1// SPDX-License-Identifier: GPL-2.0
   2/* Multipath TCP
   3 *
   4 * Copyright (c) 2017 - 2019, Intel Corporation.
   5 */
   6
   7#define pr_fmt(fmt) "MPTCP: " fmt
   8
   9#include <linux/kernel.h>
  10#include <linux/module.h>
  11#include <linux/netdevice.h>
  12#include <linux/sched/signal.h>
  13#include <linux/atomic.h>
  14#include <net/sock.h>
  15#include <net/inet_common.h>
  16#include <net/inet_hashtables.h>
  17#include <net/protocol.h>
  18#include <net/tcp_states.h>
  19#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  20#include <net/transp_v6.h>
  21#endif
  22#include <net/mptcp.h>
  23#include <net/xfrm.h>
  24#include <asm/ioctls.h>
  25#include "protocol.h"
  26#include "mib.h"
  27
  28#define CREATE_TRACE_POINTS
  29#include <trace/events/mptcp.h>
  30
  31#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  32struct mptcp6_sock {
  33	struct mptcp_sock msk;
  34	struct ipv6_pinfo np;
  35};
  36#endif
  37
  38enum {
  39	MPTCP_CMSG_TS = BIT(0),
  40	MPTCP_CMSG_INQ = BIT(1),
  41};
  42
  43static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
  44
  45static void __mptcp_destroy_sock(struct sock *sk);
  46static void mptcp_check_send_data_fin(struct sock *sk);
  47
  48DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
  49static struct net_device mptcp_napi_dev;
  50
  51/* Returns end sequence number of the receiver's advertised window */
  52static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
  53{
  54	return READ_ONCE(msk->wnd_end);
  55}
  56
  57static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
  58{
  59#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  60	if (sk->sk_prot == &tcpv6_prot)
  61		return &inet6_stream_ops;
  62#endif
  63	WARN_ON_ONCE(sk->sk_prot != &tcp_prot);
  64	return &inet_stream_ops;
  65}
  66
  67static int __mptcp_socket_create(struct mptcp_sock *msk)
  68{
  69	struct mptcp_subflow_context *subflow;
  70	struct sock *sk = (struct sock *)msk;
  71	struct socket *ssock;
  72	int err;
  73
  74	err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
  75	if (err)
  76		return err;
  77
  78	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
  79	WRITE_ONCE(msk->first, ssock->sk);
  80	subflow = mptcp_subflow_ctx(ssock->sk);
  81	list_add(&subflow->node, &msk->conn_list);
  82	sock_hold(ssock->sk);
  83	subflow->request_mptcp = 1;
  84	subflow->subflow_id = msk->subflow_id++;
  85
  86	/* This is the first subflow, always with id 0 */
  87	WRITE_ONCE(subflow->local_id, 0);
  88	mptcp_sock_graft(msk->first, sk->sk_socket);
  89	iput(SOCK_INODE(ssock));
  90
  91	return 0;
  92}
  93
  94/* If the MPC handshake is not started, returns the first subflow,
  95 * eventually allocating it.
  96 */
  97struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
  98{
  99	struct sock *sk = (struct sock *)msk;
 100	int ret;
 101
 102	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
 103		return ERR_PTR(-EINVAL);
 104
 105	if (!msk->first) {
 106		ret = __mptcp_socket_create(msk);
 107		if (ret)
 108			return ERR_PTR(ret);
 109	}
 110
 111	return msk->first;
 112}
 113
 114static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
 115{
 116	sk_drops_add(sk, skb);
 117	__kfree_skb(skb);
 118}
 119
 120static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
 121{
 122	WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
 123		   mptcp_sk(sk)->rmem_fwd_alloc + size);
 124}
 125
 126static void mptcp_rmem_charge(struct sock *sk, int size)
 127{
 128	mptcp_rmem_fwd_alloc_add(sk, -size);
 129}
 130
 131static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
 132			       struct sk_buff *from)
 133{
 134	bool fragstolen;
 135	int delta;
 136
 137	if (MPTCP_SKB_CB(from)->offset ||
 138	    !skb_try_coalesce(to, from, &fragstolen, &delta))
 139		return false;
 140
 141	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx",
 142		 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
 143		 to->len, MPTCP_SKB_CB(from)->end_seq);
 144	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
 145
 146	/* note the fwd memory can reach a negative value after accounting
 147	 * for the delta, but the later skb free will restore a non
 148	 * negative one
 149	 */
 150	atomic_add(delta, &sk->sk_rmem_alloc);
 151	mptcp_rmem_charge(sk, delta);
 152	kfree_skb_partial(from, fragstolen);
 153
 154	return true;
 155}
 156
 157static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
 158				   struct sk_buff *from)
 159{
 160	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
 161		return false;
 162
 163	return mptcp_try_coalesce((struct sock *)msk, to, from);
 164}
 165
 166static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
 167{
 168	amount >>= PAGE_SHIFT;
 169	mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
 170	__sk_mem_reduce_allocated(sk, amount);
 171}
 172
 173static void mptcp_rmem_uncharge(struct sock *sk, int size)
 174{
 175	struct mptcp_sock *msk = mptcp_sk(sk);
 176	int reclaimable;
 177
 178	mptcp_rmem_fwd_alloc_add(sk, size);
 179	reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
 180
 181	/* see sk_mem_uncharge() for the rationale behind the following schema */
 182	if (unlikely(reclaimable >= PAGE_SIZE))
 183		__mptcp_rmem_reclaim(sk, reclaimable);
 184}
 185
 186static void mptcp_rfree(struct sk_buff *skb)
 187{
 188	unsigned int len = skb->truesize;
 189	struct sock *sk = skb->sk;
 190
 191	atomic_sub(len, &sk->sk_rmem_alloc);
 192	mptcp_rmem_uncharge(sk, len);
 193}
 194
 195void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
 196{
 197	skb_orphan(skb);
 198	skb->sk = sk;
 199	skb->destructor = mptcp_rfree;
 200	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 201	mptcp_rmem_charge(sk, skb->truesize);
 202}
 203
 204/* "inspired" by tcp_data_queue_ofo(), main differences:
 205 * - use mptcp seqs
 206 * - don't cope with sacks
 207 */
 208static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
 209{
 210	struct sock *sk = (struct sock *)msk;
 211	struct rb_node **p, *parent;
 212	u64 seq, end_seq, max_seq;
 213	struct sk_buff *skb1;
 214
 215	seq = MPTCP_SKB_CB(skb)->map_seq;
 216	end_seq = MPTCP_SKB_CB(skb)->end_seq;
 217	max_seq = atomic64_read(&msk->rcv_wnd_sent);
 218
 219	pr_debug("msk=%p seq=%llx limit=%llx empty=%d", msk, seq, max_seq,
 220		 RB_EMPTY_ROOT(&msk->out_of_order_queue));
 221	if (after64(end_seq, max_seq)) {
 222		/* out of window */
 223		mptcp_drop(sk, skb);
 224		pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
 225			 (unsigned long long)end_seq - (unsigned long)max_seq,
 226			 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
 227		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
 228		return;
 229	}
 230
 231	p = &msk->out_of_order_queue.rb_node;
 232	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
 233	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
 234		rb_link_node(&skb->rbnode, NULL, p);
 235		rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 236		msk->ooo_last_skb = skb;
 237		goto end;
 238	}
 239
 240	/* with 2 subflows, adding at end of ooo queue is quite likely
 241	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
 242	 */
 243	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
 244		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 245		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 246		return;
 247	}
 248
 249	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
 250	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
 251		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 252		parent = &msk->ooo_last_skb->rbnode;
 253		p = &parent->rb_right;
 254		goto insert;
 255	}
 256
 257	/* Find place to insert this segment. Handle overlaps on the way. */
 258	parent = NULL;
 259	while (*p) {
 260		parent = *p;
 261		skb1 = rb_to_skb(parent);
 262		if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 263			p = &parent->rb_left;
 264			continue;
 265		}
 266		if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 267			if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 268				/* All the bits are present. Drop. */
 269				mptcp_drop(sk, skb);
 270				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 271				return;
 272			}
 273			if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 274				/* partial overlap:
 275				 *     |     skb      |
 276				 *  |     skb1    |
 277				 * continue traversing
 278				 */
 279			} else {
 280				/* skb's seq == skb1's seq and skb covers skb1.
 281				 * Replace skb1 with skb.
 282				 */
 283				rb_replace_node(&skb1->rbnode, &skb->rbnode,
 284						&msk->out_of_order_queue);
 285				mptcp_drop(sk, skb1);
 286				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 287				goto merge_right;
 288			}
 289		} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
 290			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 291			return;
 292		}
 293		p = &parent->rb_right;
 294	}
 295
 296insert:
 297	/* Insert segment into RB tree. */
 298	rb_link_node(&skb->rbnode, parent, p);
 299	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 300
 301merge_right:
 302	/* Remove other segments covered by skb. */
 303	while ((skb1 = skb_rb_next(skb)) != NULL) {
 304		if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
 305			break;
 306		rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
 307		mptcp_drop(sk, skb1);
 308		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 309	}
 310	/* If there is no skb after us, we are the last_skb ! */
 311	if (!skb1)
 312		msk->ooo_last_skb = skb;
 313
 314end:
 315	skb_condense(skb);
 316	mptcp_set_owner_r(skb, sk);
 317}
 318
 319static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
 320{
 321	struct mptcp_sock *msk = mptcp_sk(sk);
 322	int amt, amount;
 323
 324	if (size <= msk->rmem_fwd_alloc)
 325		return true;
 326
 327	size -= msk->rmem_fwd_alloc;
 328	amt = sk_mem_pages(size);
 329	amount = amt << PAGE_SHIFT;
 330	if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
 331		return false;
 332
 333	mptcp_rmem_fwd_alloc_add(sk, amount);
 334	return true;
 335}
 336
 337static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
 338			     struct sk_buff *skb, unsigned int offset,
 339			     size_t copy_len)
 340{
 341	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 342	struct sock *sk = (struct sock *)msk;
 343	struct sk_buff *tail;
 344	bool has_rxtstamp;
 345
 346	__skb_unlink(skb, &ssk->sk_receive_queue);
 347
 348	skb_ext_reset(skb);
 349	skb_orphan(skb);
 350
 351	/* try to fetch required memory from subflow */
 352	if (!mptcp_rmem_schedule(sk, ssk, skb->truesize))
 353		goto drop;
 354
 355	has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
 356
 357	/* the skb map_seq accounts for the skb offset:
 358	 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
 359	 * value
 360	 */
 361	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
 362	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
 363	MPTCP_SKB_CB(skb)->offset = offset;
 364	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
 365
 366	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
 367		/* in sequence */
 368		msk->bytes_received += copy_len;
 369		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
 370		tail = skb_peek_tail(&sk->sk_receive_queue);
 371		if (tail && mptcp_try_coalesce(sk, tail, skb))
 372			return true;
 373
 374		mptcp_set_owner_r(skb, sk);
 375		__skb_queue_tail(&sk->sk_receive_queue, skb);
 376		return true;
 377	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
 378		mptcp_data_queue_ofo(msk, skb);
 379		return false;
 380	}
 381
 382	/* old data, keep it simple and drop the whole pkt, sender
 383	 * will retransmit as needed, if needed.
 384	 */
 385	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 386drop:
 387	mptcp_drop(sk, skb);
 388	return false;
 389}
 390
 391static void mptcp_stop_rtx_timer(struct sock *sk)
 392{
 393	struct inet_connection_sock *icsk = inet_csk(sk);
 394
 395	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 396	mptcp_sk(sk)->timer_ival = 0;
 397}
 398
 399static void mptcp_close_wake_up(struct sock *sk)
 400{
 401	if (sock_flag(sk, SOCK_DEAD))
 402		return;
 403
 404	sk->sk_state_change(sk);
 405	if (sk->sk_shutdown == SHUTDOWN_MASK ||
 406	    sk->sk_state == TCP_CLOSE)
 407		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
 408	else
 409		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
 410}
 411
 412/* called under the msk socket lock */
 413static bool mptcp_pending_data_fin_ack(struct sock *sk)
 414{
 415	struct mptcp_sock *msk = mptcp_sk(sk);
 416
 417	return ((1 << sk->sk_state) &
 418		(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
 419	       msk->write_seq == READ_ONCE(msk->snd_una);
 420}
 421
 422static void mptcp_check_data_fin_ack(struct sock *sk)
 423{
 424	struct mptcp_sock *msk = mptcp_sk(sk);
 425
 426	/* Look for an acknowledged DATA_FIN */
 427	if (mptcp_pending_data_fin_ack(sk)) {
 428		WRITE_ONCE(msk->snd_data_fin_enable, 0);
 429
 430		switch (sk->sk_state) {
 431		case TCP_FIN_WAIT1:
 432			mptcp_set_state(sk, TCP_FIN_WAIT2);
 433			break;
 434		case TCP_CLOSING:
 435		case TCP_LAST_ACK:
 436			mptcp_set_state(sk, TCP_CLOSE);
 437			break;
 438		}
 439
 440		mptcp_close_wake_up(sk);
 441	}
 442}
 443
 444/* can be called with no lock acquired */
 445static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
 446{
 447	struct mptcp_sock *msk = mptcp_sk(sk);
 448
 449	if (READ_ONCE(msk->rcv_data_fin) &&
 450	    ((1 << inet_sk_state_load(sk)) &
 451	     (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
 452		u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
 453
 454		if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
 455			if (seq)
 456				*seq = rcv_data_fin_seq;
 457
 458			return true;
 459		}
 460	}
 461
 462	return false;
 463}
 464
 465static void mptcp_set_datafin_timeout(struct sock *sk)
 466{
 467	struct inet_connection_sock *icsk = inet_csk(sk);
 468	u32 retransmits;
 469
 470	retransmits = min_t(u32, icsk->icsk_retransmits,
 471			    ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
 472
 473	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
 474}
 475
 476static void __mptcp_set_timeout(struct sock *sk, long tout)
 477{
 478	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
 479}
 480
 481static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
 482{
 483	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 484
 485	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
 486	       inet_csk(ssk)->icsk_timeout - jiffies : 0;
 487}
 488
 489static void mptcp_set_timeout(struct sock *sk)
 490{
 491	struct mptcp_subflow_context *subflow;
 492	long tout = 0;
 493
 494	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
 495		tout = max(tout, mptcp_timeout_from_subflow(subflow));
 496	__mptcp_set_timeout(sk, tout);
 497}
 498
 499static inline bool tcp_can_send_ack(const struct sock *ssk)
 500{
 501	return !((1 << inet_sk_state_load(ssk)) &
 502	       (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
 503}
 504
 505void __mptcp_subflow_send_ack(struct sock *ssk)
 506{
 507	if (tcp_can_send_ack(ssk))
 508		tcp_send_ack(ssk);
 509}
 510
 511static void mptcp_subflow_send_ack(struct sock *ssk)
 512{
 513	bool slow;
 514
 515	slow = lock_sock_fast(ssk);
 516	__mptcp_subflow_send_ack(ssk);
 517	unlock_sock_fast(ssk, slow);
 518}
 519
 520static void mptcp_send_ack(struct mptcp_sock *msk)
 521{
 522	struct mptcp_subflow_context *subflow;
 523
 524	mptcp_for_each_subflow(msk, subflow)
 525		mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
 526}
 527
 528static void mptcp_subflow_cleanup_rbuf(struct sock *ssk)
 529{
 530	bool slow;
 531
 532	slow = lock_sock_fast(ssk);
 533	if (tcp_can_send_ack(ssk))
 534		tcp_cleanup_rbuf(ssk, 1);
 535	unlock_sock_fast(ssk, slow);
 536}
 537
 538static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
 539{
 540	const struct inet_connection_sock *icsk = inet_csk(ssk);
 541	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
 542	const struct tcp_sock *tp = tcp_sk(ssk);
 543
 544	return (ack_pending & ICSK_ACK_SCHED) &&
 545		((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
 546		  READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
 547		 (rx_empty && ack_pending &
 548			      (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
 549}
 550
 551static void mptcp_cleanup_rbuf(struct mptcp_sock *msk)
 552{
 553	int old_space = READ_ONCE(msk->old_wspace);
 554	struct mptcp_subflow_context *subflow;
 555	struct sock *sk = (struct sock *)msk;
 556	int space =  __mptcp_space(sk);
 557	bool cleanup, rx_empty;
 558
 559	cleanup = (space > 0) && (space >= (old_space << 1));
 560	rx_empty = !__mptcp_rmem(sk);
 561
 562	mptcp_for_each_subflow(msk, subflow) {
 563		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 564
 565		if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
 566			mptcp_subflow_cleanup_rbuf(ssk);
 567	}
 568}
 569
 570static bool mptcp_check_data_fin(struct sock *sk)
 571{
 572	struct mptcp_sock *msk = mptcp_sk(sk);
 573	u64 rcv_data_fin_seq;
 574	bool ret = false;
 575
 576	/* Need to ack a DATA_FIN received from a peer while this side
 577	 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
 578	 * msk->rcv_data_fin was set when parsing the incoming options
 579	 * at the subflow level and the msk lock was not held, so this
 580	 * is the first opportunity to act on the DATA_FIN and change
 581	 * the msk state.
 582	 *
 583	 * If we are caught up to the sequence number of the incoming
 584	 * DATA_FIN, send the DATA_ACK now and do state transition.  If
 585	 * not caught up, do nothing and let the recv code send DATA_ACK
 586	 * when catching up.
 587	 */
 588
 589	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
 590		WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
 591		WRITE_ONCE(msk->rcv_data_fin, 0);
 592
 593		WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
 594		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
 595
 596		switch (sk->sk_state) {
 597		case TCP_ESTABLISHED:
 598			mptcp_set_state(sk, TCP_CLOSE_WAIT);
 599			break;
 600		case TCP_FIN_WAIT1:
 601			mptcp_set_state(sk, TCP_CLOSING);
 602			break;
 603		case TCP_FIN_WAIT2:
 604			mptcp_set_state(sk, TCP_CLOSE);
 605			break;
 606		default:
 607			/* Other states not expected */
 608			WARN_ON_ONCE(1);
 609			break;
 610		}
 611
 612		ret = true;
 613		if (!__mptcp_check_fallback(msk))
 614			mptcp_send_ack(msk);
 615		mptcp_close_wake_up(sk);
 616	}
 617	return ret;
 618}
 619
 620static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
 621					   struct sock *ssk,
 622					   unsigned int *bytes)
 623{
 624	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 625	struct sock *sk = (struct sock *)msk;
 626	unsigned int moved = 0;
 627	bool more_data_avail;
 628	struct tcp_sock *tp;
 629	bool done = false;
 630	int sk_rbuf;
 631
 632	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
 633
 634	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 635		int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
 636
 637		if (unlikely(ssk_rbuf > sk_rbuf)) {
 638			WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
 639			sk_rbuf = ssk_rbuf;
 640		}
 641	}
 642
 643	pr_debug("msk=%p ssk=%p", msk, ssk);
 644	tp = tcp_sk(ssk);
 645	do {
 646		u32 map_remaining, offset;
 647		u32 seq = tp->copied_seq;
 648		struct sk_buff *skb;
 649		bool fin;
 650
 651		/* try to move as much data as available */
 652		map_remaining = subflow->map_data_len -
 653				mptcp_subflow_get_map_offset(subflow);
 654
 655		skb = skb_peek(&ssk->sk_receive_queue);
 656		if (!skb) {
 657			/* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
 658			 * a different CPU can have already processed the pending
 659			 * data, stop here or we can enter an infinite loop
 660			 */
 661			if (!moved)
 662				done = true;
 663			break;
 664		}
 665
 666		if (__mptcp_check_fallback(msk)) {
 667			/* Under fallback skbs have no MPTCP extension and TCP could
 668			 * collapse them between the dummy map creation and the
 669			 * current dequeue. Be sure to adjust the map size.
 670			 */
 671			map_remaining = skb->len;
 672			subflow->map_data_len = skb->len;
 673		}
 674
 675		offset = seq - TCP_SKB_CB(skb)->seq;
 676		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
 677		if (fin) {
 678			done = true;
 679			seq++;
 680		}
 681
 682		if (offset < skb->len) {
 683			size_t len = skb->len - offset;
 684
 685			if (tp->urg_data)
 686				done = true;
 687
 688			if (__mptcp_move_skb(msk, ssk, skb, offset, len))
 689				moved += len;
 690			seq += len;
 691
 692			if (WARN_ON_ONCE(map_remaining < len))
 693				break;
 694		} else {
 695			WARN_ON_ONCE(!fin);
 696			sk_eat_skb(ssk, skb);
 697			done = true;
 698		}
 699
 700		WRITE_ONCE(tp->copied_seq, seq);
 701		more_data_avail = mptcp_subflow_data_available(ssk);
 702
 703		if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
 704			done = true;
 705			break;
 706		}
 707	} while (more_data_avail);
 708
 709	*bytes += moved;
 710	return done;
 711}
 712
 713static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
 714{
 715	struct sock *sk = (struct sock *)msk;
 716	struct sk_buff *skb, *tail;
 717	bool moved = false;
 718	struct rb_node *p;
 719	u64 end_seq;
 720
 721	p = rb_first(&msk->out_of_order_queue);
 722	pr_debug("msk=%p empty=%d", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
 723	while (p) {
 724		skb = rb_to_skb(p);
 725		if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
 726			break;
 727
 728		p = rb_next(p);
 729		rb_erase(&skb->rbnode, &msk->out_of_order_queue);
 730
 731		if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
 732				      msk->ack_seq))) {
 733			mptcp_drop(sk, skb);
 734			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 735			continue;
 736		}
 737
 738		end_seq = MPTCP_SKB_CB(skb)->end_seq;
 739		tail = skb_peek_tail(&sk->sk_receive_queue);
 740		if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
 741			int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
 742
 743			/* skip overlapping data, if any */
 744			pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d",
 745				 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
 746				 delta);
 747			MPTCP_SKB_CB(skb)->offset += delta;
 748			MPTCP_SKB_CB(skb)->map_seq += delta;
 749			__skb_queue_tail(&sk->sk_receive_queue, skb);
 750		}
 751		msk->bytes_received += end_seq - msk->ack_seq;
 752		WRITE_ONCE(msk->ack_seq, end_seq);
 753		moved = true;
 754	}
 755	return moved;
 756}
 757
 758static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
 759{
 760	int err = sock_error(ssk);
 761	int ssk_state;
 762
 763	if (!err)
 764		return false;
 765
 766	/* only propagate errors on fallen-back sockets or
 767	 * on MPC connect
 768	 */
 769	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
 770		return false;
 771
 772	/* We need to propagate only transition to CLOSE state.
 773	 * Orphaned socket will see such state change via
 774	 * subflow_sched_work_if_closed() and that path will properly
 775	 * destroy the msk as needed.
 776	 */
 777	ssk_state = inet_sk_state_load(ssk);
 778	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
 779		mptcp_set_state(sk, ssk_state);
 780	WRITE_ONCE(sk->sk_err, -err);
 781
 782	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
 783	smp_wmb();
 784	sk_error_report(sk);
 785	return true;
 786}
 787
 788void __mptcp_error_report(struct sock *sk)
 789{
 790	struct mptcp_subflow_context *subflow;
 791	struct mptcp_sock *msk = mptcp_sk(sk);
 792
 793	mptcp_for_each_subflow(msk, subflow)
 794		if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
 795			break;
 796}
 797
 798/* In most cases we will be able to lock the mptcp socket.  If its already
 799 * owned, we need to defer to the work queue to avoid ABBA deadlock.
 800 */
 801static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
 802{
 803	struct sock *sk = (struct sock *)msk;
 804	unsigned int moved = 0;
 805
 806	__mptcp_move_skbs_from_subflow(msk, ssk, &moved);
 807	__mptcp_ofo_queue(msk);
 808	if (unlikely(ssk->sk_err)) {
 809		if (!sock_owned_by_user(sk))
 810			__mptcp_error_report(sk);
 811		else
 812			__set_bit(MPTCP_ERROR_REPORT,  &msk->cb_flags);
 813	}
 814
 815	/* If the moves have caught up with the DATA_FIN sequence number
 816	 * it's time to ack the DATA_FIN and change socket state, but
 817	 * this is not a good place to change state. Let the workqueue
 818	 * do it.
 819	 */
 820	if (mptcp_pending_data_fin(sk, NULL))
 821		mptcp_schedule_work(sk);
 822	return moved > 0;
 823}
 824
 825void mptcp_data_ready(struct sock *sk, struct sock *ssk)
 826{
 827	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 828	struct mptcp_sock *msk = mptcp_sk(sk);
 829	int sk_rbuf, ssk_rbuf;
 830
 831	/* The peer can send data while we are shutting down this
 832	 * subflow at msk destruction time, but we must avoid enqueuing
 833	 * more data to the msk receive queue
 834	 */
 835	if (unlikely(subflow->disposable))
 836		return;
 837
 838	ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
 839	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
 840	if (unlikely(ssk_rbuf > sk_rbuf))
 841		sk_rbuf = ssk_rbuf;
 842
 843	/* over limit? can't append more skbs to msk, Also, no need to wake-up*/
 844	if (__mptcp_rmem(sk) > sk_rbuf) {
 845		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
 846		return;
 847	}
 848
 849	/* Wake-up the reader only for in-sequence data */
 850	mptcp_data_lock(sk);
 851	if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
 852		sk->sk_data_ready(sk);
 853	mptcp_data_unlock(sk);
 854}
 855
 856static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
 857{
 858	mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
 859	WRITE_ONCE(msk->allow_infinite_fallback, false);
 860	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
 861}
 862
 863static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
 864{
 865	struct sock *sk = (struct sock *)msk;
 866
 867	if (sk->sk_state != TCP_ESTABLISHED)
 868		return false;
 869
 870	/* attach to msk socket only after we are sure we will deal with it
 871	 * at close time
 872	 */
 873	if (sk->sk_socket && !ssk->sk_socket)
 874		mptcp_sock_graft(ssk, sk->sk_socket);
 875
 876	mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
 877	mptcp_sockopt_sync_locked(msk, ssk);
 878	mptcp_subflow_joined(msk, ssk);
 879	mptcp_stop_tout_timer(sk);
 880	__mptcp_propagate_sndbuf(sk, ssk);
 881	return true;
 882}
 883
 884static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
 885{
 886	struct mptcp_subflow_context *tmp, *subflow;
 887	struct mptcp_sock *msk = mptcp_sk(sk);
 888
 889	list_for_each_entry_safe(subflow, tmp, join_list, node) {
 890		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 891		bool slow = lock_sock_fast(ssk);
 892
 893		list_move_tail(&subflow->node, &msk->conn_list);
 894		if (!__mptcp_finish_join(msk, ssk))
 895			mptcp_subflow_reset(ssk);
 896		unlock_sock_fast(ssk, slow);
 897	}
 898}
 899
 900static bool mptcp_rtx_timer_pending(struct sock *sk)
 901{
 902	return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
 903}
 904
 905static void mptcp_reset_rtx_timer(struct sock *sk)
 906{
 907	struct inet_connection_sock *icsk = inet_csk(sk);
 908	unsigned long tout;
 909
 910	/* prevent rescheduling on close */
 911	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
 912		return;
 913
 914	tout = mptcp_sk(sk)->timer_ival;
 915	sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
 916}
 917
 918bool mptcp_schedule_work(struct sock *sk)
 919{
 920	if (inet_sk_state_load(sk) != TCP_CLOSE &&
 921	    schedule_work(&mptcp_sk(sk)->work)) {
 922		/* each subflow already holds a reference to the sk, and the
 923		 * workqueue is invoked by a subflow, so sk can't go away here.
 924		 */
 925		sock_hold(sk);
 926		return true;
 927	}
 928	return false;
 929}
 930
 931static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
 932{
 933	struct mptcp_subflow_context *subflow;
 934
 935	msk_owned_by_me(msk);
 936
 937	mptcp_for_each_subflow(msk, subflow) {
 938		if (READ_ONCE(subflow->data_avail))
 939			return mptcp_subflow_tcp_sock(subflow);
 940	}
 941
 942	return NULL;
 943}
 944
 945static bool mptcp_skb_can_collapse_to(u64 write_seq,
 946				      const struct sk_buff *skb,
 947				      const struct mptcp_ext *mpext)
 948{
 949	if (!tcp_skb_can_collapse_to(skb))
 950		return false;
 951
 952	/* can collapse only if MPTCP level sequence is in order and this
 953	 * mapping has not been xmitted yet
 954	 */
 955	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
 956	       !mpext->frozen;
 957}
 958
 959/* we can append data to the given data frag if:
 960 * - there is space available in the backing page_frag
 961 * - the data frag tail matches the current page_frag free offset
 962 * - the data frag end sequence number matches the current write seq
 963 */
 964static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
 965				       const struct page_frag *pfrag,
 966				       const struct mptcp_data_frag *df)
 967{
 968	return df && pfrag->page == df->page &&
 969		pfrag->size - pfrag->offset > 0 &&
 970		pfrag->offset == (df->offset + df->data_len) &&
 971		df->data_seq + df->data_len == msk->write_seq;
 972}
 973
 974static void dfrag_uncharge(struct sock *sk, int len)
 975{
 976	sk_mem_uncharge(sk, len);
 977	sk_wmem_queued_add(sk, -len);
 978}
 979
 980static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
 981{
 982	int len = dfrag->data_len + dfrag->overhead;
 983
 984	list_del(&dfrag->list);
 985	dfrag_uncharge(sk, len);
 986	put_page(dfrag->page);
 987}
 988
 989/* called under both the msk socket lock and the data lock */
 990static void __mptcp_clean_una(struct sock *sk)
 991{
 992	struct mptcp_sock *msk = mptcp_sk(sk);
 993	struct mptcp_data_frag *dtmp, *dfrag;
 994	u64 snd_una;
 995
 996	snd_una = msk->snd_una;
 997	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
 998		if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
 999			break;
1000
1001		if (unlikely(dfrag == msk->first_pending)) {
1002			/* in recovery mode can see ack after the current snd head */
1003			if (WARN_ON_ONCE(!msk->recovery))
1004				break;
1005
1006			WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1007		}
1008
1009		dfrag_clear(sk, dfrag);
1010	}
1011
1012	dfrag = mptcp_rtx_head(sk);
1013	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1014		u64 delta = snd_una - dfrag->data_seq;
1015
1016		/* prevent wrap around in recovery mode */
1017		if (unlikely(delta > dfrag->already_sent)) {
1018			if (WARN_ON_ONCE(!msk->recovery))
1019				goto out;
1020			if (WARN_ON_ONCE(delta > dfrag->data_len))
1021				goto out;
1022			dfrag->already_sent += delta - dfrag->already_sent;
1023		}
1024
1025		dfrag->data_seq += delta;
1026		dfrag->offset += delta;
1027		dfrag->data_len -= delta;
1028		dfrag->already_sent -= delta;
1029
1030		dfrag_uncharge(sk, delta);
1031	}
1032
1033	/* all retransmitted data acked, recovery completed */
1034	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1035		msk->recovery = false;
1036
1037out:
1038	if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1039		if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1040			mptcp_stop_rtx_timer(sk);
1041	} else {
1042		mptcp_reset_rtx_timer(sk);
1043	}
1044
1045	if (mptcp_pending_data_fin_ack(sk))
1046		mptcp_schedule_work(sk);
1047}
1048
1049static void __mptcp_clean_una_wakeup(struct sock *sk)
1050{
1051	lockdep_assert_held_once(&sk->sk_lock.slock);
1052
1053	__mptcp_clean_una(sk);
1054	mptcp_write_space(sk);
1055}
1056
1057static void mptcp_clean_una_wakeup(struct sock *sk)
1058{
1059	mptcp_data_lock(sk);
1060	__mptcp_clean_una_wakeup(sk);
1061	mptcp_data_unlock(sk);
1062}
1063
1064static void mptcp_enter_memory_pressure(struct sock *sk)
1065{
1066	struct mptcp_subflow_context *subflow;
1067	struct mptcp_sock *msk = mptcp_sk(sk);
1068	bool first = true;
1069
1070	mptcp_for_each_subflow(msk, subflow) {
1071		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1072
1073		if (first)
1074			tcp_enter_memory_pressure(ssk);
1075		sk_stream_moderate_sndbuf(ssk);
1076
1077		first = false;
1078	}
1079	__mptcp_sync_sndbuf(sk);
1080}
1081
1082/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1083 * data
1084 */
1085static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1086{
1087	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1088					pfrag, sk->sk_allocation)))
1089		return true;
1090
1091	mptcp_enter_memory_pressure(sk);
1092	return false;
1093}
1094
1095static struct mptcp_data_frag *
1096mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1097		      int orig_offset)
1098{
1099	int offset = ALIGN(orig_offset, sizeof(long));
1100	struct mptcp_data_frag *dfrag;
1101
1102	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1103	dfrag->data_len = 0;
1104	dfrag->data_seq = msk->write_seq;
1105	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1106	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1107	dfrag->already_sent = 0;
1108	dfrag->page = pfrag->page;
1109
1110	return dfrag;
1111}
1112
1113struct mptcp_sendmsg_info {
1114	int mss_now;
1115	int size_goal;
1116	u16 limit;
1117	u16 sent;
1118	unsigned int flags;
1119	bool data_lock_held;
1120};
1121
1122static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1123				    u64 data_seq, int avail_size)
1124{
1125	u64 window_end = mptcp_wnd_end(msk);
1126	u64 mptcp_snd_wnd;
1127
1128	if (__mptcp_check_fallback(msk))
1129		return avail_size;
1130
1131	mptcp_snd_wnd = window_end - data_seq;
1132	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1133
1134	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1135		tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1136		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1137	}
1138
1139	return avail_size;
1140}
1141
1142static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1143{
1144	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1145
1146	if (!mpext)
1147		return false;
1148	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1149	return true;
1150}
1151
1152static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1153{
1154	struct sk_buff *skb;
1155
1156	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1157	if (likely(skb)) {
1158		if (likely(__mptcp_add_ext(skb, gfp))) {
1159			skb_reserve(skb, MAX_TCP_HEADER);
1160			skb->ip_summed = CHECKSUM_PARTIAL;
1161			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1162			return skb;
1163		}
1164		__kfree_skb(skb);
1165	} else {
1166		mptcp_enter_memory_pressure(sk);
1167	}
1168	return NULL;
1169}
1170
1171static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1172{
1173	struct sk_buff *skb;
1174
1175	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1176	if (!skb)
1177		return NULL;
1178
1179	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1180		tcp_skb_entail(ssk, skb);
1181		return skb;
1182	}
1183	tcp_skb_tsorted_anchor_cleanup(skb);
1184	kfree_skb(skb);
1185	return NULL;
1186}
1187
1188static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1189{
1190	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1191
1192	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1193}
1194
1195/* note: this always recompute the csum on the whole skb, even
1196 * if we just appended a single frag. More status info needed
1197 */
1198static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1199{
1200	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1201	__wsum csum = ~csum_unfold(mpext->csum);
1202	int offset = skb->len - added;
1203
1204	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1205}
1206
1207static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1208				      struct sock *ssk,
1209				      struct mptcp_ext *mpext)
1210{
1211	if (!mpext)
1212		return;
1213
1214	mpext->infinite_map = 1;
1215	mpext->data_len = 0;
1216
1217	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1218	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1219	pr_fallback(msk);
1220	mptcp_do_fallback(ssk);
1221}
1222
1223#define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1224
1225static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1226			      struct mptcp_data_frag *dfrag,
1227			      struct mptcp_sendmsg_info *info)
1228{
1229	u64 data_seq = dfrag->data_seq + info->sent;
1230	int offset = dfrag->offset + info->sent;
1231	struct mptcp_sock *msk = mptcp_sk(sk);
1232	bool zero_window_probe = false;
1233	struct mptcp_ext *mpext = NULL;
1234	bool can_coalesce = false;
1235	bool reuse_skb = true;
1236	struct sk_buff *skb;
1237	size_t copy;
1238	int i;
1239
1240	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u",
1241		 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1242
1243	if (WARN_ON_ONCE(info->sent > info->limit ||
1244			 info->limit > dfrag->data_len))
1245		return 0;
1246
1247	if (unlikely(!__tcp_can_send(ssk)))
1248		return -EAGAIN;
1249
1250	/* compute send limit */
1251	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1252		ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1253	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1254	copy = info->size_goal;
1255
1256	skb = tcp_write_queue_tail(ssk);
1257	if (skb && copy > skb->len) {
1258		/* Limit the write to the size available in the
1259		 * current skb, if any, so that we create at most a new skb.
1260		 * Explicitly tells TCP internals to avoid collapsing on later
1261		 * queue management operation, to avoid breaking the ext <->
1262		 * SSN association set here
1263		 */
1264		mpext = mptcp_get_ext(skb);
1265		if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1266			TCP_SKB_CB(skb)->eor = 1;
1267			tcp_mark_push(tcp_sk(ssk), skb);
1268			goto alloc_skb;
1269		}
1270
1271		i = skb_shinfo(skb)->nr_frags;
1272		can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1273		if (!can_coalesce && i >= READ_ONCE(sysctl_max_skb_frags)) {
1274			tcp_mark_push(tcp_sk(ssk), skb);
1275			goto alloc_skb;
1276		}
1277
1278		copy -= skb->len;
1279	} else {
1280alloc_skb:
1281		skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1282		if (!skb)
1283			return -ENOMEM;
1284
1285		i = skb_shinfo(skb)->nr_frags;
1286		reuse_skb = false;
1287		mpext = mptcp_get_ext(skb);
1288	}
1289
1290	/* Zero window and all data acked? Probe. */
1291	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1292	if (copy == 0) {
1293		u64 snd_una = READ_ONCE(msk->snd_una);
1294
1295		if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1296			tcp_remove_empty_skb(ssk);
1297			return 0;
1298		}
1299
1300		zero_window_probe = true;
1301		data_seq = snd_una - 1;
1302		copy = 1;
1303	}
1304
1305	copy = min_t(size_t, copy, info->limit - info->sent);
1306	if (!sk_wmem_schedule(ssk, copy)) {
1307		tcp_remove_empty_skb(ssk);
1308		return -ENOMEM;
1309	}
1310
1311	if (can_coalesce) {
1312		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1313	} else {
1314		get_page(dfrag->page);
1315		skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1316	}
1317
1318	skb->len += copy;
1319	skb->data_len += copy;
1320	skb->truesize += copy;
1321	sk_wmem_queued_add(ssk, copy);
1322	sk_mem_charge(ssk, copy);
1323	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1324	TCP_SKB_CB(skb)->end_seq += copy;
1325	tcp_skb_pcount_set(skb, 0);
1326
1327	/* on skb reuse we just need to update the DSS len */
1328	if (reuse_skb) {
1329		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1330		mpext->data_len += copy;
1331		goto out;
1332	}
1333
1334	memset(mpext, 0, sizeof(*mpext));
1335	mpext->data_seq = data_seq;
1336	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1337	mpext->data_len = copy;
1338	mpext->use_map = 1;
1339	mpext->dsn64 = 1;
1340
1341	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d",
1342		 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1343		 mpext->dsn64);
1344
1345	if (zero_window_probe) {
1346		mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1347		mpext->frozen = 1;
1348		if (READ_ONCE(msk->csum_enabled))
1349			mptcp_update_data_checksum(skb, copy);
1350		tcp_push_pending_frames(ssk);
1351		return 0;
1352	}
1353out:
1354	if (READ_ONCE(msk->csum_enabled))
1355		mptcp_update_data_checksum(skb, copy);
1356	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1357		mptcp_update_infinite_map(msk, ssk, mpext);
1358	trace_mptcp_sendmsg_frag(mpext);
1359	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1360	return copy;
1361}
1362
1363#define MPTCP_SEND_BURST_SIZE		((1 << 16) - \
1364					 sizeof(struct tcphdr) - \
1365					 MAX_TCP_OPTION_SPACE - \
1366					 sizeof(struct ipv6hdr) - \
1367					 sizeof(struct frag_hdr))
1368
1369struct subflow_send_info {
1370	struct sock *ssk;
1371	u64 linger_time;
1372};
1373
1374void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1375{
1376	if (!subflow->stale)
1377		return;
1378
1379	subflow->stale = 0;
1380	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1381}
1382
1383bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1384{
1385	if (unlikely(subflow->stale)) {
1386		u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1387
1388		if (subflow->stale_rcv_tstamp == rcv_tstamp)
1389			return false;
1390
1391		mptcp_subflow_set_active(subflow);
1392	}
1393	return __mptcp_subflow_active(subflow);
1394}
1395
1396#define SSK_MODE_ACTIVE	0
1397#define SSK_MODE_BACKUP	1
1398#define SSK_MODE_MAX	2
1399
1400/* implement the mptcp packet scheduler;
1401 * returns the subflow that will transmit the next DSS
1402 * additionally updates the rtx timeout
1403 */
1404struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1405{
1406	struct subflow_send_info send_info[SSK_MODE_MAX];
1407	struct mptcp_subflow_context *subflow;
1408	struct sock *sk = (struct sock *)msk;
1409	u32 pace, burst, wmem;
1410	int i, nr_active = 0;
1411	struct sock *ssk;
1412	u64 linger_time;
1413	long tout = 0;
1414
1415	/* pick the subflow with the lower wmem/wspace ratio */
1416	for (i = 0; i < SSK_MODE_MAX; ++i) {
1417		send_info[i].ssk = NULL;
1418		send_info[i].linger_time = -1;
1419	}
1420
1421	mptcp_for_each_subflow(msk, subflow) {
1422		trace_mptcp_subflow_get_send(subflow);
1423		ssk =  mptcp_subflow_tcp_sock(subflow);
1424		if (!mptcp_subflow_active(subflow))
1425			continue;
1426
1427		tout = max(tout, mptcp_timeout_from_subflow(subflow));
1428		nr_active += !subflow->backup;
1429		pace = subflow->avg_pacing_rate;
1430		if (unlikely(!pace)) {
1431			/* init pacing rate from socket */
1432			subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1433			pace = subflow->avg_pacing_rate;
1434			if (!pace)
1435				continue;
1436		}
1437
1438		linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1439		if (linger_time < send_info[subflow->backup].linger_time) {
1440			send_info[subflow->backup].ssk = ssk;
1441			send_info[subflow->backup].linger_time = linger_time;
1442		}
1443	}
1444	__mptcp_set_timeout(sk, tout);
1445
1446	/* pick the best backup if no other subflow is active */
1447	if (!nr_active)
1448		send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1449
1450	/* According to the blest algorithm, to avoid HoL blocking for the
1451	 * faster flow, we need to:
1452	 * - estimate the faster flow linger time
1453	 * - use the above to estimate the amount of byte transferred
1454	 *   by the faster flow
1455	 * - check that the amount of queued data is greter than the above,
1456	 *   otherwise do not use the picked, slower, subflow
1457	 * We select the subflow with the shorter estimated time to flush
1458	 * the queued mem, which basically ensure the above. We just need
1459	 * to check that subflow has a non empty cwin.
1460	 */
1461	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1462	if (!ssk || !sk_stream_memory_free(ssk))
1463		return NULL;
1464
1465	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1466	wmem = READ_ONCE(ssk->sk_wmem_queued);
1467	if (!burst)
1468		return ssk;
1469
1470	subflow = mptcp_subflow_ctx(ssk);
1471	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1472					   READ_ONCE(ssk->sk_pacing_rate) * burst,
1473					   burst + wmem);
1474	msk->snd_burst = burst;
1475	return ssk;
1476}
1477
1478static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1479{
1480	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1481	release_sock(ssk);
1482}
1483
1484static void mptcp_update_post_push(struct mptcp_sock *msk,
1485				   struct mptcp_data_frag *dfrag,
1486				   u32 sent)
1487{
1488	u64 snd_nxt_new = dfrag->data_seq;
1489
1490	dfrag->already_sent += sent;
1491
1492	msk->snd_burst -= sent;
1493
1494	snd_nxt_new += dfrag->already_sent;
1495
1496	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1497	 * is recovering after a failover. In that event, this re-sends
1498	 * old segments.
1499	 *
1500	 * Thus compute snd_nxt_new candidate based on
1501	 * the dfrag->data_seq that was sent and the data
1502	 * that has been handed to the subflow for transmission
1503	 * and skip update in case it was old dfrag.
1504	 */
1505	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1506		msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1507		WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1508	}
1509}
1510
1511void mptcp_check_and_set_pending(struct sock *sk)
1512{
1513	if (mptcp_send_head(sk)) {
1514		mptcp_data_lock(sk);
1515		mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1516		mptcp_data_unlock(sk);
1517	}
1518}
1519
1520static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1521				  struct mptcp_sendmsg_info *info)
1522{
1523	struct mptcp_sock *msk = mptcp_sk(sk);
1524	struct mptcp_data_frag *dfrag;
1525	int len, copied = 0, err = 0;
1526
1527	while ((dfrag = mptcp_send_head(sk))) {
1528		info->sent = dfrag->already_sent;
1529		info->limit = dfrag->data_len;
1530		len = dfrag->data_len - dfrag->already_sent;
1531		while (len > 0) {
1532			int ret = 0;
1533
1534			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1535			if (ret <= 0) {
1536				err = copied ? : ret;
1537				goto out;
1538			}
1539
1540			info->sent += ret;
1541			copied += ret;
1542			len -= ret;
1543
1544			mptcp_update_post_push(msk, dfrag, ret);
1545		}
1546		WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1547
1548		if (msk->snd_burst <= 0 ||
1549		    !sk_stream_memory_free(ssk) ||
1550		    !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1551			err = copied;
1552			goto out;
1553		}
1554		mptcp_set_timeout(sk);
1555	}
1556	err = copied;
1557
1558out:
1559	return err;
1560}
1561
1562void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1563{
1564	struct sock *prev_ssk = NULL, *ssk = NULL;
1565	struct mptcp_sock *msk = mptcp_sk(sk);
1566	struct mptcp_sendmsg_info info = {
1567				.flags = flags,
1568	};
1569	bool do_check_data_fin = false;
1570	int push_count = 1;
1571
1572	while (mptcp_send_head(sk) && (push_count > 0)) {
1573		struct mptcp_subflow_context *subflow;
1574		int ret = 0;
1575
1576		if (mptcp_sched_get_send(msk))
1577			break;
1578
1579		push_count = 0;
1580
1581		mptcp_for_each_subflow(msk, subflow) {
1582			if (READ_ONCE(subflow->scheduled)) {
1583				mptcp_subflow_set_scheduled(subflow, false);
1584
1585				prev_ssk = ssk;
1586				ssk = mptcp_subflow_tcp_sock(subflow);
1587				if (ssk != prev_ssk) {
1588					/* First check. If the ssk has changed since
1589					 * the last round, release prev_ssk
1590					 */
1591					if (prev_ssk)
1592						mptcp_push_release(prev_ssk, &info);
1593
1594					/* Need to lock the new subflow only if different
1595					 * from the previous one, otherwise we are still
1596					 * helding the relevant lock
1597					 */
1598					lock_sock(ssk);
1599				}
1600
1601				push_count++;
1602
1603				ret = __subflow_push_pending(sk, ssk, &info);
1604				if (ret <= 0) {
1605					if (ret != -EAGAIN ||
1606					    (1 << ssk->sk_state) &
1607					     (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1608						push_count--;
1609					continue;
1610				}
1611				do_check_data_fin = true;
1612			}
1613		}
1614	}
1615
1616	/* at this point we held the socket lock for the last subflow we used */
1617	if (ssk)
1618		mptcp_push_release(ssk, &info);
1619
1620	/* ensure the rtx timer is running */
1621	if (!mptcp_rtx_timer_pending(sk))
1622		mptcp_reset_rtx_timer(sk);
1623	if (do_check_data_fin)
1624		mptcp_check_send_data_fin(sk);
1625}
1626
1627static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1628{
1629	struct mptcp_sock *msk = mptcp_sk(sk);
1630	struct mptcp_sendmsg_info info = {
1631		.data_lock_held = true,
1632	};
1633	bool keep_pushing = true;
1634	struct sock *xmit_ssk;
1635	int copied = 0;
1636
1637	info.flags = 0;
1638	while (mptcp_send_head(sk) && keep_pushing) {
1639		struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1640		int ret = 0;
1641
1642		/* check for a different subflow usage only after
1643		 * spooling the first chunk of data
1644		 */
1645		if (first) {
1646			mptcp_subflow_set_scheduled(subflow, false);
1647			ret = __subflow_push_pending(sk, ssk, &info);
1648			first = false;
1649			if (ret <= 0)
1650				break;
1651			copied += ret;
1652			continue;
1653		}
1654
1655		if (mptcp_sched_get_send(msk))
1656			goto out;
1657
1658		if (READ_ONCE(subflow->scheduled)) {
1659			mptcp_subflow_set_scheduled(subflow, false);
1660			ret = __subflow_push_pending(sk, ssk, &info);
1661			if (ret <= 0)
1662				keep_pushing = false;
1663			copied += ret;
1664		}
1665
1666		mptcp_for_each_subflow(msk, subflow) {
1667			if (READ_ONCE(subflow->scheduled)) {
1668				xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1669				if (xmit_ssk != ssk) {
1670					mptcp_subflow_delegate(subflow,
1671							       MPTCP_DELEGATE_SEND);
1672					keep_pushing = false;
1673				}
1674			}
1675		}
1676	}
1677
1678out:
1679	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1680	 * not going to flush it via release_sock()
1681	 */
1682	if (copied) {
1683		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1684			 info.size_goal);
1685		if (!mptcp_rtx_timer_pending(sk))
1686			mptcp_reset_rtx_timer(sk);
1687
1688		if (msk->snd_data_fin_enable &&
1689		    msk->snd_nxt + 1 == msk->write_seq)
1690			mptcp_schedule_work(sk);
1691	}
1692}
1693
1694static int mptcp_disconnect(struct sock *sk, int flags);
1695
1696static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1697				  size_t len, int *copied_syn)
1698{
1699	unsigned int saved_flags = msg->msg_flags;
1700	struct mptcp_sock *msk = mptcp_sk(sk);
1701	struct sock *ssk;
1702	int ret;
1703
1704	/* on flags based fastopen the mptcp is supposed to create the
1705	 * first subflow right now. Otherwise we are in the defer_connect
1706	 * path, and the first subflow must be already present.
1707	 * Since the defer_connect flag is cleared after the first succsful
1708	 * fastopen attempt, no need to check for additional subflow status.
1709	 */
1710	if (msg->msg_flags & MSG_FASTOPEN) {
1711		ssk = __mptcp_nmpc_sk(msk);
1712		if (IS_ERR(ssk))
1713			return PTR_ERR(ssk);
1714	}
1715	if (!msk->first)
1716		return -EINVAL;
1717
1718	ssk = msk->first;
1719
1720	lock_sock(ssk);
1721	msg->msg_flags |= MSG_DONTWAIT;
1722	msk->fastopening = 1;
1723	ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1724	msk->fastopening = 0;
1725	msg->msg_flags = saved_flags;
1726	release_sock(ssk);
1727
1728	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1729	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1730		ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1731					    msg->msg_namelen, msg->msg_flags, 1);
1732
1733		/* Keep the same behaviour of plain TCP: zero the copied bytes in
1734		 * case of any error, except timeout or signal
1735		 */
1736		if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1737			*copied_syn = 0;
1738	} else if (ret && ret != -EINPROGRESS) {
1739		/* The disconnect() op called by tcp_sendmsg_fastopen()/
1740		 * __inet_stream_connect() can fail, due to looking check,
1741		 * see mptcp_disconnect().
1742		 * Attempt it again outside the problematic scope.
1743		 */
1744		if (!mptcp_disconnect(sk, 0))
1745			sk->sk_socket->state = SS_UNCONNECTED;
1746	}
1747	inet_clear_bit(DEFER_CONNECT, sk);
1748
1749	return ret;
1750}
1751
1752static int do_copy_data_nocache(struct sock *sk, int copy,
1753				struct iov_iter *from, char *to)
1754{
1755	if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1756		if (!copy_from_iter_full_nocache(to, copy, from))
1757			return -EFAULT;
1758	} else if (!copy_from_iter_full(to, copy, from)) {
1759		return -EFAULT;
1760	}
1761	return 0;
1762}
1763
1764/* open-code sk_stream_memory_free() plus sent limit computation to
1765 * avoid indirect calls in fast-path.
1766 * Called under the msk socket lock, so we can avoid a bunch of ONCE
1767 * annotations.
1768 */
1769static u32 mptcp_send_limit(const struct sock *sk)
1770{
1771	const struct mptcp_sock *msk = mptcp_sk(sk);
1772	u32 limit, not_sent;
1773
1774	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1775		return 0;
1776
1777	limit = mptcp_notsent_lowat(sk);
1778	if (limit == UINT_MAX)
1779		return UINT_MAX;
1780
1781	not_sent = msk->write_seq - msk->snd_nxt;
1782	if (not_sent >= limit)
1783		return 0;
1784
1785	return limit - not_sent;
1786}
1787
1788static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1789{
1790	struct mptcp_sock *msk = mptcp_sk(sk);
1791	struct page_frag *pfrag;
1792	size_t copied = 0;
1793	int ret = 0;
1794	long timeo;
1795
1796	/* silently ignore everything else */
1797	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1798
1799	lock_sock(sk);
1800
1801	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1802		     msg->msg_flags & MSG_FASTOPEN)) {
1803		int copied_syn = 0;
1804
1805		ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1806		copied += copied_syn;
1807		if (ret == -EINPROGRESS && copied_syn > 0)
1808			goto out;
1809		else if (ret)
1810			goto do_error;
1811	}
1812
1813	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1814
1815	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1816		ret = sk_stream_wait_connect(sk, &timeo);
1817		if (ret)
1818			goto do_error;
1819	}
1820
1821	ret = -EPIPE;
1822	if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1823		goto do_error;
1824
1825	pfrag = sk_page_frag(sk);
1826
1827	while (msg_data_left(msg)) {
1828		int total_ts, frag_truesize = 0;
1829		struct mptcp_data_frag *dfrag;
1830		bool dfrag_collapsed;
1831		size_t psize, offset;
1832		u32 copy_limit;
1833
1834		/* ensure fitting the notsent_lowat() constraint */
1835		copy_limit = mptcp_send_limit(sk);
1836		if (!copy_limit)
1837			goto wait_for_memory;
1838
1839		/* reuse tail pfrag, if possible, or carve a new one from the
1840		 * page allocator
1841		 */
1842		dfrag = mptcp_pending_tail(sk);
1843		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1844		if (!dfrag_collapsed) {
1845			if (!mptcp_page_frag_refill(sk, pfrag))
1846				goto wait_for_memory;
1847
1848			dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1849			frag_truesize = dfrag->overhead;
1850		}
1851
1852		/* we do not bound vs wspace, to allow a single packet.
1853		 * memory accounting will prevent execessive memory usage
1854		 * anyway
1855		 */
1856		offset = dfrag->offset + dfrag->data_len;
1857		psize = pfrag->size - offset;
1858		psize = min_t(size_t, psize, msg_data_left(msg));
1859		psize = min_t(size_t, psize, copy_limit);
1860		total_ts = psize + frag_truesize;
1861
1862		if (!sk_wmem_schedule(sk, total_ts))
1863			goto wait_for_memory;
1864
1865		ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1866					   page_address(dfrag->page) + offset);
1867		if (ret)
1868			goto do_error;
1869
1870		/* data successfully copied into the write queue */
1871		sk_forward_alloc_add(sk, -total_ts);
1872		copied += psize;
1873		dfrag->data_len += psize;
1874		frag_truesize += psize;
1875		pfrag->offset += frag_truesize;
1876		WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1877
1878		/* charge data on mptcp pending queue to the msk socket
1879		 * Note: we charge such data both to sk and ssk
1880		 */
1881		sk_wmem_queued_add(sk, frag_truesize);
1882		if (!dfrag_collapsed) {
1883			get_page(dfrag->page);
1884			list_add_tail(&dfrag->list, &msk->rtx_queue);
1885			if (!msk->first_pending)
1886				WRITE_ONCE(msk->first_pending, dfrag);
1887		}
1888		pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d", msk,
1889			 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1890			 !dfrag_collapsed);
1891
1892		continue;
1893
1894wait_for_memory:
1895		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1896		__mptcp_push_pending(sk, msg->msg_flags);
1897		ret = sk_stream_wait_memory(sk, &timeo);
1898		if (ret)
1899			goto do_error;
1900	}
1901
1902	if (copied)
1903		__mptcp_push_pending(sk, msg->msg_flags);
1904
1905out:
1906	release_sock(sk);
1907	return copied;
1908
1909do_error:
1910	if (copied)
1911		goto out;
1912
1913	copied = sk_stream_error(sk, msg->msg_flags, ret);
1914	goto out;
1915}
1916
1917static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1918				struct msghdr *msg,
1919				size_t len, int flags,
1920				struct scm_timestamping_internal *tss,
1921				int *cmsg_flags)
1922{
1923	struct sk_buff *skb, *tmp;
1924	int copied = 0;
1925
1926	skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1927		u32 offset = MPTCP_SKB_CB(skb)->offset;
1928		u32 data_len = skb->len - offset;
1929		u32 count = min_t(size_t, len - copied, data_len);
1930		int err;
1931
1932		if (!(flags & MSG_TRUNC)) {
1933			err = skb_copy_datagram_msg(skb, offset, msg, count);
1934			if (unlikely(err < 0)) {
1935				if (!copied)
1936					return err;
1937				break;
1938			}
1939		}
1940
1941		if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1942			tcp_update_recv_tstamps(skb, tss);
1943			*cmsg_flags |= MPTCP_CMSG_TS;
1944		}
1945
1946		copied += count;
1947
1948		if (count < data_len) {
1949			if (!(flags & MSG_PEEK)) {
1950				MPTCP_SKB_CB(skb)->offset += count;
1951				MPTCP_SKB_CB(skb)->map_seq += count;
1952				msk->bytes_consumed += count;
1953			}
1954			break;
1955		}
1956
1957		if (!(flags & MSG_PEEK)) {
1958			/* we will bulk release the skb memory later */
1959			skb->destructor = NULL;
1960			WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1961			__skb_unlink(skb, &msk->receive_queue);
1962			__kfree_skb(skb);
1963			msk->bytes_consumed += count;
1964		}
1965
1966		if (copied >= len)
1967			break;
1968	}
1969
1970	return copied;
1971}
1972
1973/* receive buffer autotuning.  See tcp_rcv_space_adjust for more information.
1974 *
1975 * Only difference: Use highest rtt estimate of the subflows in use.
1976 */
1977static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
1978{
1979	struct mptcp_subflow_context *subflow;
1980	struct sock *sk = (struct sock *)msk;
1981	u8 scaling_ratio = U8_MAX;
1982	u32 time, advmss = 1;
1983	u64 rtt_us, mstamp;
1984
1985	msk_owned_by_me(msk);
1986
1987	if (copied <= 0)
1988		return;
1989
1990	if (!msk->rcvspace_init)
1991		mptcp_rcv_space_init(msk, msk->first);
1992
1993	msk->rcvq_space.copied += copied;
1994
1995	mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
1996	time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
1997
1998	rtt_us = msk->rcvq_space.rtt_us;
1999	if (rtt_us && time < (rtt_us >> 3))
2000		return;
2001
2002	rtt_us = 0;
2003	mptcp_for_each_subflow(msk, subflow) {
2004		const struct tcp_sock *tp;
2005		u64 sf_rtt_us;
2006		u32 sf_advmss;
2007
2008		tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2009
2010		sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2011		sf_advmss = READ_ONCE(tp->advmss);
2012
2013		rtt_us = max(sf_rtt_us, rtt_us);
2014		advmss = max(sf_advmss, advmss);
2015		scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2016	}
2017
2018	msk->rcvq_space.rtt_us = rtt_us;
2019	msk->scaling_ratio = scaling_ratio;
2020	if (time < (rtt_us >> 3) || rtt_us == 0)
2021		return;
2022
2023	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2024		goto new_measure;
2025
2026	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2027	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2028		u64 rcvwin, grow;
2029		int rcvbuf;
2030
2031		rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2032
2033		grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2034
2035		do_div(grow, msk->rcvq_space.space);
2036		rcvwin += (grow << 1);
2037
2038		rcvbuf = min_t(u64, __tcp_space_from_win(scaling_ratio, rcvwin),
2039			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2040
2041		if (rcvbuf > sk->sk_rcvbuf) {
2042			u32 window_clamp;
2043
2044			window_clamp = __tcp_win_from_space(scaling_ratio, rcvbuf);
2045			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2046
2047			/* Make subflows follow along.  If we do not do this, we
2048			 * get drops at subflow level if skbs can't be moved to
2049			 * the mptcp rx queue fast enough (announced rcv_win can
2050			 * exceed ssk->sk_rcvbuf).
2051			 */
2052			mptcp_for_each_subflow(msk, subflow) {
2053				struct sock *ssk;
2054				bool slow;
2055
2056				ssk = mptcp_subflow_tcp_sock(subflow);
2057				slow = lock_sock_fast(ssk);
2058				WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2059				tcp_sk(ssk)->window_clamp = window_clamp;
2060				tcp_cleanup_rbuf(ssk, 1);
2061				unlock_sock_fast(ssk, slow);
2062			}
2063		}
2064	}
2065
2066	msk->rcvq_space.space = msk->rcvq_space.copied;
2067new_measure:
2068	msk->rcvq_space.copied = 0;
2069	msk->rcvq_space.time = mstamp;
2070}
2071
2072static void __mptcp_update_rmem(struct sock *sk)
2073{
2074	struct mptcp_sock *msk = mptcp_sk(sk);
2075
2076	if (!msk->rmem_released)
2077		return;
2078
2079	atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2080	mptcp_rmem_uncharge(sk, msk->rmem_released);
2081	WRITE_ONCE(msk->rmem_released, 0);
2082}
2083
2084static void __mptcp_splice_receive_queue(struct sock *sk)
2085{
2086	struct mptcp_sock *msk = mptcp_sk(sk);
2087
2088	skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2089}
2090
2091static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2092{
2093	struct sock *sk = (struct sock *)msk;
2094	unsigned int moved = 0;
2095	bool ret, done;
2096
2097	do {
2098		struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2099		bool slowpath;
2100
2101		/* we can have data pending in the subflows only if the msk
2102		 * receive buffer was full at subflow_data_ready() time,
2103		 * that is an unlikely slow path.
2104		 */
2105		if (likely(!ssk))
2106			break;
2107
2108		slowpath = lock_sock_fast(ssk);
2109		mptcp_data_lock(sk);
2110		__mptcp_update_rmem(sk);
2111		done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2112		mptcp_data_unlock(sk);
2113
2114		if (unlikely(ssk->sk_err))
2115			__mptcp_error_report(sk);
2116		unlock_sock_fast(ssk, slowpath);
2117	} while (!done);
2118
2119	/* acquire the data lock only if some input data is pending */
2120	ret = moved > 0;
2121	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2122	    !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2123		mptcp_data_lock(sk);
2124		__mptcp_update_rmem(sk);
2125		ret |= __mptcp_ofo_queue(msk);
2126		__mptcp_splice_receive_queue(sk);
2127		mptcp_data_unlock(sk);
2128	}
2129	if (ret)
2130		mptcp_check_data_fin((struct sock *)msk);
2131	return !skb_queue_empty(&msk->receive_queue);
2132}
2133
2134static unsigned int mptcp_inq_hint(const struct sock *sk)
2135{
2136	const struct mptcp_sock *msk = mptcp_sk(sk);
2137	const struct sk_buff *skb;
2138
2139	skb = skb_peek(&msk->receive_queue);
2140	if (skb) {
2141		u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2142
2143		if (hint_val >= INT_MAX)
2144			return INT_MAX;
2145
2146		return (unsigned int)hint_val;
2147	}
2148
2149	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2150		return 1;
2151
2152	return 0;
2153}
2154
2155static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2156			 int flags, int *addr_len)
2157{
2158	struct mptcp_sock *msk = mptcp_sk(sk);
2159	struct scm_timestamping_internal tss;
2160	int copied = 0, cmsg_flags = 0;
2161	int target;
2162	long timeo;
2163
2164	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2165	if (unlikely(flags & MSG_ERRQUEUE))
2166		return inet_recv_error(sk, msg, len, addr_len);
2167
2168	lock_sock(sk);
2169	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2170		copied = -ENOTCONN;
2171		goto out_err;
2172	}
2173
2174	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2175
2176	len = min_t(size_t, len, INT_MAX);
2177	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2178
2179	if (unlikely(msk->recvmsg_inq))
2180		cmsg_flags = MPTCP_CMSG_INQ;
2181
2182	while (copied < len) {
2183		int bytes_read;
2184
2185		bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2186		if (unlikely(bytes_read < 0)) {
2187			if (!copied)
2188				copied = bytes_read;
2189			goto out_err;
2190		}
2191
2192		copied += bytes_read;
2193
2194		/* be sure to advertise window change */
2195		mptcp_cleanup_rbuf(msk);
2196
2197		if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2198			continue;
2199
2200		/* only the master socket status is relevant here. The exit
2201		 * conditions mirror closely tcp_recvmsg()
2202		 */
2203		if (copied >= target)
2204			break;
2205
2206		if (copied) {
2207			if (sk->sk_err ||
2208			    sk->sk_state == TCP_CLOSE ||
2209			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2210			    !timeo ||
2211			    signal_pending(current))
2212				break;
2213		} else {
2214			if (sk->sk_err) {
2215				copied = sock_error(sk);
2216				break;
2217			}
2218
2219			if (sk->sk_shutdown & RCV_SHUTDOWN) {
2220				/* race breaker: the shutdown could be after the
2221				 * previous receive queue check
2222				 */
2223				if (__mptcp_move_skbs(msk))
2224					continue;
2225				break;
2226			}
2227
2228			if (sk->sk_state == TCP_CLOSE) {
2229				copied = -ENOTCONN;
2230				break;
2231			}
2232
2233			if (!timeo) {
2234				copied = -EAGAIN;
2235				break;
2236			}
2237
2238			if (signal_pending(current)) {
2239				copied = sock_intr_errno(timeo);
2240				break;
2241			}
2242		}
2243
2244		pr_debug("block timeout %ld", timeo);
2245		sk_wait_data(sk, &timeo, NULL);
2246	}
2247
2248out_err:
2249	if (cmsg_flags && copied >= 0) {
2250		if (cmsg_flags & MPTCP_CMSG_TS)
2251			tcp_recv_timestamp(msg, sk, &tss);
2252
2253		if (cmsg_flags & MPTCP_CMSG_INQ) {
2254			unsigned int inq = mptcp_inq_hint(sk);
2255
2256			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2257		}
2258	}
2259
2260	pr_debug("msk=%p rx queue empty=%d:%d copied=%d",
2261		 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2262		 skb_queue_empty(&msk->receive_queue), copied);
2263	if (!(flags & MSG_PEEK))
2264		mptcp_rcv_space_adjust(msk, copied);
2265
2266	release_sock(sk);
2267	return copied;
2268}
2269
2270static void mptcp_retransmit_timer(struct timer_list *t)
2271{
2272	struct inet_connection_sock *icsk = from_timer(icsk, t,
2273						       icsk_retransmit_timer);
2274	struct sock *sk = &icsk->icsk_inet.sk;
2275	struct mptcp_sock *msk = mptcp_sk(sk);
2276
2277	bh_lock_sock(sk);
2278	if (!sock_owned_by_user(sk)) {
2279		/* we need a process context to retransmit */
2280		if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2281			mptcp_schedule_work(sk);
2282	} else {
2283		/* delegate our work to tcp_release_cb() */
2284		__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2285	}
2286	bh_unlock_sock(sk);
2287	sock_put(sk);
2288}
2289
2290static void mptcp_tout_timer(struct timer_list *t)
2291{
2292	struct sock *sk = from_timer(sk, t, sk_timer);
2293
2294	mptcp_schedule_work(sk);
2295	sock_put(sk);
2296}
2297
2298/* Find an idle subflow.  Return NULL if there is unacked data at tcp
2299 * level.
2300 *
2301 * A backup subflow is returned only if that is the only kind available.
2302 */
2303struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2304{
2305	struct sock *backup = NULL, *pick = NULL;
2306	struct mptcp_subflow_context *subflow;
2307	int min_stale_count = INT_MAX;
2308
2309	mptcp_for_each_subflow(msk, subflow) {
2310		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2311
2312		if (!__mptcp_subflow_active(subflow))
2313			continue;
2314
2315		/* still data outstanding at TCP level? skip this */
2316		if (!tcp_rtx_and_write_queues_empty(ssk)) {
2317			mptcp_pm_subflow_chk_stale(msk, ssk);
2318			min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2319			continue;
2320		}
2321
2322		if (subflow->backup) {
2323			if (!backup)
2324				backup = ssk;
2325			continue;
2326		}
2327
2328		if (!pick)
2329			pick = ssk;
2330	}
2331
2332	if (pick)
2333		return pick;
2334
2335	/* use backup only if there are no progresses anywhere */
2336	return min_stale_count > 1 ? backup : NULL;
2337}
2338
2339bool __mptcp_retransmit_pending_data(struct sock *sk)
2340{
2341	struct mptcp_data_frag *cur, *rtx_head;
2342	struct mptcp_sock *msk = mptcp_sk(sk);
2343
2344	if (__mptcp_check_fallback(msk))
2345		return false;
2346
2347	/* the closing socket has some data untransmitted and/or unacked:
2348	 * some data in the mptcp rtx queue has not really xmitted yet.
2349	 * keep it simple and re-inject the whole mptcp level rtx queue
2350	 */
2351	mptcp_data_lock(sk);
2352	__mptcp_clean_una_wakeup(sk);
2353	rtx_head = mptcp_rtx_head(sk);
2354	if (!rtx_head) {
2355		mptcp_data_unlock(sk);
2356		return false;
2357	}
2358
2359	msk->recovery_snd_nxt = msk->snd_nxt;
2360	msk->recovery = true;
2361	mptcp_data_unlock(sk);
2362
2363	msk->first_pending = rtx_head;
2364	msk->snd_burst = 0;
2365
2366	/* be sure to clear the "sent status" on all re-injected fragments */
2367	list_for_each_entry(cur, &msk->rtx_queue, list) {
2368		if (!cur->already_sent)
2369			break;
2370		cur->already_sent = 0;
2371	}
2372
2373	return true;
2374}
2375
2376/* flags for __mptcp_close_ssk() */
2377#define MPTCP_CF_PUSH		BIT(1)
2378#define MPTCP_CF_FASTCLOSE	BIT(2)
2379
2380/* be sure to send a reset only if the caller asked for it, also
2381 * clean completely the subflow status when the subflow reaches
2382 * TCP_CLOSE state
2383 */
2384static void __mptcp_subflow_disconnect(struct sock *ssk,
2385				       struct mptcp_subflow_context *subflow,
2386				       unsigned int flags)
2387{
2388	if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2389	    (flags & MPTCP_CF_FASTCLOSE)) {
2390		/* The MPTCP code never wait on the subflow sockets, TCP-level
2391		 * disconnect should never fail
2392		 */
2393		WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2394		mptcp_subflow_ctx_reset(subflow);
2395	} else {
2396		tcp_shutdown(ssk, SEND_SHUTDOWN);
2397	}
2398}
2399
2400/* subflow sockets can be either outgoing (connect) or incoming
2401 * (accept).
2402 *
2403 * Outgoing subflows use in-kernel sockets.
2404 * Incoming subflows do not have their own 'struct socket' allocated,
2405 * so we need to use tcp_close() after detaching them from the mptcp
2406 * parent socket.
2407 */
2408static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2409			      struct mptcp_subflow_context *subflow,
2410			      unsigned int flags)
2411{
2412	struct mptcp_sock *msk = mptcp_sk(sk);
2413	bool dispose_it, need_push = false;
2414
2415	/* If the first subflow moved to a close state before accept, e.g. due
2416	 * to an incoming reset or listener shutdown, the subflow socket is
2417	 * already deleted by inet_child_forget() and the mptcp socket can't
2418	 * survive too.
2419	 */
2420	if (msk->in_accept_queue && msk->first == ssk &&
2421	    (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2422		/* ensure later check in mptcp_worker() will dispose the msk */
2423		sock_set_flag(sk, SOCK_DEAD);
2424		mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2425		lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2426		mptcp_subflow_drop_ctx(ssk);
2427		goto out_release;
2428	}
2429
2430	dispose_it = msk->free_first || ssk != msk->first;
2431	if (dispose_it)
2432		list_del(&subflow->node);
2433
2434	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2435
2436	if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2437		/* be sure to force the tcp_close path
2438		 * to generate the egress reset
2439		 */
2440		ssk->sk_lingertime = 0;
2441		sock_set_flag(ssk, SOCK_LINGER);
2442		subflow->send_fastclose = 1;
2443	}
2444
2445	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2446	if (!dispose_it) {
2447		__mptcp_subflow_disconnect(ssk, subflow, flags);
2448		release_sock(ssk);
2449
2450		goto out;
2451	}
2452
2453	subflow->disposable = 1;
2454
2455	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2456	 * the ssk has been already destroyed, we just need to release the
2457	 * reference owned by msk;
2458	 */
2459	if (!inet_csk(ssk)->icsk_ulp_ops) {
2460		WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2461		kfree_rcu(subflow, rcu);
2462	} else {
2463		/* otherwise tcp will dispose of the ssk and subflow ctx */
2464		__tcp_close(ssk, 0);
2465
2466		/* close acquired an extra ref */
2467		__sock_put(ssk);
2468	}
2469
2470out_release:
2471	__mptcp_subflow_error_report(sk, ssk);
2472	release_sock(ssk);
2473
2474	sock_put(ssk);
2475
2476	if (ssk == msk->first)
2477		WRITE_ONCE(msk->first, NULL);
2478
2479out:
2480	__mptcp_sync_sndbuf(sk);
2481	if (need_push)
2482		__mptcp_push_pending(sk, 0);
2483
2484	/* Catch every 'all subflows closed' scenario, including peers silently
2485	 * closing them, e.g. due to timeout.
2486	 * For established sockets, allow an additional timeout before closing,
2487	 * as the protocol can still create more subflows.
2488	 */
2489	if (list_is_singular(&msk->conn_list) && msk->first &&
2490	    inet_sk_state_load(msk->first) == TCP_CLOSE) {
2491		if (sk->sk_state != TCP_ESTABLISHED ||
2492		    msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2493			mptcp_set_state(sk, TCP_CLOSE);
2494			mptcp_close_wake_up(sk);
2495		} else {
2496			mptcp_start_tout_timer(sk);
2497		}
2498	}
2499}
2500
2501void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2502		     struct mptcp_subflow_context *subflow)
2503{
2504	if (sk->sk_state == TCP_ESTABLISHED)
2505		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2506
2507	/* subflow aborted before reaching the fully_established status
2508	 * attempt the creation of the next subflow
2509	 */
2510	mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2511
2512	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2513}
2514
2515static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2516{
2517	return 0;
2518}
2519
2520static void __mptcp_close_subflow(struct sock *sk)
2521{
2522	struct mptcp_subflow_context *subflow, *tmp;
2523	struct mptcp_sock *msk = mptcp_sk(sk);
2524
2525	might_sleep();
2526
2527	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2528		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2529
2530		if (inet_sk_state_load(ssk) != TCP_CLOSE)
2531			continue;
2532
2533		/* 'subflow_data_ready' will re-sched once rx queue is empty */
2534		if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2535			continue;
2536
2537		mptcp_close_ssk(sk, ssk, subflow);
2538	}
2539
2540}
2541
2542static bool mptcp_close_tout_expired(const struct sock *sk)
2543{
2544	if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2545	    sk->sk_state == TCP_CLOSE)
2546		return false;
2547
2548	return time_after32(tcp_jiffies32,
2549		  inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2550}
2551
2552static void mptcp_check_fastclose(struct mptcp_sock *msk)
2553{
2554	struct mptcp_subflow_context *subflow, *tmp;
2555	struct sock *sk = (struct sock *)msk;
2556
2557	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2558		return;
2559
2560	mptcp_token_destroy(msk);
2561
2562	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2563		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2564		bool slow;
2565
2566		slow = lock_sock_fast(tcp_sk);
2567		if (tcp_sk->sk_state != TCP_CLOSE) {
2568			tcp_send_active_reset(tcp_sk, GFP_ATOMIC);
2569			tcp_set_state(tcp_sk, TCP_CLOSE);
2570		}
2571		unlock_sock_fast(tcp_sk, slow);
2572	}
2573
2574	/* Mirror the tcp_reset() error propagation */
2575	switch (sk->sk_state) {
2576	case TCP_SYN_SENT:
2577		WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2578		break;
2579	case TCP_CLOSE_WAIT:
2580		WRITE_ONCE(sk->sk_err, EPIPE);
2581		break;
2582	case TCP_CLOSE:
2583		return;
2584	default:
2585		WRITE_ONCE(sk->sk_err, ECONNRESET);
2586	}
2587
2588	mptcp_set_state(sk, TCP_CLOSE);
2589	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2590	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2591	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2592
2593	/* the calling mptcp_worker will properly destroy the socket */
2594	if (sock_flag(sk, SOCK_DEAD))
2595		return;
2596
2597	sk->sk_state_change(sk);
2598	sk_error_report(sk);
2599}
2600
2601static void __mptcp_retrans(struct sock *sk)
2602{
2603	struct mptcp_sock *msk = mptcp_sk(sk);
2604	struct mptcp_subflow_context *subflow;
2605	struct mptcp_sendmsg_info info = {};
2606	struct mptcp_data_frag *dfrag;
2607	struct sock *ssk;
2608	int ret, err;
2609	u16 len = 0;
2610
2611	mptcp_clean_una_wakeup(sk);
2612
2613	/* first check ssk: need to kick "stale" logic */
2614	err = mptcp_sched_get_retrans(msk);
2615	dfrag = mptcp_rtx_head(sk);
2616	if (!dfrag) {
2617		if (mptcp_data_fin_enabled(msk)) {
2618			struct inet_connection_sock *icsk = inet_csk(sk);
2619
2620			icsk->icsk_retransmits++;
2621			mptcp_set_datafin_timeout(sk);
2622			mptcp_send_ack(msk);
2623
2624			goto reset_timer;
2625		}
2626
2627		if (!mptcp_send_head(sk))
2628			return;
2629
2630		goto reset_timer;
2631	}
2632
2633	if (err)
2634		goto reset_timer;
2635
2636	mptcp_for_each_subflow(msk, subflow) {
2637		if (READ_ONCE(subflow->scheduled)) {
2638			u16 copied = 0;
2639
2640			mptcp_subflow_set_scheduled(subflow, false);
2641
2642			ssk = mptcp_subflow_tcp_sock(subflow);
2643
2644			lock_sock(ssk);
2645
2646			/* limit retransmission to the bytes already sent on some subflows */
2647			info.sent = 0;
2648			info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2649								    dfrag->already_sent;
2650			while (info.sent < info.limit) {
2651				ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2652				if (ret <= 0)
2653					break;
2654
2655				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2656				copied += ret;
2657				info.sent += ret;
2658			}
2659			if (copied) {
2660				len = max(copied, len);
2661				tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2662					 info.size_goal);
2663				WRITE_ONCE(msk->allow_infinite_fallback, false);
2664			}
2665
2666			release_sock(ssk);
2667		}
2668	}
2669
2670	msk->bytes_retrans += len;
2671	dfrag->already_sent = max(dfrag->already_sent, len);
2672
2673reset_timer:
2674	mptcp_check_and_set_pending(sk);
2675
2676	if (!mptcp_rtx_timer_pending(sk))
2677		mptcp_reset_rtx_timer(sk);
2678}
2679
2680/* schedule the timeout timer for the relevant event: either close timeout
2681 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2682 */
2683void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2684{
2685	struct sock *sk = (struct sock *)msk;
2686	unsigned long timeout, close_timeout;
2687
2688	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2689		return;
2690
2691	close_timeout = inet_csk(sk)->icsk_mtup.probe_timestamp - tcp_jiffies32 + jiffies +
2692			mptcp_close_timeout(sk);
2693
2694	/* the close timeout takes precedence on the fail one, and here at least one of
2695	 * them is active
2696	 */
2697	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2698
2699	sk_reset_timer(sk, &sk->sk_timer, timeout);
2700}
2701
2702static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2703{
2704	struct sock *ssk = msk->first;
2705	bool slow;
2706
2707	if (!ssk)
2708		return;
2709
2710	pr_debug("MP_FAIL doesn't respond, reset the subflow");
2711
2712	slow = lock_sock_fast(ssk);
2713	mptcp_subflow_reset(ssk);
2714	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2715	unlock_sock_fast(ssk, slow);
2716}
2717
2718static void mptcp_do_fastclose(struct sock *sk)
2719{
2720	struct mptcp_subflow_context *subflow, *tmp;
2721	struct mptcp_sock *msk = mptcp_sk(sk);
2722
2723	mptcp_set_state(sk, TCP_CLOSE);
2724	mptcp_for_each_subflow_safe(msk, subflow, tmp)
2725		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2726				  subflow, MPTCP_CF_FASTCLOSE);
2727}
2728
2729static void mptcp_worker(struct work_struct *work)
2730{
2731	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2732	struct sock *sk = (struct sock *)msk;
2733	unsigned long fail_tout;
2734	int state;
2735
2736	lock_sock(sk);
2737	state = sk->sk_state;
2738	if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2739		goto unlock;
2740
2741	mptcp_check_fastclose(msk);
2742
2743	mptcp_pm_nl_work(msk);
2744
2745	mptcp_check_send_data_fin(sk);
2746	mptcp_check_data_fin_ack(sk);
2747	mptcp_check_data_fin(sk);
2748
2749	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2750		__mptcp_close_subflow(sk);
2751
2752	if (mptcp_close_tout_expired(sk)) {
2753		mptcp_do_fastclose(sk);
2754		mptcp_close_wake_up(sk);
2755	}
2756
2757	if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2758		__mptcp_destroy_sock(sk);
2759		goto unlock;
2760	}
2761
2762	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2763		__mptcp_retrans(sk);
2764
2765	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2766	if (fail_tout && time_after(jiffies, fail_tout))
2767		mptcp_mp_fail_no_response(msk);
2768
2769unlock:
2770	release_sock(sk);
2771	sock_put(sk);
2772}
2773
2774static void __mptcp_init_sock(struct sock *sk)
2775{
2776	struct mptcp_sock *msk = mptcp_sk(sk);
2777
2778	INIT_LIST_HEAD(&msk->conn_list);
2779	INIT_LIST_HEAD(&msk->join_list);
2780	INIT_LIST_HEAD(&msk->rtx_queue);
2781	INIT_WORK(&msk->work, mptcp_worker);
2782	__skb_queue_head_init(&msk->receive_queue);
2783	msk->out_of_order_queue = RB_ROOT;
2784	msk->first_pending = NULL;
2785	WRITE_ONCE(msk->rmem_fwd_alloc, 0);
2786	WRITE_ONCE(msk->rmem_released, 0);
2787	msk->timer_ival = TCP_RTO_MIN;
2788	msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2789
2790	WRITE_ONCE(msk->first, NULL);
2791	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2792	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2793	WRITE_ONCE(msk->allow_infinite_fallback, true);
2794	msk->recovery = false;
2795	msk->subflow_id = 1;
2796
2797	mptcp_pm_data_init(msk);
2798
2799	/* re-use the csk retrans timer for MPTCP-level retrans */
2800	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2801	timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2802}
2803
2804static void mptcp_ca_reset(struct sock *sk)
2805{
2806	struct inet_connection_sock *icsk = inet_csk(sk);
2807
2808	tcp_assign_congestion_control(sk);
2809	strcpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name);
2810
2811	/* no need to keep a reference to the ops, the name will suffice */
2812	tcp_cleanup_congestion_control(sk);
2813	icsk->icsk_ca_ops = NULL;
2814}
2815
2816static int mptcp_init_sock(struct sock *sk)
2817{
2818	struct net *net = sock_net(sk);
2819	int ret;
2820
2821	__mptcp_init_sock(sk);
2822
2823	if (!mptcp_is_enabled(net))
2824		return -ENOPROTOOPT;
2825
2826	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2827		return -ENOMEM;
2828
2829	ret = mptcp_init_sched(mptcp_sk(sk),
2830			       mptcp_sched_find(mptcp_get_scheduler(net)));
2831	if (ret)
2832		return ret;
2833
2834	set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2835
2836	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2837	 * propagate the correct value
2838	 */
2839	mptcp_ca_reset(sk);
2840
2841	sk_sockets_allocated_inc(sk);
2842	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2843	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2844
2845	return 0;
2846}
2847
2848static void __mptcp_clear_xmit(struct sock *sk)
2849{
2850	struct mptcp_sock *msk = mptcp_sk(sk);
2851	struct mptcp_data_frag *dtmp, *dfrag;
2852
2853	WRITE_ONCE(msk->first_pending, NULL);
2854	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2855		dfrag_clear(sk, dfrag);
2856}
2857
2858void mptcp_cancel_work(struct sock *sk)
2859{
2860	struct mptcp_sock *msk = mptcp_sk(sk);
2861
2862	if (cancel_work_sync(&msk->work))
2863		__sock_put(sk);
2864}
2865
2866void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2867{
2868	lock_sock(ssk);
2869
2870	switch (ssk->sk_state) {
2871	case TCP_LISTEN:
2872		if (!(how & RCV_SHUTDOWN))
2873			break;
2874		fallthrough;
2875	case TCP_SYN_SENT:
2876		WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2877		break;
2878	default:
2879		if (__mptcp_check_fallback(mptcp_sk(sk))) {
2880			pr_debug("Fallback");
2881			ssk->sk_shutdown |= how;
2882			tcp_shutdown(ssk, how);
2883
2884			/* simulate the data_fin ack reception to let the state
2885			 * machine move forward
2886			 */
2887			WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2888			mptcp_schedule_work(sk);
2889		} else {
2890			pr_debug("Sending DATA_FIN on subflow %p", ssk);
2891			tcp_send_ack(ssk);
2892			if (!mptcp_rtx_timer_pending(sk))
2893				mptcp_reset_rtx_timer(sk);
2894		}
2895		break;
2896	}
2897
2898	release_sock(ssk);
2899}
2900
2901void mptcp_set_state(struct sock *sk, int state)
2902{
2903	int oldstate = sk->sk_state;
2904
2905	switch (state) {
2906	case TCP_ESTABLISHED:
2907		if (oldstate != TCP_ESTABLISHED)
2908			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2909		break;
2910
2911	default:
2912		if (oldstate == TCP_ESTABLISHED)
2913			MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2914	}
2915
2916	inet_sk_state_store(sk, state);
2917}
2918
2919static const unsigned char new_state[16] = {
2920	/* current state:     new state:      action:	*/
2921	[0 /* (Invalid) */] = TCP_CLOSE,
2922	[TCP_ESTABLISHED]   = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2923	[TCP_SYN_SENT]      = TCP_CLOSE,
2924	[TCP_SYN_RECV]      = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2925	[TCP_FIN_WAIT1]     = TCP_FIN_WAIT1,
2926	[TCP_FIN_WAIT2]     = TCP_FIN_WAIT2,
2927	[TCP_TIME_WAIT]     = TCP_CLOSE,	/* should not happen ! */
2928	[TCP_CLOSE]         = TCP_CLOSE,
2929	[TCP_CLOSE_WAIT]    = TCP_LAST_ACK  | TCP_ACTION_FIN,
2930	[TCP_LAST_ACK]      = TCP_LAST_ACK,
2931	[TCP_LISTEN]        = TCP_CLOSE,
2932	[TCP_CLOSING]       = TCP_CLOSING,
2933	[TCP_NEW_SYN_RECV]  = TCP_CLOSE,	/* should not happen ! */
2934};
2935
2936static int mptcp_close_state(struct sock *sk)
2937{
2938	int next = (int)new_state[sk->sk_state];
2939	int ns = next & TCP_STATE_MASK;
2940
2941	mptcp_set_state(sk, ns);
2942
2943	return next & TCP_ACTION_FIN;
2944}
2945
2946static void mptcp_check_send_data_fin(struct sock *sk)
2947{
2948	struct mptcp_subflow_context *subflow;
2949	struct mptcp_sock *msk = mptcp_sk(sk);
2950
2951	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu",
2952		 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
2953		 msk->snd_nxt, msk->write_seq);
2954
2955	/* we still need to enqueue subflows or not really shutting down,
2956	 * skip this
2957	 */
2958	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
2959	    mptcp_send_head(sk))
2960		return;
2961
2962	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
2963
2964	mptcp_for_each_subflow(msk, subflow) {
2965		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2966
2967		mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
2968	}
2969}
2970
2971static void __mptcp_wr_shutdown(struct sock *sk)
2972{
2973	struct mptcp_sock *msk = mptcp_sk(sk);
2974
2975	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d",
2976		 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
2977		 !!mptcp_send_head(sk));
2978
2979	/* will be ignored by fallback sockets */
2980	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
2981	WRITE_ONCE(msk->snd_data_fin_enable, 1);
2982
2983	mptcp_check_send_data_fin(sk);
2984}
2985
2986static void __mptcp_destroy_sock(struct sock *sk)
2987{
2988	struct mptcp_sock *msk = mptcp_sk(sk);
2989
2990	pr_debug("msk=%p", msk);
2991
2992	might_sleep();
2993
2994	mptcp_stop_rtx_timer(sk);
2995	sk_stop_timer(sk, &sk->sk_timer);
2996	msk->pm.status = 0;
2997	mptcp_release_sched(msk);
2998
2999	sk->sk_prot->destroy(sk);
3000
3001	WARN_ON_ONCE(READ_ONCE(msk->rmem_fwd_alloc));
3002	WARN_ON_ONCE(msk->rmem_released);
3003	sk_stream_kill_queues(sk);
3004	xfrm_sk_free_policy(sk);
3005
3006	sock_put(sk);
3007}
3008
3009void __mptcp_unaccepted_force_close(struct sock *sk)
3010{
3011	sock_set_flag(sk, SOCK_DEAD);
3012	mptcp_do_fastclose(sk);
3013	__mptcp_destroy_sock(sk);
3014}
3015
3016static __poll_t mptcp_check_readable(struct sock *sk)
3017{
3018	return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3019}
3020
3021static void mptcp_check_listen_stop(struct sock *sk)
3022{
3023	struct sock *ssk;
3024
3025	if (inet_sk_state_load(sk) != TCP_LISTEN)
3026		return;
3027
3028	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3029	ssk = mptcp_sk(sk)->first;
3030	if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3031		return;
3032
3033	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3034	tcp_set_state(ssk, TCP_CLOSE);
3035	mptcp_subflow_queue_clean(sk, ssk);
3036	inet_csk_listen_stop(ssk);
3037	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3038	release_sock(ssk);
3039}
3040
3041bool __mptcp_close(struct sock *sk, long timeout)
3042{
3043	struct mptcp_subflow_context *subflow;
3044	struct mptcp_sock *msk = mptcp_sk(sk);
3045	bool do_cancel_work = false;
3046	int subflows_alive = 0;
3047
3048	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3049
3050	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3051		mptcp_check_listen_stop(sk);
3052		mptcp_set_state(sk, TCP_CLOSE);
3053		goto cleanup;
3054	}
3055
3056	if (mptcp_data_avail(msk) || timeout < 0) {
3057		/* If the msk has read data, or the caller explicitly ask it,
3058		 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3059		 */
3060		mptcp_do_fastclose(sk);
3061		timeout = 0;
3062	} else if (mptcp_close_state(sk)) {
3063		__mptcp_wr_shutdown(sk);
3064	}
3065
3066	sk_stream_wait_close(sk, timeout);
3067
3068cleanup:
3069	/* orphan all the subflows */
3070	mptcp_for_each_subflow(msk, subflow) {
3071		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3072		bool slow = lock_sock_fast_nested(ssk);
3073
3074		subflows_alive += ssk->sk_state != TCP_CLOSE;
3075
3076		/* since the close timeout takes precedence on the fail one,
3077		 * cancel the latter
3078		 */
3079		if (ssk == msk->first)
3080			subflow->fail_tout = 0;
3081
3082		/* detach from the parent socket, but allow data_ready to
3083		 * push incoming data into the mptcp stack, to properly ack it
3084		 */
3085		ssk->sk_socket = NULL;
3086		ssk->sk_wq = NULL;
3087		unlock_sock_fast(ssk, slow);
3088	}
3089	sock_orphan(sk);
3090
3091	/* all the subflows are closed, only timeout can change the msk
3092	 * state, let's not keep resources busy for no reasons
3093	 */
3094	if (subflows_alive == 0)
3095		mptcp_set_state(sk, TCP_CLOSE);
3096
3097	sock_hold(sk);
3098	pr_debug("msk=%p state=%d", sk, sk->sk_state);
3099	if (msk->token)
3100		mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3101
3102	if (sk->sk_state == TCP_CLOSE) {
3103		__mptcp_destroy_sock(sk);
3104		do_cancel_work = true;
3105	} else {
3106		mptcp_start_tout_timer(sk);
3107	}
3108
3109	return do_cancel_work;
3110}
3111
3112static void mptcp_close(struct sock *sk, long timeout)
3113{
3114	bool do_cancel_work;
3115
3116	lock_sock(sk);
3117
3118	do_cancel_work = __mptcp_close(sk, timeout);
3119	release_sock(sk);
3120	if (do_cancel_work)
3121		mptcp_cancel_work(sk);
3122
3123	sock_put(sk);
3124}
3125
3126static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3127{
3128#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3129	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3130	struct ipv6_pinfo *msk6 = inet6_sk(msk);
3131
3132	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3133	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3134
3135	if (msk6 && ssk6) {
3136		msk6->saddr = ssk6->saddr;
3137		msk6->flow_label = ssk6->flow_label;
3138	}
3139#endif
3140
3141	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3142	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3143	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3144	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3145	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3146	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3147}
3148
3149static int mptcp_disconnect(struct sock *sk, int flags)
3150{
3151	struct mptcp_sock *msk = mptcp_sk(sk);
3152
3153	/* We are on the fastopen error path. We can't call straight into the
3154	 * subflows cleanup code due to lock nesting (we are already under
3155	 * msk->firstsocket lock).
3156	 */
3157	if (msk->fastopening)
3158		return -EBUSY;
3159
3160	mptcp_check_listen_stop(sk);
3161	mptcp_set_state(sk, TCP_CLOSE);
3162
3163	mptcp_stop_rtx_timer(sk);
3164	mptcp_stop_tout_timer(sk);
3165
3166	if (msk->token)
3167		mptcp_event(MPTCP_EVENT_CLOSED, msk, NULL, GFP_KERNEL);
3168
3169	/* msk->subflow is still intact, the following will not free the first
3170	 * subflow
3171	 */
3172	mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3173	WRITE_ONCE(msk->flags, 0);
3174	msk->cb_flags = 0;
3175	msk->recovery = false;
3176	WRITE_ONCE(msk->can_ack, false);
3177	WRITE_ONCE(msk->fully_established, false);
3178	WRITE_ONCE(msk->rcv_data_fin, false);
3179	WRITE_ONCE(msk->snd_data_fin_enable, false);
3180	WRITE_ONCE(msk->rcv_fastclose, false);
3181	WRITE_ONCE(msk->use_64bit_ack, false);
3182	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3183	mptcp_pm_data_reset(msk);
3184	mptcp_ca_reset(sk);
3185	msk->bytes_consumed = 0;
3186	msk->bytes_acked = 0;
3187	msk->bytes_received = 0;
3188	msk->bytes_sent = 0;
3189	msk->bytes_retrans = 0;
3190	msk->rcvspace_init = 0;
3191
3192	WRITE_ONCE(sk->sk_shutdown, 0);
3193	sk_error_report(sk);
3194	return 0;
3195}
3196
3197#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3198static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3199{
3200	unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3201
3202	return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3203}
3204
3205static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3206{
3207	const struct ipv6_pinfo *np = inet6_sk(sk);
3208	struct ipv6_txoptions *opt;
3209	struct ipv6_pinfo *newnp;
3210
3211	newnp = inet6_sk(newsk);
3212
3213	rcu_read_lock();
3214	opt = rcu_dereference(np->opt);
3215	if (opt) {
3216		opt = ipv6_dup_options(newsk, opt);
3217		if (!opt)
3218			net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3219	}
3220	RCU_INIT_POINTER(newnp->opt, opt);
3221	rcu_read_unlock();
3222}
3223#endif
3224
3225static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3226{
3227	struct ip_options_rcu *inet_opt, *newopt = NULL;
3228	const struct inet_sock *inet = inet_sk(sk);
3229	struct inet_sock *newinet;
3230
3231	newinet = inet_sk(newsk);
3232
3233	rcu_read_lock();
3234	inet_opt = rcu_dereference(inet->inet_opt);
3235	if (inet_opt) {
3236		newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3237				      inet_opt->opt.optlen, GFP_ATOMIC);
3238		if (newopt)
3239			memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3240			       inet_opt->opt.optlen);
3241		else
3242			net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3243	}
3244	RCU_INIT_POINTER(newinet->inet_opt, newopt);
3245	rcu_read_unlock();
3246}
3247
3248struct sock *mptcp_sk_clone_init(const struct sock *sk,
3249				 const struct mptcp_options_received *mp_opt,
3250				 struct sock *ssk,
3251				 struct request_sock *req)
3252{
3253	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3254	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3255	struct mptcp_subflow_context *subflow;
3256	struct mptcp_sock *msk;
3257
3258	if (!nsk)
3259		return NULL;
3260
3261#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3262	if (nsk->sk_family == AF_INET6)
3263		inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3264#endif
3265
3266	__mptcp_init_sock(nsk);
3267
3268#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3269	if (nsk->sk_family == AF_INET6)
3270		mptcp_copy_ip6_options(nsk, sk);
3271	else
3272#endif
3273		mptcp_copy_ip_options(nsk, sk);
3274
3275	msk = mptcp_sk(nsk);
3276	WRITE_ONCE(msk->local_key, subflow_req->local_key);
3277	WRITE_ONCE(msk->token, subflow_req->token);
3278	msk->in_accept_queue = 1;
3279	WRITE_ONCE(msk->fully_established, false);
3280	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3281		WRITE_ONCE(msk->csum_enabled, true);
3282
3283	WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3284	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3285	WRITE_ONCE(msk->snd_una, msk->write_seq);
3286	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3287	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3288	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3289
3290	/* passive msk is created after the first/MPC subflow */
3291	msk->subflow_id = 2;
3292
3293	sock_reset_flag(nsk, SOCK_RCU_FREE);
3294	security_inet_csk_clone(nsk, req);
3295
3296	/* this can't race with mptcp_close(), as the msk is
3297	 * not yet exposted to user-space
3298	 */
3299	mptcp_set_state(nsk, TCP_ESTABLISHED);
3300
3301	/* The msk maintain a ref to each subflow in the connections list */
3302	WRITE_ONCE(msk->first, ssk);
3303	subflow = mptcp_subflow_ctx(ssk);
3304	list_add(&subflow->node, &msk->conn_list);
3305	sock_hold(ssk);
3306
3307	/* new mpc subflow takes ownership of the newly
3308	 * created mptcp socket
3309	 */
3310	mptcp_token_accept(subflow_req, msk);
3311
3312	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3313	 * uses the correct data
3314	 */
3315	mptcp_copy_inaddrs(nsk, ssk);
3316	__mptcp_propagate_sndbuf(nsk, ssk);
3317
3318	mptcp_rcv_space_init(msk, ssk);
3319
3320	if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3321		__mptcp_subflow_fully_established(msk, subflow, mp_opt);
3322	bh_unlock_sock(nsk);
3323
3324	/* note: the newly allocated socket refcount is 2 now */
3325	return nsk;
3326}
3327
3328void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3329{
3330	const struct tcp_sock *tp = tcp_sk(ssk);
3331
3332	msk->rcvspace_init = 1;
3333	msk->rcvq_space.copied = 0;
3334	msk->rcvq_space.rtt_us = 0;
3335
3336	msk->rcvq_space.time = tp->tcp_mstamp;
3337
3338	/* initial rcv_space offering made to peer */
3339	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3340				      TCP_INIT_CWND * tp->advmss);
3341	if (msk->rcvq_space.space == 0)
3342		msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3343}
3344
3345void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3346{
3347	struct mptcp_subflow_context *subflow, *tmp;
3348	struct sock *sk = (struct sock *)msk;
3349
3350	__mptcp_clear_xmit(sk);
3351
3352	/* join list will be eventually flushed (with rst) at sock lock release time */
3353	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3354		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3355
3356	/* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3357	mptcp_data_lock(sk);
3358	skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3359	__skb_queue_purge(&sk->sk_receive_queue);
3360	skb_rbtree_purge(&msk->out_of_order_queue);
3361	mptcp_data_unlock(sk);
3362
3363	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3364	 * inet_sock_destruct() will dispose it
3365	 */
3366	sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3367	WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3368	mptcp_token_destroy(msk);
3369	mptcp_pm_free_anno_list(msk);
3370	mptcp_free_local_addr_list(msk);
3371}
3372
3373static void mptcp_destroy(struct sock *sk)
3374{
3375	struct mptcp_sock *msk = mptcp_sk(sk);
3376
3377	/* allow the following to close even the initial subflow */
3378	msk->free_first = 1;
3379	mptcp_destroy_common(msk, 0);
3380	sk_sockets_allocated_dec(sk);
3381}
3382
3383void __mptcp_data_acked(struct sock *sk)
3384{
3385	if (!sock_owned_by_user(sk))
3386		__mptcp_clean_una(sk);
3387	else
3388		__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3389}
3390
3391void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3392{
3393	if (!mptcp_send_head(sk))
3394		return;
3395
3396	if (!sock_owned_by_user(sk))
3397		__mptcp_subflow_push_pending(sk, ssk, false);
3398	else
3399		__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3400}
3401
3402#define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3403				      BIT(MPTCP_RETRANSMIT) | \
3404				      BIT(MPTCP_FLUSH_JOIN_LIST))
3405
3406/* processes deferred events and flush wmem */
3407static void mptcp_release_cb(struct sock *sk)
3408	__must_hold(&sk->sk_lock.slock)
3409{
3410	struct mptcp_sock *msk = mptcp_sk(sk);
3411
3412	for (;;) {
3413		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3414		struct list_head join_list;
3415
3416		if (!flags)
3417			break;
3418
3419		INIT_LIST_HEAD(&join_list);
3420		list_splice_init(&msk->join_list, &join_list);
3421
3422		/* the following actions acquire the subflow socket lock
3423		 *
3424		 * 1) can't be invoked in atomic scope
3425		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3426		 *    datapath acquires the msk socket spinlock while helding
3427		 *    the subflow socket lock
3428		 */
3429		msk->cb_flags &= ~flags;
3430		spin_unlock_bh(&sk->sk_lock.slock);
3431
3432		if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3433			__mptcp_flush_join_list(sk, &join_list);
3434		if (flags & BIT(MPTCP_PUSH_PENDING))
3435			__mptcp_push_pending(sk, 0);
3436		if (flags & BIT(MPTCP_RETRANSMIT))
3437			__mptcp_retrans(sk);
3438
3439		cond_resched();
3440		spin_lock_bh(&sk->sk_lock.slock);
3441	}
3442
3443	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3444		__mptcp_clean_una_wakeup(sk);
3445	if (unlikely(msk->cb_flags)) {
3446		/* be sure to sync the msk state before taking actions
3447		 * depending on sk_state (MPTCP_ERROR_REPORT)
3448		 * On sk release avoid actions depending on the first subflow
3449		 */
3450		if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3451			__mptcp_sync_state(sk, msk->pending_state);
3452		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3453			__mptcp_error_report(sk);
3454		if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3455			__mptcp_sync_sndbuf(sk);
3456	}
3457
3458	__mptcp_update_rmem(sk);
3459}
3460
3461/* MP_JOIN client subflow must wait for 4th ack before sending any data:
3462 * TCP can't schedule delack timer before the subflow is fully established.
3463 * MPTCP uses the delack timer to do 3rd ack retransmissions
3464 */
3465static void schedule_3rdack_retransmission(struct sock *ssk)
3466{
3467	struct inet_connection_sock *icsk = inet_csk(ssk);
3468	struct tcp_sock *tp = tcp_sk(ssk);
3469	unsigned long timeout;
3470
3471	if (mptcp_subflow_ctx(ssk)->fully_established)
3472		return;
3473
3474	/* reschedule with a timeout above RTT, as we must look only for drop */
3475	if (tp->srtt_us)
3476		timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3477	else
3478		timeout = TCP_TIMEOUT_INIT;
3479	timeout += jiffies;
3480
3481	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3482	icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3483	icsk->icsk_ack.timeout = timeout;
3484	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3485}
3486
3487void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3488{
3489	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3490	struct sock *sk = subflow->conn;
3491
3492	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3493		mptcp_data_lock(sk);
3494		if (!sock_owned_by_user(sk))
3495			__mptcp_subflow_push_pending(sk, ssk, true);
3496		else
3497			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3498		mptcp_data_unlock(sk);
3499	}
3500	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3501		mptcp_data_lock(sk);
3502		if (!sock_owned_by_user(sk))
3503			__mptcp_sync_sndbuf(sk);
3504		else
3505			__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3506		mptcp_data_unlock(sk);
3507	}
3508	if (status & BIT(MPTCP_DELEGATE_ACK))
3509		schedule_3rdack_retransmission(ssk);
3510}
3511
3512static int mptcp_hash(struct sock *sk)
3513{
3514	/* should never be called,
3515	 * we hash the TCP subflows not the master socket
3516	 */
3517	WARN_ON_ONCE(1);
3518	return 0;
3519}
3520
3521static void mptcp_unhash(struct sock *sk)
3522{
3523	/* called from sk_common_release(), but nothing to do here */
3524}
3525
3526static int mptcp_get_port(struct sock *sk, unsigned short snum)
3527{
3528	struct mptcp_sock *msk = mptcp_sk(sk);
3529
3530	pr_debug("msk=%p, ssk=%p", msk, msk->first);
3531	if (WARN_ON_ONCE(!msk->first))
3532		return -EINVAL;
3533
3534	return inet_csk_get_port(msk->first, snum);
3535}
3536
3537void mptcp_finish_connect(struct sock *ssk)
3538{
3539	struct mptcp_subflow_context *subflow;
3540	struct mptcp_sock *msk;
3541	struct sock *sk;
3542
3543	subflow = mptcp_subflow_ctx(ssk);
3544	sk = subflow->conn;
3545	msk = mptcp_sk(sk);
3546
3547	pr_debug("msk=%p, token=%u", sk, subflow->token);
3548
3549	subflow->map_seq = subflow->iasn;
3550	subflow->map_subflow_seq = 1;
3551
3552	/* the socket is not connected yet, no msk/subflow ops can access/race
3553	 * accessing the field below
3554	 */
3555	WRITE_ONCE(msk->local_key, subflow->local_key);
3556
3557	mptcp_pm_new_connection(msk, ssk, 0);
3558}
3559
3560void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3561{
3562	write_lock_bh(&sk->sk_callback_lock);
3563	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3564	sk_set_socket(sk, parent);
3565	sk->sk_uid = SOCK_INODE(parent)->i_uid;
3566	write_unlock_bh(&sk->sk_callback_lock);
3567}
3568
3569bool mptcp_finish_join(struct sock *ssk)
3570{
3571	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3572	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3573	struct sock *parent = (void *)msk;
3574	bool ret = true;
3575
3576	pr_debug("msk=%p, subflow=%p", msk, subflow);
3577
3578	/* mptcp socket already closing? */
3579	if (!mptcp_is_fully_established(parent)) {
3580		subflow->reset_reason = MPTCP_RST_EMPTCP;
3581		return false;
3582	}
3583
3584	/* active subflow, already present inside the conn_list */
3585	if (!list_empty(&subflow->node)) {
3586		mptcp_subflow_joined(msk, ssk);
3587		mptcp_propagate_sndbuf(parent, ssk);
3588		return true;
3589	}
3590
3591	if (!mptcp_pm_allow_new_subflow(msk))
3592		goto err_prohibited;
3593
3594	/* If we can't acquire msk socket lock here, let the release callback
3595	 * handle it
3596	 */
3597	mptcp_data_lock(parent);
3598	if (!sock_owned_by_user(parent)) {
3599		ret = __mptcp_finish_join(msk, ssk);
3600		if (ret) {
3601			sock_hold(ssk);
3602			list_add_tail(&subflow->node, &msk->conn_list);
3603		}
3604	} else {
3605		sock_hold(ssk);
3606		list_add_tail(&subflow->node, &msk->join_list);
3607		__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3608	}
3609	mptcp_data_unlock(parent);
3610
3611	if (!ret) {
3612err_prohibited:
3613		subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3614		return false;
3615	}
3616
3617	return true;
3618}
3619
3620static void mptcp_shutdown(struct sock *sk, int how)
3621{
3622	pr_debug("sk=%p, how=%d", sk, how);
3623
3624	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3625		__mptcp_wr_shutdown(sk);
3626}
3627
3628static int mptcp_forward_alloc_get(const struct sock *sk)
3629{
3630	return READ_ONCE(sk->sk_forward_alloc) +
3631	       READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3632}
3633
3634static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3635{
3636	const struct sock *sk = (void *)msk;
3637	u64 delta;
3638
3639	if (sk->sk_state == TCP_LISTEN)
3640		return -EINVAL;
3641
3642	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3643		return 0;
3644
3645	delta = msk->write_seq - v;
3646	if (__mptcp_check_fallback(msk) && msk->first) {
3647		struct tcp_sock *tp = tcp_sk(msk->first);
3648
3649		/* the first subflow is disconnected after close - see
3650		 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3651		 * so ignore that status, too.
3652		 */
3653		if (!((1 << msk->first->sk_state) &
3654		      (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3655			delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3656	}
3657	if (delta > INT_MAX)
3658		delta = INT_MAX;
3659
3660	return (int)delta;
3661}
3662
3663static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3664{
3665	struct mptcp_sock *msk = mptcp_sk(sk);
3666	bool slow;
3667
3668	switch (cmd) {
3669	case SIOCINQ:
3670		if (sk->sk_state == TCP_LISTEN)
3671			return -EINVAL;
3672
3673		lock_sock(sk);
3674		__mptcp_move_skbs(msk);
3675		*karg = mptcp_inq_hint(sk);
3676		release_sock(sk);
3677		break;
3678	case SIOCOUTQ:
3679		slow = lock_sock_fast(sk);
3680		*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3681		unlock_sock_fast(sk, slow);
3682		break;
3683	case SIOCOUTQNSD:
3684		slow = lock_sock_fast(sk);
3685		*karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3686		unlock_sock_fast(sk, slow);
3687		break;
3688	default:
3689		return -ENOIOCTLCMD;
3690	}
3691
3692	return 0;
3693}
3694
3695static void mptcp_subflow_early_fallback(struct mptcp_sock *msk,
3696					 struct mptcp_subflow_context *subflow)
3697{
3698	subflow->request_mptcp = 0;
3699	__mptcp_do_fallback(msk);
3700}
3701
3702static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3703{
3704	struct mptcp_subflow_context *subflow;
3705	struct mptcp_sock *msk = mptcp_sk(sk);
3706	int err = -EINVAL;
3707	struct sock *ssk;
3708
3709	ssk = __mptcp_nmpc_sk(msk);
3710	if (IS_ERR(ssk))
3711		return PTR_ERR(ssk);
3712
3713	mptcp_set_state(sk, TCP_SYN_SENT);
3714	subflow = mptcp_subflow_ctx(ssk);
3715#ifdef CONFIG_TCP_MD5SIG
3716	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3717	 * TCP option space.
3718	 */
3719	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3720		mptcp_subflow_early_fallback(msk, subflow);
3721#endif
3722	if (subflow->request_mptcp && mptcp_token_new_connect(ssk)) {
3723		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3724		mptcp_subflow_early_fallback(msk, subflow);
3725	}
3726
3727	WRITE_ONCE(msk->write_seq, subflow->idsn);
3728	WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3729	if (likely(!__mptcp_check_fallback(msk)))
3730		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3731
3732	/* if reaching here via the fastopen/sendmsg path, the caller already
3733	 * acquired the subflow socket lock, too.
3734	 */
3735	if (!msk->fastopening)
3736		lock_sock(ssk);
3737
3738	/* the following mirrors closely a very small chunk of code from
3739	 * __inet_stream_connect()
3740	 */
3741	if (ssk->sk_state != TCP_CLOSE)
3742		goto out;
3743
3744	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3745		err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3746		if (err)
3747			goto out;
3748	}
3749
3750	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3751	if (err < 0)
3752		goto out;
3753
3754	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3755
3756out:
3757	if (!msk->fastopening)
3758		release_sock(ssk);
3759
3760	/* on successful connect, the msk state will be moved to established by
3761	 * subflow_finish_connect()
3762	 */
3763	if (unlikely(err)) {
3764		/* avoid leaving a dangling token in an unconnected socket */
3765		mptcp_token_destroy(msk);
3766		mptcp_set_state(sk, TCP_CLOSE);
3767		return err;
3768	}
3769
3770	mptcp_copy_inaddrs(sk, ssk);
3771	return 0;
3772}
3773
3774static struct proto mptcp_prot = {
3775	.name		= "MPTCP",
3776	.owner		= THIS_MODULE,
3777	.init		= mptcp_init_sock,
3778	.connect	= mptcp_connect,
3779	.disconnect	= mptcp_disconnect,
3780	.close		= mptcp_close,
3781	.setsockopt	= mptcp_setsockopt,
3782	.getsockopt	= mptcp_getsockopt,
3783	.shutdown	= mptcp_shutdown,
3784	.destroy	= mptcp_destroy,
3785	.sendmsg	= mptcp_sendmsg,
3786	.ioctl		= mptcp_ioctl,
3787	.recvmsg	= mptcp_recvmsg,
3788	.release_cb	= mptcp_release_cb,
3789	.hash		= mptcp_hash,
3790	.unhash		= mptcp_unhash,
3791	.get_port	= mptcp_get_port,
3792	.forward_alloc_get	= mptcp_forward_alloc_get,
3793	.stream_memory_free	= mptcp_stream_memory_free,
3794	.sockets_allocated	= &mptcp_sockets_allocated,
3795
3796	.memory_allocated	= &tcp_memory_allocated,
3797	.per_cpu_fw_alloc	= &tcp_memory_per_cpu_fw_alloc,
3798
3799	.memory_pressure	= &tcp_memory_pressure,
3800	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_wmem),
3801	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_rmem),
3802	.sysctl_mem	= sysctl_tcp_mem,
3803	.obj_size	= sizeof(struct mptcp_sock),
3804	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
3805	.no_autobind	= true,
3806};
3807
3808static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3809{
3810	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3811	struct sock *ssk, *sk = sock->sk;
3812	int err = -EINVAL;
3813
3814	lock_sock(sk);
3815	ssk = __mptcp_nmpc_sk(msk);
3816	if (IS_ERR(ssk)) {
3817		err = PTR_ERR(ssk);
3818		goto unlock;
3819	}
3820
3821	if (sk->sk_family == AF_INET)
3822		err = inet_bind_sk(ssk, uaddr, addr_len);
3823#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3824	else if (sk->sk_family == AF_INET6)
3825		err = inet6_bind_sk(ssk, uaddr, addr_len);
3826#endif
3827	if (!err)
3828		mptcp_copy_inaddrs(sk, ssk);
3829
3830unlock:
3831	release_sock(sk);
3832	return err;
3833}
3834
3835static int mptcp_listen(struct socket *sock, int backlog)
3836{
3837	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3838	struct sock *sk = sock->sk;
3839	struct sock *ssk;
3840	int err;
3841
3842	pr_debug("msk=%p", msk);
3843
3844	lock_sock(sk);
3845
3846	err = -EINVAL;
3847	if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3848		goto unlock;
3849
3850	ssk = __mptcp_nmpc_sk(msk);
3851	if (IS_ERR(ssk)) {
3852		err = PTR_ERR(ssk);
3853		goto unlock;
3854	}
3855
3856	mptcp_set_state(sk, TCP_LISTEN);
3857	sock_set_flag(sk, SOCK_RCU_FREE);
3858
3859	lock_sock(ssk);
3860	err = __inet_listen_sk(ssk, backlog);
3861	release_sock(ssk);
3862	mptcp_set_state(sk, inet_sk_state_load(ssk));
3863
3864	if (!err) {
3865		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3866		mptcp_copy_inaddrs(sk, ssk);
3867		mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3868	}
3869
3870unlock:
3871	release_sock(sk);
3872	return err;
3873}
3874
3875static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3876			       int flags, bool kern)
3877{
3878	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3879	struct sock *ssk, *newsk;
3880	int err;
3881
3882	pr_debug("msk=%p", msk);
3883
3884	/* Buggy applications can call accept on socket states other then LISTEN
3885	 * but no need to allocate the first subflow just to error out.
3886	 */
3887	ssk = READ_ONCE(msk->first);
3888	if (!ssk)
3889		return -EINVAL;
3890
3891	pr_debug("ssk=%p, listener=%p", ssk, mptcp_subflow_ctx(ssk));
3892	newsk = inet_csk_accept(ssk, flags, &err, kern);
3893	if (!newsk)
3894		return err;
3895
3896	pr_debug("newsk=%p, subflow is mptcp=%d", newsk, sk_is_mptcp(newsk));
3897	if (sk_is_mptcp(newsk)) {
3898		struct mptcp_subflow_context *subflow;
3899		struct sock *new_mptcp_sock;
3900
3901		subflow = mptcp_subflow_ctx(newsk);
3902		new_mptcp_sock = subflow->conn;
3903
3904		/* is_mptcp should be false if subflow->conn is missing, see
3905		 * subflow_syn_recv_sock()
3906		 */
3907		if (WARN_ON_ONCE(!new_mptcp_sock)) {
3908			tcp_sk(newsk)->is_mptcp = 0;
3909			goto tcpfallback;
3910		}
3911
3912		newsk = new_mptcp_sock;
3913		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3914
3915		newsk->sk_kern_sock = kern;
3916		lock_sock(newsk);
3917		__inet_accept(sock, newsock, newsk);
3918
3919		set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3920		msk = mptcp_sk(newsk);
3921		msk->in_accept_queue = 0;
3922
3923		/* set ssk->sk_socket of accept()ed flows to mptcp socket.
3924		 * This is needed so NOSPACE flag can be set from tcp stack.
3925		 */
3926		mptcp_for_each_subflow(msk, subflow) {
3927			struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3928
3929			if (!ssk->sk_socket)
3930				mptcp_sock_graft(ssk, newsock);
3931		}
3932
3933		/* Do late cleanup for the first subflow as necessary. Also
3934		 * deal with bad peers not doing a complete shutdown.
3935		 */
3936		if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3937			__mptcp_close_ssk(newsk, msk->first,
3938					  mptcp_subflow_ctx(msk->first), 0);
3939			if (unlikely(list_is_singular(&msk->conn_list)))
3940				mptcp_set_state(newsk, TCP_CLOSE);
3941		}
3942	} else {
3943tcpfallback:
3944		newsk->sk_kern_sock = kern;
3945		lock_sock(newsk);
3946		__inet_accept(sock, newsock, newsk);
3947		/* we are being invoked after accepting a non-mp-capable
3948		 * flow: sk is a tcp_sk, not an mptcp one.
3949		 *
3950		 * Hand the socket over to tcp so all further socket ops
3951		 * bypass mptcp.
3952		 */
3953		WRITE_ONCE(newsock->sk->sk_socket->ops,
3954			   mptcp_fallback_tcp_ops(newsock->sk));
3955	}
3956	release_sock(newsk);
3957
3958	return 0;
3959}
3960
3961static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
3962{
3963	struct sock *sk = (struct sock *)msk;
3964
3965	if (__mptcp_stream_is_writeable(sk, 1))
3966		return EPOLLOUT | EPOLLWRNORM;
3967
3968	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
3969	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
3970	if (__mptcp_stream_is_writeable(sk, 1))
3971		return EPOLLOUT | EPOLLWRNORM;
3972
3973	return 0;
3974}
3975
3976static __poll_t mptcp_poll(struct file *file, struct socket *sock,
3977			   struct poll_table_struct *wait)
3978{
3979	struct sock *sk = sock->sk;
3980	struct mptcp_sock *msk;
3981	__poll_t mask = 0;
3982	u8 shutdown;
3983	int state;
3984
3985	msk = mptcp_sk(sk);
3986	sock_poll_wait(file, sock, wait);
3987
3988	state = inet_sk_state_load(sk);
3989	pr_debug("msk=%p state=%d flags=%lx", msk, state, msk->flags);
3990	if (state == TCP_LISTEN) {
3991		struct sock *ssk = READ_ONCE(msk->first);
3992
3993		if (WARN_ON_ONCE(!ssk))
3994			return 0;
3995
3996		return inet_csk_listen_poll(ssk);
3997	}
3998
3999	shutdown = READ_ONCE(sk->sk_shutdown);
4000	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4001		mask |= EPOLLHUP;
4002	if (shutdown & RCV_SHUTDOWN)
4003		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4004
4005	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4006		mask |= mptcp_check_readable(sk);
4007		if (shutdown & SEND_SHUTDOWN)
4008			mask |= EPOLLOUT | EPOLLWRNORM;
4009		else
4010			mask |= mptcp_check_writeable(msk);
4011	} else if (state == TCP_SYN_SENT &&
4012		   inet_test_bit(DEFER_CONNECT, sk)) {
4013		/* cf tcp_poll() note about TFO */
4014		mask |= EPOLLOUT | EPOLLWRNORM;
4015	}
4016
4017	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4018	smp_rmb();
4019	if (READ_ONCE(sk->sk_err))
4020		mask |= EPOLLERR;
4021
4022	return mask;
4023}
4024
4025static const struct proto_ops mptcp_stream_ops = {
4026	.family		   = PF_INET,
4027	.owner		   = THIS_MODULE,
4028	.release	   = inet_release,
4029	.bind		   = mptcp_bind,
4030	.connect	   = inet_stream_connect,
4031	.socketpair	   = sock_no_socketpair,
4032	.accept		   = mptcp_stream_accept,
4033	.getname	   = inet_getname,
4034	.poll		   = mptcp_poll,
4035	.ioctl		   = inet_ioctl,
4036	.gettstamp	   = sock_gettstamp,
4037	.listen		   = mptcp_listen,
4038	.shutdown	   = inet_shutdown,
4039	.setsockopt	   = sock_common_setsockopt,
4040	.getsockopt	   = sock_common_getsockopt,
4041	.sendmsg	   = inet_sendmsg,
4042	.recvmsg	   = inet_recvmsg,
4043	.mmap		   = sock_no_mmap,
4044	.set_rcvlowat	   = mptcp_set_rcvlowat,
4045};
4046
4047static struct inet_protosw mptcp_protosw = {
4048	.type		= SOCK_STREAM,
4049	.protocol	= IPPROTO_MPTCP,
4050	.prot		= &mptcp_prot,
4051	.ops		= &mptcp_stream_ops,
4052	.flags		= INET_PROTOSW_ICSK,
4053};
4054
4055static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4056{
4057	struct mptcp_delegated_action *delegated;
4058	struct mptcp_subflow_context *subflow;
4059	int work_done = 0;
4060
4061	delegated = container_of(napi, struct mptcp_delegated_action, napi);
4062	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4063		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4064
4065		bh_lock_sock_nested(ssk);
4066		if (!sock_owned_by_user(ssk)) {
4067			mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4068		} else {
4069			/* tcp_release_cb_override already processed
4070			 * the action or will do at next release_sock().
4071			 * In both case must dequeue the subflow here - on the same
4072			 * CPU that scheduled it.
4073			 */
4074			smp_wmb();
4075			clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4076		}
4077		bh_unlock_sock(ssk);
4078		sock_put(ssk);
4079
4080		if (++work_done == budget)
4081			return budget;
4082	}
4083
4084	/* always provide a 0 'work_done' argument, so that napi_complete_done
4085	 * will not try accessing the NULL napi->dev ptr
4086	 */
4087	napi_complete_done(napi, 0);
4088	return work_done;
4089}
4090
4091void __init mptcp_proto_init(void)
4092{
4093	struct mptcp_delegated_action *delegated;
4094	int cpu;
4095
4096	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4097
4098	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4099		panic("Failed to allocate MPTCP pcpu counter\n");
4100
4101	init_dummy_netdev(&mptcp_napi_dev);
4102	for_each_possible_cpu(cpu) {
4103		delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4104		INIT_LIST_HEAD(&delegated->head);
4105		netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4106				  mptcp_napi_poll);
4107		napi_enable(&delegated->napi);
4108	}
4109
4110	mptcp_subflow_init();
4111	mptcp_pm_init();
4112	mptcp_sched_init();
4113	mptcp_token_init();
4114
4115	if (proto_register(&mptcp_prot, 1) != 0)
4116		panic("Failed to register MPTCP proto.\n");
4117
4118	inet_register_protosw(&mptcp_protosw);
4119
4120	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4121}
4122
4123#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4124static const struct proto_ops mptcp_v6_stream_ops = {
4125	.family		   = PF_INET6,
4126	.owner		   = THIS_MODULE,
4127	.release	   = inet6_release,
4128	.bind		   = mptcp_bind,
4129	.connect	   = inet_stream_connect,
4130	.socketpair	   = sock_no_socketpair,
4131	.accept		   = mptcp_stream_accept,
4132	.getname	   = inet6_getname,
4133	.poll		   = mptcp_poll,
4134	.ioctl		   = inet6_ioctl,
4135	.gettstamp	   = sock_gettstamp,
4136	.listen		   = mptcp_listen,
4137	.shutdown	   = inet_shutdown,
4138	.setsockopt	   = sock_common_setsockopt,
4139	.getsockopt	   = sock_common_getsockopt,
4140	.sendmsg	   = inet6_sendmsg,
4141	.recvmsg	   = inet6_recvmsg,
4142	.mmap		   = sock_no_mmap,
4143#ifdef CONFIG_COMPAT
4144	.compat_ioctl	   = inet6_compat_ioctl,
4145#endif
4146	.set_rcvlowat	   = mptcp_set_rcvlowat,
4147};
4148
4149static struct proto mptcp_v6_prot;
4150
4151static struct inet_protosw mptcp_v6_protosw = {
4152	.type		= SOCK_STREAM,
4153	.protocol	= IPPROTO_MPTCP,
4154	.prot		= &mptcp_v6_prot,
4155	.ops		= &mptcp_v6_stream_ops,
4156	.flags		= INET_PROTOSW_ICSK,
4157};
4158
4159int __init mptcp_proto_v6_init(void)
4160{
4161	int err;
4162
4163	mptcp_v6_prot = mptcp_prot;
4164	strcpy(mptcp_v6_prot.name, "MPTCPv6");
4165	mptcp_v6_prot.slab = NULL;
4166	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4167	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4168
4169	err = proto_register(&mptcp_v6_prot, 1);
4170	if (err)
4171		return err;
4172
4173	err = inet6_register_protosw(&mptcp_v6_protosw);
4174	if (err)
4175		proto_unregister(&mptcp_v6_prot);
4176
4177	return err;
4178}
4179#endif