Linux Audio

Check our new training course

Loading...
v6.13.7
   1// SPDX-License-Identifier: GPL-2.0-or-later
   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 * Fixes:
  22 *		Alan Cox	:	Numerous verify_area() calls
  23 *		Alan Cox	:	Set the ACK bit on a reset
  24 *		Alan Cox	:	Stopped it crashing if it closed while
  25 *					sk->inuse=1 and was trying to connect
  26 *					(tcp_err()).
  27 *		Alan Cox	:	All icmp error handling was broken
  28 *					pointers passed where wrong and the
  29 *					socket was looked up backwards. Nobody
  30 *					tested any icmp error code obviously.
  31 *		Alan Cox	:	tcp_err() now handled properly. It
  32 *					wakes people on errors. poll
  33 *					behaves and the icmp error race
  34 *					has gone by moving it into sock.c
  35 *		Alan Cox	:	tcp_send_reset() fixed to work for
  36 *					everything not just packets for
  37 *					unknown sockets.
  38 *		Alan Cox	:	tcp option processing.
  39 *		Alan Cox	:	Reset tweaked (still not 100%) [Had
  40 *					syn rule wrong]
  41 *		Herp Rosmanith  :	More reset fixes
  42 *		Alan Cox	:	No longer acks invalid rst frames.
  43 *					Acking any kind of RST is right out.
  44 *		Alan Cox	:	Sets an ignore me flag on an rst
  45 *					receive otherwise odd bits of prattle
  46 *					escape still
  47 *		Alan Cox	:	Fixed another acking RST frame bug.
  48 *					Should stop LAN workplace lockups.
  49 *		Alan Cox	: 	Some tidyups using the new skb list
  50 *					facilities
  51 *		Alan Cox	:	sk->keepopen now seems to work
  52 *		Alan Cox	:	Pulls options out correctly on accepts
  53 *		Alan Cox	:	Fixed assorted sk->rqueue->next errors
  54 *		Alan Cox	:	PSH doesn't end a TCP read. Switched a
  55 *					bit to skb ops.
  56 *		Alan Cox	:	Tidied tcp_data to avoid a potential
  57 *					nasty.
  58 *		Alan Cox	:	Added some better commenting, as the
  59 *					tcp is hard to follow
  60 *		Alan Cox	:	Removed incorrect check for 20 * psh
  61 *	Michael O'Reilly	:	ack < copied bug fix.
  62 *	Johannes Stille		:	Misc tcp fixes (not all in yet).
  63 *		Alan Cox	:	FIN with no memory -> CRASH
  64 *		Alan Cox	:	Added socket option proto entries.
  65 *					Also added awareness of them to accept.
  66 *		Alan Cox	:	Added TCP options (SOL_TCP)
  67 *		Alan Cox	:	Switched wakeup calls to callbacks,
  68 *					so the kernel can layer network
  69 *					sockets.
  70 *		Alan Cox	:	Use ip_tos/ip_ttl settings.
  71 *		Alan Cox	:	Handle FIN (more) properly (we hope).
  72 *		Alan Cox	:	RST frames sent on unsynchronised
  73 *					state ack error.
  74 *		Alan Cox	:	Put in missing check for SYN bit.
  75 *		Alan Cox	:	Added tcp_select_window() aka NET2E
  76 *					window non shrink trick.
  77 *		Alan Cox	:	Added a couple of small NET2E timer
  78 *					fixes
  79 *		Charles Hedrick :	TCP fixes
  80 *		Toomas Tamm	:	TCP window fixes
  81 *		Alan Cox	:	Small URG fix to rlogin ^C ack fight
  82 *		Charles Hedrick	:	Rewrote most of it to actually work
  83 *		Linus		:	Rewrote tcp_read() and URG handling
  84 *					completely
  85 *		Gerhard Koerting:	Fixed some missing timer handling
  86 *		Matthew Dillon  :	Reworked TCP machine states as per RFC
  87 *		Gerhard Koerting:	PC/TCP workarounds
  88 *		Adam Caldwell	:	Assorted timer/timing errors
  89 *		Matthew Dillon	:	Fixed another RST bug
  90 *		Alan Cox	:	Move to kernel side addressing changes.
  91 *		Alan Cox	:	Beginning work on TCP fastpathing
  92 *					(not yet usable)
  93 *		Arnt Gulbrandsen:	Turbocharged tcp_check() routine.
  94 *		Alan Cox	:	TCP fast path debugging
  95 *		Alan Cox	:	Window clamping
  96 *		Michael Riepe	:	Bug in tcp_check()
  97 *		Matt Dillon	:	More TCP improvements and RST bug fixes
  98 *		Matt Dillon	:	Yet more small nasties remove from the
  99 *					TCP code (Be very nice to this man if
 100 *					tcp finally works 100%) 8)
 101 *		Alan Cox	:	BSD accept semantics.
 102 *		Alan Cox	:	Reset on closedown bug.
 103 *	Peter De Schrijver	:	ENOTCONN check missing in tcp_sendto().
 104 *		Michael Pall	:	Handle poll() after URG properly in
 105 *					all cases.
 106 *		Michael Pall	:	Undo the last fix in tcp_read_urg()
 107 *					(multi URG PUSH broke rlogin).
 108 *		Michael Pall	:	Fix the multi URG PUSH problem in
 109 *					tcp_readable(), poll() after URG
 110 *					works now.
 111 *		Michael Pall	:	recv(...,MSG_OOB) never blocks in the
 112 *					BSD api.
 113 *		Alan Cox	:	Changed the semantics of sk->socket to
 114 *					fix a race and a signal problem with
 115 *					accept() and async I/O.
 116 *		Alan Cox	:	Relaxed the rules on tcp_sendto().
 117 *		Yury Shevchuk	:	Really fixed accept() blocking problem.
 118 *		Craig I. Hagan  :	Allow for BSD compatible TIME_WAIT for
 119 *					clients/servers which listen in on
 120 *					fixed ports.
 121 *		Alan Cox	:	Cleaned the above up and shrank it to
 122 *					a sensible code size.
 123 *		Alan Cox	:	Self connect lockup fix.
 124 *		Alan Cox	:	No connect to multicast.
 125 *		Ross Biro	:	Close unaccepted children on master
 126 *					socket close.
 127 *		Alan Cox	:	Reset tracing code.
 128 *		Alan Cox	:	Spurious resets on shutdown.
 129 *		Alan Cox	:	Giant 15 minute/60 second timer error
 130 *		Alan Cox	:	Small whoops in polling before an
 131 *					accept.
 132 *		Alan Cox	:	Kept the state trace facility since
 133 *					it's handy for debugging.
 134 *		Alan Cox	:	More reset handler fixes.
 135 *		Alan Cox	:	Started rewriting the code based on
 136 *					the RFC's for other useful protocol
 137 *					references see: Comer, KA9Q NOS, and
 138 *					for a reference on the difference
 139 *					between specifications and how BSD
 140 *					works see the 4.4lite source.
 141 *		A.N.Kuznetsov	:	Don't time wait on completion of tidy
 142 *					close.
 143 *		Linus Torvalds	:	Fin/Shutdown & copied_seq changes.
 144 *		Linus Torvalds	:	Fixed BSD port reuse to work first syn
 145 *		Alan Cox	:	Reimplemented timers as per the RFC
 146 *					and using multiple timers for sanity.
 147 *		Alan Cox	:	Small bug fixes, and a lot of new
 148 *					comments.
 149 *		Alan Cox	:	Fixed dual reader crash by locking
 150 *					the buffers (much like datagram.c)
 151 *		Alan Cox	:	Fixed stuck sockets in probe. A probe
 152 *					now gets fed up of retrying without
 153 *					(even a no space) answer.
 154 *		Alan Cox	:	Extracted closing code better
 155 *		Alan Cox	:	Fixed the closing state machine to
 156 *					resemble the RFC.
 157 *		Alan Cox	:	More 'per spec' fixes.
 158 *		Jorge Cwik	:	Even faster checksumming.
 159 *		Alan Cox	:	tcp_data() doesn't ack illegal PSH
 160 *					only frames. At least one pc tcp stack
 161 *					generates them.
 162 *		Alan Cox	:	Cache last socket.
 163 *		Alan Cox	:	Per route irtt.
 164 *		Matt Day	:	poll()->select() match BSD precisely on error
 165 *		Alan Cox	:	New buffers
 166 *		Marc Tamsky	:	Various sk->prot->retransmits and
 167 *					sk->retransmits misupdating fixed.
 168 *					Fixed tcp_write_timeout: stuck close,
 169 *					and TCP syn retries gets used now.
 170 *		Mark Yarvis	:	In tcp_read_wakeup(), don't send an
 171 *					ack if state is TCP_CLOSED.
 172 *		Alan Cox	:	Look up device on a retransmit - routes may
 173 *					change. Doesn't yet cope with MSS shrink right
 174 *					but it's a start!
 175 *		Marc Tamsky	:	Closing in closing fixes.
 176 *		Mike Shaver	:	RFC1122 verifications.
 177 *		Alan Cox	:	rcv_saddr errors.
 178 *		Alan Cox	:	Block double connect().
 179 *		Alan Cox	:	Small hooks for enSKIP.
 180 *		Alexey Kuznetsov:	Path MTU discovery.
 181 *		Alan Cox	:	Support soft errors.
 182 *		Alan Cox	:	Fix MTU discovery pathological case
 183 *					when the remote claims no mtu!
 184 *		Marc Tamsky	:	TCP_CLOSE fix.
 185 *		Colin (G3TNE)	:	Send a reset on syn ack replies in
 186 *					window but wrong (fixes NT lpd problems)
 187 *		Pedro Roque	:	Better TCP window handling, delayed ack.
 188 *		Joerg Reuter	:	No modification of locked buffers in
 189 *					tcp_do_retransmit()
 190 *		Eric Schenk	:	Changed receiver side silly window
 191 *					avoidance algorithm to BSD style
 192 *					algorithm. This doubles throughput
 193 *					against machines running Solaris,
 194 *					and seems to result in general
 195 *					improvement.
 196 *	Stefan Magdalinski	:	adjusted tcp_readable() to fix FIONREAD
 197 *	Willy Konynenberg	:	Transparent proxying support.
 198 *	Mike McLagan		:	Routing by source
 199 *		Keith Owens	:	Do proper merging with partial SKB's in
 200 *					tcp_do_sendmsg to avoid burstiness.
 201 *		Eric Schenk	:	Fix fast close down bug with
 202 *					shutdown() followed by close().
 203 *		Andi Kleen 	:	Make poll agree with SIGIO
 204 *	Salvatore Sanfilippo	:	Support SO_LINGER with linger == 1 and
 205 *					lingertime == 0 (RFC 793 ABORT Call)
 206 *	Hirokazu Takahashi	:	Use copy_from_user() instead of
 207 *					csum_and_copy_from_user() if possible.
 208 *
 209 * Description of States:
 210 *
 211 *	TCP_SYN_SENT		sent a connection request, waiting for ack
 212 *
 213 *	TCP_SYN_RECV		received a connection request, sent ack,
 214 *				waiting for final ack in three-way handshake.
 215 *
 216 *	TCP_ESTABLISHED		connection established
 217 *
 218 *	TCP_FIN_WAIT1		our side has shutdown, waiting to complete
 219 *				transmission of remaining buffered data
 220 *
 221 *	TCP_FIN_WAIT2		all buffered data sent, waiting for remote
 222 *				to shutdown
 223 *
 224 *	TCP_CLOSING		both sides have shutdown but we still have
 225 *				data we have to finish sending
 226 *
 227 *	TCP_TIME_WAIT		timeout to catch resent junk before entering
 228 *				closed, can only be entered from FIN_WAIT2
 229 *				or CLOSING.  Required because the other end
 230 *				may not have gotten our last ACK causing it
 231 *				to retransmit the data packet (which we ignore)
 232 *
 233 *	TCP_CLOSE_WAIT		remote side has shutdown and is waiting for
 234 *				us to finish writing our data and to shutdown
 235 *				(we have to close() to move on to LAST_ACK)
 236 *
 237 *	TCP_LAST_ACK		out side has shutdown after remote has
 238 *				shutdown.  There may still be data in our
 239 *				buffer that we have to finish sending
 240 *
 241 *	TCP_CLOSE		socket is finished
 242 */
 243
 244#define pr_fmt(fmt) "TCP: " fmt
 245
 246#include <crypto/hash.h>
 247#include <linux/kernel.h>
 248#include <linux/module.h>
 249#include <linux/types.h>
 250#include <linux/fcntl.h>
 251#include <linux/poll.h>
 252#include <linux/inet_diag.h>
 253#include <linux/init.h>
 254#include <linux/fs.h>
 255#include <linux/skbuff.h>
 256#include <linux/scatterlist.h>
 257#include <linux/splice.h>
 258#include <linux/net.h>
 259#include <linux/socket.h>
 260#include <linux/random.h>
 261#include <linux/memblock.h>
 262#include <linux/highmem.h>
 
 263#include <linux/cache.h>
 264#include <linux/err.h>
 265#include <linux/time.h>
 266#include <linux/slab.h>
 267#include <linux/errqueue.h>
 268#include <linux/static_key.h>
 269#include <linux/btf.h>
 270
 271#include <net/icmp.h>
 272#include <net/inet_common.h>
 273#include <net/tcp.h>
 274#include <net/mptcp.h>
 275#include <net/proto_memory.h>
 276#include <net/xfrm.h>
 277#include <net/ip.h>
 278#include <net/sock.h>
 279#include <net/rstreason.h>
 280
 281#include <linux/uaccess.h>
 282#include <asm/ioctls.h>
 283#include <net/busy_poll.h>
 284#include <net/hotdata.h>
 285#include <trace/events/tcp.h>
 286#include <net/rps.h>
 287
 288#include "../core/devmem.h"
 289
 290/* Track pending CMSGs. */
 291enum {
 292	TCP_CMSG_INQ = 1,
 293	TCP_CMSG_TS = 2
 294};
 295
 296DEFINE_PER_CPU(unsigned int, tcp_orphan_count);
 297EXPORT_PER_CPU_SYMBOL_GPL(tcp_orphan_count);
 298
 299DEFINE_PER_CPU(u32, tcp_tw_isn);
 300EXPORT_PER_CPU_SYMBOL_GPL(tcp_tw_isn);
 301
 302long sysctl_tcp_mem[3] __read_mostly;
 303EXPORT_SYMBOL(sysctl_tcp_mem);
 304
 305atomic_long_t tcp_memory_allocated ____cacheline_aligned_in_smp;	/* Current allocated memory. */
 306EXPORT_SYMBOL(tcp_memory_allocated);
 307DEFINE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
 308EXPORT_PER_CPU_SYMBOL_GPL(tcp_memory_per_cpu_fw_alloc);
 309
 310#if IS_ENABLED(CONFIG_SMC)
 311DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 312EXPORT_SYMBOL(tcp_have_smc);
 313#endif
 314
 315/*
 316 * Current number of TCP sockets.
 317 */
 318struct percpu_counter tcp_sockets_allocated ____cacheline_aligned_in_smp;
 319EXPORT_SYMBOL(tcp_sockets_allocated);
 320
 321/*
 322 * TCP splice context
 323 */
 324struct tcp_splice_state {
 325	struct pipe_inode_info *pipe;
 326	size_t len;
 327	unsigned int flags;
 328};
 329
 330/*
 331 * Pressure flag: try to collapse.
 332 * Technical note: it is used by multiple contexts non atomically.
 333 * All the __sk_mem_schedule() is of this nature: accounting
 334 * is strict, actions are advisory and have some latency.
 335 */
 336unsigned long tcp_memory_pressure __read_mostly;
 337EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 338
 
 
 
 
 
 339void tcp_enter_memory_pressure(struct sock *sk)
 340{
 341	unsigned long val;
 342
 343	if (READ_ONCE(tcp_memory_pressure))
 344		return;
 345	val = jiffies;
 346
 347	if (!val)
 348		val--;
 349	if (!cmpxchg(&tcp_memory_pressure, 0, val))
 350		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 351}
 352EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 353
 354void tcp_leave_memory_pressure(struct sock *sk)
 355{
 356	unsigned long val;
 357
 358	if (!READ_ONCE(tcp_memory_pressure))
 359		return;
 360	val = xchg(&tcp_memory_pressure, 0);
 361	if (val)
 362		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 363			      jiffies_to_msecs(jiffies - val));
 364}
 365EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 366
 367/* Convert seconds to retransmits based on initial and max timeout */
 368static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 369{
 370	u8 res = 0;
 371
 372	if (seconds > 0) {
 373		int period = timeout;
 374
 375		res = 1;
 376		while (seconds > period && res < 255) {
 377			res++;
 378			timeout <<= 1;
 379			if (timeout > rto_max)
 380				timeout = rto_max;
 381			period += timeout;
 382		}
 383	}
 384	return res;
 385}
 386
 387/* Convert retransmits to seconds based on initial and max timeout */
 388static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 389{
 390	int period = 0;
 391
 392	if (retrans > 0) {
 393		period = timeout;
 394		while (--retrans) {
 395			timeout <<= 1;
 396			if (timeout > rto_max)
 397				timeout = rto_max;
 398			period += timeout;
 399		}
 400	}
 401	return period;
 402}
 403
 404static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 405{
 406	u32 rate = READ_ONCE(tp->rate_delivered);
 407	u32 intv = READ_ONCE(tp->rate_interval_us);
 408	u64 rate64 = 0;
 409
 410	if (rate && intv) {
 411		rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 412		do_div(rate64, intv);
 413	}
 414	return rate64;
 415}
 416
 417/* Address-family independent initialization for a tcp_sock.
 418 *
 419 * NOTE: A lot of things set to zero explicitly by call to
 420 *       sk_alloc() so need not be done here.
 421 */
 422void tcp_init_sock(struct sock *sk)
 423{
 424	struct inet_connection_sock *icsk = inet_csk(sk);
 425	struct tcp_sock *tp = tcp_sk(sk);
 426	int rto_min_us;
 427
 428	tp->out_of_order_queue = RB_ROOT;
 429	sk->tcp_rtx_queue = RB_ROOT;
 430	tcp_init_xmit_timers(sk);
 431	INIT_LIST_HEAD(&tp->tsq_node);
 432	INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 433
 434	icsk->icsk_rto = TCP_TIMEOUT_INIT;
 435	rto_min_us = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rto_min_us);
 436	icsk->icsk_rto_min = usecs_to_jiffies(rto_min_us);
 437	icsk->icsk_delack_max = TCP_DELACK_MAX;
 438	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 439	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 440
 441	/* So many TCP implementations out there (incorrectly) count the
 442	 * initial SYN frame in their delayed-ACK and congestion control
 443	 * algorithms that we must have the following bandaid to talk
 444	 * efficiently to them.  -DaveM
 445	 */
 446	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
 447
 448	/* There's a bubble in the pipe until at least the first ACK. */
 449	tp->app_limited = ~0U;
 450	tp->rate_app_limited = 1;
 451
 452	/* See draft-stevens-tcpca-spec-01 for discussion of the
 453	 * initialization of these values.
 454	 */
 455	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 456	tp->snd_cwnd_clamp = ~0;
 457	tp->mss_cache = TCP_MSS_DEFAULT;
 458
 459	tp->reordering = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_reordering);
 460	tcp_assign_congestion_control(sk);
 461
 462	tp->tsoffset = 0;
 463	tp->rack.reo_wnd_steps = 1;
 464
 
 
 465	sk->sk_write_space = sk_stream_write_space;
 466	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 467
 468	icsk->icsk_sync_mss = tcp_sync_mss;
 469
 470	WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[1]));
 471	WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[1]));
 472	tcp_scaling_ratio_init(sk);
 473
 474	set_bit(SOCK_SUPPORT_ZC, &sk->sk_socket->flags);
 475	sk_sockets_allocated_inc(sk);
 476	xa_init_flags(&sk->sk_user_frags, XA_FLAGS_ALLOC1);
 477}
 478EXPORT_SYMBOL(tcp_init_sock);
 479
 480static void tcp_tx_timestamp(struct sock *sk, struct sockcm_cookie *sockc)
 481{
 482	struct sk_buff *skb = tcp_write_queue_tail(sk);
 483	u32 tsflags = sockc->tsflags;
 484
 485	if (tsflags && skb) {
 486		struct skb_shared_info *shinfo = skb_shinfo(skb);
 487		struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 488
 489		sock_tx_timestamp(sk, sockc, &shinfo->tx_flags);
 490		if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 491			tcb->txstamp_ack = 1;
 492		if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 493			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 494	}
 495}
 496
 497static bool tcp_stream_is_readable(struct sock *sk, int target)
 
 498{
 499	if (tcp_epollin_ready(sk, target))
 500		return true;
 501	return sk_is_readable(sk);
 502}
 503
 504/*
 505 *	Wait for a TCP event.
 506 *
 507 *	Note that we don't need to lock the socket, as the upper poll layers
 508 *	take care of normal races (between the test and the event) and we don't
 509 *	go look at any of the socket buffers directly.
 510 */
 511__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 512{
 513	__poll_t mask;
 514	struct sock *sk = sock->sk;
 515	const struct tcp_sock *tp = tcp_sk(sk);
 516	u8 shutdown;
 517	int state;
 518
 519	sock_poll_wait(file, sock, wait);
 520
 521	state = inet_sk_state_load(sk);
 522	if (state == TCP_LISTEN)
 523		return inet_csk_listen_poll(sk);
 524
 525	/* Socket is not locked. We are protected from async events
 526	 * by poll logic and correct handling of state changes
 527	 * made by other threads is impossible in any case.
 528	 */
 529
 530	mask = 0;
 531
 532	/*
 533	 * EPOLLHUP is certainly not done right. But poll() doesn't
 534	 * have a notion of HUP in just one direction, and for a
 535	 * socket the read side is more interesting.
 536	 *
 537	 * Some poll() documentation says that EPOLLHUP is incompatible
 538	 * with the EPOLLOUT/POLLWR flags, so somebody should check this
 539	 * all. But careful, it tends to be safer to return too many
 540	 * bits than too few, and you can easily break real applications
 541	 * if you don't tell them that something has hung up!
 542	 *
 543	 * Check-me.
 544	 *
 545	 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 546	 * our fs/select.c). It means that after we received EOF,
 547	 * poll always returns immediately, making impossible poll() on write()
 548	 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 549	 * if and only if shutdown has been made in both directions.
 550	 * Actually, it is interesting to look how Solaris and DUX
 551	 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 552	 * then we could set it on SND_SHUTDOWN. BTW examples given
 553	 * in Stevens' books assume exactly this behaviour, it explains
 554	 * why EPOLLHUP is incompatible with EPOLLOUT.	--ANK
 555	 *
 556	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 557	 * blocking on fresh not-connected or disconnected socket. --ANK
 558	 */
 559	shutdown = READ_ONCE(sk->sk_shutdown);
 560	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 561		mask |= EPOLLHUP;
 562	if (shutdown & RCV_SHUTDOWN)
 563		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 564
 565	/* Connected or passive Fast Open socket? */
 566	if (state != TCP_SYN_SENT &&
 567	    (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
 568		int target = sock_rcvlowat(sk, 0, INT_MAX);
 569		u16 urg_data = READ_ONCE(tp->urg_data);
 570
 571		if (unlikely(urg_data) &&
 572		    READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
 573		    !sock_flag(sk, SOCK_URGINLINE))
 574			target++;
 575
 576		if (tcp_stream_is_readable(sk, target))
 577			mask |= EPOLLIN | EPOLLRDNORM;
 578
 579		if (!(shutdown & SEND_SHUTDOWN)) {
 580			if (__sk_stream_is_writeable(sk, 1)) {
 581				mask |= EPOLLOUT | EPOLLWRNORM;
 582			} else {  /* send SIGIO later */
 583				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 584				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 585
 586				/* Race breaker. If space is freed after
 587				 * wspace test but before the flags are set,
 588				 * IO signal will be lost. Memory barrier
 589				 * pairs with the input side.
 590				 */
 591				smp_mb__after_atomic();
 592				if (__sk_stream_is_writeable(sk, 1))
 593					mask |= EPOLLOUT | EPOLLWRNORM;
 594			}
 595		} else
 596			mask |= EPOLLOUT | EPOLLWRNORM;
 597
 598		if (urg_data & TCP_URG_VALID)
 599			mask |= EPOLLPRI;
 600	} else if (state == TCP_SYN_SENT &&
 601		   inet_test_bit(DEFER_CONNECT, sk)) {
 602		/* Active TCP fastopen socket with defer_connect
 603		 * Return EPOLLOUT so application can call write()
 604		 * in order for kernel to generate SYN+data
 605		 */
 606		mask |= EPOLLOUT | EPOLLWRNORM;
 607	}
 608	/* This barrier is coupled with smp_wmb() in tcp_done_with_error() */
 609	smp_rmb();
 610	if (READ_ONCE(sk->sk_err) ||
 611	    !skb_queue_empty_lockless(&sk->sk_error_queue))
 612		mask |= EPOLLERR;
 613
 614	return mask;
 615}
 616EXPORT_SYMBOL(tcp_poll);
 617
 618int tcp_ioctl(struct sock *sk, int cmd, int *karg)
 619{
 620	struct tcp_sock *tp = tcp_sk(sk);
 621	int answ;
 622	bool slow;
 623
 624	switch (cmd) {
 625	case SIOCINQ:
 626		if (sk->sk_state == TCP_LISTEN)
 627			return -EINVAL;
 628
 629		slow = lock_sock_fast(sk);
 630		answ = tcp_inq(sk);
 631		unlock_sock_fast(sk, slow);
 632		break;
 633	case SIOCATMARK:
 634		answ = READ_ONCE(tp->urg_data) &&
 635		       READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
 636		break;
 637	case SIOCOUTQ:
 638		if (sk->sk_state == TCP_LISTEN)
 639			return -EINVAL;
 640
 641		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 642			answ = 0;
 643		else
 644			answ = READ_ONCE(tp->write_seq) - tp->snd_una;
 645		break;
 646	case SIOCOUTQNSD:
 647		if (sk->sk_state == TCP_LISTEN)
 648			return -EINVAL;
 649
 650		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 651			answ = 0;
 652		else
 653			answ = READ_ONCE(tp->write_seq) -
 654			       READ_ONCE(tp->snd_nxt);
 655		break;
 656	default:
 657		return -ENOIOCTLCMD;
 658	}
 659
 660	*karg = answ;
 661	return 0;
 662}
 663EXPORT_SYMBOL(tcp_ioctl);
 664
 665void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 666{
 667	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 668	tp->pushed_seq = tp->write_seq;
 669}
 670
 671static inline bool forced_push(const struct tcp_sock *tp)
 672{
 673	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 674}
 675
 676void tcp_skb_entail(struct sock *sk, struct sk_buff *skb)
 677{
 678	struct tcp_sock *tp = tcp_sk(sk);
 679	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 680
 
 681	tcb->seq     = tcb->end_seq = tp->write_seq;
 682	tcb->tcp_flags = TCPHDR_ACK;
 
 683	__skb_header_release(skb);
 684	tcp_add_write_queue_tail(sk, skb);
 685	sk_wmem_queued_add(sk, skb->truesize);
 686	sk_mem_charge(sk, skb->truesize);
 687	if (tp->nonagle & TCP_NAGLE_PUSH)
 688		tp->nonagle &= ~TCP_NAGLE_PUSH;
 689
 690	tcp_slow_start_after_idle_check(sk);
 691}
 692
 693static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 694{
 695	if (flags & MSG_OOB)
 696		tp->snd_up = tp->write_seq;
 697}
 698
 699/* If a not yet filled skb is pushed, do not send it if
 700 * we have data packets in Qdisc or NIC queues :
 701 * Because TX completion will happen shortly, it gives a chance
 702 * to coalesce future sendmsg() payload into this skb, without
 703 * need for a timer, and with no latency trade off.
 704 * As packets containing data payload have a bigger truesize
 705 * than pure acks (dataless) packets, the last checks prevent
 706 * autocorking if we only have an ACK in Qdisc/NIC queues,
 707 * or if TX completion was delayed after we processed ACK packet.
 708 */
 709static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 710				int size_goal)
 711{
 712	return skb->len < size_goal &&
 713	       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_autocorking) &&
 714	       !tcp_rtx_queue_empty(sk) &&
 715	       refcount_read(&sk->sk_wmem_alloc) > skb->truesize &&
 716	       tcp_skb_can_collapse_to(skb);
 717}
 718
 719void tcp_push(struct sock *sk, int flags, int mss_now,
 720	      int nonagle, int size_goal)
 721{
 722	struct tcp_sock *tp = tcp_sk(sk);
 723	struct sk_buff *skb;
 724
 725	skb = tcp_write_queue_tail(sk);
 726	if (!skb)
 727		return;
 728	if (!(flags & MSG_MORE) || forced_push(tp))
 729		tcp_mark_push(tp, skb);
 730
 731	tcp_mark_urg(tp, flags);
 732
 733	if (tcp_should_autocork(sk, skb, size_goal)) {
 734
 735		/* avoid atomic op if TSQ_THROTTLED bit is already set */
 736		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 737			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 738			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 739			smp_mb__after_atomic();
 740		}
 741		/* It is possible TX completion already happened
 742		 * before we set TSQ_THROTTLED.
 743		 */
 744		if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 745			return;
 746	}
 747
 748	if (flags & MSG_MORE)
 749		nonagle = TCP_NAGLE_CORK;
 750
 751	__tcp_push_pending_frames(sk, mss_now, nonagle);
 752}
 753
 754static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 755				unsigned int offset, size_t len)
 756{
 757	struct tcp_splice_state *tss = rd_desc->arg.data;
 758	int ret;
 759
 760	ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 761			      min(rd_desc->count, len), tss->flags);
 762	if (ret > 0)
 763		rd_desc->count -= ret;
 764	return ret;
 765}
 766
 767static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 768{
 769	/* Store TCP splice context information in read_descriptor_t. */
 770	read_descriptor_t rd_desc = {
 771		.arg.data = tss,
 772		.count	  = tss->len,
 773	};
 774
 775	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 776}
 777
 778/**
 779 *  tcp_splice_read - splice data from TCP socket to a pipe
 780 * @sock:	socket to splice from
 781 * @ppos:	position (not valid)
 782 * @pipe:	pipe to splice to
 783 * @len:	number of bytes to splice
 784 * @flags:	splice modifier flags
 785 *
 786 * Description:
 787 *    Will read pages from given socket and fill them into a pipe.
 788 *
 789 **/
 790ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 791			struct pipe_inode_info *pipe, size_t len,
 792			unsigned int flags)
 793{
 794	struct sock *sk = sock->sk;
 795	struct tcp_splice_state tss = {
 796		.pipe = pipe,
 797		.len = len,
 798		.flags = flags,
 799	};
 800	long timeo;
 801	ssize_t spliced;
 802	int ret;
 803
 804	sock_rps_record_flow(sk);
 805	/*
 806	 * We can't seek on a socket input
 807	 */
 808	if (unlikely(*ppos))
 809		return -ESPIPE;
 810
 811	ret = spliced = 0;
 812
 813	lock_sock(sk);
 814
 815	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 816	while (tss.len) {
 817		ret = __tcp_splice_read(sk, &tss);
 818		if (ret < 0)
 819			break;
 820		else if (!ret) {
 821			if (spliced)
 822				break;
 823			if (sock_flag(sk, SOCK_DONE))
 824				break;
 825			if (sk->sk_err) {
 826				ret = sock_error(sk);
 827				break;
 828			}
 829			if (sk->sk_shutdown & RCV_SHUTDOWN)
 830				break;
 831			if (sk->sk_state == TCP_CLOSE) {
 832				/*
 833				 * This occurs when user tries to read
 834				 * from never connected socket.
 835				 */
 836				ret = -ENOTCONN;
 837				break;
 838			}
 839			if (!timeo) {
 840				ret = -EAGAIN;
 841				break;
 842			}
 843			/* if __tcp_splice_read() got nothing while we have
 844			 * an skb in receive queue, we do not want to loop.
 845			 * This might happen with URG data.
 846			 */
 847			if (!skb_queue_empty(&sk->sk_receive_queue))
 848				break;
 849			ret = sk_wait_data(sk, &timeo, NULL);
 850			if (ret < 0)
 851				break;
 852			if (signal_pending(current)) {
 853				ret = sock_intr_errno(timeo);
 854				break;
 855			}
 856			continue;
 857		}
 858		tss.len -= ret;
 859		spliced += ret;
 860
 861		if (!tss.len || !timeo)
 862			break;
 863		release_sock(sk);
 864		lock_sock(sk);
 865
 866		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 867		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
 868		    signal_pending(current))
 869			break;
 870	}
 871
 872	release_sock(sk);
 873
 874	if (spliced)
 875		return spliced;
 876
 877	return ret;
 878}
 879EXPORT_SYMBOL(tcp_splice_read);
 880
 881struct sk_buff *tcp_stream_alloc_skb(struct sock *sk, gfp_t gfp,
 882				     bool force_schedule)
 883{
 884	struct sk_buff *skb;
 885
 886	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 887	if (likely(skb)) {
 888		bool mem_scheduled;
 889
 890		skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 891		if (force_schedule) {
 892			mem_scheduled = true;
 893			sk_forced_mem_schedule(sk, skb->truesize);
 894		} else {
 895			mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 896		}
 897		if (likely(mem_scheduled)) {
 898			skb_reserve(skb, MAX_TCP_HEADER);
 899			skb->ip_summed = CHECKSUM_PARTIAL;
 
 
 
 
 900			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 901			return skb;
 902		}
 903		__kfree_skb(skb);
 904	} else {
 905		sk->sk_prot->enter_memory_pressure(sk);
 906		sk_stream_moderate_sndbuf(sk);
 907	}
 908	return NULL;
 909}
 910
 911static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 912				       int large_allowed)
 913{
 914	struct tcp_sock *tp = tcp_sk(sk);
 915	u32 new_size_goal, size_goal;
 916
 917	if (!large_allowed)
 918		return mss_now;
 919
 920	/* Note : tcp_tso_autosize() will eventually split this later */
 921	new_size_goal = tcp_bound_to_half_wnd(tp, sk->sk_gso_max_size);
 
 922
 923	/* We try hard to avoid divides here */
 924	size_goal = tp->gso_segs * mss_now;
 925	if (unlikely(new_size_goal < size_goal ||
 926		     new_size_goal >= size_goal + mss_now)) {
 927		tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 928				     sk->sk_gso_max_segs);
 929		size_goal = tp->gso_segs * mss_now;
 930	}
 931
 932	return max(size_goal, mss_now);
 933}
 934
 935int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 936{
 937	int mss_now;
 938
 939	mss_now = tcp_current_mss(sk);
 940	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 941
 942	return mss_now;
 943}
 944
 945/* In some cases, sendmsg() could have added an skb to the write queue,
 946 * but failed adding payload on it. We need to remove it to consume less
 947 * memory, but more importantly be able to generate EPOLLOUT for Edge Trigger
 948 * epoll() users. Another reason is that tcp_write_xmit() does not like
 949 * finding an empty skb in the write queue.
 950 */
 951void tcp_remove_empty_skb(struct sock *sk)
 952{
 953	struct sk_buff *skb = tcp_write_queue_tail(sk);
 954
 955	if (skb && TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
 956		tcp_unlink_write_queue(skb, sk);
 957		if (tcp_write_queue_empty(sk))
 958			tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
 959		tcp_wmem_free_skb(sk, skb);
 960	}
 961}
 962
 963/* skb changing from pure zc to mixed, must charge zc */
 964static int tcp_downgrade_zcopy_pure(struct sock *sk, struct sk_buff *skb)
 965{
 966	if (unlikely(skb_zcopy_pure(skb))) {
 967		u32 extra = skb->truesize -
 968			    SKB_TRUESIZE(skb_end_offset(skb));
 
 
 969
 970		if (!sk_wmem_schedule(sk, extra))
 971			return -ENOMEM;
 
 972
 973		sk_mem_charge(sk, extra);
 974		skb_shinfo(skb)->flags &= ~SKBFL_PURE_ZEROCOPY;
 
 
 
 
 
 
 
 975	}
 976	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 977}
 
 978
 
 
 
 
 
 979
 980int tcp_wmem_schedule(struct sock *sk, int copy)
 
 
 
 
 
 
 
 981{
 982	int left;
 983
 984	if (likely(sk_wmem_schedule(sk, copy)))
 985		return copy;
 
 986
 987	/* We could be in trouble if we have nothing queued.
 988	 * Use whatever is left in sk->sk_forward_alloc and tcp_wmem[0]
 989	 * to guarantee some progress.
 990	 */
 991	left = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_wmem[0]) - sk->sk_wmem_queued;
 992	if (left > 0)
 993		sk_forced_mem_schedule(sk, min(left, copy));
 994	return min(copy, sk->sk_forward_alloc);
 995}
 
 996
 997void tcp_free_fastopen_req(struct tcp_sock *tp)
 998{
 999	if (tp->fastopen_req) {
1000		kfree(tp->fastopen_req);
1001		tp->fastopen_req = NULL;
1002	}
1003}
1004
1005int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg, int *copied,
1006			 size_t size, struct ubuf_info *uarg)
 
1007{
1008	struct tcp_sock *tp = tcp_sk(sk);
1009	struct inet_sock *inet = inet_sk(sk);
1010	struct sockaddr *uaddr = msg->msg_name;
1011	int err, flags;
1012
1013	if (!(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) &
1014	      TFO_CLIENT_ENABLE) ||
1015	    (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1016	     uaddr->sa_family == AF_UNSPEC))
1017		return -EOPNOTSUPP;
1018	if (tp->fastopen_req)
1019		return -EALREADY; /* Another Fast Open is in progress */
1020
1021	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1022				   sk->sk_allocation);
1023	if (unlikely(!tp->fastopen_req))
1024		return -ENOBUFS;
1025	tp->fastopen_req->data = msg;
1026	tp->fastopen_req->size = size;
1027	tp->fastopen_req->uarg = uarg;
1028
1029	if (inet_test_bit(DEFER_CONNECT, sk)) {
1030		err = tcp_connect(sk);
1031		/* Same failure procedure as in tcp_v4/6_connect */
1032		if (err) {
1033			tcp_set_state(sk, TCP_CLOSE);
1034			inet->inet_dport = 0;
1035			sk->sk_route_caps = 0;
1036		}
1037	}
1038	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1039	err = __inet_stream_connect(sk->sk_socket, uaddr,
1040				    msg->msg_namelen, flags, 1);
1041	/* fastopen_req could already be freed in __inet_stream_connect
1042	 * if the connection times out or gets rst
1043	 */
1044	if (tp->fastopen_req) {
1045		*copied = tp->fastopen_req->copied;
1046		tcp_free_fastopen_req(tp);
1047		inet_clear_bit(DEFER_CONNECT, sk);
1048	}
1049	return err;
1050}
1051
1052int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1053{
1054	struct tcp_sock *tp = tcp_sk(sk);
1055	struct ubuf_info *uarg = NULL;
1056	struct sk_buff *skb;
1057	struct sockcm_cookie sockc;
1058	int flags, err, copied = 0;
1059	int mss_now = 0, size_goal, copied_syn = 0;
1060	int process_backlog = 0;
1061	int zc = 0;
1062	long timeo;
1063
1064	flags = msg->msg_flags;
1065
1066	if ((flags & MSG_ZEROCOPY) && size) {
1067		if (msg->msg_ubuf) {
1068			uarg = msg->msg_ubuf;
1069			if (sk->sk_route_caps & NETIF_F_SG)
1070				zc = MSG_ZEROCOPY;
1071		} else if (sock_flag(sk, SOCK_ZEROCOPY)) {
1072			skb = tcp_write_queue_tail(sk);
1073			uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1074			if (!uarg) {
1075				err = -ENOBUFS;
1076				goto out_err;
1077			}
1078			if (sk->sk_route_caps & NETIF_F_SG)
1079				zc = MSG_ZEROCOPY;
1080			else
1081				uarg_to_msgzc(uarg)->zerocopy = 0;
1082		}
1083	} else if (unlikely(msg->msg_flags & MSG_SPLICE_PAGES) && size) {
1084		if (sk->sk_route_caps & NETIF_F_SG)
1085			zc = MSG_SPLICE_PAGES;
 
1086	}
1087
1088	if (unlikely(flags & MSG_FASTOPEN ||
1089		     inet_test_bit(DEFER_CONNECT, sk)) &&
1090	    !tp->repair) {
1091		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1092		if (err == -EINPROGRESS && copied_syn > 0)
1093			goto out;
1094		else if (err)
1095			goto out_err;
1096	}
1097
1098	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1099
1100	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1101
1102	/* Wait for a connection to finish. One exception is TCP Fast Open
1103	 * (passive side) where data is allowed to be sent before a connection
1104	 * is fully established.
1105	 */
1106	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1107	    !tcp_passive_fastopen(sk)) {
1108		err = sk_stream_wait_connect(sk, &timeo);
1109		if (err != 0)
1110			goto do_error;
1111	}
1112
1113	if (unlikely(tp->repair)) {
1114		if (tp->repair_queue == TCP_RECV_QUEUE) {
1115			copied = tcp_send_rcvq(sk, msg, size);
1116			goto out_nopush;
1117		}
1118
1119		err = -EINVAL;
1120		if (tp->repair_queue == TCP_NO_QUEUE)
1121			goto out_err;
1122
1123		/* 'common' sending to sendq */
1124	}
1125
1126	sockcm_init(&sockc, sk);
1127	if (msg->msg_controllen) {
1128		err = sock_cmsg_send(sk, msg, &sockc);
1129		if (unlikely(err)) {
1130			err = -EINVAL;
1131			goto out_err;
1132		}
1133	}
1134
1135	/* This should be in poll */
1136	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1137
1138	/* Ok commence sending. */
1139	copied = 0;
1140
1141restart:
1142	mss_now = tcp_send_mss(sk, &size_goal, flags);
1143
1144	err = -EPIPE;
1145	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1146		goto do_error;
1147
1148	while (msg_data_left(msg)) {
1149		ssize_t copy = 0;
1150
1151		skb = tcp_write_queue_tail(sk);
1152		if (skb)
1153			copy = size_goal - skb->len;
1154
1155		if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1156			bool first_skb;
1157
1158new_segment:
1159			if (!sk_stream_memory_free(sk))
1160				goto wait_for_space;
1161
1162			if (unlikely(process_backlog >= 16)) {
1163				process_backlog = 0;
1164				if (sk_flush_backlog(sk))
1165					goto restart;
1166			}
1167			first_skb = tcp_rtx_and_write_queues_empty(sk);
1168			skb = tcp_stream_alloc_skb(sk, sk->sk_allocation,
1169						   first_skb);
1170			if (!skb)
1171				goto wait_for_space;
1172
1173			process_backlog++;
 
1174
1175#ifdef CONFIG_SKB_DECRYPTED
1176			skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1177#endif
1178			tcp_skb_entail(sk, skb);
1179			copy = size_goal;
1180
1181			/* All packets are restored as if they have
1182			 * already been sent. skb_mstamp_ns isn't set to
1183			 * avoid wrong rtt estimation.
1184			 */
1185			if (tp->repair)
1186				TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1187		}
1188
1189		/* Try to append data to the end of skb. */
1190		if (copy > msg_data_left(msg))
1191			copy = msg_data_left(msg);
1192
1193		if (zc == 0) {
 
 
 
 
 
 
 
1194			bool merge = true;
1195			int i = skb_shinfo(skb)->nr_frags;
1196			struct page_frag *pfrag = sk_page_frag(sk);
1197
1198			if (!sk_page_frag_refill(sk, pfrag))
1199				goto wait_for_space;
1200
1201			if (!skb_can_coalesce(skb, i, pfrag->page,
1202					      pfrag->offset)) {
1203				if (i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1204					tcp_mark_push(tp, skb);
1205					goto new_segment;
1206				}
1207				merge = false;
1208			}
1209
1210			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1211
1212			if (unlikely(skb_zcopy_pure(skb) || skb_zcopy_managed(skb))) {
1213				if (tcp_downgrade_zcopy_pure(sk, skb))
1214					goto wait_for_space;
1215				skb_zcopy_downgrade_managed(skb);
1216			}
1217
1218			copy = tcp_wmem_schedule(sk, copy);
1219			if (!copy)
1220				goto wait_for_space;
1221
1222			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1223						       pfrag->page,
1224						       pfrag->offset,
1225						       copy);
1226			if (err)
1227				goto do_error;
1228
1229			/* Update the skb. */
1230			if (merge) {
1231				skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1232			} else {
1233				skb_fill_page_desc(skb, i, pfrag->page,
1234						   pfrag->offset, copy);
1235				page_ref_inc(pfrag->page);
1236			}
1237			pfrag->offset += copy;
1238		} else if (zc == MSG_ZEROCOPY)  {
1239			/* First append to a fragless skb builds initial
1240			 * pure zerocopy skb
1241			 */
1242			if (!skb->len)
1243				skb_shinfo(skb)->flags |= SKBFL_PURE_ZEROCOPY;
1244
1245			if (!skb_zcopy_pure(skb)) {
1246				copy = tcp_wmem_schedule(sk, copy);
1247				if (!copy)
1248					goto wait_for_space;
1249			}
1250
1251			err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1252			if (err == -EMSGSIZE || err == -EEXIST) {
1253				tcp_mark_push(tp, skb);
1254				goto new_segment;
1255			}
1256			if (err < 0)
1257				goto do_error;
1258			copy = err;
1259		} else if (zc == MSG_SPLICE_PAGES) {
1260			/* Splice in data if we can; copy if we can't. */
1261			if (tcp_downgrade_zcopy_pure(sk, skb))
1262				goto wait_for_space;
1263			copy = tcp_wmem_schedule(sk, copy);
1264			if (!copy)
1265				goto wait_for_space;
1266
1267			err = skb_splice_from_iter(skb, &msg->msg_iter, copy,
1268						   sk->sk_allocation);
1269			if (err < 0) {
1270				if (err == -EMSGSIZE) {
1271					tcp_mark_push(tp, skb);
1272					goto new_segment;
1273				}
1274				goto do_error;
1275			}
1276			copy = err;
1277
1278			if (!(flags & MSG_NO_SHARED_FRAGS))
1279				skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1280
1281			sk_wmem_queued_add(sk, copy);
1282			sk_mem_charge(sk, copy);
1283		}
1284
1285		if (!copied)
1286			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1287
1288		WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1289		TCP_SKB_CB(skb)->end_seq += copy;
1290		tcp_skb_pcount_set(skb, 0);
1291
1292		copied += copy;
1293		if (!msg_data_left(msg)) {
1294			if (unlikely(flags & MSG_EOR))
1295				TCP_SKB_CB(skb)->eor = 1;
1296			goto out;
1297		}
1298
1299		if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1300			continue;
1301
1302		if (forced_push(tp)) {
1303			tcp_mark_push(tp, skb);
1304			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1305		} else if (skb == tcp_send_head(sk))
1306			tcp_push_one(sk, mss_now);
1307		continue;
1308
1309wait_for_space:
1310		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1311		tcp_remove_empty_skb(sk);
1312		if (copied)
1313			tcp_push(sk, flags & ~MSG_MORE, mss_now,
1314				 TCP_NAGLE_PUSH, size_goal);
1315
1316		err = sk_stream_wait_memory(sk, &timeo);
1317		if (err != 0)
1318			goto do_error;
1319
1320		mss_now = tcp_send_mss(sk, &size_goal, flags);
1321	}
1322
1323out:
1324	if (copied) {
1325		tcp_tx_timestamp(sk, &sockc);
1326		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1327	}
1328out_nopush:
1329	/* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1330	if (uarg && !msg->msg_ubuf)
1331		net_zcopy_put(uarg);
1332	return copied + copied_syn;
1333
1334do_error:
1335	tcp_remove_empty_skb(sk);
 
 
1336
1337	if (copied + copied_syn)
1338		goto out;
1339out_err:
1340	/* msg->msg_ubuf is pinned by the caller so we don't take extra refs */
1341	if (uarg && !msg->msg_ubuf)
1342		net_zcopy_put_abort(uarg, true);
1343	err = sk_stream_error(sk, flags, err);
1344	/* make sure we wake any epoll edge trigger waiter */
1345	if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
 
1346		sk->sk_write_space(sk);
1347		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1348	}
1349	return err;
1350}
1351EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1352
1353int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1354{
1355	int ret;
1356
1357	lock_sock(sk);
1358	ret = tcp_sendmsg_locked(sk, msg, size);
1359	release_sock(sk);
1360
1361	return ret;
1362}
1363EXPORT_SYMBOL(tcp_sendmsg);
1364
1365void tcp_splice_eof(struct socket *sock)
1366{
1367	struct sock *sk = sock->sk;
1368	struct tcp_sock *tp = tcp_sk(sk);
1369	int mss_now, size_goal;
1370
1371	if (!tcp_write_queue_tail(sk))
1372		return;
1373
1374	lock_sock(sk);
1375	mss_now = tcp_send_mss(sk, &size_goal, 0);
1376	tcp_push(sk, 0, mss_now, tp->nonagle, size_goal);
1377	release_sock(sk);
1378}
1379EXPORT_SYMBOL_GPL(tcp_splice_eof);
1380
1381/*
1382 *	Handle reading urgent data. BSD has very simple semantics for
1383 *	this, no blocking and very strange errors 8)
1384 */
1385
1386static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1387{
1388	struct tcp_sock *tp = tcp_sk(sk);
1389
1390	/* No URG data to read. */
1391	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1392	    tp->urg_data == TCP_URG_READ)
1393		return -EINVAL;	/* Yes this is right ! */
1394
1395	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1396		return -ENOTCONN;
1397
1398	if (tp->urg_data & TCP_URG_VALID) {
1399		int err = 0;
1400		char c = tp->urg_data;
1401
1402		if (!(flags & MSG_PEEK))
1403			WRITE_ONCE(tp->urg_data, TCP_URG_READ);
1404
1405		/* Read urgent data. */
1406		msg->msg_flags |= MSG_OOB;
1407
1408		if (len > 0) {
1409			if (!(flags & MSG_TRUNC))
1410				err = memcpy_to_msg(msg, &c, 1);
1411			len = 1;
1412		} else
1413			msg->msg_flags |= MSG_TRUNC;
1414
1415		return err ? -EFAULT : len;
1416	}
1417
1418	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1419		return 0;
1420
1421	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1422	 * the available implementations agree in this case:
1423	 * this call should never block, independent of the
1424	 * blocking state of the socket.
1425	 * Mike <pall@rz.uni-karlsruhe.de>
1426	 */
1427	return -EAGAIN;
1428}
1429
1430static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1431{
1432	struct sk_buff *skb;
1433	int copied = 0, err = 0;
1434
 
 
1435	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1436		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1437		if (err)
1438			return err;
1439		copied += skb->len;
1440	}
1441
1442	skb_queue_walk(&sk->sk_write_queue, skb) {
1443		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1444		if (err)
1445			break;
1446
1447		copied += skb->len;
1448	}
1449
1450	return err ?: copied;
1451}
1452
1453/* Clean up the receive buffer for full frames taken by the user,
1454 * then send an ACK if necessary.  COPIED is the number of bytes
1455 * tcp_recvmsg has given to the user so far, it speeds up the
1456 * calculation of whether or not we must ACK for the sake of
1457 * a window update.
1458 */
1459void __tcp_cleanup_rbuf(struct sock *sk, int copied)
1460{
1461	struct tcp_sock *tp = tcp_sk(sk);
1462	bool time_to_ack = false;
1463
 
 
 
 
 
 
1464	if (inet_csk_ack_scheduled(sk)) {
1465		const struct inet_connection_sock *icsk = inet_csk(sk);
1466
1467		if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
 
 
1468		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1469		    /*
1470		     * If this read emptied read buffer, we send ACK, if
1471		     * connection is not bidirectional, user drained
1472		     * receive buffer and there was a small segment
1473		     * in queue.
1474		     */
1475		    (copied > 0 &&
1476		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1477		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1478		       !inet_csk_in_pingpong_mode(sk))) &&
1479		      !atomic_read(&sk->sk_rmem_alloc)))
1480			time_to_ack = true;
1481	}
1482
1483	/* We send an ACK if we can now advertise a non-zero window
1484	 * which has been raised "significantly".
1485	 *
1486	 * Even if window raised up to infinity, do not send window open ACK
1487	 * in states, where we will not receive more. It is useless.
1488	 */
1489	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1490		__u32 rcv_window_now = tcp_receive_window(tp);
1491
1492		/* Optimize, __tcp_select_window() is not cheap. */
1493		if (2*rcv_window_now <= tp->window_clamp) {
1494			__u32 new_window = __tcp_select_window(sk);
1495
1496			/* Send ACK now, if this read freed lots of space
1497			 * in our buffer. Certainly, new_window is new window.
1498			 * We can advertise it now, if it is not less than current one.
1499			 * "Lots" means "at least twice" here.
1500			 */
1501			if (new_window && new_window >= 2 * rcv_window_now)
1502				time_to_ack = true;
1503		}
1504	}
1505	if (time_to_ack)
1506		tcp_send_ack(sk);
1507}
1508
1509void tcp_cleanup_rbuf(struct sock *sk, int copied)
1510{
1511	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1512	struct tcp_sock *tp = tcp_sk(sk);
1513
1514	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1515	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1516	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1517	__tcp_cleanup_rbuf(sk, copied);
1518}
1519
1520static void tcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1521{
1522	__skb_unlink(skb, &sk->sk_receive_queue);
1523	if (likely(skb->destructor == sock_rfree)) {
1524		sock_rfree(skb);
1525		skb->destructor = NULL;
1526		skb->sk = NULL;
1527		return skb_attempt_defer_free(skb);
1528	}
1529	__kfree_skb(skb);
1530}
1531
1532struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1533{
1534	struct sk_buff *skb;
1535	u32 offset;
1536
1537	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1538		offset = seq - TCP_SKB_CB(skb)->seq;
1539		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1540			pr_err_once("%s: found a SYN, please report !\n", __func__);
1541			offset--;
1542		}
1543		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1544			*off = offset;
1545			return skb;
1546		}
1547		/* This looks weird, but this can happen if TCP collapsing
1548		 * splitted a fat GRO packet, while we released socket lock
1549		 * in skb_splice_bits()
1550		 */
1551		tcp_eat_recv_skb(sk, skb);
1552	}
1553	return NULL;
1554}
1555EXPORT_SYMBOL(tcp_recv_skb);
1556
1557/*
1558 * This routine provides an alternative to tcp_recvmsg() for routines
1559 * that would like to handle copying from skbuffs directly in 'sendfile'
1560 * fashion.
1561 * Note:
1562 *	- It is assumed that the socket was locked by the caller.
1563 *	- The routine does not block.
1564 *	- At present, there is no support for reading OOB data
1565 *	  or for 'peeking' the socket using this routine
1566 *	  (although both would be easy to implement).
1567 */
1568int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1569		  sk_read_actor_t recv_actor)
1570{
1571	struct sk_buff *skb;
1572	struct tcp_sock *tp = tcp_sk(sk);
1573	u32 seq = tp->copied_seq;
1574	u32 offset;
1575	int copied = 0;
1576
1577	if (sk->sk_state == TCP_LISTEN)
1578		return -ENOTCONN;
1579	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1580		if (offset < skb->len) {
1581			int used;
1582			size_t len;
1583
1584			len = skb->len - offset;
1585			/* Stop reading if we hit a patch of urgent data */
1586			if (unlikely(tp->urg_data)) {
1587				u32 urg_offset = tp->urg_seq - seq;
1588				if (urg_offset < len)
1589					len = urg_offset;
1590				if (!len)
1591					break;
1592			}
1593			used = recv_actor(desc, skb, offset, len);
1594			if (used <= 0) {
1595				if (!copied)
1596					copied = used;
1597				break;
 
 
 
 
1598			}
1599			if (WARN_ON_ONCE(used > len))
1600				used = len;
1601			seq += used;
1602			copied += used;
1603			offset += used;
1604
1605			/* If recv_actor drops the lock (e.g. TCP splice
1606			 * receive) the skb pointer might be invalid when
1607			 * getting here: tcp_collapse might have deleted it
1608			 * while aggregating skbs from the socket queue.
1609			 */
1610			skb = tcp_recv_skb(sk, seq - 1, &offset);
1611			if (!skb)
1612				break;
1613			/* TCP coalescing might have appended data to the skb.
1614			 * Try to splice more frags
1615			 */
1616			if (offset + 1 != skb->len)
1617				continue;
1618		}
1619		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1620			tcp_eat_recv_skb(sk, skb);
1621			++seq;
1622			break;
1623		}
1624		tcp_eat_recv_skb(sk, skb);
1625		if (!desc->count)
1626			break;
1627		WRITE_ONCE(tp->copied_seq, seq);
1628	}
1629	WRITE_ONCE(tp->copied_seq, seq);
1630
1631	tcp_rcv_space_adjust(sk);
1632
1633	/* Clean up data we have read: This will do ACK frames. */
1634	if (copied > 0) {
1635		tcp_recv_skb(sk, seq, &offset);
1636		tcp_cleanup_rbuf(sk, copied);
1637	}
1638	return copied;
1639}
1640EXPORT_SYMBOL(tcp_read_sock);
1641
1642int tcp_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
1643{
1644	struct sk_buff *skb;
1645	int copied = 0;
1646
1647	if (sk->sk_state == TCP_LISTEN)
1648		return -ENOTCONN;
1649
1650	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1651		u8 tcp_flags;
1652		int used;
1653
1654		__skb_unlink(skb, &sk->sk_receive_queue);
1655		WARN_ON_ONCE(!skb_set_owner_sk_safe(skb, sk));
1656		tcp_flags = TCP_SKB_CB(skb)->tcp_flags;
1657		used = recv_actor(sk, skb);
1658		if (used < 0) {
1659			if (!copied)
1660				copied = used;
1661			break;
1662		}
1663		copied += used;
1664
1665		if (tcp_flags & TCPHDR_FIN)
1666			break;
1667	}
1668	return copied;
1669}
1670EXPORT_SYMBOL(tcp_read_skb);
1671
1672void tcp_read_done(struct sock *sk, size_t len)
1673{
1674	struct tcp_sock *tp = tcp_sk(sk);
1675	u32 seq = tp->copied_seq;
1676	struct sk_buff *skb;
1677	size_t left;
1678	u32 offset;
1679
1680	if (sk->sk_state == TCP_LISTEN)
1681		return;
1682
1683	left = len;
1684	while (left && (skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1685		int used;
1686
1687		used = min_t(size_t, skb->len - offset, left);
1688		seq += used;
1689		left -= used;
1690
1691		if (skb->len > offset + used)
1692			break;
1693
1694		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1695			tcp_eat_recv_skb(sk, skb);
1696			++seq;
1697			break;
1698		}
1699		tcp_eat_recv_skb(sk, skb);
1700	}
1701	WRITE_ONCE(tp->copied_seq, seq);
1702
1703	tcp_rcv_space_adjust(sk);
1704
1705	/* Clean up data we have read: This will do ACK frames. */
1706	if (left != len)
1707		tcp_cleanup_rbuf(sk, len - left);
1708}
1709EXPORT_SYMBOL(tcp_read_done);
1710
1711int tcp_peek_len(struct socket *sock)
1712{
1713	return tcp_inq(sock->sk);
1714}
1715EXPORT_SYMBOL(tcp_peek_len);
1716
1717/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1718int tcp_set_rcvlowat(struct sock *sk, int val)
1719{
1720	int space, cap;
1721
1722	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1723		cap = sk->sk_rcvbuf >> 1;
1724	else
1725		cap = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]) >> 1;
1726	val = min(val, cap);
1727	WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1728
1729	/* Check if we need to signal EPOLLIN right now */
1730	tcp_data_ready(sk);
1731
1732	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1733		return 0;
1734
1735	space = tcp_space_from_win(sk, val);
1736	if (space > sk->sk_rcvbuf) {
1737		WRITE_ONCE(sk->sk_rcvbuf, space);
1738		WRITE_ONCE(tcp_sk(sk)->window_clamp, val);
1739	}
1740	return 0;
1741}
1742EXPORT_SYMBOL(tcp_set_rcvlowat);
1743
1744void tcp_update_recv_tstamps(struct sk_buff *skb,
1745			     struct scm_timestamping_internal *tss)
1746{
1747	if (skb->tstamp)
1748		tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1749	else
1750		tss->ts[0] = (struct timespec64) {0};
1751
1752	if (skb_hwtstamps(skb)->hwtstamp)
1753		tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1754	else
1755		tss->ts[2] = (struct timespec64) {0};
1756}
1757
1758#ifdef CONFIG_MMU
1759static const struct vm_operations_struct tcp_vm_ops = {
1760};
1761
1762int tcp_mmap(struct file *file, struct socket *sock,
1763	     struct vm_area_struct *vma)
1764{
1765	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1766		return -EPERM;
1767	vm_flags_clear(vma, VM_MAYWRITE | VM_MAYEXEC);
1768
1769	/* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1770	vm_flags_set(vma, VM_MIXEDMAP);
1771
1772	vma->vm_ops = &tcp_vm_ops;
1773	return 0;
1774}
1775EXPORT_SYMBOL(tcp_mmap);
1776
1777static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1778				       u32 *offset_frag)
1779{
1780	skb_frag_t *frag;
1781
1782	if (unlikely(offset_skb >= skb->len))
1783		return NULL;
1784
1785	offset_skb -= skb_headlen(skb);
1786	if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1787		return NULL;
1788
1789	frag = skb_shinfo(skb)->frags;
1790	while (offset_skb) {
1791		if (skb_frag_size(frag) > offset_skb) {
1792			*offset_frag = offset_skb;
1793			return frag;
1794		}
1795		offset_skb -= skb_frag_size(frag);
1796		++frag;
1797	}
1798	*offset_frag = 0;
1799	return frag;
1800}
1801
1802static bool can_map_frag(const skb_frag_t *frag)
1803{
1804	struct page *page;
1805
1806	if (skb_frag_size(frag) != PAGE_SIZE || skb_frag_off(frag))
1807		return false;
1808
1809	page = skb_frag_page(frag);
1810
1811	if (PageCompound(page) || page->mapping)
1812		return false;
1813
1814	return true;
1815}
1816
1817static int find_next_mappable_frag(const skb_frag_t *frag,
1818				   int remaining_in_skb)
1819{
1820	int offset = 0;
1821
1822	if (likely(can_map_frag(frag)))
1823		return 0;
1824
1825	while (offset < remaining_in_skb && !can_map_frag(frag)) {
1826		offset += skb_frag_size(frag);
1827		++frag;
1828	}
1829	return offset;
1830}
1831
1832static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1833					  struct tcp_zerocopy_receive *zc,
1834					  struct sk_buff *skb, u32 offset)
1835{
1836	u32 frag_offset, partial_frag_remainder = 0;
1837	int mappable_offset;
1838	skb_frag_t *frag;
1839
1840	/* worst case: skip to next skb. try to improve on this case below */
1841	zc->recv_skip_hint = skb->len - offset;
1842
1843	/* Find the frag containing this offset (and how far into that frag) */
1844	frag = skb_advance_to_frag(skb, offset, &frag_offset);
1845	if (!frag)
1846		return;
1847
1848	if (frag_offset) {
1849		struct skb_shared_info *info = skb_shinfo(skb);
1850
1851		/* We read part of the last frag, must recvmsg() rest of skb. */
1852		if (frag == &info->frags[info->nr_frags - 1])
1853			return;
1854
1855		/* Else, we must at least read the remainder in this frag. */
1856		partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1857		zc->recv_skip_hint -= partial_frag_remainder;
1858		++frag;
1859	}
1860
1861	/* partial_frag_remainder: If part way through a frag, must read rest.
1862	 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1863	 * in partial_frag_remainder.
1864	 */
1865	mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1866	zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1867}
1868
1869static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1870			      int flags, struct scm_timestamping_internal *tss,
1871			      int *cmsg_flags);
1872static int receive_fallback_to_copy(struct sock *sk,
1873				    struct tcp_zerocopy_receive *zc, int inq,
1874				    struct scm_timestamping_internal *tss)
1875{
1876	unsigned long copy_address = (unsigned long)zc->copybuf_address;
1877	struct msghdr msg = {};
1878	int err;
1879
1880	zc->length = 0;
1881	zc->recv_skip_hint = 0;
1882
1883	if (copy_address != zc->copybuf_address)
1884		return -EINVAL;
1885
1886	err = import_ubuf(ITER_DEST, (void __user *)copy_address, inq,
1887			  &msg.msg_iter);
1888	if (err)
1889		return err;
1890
1891	err = tcp_recvmsg_locked(sk, &msg, inq, MSG_DONTWAIT,
1892				 tss, &zc->msg_flags);
1893	if (err < 0)
1894		return err;
1895
1896	zc->copybuf_len = err;
1897	if (likely(zc->copybuf_len)) {
1898		struct sk_buff *skb;
1899		u32 offset;
1900
1901		skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1902		if (skb)
1903			tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1904	}
1905	return 0;
1906}
1907
1908static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1909				   struct sk_buff *skb, u32 copylen,
1910				   u32 *offset, u32 *seq)
1911{
1912	unsigned long copy_address = (unsigned long)zc->copybuf_address;
1913	struct msghdr msg = {};
1914	int err;
1915
1916	if (copy_address != zc->copybuf_address)
1917		return -EINVAL;
1918
1919	err = import_ubuf(ITER_DEST, (void __user *)copy_address, copylen,
1920			  &msg.msg_iter);
1921	if (err)
1922		return err;
1923	err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1924	if (err)
1925		return err;
1926	zc->recv_skip_hint -= copylen;
1927	*offset += copylen;
1928	*seq += copylen;
1929	return (__s32)copylen;
1930}
1931
1932static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1933				  struct sock *sk,
1934				  struct sk_buff *skb,
1935				  u32 *seq,
1936				  s32 copybuf_len,
1937				  struct scm_timestamping_internal *tss)
1938{
1939	u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1940
1941	if (!copylen)
1942		return 0;
1943	/* skb is null if inq < PAGE_SIZE. */
1944	if (skb) {
1945		offset = *seq - TCP_SKB_CB(skb)->seq;
1946	} else {
1947		skb = tcp_recv_skb(sk, *seq, &offset);
1948		if (TCP_SKB_CB(skb)->has_rxtstamp) {
1949			tcp_update_recv_tstamps(skb, tss);
1950			zc->msg_flags |= TCP_CMSG_TS;
1951		}
1952	}
1953
1954	zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1955						  seq);
1956	return zc->copybuf_len < 0 ? 0 : copylen;
1957}
1958
1959static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1960					      struct page **pending_pages,
1961					      unsigned long pages_remaining,
1962					      unsigned long *address,
1963					      u32 *length,
1964					      u32 *seq,
1965					      struct tcp_zerocopy_receive *zc,
1966					      u32 total_bytes_to_map,
1967					      int err)
1968{
1969	/* At least one page did not map. Try zapping if we skipped earlier. */
1970	if (err == -EBUSY &&
1971	    zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1972		u32 maybe_zap_len;
1973
1974		maybe_zap_len = total_bytes_to_map -  /* All bytes to map */
1975				*length + /* Mapped or pending */
1976				(pages_remaining * PAGE_SIZE); /* Failed map. */
1977		zap_page_range_single(vma, *address, maybe_zap_len, NULL);
1978		err = 0;
1979	}
1980
1981	if (!err) {
1982		unsigned long leftover_pages = pages_remaining;
1983		int bytes_mapped;
1984
1985		/* We called zap_page_range_single, try to reinsert. */
1986		err = vm_insert_pages(vma, *address,
1987				      pending_pages,
1988				      &pages_remaining);
1989		bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1990		*seq += bytes_mapped;
1991		*address += bytes_mapped;
1992	}
1993	if (err) {
1994		/* Either we were unable to zap, OR we zapped, retried an
1995		 * insert, and still had an issue. Either ways, pages_remaining
1996		 * is the number of pages we were unable to map, and we unroll
1997		 * some state we speculatively touched before.
1998		 */
1999		const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
2000
2001		*length -= bytes_not_mapped;
2002		zc->recv_skip_hint += bytes_not_mapped;
2003	}
2004	return err;
2005}
2006
2007static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
2008					struct page **pages,
2009					unsigned int pages_to_map,
2010					unsigned long *address,
2011					u32 *length,
2012					u32 *seq,
2013					struct tcp_zerocopy_receive *zc,
2014					u32 total_bytes_to_map)
2015{
2016	unsigned long pages_remaining = pages_to_map;
2017	unsigned int pages_mapped;
2018	unsigned int bytes_mapped;
2019	int err;
2020
2021	err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2022	pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2023	bytes_mapped = PAGE_SIZE * pages_mapped;
2024	/* Even if vm_insert_pages fails, it may have partially succeeded in
2025	 * mapping (some but not all of the pages).
2026	 */
2027	*seq += bytes_mapped;
2028	*address += bytes_mapped;
2029
2030	if (likely(!err))
2031		return 0;
2032
2033	/* Error: maybe zap and retry + rollback state for failed inserts. */
2034	return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2035		pages_remaining, address, length, seq, zc, total_bytes_to_map,
2036		err);
2037}
2038
2039#define TCP_VALID_ZC_MSG_FLAGS   (TCP_CMSG_TS)
2040static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2041				      struct tcp_zerocopy_receive *zc,
2042				      struct scm_timestamping_internal *tss)
2043{
2044	unsigned long msg_control_addr;
2045	struct msghdr cmsg_dummy;
2046
2047	msg_control_addr = (unsigned long)zc->msg_control;
2048	cmsg_dummy.msg_control_user = (void __user *)msg_control_addr;
2049	cmsg_dummy.msg_controllen =
2050		(__kernel_size_t)zc->msg_controllen;
2051	cmsg_dummy.msg_flags = in_compat_syscall()
2052		? MSG_CMSG_COMPAT : 0;
2053	cmsg_dummy.msg_control_is_user = true;
2054	zc->msg_flags = 0;
2055	if (zc->msg_control == msg_control_addr &&
2056	    zc->msg_controllen == cmsg_dummy.msg_controllen) {
2057		tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2058		zc->msg_control = (__u64)
2059			((uintptr_t)cmsg_dummy.msg_control_user);
2060		zc->msg_controllen =
2061			(__u64)cmsg_dummy.msg_controllen;
2062		zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2063	}
2064}
2065
2066static struct vm_area_struct *find_tcp_vma(struct mm_struct *mm,
2067					   unsigned long address,
2068					   bool *mmap_locked)
2069{
2070	struct vm_area_struct *vma = lock_vma_under_rcu(mm, address);
2071
2072	if (vma) {
2073		if (vma->vm_ops != &tcp_vm_ops) {
2074			vma_end_read(vma);
2075			return NULL;
2076		}
2077		*mmap_locked = false;
2078		return vma;
2079	}
2080
2081	mmap_read_lock(mm);
2082	vma = vma_lookup(mm, address);
2083	if (!vma || vma->vm_ops != &tcp_vm_ops) {
2084		mmap_read_unlock(mm);
2085		return NULL;
2086	}
2087	*mmap_locked = true;
2088	return vma;
2089}
2090
2091#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2092static int tcp_zerocopy_receive(struct sock *sk,
2093				struct tcp_zerocopy_receive *zc,
2094				struct scm_timestamping_internal *tss)
2095{
2096	u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2097	unsigned long address = (unsigned long)zc->address;
2098	struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2099	s32 copybuf_len = zc->copybuf_len;
2100	struct tcp_sock *tp = tcp_sk(sk);
2101	const skb_frag_t *frags = NULL;
2102	unsigned int pages_to_map = 0;
2103	struct vm_area_struct *vma;
2104	struct sk_buff *skb = NULL;
2105	u32 seq = tp->copied_seq;
2106	u32 total_bytes_to_map;
2107	int inq = tcp_inq(sk);
2108	bool mmap_locked;
2109	int ret;
2110
2111	zc->copybuf_len = 0;
2112	zc->msg_flags = 0;
2113
2114	if (address & (PAGE_SIZE - 1) || address != zc->address)
2115		return -EINVAL;
2116
2117	if (sk->sk_state == TCP_LISTEN)
2118		return -ENOTCONN;
2119
2120	sock_rps_record_flow(sk);
2121
2122	if (inq && inq <= copybuf_len)
2123		return receive_fallback_to_copy(sk, zc, inq, tss);
2124
2125	if (inq < PAGE_SIZE) {
2126		zc->length = 0;
2127		zc->recv_skip_hint = inq;
2128		if (!inq && sock_flag(sk, SOCK_DONE))
2129			return -EIO;
2130		return 0;
2131	}
2132
2133	vma = find_tcp_vma(current->mm, address, &mmap_locked);
2134	if (!vma)
2135		return -EINVAL;
 
 
2136
2137	vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2138	avail_len = min_t(u32, vma_len, inq);
2139	total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2140	if (total_bytes_to_map) {
2141		if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2142			zap_page_range_single(vma, address, total_bytes_to_map,
2143					      NULL);
2144		zc->length = total_bytes_to_map;
2145		zc->recv_skip_hint = 0;
2146	} else {
2147		zc->length = avail_len;
2148		zc->recv_skip_hint = avail_len;
2149	}
2150	ret = 0;
2151	while (length + PAGE_SIZE <= zc->length) {
2152		int mappable_offset;
2153		struct page *page;
2154
2155		if (zc->recv_skip_hint < PAGE_SIZE) {
2156			u32 offset_frag;
2157
2158			if (skb) {
2159				if (zc->recv_skip_hint > 0)
2160					break;
2161				skb = skb->next;
2162				offset = seq - TCP_SKB_CB(skb)->seq;
2163			} else {
2164				skb = tcp_recv_skb(sk, seq, &offset);
2165			}
2166
2167			if (!skb_frags_readable(skb))
2168				break;
2169
2170			if (TCP_SKB_CB(skb)->has_rxtstamp) {
2171				tcp_update_recv_tstamps(skb, tss);
2172				zc->msg_flags |= TCP_CMSG_TS;
2173			}
2174			zc->recv_skip_hint = skb->len - offset;
2175			frags = skb_advance_to_frag(skb, offset, &offset_frag);
2176			if (!frags || offset_frag)
2177				break;
 
 
 
 
 
 
 
2178		}
 
 
2179
2180		mappable_offset = find_next_mappable_frag(frags,
2181							  zc->recv_skip_hint);
2182		if (mappable_offset) {
2183			zc->recv_skip_hint = mappable_offset;
 
 
2184			break;
2185		}
2186		page = skb_frag_page(frags);
2187		if (WARN_ON_ONCE(!page))
 
2188			break;
2189
2190		prefetchw(page);
2191		pages[pages_to_map++] = page;
2192		length += PAGE_SIZE;
 
2193		zc->recv_skip_hint -= PAGE_SIZE;
2194		frags++;
2195		if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2196		    zc->recv_skip_hint < PAGE_SIZE) {
2197			/* Either full batch, or we're about to go to next skb
2198			 * (and we cannot unroll failed ops across skbs).
2199			 */
2200			ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2201							   pages_to_map,
2202							   &address, &length,
2203							   &seq, zc,
2204							   total_bytes_to_map);
2205			if (ret)
2206				goto out;
2207			pages_to_map = 0;
2208		}
2209	}
2210	if (pages_to_map) {
2211		ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2212						   &address, &length, &seq,
2213						   zc, total_bytes_to_map);
2214	}
2215out:
2216	if (mmap_locked)
2217		mmap_read_unlock(current->mm);
2218	else
2219		vma_end_read(vma);
2220	/* Try to copy straggler data. */
2221	if (!ret)
2222		copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2223
2224	if (length + copylen) {
2225		WRITE_ONCE(tp->copied_seq, seq);
2226		tcp_rcv_space_adjust(sk);
2227
2228		/* Clean up data we have read: This will do ACK frames. */
2229		tcp_recv_skb(sk, seq, &offset);
2230		tcp_cleanup_rbuf(sk, length + copylen);
2231		ret = 0;
2232		if (length == zc->length)
2233			zc->recv_skip_hint = 0;
2234	} else {
2235		if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2236			ret = -EIO;
2237	}
2238	zc->length = length;
2239	return ret;
2240}
2241#endif
2242
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2243/* Similar to __sock_recv_timestamp, but does not require an skb */
2244void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2245			struct scm_timestamping_internal *tss)
2246{
2247	int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2248	u32 tsflags = READ_ONCE(sk->sk_tsflags);
2249	bool has_timestamping = false;
2250
2251	if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2252		if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2253			if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2254				if (new_tstamp) {
2255					struct __kernel_timespec kts = {
2256						.tv_sec = tss->ts[0].tv_sec,
2257						.tv_nsec = tss->ts[0].tv_nsec,
2258					};
2259					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2260						 sizeof(kts), &kts);
2261				} else {
2262					struct __kernel_old_timespec ts_old = {
2263						.tv_sec = tss->ts[0].tv_sec,
2264						.tv_nsec = tss->ts[0].tv_nsec,
2265					};
2266					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2267						 sizeof(ts_old), &ts_old);
2268				}
2269			} else {
2270				if (new_tstamp) {
2271					struct __kernel_sock_timeval stv = {
2272						.tv_sec = tss->ts[0].tv_sec,
2273						.tv_usec = tss->ts[0].tv_nsec / 1000,
2274					};
2275					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2276						 sizeof(stv), &stv);
2277				} else {
2278					struct __kernel_old_timeval tv = {
2279						.tv_sec = tss->ts[0].tv_sec,
2280						.tv_usec = tss->ts[0].tv_nsec / 1000,
2281					};
2282					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2283						 sizeof(tv), &tv);
2284				}
2285			}
2286		}
2287
2288		if (tsflags & SOF_TIMESTAMPING_SOFTWARE &&
2289		    (tsflags & SOF_TIMESTAMPING_RX_SOFTWARE ||
2290		     !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2291			has_timestamping = true;
2292		else
2293			tss->ts[0] = (struct timespec64) {0};
2294	}
2295
2296	if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2297		if (tsflags & SOF_TIMESTAMPING_RAW_HARDWARE &&
2298		    (tsflags & SOF_TIMESTAMPING_RX_HARDWARE ||
2299		     !(tsflags & SOF_TIMESTAMPING_OPT_RX_FILTER)))
2300			has_timestamping = true;
2301		else
2302			tss->ts[2] = (struct timespec64) {0};
2303	}
2304
2305	if (has_timestamping) {
2306		tss->ts[1] = (struct timespec64) {0};
2307		if (sock_flag(sk, SOCK_TSTAMP_NEW))
2308			put_cmsg_scm_timestamping64(msg, tss);
2309		else
2310			put_cmsg_scm_timestamping(msg, tss);
2311	}
2312}
2313
2314static int tcp_inq_hint(struct sock *sk)
2315{
2316	const struct tcp_sock *tp = tcp_sk(sk);
2317	u32 copied_seq = READ_ONCE(tp->copied_seq);
2318	u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2319	int inq;
2320
2321	inq = rcv_nxt - copied_seq;
2322	if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2323		lock_sock(sk);
2324		inq = tp->rcv_nxt - tp->copied_seq;
2325		release_sock(sk);
2326	}
2327	/* After receiving a FIN, tell the user-space to continue reading
2328	 * by returning a non-zero inq.
2329	 */
2330	if (inq == 0 && sock_flag(sk, SOCK_DONE))
2331		inq = 1;
2332	return inq;
2333}
2334
2335/* batch __xa_alloc() calls and reduce xa_lock()/xa_unlock() overhead. */
2336struct tcp_xa_pool {
2337	u8		max; /* max <= MAX_SKB_FRAGS */
2338	u8		idx; /* idx <= max */
2339	__u32		tokens[MAX_SKB_FRAGS];
2340	netmem_ref	netmems[MAX_SKB_FRAGS];
2341};
2342
2343static void tcp_xa_pool_commit_locked(struct sock *sk, struct tcp_xa_pool *p)
2344{
2345	int i;
2346
2347	/* Commit part that has been copied to user space. */
2348	for (i = 0; i < p->idx; i++)
2349		__xa_cmpxchg(&sk->sk_user_frags, p->tokens[i], XA_ZERO_ENTRY,
2350			     (__force void *)p->netmems[i], GFP_KERNEL);
2351	/* Rollback what has been pre-allocated and is no longer needed. */
2352	for (; i < p->max; i++)
2353		__xa_erase(&sk->sk_user_frags, p->tokens[i]);
2354
2355	p->max = 0;
2356	p->idx = 0;
2357}
2358
2359static void tcp_xa_pool_commit(struct sock *sk, struct tcp_xa_pool *p)
2360{
2361	if (!p->max)
2362		return;
2363
2364	xa_lock_bh(&sk->sk_user_frags);
2365
2366	tcp_xa_pool_commit_locked(sk, p);
2367
2368	xa_unlock_bh(&sk->sk_user_frags);
2369}
2370
2371static int tcp_xa_pool_refill(struct sock *sk, struct tcp_xa_pool *p,
2372			      unsigned int max_frags)
2373{
2374	int err, k;
2375
2376	if (p->idx < p->max)
2377		return 0;
2378
2379	xa_lock_bh(&sk->sk_user_frags);
2380
2381	tcp_xa_pool_commit_locked(sk, p);
2382
2383	for (k = 0; k < max_frags; k++) {
2384		err = __xa_alloc(&sk->sk_user_frags, &p->tokens[k],
2385				 XA_ZERO_ENTRY, xa_limit_31b, GFP_KERNEL);
2386		if (err)
2387			break;
2388	}
2389
2390	xa_unlock_bh(&sk->sk_user_frags);
2391
2392	p->max = k;
2393	p->idx = 0;
2394	return k ? 0 : err;
2395}
2396
2397/* On error, returns the -errno. On success, returns number of bytes sent to the
2398 * user. May not consume all of @remaining_len.
2399 */
2400static int tcp_recvmsg_dmabuf(struct sock *sk, const struct sk_buff *skb,
2401			      unsigned int offset, struct msghdr *msg,
2402			      int remaining_len)
2403{
2404	struct dmabuf_cmsg dmabuf_cmsg = { 0 };
2405	struct tcp_xa_pool tcp_xa_pool;
2406	unsigned int start;
2407	int i, copy, n;
2408	int sent = 0;
2409	int err = 0;
2410
2411	tcp_xa_pool.max = 0;
2412	tcp_xa_pool.idx = 0;
2413	do {
2414		start = skb_headlen(skb);
2415
2416		if (skb_frags_readable(skb)) {
2417			err = -ENODEV;
2418			goto out;
2419		}
2420
2421		/* Copy header. */
2422		copy = start - offset;
2423		if (copy > 0) {
2424			copy = min(copy, remaining_len);
2425
2426			n = copy_to_iter(skb->data + offset, copy,
2427					 &msg->msg_iter);
2428			if (n != copy) {
2429				err = -EFAULT;
2430				goto out;
2431			}
2432
2433			offset += copy;
2434			remaining_len -= copy;
2435
2436			/* First a dmabuf_cmsg for # bytes copied to user
2437			 * buffer.
2438			 */
2439			memset(&dmabuf_cmsg, 0, sizeof(dmabuf_cmsg));
2440			dmabuf_cmsg.frag_size = copy;
2441			err = put_cmsg_notrunc(msg, SOL_SOCKET,
2442					       SO_DEVMEM_LINEAR,
2443					       sizeof(dmabuf_cmsg),
2444					       &dmabuf_cmsg);
2445			if (err)
2446				goto out;
2447
2448			sent += copy;
2449
2450			if (remaining_len == 0)
2451				goto out;
2452		}
2453
2454		/* after that, send information of dmabuf pages through a
2455		 * sequence of cmsg
2456		 */
2457		for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2458			skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2459			struct net_iov *niov;
2460			u64 frag_offset;
2461			int end;
2462
2463			/* !skb_frags_readable() should indicate that ALL the
2464			 * frags in this skb are dmabuf net_iovs. We're checking
2465			 * for that flag above, but also check individual frags
2466			 * here. If the tcp stack is not setting
2467			 * skb_frags_readable() correctly, we still don't want
2468			 * to crash here.
2469			 */
2470			if (!skb_frag_net_iov(frag)) {
2471				net_err_ratelimited("Found non-dmabuf skb with net_iov");
2472				err = -ENODEV;
2473				goto out;
2474			}
2475
2476			niov = skb_frag_net_iov(frag);
2477			end = start + skb_frag_size(frag);
2478			copy = end - offset;
2479
2480			if (copy > 0) {
2481				copy = min(copy, remaining_len);
2482
2483				frag_offset = net_iov_virtual_addr(niov) +
2484					      skb_frag_off(frag) + offset -
2485					      start;
2486				dmabuf_cmsg.frag_offset = frag_offset;
2487				dmabuf_cmsg.frag_size = copy;
2488				err = tcp_xa_pool_refill(sk, &tcp_xa_pool,
2489							 skb_shinfo(skb)->nr_frags - i);
2490				if (err)
2491					goto out;
2492
2493				/* Will perform the exchange later */
2494				dmabuf_cmsg.frag_token = tcp_xa_pool.tokens[tcp_xa_pool.idx];
2495				dmabuf_cmsg.dmabuf_id = net_iov_binding_id(niov);
2496
2497				offset += copy;
2498				remaining_len -= copy;
2499
2500				err = put_cmsg_notrunc(msg, SOL_SOCKET,
2501						       SO_DEVMEM_DMABUF,
2502						       sizeof(dmabuf_cmsg),
2503						       &dmabuf_cmsg);
2504				if (err)
2505					goto out;
2506
2507				atomic_long_inc(&niov->pp_ref_count);
2508				tcp_xa_pool.netmems[tcp_xa_pool.idx++] = skb_frag_netmem(frag);
2509
2510				sent += copy;
2511
2512				if (remaining_len == 0)
2513					goto out;
2514			}
2515			start = end;
2516		}
2517
2518		tcp_xa_pool_commit(sk, &tcp_xa_pool);
2519		if (!remaining_len)
2520			goto out;
2521
2522		/* if remaining_len is not satisfied yet, we need to go to the
2523		 * next frag in the frag_list to satisfy remaining_len.
2524		 */
2525		skb = skb_shinfo(skb)->frag_list ?: skb->next;
2526
2527		offset = offset - start;
2528	} while (skb);
2529
2530	if (remaining_len) {
2531		err = -EFAULT;
2532		goto out;
2533	}
2534
2535out:
2536	tcp_xa_pool_commit(sk, &tcp_xa_pool);
2537	if (!sent)
2538		sent = err;
2539
2540	return sent;
2541}
2542
2543/*
2544 *	This routine copies from a sock struct into the user buffer.
2545 *
2546 *	Technical note: in 2.3 we work on _locked_ socket, so that
2547 *	tricks with *seq access order and skb->users are not required.
2548 *	Probably, code can be easily improved even more.
2549 */
2550
2551static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2552			      int flags, struct scm_timestamping_internal *tss,
2553			      int *cmsg_flags)
2554{
2555	struct tcp_sock *tp = tcp_sk(sk);
2556	int last_copied_dmabuf = -1; /* uninitialized */
2557	int copied = 0;
2558	u32 peek_seq;
2559	u32 *seq;
2560	unsigned long used;
2561	int err;
2562	int target;		/* Read at least this many bytes */
2563	long timeo;
2564	struct sk_buff *skb, *last;
2565	u32 peek_offset = 0;
2566	u32 urg_hole = 0;
 
 
 
 
 
 
 
 
 
 
 
 
2567
2568	err = -ENOTCONN;
2569	if (sk->sk_state == TCP_LISTEN)
2570		goto out;
2571
2572	if (tp->recvmsg_inq) {
2573		*cmsg_flags = TCP_CMSG_INQ;
2574		msg->msg_get_inq = 1;
2575	}
2576	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2577
2578	/* Urgent data needs to be handled specially. */
2579	if (flags & MSG_OOB)
2580		goto recv_urg;
2581
2582	if (unlikely(tp->repair)) {
2583		err = -EPERM;
2584		if (!(flags & MSG_PEEK))
2585			goto out;
2586
2587		if (tp->repair_queue == TCP_SEND_QUEUE)
2588			goto recv_sndq;
2589
2590		err = -EINVAL;
2591		if (tp->repair_queue == TCP_NO_QUEUE)
2592			goto out;
2593
2594		/* 'common' recv queue MSG_PEEK-ing */
2595	}
2596
2597	seq = &tp->copied_seq;
2598	if (flags & MSG_PEEK) {
2599		peek_offset = max(sk_peek_offset(sk, flags), 0);
2600		peek_seq = tp->copied_seq + peek_offset;
2601		seq = &peek_seq;
2602	}
2603
2604	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2605
2606	do {
2607		u32 offset;
2608
2609		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2610		if (unlikely(tp->urg_data) && tp->urg_seq == *seq) {
2611			if (copied)
2612				break;
2613			if (signal_pending(current)) {
2614				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2615				break;
2616			}
2617		}
2618
2619		/* Next get a buffer. */
2620
2621		last = skb_peek_tail(&sk->sk_receive_queue);
2622		skb_queue_walk(&sk->sk_receive_queue, skb) {
2623			last = skb;
2624			/* Now that we have two receive queues this
2625			 * shouldn't happen.
2626			 */
2627			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2628				 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2629				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2630				 flags))
2631				break;
2632
2633			offset = *seq - TCP_SKB_CB(skb)->seq;
2634			if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2635				pr_err_once("%s: found a SYN, please report !\n", __func__);
2636				offset--;
2637			}
2638			if (offset < skb->len)
2639				goto found_ok_skb;
2640			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2641				goto found_fin_ok;
2642			WARN(!(flags & MSG_PEEK),
2643			     "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2644			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2645		}
2646
2647		/* Well, if we have backlog, try to process it now yet. */
2648
2649		if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2650			break;
2651
2652		if (copied) {
2653			if (!timeo ||
2654			    sk->sk_err ||
2655			    sk->sk_state == TCP_CLOSE ||
2656			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
 
2657			    signal_pending(current))
2658				break;
2659		} else {
2660			if (sock_flag(sk, SOCK_DONE))
2661				break;
2662
2663			if (sk->sk_err) {
2664				copied = sock_error(sk);
2665				break;
2666			}
2667
2668			if (sk->sk_shutdown & RCV_SHUTDOWN)
2669				break;
2670
2671			if (sk->sk_state == TCP_CLOSE) {
2672				/* This occurs when user tries to read
2673				 * from never connected socket.
2674				 */
2675				copied = -ENOTCONN;
2676				break;
2677			}
2678
2679			if (!timeo) {
2680				copied = -EAGAIN;
2681				break;
2682			}
2683
2684			if (signal_pending(current)) {
2685				copied = sock_intr_errno(timeo);
2686				break;
2687			}
2688		}
2689
 
 
2690		if (copied >= target) {
2691			/* Do not sleep, just process backlog. */
2692			__sk_flush_backlog(sk);
 
2693		} else {
2694			tcp_cleanup_rbuf(sk, copied);
2695			err = sk_wait_data(sk, &timeo, last);
2696			if (err < 0) {
2697				err = copied ? : err;
2698				goto out;
2699			}
2700		}
2701
2702		if ((flags & MSG_PEEK) &&
2703		    (peek_seq - peek_offset - copied - urg_hole != tp->copied_seq)) {
2704			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2705					    current->comm,
2706					    task_pid_nr(current));
2707			peek_seq = tp->copied_seq + peek_offset;
2708		}
2709		continue;
2710
2711found_ok_skb:
2712		/* Ok so how much can we use? */
2713		used = skb->len - offset;
2714		if (len < used)
2715			used = len;
2716
2717		/* Do we have urgent data here? */
2718		if (unlikely(tp->urg_data)) {
2719			u32 urg_offset = tp->urg_seq - *seq;
2720			if (urg_offset < used) {
2721				if (!urg_offset) {
2722					if (!sock_flag(sk, SOCK_URGINLINE)) {
2723						WRITE_ONCE(*seq, *seq + 1);
2724						urg_hole++;
2725						offset++;
2726						used--;
2727						if (!used)
2728							goto skip_copy;
2729					}
2730				} else
2731					used = urg_offset;
2732			}
2733		}
2734
2735		if (!(flags & MSG_TRUNC)) {
2736			if (last_copied_dmabuf != -1 &&
2737			    last_copied_dmabuf != !skb_frags_readable(skb))
 
 
 
2738				break;
2739
2740			if (skb_frags_readable(skb)) {
2741				err = skb_copy_datagram_msg(skb, offset, msg,
2742							    used);
2743				if (err) {
2744					/* Exception. Bailout! */
2745					if (!copied)
2746						copied = -EFAULT;
2747					break;
2748				}
2749			} else {
2750				if (!(flags & MSG_SOCK_DEVMEM)) {
2751					/* dmabuf skbs can only be received
2752					 * with the MSG_SOCK_DEVMEM flag.
2753					 */
2754					if (!copied)
2755						copied = -EFAULT;
2756
2757					break;
2758				}
2759
2760				err = tcp_recvmsg_dmabuf(sk, skb, offset, msg,
2761							 used);
2762				if (err <= 0) {
2763					if (!copied)
2764						copied = -EFAULT;
2765
2766					break;
2767				}
2768				used = err;
2769			}
2770		}
2771
2772		last_copied_dmabuf = !skb_frags_readable(skb);
2773
2774		WRITE_ONCE(*seq, *seq + used);
2775		copied += used;
2776		len -= used;
2777		if (flags & MSG_PEEK)
2778			sk_peek_offset_fwd(sk, used);
2779		else
2780			sk_peek_offset_bwd(sk, used);
2781		tcp_rcv_space_adjust(sk);
2782
2783skip_copy:
2784		if (unlikely(tp->urg_data) && after(tp->copied_seq, tp->urg_seq)) {
2785			WRITE_ONCE(tp->urg_data, 0);
2786			tcp_fast_path_check(sk);
2787		}
2788
2789		if (TCP_SKB_CB(skb)->has_rxtstamp) {
2790			tcp_update_recv_tstamps(skb, tss);
2791			*cmsg_flags |= TCP_CMSG_TS;
2792		}
2793
2794		if (used + offset < skb->len)
2795			continue;
2796
 
 
 
 
 
2797		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2798			goto found_fin_ok;
2799		if (!(flags & MSG_PEEK))
2800			tcp_eat_recv_skb(sk, skb);
2801		continue;
2802
2803found_fin_ok:
2804		/* Process the FIN. */
2805		WRITE_ONCE(*seq, *seq + 1);
2806		if (!(flags & MSG_PEEK))
2807			tcp_eat_recv_skb(sk, skb);
2808		break;
2809	} while (len > 0);
2810
2811	/* According to UNIX98, msg_name/msg_namelen are ignored
2812	 * on connected socket. I was just happy when found this 8) --ANK
2813	 */
2814
2815	/* Clean up data we have read: This will do ACK frames. */
2816	tcp_cleanup_rbuf(sk, copied);
 
 
 
 
 
 
 
 
 
 
 
 
2817	return copied;
2818
2819out:
 
2820	return err;
2821
2822recv_urg:
2823	err = tcp_recv_urg(sk, msg, len, flags);
2824	goto out;
2825
2826recv_sndq:
2827	err = tcp_peek_sndq(sk, msg, len);
2828	goto out;
2829}
2830
2831int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int flags,
2832		int *addr_len)
2833{
2834	int cmsg_flags = 0, ret;
2835	struct scm_timestamping_internal tss;
2836
2837	if (unlikely(flags & MSG_ERRQUEUE))
2838		return inet_recv_error(sk, msg, len, addr_len);
2839
2840	if (sk_can_busy_loop(sk) &&
2841	    skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2842	    sk->sk_state == TCP_ESTABLISHED)
2843		sk_busy_loop(sk, flags & MSG_DONTWAIT);
2844
2845	lock_sock(sk);
2846	ret = tcp_recvmsg_locked(sk, msg, len, flags, &tss, &cmsg_flags);
2847	release_sock(sk);
2848
2849	if ((cmsg_flags || msg->msg_get_inq) && ret >= 0) {
2850		if (cmsg_flags & TCP_CMSG_TS)
2851			tcp_recv_timestamp(msg, sk, &tss);
2852		if (msg->msg_get_inq) {
2853			msg->msg_inq = tcp_inq_hint(sk);
2854			if (cmsg_flags & TCP_CMSG_INQ)
2855				put_cmsg(msg, SOL_TCP, TCP_CM_INQ,
2856					 sizeof(msg->msg_inq), &msg->msg_inq);
2857		}
2858	}
2859	return ret;
2860}
2861EXPORT_SYMBOL(tcp_recvmsg);
2862
2863void tcp_set_state(struct sock *sk, int state)
2864{
2865	int oldstate = sk->sk_state;
2866
2867	/* We defined a new enum for TCP states that are exported in BPF
2868	 * so as not force the internal TCP states to be frozen. The
2869	 * following checks will detect if an internal state value ever
2870	 * differs from the BPF value. If this ever happens, then we will
2871	 * need to remap the internal value to the BPF value before calling
2872	 * tcp_call_bpf_2arg.
2873	 */
2874	BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2875	BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2876	BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2877	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2878	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2879	BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2880	BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2881	BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2882	BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2883	BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2884	BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2885	BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2886	BUILD_BUG_ON((int)BPF_TCP_BOUND_INACTIVE != (int)TCP_BOUND_INACTIVE);
2887	BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2888
2889	/* bpf uapi header bpf.h defines an anonymous enum with values
2890	 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2891	 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2892	 * But clang built vmlinux does not have this enum in DWARF
2893	 * since clang removes the above code before generating IR/debuginfo.
2894	 * Let us explicitly emit the type debuginfo to ensure the
2895	 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2896	 * regardless of which compiler is used.
2897	 */
2898	BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2899
2900	if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2901		tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2902
2903	switch (state) {
2904	case TCP_ESTABLISHED:
2905		if (oldstate != TCP_ESTABLISHED)
2906			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2907		break;
2908	case TCP_CLOSE_WAIT:
2909		if (oldstate == TCP_SYN_RECV)
2910			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2911		break;
2912
2913	case TCP_CLOSE:
2914		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2915			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2916
2917		sk->sk_prot->unhash(sk);
2918		if (inet_csk(sk)->icsk_bind_hash &&
2919		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2920			inet_put_port(sk);
2921		fallthrough;
2922	default:
2923		if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
2924			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2925	}
2926
2927	/* Change state AFTER socket is unhashed to avoid closed
2928	 * socket sitting in hash tables.
2929	 */
2930	inet_sk_state_store(sk, state);
2931}
2932EXPORT_SYMBOL_GPL(tcp_set_state);
2933
2934/*
2935 *	State processing on a close. This implements the state shift for
2936 *	sending our FIN frame. Note that we only send a FIN for some
2937 *	states. A shutdown() may have already sent the FIN, or we may be
2938 *	closed.
2939 */
2940
2941static const unsigned char new_state[16] = {
2942  /* current state:        new state:      action:	*/
2943  [0 /* (Invalid) */]	= TCP_CLOSE,
2944  [TCP_ESTABLISHED]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2945  [TCP_SYN_SENT]	= TCP_CLOSE,
2946  [TCP_SYN_RECV]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2947  [TCP_FIN_WAIT1]	= TCP_FIN_WAIT1,
2948  [TCP_FIN_WAIT2]	= TCP_FIN_WAIT2,
2949  [TCP_TIME_WAIT]	= TCP_CLOSE,
2950  [TCP_CLOSE]		= TCP_CLOSE,
2951  [TCP_CLOSE_WAIT]	= TCP_LAST_ACK  | TCP_ACTION_FIN,
2952  [TCP_LAST_ACK]	= TCP_LAST_ACK,
2953  [TCP_LISTEN]		= TCP_CLOSE,
2954  [TCP_CLOSING]		= TCP_CLOSING,
2955  [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
2956};
2957
2958static int tcp_close_state(struct sock *sk)
2959{
2960	int next = (int)new_state[sk->sk_state];
2961	int ns = next & TCP_STATE_MASK;
2962
2963	tcp_set_state(sk, ns);
2964
2965	return next & TCP_ACTION_FIN;
2966}
2967
2968/*
2969 *	Shutdown the sending side of a connection. Much like close except
2970 *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2971 */
2972
2973void tcp_shutdown(struct sock *sk, int how)
2974{
2975	/*	We need to grab some memory, and put together a FIN,
2976	 *	and then put it into the queue to be sent.
2977	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2978	 */
2979	if (!(how & SEND_SHUTDOWN))
2980		return;
2981
2982	/* If we've already sent a FIN, or it's a closed state, skip this. */
2983	if ((1 << sk->sk_state) &
2984	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2985	     TCPF_CLOSE_WAIT)) {
2986		/* Clear out any half completed packets.  FIN if needed. */
2987		if (tcp_close_state(sk))
2988			tcp_send_fin(sk);
2989	}
2990}
2991EXPORT_SYMBOL(tcp_shutdown);
2992
2993int tcp_orphan_count_sum(void)
2994{
2995	int i, total = 0;
2996
2997	for_each_possible_cpu(i)
2998		total += per_cpu(tcp_orphan_count, i);
2999
3000	return max(total, 0);
3001}
3002
3003static int tcp_orphan_cache;
3004static struct timer_list tcp_orphan_timer;
3005#define TCP_ORPHAN_TIMER_PERIOD msecs_to_jiffies(100)
3006
3007static void tcp_orphan_update(struct timer_list *unused)
3008{
3009	WRITE_ONCE(tcp_orphan_cache, tcp_orphan_count_sum());
3010	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
3011}
3012
3013static bool tcp_too_many_orphans(int shift)
3014{
3015	return READ_ONCE(tcp_orphan_cache) << shift >
3016		READ_ONCE(sysctl_tcp_max_orphans);
3017}
3018
3019static bool tcp_out_of_memory(const struct sock *sk)
3020{
3021	if (sk->sk_wmem_queued > SOCK_MIN_SNDBUF &&
3022	    sk_memory_allocated(sk) > sk_prot_mem_limits(sk, 2))
3023		return true;
3024	return false;
3025}
3026
3027bool tcp_check_oom(const struct sock *sk, int shift)
3028{
3029	bool too_many_orphans, out_of_socket_memory;
3030
3031	too_many_orphans = tcp_too_many_orphans(shift);
3032	out_of_socket_memory = tcp_out_of_memory(sk);
3033
3034	if (too_many_orphans)
3035		net_info_ratelimited("too many orphaned sockets\n");
3036	if (out_of_socket_memory)
3037		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
3038	return too_many_orphans || out_of_socket_memory;
3039}
3040
3041void __tcp_close(struct sock *sk, long timeout)
3042{
3043	struct sk_buff *skb;
3044	int data_was_unread = 0;
3045	int state;
3046
3047	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
 
3048
3049	if (sk->sk_state == TCP_LISTEN) {
3050		tcp_set_state(sk, TCP_CLOSE);
3051
3052		/* Special case. */
3053		inet_csk_listen_stop(sk);
3054
3055		goto adjudge_to_death;
3056	}
3057
3058	/*  We need to flush the recv. buffs.  We do this only on the
3059	 *  descriptor close, not protocol-sourced closes, because the
3060	 *  reader process may not have drained the data yet!
3061	 */
3062	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
3063		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
3064
3065		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
3066			len--;
3067		data_was_unread += len;
3068		__kfree_skb(skb);
3069	}
3070
 
 
3071	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
3072	if (sk->sk_state == TCP_CLOSE)
3073		goto adjudge_to_death;
3074
3075	/* As outlined in RFC 2525, section 2.17, we send a RST here because
3076	 * data was lost. To witness the awful effects of the old behavior of
3077	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
3078	 * GET in an FTP client, suspend the process, wait for the client to
3079	 * advertise a zero window, then kill -9 the FTP client, wheee...
3080	 * Note: timeout is always zero in such a case.
3081	 */
3082	if (unlikely(tcp_sk(sk)->repair)) {
3083		sk->sk_prot->disconnect(sk, 0);
3084	} else if (data_was_unread) {
3085		/* Unread data was tossed, zap the connection. */
3086		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
3087		tcp_set_state(sk, TCP_CLOSE);
3088		tcp_send_active_reset(sk, sk->sk_allocation,
3089				      SK_RST_REASON_TCP_ABORT_ON_CLOSE);
3090	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
3091		/* Check zero linger _after_ checking for unread data. */
3092		sk->sk_prot->disconnect(sk, 0);
3093		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
3094	} else if (tcp_close_state(sk)) {
3095		/* We FIN if the application ate all the data before
3096		 * zapping the connection.
3097		 */
3098
3099		/* RED-PEN. Formally speaking, we have broken TCP state
3100		 * machine. State transitions:
3101		 *
3102		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
3103		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (it is difficult)
3104		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
3105		 *
3106		 * are legal only when FIN has been sent (i.e. in window),
3107		 * rather than queued out of window. Purists blame.
3108		 *
3109		 * F.e. "RFC state" is ESTABLISHED,
3110		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
3111		 *
3112		 * The visible declinations are that sometimes
3113		 * we enter time-wait state, when it is not required really
3114		 * (harmless), do not send active resets, when they are
3115		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
3116		 * they look as CLOSING or LAST_ACK for Linux)
3117		 * Probably, I missed some more holelets.
3118		 * 						--ANK
3119		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
3120		 * in a single packet! (May consider it later but will
3121		 * probably need API support or TCP_CORK SYN-ACK until
3122		 * data is written and socket is closed.)
3123		 */
3124		tcp_send_fin(sk);
3125	}
3126
3127	sk_stream_wait_close(sk, timeout);
3128
3129adjudge_to_death:
3130	state = sk->sk_state;
3131	sock_hold(sk);
3132	sock_orphan(sk);
3133
3134	local_bh_disable();
3135	bh_lock_sock(sk);
3136	/* remove backlog if any, without releasing ownership. */
3137	__release_sock(sk);
3138
3139	this_cpu_inc(tcp_orphan_count);
3140
3141	/* Have we already been destroyed by a softirq or backlog? */
3142	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
3143		goto out;
3144
3145	/*	This is a (useful) BSD violating of the RFC. There is a
3146	 *	problem with TCP as specified in that the other end could
3147	 *	keep a socket open forever with no application left this end.
3148	 *	We use a 1 minute timeout (about the same as BSD) then kill
3149	 *	our end. If they send after that then tough - BUT: long enough
3150	 *	that we won't make the old 4*rto = almost no time - whoops
3151	 *	reset mistake.
3152	 *
3153	 *	Nope, it was not mistake. It is really desired behaviour
3154	 *	f.e. on http servers, when such sockets are useless, but
3155	 *	consume significant resources. Let's do it with special
3156	 *	linger2	option.					--ANK
3157	 */
3158
3159	if (sk->sk_state == TCP_FIN_WAIT2) {
3160		struct tcp_sock *tp = tcp_sk(sk);
3161		if (READ_ONCE(tp->linger2) < 0) {
3162			tcp_set_state(sk, TCP_CLOSE);
3163			tcp_send_active_reset(sk, GFP_ATOMIC,
3164					      SK_RST_REASON_TCP_ABORT_ON_LINGER);
3165			__NET_INC_STATS(sock_net(sk),
3166					LINUX_MIB_TCPABORTONLINGER);
3167		} else {
3168			const int tmo = tcp_fin_time(sk);
3169
3170			if (tmo > TCP_TIMEWAIT_LEN) {
3171				inet_csk_reset_keepalive_timer(sk,
3172						tmo - TCP_TIMEWAIT_LEN);
3173			} else {
3174				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
3175				goto out;
3176			}
3177		}
3178	}
3179	if (sk->sk_state != TCP_CLOSE) {
 
3180		if (tcp_check_oom(sk, 0)) {
3181			tcp_set_state(sk, TCP_CLOSE);
3182			tcp_send_active_reset(sk, GFP_ATOMIC,
3183					      SK_RST_REASON_TCP_ABORT_ON_MEMORY);
3184			__NET_INC_STATS(sock_net(sk),
3185					LINUX_MIB_TCPABORTONMEMORY);
3186		} else if (!check_net(sock_net(sk))) {
3187			/* Not possible to send reset; just close */
3188			tcp_set_state(sk, TCP_CLOSE);
3189		}
3190	}
3191
3192	if (sk->sk_state == TCP_CLOSE) {
3193		struct request_sock *req;
3194
3195		req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
3196						lockdep_sock_is_held(sk));
3197		/* We could get here with a non-NULL req if the socket is
3198		 * aborted (e.g., closed with unread data) before 3WHS
3199		 * finishes.
3200		 */
3201		if (req)
3202			reqsk_fastopen_remove(sk, req, false);
3203		inet_csk_destroy_sock(sk);
3204	}
3205	/* Otherwise, socket is reprieved until protocol close. */
3206
3207out:
3208	bh_unlock_sock(sk);
3209	local_bh_enable();
3210}
3211
3212void tcp_close(struct sock *sk, long timeout)
3213{
3214	lock_sock(sk);
3215	__tcp_close(sk, timeout);
3216	release_sock(sk);
3217	if (!sk->sk_net_refcnt)
3218		inet_csk_clear_xmit_timers_sync(sk);
3219	sock_put(sk);
3220}
3221EXPORT_SYMBOL(tcp_close);
3222
3223/* These states need RST on ABORT according to RFC793 */
3224
3225static inline bool tcp_need_reset(int state)
3226{
3227	return (1 << state) &
3228	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
3229		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
3230}
3231
3232static void tcp_rtx_queue_purge(struct sock *sk)
3233{
3234	struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
3235
3236	tcp_sk(sk)->highest_sack = NULL;
3237	while (p) {
3238		struct sk_buff *skb = rb_to_skb(p);
3239
3240		p = rb_next(p);
3241		/* Since we are deleting whole queue, no need to
3242		 * list_del(&skb->tcp_tsorted_anchor)
3243		 */
3244		tcp_rtx_queue_unlink(skb, sk);
3245		tcp_wmem_free_skb(sk, skb);
3246	}
3247}
3248
3249void tcp_write_queue_purge(struct sock *sk)
3250{
3251	struct sk_buff *skb;
3252
3253	tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
3254	while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
3255		tcp_skb_tsorted_anchor_cleanup(skb);
3256		tcp_wmem_free_skb(sk, skb);
3257	}
3258	tcp_rtx_queue_purge(sk);
 
 
 
 
 
3259	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
 
3260	tcp_clear_all_retrans_hints(tcp_sk(sk));
3261	tcp_sk(sk)->packets_out = 0;
3262	inet_csk(sk)->icsk_backoff = 0;
3263}
3264
3265int tcp_disconnect(struct sock *sk, int flags)
3266{
3267	struct inet_sock *inet = inet_sk(sk);
3268	struct inet_connection_sock *icsk = inet_csk(sk);
3269	struct tcp_sock *tp = tcp_sk(sk);
3270	int old_state = sk->sk_state;
3271	u32 seq;
3272
3273	if (old_state != TCP_CLOSE)
3274		tcp_set_state(sk, TCP_CLOSE);
3275
3276	/* ABORT function of RFC793 */
3277	if (old_state == TCP_LISTEN) {
3278		inet_csk_listen_stop(sk);
3279	} else if (unlikely(tp->repair)) {
3280		WRITE_ONCE(sk->sk_err, ECONNABORTED);
3281	} else if (tcp_need_reset(old_state)) {
3282		tcp_send_active_reset(sk, gfp_any(), SK_RST_REASON_TCP_STATE);
3283		WRITE_ONCE(sk->sk_err, ECONNRESET);
3284	} else if (tp->snd_nxt != tp->write_seq &&
3285		   (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK)) {
3286		/* The last check adjusts for discrepancy of Linux wrt. RFC
3287		 * states
3288		 */
3289		tcp_send_active_reset(sk, gfp_any(),
3290				      SK_RST_REASON_TCP_DISCONNECT_WITH_DATA);
3291		WRITE_ONCE(sk->sk_err, ECONNRESET);
3292	} else if (old_state == TCP_SYN_SENT)
3293		WRITE_ONCE(sk->sk_err, ECONNRESET);
3294
3295	tcp_clear_xmit_timers(sk);
3296	__skb_queue_purge(&sk->sk_receive_queue);
 
 
 
 
3297	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3298	WRITE_ONCE(tp->urg_data, 0);
3299	sk_set_peek_off(sk, -1);
3300	tcp_write_queue_purge(sk);
3301	tcp_fastopen_active_disable_ofo_check(sk);
3302	skb_rbtree_purge(&tp->out_of_order_queue);
3303
3304	inet->inet_dport = 0;
3305
3306	inet_bhash2_reset_saddr(sk);
 
3307
3308	WRITE_ONCE(sk->sk_shutdown, 0);
3309	sock_reset_flag(sk, SOCK_DONE);
3310	tp->srtt_us = 0;
3311	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
3312	tp->rcv_rtt_last_tsecr = 0;
3313
3314	seq = tp->write_seq + tp->max_window + 2;
3315	if (!seq)
3316		seq = 1;
3317	WRITE_ONCE(tp->write_seq, seq);
3318
3319	icsk->icsk_backoff = 0;
 
3320	icsk->icsk_probes_out = 0;
3321	icsk->icsk_probes_tstamp = 0;
3322	icsk->icsk_rto = TCP_TIMEOUT_INIT;
3323	icsk->icsk_rto_min = TCP_RTO_MIN;
3324	icsk->icsk_delack_max = TCP_DELACK_MAX;
3325	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
3326	tcp_snd_cwnd_set(tp, TCP_INIT_CWND);
3327	tp->snd_cwnd_cnt = 0;
3328	tp->is_cwnd_limited = 0;
3329	tp->max_packets_out = 0;
3330	tp->window_clamp = 0;
3331	tp->delivered = 0;
3332	tp->delivered_ce = 0;
3333	if (icsk->icsk_ca_initialized && icsk->icsk_ca_ops->release)
3334		icsk->icsk_ca_ops->release(sk);
3335	memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3336	icsk->icsk_ca_initialized = 0;
3337	tcp_set_ca_state(sk, TCP_CA_Open);
3338	tp->is_sack_reneg = 0;
3339	tcp_clear_retrans(tp);
3340	tp->total_retrans = 0;
3341	inet_csk_delack_init(sk);
3342	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3343	 * issue in __tcp_select_window()
3344	 */
3345	icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3346	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3347	__sk_dst_reset(sk);
3348	dst_release(unrcu_pointer(xchg(&sk->sk_rx_dst, NULL)));
 
3349	tcp_saved_syn_free(tp);
3350	tp->compressed_ack = 0;
3351	tp->segs_in = 0;
3352	tp->segs_out = 0;
3353	tp->bytes_sent = 0;
3354	tp->bytes_acked = 0;
3355	tp->bytes_received = 0;
3356	tp->bytes_retrans = 0;
3357	tp->data_segs_in = 0;
3358	tp->data_segs_out = 0;
3359	tp->duplicate_sack[0].start_seq = 0;
3360	tp->duplicate_sack[0].end_seq = 0;
3361	tp->dsack_dups = 0;
3362	tp->reord_seen = 0;
3363	tp->retrans_out = 0;
3364	tp->sacked_out = 0;
3365	tp->tlp_high_seq = 0;
3366	tp->last_oow_ack_time = 0;
3367	tp->plb_rehash = 0;
3368	/* There's a bubble in the pipe until at least the first ACK. */
3369	tp->app_limited = ~0U;
3370	tp->rate_app_limited = 1;
3371	tp->rack.mstamp = 0;
3372	tp->rack.advanced = 0;
3373	tp->rack.reo_wnd_steps = 1;
3374	tp->rack.last_delivered = 0;
3375	tp->rack.reo_wnd_persist = 0;
3376	tp->rack.dsack_seen = 0;
3377	tp->syn_data_acked = 0;
3378	tp->rx_opt.saw_tstamp = 0;
3379	tp->rx_opt.dsack = 0;
3380	tp->rx_opt.num_sacks = 0;
3381	tp->rcv_ooopack = 0;
3382
3383
3384	/* Clean up fastopen related fields */
3385	tcp_free_fastopen_req(tp);
3386	inet_clear_bit(DEFER_CONNECT, sk);
3387	tp->fastopen_client_fail = 0;
3388
3389	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3390
3391	if (sk->sk_frag.page) {
3392		put_page(sk->sk_frag.page);
3393		sk->sk_frag.page = NULL;
3394		sk->sk_frag.offset = 0;
3395	}
3396	sk_error_report(sk);
 
3397	return 0;
3398}
3399EXPORT_SYMBOL(tcp_disconnect);
3400
3401static inline bool tcp_can_repair_sock(const struct sock *sk)
3402{
3403	return sockopt_ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3404		(sk->sk_state != TCP_LISTEN);
3405}
3406
3407static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3408{
3409	struct tcp_repair_window opt;
3410
3411	if (!tp->repair)
3412		return -EPERM;
3413
3414	if (len != sizeof(opt))
3415		return -EINVAL;
3416
3417	if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3418		return -EFAULT;
3419
3420	if (opt.max_window < opt.snd_wnd)
3421		return -EINVAL;
3422
3423	if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3424		return -EINVAL;
3425
3426	if (after(opt.rcv_wup, tp->rcv_nxt))
3427		return -EINVAL;
3428
3429	tp->snd_wl1	= opt.snd_wl1;
3430	tp->snd_wnd	= opt.snd_wnd;
3431	tp->max_window	= opt.max_window;
3432
3433	tp->rcv_wnd	= opt.rcv_wnd;
3434	tp->rcv_wup	= opt.rcv_wup;
3435
3436	return 0;
3437}
3438
3439static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3440		unsigned int len)
3441{
3442	struct tcp_sock *tp = tcp_sk(sk);
3443	struct tcp_repair_opt opt;
3444	size_t offset = 0;
3445
3446	while (len >= sizeof(opt)) {
3447		if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3448			return -EFAULT;
3449
3450		offset += sizeof(opt);
3451		len -= sizeof(opt);
3452
3453		switch (opt.opt_code) {
3454		case TCPOPT_MSS:
3455			tp->rx_opt.mss_clamp = opt.opt_val;
3456			tcp_mtup_init(sk);
3457			break;
3458		case TCPOPT_WINDOW:
3459			{
3460				u16 snd_wscale = opt.opt_val & 0xFFFF;
3461				u16 rcv_wscale = opt.opt_val >> 16;
3462
3463				if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3464					return -EFBIG;
3465
3466				tp->rx_opt.snd_wscale = snd_wscale;
3467				tp->rx_opt.rcv_wscale = rcv_wscale;
3468				tp->rx_opt.wscale_ok = 1;
3469			}
3470			break;
3471		case TCPOPT_SACK_PERM:
3472			if (opt.opt_val != 0)
3473				return -EINVAL;
3474
3475			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3476			break;
3477		case TCPOPT_TIMESTAMP:
3478			if (opt.opt_val != 0)
3479				return -EINVAL;
3480
3481			tp->rx_opt.tstamp_ok = 1;
3482			break;
3483		}
3484	}
3485
3486	return 0;
3487}
3488
3489DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3490EXPORT_SYMBOL(tcp_tx_delay_enabled);
3491
3492static void tcp_enable_tx_delay(void)
3493{
3494	if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3495		static int __tcp_tx_delay_enabled = 0;
3496
3497		if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3498			static_branch_enable(&tcp_tx_delay_enabled);
3499			pr_info("TCP_TX_DELAY enabled\n");
3500		}
3501	}
3502}
3503
3504/* When set indicates to always queue non-full frames.  Later the user clears
3505 * this option and we transmit any pending partial frames in the queue.  This is
3506 * meant to be used alongside sendfile() to get properly filled frames when the
3507 * user (for example) must write out headers with a write() call first and then
3508 * use sendfile to send out the data parts.
3509 *
3510 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3511 * TCP_NODELAY.
3512 */
3513void __tcp_sock_set_cork(struct sock *sk, bool on)
3514{
3515	struct tcp_sock *tp = tcp_sk(sk);
3516
3517	if (on) {
3518		tp->nonagle |= TCP_NAGLE_CORK;
3519	} else {
3520		tp->nonagle &= ~TCP_NAGLE_CORK;
3521		if (tp->nonagle & TCP_NAGLE_OFF)
3522			tp->nonagle |= TCP_NAGLE_PUSH;
3523		tcp_push_pending_frames(sk);
3524	}
3525}
3526
3527void tcp_sock_set_cork(struct sock *sk, bool on)
3528{
3529	lock_sock(sk);
3530	__tcp_sock_set_cork(sk, on);
3531	release_sock(sk);
3532}
3533EXPORT_SYMBOL(tcp_sock_set_cork);
3534
3535/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3536 * remembered, but it is not activated until cork is cleared.
3537 *
3538 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3539 * even TCP_CORK for currently queued segments.
3540 */
3541void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3542{
3543	if (on) {
3544		tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3545		tcp_push_pending_frames(sk);
3546	} else {
3547		tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3548	}
3549}
3550
3551void tcp_sock_set_nodelay(struct sock *sk)
3552{
3553	lock_sock(sk);
3554	__tcp_sock_set_nodelay(sk, true);
3555	release_sock(sk);
3556}
3557EXPORT_SYMBOL(tcp_sock_set_nodelay);
3558
3559static void __tcp_sock_set_quickack(struct sock *sk, int val)
3560{
3561	if (!val) {
3562		inet_csk_enter_pingpong_mode(sk);
3563		return;
3564	}
3565
3566	inet_csk_exit_pingpong_mode(sk);
3567	if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3568	    inet_csk_ack_scheduled(sk)) {
3569		inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3570		tcp_cleanup_rbuf(sk, 1);
3571		if (!(val & 1))
3572			inet_csk_enter_pingpong_mode(sk);
3573	}
3574}
3575
3576void tcp_sock_set_quickack(struct sock *sk, int val)
3577{
3578	lock_sock(sk);
3579	__tcp_sock_set_quickack(sk, val);
3580	release_sock(sk);
3581}
3582EXPORT_SYMBOL(tcp_sock_set_quickack);
3583
3584int tcp_sock_set_syncnt(struct sock *sk, int val)
3585{
3586	if (val < 1 || val > MAX_TCP_SYNCNT)
3587		return -EINVAL;
3588
3589	WRITE_ONCE(inet_csk(sk)->icsk_syn_retries, val);
3590	return 0;
3591}
3592EXPORT_SYMBOL(tcp_sock_set_syncnt);
3593
3594int tcp_sock_set_user_timeout(struct sock *sk, int val)
3595{
3596	/* Cap the max time in ms TCP will retry or probe the window
3597	 * before giving up and aborting (ETIMEDOUT) a connection.
3598	 */
3599	if (val < 0)
3600		return -EINVAL;
3601
3602	WRITE_ONCE(inet_csk(sk)->icsk_user_timeout, val);
3603	return 0;
3604}
3605EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3606
3607int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3608{
3609	struct tcp_sock *tp = tcp_sk(sk);
3610
3611	if (val < 1 || val > MAX_TCP_KEEPIDLE)
3612		return -EINVAL;
3613
3614	/* Paired with WRITE_ONCE() in keepalive_time_when() */
3615	WRITE_ONCE(tp->keepalive_time, val * HZ);
3616	if (sock_flag(sk, SOCK_KEEPOPEN) &&
3617	    !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3618		u32 elapsed = keepalive_time_elapsed(tp);
3619
3620		if (tp->keepalive_time > elapsed)
3621			elapsed = tp->keepalive_time - elapsed;
3622		else
3623			elapsed = 0;
3624		inet_csk_reset_keepalive_timer(sk, elapsed);
3625	}
3626
3627	return 0;
3628}
3629
3630int tcp_sock_set_keepidle(struct sock *sk, int val)
3631{
3632	int err;
3633
3634	lock_sock(sk);
3635	err = tcp_sock_set_keepidle_locked(sk, val);
3636	release_sock(sk);
3637	return err;
3638}
3639EXPORT_SYMBOL(tcp_sock_set_keepidle);
3640
3641int tcp_sock_set_keepintvl(struct sock *sk, int val)
3642{
3643	if (val < 1 || val > MAX_TCP_KEEPINTVL)
3644		return -EINVAL;
3645
3646	WRITE_ONCE(tcp_sk(sk)->keepalive_intvl, val * HZ);
3647	return 0;
3648}
3649EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3650
3651int tcp_sock_set_keepcnt(struct sock *sk, int val)
3652{
3653	if (val < 1 || val > MAX_TCP_KEEPCNT)
3654		return -EINVAL;
3655
3656	/* Paired with READ_ONCE() in keepalive_probes() */
3657	WRITE_ONCE(tcp_sk(sk)->keepalive_probes, val);
3658	return 0;
3659}
3660EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3661
3662int tcp_set_window_clamp(struct sock *sk, int val)
3663{
3664	struct tcp_sock *tp = tcp_sk(sk);
3665
3666	if (!val) {
3667		if (sk->sk_state != TCP_CLOSE)
3668			return -EINVAL;
3669		WRITE_ONCE(tp->window_clamp, 0);
3670	} else {
3671		u32 new_rcv_ssthresh, old_window_clamp = tp->window_clamp;
3672		u32 new_window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3673						SOCK_MIN_RCVBUF / 2 : val;
3674
3675		if (new_window_clamp == old_window_clamp)
3676			return 0;
3677
3678		WRITE_ONCE(tp->window_clamp, new_window_clamp);
3679		if (new_window_clamp < old_window_clamp) {
3680			/* need to apply the reserved mem provisioning only
3681			 * when shrinking the window clamp
3682			 */
3683			__tcp_adjust_rcv_ssthresh(sk, tp->window_clamp);
3684
3685		} else {
3686			new_rcv_ssthresh = min(tp->rcv_wnd, tp->window_clamp);
3687			tp->rcv_ssthresh = max(new_rcv_ssthresh,
3688					       tp->rcv_ssthresh);
3689		}
3690	}
3691	return 0;
3692}
3693
3694/*
3695 *	Socket option code for TCP.
3696 */
3697int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3698		      sockptr_t optval, unsigned int optlen)
3699{
3700	struct tcp_sock *tp = tcp_sk(sk);
3701	struct inet_connection_sock *icsk = inet_csk(sk);
3702	struct net *net = sock_net(sk);
3703	int val;
3704	int err = 0;
3705
3706	/* These are data/string values, all the others are ints */
3707	switch (optname) {
3708	case TCP_CONGESTION: {
3709		char name[TCP_CA_NAME_MAX];
3710
3711		if (optlen < 1)
3712			return -EINVAL;
3713
3714		val = strncpy_from_sockptr(name, optval,
3715					min_t(long, TCP_CA_NAME_MAX-1, optlen));
3716		if (val < 0)
3717			return -EFAULT;
3718		name[val] = 0;
3719
3720		sockopt_lock_sock(sk);
3721		err = tcp_set_congestion_control(sk, name, !has_current_bpf_ctx(),
3722						 sockopt_ns_capable(sock_net(sk)->user_ns,
3723								    CAP_NET_ADMIN));
3724		sockopt_release_sock(sk);
3725		return err;
3726	}
3727	case TCP_ULP: {
3728		char name[TCP_ULP_NAME_MAX];
3729
3730		if (optlen < 1)
3731			return -EINVAL;
3732
3733		val = strncpy_from_sockptr(name, optval,
3734					min_t(long, TCP_ULP_NAME_MAX - 1,
3735					      optlen));
3736		if (val < 0)
3737			return -EFAULT;
3738		name[val] = 0;
3739
3740		sockopt_lock_sock(sk);
3741		err = tcp_set_ulp(sk, name);
3742		sockopt_release_sock(sk);
3743		return err;
3744	}
3745	case TCP_FASTOPEN_KEY: {
3746		__u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3747		__u8 *backup_key = NULL;
3748
3749		/* Allow a backup key as well to facilitate key rotation
3750		 * First key is the active one.
3751		 */
3752		if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3753		    optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3754			return -EINVAL;
3755
3756		if (copy_from_sockptr(key, optval, optlen))
3757			return -EFAULT;
3758
3759		if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3760			backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3761
3762		return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3763	}
3764	default:
3765		/* fallthru */
3766		break;
3767	}
3768
3769	if (optlen < sizeof(int))
3770		return -EINVAL;
3771
3772	if (copy_from_sockptr(&val, optval, sizeof(val)))
3773		return -EFAULT;
3774
3775	/* Handle options that can be set without locking the socket. */
3776	switch (optname) {
3777	case TCP_SYNCNT:
3778		return tcp_sock_set_syncnt(sk, val);
3779	case TCP_USER_TIMEOUT:
3780		return tcp_sock_set_user_timeout(sk, val);
3781	case TCP_KEEPINTVL:
3782		return tcp_sock_set_keepintvl(sk, val);
3783	case TCP_KEEPCNT:
3784		return tcp_sock_set_keepcnt(sk, val);
3785	case TCP_LINGER2:
3786		if (val < 0)
3787			WRITE_ONCE(tp->linger2, -1);
3788		else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3789			WRITE_ONCE(tp->linger2, TCP_FIN_TIMEOUT_MAX);
3790		else
3791			WRITE_ONCE(tp->linger2, val * HZ);
3792		return 0;
3793	case TCP_DEFER_ACCEPT:
3794		/* Translate value in seconds to number of retransmits */
3795		WRITE_ONCE(icsk->icsk_accept_queue.rskq_defer_accept,
3796			   secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3797					   TCP_RTO_MAX / HZ));
3798		return 0;
3799	}
3800
3801	sockopt_lock_sock(sk);
3802
3803	switch (optname) {
3804	case TCP_MAXSEG:
3805		/* Values greater than interface MTU won't take effect. However
3806		 * at the point when this call is done we typically don't yet
3807		 * know which interface is going to be used
3808		 */
3809		if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3810			err = -EINVAL;
3811			break;
3812		}
3813		tp->rx_opt.user_mss = val;
3814		break;
3815
3816	case TCP_NODELAY:
3817		__tcp_sock_set_nodelay(sk, val);
 
 
 
 
 
 
 
 
 
 
 
 
 
3818		break;
3819
3820	case TCP_THIN_LINEAR_TIMEOUTS:
3821		if (val < 0 || val > 1)
3822			err = -EINVAL;
3823		else
3824			tp->thin_lto = val;
3825		break;
3826
3827	case TCP_THIN_DUPACK:
3828		if (val < 0 || val > 1)
3829			err = -EINVAL;
3830		break;
3831
3832	case TCP_REPAIR:
3833		if (!tcp_can_repair_sock(sk))
3834			err = -EPERM;
3835		else if (val == TCP_REPAIR_ON) {
3836			tp->repair = 1;
3837			sk->sk_reuse = SK_FORCE_REUSE;
3838			tp->repair_queue = TCP_NO_QUEUE;
3839		} else if (val == TCP_REPAIR_OFF) {
3840			tp->repair = 0;
3841			sk->sk_reuse = SK_NO_REUSE;
3842			tcp_send_window_probe(sk);
3843		} else if (val == TCP_REPAIR_OFF_NO_WP) {
3844			tp->repair = 0;
3845			sk->sk_reuse = SK_NO_REUSE;
3846		} else
3847			err = -EINVAL;
3848
3849		break;
3850
3851	case TCP_REPAIR_QUEUE:
3852		if (!tp->repair)
3853			err = -EPERM;
3854		else if ((unsigned int)val < TCP_QUEUES_NR)
3855			tp->repair_queue = val;
3856		else
3857			err = -EINVAL;
3858		break;
3859
3860	case TCP_QUEUE_SEQ:
3861		if (sk->sk_state != TCP_CLOSE) {
3862			err = -EPERM;
3863		} else if (tp->repair_queue == TCP_SEND_QUEUE) {
3864			if (!tcp_rtx_queue_empty(sk))
3865				err = -EPERM;
3866			else
3867				WRITE_ONCE(tp->write_seq, val);
3868		} else if (tp->repair_queue == TCP_RECV_QUEUE) {
3869			if (tp->rcv_nxt != tp->copied_seq) {
3870				err = -EPERM;
3871			} else {
3872				WRITE_ONCE(tp->rcv_nxt, val);
3873				WRITE_ONCE(tp->copied_seq, val);
3874			}
3875		} else {
3876			err = -EINVAL;
3877		}
3878		break;
3879
3880	case TCP_REPAIR_OPTIONS:
3881		if (!tp->repair)
3882			err = -EINVAL;
3883		else if (sk->sk_state == TCP_ESTABLISHED && !tp->bytes_sent)
3884			err = tcp_repair_options_est(sk, optval, optlen);
 
 
3885		else
3886			err = -EPERM;
3887		break;
3888
3889	case TCP_CORK:
3890		__tcp_sock_set_cork(sk, val);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3891		break;
3892
3893	case TCP_KEEPIDLE:
3894		err = tcp_sock_set_keepidle_locked(sk, val);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3895		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3896	case TCP_SAVE_SYN:
3897		/* 0: disable, 1: enable, 2: start from ether_header */
3898		if (val < 0 || val > 2)
3899			err = -EINVAL;
3900		else
3901			tp->save_syn = val;
3902		break;
3903
3904	case TCP_WINDOW_CLAMP:
3905		err = tcp_set_window_clamp(sk, val);
 
 
 
 
 
3906		break;
3907
3908	case TCP_QUICKACK:
3909		__tcp_sock_set_quickack(sk, val);
 
 
 
3910		break;
3911
3912	case TCP_AO_REPAIR:
3913		if (!tcp_can_repair_sock(sk)) {
3914			err = -EPERM;
3915			break;
3916		}
3917		err = tcp_ao_set_repair(sk, optval, optlen);
3918		break;
3919#ifdef CONFIG_TCP_AO
3920	case TCP_AO_ADD_KEY:
3921	case TCP_AO_DEL_KEY:
3922	case TCP_AO_INFO: {
3923		/* If this is the first TCP-AO setsockopt() on the socket,
3924		 * sk_state has to be LISTEN or CLOSE. Allow TCP_REPAIR
3925		 * in any state.
3926		 */
3927		if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
3928			goto ao_parse;
3929		if (rcu_dereference_protected(tcp_sk(sk)->ao_info,
3930					      lockdep_sock_is_held(sk)))
3931			goto ao_parse;
3932		if (tp->repair)
3933			goto ao_parse;
3934		err = -EISCONN;
3935		break;
3936ao_parse:
3937		err = tp->af_specific->ao_parse(sk, optname, optval, optlen);
 
 
 
 
 
 
 
 
 
 
 
 
 
3938		break;
3939	}
3940#endif
3941#ifdef CONFIG_TCP_MD5SIG
3942	case TCP_MD5SIG:
3943	case TCP_MD5SIG_EXT:
3944		err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
 
 
 
3945		break;
3946#endif
 
 
 
 
 
 
 
 
 
 
3947	case TCP_FASTOPEN:
3948		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3949		    TCPF_LISTEN))) {
3950			tcp_fastopen_init_key_once(net);
3951
3952			fastopen_queue_tune(sk, val);
3953		} else {
3954			err = -EINVAL;
3955		}
3956		break;
3957	case TCP_FASTOPEN_CONNECT:
3958		if (val > 1 || val < 0) {
3959			err = -EINVAL;
3960		} else if (READ_ONCE(net->ipv4.sysctl_tcp_fastopen) &
3961			   TFO_CLIENT_ENABLE) {
3962			if (sk->sk_state == TCP_CLOSE)
3963				tp->fastopen_connect = val;
3964			else
3965				err = -EINVAL;
3966		} else {
3967			err = -EOPNOTSUPP;
3968		}
3969		break;
3970	case TCP_FASTOPEN_NO_COOKIE:
3971		if (val > 1 || val < 0)
3972			err = -EINVAL;
3973		else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3974			err = -EINVAL;
3975		else
3976			tp->fastopen_no_cookie = val;
3977		break;
3978	case TCP_TIMESTAMP:
3979		if (!tp->repair) {
3980			err = -EPERM;
3981			break;
3982		}
3983		/* val is an opaque field,
3984		 * and low order bit contains usec_ts enable bit.
3985		 * Its a best effort, and we do not care if user makes an error.
3986		 */
3987		tp->tcp_usec_ts = val & 1;
3988		WRITE_ONCE(tp->tsoffset, val - tcp_clock_ts(tp->tcp_usec_ts));
3989		break;
3990	case TCP_REPAIR_WINDOW:
3991		err = tcp_repair_set_window(tp, optval, optlen);
3992		break;
3993	case TCP_NOTSENT_LOWAT:
3994		WRITE_ONCE(tp->notsent_lowat, val);
3995		sk->sk_write_space(sk);
3996		break;
3997	case TCP_INQ:
3998		if (val > 1 || val < 0)
3999			err = -EINVAL;
4000		else
4001			tp->recvmsg_inq = val;
4002		break;
4003	case TCP_TX_DELAY:
4004		if (val)
4005			tcp_enable_tx_delay();
4006		WRITE_ONCE(tp->tcp_tx_delay, val);
4007		break;
4008	default:
4009		err = -ENOPROTOOPT;
4010		break;
4011	}
4012
4013	sockopt_release_sock(sk);
4014	return err;
4015}
4016
4017int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
4018		   unsigned int optlen)
4019{
4020	const struct inet_connection_sock *icsk = inet_csk(sk);
4021
4022	if (level != SOL_TCP)
4023		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4024		return READ_ONCE(icsk->icsk_af_ops)->setsockopt(sk, level, optname,
4025								optval, optlen);
4026	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
4027}
4028EXPORT_SYMBOL(tcp_setsockopt);
4029
 
 
 
 
 
 
 
 
 
 
 
 
4030static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
4031				      struct tcp_info *info)
4032{
4033	u64 stats[__TCP_CHRONO_MAX], total = 0;
4034	enum tcp_chrono i;
4035
4036	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
4037		stats[i] = tp->chrono_stat[i - 1];
4038		if (i == tp->chrono_type)
4039			stats[i] += tcp_jiffies32 - tp->chrono_start;
4040		stats[i] *= USEC_PER_SEC / HZ;
4041		total += stats[i];
4042	}
4043
4044	info->tcpi_busy_time = total;
4045	info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
4046	info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
4047}
4048
4049/* Return information about state of tcp endpoint in API format. */
4050void tcp_get_info(struct sock *sk, struct tcp_info *info)
4051{
4052	const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
4053	const struct inet_connection_sock *icsk = inet_csk(sk);
4054	unsigned long rate;
4055	u32 now;
4056	u64 rate64;
4057	bool slow;
4058
4059	memset(info, 0, sizeof(*info));
4060	if (sk->sk_type != SOCK_STREAM)
4061		return;
4062
4063	info->tcpi_state = inet_sk_state_load(sk);
4064
4065	/* Report meaningful fields for all TCP states, including listeners */
4066	rate = READ_ONCE(sk->sk_pacing_rate);
4067	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4068	info->tcpi_pacing_rate = rate64;
4069
4070	rate = READ_ONCE(sk->sk_max_pacing_rate);
4071	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4072	info->tcpi_max_pacing_rate = rate64;
4073
4074	info->tcpi_reordering = tp->reordering;
4075	info->tcpi_snd_cwnd = tcp_snd_cwnd(tp);
4076
4077	if (info->tcpi_state == TCP_LISTEN) {
4078		/* listeners aliased fields :
4079		 * tcpi_unacked -> Number of children ready for accept()
4080		 * tcpi_sacked  -> max backlog
4081		 */
4082		info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
4083		info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
4084		return;
4085	}
4086
4087	slow = lock_sock_fast(sk);
4088
4089	info->tcpi_ca_state = icsk->icsk_ca_state;
4090	info->tcpi_retransmits = icsk->icsk_retransmits;
4091	info->tcpi_probes = icsk->icsk_probes_out;
4092	info->tcpi_backoff = icsk->icsk_backoff;
4093
4094	if (tp->rx_opt.tstamp_ok)
4095		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
4096	if (tcp_is_sack(tp))
4097		info->tcpi_options |= TCPI_OPT_SACK;
4098	if (tp->rx_opt.wscale_ok) {
4099		info->tcpi_options |= TCPI_OPT_WSCALE;
4100		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
4101		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
4102	}
4103
4104	if (tp->ecn_flags & TCP_ECN_OK)
4105		info->tcpi_options |= TCPI_OPT_ECN;
4106	if (tp->ecn_flags & TCP_ECN_SEEN)
4107		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
4108	if (tp->syn_data_acked)
4109		info->tcpi_options |= TCPI_OPT_SYN_DATA;
4110	if (tp->tcp_usec_ts)
4111		info->tcpi_options |= TCPI_OPT_USEC_TS;
4112
4113	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
4114	info->tcpi_ato = jiffies_to_usecs(min_t(u32, icsk->icsk_ack.ato,
4115						tcp_delack_max(sk)));
4116	info->tcpi_snd_mss = tp->mss_cache;
4117	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
4118
4119	info->tcpi_unacked = tp->packets_out;
4120	info->tcpi_sacked = tp->sacked_out;
4121
4122	info->tcpi_lost = tp->lost_out;
4123	info->tcpi_retrans = tp->retrans_out;
4124
4125	now = tcp_jiffies32;
4126	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
4127	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
4128	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
4129
4130	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
4131	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
4132	info->tcpi_rtt = tp->srtt_us >> 3;
4133	info->tcpi_rttvar = tp->mdev_us >> 2;
4134	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
4135	info->tcpi_advmss = tp->advmss;
4136
4137	info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
4138	info->tcpi_rcv_space = tp->rcvq_space.space;
4139
4140	info->tcpi_total_retrans = tp->total_retrans;
4141
4142	info->tcpi_bytes_acked = tp->bytes_acked;
4143	info->tcpi_bytes_received = tp->bytes_received;
4144	info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
4145	tcp_get_info_chrono_stats(tp, info);
4146
4147	info->tcpi_segs_out = tp->segs_out;
4148
4149	/* segs_in and data_segs_in can be updated from tcp_segs_in() from BH */
4150	info->tcpi_segs_in = READ_ONCE(tp->segs_in);
4151	info->tcpi_data_segs_in = READ_ONCE(tp->data_segs_in);
4152
4153	info->tcpi_min_rtt = tcp_min_rtt(tp);
 
4154	info->tcpi_data_segs_out = tp->data_segs_out;
4155
4156	info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
4157	rate64 = tcp_compute_delivery_rate(tp);
4158	if (rate64)
4159		info->tcpi_delivery_rate = rate64;
4160	info->tcpi_delivered = tp->delivered;
4161	info->tcpi_delivered_ce = tp->delivered_ce;
4162	info->tcpi_bytes_sent = tp->bytes_sent;
4163	info->tcpi_bytes_retrans = tp->bytes_retrans;
4164	info->tcpi_dsack_dups = tp->dsack_dups;
4165	info->tcpi_reord_seen = tp->reord_seen;
4166	info->tcpi_rcv_ooopack = tp->rcv_ooopack;
4167	info->tcpi_snd_wnd = tp->snd_wnd;
4168	info->tcpi_rcv_wnd = tp->rcv_wnd;
4169	info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash;
4170	info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
4171
4172	info->tcpi_total_rto = tp->total_rto;
4173	info->tcpi_total_rto_recoveries = tp->total_rto_recoveries;
4174	info->tcpi_total_rto_time = tp->total_rto_time;
4175	if (tp->rto_stamp)
4176		info->tcpi_total_rto_time += tcp_clock_ms() - tp->rto_stamp;
4177
4178	unlock_sock_fast(sk, slow);
4179}
4180EXPORT_SYMBOL_GPL(tcp_get_info);
4181
4182static size_t tcp_opt_stats_get_size(void)
4183{
4184	return
4185		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
4186		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
4187		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
4188		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
4189		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
4190		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
4191		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
4192		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
4193		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
4194		nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
4195		nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
4196		nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
4197		nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
4198		nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
4199		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
4200		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
4201		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
4202		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
4203		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
4204		nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
4205		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
4206		nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
4207		nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
4208		nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
4209		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
4210		nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
4211		nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */
4212		0;
4213}
4214
4215/* Returns TTL or hop limit of an incoming packet from skb. */
4216static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
4217{
4218	if (skb->protocol == htons(ETH_P_IP))
4219		return ip_hdr(skb)->ttl;
4220	else if (skb->protocol == htons(ETH_P_IPV6))
4221		return ipv6_hdr(skb)->hop_limit;
4222	else
4223		return 0;
4224}
4225
4226struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
4227					       const struct sk_buff *orig_skb,
4228					       const struct sk_buff *ack_skb)
4229{
4230	const struct tcp_sock *tp = tcp_sk(sk);
4231	struct sk_buff *stats;
4232	struct tcp_info info;
4233	unsigned long rate;
4234	u64 rate64;
4235
4236	stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
4237	if (!stats)
4238		return NULL;
4239
4240	tcp_get_info_chrono_stats(tp, &info);
4241	nla_put_u64_64bit(stats, TCP_NLA_BUSY,
4242			  info.tcpi_busy_time, TCP_NLA_PAD);
4243	nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
4244			  info.tcpi_rwnd_limited, TCP_NLA_PAD);
4245	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
4246			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
4247	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
4248			  tp->data_segs_out, TCP_NLA_PAD);
4249	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
4250			  tp->total_retrans, TCP_NLA_PAD);
4251
4252	rate = READ_ONCE(sk->sk_pacing_rate);
4253	rate64 = (rate != ~0UL) ? rate : ~0ULL;
4254	nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
4255
4256	rate64 = tcp_compute_delivery_rate(tp);
4257	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
4258
4259	nla_put_u32(stats, TCP_NLA_SND_CWND, tcp_snd_cwnd(tp));
4260	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
4261	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
4262
4263	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
4264	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
4265	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
4266	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
4267	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
4268
4269	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
4270	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
4271
4272	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
4273			  TCP_NLA_PAD);
4274	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
4275			  TCP_NLA_PAD);
4276	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
4277	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
4278	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
4279	nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
4280	nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
4281		    max_t(int, 0, tp->write_seq - tp->snd_nxt));
4282	nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
4283			  TCP_NLA_PAD);
4284	if (ack_skb)
4285		nla_put_u8(stats, TCP_NLA_TTL,
4286			   tcp_skb_ttl_or_hop_limit(ack_skb));
4287
4288	nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash);
4289	return stats;
4290}
4291
4292int do_tcp_getsockopt(struct sock *sk, int level,
4293		      int optname, sockptr_t optval, sockptr_t optlen)
4294{
4295	struct inet_connection_sock *icsk = inet_csk(sk);
4296	struct tcp_sock *tp = tcp_sk(sk);
4297	struct net *net = sock_net(sk);
4298	int val, len;
4299
4300	if (copy_from_sockptr(&len, optlen, sizeof(int)))
4301		return -EFAULT;
4302
 
 
4303	if (len < 0)
4304		return -EINVAL;
4305
4306	len = min_t(unsigned int, len, sizeof(int));
4307
4308	switch (optname) {
4309	case TCP_MAXSEG:
4310		val = tp->mss_cache;
4311		if (tp->rx_opt.user_mss &&
4312		    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
4313			val = tp->rx_opt.user_mss;
4314		if (tp->repair)
4315			val = tp->rx_opt.mss_clamp;
4316		break;
4317	case TCP_NODELAY:
4318		val = !!(tp->nonagle&TCP_NAGLE_OFF);
4319		break;
4320	case TCP_CORK:
4321		val = !!(tp->nonagle&TCP_NAGLE_CORK);
4322		break;
4323	case TCP_KEEPIDLE:
4324		val = keepalive_time_when(tp) / HZ;
4325		break;
4326	case TCP_KEEPINTVL:
4327		val = keepalive_intvl_when(tp) / HZ;
4328		break;
4329	case TCP_KEEPCNT:
4330		val = keepalive_probes(tp);
4331		break;
4332	case TCP_SYNCNT:
4333		val = READ_ONCE(icsk->icsk_syn_retries) ? :
4334			READ_ONCE(net->ipv4.sysctl_tcp_syn_retries);
4335		break;
4336	case TCP_LINGER2:
4337		val = READ_ONCE(tp->linger2);
4338		if (val >= 0)
4339			val = (val ? : READ_ONCE(net->ipv4.sysctl_tcp_fin_timeout)) / HZ;
4340		break;
4341	case TCP_DEFER_ACCEPT:
4342		val = READ_ONCE(icsk->icsk_accept_queue.rskq_defer_accept);
4343		val = retrans_to_secs(val, TCP_TIMEOUT_INIT / HZ,
4344				      TCP_RTO_MAX / HZ);
4345		break;
4346	case TCP_WINDOW_CLAMP:
4347		val = READ_ONCE(tp->window_clamp);
4348		break;
4349	case TCP_INFO: {
4350		struct tcp_info info;
4351
4352		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4353			return -EFAULT;
4354
4355		tcp_get_info(sk, &info);
4356
4357		len = min_t(unsigned int, len, sizeof(info));
4358		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4359			return -EFAULT;
4360		if (copy_to_sockptr(optval, &info, len))
4361			return -EFAULT;
4362		return 0;
4363	}
4364	case TCP_CC_INFO: {
4365		const struct tcp_congestion_ops *ca_ops;
4366		union tcp_cc_info info;
4367		size_t sz = 0;
4368		int attr;
4369
4370		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4371			return -EFAULT;
4372
4373		ca_ops = icsk->icsk_ca_ops;
4374		if (ca_ops && ca_ops->get_info)
4375			sz = ca_ops->get_info(sk, ~0U, &attr, &info);
4376
4377		len = min_t(unsigned int, len, sz);
4378		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4379			return -EFAULT;
4380		if (copy_to_sockptr(optval, &info, len))
4381			return -EFAULT;
4382		return 0;
4383	}
4384	case TCP_QUICKACK:
4385		val = !inet_csk_in_pingpong_mode(sk);
4386		break;
4387
4388	case TCP_CONGESTION:
4389		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4390			return -EFAULT;
4391		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4392		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4393			return -EFAULT;
4394		if (copy_to_sockptr(optval, icsk->icsk_ca_ops->name, len))
4395			return -EFAULT;
4396		return 0;
4397
4398	case TCP_ULP:
4399		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4400			return -EFAULT;
4401		len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4402		if (!icsk->icsk_ulp_ops) {
4403			len = 0;
4404			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4405				return -EFAULT;
4406			return 0;
4407		}
4408		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4409			return -EFAULT;
4410		if (copy_to_sockptr(optval, icsk->icsk_ulp_ops->name, len))
4411			return -EFAULT;
4412		return 0;
4413
4414	case TCP_FASTOPEN_KEY: {
4415		u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4416		unsigned int key_len;
 
4417
4418		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4419			return -EFAULT;
4420
4421		key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4422				TCP_FASTOPEN_KEY_LENGTH;
 
 
 
 
 
 
 
4423		len = min_t(unsigned int, len, key_len);
4424		if (copy_to_sockptr(optlen, &len, sizeof(int)))
4425			return -EFAULT;
4426		if (copy_to_sockptr(optval, key, len))
4427			return -EFAULT;
4428		return 0;
4429	}
4430	case TCP_THIN_LINEAR_TIMEOUTS:
4431		val = tp->thin_lto;
4432		break;
4433
4434	case TCP_THIN_DUPACK:
4435		val = 0;
4436		break;
4437
4438	case TCP_REPAIR:
4439		val = tp->repair;
4440		break;
4441
4442	case TCP_REPAIR_QUEUE:
4443		if (tp->repair)
4444			val = tp->repair_queue;
4445		else
4446			return -EINVAL;
4447		break;
4448
4449	case TCP_REPAIR_WINDOW: {
4450		struct tcp_repair_window opt;
4451
4452		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4453			return -EFAULT;
4454
4455		if (len != sizeof(opt))
4456			return -EINVAL;
4457
4458		if (!tp->repair)
4459			return -EPERM;
4460
4461		opt.snd_wl1	= tp->snd_wl1;
4462		opt.snd_wnd	= tp->snd_wnd;
4463		opt.max_window	= tp->max_window;
4464		opt.rcv_wnd	= tp->rcv_wnd;
4465		opt.rcv_wup	= tp->rcv_wup;
4466
4467		if (copy_to_sockptr(optval, &opt, len))
4468			return -EFAULT;
4469		return 0;
4470	}
4471	case TCP_QUEUE_SEQ:
4472		if (tp->repair_queue == TCP_SEND_QUEUE)
4473			val = tp->write_seq;
4474		else if (tp->repair_queue == TCP_RECV_QUEUE)
4475			val = tp->rcv_nxt;
4476		else
4477			return -EINVAL;
4478		break;
4479
4480	case TCP_USER_TIMEOUT:
4481		val = READ_ONCE(icsk->icsk_user_timeout);
4482		break;
4483
4484	case TCP_FASTOPEN:
4485		val = READ_ONCE(icsk->icsk_accept_queue.fastopenq.max_qlen);
4486		break;
4487
4488	case TCP_FASTOPEN_CONNECT:
4489		val = tp->fastopen_connect;
4490		break;
4491
4492	case TCP_FASTOPEN_NO_COOKIE:
4493		val = tp->fastopen_no_cookie;
4494		break;
4495
4496	case TCP_TX_DELAY:
4497		val = READ_ONCE(tp->tcp_tx_delay);
4498		break;
4499
4500	case TCP_TIMESTAMP:
4501		val = tcp_clock_ts(tp->tcp_usec_ts) + READ_ONCE(tp->tsoffset);
4502		if (tp->tcp_usec_ts)
4503			val |= 1;
4504		else
4505			val &= ~1;
4506		break;
4507	case TCP_NOTSENT_LOWAT:
4508		val = READ_ONCE(tp->notsent_lowat);
4509		break;
4510	case TCP_INQ:
4511		val = tp->recvmsg_inq;
4512		break;
4513	case TCP_SAVE_SYN:
4514		val = tp->save_syn;
4515		break;
4516	case TCP_SAVED_SYN: {
4517		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4518			return -EFAULT;
4519
4520		sockopt_lock_sock(sk);
4521		if (tp->saved_syn) {
4522			if (len < tcp_saved_syn_len(tp->saved_syn)) {
4523				len = tcp_saved_syn_len(tp->saved_syn);
4524				if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4525					sockopt_release_sock(sk);
4526					return -EFAULT;
4527				}
4528				sockopt_release_sock(sk);
4529				return -EINVAL;
4530			}
4531			len = tcp_saved_syn_len(tp->saved_syn);
4532			if (copy_to_sockptr(optlen, &len, sizeof(int))) {
4533				sockopt_release_sock(sk);
4534				return -EFAULT;
4535			}
4536			if (copy_to_sockptr(optval, tp->saved_syn->data, len)) {
4537				sockopt_release_sock(sk);
4538				return -EFAULT;
4539			}
4540			tcp_saved_syn_free(tp);
4541			sockopt_release_sock(sk);
4542		} else {
4543			sockopt_release_sock(sk);
4544			len = 0;
4545			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4546				return -EFAULT;
4547		}
4548		return 0;
4549	}
4550#ifdef CONFIG_MMU
4551	case TCP_ZEROCOPY_RECEIVE: {
4552		struct scm_timestamping_internal tss;
4553		struct tcp_zerocopy_receive zc = {};
4554		int err;
4555
4556		if (copy_from_sockptr(&len, optlen, sizeof(int)))
4557			return -EFAULT;
4558		if (len < 0 ||
4559		    len < offsetofend(struct tcp_zerocopy_receive, length))
4560			return -EINVAL;
4561		if (unlikely(len > sizeof(zc))) {
4562			err = check_zeroed_sockptr(optval, sizeof(zc),
4563						   len - sizeof(zc));
4564			if (err < 1)
4565				return err == 0 ? -EINVAL : err;
4566			len = sizeof(zc);
4567			if (copy_to_sockptr(optlen, &len, sizeof(int)))
4568				return -EFAULT;
4569		}
4570		if (copy_from_sockptr(&zc, optval, len))
4571			return -EFAULT;
4572		if (zc.reserved)
4573			return -EINVAL;
4574		if (zc.msg_flags &  ~(TCP_VALID_ZC_MSG_FLAGS))
4575			return -EINVAL;
4576		sockopt_lock_sock(sk);
4577		err = tcp_zerocopy_receive(sk, &zc, &tss);
4578		err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4579							  &zc, &len, err);
4580		sockopt_release_sock(sk);
4581		if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4582			goto zerocopy_rcv_cmsg;
4583		switch (len) {
4584		case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4585			goto zerocopy_rcv_cmsg;
4586		case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4587		case offsetofend(struct tcp_zerocopy_receive, msg_control):
4588		case offsetofend(struct tcp_zerocopy_receive, flags):
4589		case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4590		case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4591		case offsetofend(struct tcp_zerocopy_receive, err):
4592			goto zerocopy_rcv_sk_err;
4593		case offsetofend(struct tcp_zerocopy_receive, inq):
4594			goto zerocopy_rcv_inq;
4595		case offsetofend(struct tcp_zerocopy_receive, length):
4596		default:
4597			goto zerocopy_rcv_out;
4598		}
4599zerocopy_rcv_cmsg:
4600		if (zc.msg_flags & TCP_CMSG_TS)
4601			tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4602		else
4603			zc.msg_flags = 0;
4604zerocopy_rcv_sk_err:
4605		if (!err)
4606			zc.err = sock_error(sk);
4607zerocopy_rcv_inq:
4608		zc.inq = tcp_inq_hint(sk);
4609zerocopy_rcv_out:
4610		if (!err && copy_to_sockptr(optval, &zc, len))
4611			err = -EFAULT;
4612		return err;
4613	}
4614#endif
4615	case TCP_AO_REPAIR:
4616		if (!tcp_can_repair_sock(sk))
4617			return -EPERM;
4618		return tcp_ao_get_repair(sk, optval, optlen);
4619	case TCP_AO_GET_KEYS:
4620	case TCP_AO_INFO: {
4621		int err;
4622
4623		sockopt_lock_sock(sk);
4624		if (optname == TCP_AO_GET_KEYS)
4625			err = tcp_ao_get_mkts(sk, optval, optlen);
4626		else
4627			err = tcp_ao_get_sock_info(sk, optval, optlen);
4628		sockopt_release_sock(sk);
4629
4630		return err;
4631	}
4632	case TCP_IS_MPTCP:
4633		val = 0;
4634		break;
4635	default:
4636		return -ENOPROTOOPT;
4637	}
4638
4639	if (copy_to_sockptr(optlen, &len, sizeof(int)))
4640		return -EFAULT;
4641	if (copy_to_sockptr(optval, &val, len))
4642		return -EFAULT;
4643	return 0;
4644}
4645
4646bool tcp_bpf_bypass_getsockopt(int level, int optname)
4647{
4648	/* TCP do_tcp_getsockopt has optimized getsockopt implementation
4649	 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4650	 */
4651	if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4652		return true;
4653
4654	return false;
4655}
4656EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4657
4658int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4659		   int __user *optlen)
4660{
4661	struct inet_connection_sock *icsk = inet_csk(sk);
4662
4663	if (level != SOL_TCP)
4664		/* Paired with WRITE_ONCE() in do_ipv6_setsockopt() and tcp_v6_connect() */
4665		return READ_ONCE(icsk->icsk_af_ops)->getsockopt(sk, level, optname,
4666								optval, optlen);
4667	return do_tcp_getsockopt(sk, level, optname, USER_SOCKPTR(optval),
4668				 USER_SOCKPTR(optlen));
4669}
4670EXPORT_SYMBOL(tcp_getsockopt);
4671
4672#ifdef CONFIG_TCP_MD5SIG
4673int tcp_md5_sigpool_id = -1;
4674EXPORT_SYMBOL_GPL(tcp_md5_sigpool_id);
4675
4676int tcp_md5_alloc_sigpool(void)
4677{
4678	size_t scratch_size;
4679	int ret;
4680
4681	scratch_size = sizeof(union tcp_md5sum_block) + sizeof(struct tcphdr);
4682	ret = tcp_sigpool_alloc_ahash("md5", scratch_size);
4683	if (ret >= 0) {
4684		/* As long as any md5 sigpool was allocated, the return
4685		 * id would stay the same. Re-write the id only for the case
4686		 * when previously all MD5 keys were deleted and this call
4687		 * allocates the first MD5 key, which may return a different
4688		 * sigpool id than was used previously.
4689		 */
4690		WRITE_ONCE(tcp_md5_sigpool_id, ret); /* Avoids the compiler potentially being smart here */
4691		return 0;
4692	}
4693	return ret;
4694}
 
 
4695
4696void tcp_md5_release_sigpool(void)
4697{
4698	tcp_sigpool_release(READ_ONCE(tcp_md5_sigpool_id));
4699}
4700
4701void tcp_md5_add_sigpool(void)
4702{
4703	tcp_sigpool_get(READ_ONCE(tcp_md5_sigpool_id));
4704}
4705
4706int tcp_md5_hash_key(struct tcp_sigpool *hp,
4707		     const struct tcp_md5sig_key *key)
4708{
4709	u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4710	struct scatterlist sg;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4711
4712	sg_init_one(&sg, key->key, keylen);
4713	ahash_request_set_crypt(hp->req, &sg, NULL, keylen);
4714
4715	/* We use data_race() because tcp_md5_do_add() might change
4716	 * key->key under us
 
 
4717	 */
4718	return data_race(crypto_ahash_update(hp->req));
 
4719}
4720EXPORT_SYMBOL(tcp_md5_hash_key);
4721
4722/* Called with rcu_read_lock() */
4723static enum skb_drop_reason
4724tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4725		     const void *saddr, const void *daddr,
4726		     int family, int l3index, const __u8 *hash_location)
4727{
4728	/* This gets called for each TCP segment that has TCP-MD5 option.
4729	 * We have 3 drop cases:
4730	 * o No MD5 hash and one expected.
4731	 * o MD5 hash and we're not expecting one.
4732	 * o MD5 hash and its wrong.
4733	 */
4734	const struct tcp_sock *tp = tcp_sk(sk);
4735	struct tcp_md5sig_key *key;
4736	u8 newhash[16];
4737	int genhash;
4738
4739	key = tcp_md5_do_lookup(sk, l3index, saddr, family);
 
 
 
 
4740
4741	if (!key && hash_location) {
4742		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5UNEXPECTED);
4743		trace_tcp_hash_md5_unexpected(sk, skb);
4744		return SKB_DROP_REASON_TCP_MD5UNEXPECTED;
4745	}
4746
4747	/* Check the signature.
4748	 * To support dual stack listeners, we need to handle
4749	 * IPv4-mapped case.
4750	 */
4751	if (family == AF_INET)
4752		genhash = tcp_v4_md5_hash_skb(newhash, key, NULL, skb);
4753	else
4754		genhash = tp->af_specific->calc_md5_hash(newhash, key,
4755							 NULL, skb);
4756	if (genhash || memcmp(hash_location, newhash, 16) != 0) {
4757		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5FAILURE);
4758		trace_tcp_hash_md5_mismatch(sk, skb);
4759		return SKB_DROP_REASON_TCP_MD5FAILURE;
4760	}
4761	return SKB_NOT_DROPPED_YET;
4762}
4763#else
4764static inline enum skb_drop_reason
4765tcp_inbound_md5_hash(const struct sock *sk, const struct sk_buff *skb,
4766		     const void *saddr, const void *daddr,
4767		     int family, int l3index, const __u8 *hash_location)
4768{
4769	return SKB_NOT_DROPPED_YET;
4770}
 
4771
4772#endif
4773
4774/* Called with rcu_read_lock() */
4775enum skb_drop_reason
4776tcp_inbound_hash(struct sock *sk, const struct request_sock *req,
4777		 const struct sk_buff *skb,
4778		 const void *saddr, const void *daddr,
4779		 int family, int dif, int sdif)
4780{
4781	const struct tcphdr *th = tcp_hdr(skb);
4782	const struct tcp_ao_hdr *aoh;
4783	const __u8 *md5_location;
4784	int l3index;
4785
4786	/* Invalid option or two times meet any of auth options */
4787	if (tcp_parse_auth_options(th, &md5_location, &aoh)) {
4788		trace_tcp_hash_bad_header(sk, skb);
4789		return SKB_DROP_REASON_TCP_AUTH_HDR;
4790	}
4791
4792	if (req) {
4793		if (tcp_rsk_used_ao(req) != !!aoh) {
4794			u8 keyid, rnext, maclen;
4795
4796			if (aoh) {
4797				keyid = aoh->keyid;
4798				rnext = aoh->rnext_keyid;
4799				maclen = tcp_ao_hdr_maclen(aoh);
4800			} else {
4801				keyid = rnext = maclen = 0;
4802			}
4803
4804			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAOBAD);
4805			trace_tcp_ao_handshake_failure(sk, skb, keyid, rnext, maclen);
4806			return SKB_DROP_REASON_TCP_AOFAILURE;
4807		}
4808	}
 
 
 
 
4809
4810	/* sdif set, means packet ingressed via a device
4811	 * in an L3 domain and dif is set to the l3mdev
4812	 */
4813	l3index = sdif ? dif : 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4814
4815	/* Fast path: unsigned segments */
4816	if (likely(!md5_location && !aoh)) {
4817		/* Drop if there's TCP-MD5 or TCP-AO key with any rcvid/sndid
4818		 * for the remote peer. On TCP-AO established connection
4819		 * the last key is impossible to remove, so there's
4820		 * always at least one current_key.
4821		 */
4822		if (tcp_ao_required(sk, saddr, family, l3index, true)) {
4823			trace_tcp_hash_ao_required(sk, skb);
4824			return SKB_DROP_REASON_TCP_AONOTFOUND;
4825		}
4826		if (unlikely(tcp_md5_do_lookup(sk, l3index, saddr, family))) {
4827			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMD5NOTFOUND);
4828			trace_tcp_hash_md5_required(sk, skb);
4829			return SKB_DROP_REASON_TCP_MD5NOTFOUND;
4830		}
4831		return SKB_NOT_DROPPED_YET;
4832	}
4833
4834	if (aoh)
4835		return tcp_inbound_ao_hash(sk, skb, family, req, l3index, aoh);
 
4836
4837	return tcp_inbound_md5_hash(sk, skb, saddr, daddr, family,
4838				    l3index, md5_location);
 
4839}
4840EXPORT_SYMBOL_GPL(tcp_inbound_hash);
 
 
4841
4842void tcp_done(struct sock *sk)
4843{
4844	struct request_sock *req;
4845
4846	/* We might be called with a new socket, after
4847	 * inet_csk_prepare_forced_close() has been called
4848	 * so we can not use lockdep_sock_is_held(sk)
4849	 */
4850	req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4851
4852	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4853		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4854
4855	tcp_set_state(sk, TCP_CLOSE);
4856	tcp_clear_xmit_timers(sk);
4857	if (req)
4858		reqsk_fastopen_remove(sk, req, false);
4859
4860	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
4861
4862	if (!sock_flag(sk, SOCK_DEAD))
4863		sk->sk_state_change(sk);
4864	else
4865		inet_csk_destroy_sock(sk);
4866}
4867EXPORT_SYMBOL_GPL(tcp_done);
4868
4869int tcp_abort(struct sock *sk, int err)
4870{
4871	int state = inet_sk_state_load(sk);
4872
4873	if (state == TCP_NEW_SYN_RECV) {
4874		struct request_sock *req = inet_reqsk(sk);
4875
4876		local_bh_disable();
4877		inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4878		local_bh_enable();
4879		return 0;
4880	}
4881	if (state == TCP_TIME_WAIT) {
4882		struct inet_timewait_sock *tw = inet_twsk(sk);
4883
4884		refcount_inc(&tw->tw_refcnt);
4885		local_bh_disable();
4886		inet_twsk_deschedule_put(tw);
4887		local_bh_enable();
4888		return 0;
4889	}
4890
4891	/* BPF context ensures sock locking. */
4892	if (!has_current_bpf_ctx())
4893		/* Don't race with userspace socket closes such as tcp_close. */
4894		lock_sock(sk);
4895
4896	/* Avoid closing the same socket twice. */
4897	if (sk->sk_state == TCP_CLOSE) {
4898		if (!has_current_bpf_ctx())
4899			release_sock(sk);
4900		return -ENOENT;
4901	}
4902
4903	if (sk->sk_state == TCP_LISTEN) {
4904		tcp_set_state(sk, TCP_CLOSE);
4905		inet_csk_listen_stop(sk);
4906	}
4907
4908	/* Don't race with BH socket closes such as inet_csk_listen_stop. */
4909	local_bh_disable();
4910	bh_lock_sock(sk);
4911
4912	if (tcp_need_reset(sk->sk_state))
4913		tcp_send_active_reset(sk, GFP_ATOMIC,
4914				      SK_RST_REASON_TCP_STATE);
4915	tcp_done_with_error(sk, err);
 
 
 
 
 
4916
4917	bh_unlock_sock(sk);
4918	local_bh_enable();
4919	if (!has_current_bpf_ctx())
4920		release_sock(sk);
4921	return 0;
4922}
4923EXPORT_SYMBOL_GPL(tcp_abort);
4924
4925extern struct tcp_congestion_ops tcp_reno;
4926
4927static __initdata unsigned long thash_entries;
4928static int __init set_thash_entries(char *str)
4929{
4930	ssize_t ret;
4931
4932	if (!str)
4933		return 0;
4934
4935	ret = kstrtoul(str, 0, &thash_entries);
4936	if (ret)
4937		return 0;
4938
4939	return 1;
4940}
4941__setup("thash_entries=", set_thash_entries);
4942
4943static void __init tcp_init_mem(void)
4944{
4945	unsigned long limit = nr_free_buffer_pages() / 16;
4946
4947	limit = max(limit, 128UL);
4948	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
4949	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
4950	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
4951}
4952
4953static void __init tcp_struct_check(void)
4954{
4955	/* TX read-mostly hotpath cache lines */
4956	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, max_window);
4957	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, rcv_ssthresh);
4958	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, reordering);
4959	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, notsent_lowat);
4960	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, gso_segs);
4961	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, lost_skb_hint);
4962	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_tx, retransmit_skb_hint);
4963	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_tx, 40);
4964
4965	/* TXRX read-mostly hotpath cache lines */
4966	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, tsoffset);
4967	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_wnd);
4968	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, mss_cache);
4969	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, snd_cwnd);
4970	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, prr_out);
4971	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, lost_out);
4972	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, sacked_out);
4973	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_txrx, scaling_ratio);
4974	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_txrx, 32);
4975
4976	/* RX read-mostly hotpath cache lines */
4977	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, copied_seq);
4978	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rcv_tstamp);
4979	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_wl1);
4980	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, tlp_high_seq);
4981	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rttvar_us);
4982	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, retrans_out);
4983	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, advmss);
4984	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, urg_data);
4985	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, lost);
4986	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, rtt_min);
4987	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, out_of_order_queue);
4988	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_read_rx, snd_ssthresh);
4989	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_read_rx, 69);
4990
4991	/* TX read-write hotpath cache lines */
4992	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, segs_out);
4993	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, data_segs_out);
4994	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, bytes_sent);
4995	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, snd_sml);
4996	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_start);
4997	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, chrono_stat);
4998	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, write_seq);
4999	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, pushed_seq);
5000	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, lsndtime);
5001	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, mdev_us);
5002	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tcp_wstamp_ns);
5003	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, rtt_seq);
5004	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, tsorted_sent_queue);
5005	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, highest_sack);
5006	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_tx, ecn_flags);
5007	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_tx, 89);
5008
5009	/* TXRX read-write hotpath cache lines */
5010	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, pred_flags);
5011	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_clock_cache);
5012	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, tcp_mstamp);
5013	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_nxt);
5014	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_nxt);
5015	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_una);
5016	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, window_clamp);
5017	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, srtt_us);
5018	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, packets_out);
5019	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, snd_up);
5020	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered);
5021	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, delivered_ce);
5022	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, app_limited);
5023	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rcv_wnd);
5024	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_txrx, rx_opt);
5025
5026	/* 32bit arches with 8byte alignment on u64 fields might need padding
5027	 * before tcp_clock_cache.
5028	 */
5029	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_txrx, 92 + 4);
5030
5031	/* RX read-write hotpath cache lines */
5032	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_received);
5033	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, segs_in);
5034	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, data_segs_in);
5035	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_wup);
5036	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, max_packets_out);
5037	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, cwnd_usage_seq);
5038	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_delivered);
5039	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rate_interval_us);
5040	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_last_tsecr);
5041	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, first_tx_mstamp);
5042	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, delivered_mstamp);
5043	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, bytes_acked);
5044	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcv_rtt_est);
5045	CACHELINE_ASSERT_GROUP_MEMBER(struct tcp_sock, tcp_sock_write_rx, rcvq_space);
5046	CACHELINE_ASSERT_GROUP_SIZE(struct tcp_sock, tcp_sock_write_rx, 99);
5047}
5048
5049void __init tcp_init(void)
5050{
5051	int max_rshare, max_wshare, cnt;
5052	unsigned long limit;
5053	unsigned int i;
5054
5055	BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
5056	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
5057		     sizeof_field(struct sk_buff, cb));
5058
5059	tcp_struct_check();
5060
5061	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
5062
5063	timer_setup(&tcp_orphan_timer, tcp_orphan_update, TIMER_DEFERRABLE);
5064	mod_timer(&tcp_orphan_timer, jiffies + TCP_ORPHAN_TIMER_PERIOD);
5065
5066	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
5067			    thash_entries, 21,  /* one slot per 2 MB*/
5068			    0, 64 * 1024);
5069	tcp_hashinfo.bind_bucket_cachep =
5070		kmem_cache_create("tcp_bind_bucket",
5071				  sizeof(struct inet_bind_bucket), 0,
5072				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5073				  SLAB_ACCOUNT,
5074				  NULL);
5075	tcp_hashinfo.bind2_bucket_cachep =
5076		kmem_cache_create("tcp_bind2_bucket",
5077				  sizeof(struct inet_bind2_bucket), 0,
5078				  SLAB_HWCACHE_ALIGN | SLAB_PANIC |
5079				  SLAB_ACCOUNT,
5080				  NULL);
5081
5082	/* Size and allocate the main established and bind bucket
5083	 * hash tables.
5084	 *
5085	 * The methodology is similar to that of the buffer cache.
5086	 */
5087	tcp_hashinfo.ehash =
5088		alloc_large_system_hash("TCP established",
5089					sizeof(struct inet_ehash_bucket),
5090					thash_entries,
5091					17, /* one slot per 128 KB of memory */
5092					0,
5093					NULL,
5094					&tcp_hashinfo.ehash_mask,
5095					0,
5096					thash_entries ? 0 : 512 * 1024);
5097	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
5098		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
5099
5100	if (inet_ehash_locks_alloc(&tcp_hashinfo))
5101		panic("TCP: failed to alloc ehash_locks");
5102	tcp_hashinfo.bhash =
5103		alloc_large_system_hash("TCP bind",
5104					2 * sizeof(struct inet_bind_hashbucket),
5105					tcp_hashinfo.ehash_mask + 1,
5106					17, /* one slot per 128 KB of memory */
5107					0,
5108					&tcp_hashinfo.bhash_size,
5109					NULL,
5110					0,
5111					64 * 1024);
5112	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
5113	tcp_hashinfo.bhash2 = tcp_hashinfo.bhash + tcp_hashinfo.bhash_size;
5114	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
5115		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
5116		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
5117		spin_lock_init(&tcp_hashinfo.bhash2[i].lock);
5118		INIT_HLIST_HEAD(&tcp_hashinfo.bhash2[i].chain);
5119	}
5120
5121	tcp_hashinfo.pernet = false;
5122
5123	cnt = tcp_hashinfo.ehash_mask + 1;
5124	sysctl_tcp_max_orphans = cnt / 2;
5125
5126	tcp_init_mem();
5127	/* Set per-socket limits to no more than 1/128 the pressure threshold */
5128	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
5129	max_wshare = min(4UL*1024*1024, limit);
5130	max_rshare = min(6UL*1024*1024, limit);
5131
5132	init_net.ipv4.sysctl_tcp_wmem[0] = PAGE_SIZE;
5133	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
5134	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
5135
5136	init_net.ipv4.sysctl_tcp_rmem[0] = PAGE_SIZE;
5137	init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
5138	init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
5139
5140	pr_info("Hash tables configured (established %u bind %u)\n",
5141		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
5142
5143	tcp_v4_init();
5144	tcp_metrics_init();
5145	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
5146	tcp_tasklet_init();
5147	mptcp_init();
5148}
v5.4
   1// SPDX-License-Identifier: GPL-2.0-or-later
   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 * Fixes:
  22 *		Alan Cox	:	Numerous verify_area() calls
  23 *		Alan Cox	:	Set the ACK bit on a reset
  24 *		Alan Cox	:	Stopped it crashing if it closed while
  25 *					sk->inuse=1 and was trying to connect
  26 *					(tcp_err()).
  27 *		Alan Cox	:	All icmp error handling was broken
  28 *					pointers passed where wrong and the
  29 *					socket was looked up backwards. Nobody
  30 *					tested any icmp error code obviously.
  31 *		Alan Cox	:	tcp_err() now handled properly. It
  32 *					wakes people on errors. poll
  33 *					behaves and the icmp error race
  34 *					has gone by moving it into sock.c
  35 *		Alan Cox	:	tcp_send_reset() fixed to work for
  36 *					everything not just packets for
  37 *					unknown sockets.
  38 *		Alan Cox	:	tcp option processing.
  39 *		Alan Cox	:	Reset tweaked (still not 100%) [Had
  40 *					syn rule wrong]
  41 *		Herp Rosmanith  :	More reset fixes
  42 *		Alan Cox	:	No longer acks invalid rst frames.
  43 *					Acking any kind of RST is right out.
  44 *		Alan Cox	:	Sets an ignore me flag on an rst
  45 *					receive otherwise odd bits of prattle
  46 *					escape still
  47 *		Alan Cox	:	Fixed another acking RST frame bug.
  48 *					Should stop LAN workplace lockups.
  49 *		Alan Cox	: 	Some tidyups using the new skb list
  50 *					facilities
  51 *		Alan Cox	:	sk->keepopen now seems to work
  52 *		Alan Cox	:	Pulls options out correctly on accepts
  53 *		Alan Cox	:	Fixed assorted sk->rqueue->next errors
  54 *		Alan Cox	:	PSH doesn't end a TCP read. Switched a
  55 *					bit to skb ops.
  56 *		Alan Cox	:	Tidied tcp_data to avoid a potential
  57 *					nasty.
  58 *		Alan Cox	:	Added some better commenting, as the
  59 *					tcp is hard to follow
  60 *		Alan Cox	:	Removed incorrect check for 20 * psh
  61 *	Michael O'Reilly	:	ack < copied bug fix.
  62 *	Johannes Stille		:	Misc tcp fixes (not all in yet).
  63 *		Alan Cox	:	FIN with no memory -> CRASH
  64 *		Alan Cox	:	Added socket option proto entries.
  65 *					Also added awareness of them to accept.
  66 *		Alan Cox	:	Added TCP options (SOL_TCP)
  67 *		Alan Cox	:	Switched wakeup calls to callbacks,
  68 *					so the kernel can layer network
  69 *					sockets.
  70 *		Alan Cox	:	Use ip_tos/ip_ttl settings.
  71 *		Alan Cox	:	Handle FIN (more) properly (we hope).
  72 *		Alan Cox	:	RST frames sent on unsynchronised
  73 *					state ack error.
  74 *		Alan Cox	:	Put in missing check for SYN bit.
  75 *		Alan Cox	:	Added tcp_select_window() aka NET2E
  76 *					window non shrink trick.
  77 *		Alan Cox	:	Added a couple of small NET2E timer
  78 *					fixes
  79 *		Charles Hedrick :	TCP fixes
  80 *		Toomas Tamm	:	TCP window fixes
  81 *		Alan Cox	:	Small URG fix to rlogin ^C ack fight
  82 *		Charles Hedrick	:	Rewrote most of it to actually work
  83 *		Linus		:	Rewrote tcp_read() and URG handling
  84 *					completely
  85 *		Gerhard Koerting:	Fixed some missing timer handling
  86 *		Matthew Dillon  :	Reworked TCP machine states as per RFC
  87 *		Gerhard Koerting:	PC/TCP workarounds
  88 *		Adam Caldwell	:	Assorted timer/timing errors
  89 *		Matthew Dillon	:	Fixed another RST bug
  90 *		Alan Cox	:	Move to kernel side addressing changes.
  91 *		Alan Cox	:	Beginning work on TCP fastpathing
  92 *					(not yet usable)
  93 *		Arnt Gulbrandsen:	Turbocharged tcp_check() routine.
  94 *		Alan Cox	:	TCP fast path debugging
  95 *		Alan Cox	:	Window clamping
  96 *		Michael Riepe	:	Bug in tcp_check()
  97 *		Matt Dillon	:	More TCP improvements and RST bug fixes
  98 *		Matt Dillon	:	Yet more small nasties remove from the
  99 *					TCP code (Be very nice to this man if
 100 *					tcp finally works 100%) 8)
 101 *		Alan Cox	:	BSD accept semantics.
 102 *		Alan Cox	:	Reset on closedown bug.
 103 *	Peter De Schrijver	:	ENOTCONN check missing in tcp_sendto().
 104 *		Michael Pall	:	Handle poll() after URG properly in
 105 *					all cases.
 106 *		Michael Pall	:	Undo the last fix in tcp_read_urg()
 107 *					(multi URG PUSH broke rlogin).
 108 *		Michael Pall	:	Fix the multi URG PUSH problem in
 109 *					tcp_readable(), poll() after URG
 110 *					works now.
 111 *		Michael Pall	:	recv(...,MSG_OOB) never blocks in the
 112 *					BSD api.
 113 *		Alan Cox	:	Changed the semantics of sk->socket to
 114 *					fix a race and a signal problem with
 115 *					accept() and async I/O.
 116 *		Alan Cox	:	Relaxed the rules on tcp_sendto().
 117 *		Yury Shevchuk	:	Really fixed accept() blocking problem.
 118 *		Craig I. Hagan  :	Allow for BSD compatible TIME_WAIT for
 119 *					clients/servers which listen in on
 120 *					fixed ports.
 121 *		Alan Cox	:	Cleaned the above up and shrank it to
 122 *					a sensible code size.
 123 *		Alan Cox	:	Self connect lockup fix.
 124 *		Alan Cox	:	No connect to multicast.
 125 *		Ross Biro	:	Close unaccepted children on master
 126 *					socket close.
 127 *		Alan Cox	:	Reset tracing code.
 128 *		Alan Cox	:	Spurious resets on shutdown.
 129 *		Alan Cox	:	Giant 15 minute/60 second timer error
 130 *		Alan Cox	:	Small whoops in polling before an
 131 *					accept.
 132 *		Alan Cox	:	Kept the state trace facility since
 133 *					it's handy for debugging.
 134 *		Alan Cox	:	More reset handler fixes.
 135 *		Alan Cox	:	Started rewriting the code based on
 136 *					the RFC's for other useful protocol
 137 *					references see: Comer, KA9Q NOS, and
 138 *					for a reference on the difference
 139 *					between specifications and how BSD
 140 *					works see the 4.4lite source.
 141 *		A.N.Kuznetsov	:	Don't time wait on completion of tidy
 142 *					close.
 143 *		Linus Torvalds	:	Fin/Shutdown & copied_seq changes.
 144 *		Linus Torvalds	:	Fixed BSD port reuse to work first syn
 145 *		Alan Cox	:	Reimplemented timers as per the RFC
 146 *					and using multiple timers for sanity.
 147 *		Alan Cox	:	Small bug fixes, and a lot of new
 148 *					comments.
 149 *		Alan Cox	:	Fixed dual reader crash by locking
 150 *					the buffers (much like datagram.c)
 151 *		Alan Cox	:	Fixed stuck sockets in probe. A probe
 152 *					now gets fed up of retrying without
 153 *					(even a no space) answer.
 154 *		Alan Cox	:	Extracted closing code better
 155 *		Alan Cox	:	Fixed the closing state machine to
 156 *					resemble the RFC.
 157 *		Alan Cox	:	More 'per spec' fixes.
 158 *		Jorge Cwik	:	Even faster checksumming.
 159 *		Alan Cox	:	tcp_data() doesn't ack illegal PSH
 160 *					only frames. At least one pc tcp stack
 161 *					generates them.
 162 *		Alan Cox	:	Cache last socket.
 163 *		Alan Cox	:	Per route irtt.
 164 *		Matt Day	:	poll()->select() match BSD precisely on error
 165 *		Alan Cox	:	New buffers
 166 *		Marc Tamsky	:	Various sk->prot->retransmits and
 167 *					sk->retransmits misupdating fixed.
 168 *					Fixed tcp_write_timeout: stuck close,
 169 *					and TCP syn retries gets used now.
 170 *		Mark Yarvis	:	In tcp_read_wakeup(), don't send an
 171 *					ack if state is TCP_CLOSED.
 172 *		Alan Cox	:	Look up device on a retransmit - routes may
 173 *					change. Doesn't yet cope with MSS shrink right
 174 *					but it's a start!
 175 *		Marc Tamsky	:	Closing in closing fixes.
 176 *		Mike Shaver	:	RFC1122 verifications.
 177 *		Alan Cox	:	rcv_saddr errors.
 178 *		Alan Cox	:	Block double connect().
 179 *		Alan Cox	:	Small hooks for enSKIP.
 180 *		Alexey Kuznetsov:	Path MTU discovery.
 181 *		Alan Cox	:	Support soft errors.
 182 *		Alan Cox	:	Fix MTU discovery pathological case
 183 *					when the remote claims no mtu!
 184 *		Marc Tamsky	:	TCP_CLOSE fix.
 185 *		Colin (G3TNE)	:	Send a reset on syn ack replies in
 186 *					window but wrong (fixes NT lpd problems)
 187 *		Pedro Roque	:	Better TCP window handling, delayed ack.
 188 *		Joerg Reuter	:	No modification of locked buffers in
 189 *					tcp_do_retransmit()
 190 *		Eric Schenk	:	Changed receiver side silly window
 191 *					avoidance algorithm to BSD style
 192 *					algorithm. This doubles throughput
 193 *					against machines running Solaris,
 194 *					and seems to result in general
 195 *					improvement.
 196 *	Stefan Magdalinski	:	adjusted tcp_readable() to fix FIONREAD
 197 *	Willy Konynenberg	:	Transparent proxying support.
 198 *	Mike McLagan		:	Routing by source
 199 *		Keith Owens	:	Do proper merging with partial SKB's in
 200 *					tcp_do_sendmsg to avoid burstiness.
 201 *		Eric Schenk	:	Fix fast close down bug with
 202 *					shutdown() followed by close().
 203 *		Andi Kleen 	:	Make poll agree with SIGIO
 204 *	Salvatore Sanfilippo	:	Support SO_LINGER with linger == 1 and
 205 *					lingertime == 0 (RFC 793 ABORT Call)
 206 *	Hirokazu Takahashi	:	Use copy_from_user() instead of
 207 *					csum_and_copy_from_user() if possible.
 208 *
 209 * Description of States:
 210 *
 211 *	TCP_SYN_SENT		sent a connection request, waiting for ack
 212 *
 213 *	TCP_SYN_RECV		received a connection request, sent ack,
 214 *				waiting for final ack in three-way handshake.
 215 *
 216 *	TCP_ESTABLISHED		connection established
 217 *
 218 *	TCP_FIN_WAIT1		our side has shutdown, waiting to complete
 219 *				transmission of remaining buffered data
 220 *
 221 *	TCP_FIN_WAIT2		all buffered data sent, waiting for remote
 222 *				to shutdown
 223 *
 224 *	TCP_CLOSING		both sides have shutdown but we still have
 225 *				data we have to finish sending
 226 *
 227 *	TCP_TIME_WAIT		timeout to catch resent junk before entering
 228 *				closed, can only be entered from FIN_WAIT2
 229 *				or CLOSING.  Required because the other end
 230 *				may not have gotten our last ACK causing it
 231 *				to retransmit the data packet (which we ignore)
 232 *
 233 *	TCP_CLOSE_WAIT		remote side has shutdown and is waiting for
 234 *				us to finish writing our data and to shutdown
 235 *				(we have to close() to move on to LAST_ACK)
 236 *
 237 *	TCP_LAST_ACK		out side has shutdown after remote has
 238 *				shutdown.  There may still be data in our
 239 *				buffer that we have to finish sending
 240 *
 241 *	TCP_CLOSE		socket is finished
 242 */
 243
 244#define pr_fmt(fmt) "TCP: " fmt
 245
 246#include <crypto/hash.h>
 247#include <linux/kernel.h>
 248#include <linux/module.h>
 249#include <linux/types.h>
 250#include <linux/fcntl.h>
 251#include <linux/poll.h>
 252#include <linux/inet_diag.h>
 253#include <linux/init.h>
 254#include <linux/fs.h>
 255#include <linux/skbuff.h>
 256#include <linux/scatterlist.h>
 257#include <linux/splice.h>
 258#include <linux/net.h>
 259#include <linux/socket.h>
 260#include <linux/random.h>
 261#include <linux/memblock.h>
 262#include <linux/highmem.h>
 263#include <linux/swap.h>
 264#include <linux/cache.h>
 265#include <linux/err.h>
 266#include <linux/time.h>
 267#include <linux/slab.h>
 268#include <linux/errqueue.h>
 269#include <linux/static_key.h>
 
 270
 271#include <net/icmp.h>
 272#include <net/inet_common.h>
 273#include <net/tcp.h>
 
 
 274#include <net/xfrm.h>
 275#include <net/ip.h>
 276#include <net/sock.h>
 
 277
 278#include <linux/uaccess.h>
 279#include <asm/ioctls.h>
 280#include <net/busy_poll.h>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 281
 282struct percpu_counter tcp_orphan_count;
 283EXPORT_SYMBOL_GPL(tcp_orphan_count);
 284
 285long sysctl_tcp_mem[3] __read_mostly;
 286EXPORT_SYMBOL(sysctl_tcp_mem);
 287
 288atomic_long_t tcp_memory_allocated;	/* Current allocated memory. */
 289EXPORT_SYMBOL(tcp_memory_allocated);
 
 
 290
 291#if IS_ENABLED(CONFIG_SMC)
 292DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
 293EXPORT_SYMBOL(tcp_have_smc);
 294#endif
 295
 296/*
 297 * Current number of TCP sockets.
 298 */
 299struct percpu_counter tcp_sockets_allocated;
 300EXPORT_SYMBOL(tcp_sockets_allocated);
 301
 302/*
 303 * TCP splice context
 304 */
 305struct tcp_splice_state {
 306	struct pipe_inode_info *pipe;
 307	size_t len;
 308	unsigned int flags;
 309};
 310
 311/*
 312 * Pressure flag: try to collapse.
 313 * Technical note: it is used by multiple contexts non atomically.
 314 * All the __sk_mem_schedule() is of this nature: accounting
 315 * is strict, actions are advisory and have some latency.
 316 */
 317unsigned long tcp_memory_pressure __read_mostly;
 318EXPORT_SYMBOL_GPL(tcp_memory_pressure);
 319
 320DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
 321EXPORT_SYMBOL(tcp_rx_skb_cache_key);
 322
 323DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
 324
 325void tcp_enter_memory_pressure(struct sock *sk)
 326{
 327	unsigned long val;
 328
 329	if (READ_ONCE(tcp_memory_pressure))
 330		return;
 331	val = jiffies;
 332
 333	if (!val)
 334		val--;
 335	if (!cmpxchg(&tcp_memory_pressure, 0, val))
 336		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
 337}
 338EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
 339
 340void tcp_leave_memory_pressure(struct sock *sk)
 341{
 342	unsigned long val;
 343
 344	if (!READ_ONCE(tcp_memory_pressure))
 345		return;
 346	val = xchg(&tcp_memory_pressure, 0);
 347	if (val)
 348		NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
 349			      jiffies_to_msecs(jiffies - val));
 350}
 351EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
 352
 353/* Convert seconds to retransmits based on initial and max timeout */
 354static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
 355{
 356	u8 res = 0;
 357
 358	if (seconds > 0) {
 359		int period = timeout;
 360
 361		res = 1;
 362		while (seconds > period && res < 255) {
 363			res++;
 364			timeout <<= 1;
 365			if (timeout > rto_max)
 366				timeout = rto_max;
 367			period += timeout;
 368		}
 369	}
 370	return res;
 371}
 372
 373/* Convert retransmits to seconds based on initial and max timeout */
 374static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
 375{
 376	int period = 0;
 377
 378	if (retrans > 0) {
 379		period = timeout;
 380		while (--retrans) {
 381			timeout <<= 1;
 382			if (timeout > rto_max)
 383				timeout = rto_max;
 384			period += timeout;
 385		}
 386	}
 387	return period;
 388}
 389
 390static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
 391{
 392	u32 rate = READ_ONCE(tp->rate_delivered);
 393	u32 intv = READ_ONCE(tp->rate_interval_us);
 394	u64 rate64 = 0;
 395
 396	if (rate && intv) {
 397		rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
 398		do_div(rate64, intv);
 399	}
 400	return rate64;
 401}
 402
 403/* Address-family independent initialization for a tcp_sock.
 404 *
 405 * NOTE: A lot of things set to zero explicitly by call to
 406 *       sk_alloc() so need not be done here.
 407 */
 408void tcp_init_sock(struct sock *sk)
 409{
 410	struct inet_connection_sock *icsk = inet_csk(sk);
 411	struct tcp_sock *tp = tcp_sk(sk);
 
 412
 413	tp->out_of_order_queue = RB_ROOT;
 414	sk->tcp_rtx_queue = RB_ROOT;
 415	tcp_init_xmit_timers(sk);
 416	INIT_LIST_HEAD(&tp->tsq_node);
 417	INIT_LIST_HEAD(&tp->tsorted_sent_queue);
 418
 419	icsk->icsk_rto = TCP_TIMEOUT_INIT;
 
 
 
 420	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
 421	minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
 422
 423	/* So many TCP implementations out there (incorrectly) count the
 424	 * initial SYN frame in their delayed-ACK and congestion control
 425	 * algorithms that we must have the following bandaid to talk
 426	 * efficiently to them.  -DaveM
 427	 */
 428	tp->snd_cwnd = TCP_INIT_CWND;
 429
 430	/* There's a bubble in the pipe until at least the first ACK. */
 431	tp->app_limited = ~0U;
 
 432
 433	/* See draft-stevens-tcpca-spec-01 for discussion of the
 434	 * initialization of these values.
 435	 */
 436	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 437	tp->snd_cwnd_clamp = ~0;
 438	tp->mss_cache = TCP_MSS_DEFAULT;
 439
 440	tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
 441	tcp_assign_congestion_control(sk);
 442
 443	tp->tsoffset = 0;
 444	tp->rack.reo_wnd_steps = 1;
 445
 446	sk->sk_state = TCP_CLOSE;
 447
 448	sk->sk_write_space = sk_stream_write_space;
 449	sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
 450
 451	icsk->icsk_sync_mss = tcp_sync_mss;
 452
 453	WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
 454	WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
 
 455
 
 456	sk_sockets_allocated_inc(sk);
 457	sk->sk_route_forced_caps = NETIF_F_GSO;
 458}
 459EXPORT_SYMBOL(tcp_init_sock);
 460
 461static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
 462{
 463	struct sk_buff *skb = tcp_write_queue_tail(sk);
 
 464
 465	if (tsflags && skb) {
 466		struct skb_shared_info *shinfo = skb_shinfo(skb);
 467		struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 468
 469		sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
 470		if (tsflags & SOF_TIMESTAMPING_TX_ACK)
 471			tcb->txstamp_ack = 1;
 472		if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
 473			shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
 474	}
 475}
 476
 477static inline bool tcp_stream_is_readable(const struct tcp_sock *tp,
 478					  int target, struct sock *sk)
 479{
 480	return (READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->copied_seq) >= target) ||
 481		(sk->sk_prot->stream_memory_read ?
 482		sk->sk_prot->stream_memory_read(sk) : false);
 483}
 484
 485/*
 486 *	Wait for a TCP event.
 487 *
 488 *	Note that we don't need to lock the socket, as the upper poll layers
 489 *	take care of normal races (between the test and the event) and we don't
 490 *	go look at any of the socket buffers directly.
 491 */
 492__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
 493{
 494	__poll_t mask;
 495	struct sock *sk = sock->sk;
 496	const struct tcp_sock *tp = tcp_sk(sk);
 
 497	int state;
 498
 499	sock_poll_wait(file, sock, wait);
 500
 501	state = inet_sk_state_load(sk);
 502	if (state == TCP_LISTEN)
 503		return inet_csk_listen_poll(sk);
 504
 505	/* Socket is not locked. We are protected from async events
 506	 * by poll logic and correct handling of state changes
 507	 * made by other threads is impossible in any case.
 508	 */
 509
 510	mask = 0;
 511
 512	/*
 513	 * EPOLLHUP is certainly not done right. But poll() doesn't
 514	 * have a notion of HUP in just one direction, and for a
 515	 * socket the read side is more interesting.
 516	 *
 517	 * Some poll() documentation says that EPOLLHUP is incompatible
 518	 * with the EPOLLOUT/POLLWR flags, so somebody should check this
 519	 * all. But careful, it tends to be safer to return too many
 520	 * bits than too few, and you can easily break real applications
 521	 * if you don't tell them that something has hung up!
 522	 *
 523	 * Check-me.
 524	 *
 525	 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
 526	 * our fs/select.c). It means that after we received EOF,
 527	 * poll always returns immediately, making impossible poll() on write()
 528	 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
 529	 * if and only if shutdown has been made in both directions.
 530	 * Actually, it is interesting to look how Solaris and DUX
 531	 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
 532	 * then we could set it on SND_SHUTDOWN. BTW examples given
 533	 * in Stevens' books assume exactly this behaviour, it explains
 534	 * why EPOLLHUP is incompatible with EPOLLOUT.	--ANK
 535	 *
 536	 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
 537	 * blocking on fresh not-connected or disconnected socket. --ANK
 538	 */
 539	if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
 
 540		mask |= EPOLLHUP;
 541	if (sk->sk_shutdown & RCV_SHUTDOWN)
 542		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
 543
 544	/* Connected or passive Fast Open socket? */
 545	if (state != TCP_SYN_SENT &&
 546	    (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
 547		int target = sock_rcvlowat(sk, 0, INT_MAX);
 
 548
 549		if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
 550		    !sock_flag(sk, SOCK_URGINLINE) &&
 551		    tp->urg_data)
 552			target++;
 553
 554		if (tcp_stream_is_readable(tp, target, sk))
 555			mask |= EPOLLIN | EPOLLRDNORM;
 556
 557		if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
 558			if (sk_stream_is_writeable(sk)) {
 559				mask |= EPOLLOUT | EPOLLWRNORM;
 560			} else {  /* send SIGIO later */
 561				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 562				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 563
 564				/* Race breaker. If space is freed after
 565				 * wspace test but before the flags are set,
 566				 * IO signal will be lost. Memory barrier
 567				 * pairs with the input side.
 568				 */
 569				smp_mb__after_atomic();
 570				if (sk_stream_is_writeable(sk))
 571					mask |= EPOLLOUT | EPOLLWRNORM;
 572			}
 573		} else
 574			mask |= EPOLLOUT | EPOLLWRNORM;
 575
 576		if (tp->urg_data & TCP_URG_VALID)
 577			mask |= EPOLLPRI;
 578	} else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
 
 579		/* Active TCP fastopen socket with defer_connect
 580		 * Return EPOLLOUT so application can call write()
 581		 * in order for kernel to generate SYN+data
 582		 */
 583		mask |= EPOLLOUT | EPOLLWRNORM;
 584	}
 585	/* This barrier is coupled with smp_wmb() in tcp_reset() */
 586	smp_rmb();
 587	if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
 
 588		mask |= EPOLLERR;
 589
 590	return mask;
 591}
 592EXPORT_SYMBOL(tcp_poll);
 593
 594int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
 595{
 596	struct tcp_sock *tp = tcp_sk(sk);
 597	int answ;
 598	bool slow;
 599
 600	switch (cmd) {
 601	case SIOCINQ:
 602		if (sk->sk_state == TCP_LISTEN)
 603			return -EINVAL;
 604
 605		slow = lock_sock_fast(sk);
 606		answ = tcp_inq(sk);
 607		unlock_sock_fast(sk, slow);
 608		break;
 609	case SIOCATMARK:
 610		answ = tp->urg_data &&
 611		       READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
 612		break;
 613	case SIOCOUTQ:
 614		if (sk->sk_state == TCP_LISTEN)
 615			return -EINVAL;
 616
 617		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 618			answ = 0;
 619		else
 620			answ = READ_ONCE(tp->write_seq) - tp->snd_una;
 621		break;
 622	case SIOCOUTQNSD:
 623		if (sk->sk_state == TCP_LISTEN)
 624			return -EINVAL;
 625
 626		if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
 627			answ = 0;
 628		else
 629			answ = READ_ONCE(tp->write_seq) -
 630			       READ_ONCE(tp->snd_nxt);
 631		break;
 632	default:
 633		return -ENOIOCTLCMD;
 634	}
 635
 636	return put_user(answ, (int __user *)arg);
 
 637}
 638EXPORT_SYMBOL(tcp_ioctl);
 639
 640static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
 641{
 642	TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
 643	tp->pushed_seq = tp->write_seq;
 644}
 645
 646static inline bool forced_push(const struct tcp_sock *tp)
 647{
 648	return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
 649}
 650
 651static void skb_entail(struct sock *sk, struct sk_buff *skb)
 652{
 653	struct tcp_sock *tp = tcp_sk(sk);
 654	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 655
 656	skb->csum    = 0;
 657	tcb->seq     = tcb->end_seq = tp->write_seq;
 658	tcb->tcp_flags = TCPHDR_ACK;
 659	tcb->sacked  = 0;
 660	__skb_header_release(skb);
 661	tcp_add_write_queue_tail(sk, skb);
 662	sk_wmem_queued_add(sk, skb->truesize);
 663	sk_mem_charge(sk, skb->truesize);
 664	if (tp->nonagle & TCP_NAGLE_PUSH)
 665		tp->nonagle &= ~TCP_NAGLE_PUSH;
 666
 667	tcp_slow_start_after_idle_check(sk);
 668}
 669
 670static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
 671{
 672	if (flags & MSG_OOB)
 673		tp->snd_up = tp->write_seq;
 674}
 675
 676/* If a not yet filled skb is pushed, do not send it if
 677 * we have data packets in Qdisc or NIC queues :
 678 * Because TX completion will happen shortly, it gives a chance
 679 * to coalesce future sendmsg() payload into this skb, without
 680 * need for a timer, and with no latency trade off.
 681 * As packets containing data payload have a bigger truesize
 682 * than pure acks (dataless) packets, the last checks prevent
 683 * autocorking if we only have an ACK in Qdisc/NIC queues,
 684 * or if TX completion was delayed after we processed ACK packet.
 685 */
 686static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
 687				int size_goal)
 688{
 689	return skb->len < size_goal &&
 690	       sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
 691	       !tcp_rtx_queue_empty(sk) &&
 692	       refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
 
 693}
 694
 695static void tcp_push(struct sock *sk, int flags, int mss_now,
 696		     int nonagle, int size_goal)
 697{
 698	struct tcp_sock *tp = tcp_sk(sk);
 699	struct sk_buff *skb;
 700
 701	skb = tcp_write_queue_tail(sk);
 702	if (!skb)
 703		return;
 704	if (!(flags & MSG_MORE) || forced_push(tp))
 705		tcp_mark_push(tp, skb);
 706
 707	tcp_mark_urg(tp, flags);
 708
 709	if (tcp_should_autocork(sk, skb, size_goal)) {
 710
 711		/* avoid atomic op if TSQ_THROTTLED bit is already set */
 712		if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
 713			NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
 714			set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
 
 715		}
 716		/* It is possible TX completion already happened
 717		 * before we set TSQ_THROTTLED.
 718		 */
 719		if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
 720			return;
 721	}
 722
 723	if (flags & MSG_MORE)
 724		nonagle = TCP_NAGLE_CORK;
 725
 726	__tcp_push_pending_frames(sk, mss_now, nonagle);
 727}
 728
 729static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 730				unsigned int offset, size_t len)
 731{
 732	struct tcp_splice_state *tss = rd_desc->arg.data;
 733	int ret;
 734
 735	ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
 736			      min(rd_desc->count, len), tss->flags);
 737	if (ret > 0)
 738		rd_desc->count -= ret;
 739	return ret;
 740}
 741
 742static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
 743{
 744	/* Store TCP splice context information in read_descriptor_t. */
 745	read_descriptor_t rd_desc = {
 746		.arg.data = tss,
 747		.count	  = tss->len,
 748	};
 749
 750	return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
 751}
 752
 753/**
 754 *  tcp_splice_read - splice data from TCP socket to a pipe
 755 * @sock:	socket to splice from
 756 * @ppos:	position (not valid)
 757 * @pipe:	pipe to splice to
 758 * @len:	number of bytes to splice
 759 * @flags:	splice modifier flags
 760 *
 761 * Description:
 762 *    Will read pages from given socket and fill them into a pipe.
 763 *
 764 **/
 765ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
 766			struct pipe_inode_info *pipe, size_t len,
 767			unsigned int flags)
 768{
 769	struct sock *sk = sock->sk;
 770	struct tcp_splice_state tss = {
 771		.pipe = pipe,
 772		.len = len,
 773		.flags = flags,
 774	};
 775	long timeo;
 776	ssize_t spliced;
 777	int ret;
 778
 779	sock_rps_record_flow(sk);
 780	/*
 781	 * We can't seek on a socket input
 782	 */
 783	if (unlikely(*ppos))
 784		return -ESPIPE;
 785
 786	ret = spliced = 0;
 787
 788	lock_sock(sk);
 789
 790	timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
 791	while (tss.len) {
 792		ret = __tcp_splice_read(sk, &tss);
 793		if (ret < 0)
 794			break;
 795		else if (!ret) {
 796			if (spliced)
 797				break;
 798			if (sock_flag(sk, SOCK_DONE))
 799				break;
 800			if (sk->sk_err) {
 801				ret = sock_error(sk);
 802				break;
 803			}
 804			if (sk->sk_shutdown & RCV_SHUTDOWN)
 805				break;
 806			if (sk->sk_state == TCP_CLOSE) {
 807				/*
 808				 * This occurs when user tries to read
 809				 * from never connected socket.
 810				 */
 811				ret = -ENOTCONN;
 812				break;
 813			}
 814			if (!timeo) {
 815				ret = -EAGAIN;
 816				break;
 817			}
 818			/* if __tcp_splice_read() got nothing while we have
 819			 * an skb in receive queue, we do not want to loop.
 820			 * This might happen with URG data.
 821			 */
 822			if (!skb_queue_empty(&sk->sk_receive_queue))
 823				break;
 824			sk_wait_data(sk, &timeo, NULL);
 
 
 825			if (signal_pending(current)) {
 826				ret = sock_intr_errno(timeo);
 827				break;
 828			}
 829			continue;
 830		}
 831		tss.len -= ret;
 832		spliced += ret;
 833
 834		if (!timeo)
 835			break;
 836		release_sock(sk);
 837		lock_sock(sk);
 838
 839		if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
 840		    (sk->sk_shutdown & RCV_SHUTDOWN) ||
 841		    signal_pending(current))
 842			break;
 843	}
 844
 845	release_sock(sk);
 846
 847	if (spliced)
 848		return spliced;
 849
 850	return ret;
 851}
 852EXPORT_SYMBOL(tcp_splice_read);
 853
 854struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
 855				    bool force_schedule)
 856{
 857	struct sk_buff *skb;
 858
 859	if (likely(!size)) {
 860		skb = sk->sk_tx_skb_cache;
 861		if (skb) {
 862			skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
 863			sk->sk_tx_skb_cache = NULL;
 864			pskb_trim(skb, 0);
 865			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 866			skb_shinfo(skb)->tx_flags = 0;
 867			memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
 868			return skb;
 869		}
 870	}
 871	/* The TCP header must be at least 32-bit aligned.  */
 872	size = ALIGN(size, 4);
 873
 874	if (unlikely(tcp_under_memory_pressure(sk)))
 875		sk_mem_reclaim_partial(sk);
 876
 877	skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
 878	if (likely(skb)) {
 879		bool mem_scheduled;
 880
 
 881		if (force_schedule) {
 882			mem_scheduled = true;
 883			sk_forced_mem_schedule(sk, skb->truesize);
 884		} else {
 885			mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
 886		}
 887		if (likely(mem_scheduled)) {
 888			skb_reserve(skb, sk->sk_prot->max_header);
 889			/*
 890			 * Make sure that we have exactly size bytes
 891			 * available to the caller, no more, no less.
 892			 */
 893			skb->reserved_tailroom = skb->end - skb->tail - size;
 894			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
 895			return skb;
 896		}
 897		__kfree_skb(skb);
 898	} else {
 899		sk->sk_prot->enter_memory_pressure(sk);
 900		sk_stream_moderate_sndbuf(sk);
 901	}
 902	return NULL;
 903}
 904
 905static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
 906				       int large_allowed)
 907{
 908	struct tcp_sock *tp = tcp_sk(sk);
 909	u32 new_size_goal, size_goal;
 910
 911	if (!large_allowed)
 912		return mss_now;
 913
 914	/* Note : tcp_tso_autosize() will eventually split this later */
 915	new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
 916	new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
 917
 918	/* We try hard to avoid divides here */
 919	size_goal = tp->gso_segs * mss_now;
 920	if (unlikely(new_size_goal < size_goal ||
 921		     new_size_goal >= size_goal + mss_now)) {
 922		tp->gso_segs = min_t(u16, new_size_goal / mss_now,
 923				     sk->sk_gso_max_segs);
 924		size_goal = tp->gso_segs * mss_now;
 925	}
 926
 927	return max(size_goal, mss_now);
 928}
 929
 930static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
 931{
 932	int mss_now;
 933
 934	mss_now = tcp_current_mss(sk);
 935	*size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
 936
 937	return mss_now;
 938}
 939
 940/* In some cases, both sendpage() and sendmsg() could have added
 941 * an skb to the write queue, but failed adding payload on it.
 942 * We need to remove it to consume less memory, but more
 943 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
 944 * users.
 945 */
 946static void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
 947{
 948	if (skb && !skb->len) {
 
 
 949		tcp_unlink_write_queue(skb, sk);
 950		if (tcp_write_queue_empty(sk))
 951			tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
 952		sk_wmem_free_skb(sk, skb);
 953	}
 954}
 955
 956ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
 957			 size_t size, int flags)
 958{
 959	struct tcp_sock *tp = tcp_sk(sk);
 960	int mss_now, size_goal;
 961	int err;
 962	ssize_t copied;
 963	long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 964
 965	if (IS_ENABLED(CONFIG_DEBUG_VM) &&
 966	    WARN_ONCE(PageSlab(page), "page must not be a Slab one"))
 967		return -EINVAL;
 968
 969	/* Wait for a connection to finish. One exception is TCP Fast Open
 970	 * (passive side) where data is allowed to be sent before a connection
 971	 * is fully established.
 972	 */
 973	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
 974	    !tcp_passive_fastopen(sk)) {
 975		err = sk_stream_wait_connect(sk, &timeo);
 976		if (err != 0)
 977			goto out_err;
 978	}
 979
 980	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 981
 982	mss_now = tcp_send_mss(sk, &size_goal, flags);
 983	copied = 0;
 984
 985	err = -EPIPE;
 986	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
 987		goto out_err;
 988
 989	while (size > 0) {
 990		struct sk_buff *skb = tcp_write_queue_tail(sk);
 991		int copy, i;
 992		bool can_coalesce;
 993
 994		if (!skb || (copy = size_goal - skb->len) <= 0 ||
 995		    !tcp_skb_can_collapse_to(skb)) {
 996new_segment:
 997			if (!sk_stream_memory_free(sk))
 998				goto wait_for_sndbuf;
 999
1000			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1001					tcp_rtx_and_write_queues_empty(sk));
1002			if (!skb)
1003				goto wait_for_memory;
1004
1005#ifdef CONFIG_TLS_DEVICE
1006			skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
1007#endif
1008			skb_entail(sk, skb);
1009			copy = size_goal;
1010		}
1011
1012		if (copy > size)
1013			copy = size;
1014
1015		i = skb_shinfo(skb)->nr_frags;
1016		can_coalesce = skb_can_coalesce(skb, i, page, offset);
1017		if (!can_coalesce && i >= sysctl_max_skb_frags) {
1018			tcp_mark_push(tp, skb);
1019			goto new_segment;
1020		}
1021		if (!sk_wmem_schedule(sk, copy))
1022			goto wait_for_memory;
1023
1024		if (can_coalesce) {
1025			skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1026		} else {
1027			get_page(page);
1028			skb_fill_page_desc(skb, i, page, offset, copy);
1029		}
1030
1031		if (!(flags & MSG_NO_SHARED_FRAGS))
1032			skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
1033
1034		skb->len += copy;
1035		skb->data_len += copy;
1036		skb->truesize += copy;
1037		sk_wmem_queued_add(sk, copy);
1038		sk_mem_charge(sk, copy);
1039		skb->ip_summed = CHECKSUM_PARTIAL;
1040		WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1041		TCP_SKB_CB(skb)->end_seq += copy;
1042		tcp_skb_pcount_set(skb, 0);
1043
1044		if (!copied)
1045			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1046
1047		copied += copy;
1048		offset += copy;
1049		size -= copy;
1050		if (!size)
1051			goto out;
1052
1053		if (skb->len < size_goal || (flags & MSG_OOB))
1054			continue;
1055
1056		if (forced_push(tp)) {
1057			tcp_mark_push(tp, skb);
1058			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1059		} else if (skb == tcp_send_head(sk))
1060			tcp_push_one(sk, mss_now);
1061		continue;
1062
1063wait_for_sndbuf:
1064		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1065wait_for_memory:
1066		tcp_push(sk, flags & ~MSG_MORE, mss_now,
1067			 TCP_NAGLE_PUSH, size_goal);
1068
1069		err = sk_stream_wait_memory(sk, &timeo);
1070		if (err != 0)
1071			goto do_error;
1072
1073		mss_now = tcp_send_mss(sk, &size_goal, flags);
1074	}
1075
1076out:
1077	if (copied) {
1078		tcp_tx_timestamp(sk, sk->sk_tsflags);
1079		if (!(flags & MSG_SENDPAGE_NOTLAST))
1080			tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1081	}
1082	return copied;
1083
1084do_error:
1085	tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1086	if (copied)
1087		goto out;
1088out_err:
1089	/* make sure we wake any epoll edge trigger waiter */
1090	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1091		     err == -EAGAIN)) {
1092		sk->sk_write_space(sk);
1093		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1094	}
1095	return sk_stream_error(sk, flags, err);
1096}
1097EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1098
1099int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1100			size_t size, int flags)
1101{
1102	if (!(sk->sk_route_caps & NETIF_F_SG))
1103		return sock_no_sendpage_locked(sk, page, offset, size, flags);
1104
1105	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1106
1107	return do_tcp_sendpages(sk, page, offset, size, flags);
1108}
1109EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1110
1111int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1112		 size_t size, int flags)
1113{
1114	int ret;
1115
1116	lock_sock(sk);
1117	ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1118	release_sock(sk);
1119
1120	return ret;
 
 
 
 
 
 
 
1121}
1122EXPORT_SYMBOL(tcp_sendpage);
1123
1124void tcp_free_fastopen_req(struct tcp_sock *tp)
1125{
1126	if (tp->fastopen_req) {
1127		kfree(tp->fastopen_req);
1128		tp->fastopen_req = NULL;
1129	}
1130}
1131
1132static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1133				int *copied, size_t size,
1134				struct ubuf_info *uarg)
1135{
1136	struct tcp_sock *tp = tcp_sk(sk);
1137	struct inet_sock *inet = inet_sk(sk);
1138	struct sockaddr *uaddr = msg->msg_name;
1139	int err, flags;
1140
1141	if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
 
1142	    (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1143	     uaddr->sa_family == AF_UNSPEC))
1144		return -EOPNOTSUPP;
1145	if (tp->fastopen_req)
1146		return -EALREADY; /* Another Fast Open is in progress */
1147
1148	tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1149				   sk->sk_allocation);
1150	if (unlikely(!tp->fastopen_req))
1151		return -ENOBUFS;
1152	tp->fastopen_req->data = msg;
1153	tp->fastopen_req->size = size;
1154	tp->fastopen_req->uarg = uarg;
1155
1156	if (inet->defer_connect) {
1157		err = tcp_connect(sk);
1158		/* Same failure procedure as in tcp_v4/6_connect */
1159		if (err) {
1160			tcp_set_state(sk, TCP_CLOSE);
1161			inet->inet_dport = 0;
1162			sk->sk_route_caps = 0;
1163		}
1164	}
1165	flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1166	err = __inet_stream_connect(sk->sk_socket, uaddr,
1167				    msg->msg_namelen, flags, 1);
1168	/* fastopen_req could already be freed in __inet_stream_connect
1169	 * if the connection times out or gets rst
1170	 */
1171	if (tp->fastopen_req) {
1172		*copied = tp->fastopen_req->copied;
1173		tcp_free_fastopen_req(tp);
1174		inet->defer_connect = 0;
1175	}
1176	return err;
1177}
1178
1179int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1180{
1181	struct tcp_sock *tp = tcp_sk(sk);
1182	struct ubuf_info *uarg = NULL;
1183	struct sk_buff *skb;
1184	struct sockcm_cookie sockc;
1185	int flags, err, copied = 0;
1186	int mss_now = 0, size_goal, copied_syn = 0;
1187	int process_backlog = 0;
1188	bool zc = false;
1189	long timeo;
1190
1191	flags = msg->msg_flags;
1192
1193	if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1194		skb = tcp_write_queue_tail(sk);
1195		uarg = sock_zerocopy_realloc(sk, size, skb_zcopy(skb));
1196		if (!uarg) {
1197			err = -ENOBUFS;
1198			goto out_err;
 
 
 
 
 
 
 
 
 
 
1199		}
1200
1201		zc = sk->sk_route_caps & NETIF_F_SG;
1202		if (!zc)
1203			uarg->zerocopy = 0;
1204	}
1205
1206	if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
 
1207	    !tp->repair) {
1208		err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1209		if (err == -EINPROGRESS && copied_syn > 0)
1210			goto out;
1211		else if (err)
1212			goto out_err;
1213	}
1214
1215	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1216
1217	tcp_rate_check_app_limited(sk);  /* is sending application-limited? */
1218
1219	/* Wait for a connection to finish. One exception is TCP Fast Open
1220	 * (passive side) where data is allowed to be sent before a connection
1221	 * is fully established.
1222	 */
1223	if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1224	    !tcp_passive_fastopen(sk)) {
1225		err = sk_stream_wait_connect(sk, &timeo);
1226		if (err != 0)
1227			goto do_error;
1228	}
1229
1230	if (unlikely(tp->repair)) {
1231		if (tp->repair_queue == TCP_RECV_QUEUE) {
1232			copied = tcp_send_rcvq(sk, msg, size);
1233			goto out_nopush;
1234		}
1235
1236		err = -EINVAL;
1237		if (tp->repair_queue == TCP_NO_QUEUE)
1238			goto out_err;
1239
1240		/* 'common' sending to sendq */
1241	}
1242
1243	sockcm_init(&sockc, sk);
1244	if (msg->msg_controllen) {
1245		err = sock_cmsg_send(sk, msg, &sockc);
1246		if (unlikely(err)) {
1247			err = -EINVAL;
1248			goto out_err;
1249		}
1250	}
1251
1252	/* This should be in poll */
1253	sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1254
1255	/* Ok commence sending. */
1256	copied = 0;
1257
1258restart:
1259	mss_now = tcp_send_mss(sk, &size_goal, flags);
1260
1261	err = -EPIPE;
1262	if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1263		goto do_error;
1264
1265	while (msg_data_left(msg)) {
1266		int copy = 0;
1267
1268		skb = tcp_write_queue_tail(sk);
1269		if (skb)
1270			copy = size_goal - skb->len;
1271
1272		if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1273			bool first_skb;
1274
1275new_segment:
1276			if (!sk_stream_memory_free(sk))
1277				goto wait_for_sndbuf;
1278
1279			if (unlikely(process_backlog >= 16)) {
1280				process_backlog = 0;
1281				if (sk_flush_backlog(sk))
1282					goto restart;
1283			}
1284			first_skb = tcp_rtx_and_write_queues_empty(sk);
1285			skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1286						  first_skb);
1287			if (!skb)
1288				goto wait_for_memory;
1289
1290			process_backlog++;
1291			skb->ip_summed = CHECKSUM_PARTIAL;
1292
1293			skb_entail(sk, skb);
 
 
 
1294			copy = size_goal;
1295
1296			/* All packets are restored as if they have
1297			 * already been sent. skb_mstamp_ns isn't set to
1298			 * avoid wrong rtt estimation.
1299			 */
1300			if (tp->repair)
1301				TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1302		}
1303
1304		/* Try to append data to the end of skb. */
1305		if (copy > msg_data_left(msg))
1306			copy = msg_data_left(msg);
1307
1308		/* Where to copy to? */
1309		if (skb_availroom(skb) > 0 && !zc) {
1310			/* We have some space in skb head. Superb! */
1311			copy = min_t(int, copy, skb_availroom(skb));
1312			err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1313			if (err)
1314				goto do_fault;
1315		} else if (!zc) {
1316			bool merge = true;
1317			int i = skb_shinfo(skb)->nr_frags;
1318			struct page_frag *pfrag = sk_page_frag(sk);
1319
1320			if (!sk_page_frag_refill(sk, pfrag))
1321				goto wait_for_memory;
1322
1323			if (!skb_can_coalesce(skb, i, pfrag->page,
1324					      pfrag->offset)) {
1325				if (i >= sysctl_max_skb_frags) {
1326					tcp_mark_push(tp, skb);
1327					goto new_segment;
1328				}
1329				merge = false;
1330			}
1331
1332			copy = min_t(int, copy, pfrag->size - pfrag->offset);
1333
1334			if (!sk_wmem_schedule(sk, copy))
1335				goto wait_for_memory;
 
 
 
 
 
 
 
1336
1337			err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1338						       pfrag->page,
1339						       pfrag->offset,
1340						       copy);
1341			if (err)
1342				goto do_error;
1343
1344			/* Update the skb. */
1345			if (merge) {
1346				skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1347			} else {
1348				skb_fill_page_desc(skb, i, pfrag->page,
1349						   pfrag->offset, copy);
1350				page_ref_inc(pfrag->page);
1351			}
1352			pfrag->offset += copy;
1353		} else {
 
 
 
 
 
 
 
 
 
 
 
 
1354			err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1355			if (err == -EMSGSIZE || err == -EEXIST) {
1356				tcp_mark_push(tp, skb);
1357				goto new_segment;
1358			}
1359			if (err < 0)
1360				goto do_error;
1361			copy = err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1362		}
1363
1364		if (!copied)
1365			TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1366
1367		WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1368		TCP_SKB_CB(skb)->end_seq += copy;
1369		tcp_skb_pcount_set(skb, 0);
1370
1371		copied += copy;
1372		if (!msg_data_left(msg)) {
1373			if (unlikely(flags & MSG_EOR))
1374				TCP_SKB_CB(skb)->eor = 1;
1375			goto out;
1376		}
1377
1378		if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1379			continue;
1380
1381		if (forced_push(tp)) {
1382			tcp_mark_push(tp, skb);
1383			__tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1384		} else if (skb == tcp_send_head(sk))
1385			tcp_push_one(sk, mss_now);
1386		continue;
1387
1388wait_for_sndbuf:
1389		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1390wait_for_memory:
1391		if (copied)
1392			tcp_push(sk, flags & ~MSG_MORE, mss_now,
1393				 TCP_NAGLE_PUSH, size_goal);
1394
1395		err = sk_stream_wait_memory(sk, &timeo);
1396		if (err != 0)
1397			goto do_error;
1398
1399		mss_now = tcp_send_mss(sk, &size_goal, flags);
1400	}
1401
1402out:
1403	if (copied) {
1404		tcp_tx_timestamp(sk, sockc.tsflags);
1405		tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1406	}
1407out_nopush:
1408	sock_zerocopy_put(uarg);
 
 
1409	return copied + copied_syn;
1410
1411do_error:
1412	skb = tcp_write_queue_tail(sk);
1413do_fault:
1414	tcp_remove_empty_skb(sk, skb);
1415
1416	if (copied + copied_syn)
1417		goto out;
1418out_err:
1419	sock_zerocopy_put_abort(uarg, true);
 
 
1420	err = sk_stream_error(sk, flags, err);
1421	/* make sure we wake any epoll edge trigger waiter */
1422	if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1423		     err == -EAGAIN)) {
1424		sk->sk_write_space(sk);
1425		tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1426	}
1427	return err;
1428}
1429EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1430
1431int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1432{
1433	int ret;
1434
1435	lock_sock(sk);
1436	ret = tcp_sendmsg_locked(sk, msg, size);
1437	release_sock(sk);
1438
1439	return ret;
1440}
1441EXPORT_SYMBOL(tcp_sendmsg);
1442
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1443/*
1444 *	Handle reading urgent data. BSD has very simple semantics for
1445 *	this, no blocking and very strange errors 8)
1446 */
1447
1448static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1449{
1450	struct tcp_sock *tp = tcp_sk(sk);
1451
1452	/* No URG data to read. */
1453	if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1454	    tp->urg_data == TCP_URG_READ)
1455		return -EINVAL;	/* Yes this is right ! */
1456
1457	if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1458		return -ENOTCONN;
1459
1460	if (tp->urg_data & TCP_URG_VALID) {
1461		int err = 0;
1462		char c = tp->urg_data;
1463
1464		if (!(flags & MSG_PEEK))
1465			tp->urg_data = TCP_URG_READ;
1466
1467		/* Read urgent data. */
1468		msg->msg_flags |= MSG_OOB;
1469
1470		if (len > 0) {
1471			if (!(flags & MSG_TRUNC))
1472				err = memcpy_to_msg(msg, &c, 1);
1473			len = 1;
1474		} else
1475			msg->msg_flags |= MSG_TRUNC;
1476
1477		return err ? -EFAULT : len;
1478	}
1479
1480	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1481		return 0;
1482
1483	/* Fixed the recv(..., MSG_OOB) behaviour.  BSD docs and
1484	 * the available implementations agree in this case:
1485	 * this call should never block, independent of the
1486	 * blocking state of the socket.
1487	 * Mike <pall@rz.uni-karlsruhe.de>
1488	 */
1489	return -EAGAIN;
1490}
1491
1492static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1493{
1494	struct sk_buff *skb;
1495	int copied = 0, err = 0;
1496
1497	/* XXX -- need to support SO_PEEK_OFF */
1498
1499	skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1500		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1501		if (err)
1502			return err;
1503		copied += skb->len;
1504	}
1505
1506	skb_queue_walk(&sk->sk_write_queue, skb) {
1507		err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1508		if (err)
1509			break;
1510
1511		copied += skb->len;
1512	}
1513
1514	return err ?: copied;
1515}
1516
1517/* Clean up the receive buffer for full frames taken by the user,
1518 * then send an ACK if necessary.  COPIED is the number of bytes
1519 * tcp_recvmsg has given to the user so far, it speeds up the
1520 * calculation of whether or not we must ACK for the sake of
1521 * a window update.
1522 */
1523static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1524{
1525	struct tcp_sock *tp = tcp_sk(sk);
1526	bool time_to_ack = false;
1527
1528	struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1529
1530	WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1531	     "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1532	     tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1533
1534	if (inet_csk_ack_scheduled(sk)) {
1535		const struct inet_connection_sock *icsk = inet_csk(sk);
1536		   /* Delayed ACKs frequently hit locked sockets during bulk
1537		    * receive. */
1538		if (icsk->icsk_ack.blocked ||
1539		    /* Once-per-two-segments ACK was not sent by tcp_input.c */
1540		    tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1541		    /*
1542		     * If this read emptied read buffer, we send ACK, if
1543		     * connection is not bidirectional, user drained
1544		     * receive buffer and there was a small segment
1545		     * in queue.
1546		     */
1547		    (copied > 0 &&
1548		     ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1549		      ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1550		       !inet_csk_in_pingpong_mode(sk))) &&
1551		      !atomic_read(&sk->sk_rmem_alloc)))
1552			time_to_ack = true;
1553	}
1554
1555	/* We send an ACK if we can now advertise a non-zero window
1556	 * which has been raised "significantly".
1557	 *
1558	 * Even if window raised up to infinity, do not send window open ACK
1559	 * in states, where we will not receive more. It is useless.
1560	 */
1561	if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1562		__u32 rcv_window_now = tcp_receive_window(tp);
1563
1564		/* Optimize, __tcp_select_window() is not cheap. */
1565		if (2*rcv_window_now <= tp->window_clamp) {
1566			__u32 new_window = __tcp_select_window(sk);
1567
1568			/* Send ACK now, if this read freed lots of space
1569			 * in our buffer. Certainly, new_window is new window.
1570			 * We can advertise it now, if it is not less than current one.
1571			 * "Lots" means "at least twice" here.
1572			 */
1573			if (new_window && new_window >= 2 * rcv_window_now)
1574				time_to_ack = true;
1575		}
1576	}
1577	if (time_to_ack)
1578		tcp_send_ack(sk);
1579}
1580
1581static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1582{
1583	struct sk_buff *skb;
1584	u32 offset;
1585
1586	while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1587		offset = seq - TCP_SKB_CB(skb)->seq;
1588		if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1589			pr_err_once("%s: found a SYN, please report !\n", __func__);
1590			offset--;
1591		}
1592		if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1593			*off = offset;
1594			return skb;
1595		}
1596		/* This looks weird, but this can happen if TCP collapsing
1597		 * splitted a fat GRO packet, while we released socket lock
1598		 * in skb_splice_bits()
1599		 */
1600		sk_eat_skb(sk, skb);
1601	}
1602	return NULL;
1603}
 
1604
1605/*
1606 * This routine provides an alternative to tcp_recvmsg() for routines
1607 * that would like to handle copying from skbuffs directly in 'sendfile'
1608 * fashion.
1609 * Note:
1610 *	- It is assumed that the socket was locked by the caller.
1611 *	- The routine does not block.
1612 *	- At present, there is no support for reading OOB data
1613 *	  or for 'peeking' the socket using this routine
1614 *	  (although both would be easy to implement).
1615 */
1616int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1617		  sk_read_actor_t recv_actor)
1618{
1619	struct sk_buff *skb;
1620	struct tcp_sock *tp = tcp_sk(sk);
1621	u32 seq = tp->copied_seq;
1622	u32 offset;
1623	int copied = 0;
1624
1625	if (sk->sk_state == TCP_LISTEN)
1626		return -ENOTCONN;
1627	while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1628		if (offset < skb->len) {
1629			int used;
1630			size_t len;
1631
1632			len = skb->len - offset;
1633			/* Stop reading if we hit a patch of urgent data */
1634			if (tp->urg_data) {
1635				u32 urg_offset = tp->urg_seq - seq;
1636				if (urg_offset < len)
1637					len = urg_offset;
1638				if (!len)
1639					break;
1640			}
1641			used = recv_actor(desc, skb, offset, len);
1642			if (used <= 0) {
1643				if (!copied)
1644					copied = used;
1645				break;
1646			} else if (used <= len) {
1647				seq += used;
1648				copied += used;
1649				offset += used;
1650			}
 
 
 
 
 
 
1651			/* If recv_actor drops the lock (e.g. TCP splice
1652			 * receive) the skb pointer might be invalid when
1653			 * getting here: tcp_collapse might have deleted it
1654			 * while aggregating skbs from the socket queue.
1655			 */
1656			skb = tcp_recv_skb(sk, seq - 1, &offset);
1657			if (!skb)
1658				break;
1659			/* TCP coalescing might have appended data to the skb.
1660			 * Try to splice more frags
1661			 */
1662			if (offset + 1 != skb->len)
1663				continue;
1664		}
1665		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1666			sk_eat_skb(sk, skb);
1667			++seq;
1668			break;
1669		}
1670		sk_eat_skb(sk, skb);
1671		if (!desc->count)
1672			break;
1673		WRITE_ONCE(tp->copied_seq, seq);
1674	}
1675	WRITE_ONCE(tp->copied_seq, seq);
1676
1677	tcp_rcv_space_adjust(sk);
1678
1679	/* Clean up data we have read: This will do ACK frames. */
1680	if (copied > 0) {
1681		tcp_recv_skb(sk, seq, &offset);
1682		tcp_cleanup_rbuf(sk, copied);
1683	}
1684	return copied;
1685}
1686EXPORT_SYMBOL(tcp_read_sock);
1687
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1688int tcp_peek_len(struct socket *sock)
1689{
1690	return tcp_inq(sock->sk);
1691}
1692EXPORT_SYMBOL(tcp_peek_len);
1693
1694/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1695int tcp_set_rcvlowat(struct sock *sk, int val)
1696{
1697	int cap;
1698
1699	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1700		cap = sk->sk_rcvbuf >> 1;
1701	else
1702		cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1703	val = min(val, cap);
1704	WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1705
1706	/* Check if we need to signal EPOLLIN right now */
1707	tcp_data_ready(sk);
1708
1709	if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1710		return 0;
1711
1712	val <<= 1;
1713	if (val > sk->sk_rcvbuf) {
1714		WRITE_ONCE(sk->sk_rcvbuf, val);
1715		tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1716	}
1717	return 0;
1718}
1719EXPORT_SYMBOL(tcp_set_rcvlowat);
1720
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1721#ifdef CONFIG_MMU
1722static const struct vm_operations_struct tcp_vm_ops = {
1723};
1724
1725int tcp_mmap(struct file *file, struct socket *sock,
1726	     struct vm_area_struct *vma)
1727{
1728	if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1729		return -EPERM;
1730	vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1731
1732	/* Instruct vm_insert_page() to not down_read(mmap_sem) */
1733	vma->vm_flags |= VM_MIXEDMAP;
1734
1735	vma->vm_ops = &tcp_vm_ops;
1736	return 0;
1737}
1738EXPORT_SYMBOL(tcp_mmap);
1739
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1740static int tcp_zerocopy_receive(struct sock *sk,
1741				struct tcp_zerocopy_receive *zc)
 
1742{
 
1743	unsigned long address = (unsigned long)zc->address;
 
 
 
1744	const skb_frag_t *frags = NULL;
1745	u32 length = 0, seq, offset;
1746	struct vm_area_struct *vma;
1747	struct sk_buff *skb = NULL;
1748	struct tcp_sock *tp;
1749	int inq;
 
 
1750	int ret;
1751
 
 
 
1752	if (address & (PAGE_SIZE - 1) || address != zc->address)
1753		return -EINVAL;
1754
1755	if (sk->sk_state == TCP_LISTEN)
1756		return -ENOTCONN;
1757
1758	sock_rps_record_flow(sk);
1759
1760	down_read(&current->mm->mmap_sem);
 
 
 
 
 
 
 
 
 
1761
1762	ret = -EINVAL;
1763	vma = find_vma(current->mm, address);
1764	if (!vma || vma->vm_start > address || vma->vm_ops != &tcp_vm_ops)
1765		goto out;
1766	zc->length = min_t(unsigned long, zc->length, vma->vm_end - address);
1767
1768	tp = tcp_sk(sk);
1769	seq = tp->copied_seq;
1770	inq = tcp_inq(sk);
1771	zc->length = min_t(u32, zc->length, inq);
1772	zc->length &= ~(PAGE_SIZE - 1);
1773	if (zc->length) {
1774		zap_page_range(vma, address, zc->length);
 
1775		zc->recv_skip_hint = 0;
1776	} else {
1777		zc->recv_skip_hint = inq;
 
1778	}
1779	ret = 0;
1780	while (length + PAGE_SIZE <= zc->length) {
 
 
 
1781		if (zc->recv_skip_hint < PAGE_SIZE) {
 
 
1782			if (skb) {
 
 
1783				skb = skb->next;
1784				offset = seq - TCP_SKB_CB(skb)->seq;
1785			} else {
1786				skb = tcp_recv_skb(sk, seq, &offset);
1787			}
1788
 
 
 
 
 
 
 
1789			zc->recv_skip_hint = skb->len - offset;
1790			offset -= skb_headlen(skb);
1791			if ((int)offset < 0 || skb_has_frag_list(skb))
1792				break;
1793			frags = skb_shinfo(skb)->frags;
1794			while (offset) {
1795				if (skb_frag_size(frags) > offset)
1796					goto out;
1797				offset -= skb_frag_size(frags);
1798				frags++;
1799			}
1800		}
1801		if (skb_frag_size(frags) != PAGE_SIZE || skb_frag_off(frags)) {
1802			int remaining = zc->recv_skip_hint;
1803
1804			while (remaining && (skb_frag_size(frags) != PAGE_SIZE ||
1805					     skb_frag_off(frags))) {
1806				remaining -= skb_frag_size(frags);
1807				frags++;
1808			}
1809			zc->recv_skip_hint -= remaining;
1810			break;
1811		}
1812		ret = vm_insert_page(vma, address + length,
1813				     skb_frag_page(frags));
1814		if (ret)
1815			break;
 
 
 
1816		length += PAGE_SIZE;
1817		seq += PAGE_SIZE;
1818		zc->recv_skip_hint -= PAGE_SIZE;
1819		frags++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1820	}
1821out:
1822	up_read(&current->mm->mmap_sem);
1823	if (length) {
 
 
 
 
 
 
 
1824		WRITE_ONCE(tp->copied_seq, seq);
1825		tcp_rcv_space_adjust(sk);
1826
1827		/* Clean up data we have read: This will do ACK frames. */
1828		tcp_recv_skb(sk, seq, &offset);
1829		tcp_cleanup_rbuf(sk, length);
1830		ret = 0;
1831		if (length == zc->length)
1832			zc->recv_skip_hint = 0;
1833	} else {
1834		if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
1835			ret = -EIO;
1836	}
1837	zc->length = length;
1838	return ret;
1839}
1840#endif
1841
1842static void tcp_update_recv_tstamps(struct sk_buff *skb,
1843				    struct scm_timestamping_internal *tss)
1844{
1845	if (skb->tstamp)
1846		tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1847	else
1848		tss->ts[0] = (struct timespec64) {0};
1849
1850	if (skb_hwtstamps(skb)->hwtstamp)
1851		tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1852	else
1853		tss->ts[2] = (struct timespec64) {0};
1854}
1855
1856/* Similar to __sock_recv_timestamp, but does not require an skb */
1857static void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
1858			       struct scm_timestamping_internal *tss)
1859{
1860	int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
 
1861	bool has_timestamping = false;
1862
1863	if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
1864		if (sock_flag(sk, SOCK_RCVTSTAMP)) {
1865			if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
1866				if (new_tstamp) {
1867					struct __kernel_timespec kts = {tss->ts[0].tv_sec, tss->ts[0].tv_nsec};
1868
 
 
1869					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
1870						 sizeof(kts), &kts);
1871				} else {
1872					struct timespec ts_old = timespec64_to_timespec(tss->ts[0]);
1873
 
 
1874					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
1875						 sizeof(ts_old), &ts_old);
1876				}
1877			} else {
1878				if (new_tstamp) {
1879					struct __kernel_sock_timeval stv;
1880
1881					stv.tv_sec = tss->ts[0].tv_sec;
1882					stv.tv_usec = tss->ts[0].tv_nsec / 1000;
1883					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
1884						 sizeof(stv), &stv);
1885				} else {
1886					struct __kernel_old_timeval tv;
1887
1888					tv.tv_sec = tss->ts[0].tv_sec;
1889					tv.tv_usec = tss->ts[0].tv_nsec / 1000;
1890					put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
1891						 sizeof(tv), &tv);
1892				}
1893			}
1894		}
1895
1896		if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
 
 
1897			has_timestamping = true;
1898		else
1899			tss->ts[0] = (struct timespec64) {0};
1900	}
1901
1902	if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
1903		if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
 
 
1904			has_timestamping = true;
1905		else
1906			tss->ts[2] = (struct timespec64) {0};
1907	}
1908
1909	if (has_timestamping) {
1910		tss->ts[1] = (struct timespec64) {0};
1911		if (sock_flag(sk, SOCK_TSTAMP_NEW))
1912			put_cmsg_scm_timestamping64(msg, tss);
1913		else
1914			put_cmsg_scm_timestamping(msg, tss);
1915	}
1916}
1917
1918static int tcp_inq_hint(struct sock *sk)
1919{
1920	const struct tcp_sock *tp = tcp_sk(sk);
1921	u32 copied_seq = READ_ONCE(tp->copied_seq);
1922	u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
1923	int inq;
1924
1925	inq = rcv_nxt - copied_seq;
1926	if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
1927		lock_sock(sk);
1928		inq = tp->rcv_nxt - tp->copied_seq;
1929		release_sock(sk);
1930	}
1931	/* After receiving a FIN, tell the user-space to continue reading
1932	 * by returning a non-zero inq.
1933	 */
1934	if (inq == 0 && sock_flag(sk, SOCK_DONE))
1935		inq = 1;
1936	return inq;
1937}
1938
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1939/*
1940 *	This routine copies from a sock struct into the user buffer.
1941 *
1942 *	Technical note: in 2.3 we work on _locked_ socket, so that
1943 *	tricks with *seq access order and skb->users are not required.
1944 *	Probably, code can be easily improved even more.
1945 */
1946
1947int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1948		int flags, int *addr_len)
 
1949{
1950	struct tcp_sock *tp = tcp_sk(sk);
 
1951	int copied = 0;
1952	u32 peek_seq;
1953	u32 *seq;
1954	unsigned long used;
1955	int err, inq;
1956	int target;		/* Read at least this many bytes */
1957	long timeo;
1958	struct sk_buff *skb, *last;
 
1959	u32 urg_hole = 0;
1960	struct scm_timestamping_internal tss;
1961	bool has_tss = false;
1962	bool has_cmsg;
1963
1964	if (unlikely(flags & MSG_ERRQUEUE))
1965		return inet_recv_error(sk, msg, len, addr_len);
1966
1967	if (sk_can_busy_loop(sk) && skb_queue_empty_lockless(&sk->sk_receive_queue) &&
1968	    (sk->sk_state == TCP_ESTABLISHED))
1969		sk_busy_loop(sk, nonblock);
1970
1971	lock_sock(sk);
1972
1973	err = -ENOTCONN;
1974	if (sk->sk_state == TCP_LISTEN)
1975		goto out;
1976
1977	has_cmsg = tp->recvmsg_inq;
1978	timeo = sock_rcvtimeo(sk, nonblock);
 
 
 
1979
1980	/* Urgent data needs to be handled specially. */
1981	if (flags & MSG_OOB)
1982		goto recv_urg;
1983
1984	if (unlikely(tp->repair)) {
1985		err = -EPERM;
1986		if (!(flags & MSG_PEEK))
1987			goto out;
1988
1989		if (tp->repair_queue == TCP_SEND_QUEUE)
1990			goto recv_sndq;
1991
1992		err = -EINVAL;
1993		if (tp->repair_queue == TCP_NO_QUEUE)
1994			goto out;
1995
1996		/* 'common' recv queue MSG_PEEK-ing */
1997	}
1998
1999	seq = &tp->copied_seq;
2000	if (flags & MSG_PEEK) {
2001		peek_seq = tp->copied_seq;
 
2002		seq = &peek_seq;
2003	}
2004
2005	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2006
2007	do {
2008		u32 offset;
2009
2010		/* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2011		if (tp->urg_data && tp->urg_seq == *seq) {
2012			if (copied)
2013				break;
2014			if (signal_pending(current)) {
2015				copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2016				break;
2017			}
2018		}
2019
2020		/* Next get a buffer. */
2021
2022		last = skb_peek_tail(&sk->sk_receive_queue);
2023		skb_queue_walk(&sk->sk_receive_queue, skb) {
2024			last = skb;
2025			/* Now that we have two receive queues this
2026			 * shouldn't happen.
2027			 */
2028			if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2029				 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2030				 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2031				 flags))
2032				break;
2033
2034			offset = *seq - TCP_SKB_CB(skb)->seq;
2035			if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2036				pr_err_once("%s: found a SYN, please report !\n", __func__);
2037				offset--;
2038			}
2039			if (offset < skb->len)
2040				goto found_ok_skb;
2041			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2042				goto found_fin_ok;
2043			WARN(!(flags & MSG_PEEK),
2044			     "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2045			     *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2046		}
2047
2048		/* Well, if we have backlog, try to process it now yet. */
2049
2050		if (copied >= target && !sk->sk_backlog.tail)
2051			break;
2052
2053		if (copied) {
2054			if (sk->sk_err ||
 
2055			    sk->sk_state == TCP_CLOSE ||
2056			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2057			    !timeo ||
2058			    signal_pending(current))
2059				break;
2060		} else {
2061			if (sock_flag(sk, SOCK_DONE))
2062				break;
2063
2064			if (sk->sk_err) {
2065				copied = sock_error(sk);
2066				break;
2067			}
2068
2069			if (sk->sk_shutdown & RCV_SHUTDOWN)
2070				break;
2071
2072			if (sk->sk_state == TCP_CLOSE) {
2073				/* This occurs when user tries to read
2074				 * from never connected socket.
2075				 */
2076				copied = -ENOTCONN;
2077				break;
2078			}
2079
2080			if (!timeo) {
2081				copied = -EAGAIN;
2082				break;
2083			}
2084
2085			if (signal_pending(current)) {
2086				copied = sock_intr_errno(timeo);
2087				break;
2088			}
2089		}
2090
2091		tcp_cleanup_rbuf(sk, copied);
2092
2093		if (copied >= target) {
2094			/* Do not sleep, just process backlog. */
2095			release_sock(sk);
2096			lock_sock(sk);
2097		} else {
2098			sk_wait_data(sk, &timeo, last);
 
 
 
 
 
2099		}
2100
2101		if ((flags & MSG_PEEK) &&
2102		    (peek_seq - copied - urg_hole != tp->copied_seq)) {
2103			net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2104					    current->comm,
2105					    task_pid_nr(current));
2106			peek_seq = tp->copied_seq;
2107		}
2108		continue;
2109
2110found_ok_skb:
2111		/* Ok so how much can we use? */
2112		used = skb->len - offset;
2113		if (len < used)
2114			used = len;
2115
2116		/* Do we have urgent data here? */
2117		if (tp->urg_data) {
2118			u32 urg_offset = tp->urg_seq - *seq;
2119			if (urg_offset < used) {
2120				if (!urg_offset) {
2121					if (!sock_flag(sk, SOCK_URGINLINE)) {
2122						WRITE_ONCE(*seq, *seq + 1);
2123						urg_hole++;
2124						offset++;
2125						used--;
2126						if (!used)
2127							goto skip_copy;
2128					}
2129				} else
2130					used = urg_offset;
2131			}
2132		}
2133
2134		if (!(flags & MSG_TRUNC)) {
2135			err = skb_copy_datagram_msg(skb, offset, msg, used);
2136			if (err) {
2137				/* Exception. Bailout! */
2138				if (!copied)
2139					copied = -EFAULT;
2140				break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2141			}
2142		}
2143
 
 
2144		WRITE_ONCE(*seq, *seq + used);
2145		copied += used;
2146		len -= used;
2147
 
 
 
2148		tcp_rcv_space_adjust(sk);
2149
2150skip_copy:
2151		if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2152			tp->urg_data = 0;
2153			tcp_fast_path_check(sk);
2154		}
 
 
 
 
 
 
2155		if (used + offset < skb->len)
2156			continue;
2157
2158		if (TCP_SKB_CB(skb)->has_rxtstamp) {
2159			tcp_update_recv_tstamps(skb, &tss);
2160			has_tss = true;
2161			has_cmsg = true;
2162		}
2163		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2164			goto found_fin_ok;
2165		if (!(flags & MSG_PEEK))
2166			sk_eat_skb(sk, skb);
2167		continue;
2168
2169found_fin_ok:
2170		/* Process the FIN. */
2171		WRITE_ONCE(*seq, *seq + 1);
2172		if (!(flags & MSG_PEEK))
2173			sk_eat_skb(sk, skb);
2174		break;
2175	} while (len > 0);
2176
2177	/* According to UNIX98, msg_name/msg_namelen are ignored
2178	 * on connected socket. I was just happy when found this 8) --ANK
2179	 */
2180
2181	/* Clean up data we have read: This will do ACK frames. */
2182	tcp_cleanup_rbuf(sk, copied);
2183
2184	release_sock(sk);
2185
2186	if (has_cmsg) {
2187		if (has_tss)
2188			tcp_recv_timestamp(msg, sk, &tss);
2189		if (tp->recvmsg_inq) {
2190			inq = tcp_inq_hint(sk);
2191			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2192		}
2193	}
2194
2195	return copied;
2196
2197out:
2198	release_sock(sk);
2199	return err;
2200
2201recv_urg:
2202	err = tcp_recv_urg(sk, msg, len, flags);
2203	goto out;
2204
2205recv_sndq:
2206	err = tcp_peek_sndq(sk, msg, len);
2207	goto out;
2208}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2209EXPORT_SYMBOL(tcp_recvmsg);
2210
2211void tcp_set_state(struct sock *sk, int state)
2212{
2213	int oldstate = sk->sk_state;
2214
2215	/* We defined a new enum for TCP states that are exported in BPF
2216	 * so as not force the internal TCP states to be frozen. The
2217	 * following checks will detect if an internal state value ever
2218	 * differs from the BPF value. If this ever happens, then we will
2219	 * need to remap the internal value to the BPF value before calling
2220	 * tcp_call_bpf_2arg.
2221	 */
2222	BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2223	BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2224	BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2225	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2226	BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2227	BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2228	BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2229	BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2230	BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2231	BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2232	BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2233	BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
 
2234	BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2235
 
 
 
 
 
 
 
 
 
 
 
2236	if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2237		tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2238
2239	switch (state) {
2240	case TCP_ESTABLISHED:
2241		if (oldstate != TCP_ESTABLISHED)
2242			TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2243		break;
 
 
 
 
2244
2245	case TCP_CLOSE:
2246		if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2247			TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2248
2249		sk->sk_prot->unhash(sk);
2250		if (inet_csk(sk)->icsk_bind_hash &&
2251		    !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2252			inet_put_port(sk);
2253		/* fall through */
2254	default:
2255		if (oldstate == TCP_ESTABLISHED)
2256			TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2257	}
2258
2259	/* Change state AFTER socket is unhashed to avoid closed
2260	 * socket sitting in hash tables.
2261	 */
2262	inet_sk_state_store(sk, state);
2263}
2264EXPORT_SYMBOL_GPL(tcp_set_state);
2265
2266/*
2267 *	State processing on a close. This implements the state shift for
2268 *	sending our FIN frame. Note that we only send a FIN for some
2269 *	states. A shutdown() may have already sent the FIN, or we may be
2270 *	closed.
2271 */
2272
2273static const unsigned char new_state[16] = {
2274  /* current state:        new state:      action:	*/
2275  [0 /* (Invalid) */]	= TCP_CLOSE,
2276  [TCP_ESTABLISHED]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2277  [TCP_SYN_SENT]	= TCP_CLOSE,
2278  [TCP_SYN_RECV]	= TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2279  [TCP_FIN_WAIT1]	= TCP_FIN_WAIT1,
2280  [TCP_FIN_WAIT2]	= TCP_FIN_WAIT2,
2281  [TCP_TIME_WAIT]	= TCP_CLOSE,
2282  [TCP_CLOSE]		= TCP_CLOSE,
2283  [TCP_CLOSE_WAIT]	= TCP_LAST_ACK  | TCP_ACTION_FIN,
2284  [TCP_LAST_ACK]	= TCP_LAST_ACK,
2285  [TCP_LISTEN]		= TCP_CLOSE,
2286  [TCP_CLOSING]		= TCP_CLOSING,
2287  [TCP_NEW_SYN_RECV]	= TCP_CLOSE,	/* should not happen ! */
2288};
2289
2290static int tcp_close_state(struct sock *sk)
2291{
2292	int next = (int)new_state[sk->sk_state];
2293	int ns = next & TCP_STATE_MASK;
2294
2295	tcp_set_state(sk, ns);
2296
2297	return next & TCP_ACTION_FIN;
2298}
2299
2300/*
2301 *	Shutdown the sending side of a connection. Much like close except
2302 *	that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2303 */
2304
2305void tcp_shutdown(struct sock *sk, int how)
2306{
2307	/*	We need to grab some memory, and put together a FIN,
2308	 *	and then put it into the queue to be sent.
2309	 *		Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2310	 */
2311	if (!(how & SEND_SHUTDOWN))
2312		return;
2313
2314	/* If we've already sent a FIN, or it's a closed state, skip this. */
2315	if ((1 << sk->sk_state) &
2316	    (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2317	     TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2318		/* Clear out any half completed packets.  FIN if needed. */
2319		if (tcp_close_state(sk))
2320			tcp_send_fin(sk);
2321	}
2322}
2323EXPORT_SYMBOL(tcp_shutdown);
2324
2325bool tcp_check_oom(struct sock *sk, int shift)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2326{
2327	bool too_many_orphans, out_of_socket_memory;
2328
2329	too_many_orphans = tcp_too_many_orphans(sk, shift);
2330	out_of_socket_memory = tcp_out_of_memory(sk);
2331
2332	if (too_many_orphans)
2333		net_info_ratelimited("too many orphaned sockets\n");
2334	if (out_of_socket_memory)
2335		net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2336	return too_many_orphans || out_of_socket_memory;
2337}
2338
2339void tcp_close(struct sock *sk, long timeout)
2340{
2341	struct sk_buff *skb;
2342	int data_was_unread = 0;
2343	int state;
2344
2345	lock_sock(sk);
2346	sk->sk_shutdown = SHUTDOWN_MASK;
2347
2348	if (sk->sk_state == TCP_LISTEN) {
2349		tcp_set_state(sk, TCP_CLOSE);
2350
2351		/* Special case. */
2352		inet_csk_listen_stop(sk);
2353
2354		goto adjudge_to_death;
2355	}
2356
2357	/*  We need to flush the recv. buffs.  We do this only on the
2358	 *  descriptor close, not protocol-sourced closes, because the
2359	 *  reader process may not have drained the data yet!
2360	 */
2361	while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2362		u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2363
2364		if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2365			len--;
2366		data_was_unread += len;
2367		__kfree_skb(skb);
2368	}
2369
2370	sk_mem_reclaim(sk);
2371
2372	/* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2373	if (sk->sk_state == TCP_CLOSE)
2374		goto adjudge_to_death;
2375
2376	/* As outlined in RFC 2525, section 2.17, we send a RST here because
2377	 * data was lost. To witness the awful effects of the old behavior of
2378	 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2379	 * GET in an FTP client, suspend the process, wait for the client to
2380	 * advertise a zero window, then kill -9 the FTP client, wheee...
2381	 * Note: timeout is always zero in such a case.
2382	 */
2383	if (unlikely(tcp_sk(sk)->repair)) {
2384		sk->sk_prot->disconnect(sk, 0);
2385	} else if (data_was_unread) {
2386		/* Unread data was tossed, zap the connection. */
2387		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2388		tcp_set_state(sk, TCP_CLOSE);
2389		tcp_send_active_reset(sk, sk->sk_allocation);
 
2390	} else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2391		/* Check zero linger _after_ checking for unread data. */
2392		sk->sk_prot->disconnect(sk, 0);
2393		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2394	} else if (tcp_close_state(sk)) {
2395		/* We FIN if the application ate all the data before
2396		 * zapping the connection.
2397		 */
2398
2399		/* RED-PEN. Formally speaking, we have broken TCP state
2400		 * machine. State transitions:
2401		 *
2402		 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2403		 * TCP_SYN_RECV	-> TCP_FIN_WAIT1 (forget it, it's impossible)
2404		 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2405		 *
2406		 * are legal only when FIN has been sent (i.e. in window),
2407		 * rather than queued out of window. Purists blame.
2408		 *
2409		 * F.e. "RFC state" is ESTABLISHED,
2410		 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2411		 *
2412		 * The visible declinations are that sometimes
2413		 * we enter time-wait state, when it is not required really
2414		 * (harmless), do not send active resets, when they are
2415		 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2416		 * they look as CLOSING or LAST_ACK for Linux)
2417		 * Probably, I missed some more holelets.
2418		 * 						--ANK
2419		 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2420		 * in a single packet! (May consider it later but will
2421		 * probably need API support or TCP_CORK SYN-ACK until
2422		 * data is written and socket is closed.)
2423		 */
2424		tcp_send_fin(sk);
2425	}
2426
2427	sk_stream_wait_close(sk, timeout);
2428
2429adjudge_to_death:
2430	state = sk->sk_state;
2431	sock_hold(sk);
2432	sock_orphan(sk);
2433
2434	local_bh_disable();
2435	bh_lock_sock(sk);
2436	/* remove backlog if any, without releasing ownership. */
2437	__release_sock(sk);
2438
2439	percpu_counter_inc(sk->sk_prot->orphan_count);
2440
2441	/* Have we already been destroyed by a softirq or backlog? */
2442	if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2443		goto out;
2444
2445	/*	This is a (useful) BSD violating of the RFC. There is a
2446	 *	problem with TCP as specified in that the other end could
2447	 *	keep a socket open forever with no application left this end.
2448	 *	We use a 1 minute timeout (about the same as BSD) then kill
2449	 *	our end. If they send after that then tough - BUT: long enough
2450	 *	that we won't make the old 4*rto = almost no time - whoops
2451	 *	reset mistake.
2452	 *
2453	 *	Nope, it was not mistake. It is really desired behaviour
2454	 *	f.e. on http servers, when such sockets are useless, but
2455	 *	consume significant resources. Let's do it with special
2456	 *	linger2	option.					--ANK
2457	 */
2458
2459	if (sk->sk_state == TCP_FIN_WAIT2) {
2460		struct tcp_sock *tp = tcp_sk(sk);
2461		if (tp->linger2 < 0) {
2462			tcp_set_state(sk, TCP_CLOSE);
2463			tcp_send_active_reset(sk, GFP_ATOMIC);
 
2464			__NET_INC_STATS(sock_net(sk),
2465					LINUX_MIB_TCPABORTONLINGER);
2466		} else {
2467			const int tmo = tcp_fin_time(sk);
2468
2469			if (tmo > TCP_TIMEWAIT_LEN) {
2470				inet_csk_reset_keepalive_timer(sk,
2471						tmo - TCP_TIMEWAIT_LEN);
2472			} else {
2473				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2474				goto out;
2475			}
2476		}
2477	}
2478	if (sk->sk_state != TCP_CLOSE) {
2479		sk_mem_reclaim(sk);
2480		if (tcp_check_oom(sk, 0)) {
2481			tcp_set_state(sk, TCP_CLOSE);
2482			tcp_send_active_reset(sk, GFP_ATOMIC);
 
2483			__NET_INC_STATS(sock_net(sk),
2484					LINUX_MIB_TCPABORTONMEMORY);
2485		} else if (!check_net(sock_net(sk))) {
2486			/* Not possible to send reset; just close */
2487			tcp_set_state(sk, TCP_CLOSE);
2488		}
2489	}
2490
2491	if (sk->sk_state == TCP_CLOSE) {
2492		struct request_sock *req;
2493
2494		req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2495						lockdep_sock_is_held(sk));
2496		/* We could get here with a non-NULL req if the socket is
2497		 * aborted (e.g., closed with unread data) before 3WHS
2498		 * finishes.
2499		 */
2500		if (req)
2501			reqsk_fastopen_remove(sk, req, false);
2502		inet_csk_destroy_sock(sk);
2503	}
2504	/* Otherwise, socket is reprieved until protocol close. */
2505
2506out:
2507	bh_unlock_sock(sk);
2508	local_bh_enable();
 
 
 
 
 
 
2509	release_sock(sk);
 
 
2510	sock_put(sk);
2511}
2512EXPORT_SYMBOL(tcp_close);
2513
2514/* These states need RST on ABORT according to RFC793 */
2515
2516static inline bool tcp_need_reset(int state)
2517{
2518	return (1 << state) &
2519	       (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2520		TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2521}
2522
2523static void tcp_rtx_queue_purge(struct sock *sk)
2524{
2525	struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2526
 
2527	while (p) {
2528		struct sk_buff *skb = rb_to_skb(p);
2529
2530		p = rb_next(p);
2531		/* Since we are deleting whole queue, no need to
2532		 * list_del(&skb->tcp_tsorted_anchor)
2533		 */
2534		tcp_rtx_queue_unlink(skb, sk);
2535		sk_wmem_free_skb(sk, skb);
2536	}
2537}
2538
2539void tcp_write_queue_purge(struct sock *sk)
2540{
2541	struct sk_buff *skb;
2542
2543	tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2544	while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2545		tcp_skb_tsorted_anchor_cleanup(skb);
2546		sk_wmem_free_skb(sk, skb);
2547	}
2548	tcp_rtx_queue_purge(sk);
2549	skb = sk->sk_tx_skb_cache;
2550	if (skb) {
2551		__kfree_skb(skb);
2552		sk->sk_tx_skb_cache = NULL;
2553	}
2554	INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2555	sk_mem_reclaim(sk);
2556	tcp_clear_all_retrans_hints(tcp_sk(sk));
2557	tcp_sk(sk)->packets_out = 0;
2558	inet_csk(sk)->icsk_backoff = 0;
2559}
2560
2561int tcp_disconnect(struct sock *sk, int flags)
2562{
2563	struct inet_sock *inet = inet_sk(sk);
2564	struct inet_connection_sock *icsk = inet_csk(sk);
2565	struct tcp_sock *tp = tcp_sk(sk);
2566	int old_state = sk->sk_state;
2567	u32 seq;
2568
2569	if (old_state != TCP_CLOSE)
2570		tcp_set_state(sk, TCP_CLOSE);
2571
2572	/* ABORT function of RFC793 */
2573	if (old_state == TCP_LISTEN) {
2574		inet_csk_listen_stop(sk);
2575	} else if (unlikely(tp->repair)) {
2576		sk->sk_err = ECONNABORTED;
2577	} else if (tcp_need_reset(old_state) ||
2578		   (tp->snd_nxt != tp->write_seq &&
2579		    (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
 
 
2580		/* The last check adjusts for discrepancy of Linux wrt. RFC
2581		 * states
2582		 */
2583		tcp_send_active_reset(sk, gfp_any());
2584		sk->sk_err = ECONNRESET;
 
2585	} else if (old_state == TCP_SYN_SENT)
2586		sk->sk_err = ECONNRESET;
2587
2588	tcp_clear_xmit_timers(sk);
2589	__skb_queue_purge(&sk->sk_receive_queue);
2590	if (sk->sk_rx_skb_cache) {
2591		__kfree_skb(sk->sk_rx_skb_cache);
2592		sk->sk_rx_skb_cache = NULL;
2593	}
2594	WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2595	tp->urg_data = 0;
 
2596	tcp_write_queue_purge(sk);
2597	tcp_fastopen_active_disable_ofo_check(sk);
2598	skb_rbtree_purge(&tp->out_of_order_queue);
2599
2600	inet->inet_dport = 0;
2601
2602	if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2603		inet_reset_saddr(sk);
2604
2605	sk->sk_shutdown = 0;
2606	sock_reset_flag(sk, SOCK_DONE);
2607	tp->srtt_us = 0;
2608	tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2609	tp->rcv_rtt_last_tsecr = 0;
2610
2611	seq = tp->write_seq + tp->max_window + 2;
2612	if (!seq)
2613		seq = 1;
2614	WRITE_ONCE(tp->write_seq, seq);
2615
2616	icsk->icsk_backoff = 0;
2617	tp->snd_cwnd = 2;
2618	icsk->icsk_probes_out = 0;
 
2619	icsk->icsk_rto = TCP_TIMEOUT_INIT;
 
 
2620	tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2621	tp->snd_cwnd = TCP_INIT_CWND;
2622	tp->snd_cwnd_cnt = 0;
 
 
2623	tp->window_clamp = 0;
 
2624	tp->delivered_ce = 0;
 
 
 
 
2625	tcp_set_ca_state(sk, TCP_CA_Open);
2626	tp->is_sack_reneg = 0;
2627	tcp_clear_retrans(tp);
 
2628	inet_csk_delack_init(sk);
2629	/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2630	 * issue in __tcp_select_window()
2631	 */
2632	icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
2633	memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2634	__sk_dst_reset(sk);
2635	dst_release(sk->sk_rx_dst);
2636	sk->sk_rx_dst = NULL;
2637	tcp_saved_syn_free(tp);
2638	tp->compressed_ack = 0;
 
 
2639	tp->bytes_sent = 0;
2640	tp->bytes_acked = 0;
2641	tp->bytes_received = 0;
2642	tp->bytes_retrans = 0;
 
 
2643	tp->duplicate_sack[0].start_seq = 0;
2644	tp->duplicate_sack[0].end_seq = 0;
2645	tp->dsack_dups = 0;
2646	tp->reord_seen = 0;
2647	tp->retrans_out = 0;
2648	tp->sacked_out = 0;
2649	tp->tlp_high_seq = 0;
2650	tp->last_oow_ack_time = 0;
 
2651	/* There's a bubble in the pipe until at least the first ACK. */
2652	tp->app_limited = ~0U;
 
2653	tp->rack.mstamp = 0;
2654	tp->rack.advanced = 0;
2655	tp->rack.reo_wnd_steps = 1;
2656	tp->rack.last_delivered = 0;
2657	tp->rack.reo_wnd_persist = 0;
2658	tp->rack.dsack_seen = 0;
2659	tp->syn_data_acked = 0;
2660	tp->rx_opt.saw_tstamp = 0;
2661	tp->rx_opt.dsack = 0;
2662	tp->rx_opt.num_sacks = 0;
2663	tp->rcv_ooopack = 0;
2664
2665
2666	/* Clean up fastopen related fields */
2667	tcp_free_fastopen_req(tp);
2668	inet->defer_connect = 0;
 
2669
2670	WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2671
2672	if (sk->sk_frag.page) {
2673		put_page(sk->sk_frag.page);
2674		sk->sk_frag.page = NULL;
2675		sk->sk_frag.offset = 0;
2676	}
2677
2678	sk->sk_error_report(sk);
2679	return 0;
2680}
2681EXPORT_SYMBOL(tcp_disconnect);
2682
2683static inline bool tcp_can_repair_sock(const struct sock *sk)
2684{
2685	return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2686		(sk->sk_state != TCP_LISTEN);
2687}
2688
2689static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2690{
2691	struct tcp_repair_window opt;
2692
2693	if (!tp->repair)
2694		return -EPERM;
2695
2696	if (len != sizeof(opt))
2697		return -EINVAL;
2698
2699	if (copy_from_user(&opt, optbuf, sizeof(opt)))
2700		return -EFAULT;
2701
2702	if (opt.max_window < opt.snd_wnd)
2703		return -EINVAL;
2704
2705	if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2706		return -EINVAL;
2707
2708	if (after(opt.rcv_wup, tp->rcv_nxt))
2709		return -EINVAL;
2710
2711	tp->snd_wl1	= opt.snd_wl1;
2712	tp->snd_wnd	= opt.snd_wnd;
2713	tp->max_window	= opt.max_window;
2714
2715	tp->rcv_wnd	= opt.rcv_wnd;
2716	tp->rcv_wup	= opt.rcv_wup;
2717
2718	return 0;
2719}
2720
2721static int tcp_repair_options_est(struct sock *sk,
2722		struct tcp_repair_opt __user *optbuf, unsigned int len)
2723{
2724	struct tcp_sock *tp = tcp_sk(sk);
2725	struct tcp_repair_opt opt;
 
2726
2727	while (len >= sizeof(opt)) {
2728		if (copy_from_user(&opt, optbuf, sizeof(opt)))
2729			return -EFAULT;
2730
2731		optbuf++;
2732		len -= sizeof(opt);
2733
2734		switch (opt.opt_code) {
2735		case TCPOPT_MSS:
2736			tp->rx_opt.mss_clamp = opt.opt_val;
2737			tcp_mtup_init(sk);
2738			break;
2739		case TCPOPT_WINDOW:
2740			{
2741				u16 snd_wscale = opt.opt_val & 0xFFFF;
2742				u16 rcv_wscale = opt.opt_val >> 16;
2743
2744				if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
2745					return -EFBIG;
2746
2747				tp->rx_opt.snd_wscale = snd_wscale;
2748				tp->rx_opt.rcv_wscale = rcv_wscale;
2749				tp->rx_opt.wscale_ok = 1;
2750			}
2751			break;
2752		case TCPOPT_SACK_PERM:
2753			if (opt.opt_val != 0)
2754				return -EINVAL;
2755
2756			tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2757			break;
2758		case TCPOPT_TIMESTAMP:
2759			if (opt.opt_val != 0)
2760				return -EINVAL;
2761
2762			tp->rx_opt.tstamp_ok = 1;
2763			break;
2764		}
2765	}
2766
2767	return 0;
2768}
2769
2770DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
2771EXPORT_SYMBOL(tcp_tx_delay_enabled);
2772
2773static void tcp_enable_tx_delay(void)
2774{
2775	if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
2776		static int __tcp_tx_delay_enabled = 0;
2777
2778		if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
2779			static_branch_enable(&tcp_tx_delay_enabled);
2780			pr_info("TCP_TX_DELAY enabled\n");
2781		}
2782	}
2783}
2784
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2785/*
2786 *	Socket option code for TCP.
2787 */
2788static int do_tcp_setsockopt(struct sock *sk, int level,
2789		int optname, char __user *optval, unsigned int optlen)
2790{
2791	struct tcp_sock *tp = tcp_sk(sk);
2792	struct inet_connection_sock *icsk = inet_csk(sk);
2793	struct net *net = sock_net(sk);
2794	int val;
2795	int err = 0;
2796
2797	/* These are data/string values, all the others are ints */
2798	switch (optname) {
2799	case TCP_CONGESTION: {
2800		char name[TCP_CA_NAME_MAX];
2801
2802		if (optlen < 1)
2803			return -EINVAL;
2804
2805		val = strncpy_from_user(name, optval,
2806					min_t(long, TCP_CA_NAME_MAX-1, optlen));
2807		if (val < 0)
2808			return -EFAULT;
2809		name[val] = 0;
2810
2811		lock_sock(sk);
2812		err = tcp_set_congestion_control(sk, name, true, true,
2813						 ns_capable(sock_net(sk)->user_ns,
2814							    CAP_NET_ADMIN));
2815		release_sock(sk);
2816		return err;
2817	}
2818	case TCP_ULP: {
2819		char name[TCP_ULP_NAME_MAX];
2820
2821		if (optlen < 1)
2822			return -EINVAL;
2823
2824		val = strncpy_from_user(name, optval,
2825					min_t(long, TCP_ULP_NAME_MAX - 1,
2826					      optlen));
2827		if (val < 0)
2828			return -EFAULT;
2829		name[val] = 0;
2830
2831		lock_sock(sk);
2832		err = tcp_set_ulp(sk, name);
2833		release_sock(sk);
2834		return err;
2835	}
2836	case TCP_FASTOPEN_KEY: {
2837		__u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
2838		__u8 *backup_key = NULL;
2839
2840		/* Allow a backup key as well to facilitate key rotation
2841		 * First key is the active one.
2842		 */
2843		if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
2844		    optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
2845			return -EINVAL;
2846
2847		if (copy_from_user(key, optval, optlen))
2848			return -EFAULT;
2849
2850		if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
2851			backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
2852
2853		return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
2854	}
2855	default:
2856		/* fallthru */
2857		break;
2858	}
2859
2860	if (optlen < sizeof(int))
2861		return -EINVAL;
2862
2863	if (get_user(val, (int __user *)optval))
2864		return -EFAULT;
2865
2866	lock_sock(sk);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2867
2868	switch (optname) {
2869	case TCP_MAXSEG:
2870		/* Values greater than interface MTU won't take effect. However
2871		 * at the point when this call is done we typically don't yet
2872		 * know which interface is going to be used
2873		 */
2874		if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
2875			err = -EINVAL;
2876			break;
2877		}
2878		tp->rx_opt.user_mss = val;
2879		break;
2880
2881	case TCP_NODELAY:
2882		if (val) {
2883			/* TCP_NODELAY is weaker than TCP_CORK, so that
2884			 * this option on corked socket is remembered, but
2885			 * it is not activated until cork is cleared.
2886			 *
2887			 * However, when TCP_NODELAY is set we make
2888			 * an explicit push, which overrides even TCP_CORK
2889			 * for currently queued segments.
2890			 */
2891			tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2892			tcp_push_pending_frames(sk);
2893		} else {
2894			tp->nonagle &= ~TCP_NAGLE_OFF;
2895		}
2896		break;
2897
2898	case TCP_THIN_LINEAR_TIMEOUTS:
2899		if (val < 0 || val > 1)
2900			err = -EINVAL;
2901		else
2902			tp->thin_lto = val;
2903		break;
2904
2905	case TCP_THIN_DUPACK:
2906		if (val < 0 || val > 1)
2907			err = -EINVAL;
2908		break;
2909
2910	case TCP_REPAIR:
2911		if (!tcp_can_repair_sock(sk))
2912			err = -EPERM;
2913		else if (val == TCP_REPAIR_ON) {
2914			tp->repair = 1;
2915			sk->sk_reuse = SK_FORCE_REUSE;
2916			tp->repair_queue = TCP_NO_QUEUE;
2917		} else if (val == TCP_REPAIR_OFF) {
2918			tp->repair = 0;
2919			sk->sk_reuse = SK_NO_REUSE;
2920			tcp_send_window_probe(sk);
2921		} else if (val == TCP_REPAIR_OFF_NO_WP) {
2922			tp->repair = 0;
2923			sk->sk_reuse = SK_NO_REUSE;
2924		} else
2925			err = -EINVAL;
2926
2927		break;
2928
2929	case TCP_REPAIR_QUEUE:
2930		if (!tp->repair)
2931			err = -EPERM;
2932		else if ((unsigned int)val < TCP_QUEUES_NR)
2933			tp->repair_queue = val;
2934		else
2935			err = -EINVAL;
2936		break;
2937
2938	case TCP_QUEUE_SEQ:
2939		if (sk->sk_state != TCP_CLOSE)
2940			err = -EPERM;
2941		else if (tp->repair_queue == TCP_SEND_QUEUE)
2942			WRITE_ONCE(tp->write_seq, val);
2943		else if (tp->repair_queue == TCP_RECV_QUEUE)
2944			WRITE_ONCE(tp->rcv_nxt, val);
2945		else
 
 
 
 
 
 
 
 
2946			err = -EINVAL;
 
2947		break;
2948
2949	case TCP_REPAIR_OPTIONS:
2950		if (!tp->repair)
2951			err = -EINVAL;
2952		else if (sk->sk_state == TCP_ESTABLISHED)
2953			err = tcp_repair_options_est(sk,
2954					(struct tcp_repair_opt __user *)optval,
2955					optlen);
2956		else
2957			err = -EPERM;
2958		break;
2959
2960	case TCP_CORK:
2961		/* When set indicates to always queue non-full frames.
2962		 * Later the user clears this option and we transmit
2963		 * any pending partial frames in the queue.  This is
2964		 * meant to be used alongside sendfile() to get properly
2965		 * filled frames when the user (for example) must write
2966		 * out headers with a write() call first and then use
2967		 * sendfile to send out the data parts.
2968		 *
2969		 * TCP_CORK can be set together with TCP_NODELAY and it is
2970		 * stronger than TCP_NODELAY.
2971		 */
2972		if (val) {
2973			tp->nonagle |= TCP_NAGLE_CORK;
2974		} else {
2975			tp->nonagle &= ~TCP_NAGLE_CORK;
2976			if (tp->nonagle&TCP_NAGLE_OFF)
2977				tp->nonagle |= TCP_NAGLE_PUSH;
2978			tcp_push_pending_frames(sk);
2979		}
2980		break;
2981
2982	case TCP_KEEPIDLE:
2983		if (val < 1 || val > MAX_TCP_KEEPIDLE)
2984			err = -EINVAL;
2985		else {
2986			tp->keepalive_time = val * HZ;
2987			if (sock_flag(sk, SOCK_KEEPOPEN) &&
2988			    !((1 << sk->sk_state) &
2989			      (TCPF_CLOSE | TCPF_LISTEN))) {
2990				u32 elapsed = keepalive_time_elapsed(tp);
2991				if (tp->keepalive_time > elapsed)
2992					elapsed = tp->keepalive_time - elapsed;
2993				else
2994					elapsed = 0;
2995				inet_csk_reset_keepalive_timer(sk, elapsed);
2996			}
2997		}
2998		break;
2999	case TCP_KEEPINTVL:
3000		if (val < 1 || val > MAX_TCP_KEEPINTVL)
3001			err = -EINVAL;
3002		else
3003			tp->keepalive_intvl = val * HZ;
3004		break;
3005	case TCP_KEEPCNT:
3006		if (val < 1 || val > MAX_TCP_KEEPCNT)
3007			err = -EINVAL;
3008		else
3009			tp->keepalive_probes = val;
3010		break;
3011	case TCP_SYNCNT:
3012		if (val < 1 || val > MAX_TCP_SYNCNT)
3013			err = -EINVAL;
3014		else
3015			icsk->icsk_syn_retries = val;
3016		break;
3017
3018	case TCP_SAVE_SYN:
3019		if (val < 0 || val > 1)
 
3020			err = -EINVAL;
3021		else
3022			tp->save_syn = val;
3023		break;
3024
3025	case TCP_LINGER2:
3026		if (val < 0)
3027			tp->linger2 = -1;
3028		else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
3029			tp->linger2 = 0;
3030		else
3031			tp->linger2 = val * HZ;
3032		break;
3033
3034	case TCP_DEFER_ACCEPT:
3035		/* Translate value in seconds to number of retransmits */
3036		icsk->icsk_accept_queue.rskq_defer_accept =
3037			secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3038					TCP_RTO_MAX / HZ);
3039		break;
3040
3041	case TCP_WINDOW_CLAMP:
3042		if (!val) {
3043			if (sk->sk_state != TCP_CLOSE) {
3044				err = -EINVAL;
3045				break;
3046			}
3047			tp->window_clamp = 0;
3048		} else
3049			tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3050						SOCK_MIN_RCVBUF / 2 : val;
 
 
 
 
 
 
 
 
 
 
 
 
 
3051		break;
3052
3053	case TCP_QUICKACK:
3054		if (!val) {
3055			inet_csk_enter_pingpong_mode(sk);
3056		} else {
3057			inet_csk_exit_pingpong_mode(sk);
3058			if ((1 << sk->sk_state) &
3059			    (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3060			    inet_csk_ack_scheduled(sk)) {
3061				icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
3062				tcp_cleanup_rbuf(sk, 1);
3063				if (!(val & 1))
3064					inet_csk_enter_pingpong_mode(sk);
3065			}
3066		}
3067		break;
3068
 
3069#ifdef CONFIG_TCP_MD5SIG
3070	case TCP_MD5SIG:
3071	case TCP_MD5SIG_EXT:
3072		if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
3073			err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3074		else
3075			err = -EINVAL;
3076		break;
3077#endif
3078	case TCP_USER_TIMEOUT:
3079		/* Cap the max time in ms TCP will retry or probe the window
3080		 * before giving up and aborting (ETIMEDOUT) a connection.
3081		 */
3082		if (val < 0)
3083			err = -EINVAL;
3084		else
3085			icsk->icsk_user_timeout = val;
3086		break;
3087
3088	case TCP_FASTOPEN:
3089		if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3090		    TCPF_LISTEN))) {
3091			tcp_fastopen_init_key_once(net);
3092
3093			fastopen_queue_tune(sk, val);
3094		} else {
3095			err = -EINVAL;
3096		}
3097		break;
3098	case TCP_FASTOPEN_CONNECT:
3099		if (val > 1 || val < 0) {
3100			err = -EINVAL;
3101		} else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
 
3102			if (sk->sk_state == TCP_CLOSE)
3103				tp->fastopen_connect = val;
3104			else
3105				err = -EINVAL;
3106		} else {
3107			err = -EOPNOTSUPP;
3108		}
3109		break;
3110	case TCP_FASTOPEN_NO_COOKIE:
3111		if (val > 1 || val < 0)
3112			err = -EINVAL;
3113		else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3114			err = -EINVAL;
3115		else
3116			tp->fastopen_no_cookie = val;
3117		break;
3118	case TCP_TIMESTAMP:
3119		if (!tp->repair)
3120			err = -EPERM;
3121		else
3122			tp->tsoffset = val - tcp_time_stamp_raw();
 
 
 
 
 
 
3123		break;
3124	case TCP_REPAIR_WINDOW:
3125		err = tcp_repair_set_window(tp, optval, optlen);
3126		break;
3127	case TCP_NOTSENT_LOWAT:
3128		tp->notsent_lowat = val;
3129		sk->sk_write_space(sk);
3130		break;
3131	case TCP_INQ:
3132		if (val > 1 || val < 0)
3133			err = -EINVAL;
3134		else
3135			tp->recvmsg_inq = val;
3136		break;
3137	case TCP_TX_DELAY:
3138		if (val)
3139			tcp_enable_tx_delay();
3140		tp->tcp_tx_delay = val;
3141		break;
3142	default:
3143		err = -ENOPROTOOPT;
3144		break;
3145	}
3146
3147	release_sock(sk);
3148	return err;
3149}
3150
3151int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
3152		   unsigned int optlen)
3153{
3154	const struct inet_connection_sock *icsk = inet_csk(sk);
3155
3156	if (level != SOL_TCP)
3157		return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3158						     optval, optlen);
 
3159	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3160}
3161EXPORT_SYMBOL(tcp_setsockopt);
3162
3163#ifdef CONFIG_COMPAT
3164int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
3165			  char __user *optval, unsigned int optlen)
3166{
3167	if (level != SOL_TCP)
3168		return inet_csk_compat_setsockopt(sk, level, optname,
3169						  optval, optlen);
3170	return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3171}
3172EXPORT_SYMBOL(compat_tcp_setsockopt);
3173#endif
3174
3175static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3176				      struct tcp_info *info)
3177{
3178	u64 stats[__TCP_CHRONO_MAX], total = 0;
3179	enum tcp_chrono i;
3180
3181	for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3182		stats[i] = tp->chrono_stat[i - 1];
3183		if (i == tp->chrono_type)
3184			stats[i] += tcp_jiffies32 - tp->chrono_start;
3185		stats[i] *= USEC_PER_SEC / HZ;
3186		total += stats[i];
3187	}
3188
3189	info->tcpi_busy_time = total;
3190	info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3191	info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3192}
3193
3194/* Return information about state of tcp endpoint in API format. */
3195void tcp_get_info(struct sock *sk, struct tcp_info *info)
3196{
3197	const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3198	const struct inet_connection_sock *icsk = inet_csk(sk);
3199	unsigned long rate;
3200	u32 now;
3201	u64 rate64;
3202	bool slow;
3203
3204	memset(info, 0, sizeof(*info));
3205	if (sk->sk_type != SOCK_STREAM)
3206		return;
3207
3208	info->tcpi_state = inet_sk_state_load(sk);
3209
3210	/* Report meaningful fields for all TCP states, including listeners */
3211	rate = READ_ONCE(sk->sk_pacing_rate);
3212	rate64 = (rate != ~0UL) ? rate : ~0ULL;
3213	info->tcpi_pacing_rate = rate64;
3214
3215	rate = READ_ONCE(sk->sk_max_pacing_rate);
3216	rate64 = (rate != ~0UL) ? rate : ~0ULL;
3217	info->tcpi_max_pacing_rate = rate64;
3218
3219	info->tcpi_reordering = tp->reordering;
3220	info->tcpi_snd_cwnd = tp->snd_cwnd;
3221
3222	if (info->tcpi_state == TCP_LISTEN) {
3223		/* listeners aliased fields :
3224		 * tcpi_unacked -> Number of children ready for accept()
3225		 * tcpi_sacked  -> max backlog
3226		 */
3227		info->tcpi_unacked = sk->sk_ack_backlog;
3228		info->tcpi_sacked = sk->sk_max_ack_backlog;
3229		return;
3230	}
3231
3232	slow = lock_sock_fast(sk);
3233
3234	info->tcpi_ca_state = icsk->icsk_ca_state;
3235	info->tcpi_retransmits = icsk->icsk_retransmits;
3236	info->tcpi_probes = icsk->icsk_probes_out;
3237	info->tcpi_backoff = icsk->icsk_backoff;
3238
3239	if (tp->rx_opt.tstamp_ok)
3240		info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3241	if (tcp_is_sack(tp))
3242		info->tcpi_options |= TCPI_OPT_SACK;
3243	if (tp->rx_opt.wscale_ok) {
3244		info->tcpi_options |= TCPI_OPT_WSCALE;
3245		info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3246		info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3247	}
3248
3249	if (tp->ecn_flags & TCP_ECN_OK)
3250		info->tcpi_options |= TCPI_OPT_ECN;
3251	if (tp->ecn_flags & TCP_ECN_SEEN)
3252		info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3253	if (tp->syn_data_acked)
3254		info->tcpi_options |= TCPI_OPT_SYN_DATA;
 
 
3255
3256	info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3257	info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
 
3258	info->tcpi_snd_mss = tp->mss_cache;
3259	info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3260
3261	info->tcpi_unacked = tp->packets_out;
3262	info->tcpi_sacked = tp->sacked_out;
3263
3264	info->tcpi_lost = tp->lost_out;
3265	info->tcpi_retrans = tp->retrans_out;
3266
3267	now = tcp_jiffies32;
3268	info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3269	info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3270	info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3271
3272	info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3273	info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3274	info->tcpi_rtt = tp->srtt_us >> 3;
3275	info->tcpi_rttvar = tp->mdev_us >> 2;
3276	info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3277	info->tcpi_advmss = tp->advmss;
3278
3279	info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3280	info->tcpi_rcv_space = tp->rcvq_space.space;
3281
3282	info->tcpi_total_retrans = tp->total_retrans;
3283
3284	info->tcpi_bytes_acked = tp->bytes_acked;
3285	info->tcpi_bytes_received = tp->bytes_received;
3286	info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3287	tcp_get_info_chrono_stats(tp, info);
3288
3289	info->tcpi_segs_out = tp->segs_out;
3290	info->tcpi_segs_in = tp->segs_in;
 
 
 
3291
3292	info->tcpi_min_rtt = tcp_min_rtt(tp);
3293	info->tcpi_data_segs_in = tp->data_segs_in;
3294	info->tcpi_data_segs_out = tp->data_segs_out;
3295
3296	info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3297	rate64 = tcp_compute_delivery_rate(tp);
3298	if (rate64)
3299		info->tcpi_delivery_rate = rate64;
3300	info->tcpi_delivered = tp->delivered;
3301	info->tcpi_delivered_ce = tp->delivered_ce;
3302	info->tcpi_bytes_sent = tp->bytes_sent;
3303	info->tcpi_bytes_retrans = tp->bytes_retrans;
3304	info->tcpi_dsack_dups = tp->dsack_dups;
3305	info->tcpi_reord_seen = tp->reord_seen;
3306	info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3307	info->tcpi_snd_wnd = tp->snd_wnd;
 
 
 
 
 
 
 
 
 
 
3308	unlock_sock_fast(sk, slow);
3309}
3310EXPORT_SYMBOL_GPL(tcp_get_info);
3311
3312static size_t tcp_opt_stats_get_size(void)
3313{
3314	return
3315		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3316		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3317		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3318		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3319		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3320		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3321		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3322		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3323		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3324		nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3325		nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3326		nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3327		nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3328		nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3329		nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3330		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3331		nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3332		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3333		nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3334		nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3335		nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3336		nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
 
 
 
 
 
3337		0;
3338}
3339
3340struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
 
 
 
 
 
 
 
 
 
 
 
 
 
3341{
3342	const struct tcp_sock *tp = tcp_sk(sk);
3343	struct sk_buff *stats;
3344	struct tcp_info info;
3345	unsigned long rate;
3346	u64 rate64;
3347
3348	stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3349	if (!stats)
3350		return NULL;
3351
3352	tcp_get_info_chrono_stats(tp, &info);
3353	nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3354			  info.tcpi_busy_time, TCP_NLA_PAD);
3355	nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3356			  info.tcpi_rwnd_limited, TCP_NLA_PAD);
3357	nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3358			  info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3359	nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3360			  tp->data_segs_out, TCP_NLA_PAD);
3361	nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3362			  tp->total_retrans, TCP_NLA_PAD);
3363
3364	rate = READ_ONCE(sk->sk_pacing_rate);
3365	rate64 = (rate != ~0UL) ? rate : ~0ULL;
3366	nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3367
3368	rate64 = tcp_compute_delivery_rate(tp);
3369	nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3370
3371	nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3372	nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3373	nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3374
3375	nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3376	nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3377	nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3378	nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3379	nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3380
3381	nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3382	nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3383
3384	nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3385			  TCP_NLA_PAD);
3386	nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3387			  TCP_NLA_PAD);
3388	nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3389	nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3390	nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
 
 
 
 
 
 
 
 
3391
 
3392	return stats;
3393}
3394
3395static int do_tcp_getsockopt(struct sock *sk, int level,
3396		int optname, char __user *optval, int __user *optlen)
3397{
3398	struct inet_connection_sock *icsk = inet_csk(sk);
3399	struct tcp_sock *tp = tcp_sk(sk);
3400	struct net *net = sock_net(sk);
3401	int val, len;
3402
3403	if (get_user(len, optlen))
3404		return -EFAULT;
3405
3406	len = min_t(unsigned int, len, sizeof(int));
3407
3408	if (len < 0)
3409		return -EINVAL;
3410
 
 
3411	switch (optname) {
3412	case TCP_MAXSEG:
3413		val = tp->mss_cache;
3414		if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
 
3415			val = tp->rx_opt.user_mss;
3416		if (tp->repair)
3417			val = tp->rx_opt.mss_clamp;
3418		break;
3419	case TCP_NODELAY:
3420		val = !!(tp->nonagle&TCP_NAGLE_OFF);
3421		break;
3422	case TCP_CORK:
3423		val = !!(tp->nonagle&TCP_NAGLE_CORK);
3424		break;
3425	case TCP_KEEPIDLE:
3426		val = keepalive_time_when(tp) / HZ;
3427		break;
3428	case TCP_KEEPINTVL:
3429		val = keepalive_intvl_when(tp) / HZ;
3430		break;
3431	case TCP_KEEPCNT:
3432		val = keepalive_probes(tp);
3433		break;
3434	case TCP_SYNCNT:
3435		val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
 
3436		break;
3437	case TCP_LINGER2:
3438		val = tp->linger2;
3439		if (val >= 0)
3440			val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3441		break;
3442	case TCP_DEFER_ACCEPT:
3443		val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3444				      TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
 
3445		break;
3446	case TCP_WINDOW_CLAMP:
3447		val = tp->window_clamp;
3448		break;
3449	case TCP_INFO: {
3450		struct tcp_info info;
3451
3452		if (get_user(len, optlen))
3453			return -EFAULT;
3454
3455		tcp_get_info(sk, &info);
3456
3457		len = min_t(unsigned int, len, sizeof(info));
3458		if (put_user(len, optlen))
3459			return -EFAULT;
3460		if (copy_to_user(optval, &info, len))
3461			return -EFAULT;
3462		return 0;
3463	}
3464	case TCP_CC_INFO: {
3465		const struct tcp_congestion_ops *ca_ops;
3466		union tcp_cc_info info;
3467		size_t sz = 0;
3468		int attr;
3469
3470		if (get_user(len, optlen))
3471			return -EFAULT;
3472
3473		ca_ops = icsk->icsk_ca_ops;
3474		if (ca_ops && ca_ops->get_info)
3475			sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3476
3477		len = min_t(unsigned int, len, sz);
3478		if (put_user(len, optlen))
3479			return -EFAULT;
3480		if (copy_to_user(optval, &info, len))
3481			return -EFAULT;
3482		return 0;
3483	}
3484	case TCP_QUICKACK:
3485		val = !inet_csk_in_pingpong_mode(sk);
3486		break;
3487
3488	case TCP_CONGESTION:
3489		if (get_user(len, optlen))
3490			return -EFAULT;
3491		len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
3492		if (put_user(len, optlen))
3493			return -EFAULT;
3494		if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
3495			return -EFAULT;
3496		return 0;
3497
3498	case TCP_ULP:
3499		if (get_user(len, optlen))
3500			return -EFAULT;
3501		len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
3502		if (!icsk->icsk_ulp_ops) {
3503			if (put_user(0, optlen))
 
3504				return -EFAULT;
3505			return 0;
3506		}
3507		if (put_user(len, optlen))
3508			return -EFAULT;
3509		if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
3510			return -EFAULT;
3511		return 0;
3512
3513	case TCP_FASTOPEN_KEY: {
3514		__u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3515		struct tcp_fastopen_context *ctx;
3516		unsigned int key_len = 0;
3517
3518		if (get_user(len, optlen))
3519			return -EFAULT;
3520
3521		rcu_read_lock();
3522		ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
3523		if (ctx) {
3524			key_len = tcp_fastopen_context_len(ctx) *
3525					TCP_FASTOPEN_KEY_LENGTH;
3526			memcpy(&key[0], &ctx->key[0], key_len);
3527		}
3528		rcu_read_unlock();
3529
3530		len = min_t(unsigned int, len, key_len);
3531		if (put_user(len, optlen))
3532			return -EFAULT;
3533		if (copy_to_user(optval, key, len))
3534			return -EFAULT;
3535		return 0;
3536	}
3537	case TCP_THIN_LINEAR_TIMEOUTS:
3538		val = tp->thin_lto;
3539		break;
3540
3541	case TCP_THIN_DUPACK:
3542		val = 0;
3543		break;
3544
3545	case TCP_REPAIR:
3546		val = tp->repair;
3547		break;
3548
3549	case TCP_REPAIR_QUEUE:
3550		if (tp->repair)
3551			val = tp->repair_queue;
3552		else
3553			return -EINVAL;
3554		break;
3555
3556	case TCP_REPAIR_WINDOW: {
3557		struct tcp_repair_window opt;
3558
3559		if (get_user(len, optlen))
3560			return -EFAULT;
3561
3562		if (len != sizeof(opt))
3563			return -EINVAL;
3564
3565		if (!tp->repair)
3566			return -EPERM;
3567
3568		opt.snd_wl1	= tp->snd_wl1;
3569		opt.snd_wnd	= tp->snd_wnd;
3570		opt.max_window	= tp->max_window;
3571		opt.rcv_wnd	= tp->rcv_wnd;
3572		opt.rcv_wup	= tp->rcv_wup;
3573
3574		if (copy_to_user(optval, &opt, len))
3575			return -EFAULT;
3576		return 0;
3577	}
3578	case TCP_QUEUE_SEQ:
3579		if (tp->repair_queue == TCP_SEND_QUEUE)
3580			val = tp->write_seq;
3581		else if (tp->repair_queue == TCP_RECV_QUEUE)
3582			val = tp->rcv_nxt;
3583		else
3584			return -EINVAL;
3585		break;
3586
3587	case TCP_USER_TIMEOUT:
3588		val = icsk->icsk_user_timeout;
3589		break;
3590
3591	case TCP_FASTOPEN:
3592		val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3593		break;
3594
3595	case TCP_FASTOPEN_CONNECT:
3596		val = tp->fastopen_connect;
3597		break;
3598
3599	case TCP_FASTOPEN_NO_COOKIE:
3600		val = tp->fastopen_no_cookie;
3601		break;
3602
3603	case TCP_TX_DELAY:
3604		val = tp->tcp_tx_delay;
3605		break;
3606
3607	case TCP_TIMESTAMP:
3608		val = tcp_time_stamp_raw() + tp->tsoffset;
 
 
 
 
3609		break;
3610	case TCP_NOTSENT_LOWAT:
3611		val = tp->notsent_lowat;
3612		break;
3613	case TCP_INQ:
3614		val = tp->recvmsg_inq;
3615		break;
3616	case TCP_SAVE_SYN:
3617		val = tp->save_syn;
3618		break;
3619	case TCP_SAVED_SYN: {
3620		if (get_user(len, optlen))
3621			return -EFAULT;
3622
3623		lock_sock(sk);
3624		if (tp->saved_syn) {
3625			if (len < tp->saved_syn[0]) {
3626				if (put_user(tp->saved_syn[0], optlen)) {
3627					release_sock(sk);
 
3628					return -EFAULT;
3629				}
3630				release_sock(sk);
3631				return -EINVAL;
3632			}
3633			len = tp->saved_syn[0];
3634			if (put_user(len, optlen)) {
3635				release_sock(sk);
3636				return -EFAULT;
3637			}
3638			if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3639				release_sock(sk);
3640				return -EFAULT;
3641			}
3642			tcp_saved_syn_free(tp);
3643			release_sock(sk);
3644		} else {
3645			release_sock(sk);
3646			len = 0;
3647			if (put_user(len, optlen))
3648				return -EFAULT;
3649		}
3650		return 0;
3651	}
3652#ifdef CONFIG_MMU
3653	case TCP_ZEROCOPY_RECEIVE: {
3654		struct tcp_zerocopy_receive zc;
 
3655		int err;
3656
3657		if (get_user(len, optlen))
3658			return -EFAULT;
3659		if (len != sizeof(zc))
 
3660			return -EINVAL;
3661		if (copy_from_user(&zc, optval, len))
 
 
 
 
 
 
 
 
 
3662			return -EFAULT;
3663		lock_sock(sk);
3664		err = tcp_zerocopy_receive(sk, &zc);
3665		release_sock(sk);
3666		if (!err && copy_to_user(optval, &zc, len))
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3667			err = -EFAULT;
3668		return err;
3669	}
3670#endif
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3671	default:
3672		return -ENOPROTOOPT;
3673	}
3674
3675	if (put_user(len, optlen))
3676		return -EFAULT;
3677	if (copy_to_user(optval, &val, len))
3678		return -EFAULT;
3679	return 0;
3680}
3681
 
 
 
 
 
 
 
 
 
 
 
 
3682int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3683		   int __user *optlen)
3684{
3685	struct inet_connection_sock *icsk = inet_csk(sk);
3686
3687	if (level != SOL_TCP)
3688		return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3689						     optval, optlen);
3690	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
 
 
3691}
3692EXPORT_SYMBOL(tcp_getsockopt);
3693
3694#ifdef CONFIG_COMPAT
3695int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3696			  char __user *optval, int __user *optlen)
 
 
3697{
3698	if (level != SOL_TCP)
3699		return inet_csk_compat_getsockopt(sk, level, optname,
3700						  optval, optlen);
3701	return do_tcp_getsockopt(sk, level, optname, optval, optlen);
 
 
 
 
 
 
 
 
 
 
 
 
3702}
3703EXPORT_SYMBOL(compat_tcp_getsockopt);
3704#endif
3705
3706#ifdef CONFIG_TCP_MD5SIG
3707static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3708static DEFINE_MUTEX(tcp_md5sig_mutex);
3709static bool tcp_md5sig_pool_populated = false;
3710
3711static void __tcp_alloc_md5sig_pool(void)
3712{
3713	struct crypto_ahash *hash;
3714	int cpu;
3715
3716	hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3717	if (IS_ERR(hash))
3718		return;
3719
3720	for_each_possible_cpu(cpu) {
3721		void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3722		struct ahash_request *req;
3723
3724		if (!scratch) {
3725			scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3726					       sizeof(struct tcphdr),
3727					       GFP_KERNEL,
3728					       cpu_to_node(cpu));
3729			if (!scratch)
3730				return;
3731			per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3732		}
3733		if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3734			continue;
3735
3736		req = ahash_request_alloc(hash, GFP_KERNEL);
3737		if (!req)
3738			return;
3739
3740		ahash_request_set_callback(req, 0, NULL, NULL);
 
3741
3742		per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3743	}
3744	/* before setting tcp_md5sig_pool_populated, we must commit all writes
3745	 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3746	 */
3747	smp_wmb();
3748	tcp_md5sig_pool_populated = true;
3749}
 
3750
3751bool tcp_alloc_md5sig_pool(void)
3752{
3753	if (unlikely(!tcp_md5sig_pool_populated)) {
3754		mutex_lock(&tcp_md5sig_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
3755
3756		if (!tcp_md5sig_pool_populated) {
3757			__tcp_alloc_md5sig_pool();
3758			if (tcp_md5sig_pool_populated)
3759				static_branch_inc(&tcp_md5_needed);
3760		}
3761
3762		mutex_unlock(&tcp_md5sig_mutex);
 
 
 
3763	}
3764	return tcp_md5sig_pool_populated;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3765}
3766EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3767
 
3768
3769/**
3770 *	tcp_get_md5sig_pool - get md5sig_pool for this user
3771 *
3772 *	We use percpu structure, so if we succeed, we exit with preemption
3773 *	and BH disabled, to make sure another thread or softirq handling
3774 *	wont try to get same context.
3775 */
3776struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3777{
3778	local_bh_disable();
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3779
3780	if (tcp_md5sig_pool_populated) {
3781		/* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3782		smp_rmb();
3783		return this_cpu_ptr(&tcp_md5sig_pool);
3784	}
3785	local_bh_enable();
3786	return NULL;
3787}
3788EXPORT_SYMBOL(tcp_get_md5sig_pool);
3789
3790int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3791			  const struct sk_buff *skb, unsigned int header_len)
3792{
3793	struct scatterlist sg;
3794	const struct tcphdr *tp = tcp_hdr(skb);
3795	struct ahash_request *req = hp->md5_req;
3796	unsigned int i;
3797	const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3798					   skb_headlen(skb) - header_len : 0;
3799	const struct skb_shared_info *shi = skb_shinfo(skb);
3800	struct sk_buff *frag_iter;
3801
3802	sg_init_table(&sg, 1);
3803
3804	sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3805	ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3806	if (crypto_ahash_update(req))
3807		return 1;
3808
3809	for (i = 0; i < shi->nr_frags; ++i) {
3810		const skb_frag_t *f = &shi->frags[i];
3811		unsigned int offset = skb_frag_off(f);
3812		struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3813
3814		sg_set_page(&sg, page, skb_frag_size(f),
3815			    offset_in_page(offset));
3816		ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3817		if (crypto_ahash_update(req))
3818			return 1;
3819	}
3820
3821	skb_walk_frags(skb, frag_iter)
3822		if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3823			return 1;
3824
3825	return 0;
3826}
3827EXPORT_SYMBOL(tcp_md5_hash_skb_data);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3828
3829int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3830{
3831	struct scatterlist sg;
3832
3833	sg_init_one(&sg, key->key, key->keylen);
3834	ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3835	return crypto_ahash_update(hp->md5_req);
3836}
3837EXPORT_SYMBOL(tcp_md5_hash_key);
3838
3839#endif
3840
3841void tcp_done(struct sock *sk)
3842{
3843	struct request_sock *req;
3844
3845	/* We might be called with a new socket, after
3846	 * inet_csk_prepare_forced_close() has been called
3847	 * so we can not use lockdep_sock_is_held(sk)
3848	 */
3849	req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
3850
3851	if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3852		TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3853
3854	tcp_set_state(sk, TCP_CLOSE);
3855	tcp_clear_xmit_timers(sk);
3856	if (req)
3857		reqsk_fastopen_remove(sk, req, false);
3858
3859	sk->sk_shutdown = SHUTDOWN_MASK;
3860
3861	if (!sock_flag(sk, SOCK_DEAD))
3862		sk->sk_state_change(sk);
3863	else
3864		inet_csk_destroy_sock(sk);
3865}
3866EXPORT_SYMBOL_GPL(tcp_done);
3867
3868int tcp_abort(struct sock *sk, int err)
3869{
3870	if (!sk_fullsock(sk)) {
3871		if (sk->sk_state == TCP_NEW_SYN_RECV) {
3872			struct request_sock *req = inet_reqsk(sk);
3873
3874			local_bh_disable();
3875			inet_csk_reqsk_queue_drop(req->rsk_listener, req);
3876			local_bh_enable();
3877			return 0;
3878		}
3879		return -EOPNOTSUPP;
 
 
 
 
 
 
 
 
3880	}
3881
3882	/* Don't race with userspace socket closes such as tcp_close. */
3883	lock_sock(sk);
 
 
 
 
 
 
 
 
 
3884
3885	if (sk->sk_state == TCP_LISTEN) {
3886		tcp_set_state(sk, TCP_CLOSE);
3887		inet_csk_listen_stop(sk);
3888	}
3889
3890	/* Don't race with BH socket closes such as inet_csk_listen_stop. */
3891	local_bh_disable();
3892	bh_lock_sock(sk);
3893
3894	if (!sock_flag(sk, SOCK_DEAD)) {
3895		sk->sk_err = err;
3896		/* This barrier is coupled with smp_rmb() in tcp_poll() */
3897		smp_wmb();
3898		sk->sk_error_report(sk);
3899		if (tcp_need_reset(sk->sk_state))
3900			tcp_send_active_reset(sk, GFP_ATOMIC);
3901		tcp_done(sk);
3902	}
3903
3904	bh_unlock_sock(sk);
3905	local_bh_enable();
3906	tcp_write_queue_purge(sk);
3907	release_sock(sk);
3908	return 0;
3909}
3910EXPORT_SYMBOL_GPL(tcp_abort);
3911
3912extern struct tcp_congestion_ops tcp_reno;
3913
3914static __initdata unsigned long thash_entries;
3915static int __init set_thash_entries(char *str)
3916{
3917	ssize_t ret;
3918
3919	if (!str)
3920		return 0;
3921
3922	ret = kstrtoul(str, 0, &thash_entries);
3923	if (ret)
3924		return 0;
3925
3926	return 1;
3927}
3928__setup("thash_entries=", set_thash_entries);
3929
3930static void __init tcp_init_mem(void)
3931{
3932	unsigned long limit = nr_free_buffer_pages() / 16;
3933
3934	limit = max(limit, 128UL);
3935	sysctl_tcp_mem[0] = limit / 4 * 3;		/* 4.68 % */
3936	sysctl_tcp_mem[1] = limit;			/* 6.25 % */
3937	sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2;	/* 9.37 % */
3938}
3939
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3940void __init tcp_init(void)
3941{
3942	int max_rshare, max_wshare, cnt;
3943	unsigned long limit;
3944	unsigned int i;
3945
3946	BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
3947	BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3948		     FIELD_SIZEOF(struct sk_buff, cb));
 
 
3949
3950	percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3951	percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3952	inet_hashinfo_init(&tcp_hashinfo);
 
 
3953	inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
3954			    thash_entries, 21,  /* one slot per 2 MB*/
3955			    0, 64 * 1024);
3956	tcp_hashinfo.bind_bucket_cachep =
3957		kmem_cache_create("tcp_bind_bucket",
3958				  sizeof(struct inet_bind_bucket), 0,
3959				  SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
 
 
 
 
 
 
 
 
3960
3961	/* Size and allocate the main established and bind bucket
3962	 * hash tables.
3963	 *
3964	 * The methodology is similar to that of the buffer cache.
3965	 */
3966	tcp_hashinfo.ehash =
3967		alloc_large_system_hash("TCP established",
3968					sizeof(struct inet_ehash_bucket),
3969					thash_entries,
3970					17, /* one slot per 128 KB of memory */
3971					0,
3972					NULL,
3973					&tcp_hashinfo.ehash_mask,
3974					0,
3975					thash_entries ? 0 : 512 * 1024);
3976	for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3977		INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3978
3979	if (inet_ehash_locks_alloc(&tcp_hashinfo))
3980		panic("TCP: failed to alloc ehash_locks");
3981	tcp_hashinfo.bhash =
3982		alloc_large_system_hash("TCP bind",
3983					sizeof(struct inet_bind_hashbucket),
3984					tcp_hashinfo.ehash_mask + 1,
3985					17, /* one slot per 128 KB of memory */
3986					0,
3987					&tcp_hashinfo.bhash_size,
3988					NULL,
3989					0,
3990					64 * 1024);
3991	tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
 
3992	for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3993		spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3994		INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
 
 
3995	}
3996
 
3997
3998	cnt = tcp_hashinfo.ehash_mask + 1;
3999	sysctl_tcp_max_orphans = cnt / 2;
4000
4001	tcp_init_mem();
4002	/* Set per-socket limits to no more than 1/128 the pressure threshold */
4003	limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4004	max_wshare = min(4UL*1024*1024, limit);
4005	max_rshare = min(6UL*1024*1024, limit);
4006
4007	init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4008	init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4009	init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4010
4011	init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4012	init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4013	init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4014
4015	pr_info("Hash tables configured (established %u bind %u)\n",
4016		tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4017
4018	tcp_v4_init();
4019	tcp_metrics_init();
4020	BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4021	tcp_tasklet_init();
 
4022}