048	return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2049}
2050
2051/* Returns the portion of skb which can be sent right away */
2052static unsigned int tcp_mss_split_point(const struct sock *sk,
2053					const struct sk_buff *skb,
2054					unsigned int mss_now,
2055					unsigned int max_segs,
2056					int nonagle)
2057{
2058	const struct tcp_sock *tp = tcp_sk(sk);
2059	u32 partial, needed, window, max_len;
2060
2061	window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2062	max_len = mss_now * max_segs;
2063
2064	if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2065		return max_len;
2066
2067	needed = min(skb->len, window);
2068
2069	if (max_len <= needed)
2070		return max_len;
2071
2072	partial = needed % mss_now;
2073	/* If last segment is not a full MSS, check if Nagle rules allow us
2074	 * to include this last segment in this skb.
2075	 * Otherwise, we'll split the skb at last MSS boundary
2076	 */
2077	if (tcp_nagle_check(partial != 0, tp, nonagle))
2078		return needed - partial;
2079
2080	return needed;
2081}
2082
2083/* Can at least one segment of SKB be sent right now, according to the
2084 * congestion window rules?  If so, return how many segments are allowed.
2085 */
2086static u32 tcp_cwnd_test(const struct tcp_sock *tp)
 
2087{
2088	u32 in_flight, cwnd, halfcwnd;
2089
 
 
 
 
 
2090	in_flight = tcp_packets_in_flight(tp);
2091	cwnd = tcp_snd_cwnd(tp);
2092	if (in_flight >= cwnd)
2093		return 0;
2094
2095	/* For better scheduling, ensure we have at least
2096	 * 2 GSO packets in flight.
2097	 */
2098	halfcwnd = max(cwnd >> 1, 1U);
2099	return min(halfcwnd, cwnd - in_flight);
2100}
2101
2102/* Initialize TSO state of a skb.
2103 * This must be invoked the first time we consider transmitting
2104 * SKB onto the wire.
2105 */
2106static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2107{
2108	int tso_segs = tcp_skb_pcount(skb);
2109
2110	if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now))
2111		return tcp_set_skb_tso_segs(skb, mss_now);
2112
 
2113	return tso_segs;
2114}
2115
2116
2117/* Return true if the Nagle test allows this packet to be
2118 * sent now.
2119 */
2120static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2121				  unsigned int cur_mss, int nonagle)
2122{
2123	/* Nagle rule does not apply to frames, which sit in the middle of the
2124	 * write_queue (they have no chances to get new data).
2125	 *
2126	 * This is implemented in the callers, where they modify the 'nonagle'
2127	 * argument based upon the location of SKB in the send queue.
2128	 */
2129	if (nonagle & TCP_NAGLE_PUSH)
2130		return true;
2131
2132	/* Don't use the nagle rule for urgent data (or for the final FIN). */
2133	if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2134		return true;
2135
2136	if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2137		return true;
2138
2139	return false;
2140}
2141
2142/* Does at least the first segment of SKB fit into the send window? */
2143static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2144			     const struct sk_buff *skb,
2145			     unsigned int cur_mss)
2146{
2147	u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2148
2149	if (skb->len > cur_mss)
2150		end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2151
2152	return !after(end_seq, tcp_wnd_end(tp));
2153}
2154
2155/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2156 * which is put after SKB on the list.  It is very much like
2157 * tcp_fragment() except that it may make several kinds of assumptions
2158 * in order to speed up the splitting operation.  In particular, we
2159 * know that all the data is in scatter-gather pages, and that the
2160 * packet has never been sent out before (and thus is not cloned).
2161 */
2162static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2163			unsigned int mss_now, gfp_t gfp)
2164{
2165	int nlen = skb->len - len;
2166	struct sk_buff *buff;
2167	u8 flags;
2168
2169	/* All of a TSO frame must be composed of paged data.  */
2170	DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
 
 
2171
2172	buff = tcp_stream_alloc_skb(sk, gfp, true);
2173	if (unlikely(!buff))
2174		return -ENOMEM;
2175	skb_copy_decrypted(buff, skb);
2176	mptcp_skb_ext_copy(buff, skb);
2177
2178	sk_wmem_queued_add(sk, buff->truesize);
2179	sk_mem_charge(sk, buff->truesize);
2180	buff->truesize += nlen;
2181	skb->truesize -= nlen;
2182
2183	/* Correct the sequence numbers. */
2184	TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2185	TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2186	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2187
2188	/* PSH and FIN should only be set in the second packet. */
2189	flags = TCP_SKB_CB(skb)->tcp_flags;
2190	TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2191	TCP_SKB_CB(buff)->tcp_flags = flags;
2192
2193	tcp_skb_fragment_eor(skb, buff);
2194
2195	skb_split(skb, buff, len);
2196	tcp_fragment_tstamp(skb, buff);
2197
2198	/* Fix up tso_factor for both original and new SKB.  */
2199	tcp_set_skb_tso_segs(skb, mss_now);
2200	tcp_set_skb_tso_segs(buff, mss_now);
2201
2202	/* Link BUFF into the send queue. */
2203	__skb_header_release(buff);
2204	tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2205
2206	return 0;
2207}
2208
2209/* Try to defer sending, if possible, in order to minimize the amount
2210 * of TSO splitting we do.  View it as a kind of TSO Nagle test.
2211 *
2212 * This algorithm is from John Heffner.
2213 */
2214static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2215				 bool *is_cwnd_limited,
2216				 bool *is_rwnd_limited,
2217				 u32 max_segs)
2218{
2219	const struct inet_connection_sock *icsk = inet_csk(sk);
2220	u32 send_win, cong_win, limit, in_flight;
2221	struct tcp_sock *tp = tcp_sk(sk);
2222	struct sk_buff *head;
2223	int win_divisor;
2224	s64 delta;
2225
2226	if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2227		goto send_now;
2228
2229	/* Avoid bursty behavior by allowing defer
2230	 * only if the last write was recent (1 ms).
2231	 * Note that tp->tcp_wstamp_ns can be in the future if we have
2232	 * packets waiting in a qdisc or device for EDT delivery.
2233	 */
2234	delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2235	if (delta > 0)
2236		goto send_now;
2237
2238	in_flight = tcp_packets_in_flight(tp);
2239
2240	BUG_ON(tcp_skb_pcount(skb) <= 1);
2241	BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2242
2243	send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2244
2245	/* From in_flight test above, we know that cwnd > in_flight.  */
2246	cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2247
2248	limit = min(send_win, cong_win);
2249
2250	/* If a full-sized TSO skb can be sent, do it. */
2251	if (limit >= max_segs * tp->mss_cache)
2252		goto send_now;
2253
2254	/* Middle in queue won't get any more data, full sendable already? */
2255	if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2256		goto send_now;
2257
2258	win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2259	if (win_divisor) {
2260		u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2261
2262		/* If at least some fraction of a window is available,
2263		 * just use it.
2264		 */
2265		chunk /= win_divisor;
2266		if (limit >= chunk)
2267			goto send_now;
2268	} else {
2269		/* Different approach, try not to defer past a single
2270		 * ACK.  Receiver should ACK every other full sized
2271		 * frame, so if we have space for more than 3 frames
2272		 * then send now.
2273		 */
2274		if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2275			goto send_now;
2276	}
2277
2278	/* TODO : use tsorted_sent_queue ? */
2279	head = tcp_rtx_queue_head(sk);
2280	if (!head)
2281		goto send_now;
2282	delta = tp->tcp_clock_cache - head->tstamp;
2283	/* If next ACK is likely to come too late (half srtt), do not defer */
2284	if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2285		goto send_now;
2286
2287	/* Ok, it looks like it is advisable to defer.
2288	 * Three cases are tracked :
2289	 * 1) We are cwnd-limited
2290	 * 2) We are rwnd-limited
2291	 * 3) We are application limited.
2292	 */
2293	if (cong_win < send_win) {
2294		if (cong_win <= skb->len) {
2295			*is_cwnd_limited = true;
2296			return true;
2297		}
2298	} else {
2299		if (send_win <= skb->len) {
2300			*is_rwnd_limited = true;
2301			return true;
2302		}
2303	}
2304
2305	/* If this packet won't get more data, do not wait. */
2306	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2307	    TCP_SKB_CB(skb)->eor)
2308		goto send_now;
2309
2310	return true;
2311
2312send_now:
2313	return false;
2314}
2315
2316static inline void tcp_mtu_check_reprobe(struct sock *sk)
2317{
2318	struct inet_connection_sock *icsk = inet_csk(sk);
2319	struct tcp_sock *tp = tcp_sk(sk);
2320	struct net *net = sock_net(sk);
2321	u32 interval;
2322	s32 delta;
2323
2324	interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2325	delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2326	if (unlikely(delta >= interval * HZ)) {
2327		int mss = tcp_current_mss(sk);
2328
2329		/* Update current search range */
2330		icsk->icsk_mtup.probe_size = 0;
2331		icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2332			sizeof(struct tcphdr) +
2333			icsk->icsk_af_ops->net_header_len;
2334		icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2335
2336		/* Update probe time stamp */
2337		icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2338	}
2339}
2340
2341static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2342{
2343	struct sk_buff *skb, *next;
2344
2345	skb = tcp_send_head(sk);
2346	tcp_for_write_queue_from_safe(skb, next, sk) {
2347		if (len <= skb->len)
2348			break;
2349
2350		if (tcp_has_tx_tstamp(skb) || !tcp_skb_can_collapse(skb, next))
 
 
2351			return false;
2352
2353		len -= skb->len;
2354	}
2355
2356	return true;
2357}
2358
2359static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2360			     int probe_size)
2361{
2362	skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2363	int i, todo, len = 0, nr_frags = 0;
2364	const struct sk_buff *skb;
2365
2366	if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2367		return -ENOMEM;
2368
2369	skb_queue_walk(&sk->sk_write_queue, skb) {
2370		const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2371
2372		if (skb_headlen(skb))
2373			return -EINVAL;
2374
2375		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2376			if (len >= probe_size)
2377				goto commit;
2378			todo = min_t(int, skb_frag_size(fragfrom),
2379				     probe_size - len);
2380			len += todo;
2381			if (lastfrag &&
2382			    skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2383			    skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2384						      skb_frag_size(lastfrag)) {
2385				skb_frag_size_add(lastfrag, todo);
2386				continue;
2387			}
2388			if (unlikely(nr_frags == MAX_SKB_FRAGS))
2389				return -E2BIG;
2390			skb_frag_page_copy(fragto, fragfrom);
2391			skb_frag_off_copy(fragto, fragfrom);
2392			skb_frag_size_set(fragto, todo);
2393			nr_frags++;
2394			lastfrag = fragto++;
2395		}
2396	}
2397commit:
2398	WARN_ON_ONCE(len != probe_size);
2399	for (i = 0; i < nr_frags; i++)
2400		skb_frag_ref(to, i);
2401
2402	skb_shinfo(to)->nr_frags = nr_frags;
2403	to->truesize += probe_size;
2404	to->len += probe_size;
2405	to->data_len += probe_size;
2406	__skb_header_release(to);
2407	return 0;
2408}
2409
2410/* tcp_mtu_probe() and tcp_grow_skb() can both eat an skb (src) if
2411 * all its payload was moved to another one (dst).
2412 * Make sure to transfer tcp_flags, eor, and tstamp.
2413 */
2414static void tcp_eat_one_skb(struct sock *sk,
2415			    struct sk_buff *dst,
2416			    struct sk_buff *src)
2417{
2418	TCP_SKB_CB(dst)->tcp_flags |= TCP_SKB_CB(src)->tcp_flags;
2419	TCP_SKB_CB(dst)->eor = TCP_SKB_CB(src)->eor;
2420	tcp_skb_collapse_tstamp(dst, src);
2421	tcp_unlink_write_queue(src, sk);
2422	tcp_wmem_free_skb(sk, src);
2423}
2424
2425/* Create a new MTU probe if we are ready.
2426 * MTU probe is regularly attempting to increase the path MTU by
2427 * deliberately sending larger packets.  This discovers routing
2428 * changes resulting in larger path MTUs.
2429 *
2430 * Returns 0 if we should wait to probe (no cwnd available),
2431 *         1 if a probe was sent,
2432 *         -1 otherwise
2433 */
2434static int tcp_mtu_probe(struct sock *sk)
2435{
2436	struct inet_connection_sock *icsk = inet_csk(sk);
2437	struct tcp_sock *tp = tcp_sk(sk);
2438	struct sk_buff *skb, *nskb, *next;
2439	struct net *net = sock_net(sk);
2440	int probe_size;
2441	int size_needed;
2442	int copy, len;
2443	int mss_now;
2444	int interval;
2445
2446	/* Not currently probing/verifying,
2447	 * not in recovery,
2448	 * have enough cwnd, and
2449	 * not SACKing (the variable headers throw things off)
2450	 */
2451	if (likely(!icsk->icsk_mtup.enabled ||
2452		   icsk->icsk_mtup.probe_size ||
2453		   inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2454		   tcp_snd_cwnd(tp) < 11 ||
2455		   tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2456		return -1;
2457
2458	/* Use binary search for probe_size between tcp_mss_base,
2459	 * and current mss_clamp. if (search_high - search_low)
2460	 * smaller than a threshold, backoff from probing.
2461	 */
2462	mss_now = tcp_current_mss(sk);
2463	probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2464				    icsk->icsk_mtup.search_low) >> 1);
2465	size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2466	interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2467	/* When misfortune happens, we are reprobing actively,
2468	 * and then reprobe timer has expired. We stick with current
2469	 * probing process by not resetting search range to its orignal.
2470	 */
2471	if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2472	    interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2473		/* Check whether enough time has elaplased for
2474		 * another round of probing.
2475		 */
2476		tcp_mtu_check_reprobe(sk);
2477		return -1;
2478	}
2479
2480	/* Have enough data in the send queue to probe? */
2481	if (tp->write_seq - tp->snd_nxt < size_needed)
2482		return -1;
2483
2484	if (tp->snd_wnd < size_needed)
2485		return -1;
2486	if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2487		return 0;
2488
2489	/* Do we need to wait to drain cwnd? With none in flight, don't stall */
2490	if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2491		if (!tcp_packets_in_flight(tp))
2492			return -1;
2493		else
2494			return 0;
2495	}
2496
2497	if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2498		return -1;
2499
2500	/* We're allowed to probe.  Build it now. */
2501	nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2502	if (!nskb)
2503		return -1;
2504
2505	/* build the payload, and be prepared to abort if this fails. */
2506	if (tcp_clone_payload(sk, nskb, probe_size)) {
2507		tcp_skb_tsorted_anchor_cleanup(nskb);
2508		consume_skb(nskb);
2509		return -1;
2510	}
2511	sk_wmem_queued_add(sk, nskb->truesize);
2512	sk_mem_charge(sk, nskb->truesize);
2513
2514	skb = tcp_send_head(sk);
2515	skb_copy_decrypted(nskb, skb);
2516	mptcp_skb_ext_copy(nskb, skb);
2517
2518	TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2519	TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2520	TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2521
2522	tcp_insert_write_queue_before(nskb, skb, sk);
2523	tcp_highest_sack_replace(sk, skb, nskb);
2524
2525	len = 0;
2526	tcp_for_write_queue_from_safe(skb, next, sk) {
2527		copy = min_t(int, skb->len, probe_size - len);
 
2528
2529		if (skb->len <= copy) {
2530			tcp_eat_one_skb(sk, nskb, skb);
 
 
 
 
 
 
 
 
 
2531		} else {
2532			TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2533						   ~(TCPHDR_FIN|TCPHDR_PSH);
2534			__pskb_trim_head(skb, copy);
2535			tcp_set_skb_tso_segs(skb, mss_now);
 
 
 
 
2536			TCP_SKB_CB(skb)->seq += copy;
2537		}
2538
2539		len += copy;
2540
2541		if (len >= probe_size)
2542			break;
2543	}
2544	tcp_init_tso_segs(nskb, nskb->len);
2545
2546	/* We're ready to send.  If this fails, the probe will
2547	 * be resegmented into mss-sized pieces by tcp_write_xmit().
2548	 */
2549	if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2550		/* Decrement cwnd here because we are sending
2551		 * effectively two packets. */
2552		tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2553		tcp_event_new_data_sent(sk, nskb);
2554
2555		icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2556		tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2557		tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2558
2559		return 1;
2560	}
2561
2562	return -1;
2563}
2564
2565static bool tcp_pacing_check(struct sock *sk)
2566{
2567	struct tcp_sock *tp = tcp_sk(sk);
2568
2569	if (!tcp_needs_internal_pacing(sk))
2570		return false;
2571
2572	if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2573		return false;
2574
2575	if (!hrtimer_is_queued(&tp->pacing_timer)) {
2576		hrtimer_start(&tp->pacing_timer,
2577			      ns_to_ktime(tp->tcp_wstamp_ns),
2578			      HRTIMER_MODE_ABS_PINNED_SOFT);
2579		sock_hold(sk);
2580	}
2581	return true;
2582}
2583
2584static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2585{
2586	const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2587
2588	/* No skb in the rtx queue. */
2589	if (!node)
2590		return true;
2591
2592	/* Only one skb in rtx queue. */
2593	return !node->rb_left && !node->rb_right;
2594}
2595
2596/* TCP Small Queues :
2597 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2598 * (These limits are doubled for retransmits)
2599 * This allows for :
2600 *  - better RTT estimation and ACK scheduling
2601 *  - faster recovery
2602 *  - high rates
2603 * Alas, some drivers / subsystems require a fair amount
2604 * of queued bytes to ensure line rate.
2605 * One example is wifi aggregation (802.11 AMPDU)
2606 */
2607static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2608				  unsigned int factor)
2609{
2610	unsigned long limit;
2611
2612	limit = max_t(unsigned long,
2613		      2 * skb->truesize,
2614		      READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2615	if (sk->sk_pacing_status == SK_PACING_NONE)
2616		limit = min_t(unsigned long, limit,
2617			      READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2618	limit <<= factor;
2619
2620	if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2621	    tcp_sk(sk)->tcp_tx_delay) {
2622		u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2623				  tcp_sk(sk)->tcp_tx_delay;
2624
2625		/* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2626		 * approximate our needs assuming an ~100% skb->truesize overhead.
2627		 * USEC_PER_SEC is approximated by 2^20.
2628		 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2629		 */
2630		extra_bytes >>= (20 - 1);
2631		limit += extra_bytes;
2632	}
2633	if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2634		/* Always send skb if rtx queue is empty or has one skb.
2635		 * No need to wait for TX completion to call us back,
2636		 * after softirq/tasklet schedule.
2637		 * This helps when TX completions are delayed too much.
2638		 */
2639		if (tcp_rtx_queue_empty_or_single_skb(sk))
2640			return false;
2641
2642		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2643		/* It is possible TX completion already happened
2644		 * before we set TSQ_THROTTLED, so we must
2645		 * test again the condition.
2646		 */
2647		smp_mb__after_atomic();
2648		if (refcount_read(&sk->sk_wmem_alloc) > limit)
2649			return true;
2650	}
2651	return false;
2652}
2653
2654static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2655{
2656	const u32 now = tcp_jiffies32;
2657	enum tcp_chrono old = tp->chrono_type;
2658
2659	if (old > TCP_CHRONO_UNSPEC)
2660		tp->chrono_stat[old - 1] += now - tp->chrono_start;
2661	tp->chrono_start = now;
2662	tp->chrono_type = new;
2663}
2664
2665void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2666{
2667	struct tcp_sock *tp = tcp_sk(sk);
2668
2669	/* If there are multiple conditions worthy of tracking in a
2670	 * chronograph then the highest priority enum takes precedence
2671	 * over the other conditions. So that if something "more interesting"
2672	 * starts happening, stop the previous chrono and start a new one.
2673	 */
2674	if (type > tp->chrono_type)
2675		tcp_chrono_set(tp, type);
2676}
2677
2678void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2679{
2680	struct tcp_sock *tp = tcp_sk(sk);
2681
2682
2683	/* There are multiple conditions worthy of tracking in a
2684	 * chronograph, so that the highest priority enum takes
2685	 * precedence over the other conditions (see tcp_chrono_start).
2686	 * If a condition stops, we only stop chrono tracking if
2687	 * it's the "most interesting" or current chrono we are
2688	 * tracking and starts busy chrono if we have pending data.
2689	 */
2690	if (tcp_rtx_and_write_queues_empty(sk))
2691		tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2692	else if (type == tp->chrono_type)
2693		tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2694}
2695
2696/* First skb in the write queue is smaller than ideal packet size.
2697 * Check if we can move payload from the second skb in the queue.
2698 */
2699static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount)
2700{
2701	struct sk_buff *next_skb = skb->next;
2702	unsigned int nlen;
2703
2704	if (tcp_skb_is_last(sk, skb))
2705		return;
2706
2707	if (!tcp_skb_can_collapse(skb, next_skb))
2708		return;
2709
2710	nlen = min_t(u32, amount, next_skb->len);
2711	if (!nlen || !skb_shift(skb, next_skb, nlen))
2712		return;
2713
2714	TCP_SKB_CB(skb)->end_seq += nlen;
2715	TCP_SKB_CB(next_skb)->seq += nlen;
2716
2717	if (!next_skb->len) {
2718		/* In case FIN is set, we need to update end_seq */
2719		TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2720
2721		tcp_eat_one_skb(sk, skb, next_skb);
2722	}
2723}
2724
2725/* This routine writes packets to the network.  It advances the
2726 * send_head.  This happens as incoming acks open up the remote
2727 * window for us.
2728 *
2729 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2730 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2731 * account rare use of URG, this is not a big flaw.
2732 *
2733 * Send at most one packet when push_one > 0. Temporarily ignore
2734 * cwnd limit to force at most one packet out when push_one == 2.
2735
2736 * Returns true, if no segments are in flight and we have queued segments,
2737 * but cannot send anything now because of SWS or another problem.
2738 */
2739static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2740			   int push_one, gfp_t gfp)
2741{
2742	struct tcp_sock *tp = tcp_sk(sk);
2743	struct sk_buff *skb;
2744	unsigned int tso_segs, sent_pkts;
2745	u32 cwnd_quota, max_segs;
2746	int result;
2747	bool is_cwnd_limited = false, is_rwnd_limited = false;
 
2748
2749	sent_pkts = 0;
2750
2751	tcp_mstamp_refresh(tp);
2752	if (!push_one) {
2753		/* Do MTU probing. */
2754		result = tcp_mtu_probe(sk);
2755		if (!result) {
2756			return false;
2757		} else if (result > 0) {
2758			sent_pkts = 1;
2759		}
2760	}
2761
2762	max_segs = tcp_tso_segs(sk, mss_now);
2763	while ((skb = tcp_send_head(sk))) {
2764		unsigned int limit;
2765		int missing_bytes;
2766
2767		if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2768			/* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2769			tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2770			skb_set_delivery_time(skb, tp->tcp_wstamp_ns, SKB_CLOCK_MONOTONIC);
2771			list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2772			tcp_init_tso_segs(skb, mss_now);
2773			goto repair; /* Skip network transmission */
2774		}
2775
2776		if (tcp_pacing_check(sk))
2777			break;
2778
2779		cwnd_quota = tcp_cwnd_test(tp);
 
 
 
2780		if (!cwnd_quota) {
2781			if (push_one == 2)
2782				/* Force out a loss probe pkt. */
2783				cwnd_quota = 1;
2784			else
2785				break;
2786		}
2787		cwnd_quota = min(cwnd_quota, max_segs);
2788		missing_bytes = cwnd_quota * mss_now - skb->len;
2789		if (missing_bytes > 0)
2790			tcp_grow_skb(sk, skb, missing_bytes);
2791
2792		tso_segs = tcp_set_skb_tso_segs(skb, mss_now);
2793
2794		if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2795			is_rwnd_limited = true;
2796			break;
2797		}
2798
2799		if (tso_segs == 1) {
2800			if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2801						     (tcp_skb_is_last(sk, skb) ?
2802						      nonagle : TCP_NAGLE_PUSH))))
2803				break;
2804		} else {
2805			if (!push_one &&
2806			    tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2807						 &is_rwnd_limited, max_segs))
2808				break;
2809		}
2810
2811		limit = mss_now;
2812		if (tso_segs > 1 && !tcp_urg_mode(tp))
2813			limit = tcp_mss_split_point(sk, skb, mss_now,
2814						    cwnd_quota,
 
 
2815						    nonagle);
2816
2817		if (skb->len > limit &&
2818		    unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2819			break;
2820
2821		if (tcp_small_queue_check(sk, skb, 0))
2822			break;
2823
2824		/* Argh, we hit an empty skb(), presumably a thread
2825		 * is sleeping in sendmsg()/sk_stream_wait_memory().
2826		 * We do not want to send a pure-ack packet and have
2827		 * a strange looking rtx queue with empty packet(s).
2828		 */
2829		if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2830			break;
2831
2832		if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2833			break;
2834
2835repair:
2836		/* Advance the send_head.  This one is sent out.
2837		 * This call will increment packets_out.
2838		 */
2839		tcp_event_new_data_sent(sk, skb);
2840
2841		tcp_minshall_update(tp, mss_now, skb);
2842		sent_pkts += tcp_skb_pcount(skb);
2843
2844		if (push_one)
2845			break;
2846	}
2847
2848	if (is_rwnd_limited)
2849		tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2850	else
2851		tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2852
2853	is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2854	if (likely(sent_pkts || is_cwnd_limited))
2855		tcp_cwnd_validate(sk, is_cwnd_limited);
2856
2857	if (likely(sent_pkts)) {
2858		if (tcp_in_cwnd_reduction(sk))
2859			tp->prr_out += sent_pkts;
2860
2861		/* Send one loss probe per tail loss episode. */
2862		if (push_one != 2)
2863			tcp_schedule_loss_probe(sk, false);
2864		return false;
2865	}
2866	return !tp->packets_out && !tcp_write_queue_empty(sk);
2867}
2868
2869bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2870{
2871	struct inet_connection_sock *icsk = inet_csk(sk);
2872	struct tcp_sock *tp = tcp_sk(sk);
2873	u32 timeout, timeout_us, rto_delta_us;
2874	int early_retrans;
2875
2876	/* Don't do any loss probe on a Fast Open connection before 3WHS
2877	 * finishes.
2878	 */
2879	if (rcu_access_pointer(tp->fastopen_rsk))
2880		return false;
2881
2882	early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2883	/* Schedule a loss probe in 2*RTT for SACK capable connections
2884	 * not in loss recovery, that are either limited by cwnd or application.
2885	 */
2886	if ((early_retrans != 3 && early_retrans != 4) ||
2887	    !tp->packets_out || !tcp_is_sack(tp) ||
2888	    (icsk->icsk_ca_state != TCP_CA_Open &&
2889	     icsk->icsk_ca_state != TCP_CA_CWR))
2890		return false;
2891
2892	/* Probe timeout is 2*rtt. Add minimum RTO to account
2893	 * for delayed ack when there's one outstanding packet. If no RTT
2894	 * sample is available then probe after TCP_TIMEOUT_INIT.
2895	 */
2896	if (tp->srtt_us) {
2897		timeout_us = tp->srtt_us >> 2;
2898		if (tp->packets_out == 1)
2899			timeout_us += tcp_rto_min_us(sk);
2900		else
2901			timeout_us += TCP_TIMEOUT_MIN_US;
2902		timeout = usecs_to_jiffies(timeout_us);
2903	} else {
2904		timeout = TCP_TIMEOUT_INIT;
2905	}
2906
2907	/* If the RTO formula yields an earlier time, then use that time. */
2908	rto_delta_us = advancing_rto ?
2909			jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2910			tcp_rto_delta_us(sk);  /* How far in future is RTO? */
2911	if (rto_delta_us > 0)
2912		timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2913
2914	tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2915	return true;
2916}
2917
2918/* Thanks to skb fast clones, we can detect if a prior transmit of
2919 * a packet is still in a qdisc or driver queue.
2920 * In this case, there is very little point doing a retransmit !
2921 */
2922static bool skb_still_in_host_queue(struct sock *sk,
2923				    const struct sk_buff *skb)
2924{
2925	if (unlikely(skb_fclone_busy(sk, skb))) {
2926		set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2927		smp_mb__after_atomic();
2928		if (skb_fclone_busy(sk, skb)) {
2929			NET_INC_STATS(sock_net(sk),
2930				      LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2931			return true;
2932		}
2933	}
2934	return false;
2935}
2936
2937/* When probe timeout (PTO) fires, try send a new segment if possible, else
2938 * retransmit the last segment.
2939 */
2940void tcp_send_loss_probe(struct sock *sk)
2941{
2942	struct tcp_sock *tp = tcp_sk(sk);
2943	struct sk_buff *skb;
2944	int pcount;
2945	int mss = tcp_current_mss(sk);
2946
2947	/* At most one outstanding TLP */
2948	if (tp->tlp_high_seq)
2949		goto rearm_timer;
2950
2951	tp->tlp_retrans = 0;
2952	skb = tcp_send_head(sk);
2953	if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2954		pcount = tp->packets_out;
2955		tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2956		if (tp->packets_out > pcount)
2957			goto probe_sent;
2958		goto rearm_timer;
2959	}
2960	skb = skb_rb_last(&sk->tcp_rtx_queue);
2961	if (unlikely(!skb)) {
2962		tcp_warn_once(sk, tp->packets_out, "invalid inflight: ");
2963		smp_store_release(&inet_csk(sk)->icsk_pending, 0);
 
 
2964		return;
2965	}
2966
2967	if (skb_still_in_host_queue(sk, skb))
2968		goto rearm_timer;
2969
2970	pcount = tcp_skb_pcount(skb);
2971	if (WARN_ON(!pcount))
2972		goto rearm_timer;
2973
2974	if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2975		if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2976					  (pcount - 1) * mss, mss,
2977					  GFP_ATOMIC)))
2978			goto rearm_timer;
2979		skb = skb_rb_next(skb);
2980	}
2981
2982	if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2983		goto rearm_timer;
2984
2985	if (__tcp_retransmit_skb(sk, skb, 1))
2986		goto rearm_timer;
2987
2988	tp->tlp_retrans = 1;
2989
2990probe_sent:
2991	/* Record snd_nxt for loss detection. */
2992	tp->tlp_high_seq = tp->snd_nxt;
2993
2994	NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2995	/* Reset s.t. tcp_rearm_rto will restart timer from now */
2996	smp_store_release(&inet_csk(sk)->icsk_pending, 0);
2997rearm_timer:
2998	tcp_rearm_rto(sk);
2999}
3000
3001/* Push out any pending frames which were held back due to
3002 * TCP_CORK or attempt at coalescing tiny packets.
3003 * The socket must be locked by the caller.
3004 */
3005void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
3006			       int nonagle)
3007{
3008	/* If we are closed, the bytes will have to remain here.
3009	 * In time closedown will finish, we empty the write queue and
3010	 * all will be happy.
3011	 */
3012	if (unlikely(sk->sk_state == TCP_CLOSE))
3013		return;
3014
3015	if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
3016			   sk_gfp_mask(sk, GFP_ATOMIC)))
3017		tcp_check_probe_timer(sk);
3018}
3019
3020/* Send _single_ skb sitting at the send head. This function requires
3021 * true push pending frames to setup probe timer etc.
3022 */
3023void tcp_push_one(struct sock *sk, unsigned int mss_now)
3024{
3025	struct sk_buff *skb = tcp_send_head(sk);
3026
3027	BUG_ON(!skb || skb->len < mss_now);
3028
3029	tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
3030}
3031
3032/* This function returns the amount that we can raise the
3033 * usable window based on the following constraints
3034 *
3035 * 1. The window can never be shrunk once it is offered (RFC 793)
3036 * 2. We limit memory per socket
3037 *
3038 * RFC 1122:
3039 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3040 *  RECV.NEXT + RCV.WIN fixed until:
3041 *  RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3042 *
3043 * i.e. don't raise the right edge of the window until you can raise
3044 * it at least MSS bytes.
3045 *
3046 * Unfortunately, the recommended algorithm breaks header prediction,
3047 * since header prediction assumes th->window stays fixed.
3048 *
3049 * Strictly speaking, keeping th->window fixed violates the receiver
3050 * side SWS prevention criteria. The problem is that under this rule
3051 * a stream of single byte packets will cause the right side of the
3052 * window to always advance by a single byte.
3053 *
3054 * Of course, if the sender implements sender side SWS prevention
3055 * then this will not be a problem.
3056 *
3057 * BSD seems to make the following compromise:
3058 *
3059 *	If the free space is less than the 1/4 of the maximum
3060 *	space available and the free space is less than 1/2 mss,
3061 *	then set the window to 0.
3062 *	[ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3063 *	Otherwise, just prevent the window from shrinking
3064 *	and from being larger than the largest representable value.
3065 *
3066 * This prevents incremental opening of the window in the regime
3067 * where TCP is limited by the speed of the reader side taking
3068 * data out of the TCP receive queue. It does nothing about
3069 * those cases where the window is constrained on the sender side
3070 * because the pipeline is full.
3071 *
3072 * BSD also seems to "accidentally" limit itself to windows that are a
3073 * multiple of MSS, at least until the free space gets quite small.
3074 * This would appear to be a side effect of the mbuf implementation.
3075 * Combining these two algorithms results in the observed behavior
3076 * of having a fixed window size at almost all times.
3077 *
3078 * Below we obtain similar behavior by forcing the offered window to
3079 * a multiple of the mss when it is feasible to do so.
3080 *
3081 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3082 * Regular options like TIMESTAMP are taken into account.
3083 */
3084u32 __tcp_select_window(struct sock *sk)
3085{
3086	struct inet_connection_sock *icsk = inet_csk(sk);
3087	struct tcp_sock *tp = tcp_sk(sk);
3088	struct net *net = sock_net(sk);
3089	/* MSS for the peer's data.  Previous versions used mss_clamp
3090	 * here.  I don't know if the value based on our guesses
3091	 * of peer's MSS is better for the performance.  It's more correct
3092	 * but may be worse for the performance because of rcv_mss
3093	 * fluctuations.  --SAW  1998/11/1
3094	 */
3095	int mss = icsk->icsk_ack.rcv_mss;
3096	int free_space = tcp_space(sk);
3097	int allowed_space = tcp_full_space(sk);
3098	int full_space, window;
3099
3100	if (sk_is_mptcp(sk))
3101		mptcp_space(sk, &free_space, &allowed_space);
3102
3103	full_space = min_t(int, tp->window_clamp, allowed_space);
3104
3105	if (unlikely(mss > full_space)) {
3106		mss = full_space;
3107		if (mss <= 0)
3108			return 0;
3109	}
3110
3111	/* Only allow window shrink if the sysctl is enabled and we have
3112	 * a non-zero scaling factor in effect.
3113	 */
3114	if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3115		goto shrink_window_allowed;
3116
3117	/* do not allow window to shrink */
3118
3119	if (free_space < (full_space >> 1)) {
3120		icsk->icsk_ack.quick = 0;
3121
3122		if (tcp_under_memory_pressure(sk))
3123			tcp_adjust_rcv_ssthresh(sk);
3124
3125		/* free_space might become our new window, make sure we don't
3126		 * increase it due to wscale.
3127		 */
3128		free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3129
3130		/* if free space is less than mss estimate, or is below 1/16th
3131		 * of the maximum allowed, try to move to zero-window, else
3132		 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3133		 * new incoming data is dropped due to memory limits.
3134		 * With large window, mss test triggers way too late in order
3135		 * to announce zero window in time before rmem limit kicks in.
3136		 */
3137		if (free_space < (allowed_space >> 4) || free_space < mss)
3138			return 0;
3139	}
3140
3141	if (free_space > tp->rcv_ssthresh)
3142		free_space = tp->rcv_ssthresh;
3143
3144	/* Don't do rounding if we are using window scaling, since the
3145	 * scaled window will not line up with the MSS boundary anyway.
3146	 */
3147	if (tp->rx_opt.rcv_wscale) {
3148		window = free_space;
3149
3150		/* Advertise enough space so that it won't get scaled away.
3151		 * Import case: prevent zero window announcement if
3152		 * 1<<rcv_wscale > mss.
3153		 */
3154		window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3155	} else {
3156		window = tp->rcv_wnd;
3157		/* Get the largest window that is a nice multiple of mss.
3158		 * Window clamp already applied above.
3159		 * If our current window offering is within 1 mss of the
3160		 * free space we just keep it. This prevents the divide
3161		 * and multiply from happening most of the time.
3162		 * We also don't do any window rounding when the free space
3163		 * is too small.
3164		 */
3165		if (window <= free_space - mss || window > free_space)
3166			window = rounddown(free_space, mss);
3167		else if (mss == full_space &&
3168			 free_space > window + (full_space >> 1))
3169			window = free_space;
3170	}
3171
3172	return window;
3173
3174shrink_window_allowed:
3175	/* new window should always be an exact multiple of scaling factor */
3176	free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3177
3178	if (free_space < (full_space >> 1)) {
3179		icsk->icsk_ack.quick = 0;
3180
3181		if (tcp_under_memory_pressure(sk))
3182			tcp_adjust_rcv_ssthresh(sk);
3183
3184		/* if free space is too low, return a zero window */
3185		if (free_space < (allowed_space >> 4) || free_space < mss ||
3186			free_space < (1 << tp->rx_opt.rcv_wscale))
3187			return 0;
3188	}
3189
3190	if (free_space > tp->rcv_ssthresh) {
3191		free_space = tp->rcv_ssthresh;
3192		/* new window should always be an exact multiple of scaling factor
3193		 *
3194		 * For this case, we ALIGN "up" (increase free_space) because
3195		 * we know free_space is not zero here, it has been reduced from
3196		 * the memory-based limit, and rcv_ssthresh is not a hard limit
3197		 * (unlike sk_rcvbuf).
3198		 */
3199		free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3200	}
3201
3202	return free_space;
3203}
3204
3205void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3206			     const struct sk_buff *next_skb)
3207{
3208	if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3209		const struct skb_shared_info *next_shinfo =
3210			skb_shinfo(next_skb);
3211		struct skb_shared_info *shinfo = skb_shinfo(skb);
3212
3213		shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3214		shinfo->tskey = next_shinfo->tskey;
3215		TCP_SKB_CB(skb)->txstamp_ack |=
3216			TCP_SKB_CB(next_skb)->txstamp_ack;
3217	}
3218}
3219
3220/* Collapses two adjacent SKB's during retransmission. */
3221static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3222{
3223	struct tcp_sock *tp = tcp_sk(sk);
3224	struct sk_buff *next_skb = skb_rb_next(skb);
3225	int next_skb_size;
3226
3227	next_skb_size = next_skb->len;
3228
3229	BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3230
3231	if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3232		return false;
3233
3234	tcp_highest_sack_replace(sk, next_skb, skb);
3235
3236	/* Update sequence range on original skb. */
3237	TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3238
3239	/* Merge over control information. This moves PSH/FIN etc. over */
3240	TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3241
3242	/* All done, get rid of second SKB and account for it so
3243	 * packet counting does not break.
3244	 */
3245	TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3246	TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3247
3248	/* changed transmit queue under us so clear hints */
3249	tcp_clear_retrans_hints_partial(tp);
3250	if (next_skb == tp->retransmit_skb_hint)
3251		tp->retransmit_skb_hint = skb;
3252
3253	tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3254
3255	tcp_skb_collapse_tstamp(skb, next_skb);
3256
3257	tcp_rtx_queue_unlink_and_free(next_skb, sk);
3258	return true;
3259}
3260
3261/* Check if coalescing SKBs is legal. */
3262static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3263{
3264	if (tcp_skb_pcount(skb) > 1)
3265		return false;
3266	if (skb_cloned(skb))
3267		return false;
3268	if (!skb_frags_readable(skb))
3269		return false;
3270	/* Some heuristics for collapsing over SACK'd could be invented */
3271	if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3272		return false;
3273
3274	return true;
3275}
3276
3277/* Collapse packets in the retransmit queue to make to create
3278 * less packets on the wire. This is only done on retransmission.
3279 */
3280static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3281				     int space)
3282{
3283	struct tcp_sock *tp = tcp_sk(sk);
3284	struct sk_buff *skb = to, *tmp;
3285	bool first = true;
3286
3287	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3288		return;
3289	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3290		return;
3291
3292	skb_rbtree_walk_from_safe(skb, tmp) {
3293		if (!tcp_can_collapse(sk, skb))
3294			break;
3295
3296		if (!tcp_skb_can_collapse(to, skb))
3297			break;
3298
3299		space -= skb->len;
3300
3301		if (first) {
3302			first = false;
3303			continue;
3304		}
3305
3306		if (space < 0)
3307			break;
3308
3309		if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3310			break;
3311
3312		if (!tcp_collapse_retrans(sk, to))
3313			break;
3314	}
3315}
3316
3317/* This retransmits one SKB.  Policy decisions and retransmit queue
3318 * state updates are done by the caller.  Returns non-zero if an
3319 * error occurred which prevented the send.
3320 */
3321int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3322{
3323	struct inet_connection_sock *icsk = inet_csk(sk);
3324	struct tcp_sock *tp = tcp_sk(sk);
3325	unsigned int cur_mss;
3326	int diff, len, err;
3327	int avail_wnd;
3328
3329	/* Inconclusive MTU probe */
3330	if (icsk->icsk_mtup.probe_size)
3331		icsk->icsk_mtup.probe_size = 0;
3332
3333	if (skb_still_in_host_queue(sk, skb))
3334		return -EBUSY;
3335
3336start:
3337	if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3338		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3339			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3340			TCP_SKB_CB(skb)->seq++;
3341			goto start;
3342		}
3343		if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3344			WARN_ON_ONCE(1);
3345			return -EINVAL;
3346		}
3347		if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3348			return -ENOMEM;
3349	}
3350
3351	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3352		return -EHOSTUNREACH; /* Routing failure or similar. */
3353
3354	cur_mss = tcp_current_mss(sk);
3355	avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3356
3357	/* If receiver has shrunk his window, and skb is out of
3358	 * new window, do not retransmit it. The exception is the
3359	 * case, when window is shrunk to zero. In this case
3360	 * our retransmit of one segment serves as a zero window probe.
3361	 */
3362	if (avail_wnd <= 0) {
3363		if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3364			return -EAGAIN;
3365		avail_wnd = cur_mss;
3366	}
3367
3368	len = cur_mss * segs;
3369	if (len > avail_wnd) {
3370		len = rounddown(avail_wnd, cur_mss);
3371		if (!len)
3372			len = avail_wnd;
3373	}
3374	if (skb->len > len) {
3375		if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3376				 cur_mss, GFP_ATOMIC))
3377			return -ENOMEM; /* We'll try again later. */
3378	} else {
3379		if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3380			return -ENOMEM;
3381
3382		diff = tcp_skb_pcount(skb);
3383		tcp_set_skb_tso_segs(skb, cur_mss);
3384		diff -= tcp_skb_pcount(skb);
3385		if (diff)
3386			tcp_adjust_pcount(sk, skb, diff);
3387		avail_wnd = min_t(int, avail_wnd, cur_mss);
3388		if (skb->len < avail_wnd)
3389			tcp_retrans_try_collapse(sk, skb, avail_wnd);
3390	}
3391
3392	/* RFC3168, section 6.1.1.1. ECN fallback */
3393	if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3394		tcp_ecn_clear_syn(sk, skb);
3395
3396	/* Update global and local TCP statistics. */
3397	segs = tcp_skb_pcount(skb);
3398	TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3399	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3400		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3401	tp->total_retrans += segs;
3402	tp->bytes_retrans += skb->len;
3403
3404	/* make sure skb->data is aligned on arches that require it
3405	 * and check if ack-trimming & collapsing extended the headroom
3406	 * beyond what csum_start can cover.
3407	 */
3408	if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3409		     skb_headroom(skb) >= 0xFFFF)) {
3410		struct sk_buff *nskb;
3411
3412		tcp_skb_tsorted_save(skb) {
3413			nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3414			if (nskb) {
3415				nskb->dev = NULL;
3416				err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3417			} else {
3418				err = -ENOBUFS;
3419			}
3420		} tcp_skb_tsorted_restore(skb);
3421
3422		if (!err) {
3423			tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3424			tcp_rate_skb_sent(sk, skb);
3425		}
3426	} else {
3427		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3428	}
3429
 
 
 
 
 
3430	if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3431		tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3432				  TCP_SKB_CB(skb)->seq, segs, err);
3433
3434	if (likely(!err)) {
3435		trace_tcp_retransmit_skb(sk, skb);
3436	} else if (err != -EBUSY) {
3437		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3438	}
3439
3440	/* To avoid taking spuriously low RTT samples based on a timestamp
3441	 * for a transmit that never happened, always mark EVER_RETRANS
3442	 */
3443	TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3444
3445	return err;
3446}
3447
3448int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3449{
3450	struct tcp_sock *tp = tcp_sk(sk);
3451	int err = __tcp_retransmit_skb(sk, skb, segs);
3452
3453	if (err == 0) {
3454#if FASTRETRANS_DEBUG > 0
3455		if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3456			net_dbg_ratelimited("retrans_out leaked\n");
3457		}
3458#endif
3459		TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3460		tp->retrans_out += tcp_skb_pcount(skb);
3461	}
3462
3463	/* Save stamp of the first (attempted) retransmit. */
3464	if (!tp->retrans_stamp)
3465		tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb);
3466
3467	if (tp->undo_retrans < 0)
3468		tp->undo_retrans = 0;
3469	tp->undo_retrans += tcp_skb_pcount(skb);
3470	return err;
3471}
3472
3473/* This gets called after a retransmit timeout, and the initially
3474 * retransmitted data is acknowledged.  It tries to continue
3475 * resending the rest of the retransmit queue, until either
3476 * we've sent it all or the congestion window limit is reached.
3477 */
3478void tcp_xmit_retransmit_queue(struct sock *sk)
3479{
3480	const struct inet_connection_sock *icsk = inet_csk(sk);
3481	struct sk_buff *skb, *rtx_head, *hole = NULL;
3482	struct tcp_sock *tp = tcp_sk(sk);
3483	bool rearm_timer = false;
3484	u32 max_segs;
3485	int mib_idx;
3486
3487	if (!tp->packets_out)
3488		return;
3489
3490	rtx_head = tcp_rtx_queue_head(sk);
3491	skb = tp->retransmit_skb_hint ?: rtx_head;
3492	max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3493	skb_rbtree_walk_from(skb) {
3494		__u8 sacked;
3495		int segs;
3496
3497		if (tcp_pacing_check(sk))
3498			break;
3499
3500		/* we could do better than to assign each time */
3501		if (!hole)
3502			tp->retransmit_skb_hint = skb;
3503
3504		segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3505		if (segs <= 0)
3506			break;
3507		sacked = TCP_SKB_CB(skb)->sacked;
3508		/* In case tcp_shift_skb_data() have aggregated large skbs,
3509		 * we need to make sure not sending too bigs TSO packets
3510		 */
3511		segs = min_t(int, segs, max_segs);
3512
3513		if (tp->retrans_out >= tp->lost_out) {
3514			break;
3515		} else if (!(sacked & TCPCB_LOST)) {
3516			if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3517				hole = skb;
3518			continue;
3519
3520		} else {
3521			if (icsk->icsk_ca_state != TCP_CA_Loss)
3522				mib_idx = LINUX_MIB_TCPFASTRETRANS;
3523			else
3524				mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3525		}
3526
3527		if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3528			continue;
3529
3530		if (tcp_small_queue_check(sk, skb, 1))
3531			break;
3532
3533		if (tcp_retransmit_skb(sk, skb, segs))
3534			break;
3535
3536		NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3537
3538		if (tcp_in_cwnd_reduction(sk))
3539			tp->prr_out += tcp_skb_pcount(skb);
3540
3541		if (skb == rtx_head &&
3542		    icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3543			rearm_timer = true;
3544
3545	}
3546	if (rearm_timer)
3547		tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3548				     inet_csk(sk)->icsk_rto,
3549				     TCP_RTO_MAX);
3550}
3551
3552/* We allow to exceed memory limits for FIN packets to expedite
3553 * connection tear down and (memory) recovery.
3554 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3555 * or even be forced to close flow without any FIN.
3556 * In general, we want to allow one skb per socket to avoid hangs
3557 * with edge trigger epoll()
3558 */
3559void sk_forced_mem_schedule(struct sock *sk, int size)
3560{
3561	int delta, amt;
3562
3563	delta = size - sk->sk_forward_alloc;
3564	if (delta <= 0)
3565		return;
3566	amt = sk_mem_pages(delta);
3567	sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3568	sk_memory_allocated_add(sk, amt);
3569
3570	if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3571		mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3572					gfp_memcg_charge() | __GFP_NOFAIL);
3573}
3574
3575/* Send a FIN. The caller locks the socket for us.
3576 * We should try to send a FIN packet really hard, but eventually give up.
3577 */
3578void tcp_send_fin(struct sock *sk)
3579{
3580	struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3581	struct tcp_sock *tp = tcp_sk(sk);
3582
3583	/* Optimization, tack on the FIN if we have one skb in write queue and
3584	 * this skb was not yet sent, or we are under memory pressure.
3585	 * Note: in the latter case, FIN packet will be sent after a timeout,
3586	 * as TCP stack thinks it has already been transmitted.
3587	 */
3588	tskb = tail;
3589	if (!tskb && tcp_under_memory_pressure(sk))
3590		tskb = skb_rb_last(&sk->tcp_rtx_queue);
3591
3592	if (tskb) {
3593		TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3594		TCP_SKB_CB(tskb)->end_seq++;
3595		tp->write_seq++;
3596		if (!tail) {
3597			/* This means tskb was already sent.
3598			 * Pretend we included the FIN on previous transmit.
3599			 * We need to set tp->snd_nxt to the value it would have
3600			 * if FIN had been sent. This is because retransmit path
3601			 * does not change tp->snd_nxt.
3602			 */
3603			WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3604			return;
3605		}
3606	} else {
3607		skb = alloc_skb_fclone(MAX_TCP_HEADER,
3608				       sk_gfp_mask(sk, GFP_ATOMIC |
3609						       __GFP_NOWARN));
3610		if (unlikely(!skb))
3611			return;
3612
3613		INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3614		skb_reserve(skb, MAX_TCP_HEADER);
3615		sk_forced_mem_schedule(sk, skb->truesize);
3616		/* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3617		tcp_init_nondata_skb(skb, tp->write_seq,
3618				     TCPHDR_ACK | TCPHDR_FIN);
3619		tcp_queue_skb(sk, skb);
3620	}
3621	__tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3622}
3623
3624/* We get here when a process closes a file descriptor (either due to
3625 * an explicit close() or as a byproduct of exit()'ing) and there
3626 * was unread data in the receive queue.  This behavior is recommended
3627 * by RFC 2525, section 2.17.  -DaveM
3628 */
3629void tcp_send_active_reset(struct sock *sk, gfp_t priority,
3630			   enum sk_rst_reason reason)
3631{
3632	struct sk_buff *skb;
3633
3634	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3635
3636	/* NOTE: No TCP options attached and we never retransmit this. */
3637	skb = alloc_skb(MAX_TCP_HEADER, priority);
3638	if (!skb) {
3639		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3640		return;
3641	}
3642
3643	/* Reserve space for headers and prepare control bits. */
3644	skb_reserve(skb, MAX_TCP_HEADER);
3645	tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3646			     TCPHDR_ACK | TCPHDR_RST);
3647	tcp_mstamp_refresh(tcp_sk(sk));
3648	/* Send it off. */
3649	if (tcp_transmit_skb(sk, skb, 0, priority))
3650		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3651
3652	/* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3653	 * skb here is different to the troublesome skb, so use NULL
3654	 */
3655	trace_tcp_send_reset(sk, NULL, reason);
3656}
3657
3658/* Send a crossed SYN-ACK during socket establishment.
3659 * WARNING: This routine must only be called when we have already sent
3660 * a SYN packet that crossed the incoming SYN that caused this routine
3661 * to get called. If this assumption fails then the initial rcv_wnd
3662 * and rcv_wscale values will not be correct.
3663 */
3664int tcp_send_synack(struct sock *sk)
3665{
3666	struct sk_buff *skb;
3667
3668	skb = tcp_rtx_queue_head(sk);
3669	if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3670		pr_err("%s: wrong queue state\n", __func__);
3671		return -EFAULT;
3672	}
3673	if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3674		if (skb_cloned(skb)) {
3675			struct sk_buff *nskb;
3676
3677			tcp_skb_tsorted_save(skb) {
3678				nskb = skb_copy(skb, GFP_ATOMIC);
3679			} tcp_skb_tsorted_restore(skb);
3680			if (!nskb)
3681				return -ENOMEM;
3682			INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3683			tcp_highest_sack_replace(sk, skb, nskb);
3684			tcp_rtx_queue_unlink_and_free(skb, sk);
3685			__skb_header_release(nskb);
3686			tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3687			sk_wmem_queued_add(sk, nskb->truesize);
3688			sk_mem_charge(sk, nskb->truesize);
3689			skb = nskb;
3690		}
3691
3692		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3693		tcp_ecn_send_synack(sk, skb);
3694	}
3695	return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3696}
3697
3698/**
3699 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3700 * @sk: listener socket
3701 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3702 *       should not use it again.
3703 * @req: request_sock pointer
3704 * @foc: cookie for tcp fast open
3705 * @synack_type: Type of synack to prepare
3706 * @syn_skb: SYN packet just received.  It could be NULL for rtx case.
3707 */
3708struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3709				struct request_sock *req,
3710				struct tcp_fastopen_cookie *foc,
3711				enum tcp_synack_type synack_type,
3712				struct sk_buff *syn_skb)
3713{
3714	struct inet_request_sock *ireq = inet_rsk(req);
3715	const struct tcp_sock *tp = tcp_sk(sk);
 
3716	struct tcp_out_options opts;
3717	struct tcp_key key = {};
3718	struct sk_buff *skb;
3719	int tcp_header_size;
3720	struct tcphdr *th;
3721	int mss;
3722	u64 now;
3723
3724	skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3725	if (unlikely(!skb)) {
3726		dst_release(dst);
3727		return NULL;
3728	}
3729	/* Reserve space for headers. */
3730	skb_reserve(skb, MAX_TCP_HEADER);
3731
3732	switch (synack_type) {
3733	case TCP_SYNACK_NORMAL:
3734		skb_set_owner_edemux(skb, req_to_sk(req));
3735		break;
3736	case TCP_SYNACK_COOKIE:
3737		/* Under synflood, we do not attach skb to a socket,
3738		 * to avoid false sharing.
3739		 */
3740		break;
3741	case TCP_SYNACK_FASTOPEN:
3742		/* sk is a const pointer, because we want to express multiple
3743		 * cpu might call us concurrently.
3744		 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3745		 */
3746		skb_set_owner_w(skb, (struct sock *)sk);
3747		break;
3748	}
3749	skb_dst_set(skb, dst);
3750
3751	mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3752
3753	memset(&opts, 0, sizeof(opts));
3754	now = tcp_clock_ns();
3755#ifdef CONFIG_SYN_COOKIES
3756	if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3757		skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3758				      SKB_CLOCK_MONOTONIC);
3759	else
3760#endif
3761	{
3762		skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
3763		if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3764			tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3765	}
3766
3767#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3768	rcu_read_lock();
 
3769#endif
3770	if (tcp_rsk_used_ao(req)) {
3771#ifdef CONFIG_TCP_AO
3772		struct tcp_ao_key *ao_key = NULL;
3773		u8 keyid = tcp_rsk(req)->ao_keyid;
3774		u8 rnext = tcp_rsk(req)->ao_rcv_next;
3775
3776		ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3777							    keyid, -1);
3778		/* If there is no matching key - avoid sending anything,
3779		 * especially usigned segments. It could try harder and lookup
3780		 * for another peer-matching key, but the peer has requested
3781		 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3782		 */
3783		if (unlikely(!ao_key)) {
3784			trace_tcp_ao_synack_no_key(sk, keyid, rnext);
3785			rcu_read_unlock();
3786			kfree_skb(skb);
3787			net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3788					     keyid);
3789			return NULL;
3790		}
3791		key.ao_key = ao_key;
3792		key.type = TCP_KEY_AO;
3793#endif
3794	} else {
3795#ifdef CONFIG_TCP_MD5SIG
3796		key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3797					req_to_sk(req));
3798		if (key.md5_key)
3799			key.type = TCP_KEY_MD5;
3800#endif
3801	}
3802	skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3803	/* bpf program will be interested in the tcp_flags */
3804	TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3805	tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts,
3806					     &key, foc, synack_type, syn_skb)
3807					+ sizeof(*th);
3808
3809	skb_push(skb, tcp_header_size);
3810	skb_reset_transport_header(skb);
3811
3812	th = (struct tcphdr *)skb->data;
3813	memset(th, 0, sizeof(struct tcphdr));
3814	th->syn = 1;
3815	th->ack = 1;
3816	tcp_ecn_make_synack(req, th);
3817	th->source = htons(ireq->ir_num);
3818	th->dest = ireq->ir_rmt_port;
3819	skb->mark = ireq->ir_mark;
3820	skb->ip_summed = CHECKSUM_PARTIAL;
3821	th->seq = htonl(tcp_rsk(req)->snt_isn);
3822	/* XXX data is queued and acked as is. No buffer/window check */
3823	th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3824
3825	/* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3826	th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3827	tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key);
3828	th->doff = (tcp_header_size >> 2);
3829	TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3830
 
3831	/* Okay, we have all we need - do the md5 hash if needed */
3832	if (tcp_key_is_md5(&key)) {
3833#ifdef CONFIG_TCP_MD5SIG
3834		tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3835					key.md5_key, req_to_sk(req), skb);
3836#endif
3837	} else if (tcp_key_is_ao(&key)) {
3838#ifdef CONFIG_TCP_AO
3839		tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
3840					key.ao_key, req, skb,
3841					opts.hash_location - (u8 *)th, 0);
3842#endif
3843	}
3844#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3845	rcu_read_unlock();
3846#endif
3847
3848	bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3849				synack_type, &opts);
3850
3851	skb_set_delivery_time(skb, now, SKB_CLOCK_MONOTONIC);
3852	tcp_add_tx_delay(skb, tp);
3853
3854	return skb;
3855}
3856EXPORT_SYMBOL(tcp_make_synack);
3857
3858static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3859{
3860	struct inet_connection_sock *icsk = inet_csk(sk);
3861	const struct tcp_congestion_ops *ca;
3862	u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3863
3864	if (ca_key == TCP_CA_UNSPEC)
3865		return;
3866
3867	rcu_read_lock();
3868	ca = tcp_ca_find_key(ca_key);
3869	if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3870		bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3871		icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3872		icsk->icsk_ca_ops = ca;
3873	}
3874	rcu_read_unlock();
3875}
3876
3877/* Do all connect socket setups that can be done AF independent. */
3878static void tcp_connect_init(struct sock *sk)
3879{
3880	const struct dst_entry *dst = __sk_dst_get(sk);
3881	struct tcp_sock *tp = tcp_sk(sk);
3882	__u8 rcv_wscale;
3883	u32 rcv_wnd;
3884
3885	/* We'll fix this up when we get a response from the other end.
3886	 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3887	 */
3888	tp->tcp_header_len = sizeof(struct tcphdr);
3889	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3890		tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3891
3892	tcp_ao_connect_init(sk);
 
 
 
3893
3894	/* If user gave his TCP_MAXSEG, record it to clamp */
3895	if (tp->rx_opt.user_mss)
3896		tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3897	tp->max_window = 0;
3898	tcp_mtup_init(sk);
3899	tcp_sync_mss(sk, dst_mtu(dst));
3900
3901	tcp_ca_dst_init(sk, dst);
3902
3903	if (!tp->window_clamp)
3904		WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW));
3905	tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3906
3907	tcp_initialize_rcv_mss(sk);
3908
3909	/* limit the window selection if the user enforce a smaller rx buffer */
3910	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3911	    (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3912		WRITE_ONCE(tp->window_clamp, tcp_full_space(sk));
3913
3914	rcv_wnd = tcp_rwnd_init_bpf(sk);
3915	if (rcv_wnd == 0)
3916		rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3917
3918	tcp_select_initial_window(sk, tcp_full_space(sk),
3919				  tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3920				  &tp->rcv_wnd,
3921				  &tp->window_clamp,
3922				  READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3923				  &rcv_wscale,
3924				  rcv_wnd);
3925
3926	tp->rx_opt.rcv_wscale = rcv_wscale;
3927	tp->rcv_ssthresh = tp->rcv_wnd;
3928
3929	WRITE_ONCE(sk->sk_err, 0);
3930	sock_reset_flag(sk, SOCK_DONE);
3931	tp->snd_wnd = 0;
3932	tcp_init_wl(tp, 0);
3933	tcp_write_queue_purge(sk);
3934	tp->snd_una = tp->write_seq;
3935	tp->snd_sml = tp->write_seq;
3936	tp->snd_up = tp->write_seq;
3937	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3938
3939	if (likely(!tp->repair))
3940		tp->rcv_nxt = 0;
3941	else
3942		tp->rcv_tstamp = tcp_jiffies32;
3943	tp->rcv_wup = tp->rcv_nxt;
3944	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3945
3946	inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3947	inet_csk(sk)->icsk_retransmits = 0;
3948	tcp_clear_retrans(tp);
3949}
3950
3951static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3952{
3953	struct tcp_sock *tp = tcp_sk(sk);
3954	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3955
3956	tcb->end_seq += skb->len;
3957	__skb_header_release(skb);
3958	sk_wmem_queued_add(sk, skb->truesize);
3959	sk_mem_charge(sk, skb->truesize);
3960	WRITE_ONCE(tp->write_seq, tcb->end_seq);
3961	tp->packets_out += tcp_skb_pcount(skb);
3962}
3963
3964/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3965 * queue a data-only packet after the regular SYN, such that regular SYNs
3966 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3967 * only the SYN sequence, the data are retransmitted in the first ACK.
3968 * If cookie is not cached or other error occurs, falls back to send a
3969 * regular SYN with Fast Open cookie request option.
3970 */
3971static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3972{
3973	struct inet_connection_sock *icsk = inet_csk(sk);
3974	struct tcp_sock *tp = tcp_sk(sk);
3975	struct tcp_fastopen_request *fo = tp->fastopen_req;
3976	struct page_frag *pfrag = sk_page_frag(sk);
3977	struct sk_buff *syn_data;
3978	int space, err = 0;
3979
3980	tp->rx_opt.mss_clamp = tp->advmss;  /* If MSS is not cached */
3981	if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3982		goto fallback;
3983
3984	/* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3985	 * user-MSS. Reserve maximum option space for middleboxes that add
3986	 * private TCP options. The cost is reduced data space in SYN :(
3987	 */
3988	tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3989	/* Sync mss_cache after updating the mss_clamp */
3990	tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3991
3992	space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3993		MAX_TCP_OPTION_SPACE;
3994
3995	space = min_t(size_t, space, fo->size);
3996
3997	if (space &&
3998	    !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
3999				  pfrag, sk->sk_allocation))
4000		goto fallback;
4001	syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
4002	if (!syn_data)
4003		goto fallback;
4004	memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
4005	if (space) {
4006		space = min_t(size_t, space, pfrag->size - pfrag->offset);
4007		space = tcp_wmem_schedule(sk, space);
4008	}
4009	if (space) {
4010		space = copy_page_from_iter(pfrag->page, pfrag->offset,
4011					    space, &fo->data->msg_iter);
4012		if (unlikely(!space)) {
4013			tcp_skb_tsorted_anchor_cleanup(syn_data);
4014			kfree_skb(syn_data);
4015			goto fallback;
4016		}
4017		skb_fill_page_desc(syn_data, 0, pfrag->page,
4018				   pfrag->offset, space);
4019		page_ref_inc(pfrag->page);
4020		pfrag->offset += space;
4021		skb_len_add(syn_data, space);
4022		skb_zcopy_set(syn_data, fo->uarg, NULL);
4023	}
4024	/* No more data pending in inet_wait_for_connect() */
4025	if (space == fo->size)
4026		fo->data = NULL;
4027	fo->copied = space;
4028
4029	tcp_connect_queue_skb(sk, syn_data);
4030	if (syn_data->len)
4031		tcp_chrono_start(sk, TCP_CHRONO_BUSY);
4032
4033	err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
4034
4035	skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, SKB_CLOCK_MONOTONIC);
4036
4037	/* Now full SYN+DATA was cloned and sent (or not),
4038	 * remove the SYN from the original skb (syn_data)
4039	 * we keep in write queue in case of a retransmit, as we
4040	 * also have the SYN packet (with no data) in the same queue.
4041	 */
4042	TCP_SKB_CB(syn_data)->seq++;
4043	TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4044	if (!err) {
4045		tp->syn_data = (fo->copied > 0);
4046		tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
4047		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4048		goto done;
4049	}
4050
4051	/* data was not sent, put it in write_queue */
4052	__skb_queue_tail(&sk->sk_write_queue, syn_data);
4053	tp->packets_out -= tcp_skb_pcount(syn_data);
4054
4055fallback:
4056	/* Send a regular SYN with Fast Open cookie request option */
4057	if (fo->cookie.len > 0)
4058		fo->cookie.len = 0;
4059	err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
4060	if (err)
4061		tp->syn_fastopen = 0;
4062done:
4063	fo->cookie.len = -1;  /* Exclude Fast Open option for SYN retries */
4064	return err;
4065}
4066
4067/* Build a SYN and send it off. */
4068int tcp_connect(struct sock *sk)
4069{
4070	struct tcp_sock *tp = tcp_sk(sk);
4071	struct sk_buff *buff;
4072	int err;
4073
4074	tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
4075
4076#if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4077	/* Has to be checked late, after setting daddr/saddr/ops.
4078	 * Return error if the peer has both a md5 and a tcp-ao key
4079	 * configured as this is ambiguous.
4080	 */
4081	if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4082					       lockdep_sock_is_held(sk)))) {
4083		bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4084		bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4085		struct tcp_ao_info *ao_info;
4086
4087		ao_info = rcu_dereference_check(tp->ao_info,
4088						lockdep_sock_is_held(sk));
4089		if (ao_info) {
4090			/* This is an extra check: tcp_ao_required() in
4091			 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4092			 * md5 keys on ao_required socket.
4093			 */
4094			needs_ao |= ao_info->ao_required;
4095			WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4096		}
4097		if (needs_md5 && needs_ao)
4098			return -EKEYREJECTED;
4099
4100		/* If we have a matching md5 key and no matching tcp-ao key
4101		 * then free up ao_info if allocated.
4102		 */
4103		if (needs_md5) {
4104			tcp_ao_destroy_sock(sk, false);
4105		} else if (needs_ao) {
4106			tcp_clear_md5_list(sk);
4107			kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4108						  lockdep_sock_is_held(sk)));
4109		}
4110	}
4111#endif
4112#ifdef CONFIG_TCP_AO
4113	if (unlikely(rcu_dereference_protected(tp->ao_info,
4114					       lockdep_sock_is_held(sk)))) {
4115		/* Don't allow connecting if ao is configured but no
4116		 * matching key is found.
4117		 */
4118		if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4119			return -EKEYREJECTED;
4120	}
4121#endif
4122
4123	if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4124		return -EHOSTUNREACH; /* Routing failure or similar. */
4125
4126	tcp_connect_init(sk);
4127
4128	if (unlikely(tp->repair)) {
4129		tcp_finish_connect(sk, NULL);
4130		return 0;
4131	}
4132
4133	buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
4134	if (unlikely(!buff))
4135		return -ENOBUFS;
4136
4137	/* SYN eats a sequence byte, write_seq updated by
4138	 * tcp_connect_queue_skb().
4139	 */
4140	tcp_init_nondata_skb(buff, tp->write_seq, TCPHDR_SYN);
4141	tcp_mstamp_refresh(tp);
4142	tp->retrans_stamp = tcp_time_stamp_ts(tp);
4143	tcp_connect_queue_skb(sk, buff);
4144	tcp_ecn_send_syn(sk, buff);
4145	tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
4146
4147	/* Send off SYN; include data in Fast Open. */
4148	err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
4149	      tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
4150	if (err == -ECONNREFUSED)
4151		return err;
4152
4153	/* We change tp->snd_nxt after the tcp_transmit_skb() call
4154	 * in order to make this packet get counted in tcpOutSegs.
4155	 */
4156	WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4157	tp->pushed_seq = tp->write_seq;
4158	buff = tcp_send_head(sk);
4159	if (unlikely(buff)) {
4160		WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4161		tp->pushed_seq	= TCP_SKB_CB(buff)->seq;
4162	}
4163	TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4164
4165	/* Timer for repeating the SYN until an answer. */
4166	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4167				  inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
4168	return 0;
4169}
4170EXPORT_SYMBOL(tcp_connect);
4171
4172u32 tcp_delack_max(const struct sock *sk)
4173{
4174	u32 delack_from_rto_min = max(tcp_rto_min(sk), 2) - 1;
4175
4176	return min(inet_csk(sk)->icsk_delack_max, delack_from_rto_min);
4177}
4178
4179/* Send out a delayed ack, the caller does the policy checking
4180 * to see if we should even be here.  See tcp_input.c:tcp_ack_snd_check()
4181 * for details.
4182 */
4183void tcp_send_delayed_ack(struct sock *sk)
4184{
4185	struct inet_connection_sock *icsk = inet_csk(sk);
4186	int ato = icsk->icsk_ack.ato;
4187	unsigned long timeout;
4188
4189	if (ato > TCP_DELACK_MIN) {
4190		const struct tcp_sock *tp = tcp_sk(sk);
4191		int max_ato = HZ / 2;
4192
4193		if (inet_csk_in_pingpong_mode(sk) ||
4194		    (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4195			max_ato = TCP_DELACK_MAX;
4196
4197		/* Slow path, intersegment interval is "high". */
4198
4199		/* If some rtt estimate is known, use it to bound delayed ack.
4200		 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4201		 * directly.
4202		 */
4203		if (tp->srtt_us) {
4204			int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4205					TCP_DELACK_MIN);
4206
4207			if (rtt < max_ato)
4208				max_ato = rtt;
4209		}
4210
4211		ato = min(ato, max_ato);
4212	}
4213
4214	ato = min_t(u32, ato, tcp_delack_max(sk));
4215
4216	/* Stay within the limit we were given */
4217	timeout = jiffies + ato;
4218
4219	/* Use new timeout only if there wasn't a older one earlier. */
4220	if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4221		/* If delack timer is about to expire, send ACK now. */
4222		if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
4223			tcp_send_ack(sk);
4224			return;
4225		}
4226
4227		if (!time_before(timeout, icsk->icsk_ack.timeout))
4228			timeout = icsk->icsk_ack.timeout;
4229	}
4230	smp_store_release(&icsk->icsk_ack.pending,
4231			  icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
4232	icsk->icsk_ack.timeout = timeout;
4233	sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4234}
4235
4236/* This routine sends an ack and also updates the window. */
4237void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4238{
4239	struct sk_buff *buff;
4240
4241	/* If we have been reset, we may not send again. */
4242	if (sk->sk_state == TCP_CLOSE)
4243		return;
4244
4245	/* We are not putting this on the write queue, so
4246	 * tcp_transmit_skb() will set the ownership to this
4247	 * sock.
4248	 */
4249	buff = alloc_skb(MAX_TCP_HEADER,
4250			 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4251	if (unlikely(!buff)) {
4252		struct inet_connection_sock *icsk = inet_csk(sk);
4253		unsigned long delay;
4254
4255		delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4256		if (delay < TCP_RTO_MAX)
4257			icsk->icsk_ack.retry++;
4258		inet_csk_schedule_ack(sk);
4259		icsk->icsk_ack.ato = TCP_ATO_MIN;
4260		inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
4261		return;
4262	}
4263
4264	/* Reserve space for headers and prepare control bits. */
4265	skb_reserve(buff, MAX_TCP_HEADER);
4266	tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4267
4268	/* We do not want pure acks influencing TCP Small Queues or fq/pacing
4269	 * too much.
4270	 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4271	 */
4272	skb_set_tcp_pure_ack(buff);
4273
4274	/* Send it off, this clears delayed acks for us. */
4275	__tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4276}
4277EXPORT_SYMBOL_GPL(__tcp_send_ack);
4278
4279void tcp_send_ack(struct sock *sk)
4280{
4281	__tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4282}
4283
4284/* This routine sends a packet with an out of date sequence
4285 * number. It assumes the other end will try to ack it.
4286 *
4287 * Question: what should we make while urgent mode?
4288 * 4.4BSD forces sending single byte of data. We cannot send
4289 * out of window data, because we have SND.NXT==SND.MAX...
4290 *
4291 * Current solution: to send TWO zero-length segments in urgent mode:
4292 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4293 * out-of-date with SND.UNA-1 to probe window.
4294 */
4295static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4296{
4297	struct tcp_sock *tp = tcp_sk(sk);
4298	struct sk_buff *skb;
4299
4300	/* We don't queue it, tcp_transmit_skb() sets ownership. */
4301	skb = alloc_skb(MAX_TCP_HEADER,
4302			sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4303	if (!skb)
4304		return -1;
4305
4306	/* Reserve space for headers and set control bits. */
4307	skb_reserve(skb, MAX_TCP_HEADER);
4308	/* Use a previous sequence.  This should cause the other
4309	 * end to send an ack.  Don't queue or clone SKB, just
4310	 * send it.
4311	 */
4312	tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4313	NET_INC_STATS(sock_net(sk), mib);
4314	return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4315}
4316
4317/* Called from setsockopt( ... TCP_REPAIR ) */
4318void tcp_send_window_probe(struct sock *sk)
4319{
4320	if (sk->sk_state == TCP_ESTABLISHED) {
4321		tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4322		tcp_mstamp_refresh(tcp_sk(sk));
4323		tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4324	}
4325}
4326
4327/* Initiate keepalive or window probe from timer. */
4328int tcp_write_wakeup(struct sock *sk, int mib)
4329{
4330	struct tcp_sock *tp = tcp_sk(sk);
4331	struct sk_buff *skb;
4332
4333	if (sk->sk_state == TCP_CLOSE)
4334		return -1;
4335
4336	skb = tcp_send_head(sk);
4337	if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4338		int err;
4339		unsigned int mss = tcp_current_mss(sk);
4340		unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4341
4342		if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4343			tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4344
4345		/* We are probing the opening of a window
4346		 * but the window size is != 0
4347		 * must have been a result SWS avoidance ( sender )
4348		 */
4349		if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4350		    skb->len > mss) {
4351			seg_size = min(seg_size, mss);
4352			TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4353			if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4354					 skb, seg_size, mss, GFP_ATOMIC))
4355				return -1;
4356		} else if (!tcp_skb_pcount(skb))
4357			tcp_set_skb_tso_segs(skb, mss);
4358
4359		TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4360		err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4361		if (!err)
4362			tcp_event_new_data_sent(sk, skb);
4363		return err;
4364	} else {
4365		if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4366			tcp_xmit_probe_skb(sk, 1, mib);
4367		return tcp_xmit_probe_skb(sk, 0, mib);
4368	}
4369}
4370
4371/* A window probe timeout has occurred.  If window is not closed send
4372 * a partial packet else a zero probe.
4373 */
4374void tcp_send_probe0(struct sock *sk)
4375{
4376	struct inet_connection_sock *icsk = inet_csk(sk);
4377	struct tcp_sock *tp = tcp_sk(sk);
4378	struct net *net = sock_net(sk);
4379	unsigned long timeout;
4380	int err;
4381
4382	err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4383
4384	if (tp->packets_out || tcp_write_queue_empty(sk)) {
4385		/* Cancel probe timer, if it is not required. */
4386		icsk->icsk_probes_out = 0;
4387		icsk->icsk_backoff = 0;
4388		icsk->icsk_probes_tstamp = 0;
4389		return;
4390	}
4391
4392	icsk->icsk_probes_out++;
4393	if (err <= 0) {
4394		if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4395			icsk->icsk_backoff++;
4396		timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4397	} else {
4398		/* If packet was not sent due to local congestion,
4399		 * Let senders fight for local resources conservatively.
4400		 */
4401		timeout = TCP_RESOURCE_PROBE_INTERVAL;
4402	}
4403
4404	timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4405	tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4406}
4407
4408int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4409{
4410	const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4411	struct flowi fl;
4412	int res;
4413
4414	/* Paired with WRITE_ONCE() in sock_setsockopt() */
4415	if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4416		WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4417	res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4418				  NULL);
4419	if (!res) {
4420		TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4421		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4422		if (unlikely(tcp_passive_fastopen(sk))) {
4423			/* sk has const attribute because listeners are lockless.
4424			 * However in this case, we are dealing with a passive fastopen
4425			 * socket thus we can change total_retrans value.
4426			 */
4427			tcp_sk_rw(sk)->total_retrans++;
4428		}
4429		trace_tcp_retransmit_synack(sk, req);
4430	}
4431	return res;
4432}
4433EXPORT_SYMBOL(tcp_rtx_synack);