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/hotdata.h>
  24#include <net/xfrm.h>
  25#include <asm/ioctls.h>
  26#include "protocol.h"
  27#include "mib.h"
  28
  29#define CREATE_TRACE_POINTS
  30#include <trace/events/mptcp.h>
  31
  32#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  33struct mptcp6_sock {
  34	struct mptcp_sock msk;
  35	struct ipv6_pinfo np;
  36};
  37#endif
  38
  39enum {
  40	MPTCP_CMSG_TS = BIT(0),
  41	MPTCP_CMSG_INQ = BIT(1),
  42};
  43
  44static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
  45
  46static void __mptcp_destroy_sock(struct sock *sk);
  47static void mptcp_check_send_data_fin(struct sock *sk);
  48
  49DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions);
  50static struct net_device mptcp_napi_dev;
  51
  52/* Returns end sequence number of the receiver's advertised window */
  53static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
  54{
  55	return READ_ONCE(msk->wnd_end);
  56}
  57
  58static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
  59{
  60#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  61	if (sk->sk_prot == &tcpv6_prot)
  62		return &inet6_stream_ops;
  63#endif
  64	WARN_ON_ONCE(sk->sk_prot != &tcp_prot);
  65	return &inet_stream_ops;
  66}
  67
  68static int __mptcp_socket_create(struct mptcp_sock *msk)
  69{
  70	struct mptcp_subflow_context *subflow;
  71	struct sock *sk = (struct sock *)msk;
  72	struct socket *ssock;
  73	int err;
  74
  75	err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
  76	if (err)
  77		return err;
  78
  79	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
  80	WRITE_ONCE(msk->first, ssock->sk);
  81	subflow = mptcp_subflow_ctx(ssock->sk);
  82	list_add(&subflow->node, &msk->conn_list);
  83	sock_hold(ssock->sk);
  84	subflow->request_mptcp = 1;
  85	subflow->subflow_id = msk->subflow_id++;
  86
  87	/* This is the first subflow, always with id 0 */
  88	WRITE_ONCE(subflow->local_id, 0);
  89	mptcp_sock_graft(msk->first, sk->sk_socket);
  90	iput(SOCK_INODE(ssock));
  91
  92	return 0;
  93}
  94
  95/* If the MPC handshake is not started, returns the first subflow,
  96 * eventually allocating it.
  97 */
  98struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
  99{
 100	struct sock *sk = (struct sock *)msk;
 101	int ret;
 102
 103	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
 104		return ERR_PTR(-EINVAL);
 105
 106	if (!msk->first) {
 107		ret = __mptcp_socket_create(msk);
 108		if (ret)
 109			return ERR_PTR(ret);
 110	}
 111
 112	return msk->first;
 113}
 114
 115static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
 116{
 117	sk_drops_add(sk, skb);
 118	__kfree_skb(skb);
 119}
 120
 121static void mptcp_rmem_fwd_alloc_add(struct sock *sk, int size)
 122{
 123	WRITE_ONCE(mptcp_sk(sk)->rmem_fwd_alloc,
 124		   mptcp_sk(sk)->rmem_fwd_alloc + size);
 125}
 126
 127static void mptcp_rmem_charge(struct sock *sk, int size)
 128{
 129	mptcp_rmem_fwd_alloc_add(sk, -size);
 130}
 131
 132static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
 133			       struct sk_buff *from)
 134{
 135	bool fragstolen;
 136	int delta;
 137
 138	if (MPTCP_SKB_CB(from)->offset ||
 139	    ((to->len + from->len) > (sk->sk_rcvbuf >> 3)) ||
 140	    !skb_try_coalesce(to, from, &fragstolen, &delta))
 141		return false;
 142
 143	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n",
 144		 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
 145		 to->len, MPTCP_SKB_CB(from)->end_seq);
 146	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
 147
 148	/* note the fwd memory can reach a negative value after accounting
 149	 * for the delta, but the later skb free will restore a non
 150	 * negative one
 151	 */
 152	atomic_add(delta, &sk->sk_rmem_alloc);
 153	mptcp_rmem_charge(sk, delta);
 154	kfree_skb_partial(from, fragstolen);
 155
 156	return true;
 157}
 158
 159static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
 160				   struct sk_buff *from)
 161{
 162	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
 163		return false;
 164
 165	return mptcp_try_coalesce((struct sock *)msk, to, from);
 166}
 167
 168static void __mptcp_rmem_reclaim(struct sock *sk, int amount)
 169{
 170	amount >>= PAGE_SHIFT;
 171	mptcp_rmem_charge(sk, amount << PAGE_SHIFT);
 172	__sk_mem_reduce_allocated(sk, amount);
 173}
 174
 175static void mptcp_rmem_uncharge(struct sock *sk, int size)
 176{
 177	struct mptcp_sock *msk = mptcp_sk(sk);
 178	int reclaimable;
 179
 180	mptcp_rmem_fwd_alloc_add(sk, size);
 181	reclaimable = msk->rmem_fwd_alloc - sk_unused_reserved_mem(sk);
 182
 183	/* see sk_mem_uncharge() for the rationale behind the following schema */
 184	if (unlikely(reclaimable >= PAGE_SIZE))
 185		__mptcp_rmem_reclaim(sk, reclaimable);
 186}
 187
 188static void mptcp_rfree(struct sk_buff *skb)
 189{
 190	unsigned int len = skb->truesize;
 191	struct sock *sk = skb->sk;
 192
 193	atomic_sub(len, &sk->sk_rmem_alloc);
 194	mptcp_rmem_uncharge(sk, len);
 195}
 196
 197void mptcp_set_owner_r(struct sk_buff *skb, struct sock *sk)
 198{
 199	skb_orphan(skb);
 200	skb->sk = sk;
 201	skb->destructor = mptcp_rfree;
 202	atomic_add(skb->truesize, &sk->sk_rmem_alloc);
 203	mptcp_rmem_charge(sk, skb->truesize);
 204}
 205
 206/* "inspired" by tcp_data_queue_ofo(), main differences:
 207 * - use mptcp seqs
 208 * - don't cope with sacks
 209 */
 210static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
 211{
 212	struct sock *sk = (struct sock *)msk;
 213	struct rb_node **p, *parent;
 214	u64 seq, end_seq, max_seq;
 215	struct sk_buff *skb1;
 216
 217	seq = MPTCP_SKB_CB(skb)->map_seq;
 218	end_seq = MPTCP_SKB_CB(skb)->end_seq;
 219	max_seq = atomic64_read(&msk->rcv_wnd_sent);
 220
 221	pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq,
 222		 RB_EMPTY_ROOT(&msk->out_of_order_queue));
 223	if (after64(end_seq, max_seq)) {
 224		/* out of window */
 225		mptcp_drop(sk, skb);
 226		pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
 227			 (unsigned long long)end_seq - (unsigned long)max_seq,
 228			 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
 229		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
 230		return;
 231	}
 232
 233	p = &msk->out_of_order_queue.rb_node;
 234	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
 235	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
 236		rb_link_node(&skb->rbnode, NULL, p);
 237		rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 238		msk->ooo_last_skb = skb;
 239		goto end;
 240	}
 241
 242	/* with 2 subflows, adding at end of ooo queue is quite likely
 243	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
 244	 */
 245	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
 246		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 247		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 248		return;
 249	}
 250
 251	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
 252	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
 253		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 254		parent = &msk->ooo_last_skb->rbnode;
 255		p = &parent->rb_right;
 256		goto insert;
 257	}
 258
 259	/* Find place to insert this segment. Handle overlaps on the way. */
 260	parent = NULL;
 261	while (*p) {
 262		parent = *p;
 263		skb1 = rb_to_skb(parent);
 264		if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 265			p = &parent->rb_left;
 266			continue;
 267		}
 268		if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 269			if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 270				/* All the bits are present. Drop. */
 271				mptcp_drop(sk, skb);
 272				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 273				return;
 274			}
 275			if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 276				/* partial overlap:
 277				 *     |     skb      |
 278				 *  |     skb1    |
 279				 * continue traversing
 280				 */
 281			} else {
 282				/* skb's seq == skb1's seq and skb covers skb1.
 283				 * Replace skb1 with skb.
 284				 */
 285				rb_replace_node(&skb1->rbnode, &skb->rbnode,
 286						&msk->out_of_order_queue);
 287				mptcp_drop(sk, skb1);
 288				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 289				goto merge_right;
 290			}
 291		} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
 292			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 293			return;
 294		}
 295		p = &parent->rb_right;
 296	}
 297
 298insert:
 299	/* Insert segment into RB tree. */
 300	rb_link_node(&skb->rbnode, parent, p);
 301	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 302
 303merge_right:
 304	/* Remove other segments covered by skb. */
 305	while ((skb1 = skb_rb_next(skb)) != NULL) {
 306		if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
 307			break;
 308		rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
 309		mptcp_drop(sk, skb1);
 310		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 311	}
 312	/* If there is no skb after us, we are the last_skb ! */
 313	if (!skb1)
 314		msk->ooo_last_skb = skb;
 315
 316end:
 317	skb_condense(skb);
 318	mptcp_set_owner_r(skb, sk);
 319}
 320
 321static bool mptcp_rmem_schedule(struct sock *sk, struct sock *ssk, int size)
 322{
 323	struct mptcp_sock *msk = mptcp_sk(sk);
 324	int amt, amount;
 325
 326	if (size <= msk->rmem_fwd_alloc)
 327		return true;
 328
 329	size -= msk->rmem_fwd_alloc;
 330	amt = sk_mem_pages(size);
 331	amount = amt << PAGE_SHIFT;
 332	if (!__sk_mem_raise_allocated(sk, size, amt, SK_MEM_RECV))
 333		return false;
 334
 335	mptcp_rmem_fwd_alloc_add(sk, amount);
 336	return true;
 337}
 338
 339static bool __mptcp_move_skb(struct mptcp_sock *msk, struct sock *ssk,
 340			     struct sk_buff *skb, unsigned int offset,
 341			     size_t copy_len)
 342{
 343	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 344	struct sock *sk = (struct sock *)msk;
 345	struct sk_buff *tail;
 346	bool has_rxtstamp;
 347
 348	__skb_unlink(skb, &ssk->sk_receive_queue);
 349
 350	skb_ext_reset(skb);
 351	skb_orphan(skb);
 352
 353	/* try to fetch required memory from subflow */
 354	if (!mptcp_rmem_schedule(sk, ssk, skb->truesize)) {
 355		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RCVPRUNED);
 356		goto drop;
 357	}
 358
 359	has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
 360
 361	/* the skb map_seq accounts for the skb offset:
 362	 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
 363	 * value
 364	 */
 365	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
 366	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
 367	MPTCP_SKB_CB(skb)->offset = offset;
 368	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
 369
 370	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
 371		/* in sequence */
 372		msk->bytes_received += copy_len;
 373		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
 374		tail = skb_peek_tail(&sk->sk_receive_queue);
 375		if (tail && mptcp_try_coalesce(sk, tail, skb))
 376			return true;
 377
 378		mptcp_set_owner_r(skb, sk);
 379		__skb_queue_tail(&sk->sk_receive_queue, skb);
 380		return true;
 381	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
 382		mptcp_data_queue_ofo(msk, skb);
 383		return false;
 384	}
 385
 386	/* old data, keep it simple and drop the whole pkt, sender
 387	 * will retransmit as needed, if needed.
 388	 */
 389	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 390drop:
 391	mptcp_drop(sk, skb);
 392	return false;
 393}
 394
 395static void mptcp_stop_rtx_timer(struct sock *sk)
 396{
 397	struct inet_connection_sock *icsk = inet_csk(sk);
 398
 399	sk_stop_timer(sk, &icsk->icsk_retransmit_timer);
 400	mptcp_sk(sk)->timer_ival = 0;
 401}
 402
 403static void mptcp_close_wake_up(struct sock *sk)
 404{
 405	if (sock_flag(sk, SOCK_DEAD))
 406		return;
 407
 408	sk->sk_state_change(sk);
 409	if (sk->sk_shutdown == SHUTDOWN_MASK ||
 410	    sk->sk_state == TCP_CLOSE)
 411		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
 412	else
 413		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
 414}
 415
 416/* called under the msk socket lock */
 417static bool mptcp_pending_data_fin_ack(struct sock *sk)
 418{
 419	struct mptcp_sock *msk = mptcp_sk(sk);
 420
 421	return ((1 << sk->sk_state) &
 422		(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
 423	       msk->write_seq == READ_ONCE(msk->snd_una);
 424}
 425
 426static void mptcp_check_data_fin_ack(struct sock *sk)
 427{
 428	struct mptcp_sock *msk = mptcp_sk(sk);
 429
 430	/* Look for an acknowledged DATA_FIN */
 431	if (mptcp_pending_data_fin_ack(sk)) {
 432		WRITE_ONCE(msk->snd_data_fin_enable, 0);
 433
 434		switch (sk->sk_state) {
 435		case TCP_FIN_WAIT1:
 436			mptcp_set_state(sk, TCP_FIN_WAIT2);
 437			break;
 438		case TCP_CLOSING:
 439		case TCP_LAST_ACK:
 440			mptcp_set_state(sk, TCP_CLOSE);
 441			break;
 442		}
 443
 444		mptcp_close_wake_up(sk);
 445	}
 446}
 447
 448/* can be called with no lock acquired */
 449static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
 450{
 451	struct mptcp_sock *msk = mptcp_sk(sk);
 452
 453	if (READ_ONCE(msk->rcv_data_fin) &&
 454	    ((1 << inet_sk_state_load(sk)) &
 455	     (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
 456		u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
 457
 458		if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
 459			if (seq)
 460				*seq = rcv_data_fin_seq;
 461
 462			return true;
 463		}
 464	}
 465
 466	return false;
 467}
 468
 469static void mptcp_set_datafin_timeout(struct sock *sk)
 470{
 471	struct inet_connection_sock *icsk = inet_csk(sk);
 472	u32 retransmits;
 473
 474	retransmits = min_t(u32, icsk->icsk_retransmits,
 475			    ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
 476
 477	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
 478}
 479
 480static void __mptcp_set_timeout(struct sock *sk, long tout)
 481{
 482	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
 483}
 484
 485static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
 486{
 487	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 488
 489	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
 490	       inet_csk(ssk)->icsk_timeout - jiffies : 0;
 491}
 492
 493static void mptcp_set_timeout(struct sock *sk)
 494{
 495	struct mptcp_subflow_context *subflow;
 496	long tout = 0;
 497
 498	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
 499		tout = max(tout, mptcp_timeout_from_subflow(subflow));
 500	__mptcp_set_timeout(sk, tout);
 501}
 502
 503static inline bool tcp_can_send_ack(const struct sock *ssk)
 504{
 505	return !((1 << inet_sk_state_load(ssk)) &
 506	       (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
 507}
 508
 509void __mptcp_subflow_send_ack(struct sock *ssk)
 510{
 511	if (tcp_can_send_ack(ssk))
 512		tcp_send_ack(ssk);
 513}
 514
 515static void mptcp_subflow_send_ack(struct sock *ssk)
 516{
 517	bool slow;
 518
 519	slow = lock_sock_fast(ssk);
 520	__mptcp_subflow_send_ack(ssk);
 521	unlock_sock_fast(ssk, slow);
 522}
 523
 524static void mptcp_send_ack(struct mptcp_sock *msk)
 525{
 526	struct mptcp_subflow_context *subflow;
 527
 528	mptcp_for_each_subflow(msk, subflow)
 529		mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
 530}
 531
 532static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied)
 533{
 534	bool slow;
 535
 536	slow = lock_sock_fast(ssk);
 537	if (tcp_can_send_ack(ssk))
 538		tcp_cleanup_rbuf(ssk, copied);
 539	unlock_sock_fast(ssk, slow);
 540}
 541
 542static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
 543{
 544	const struct inet_connection_sock *icsk = inet_csk(ssk);
 545	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
 546	const struct tcp_sock *tp = tcp_sk(ssk);
 547
 548	return (ack_pending & ICSK_ACK_SCHED) &&
 549		((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
 550		  READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
 551		 (rx_empty && ack_pending &
 552			      (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
 553}
 554
 555static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied)
 556{
 557	int old_space = READ_ONCE(msk->old_wspace);
 558	struct mptcp_subflow_context *subflow;
 559	struct sock *sk = (struct sock *)msk;
 560	int space =  __mptcp_space(sk);
 561	bool cleanup, rx_empty;
 562
 563	cleanup = (space > 0) && (space >= (old_space << 1)) && copied;
 564	rx_empty = !__mptcp_rmem(sk) && copied;
 565
 566	mptcp_for_each_subflow(msk, subflow) {
 567		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 568
 569		if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
 570			mptcp_subflow_cleanup_rbuf(ssk, copied);
 571	}
 572}
 573
 574static bool mptcp_check_data_fin(struct sock *sk)
 575{
 576	struct mptcp_sock *msk = mptcp_sk(sk);
 577	u64 rcv_data_fin_seq;
 578	bool ret = false;
 579
 580	/* Need to ack a DATA_FIN received from a peer while this side
 581	 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
 582	 * msk->rcv_data_fin was set when parsing the incoming options
 583	 * at the subflow level and the msk lock was not held, so this
 584	 * is the first opportunity to act on the DATA_FIN and change
 585	 * the msk state.
 586	 *
 587	 * If we are caught up to the sequence number of the incoming
 588	 * DATA_FIN, send the DATA_ACK now and do state transition.  If
 589	 * not caught up, do nothing and let the recv code send DATA_ACK
 590	 * when catching up.
 591	 */
 592
 593	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
 594		WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
 595		WRITE_ONCE(msk->rcv_data_fin, 0);
 596
 597		WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
 598		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
 599
 600		switch (sk->sk_state) {
 601		case TCP_ESTABLISHED:
 602			mptcp_set_state(sk, TCP_CLOSE_WAIT);
 603			break;
 604		case TCP_FIN_WAIT1:
 605			mptcp_set_state(sk, TCP_CLOSING);
 606			break;
 607		case TCP_FIN_WAIT2:
 608			mptcp_set_state(sk, TCP_CLOSE);
 609			break;
 610		default:
 611			/* Other states not expected */
 612			WARN_ON_ONCE(1);
 613			break;
 614		}
 615
 616		ret = true;
 617		if (!__mptcp_check_fallback(msk))
 618			mptcp_send_ack(msk);
 619		mptcp_close_wake_up(sk);
 620	}
 621	return ret;
 622}
 623
 624static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk)
 625{
 626	if (READ_ONCE(msk->allow_infinite_fallback)) {
 627		MPTCP_INC_STATS(sock_net(ssk),
 628				MPTCP_MIB_DSSCORRUPTIONFALLBACK);
 629		mptcp_do_fallback(ssk);
 630	} else {
 631		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET);
 632		mptcp_subflow_reset(ssk);
 633	}
 634}
 635
 636static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
 637					   struct sock *ssk,
 638					   unsigned int *bytes)
 639{
 640	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 641	struct sock *sk = (struct sock *)msk;
 642	unsigned int moved = 0;
 643	bool more_data_avail;
 644	struct tcp_sock *tp;
 645	bool done = false;
 646	int sk_rbuf;
 647
 648	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
 649
 650	if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
 651		int ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
 652
 653		if (unlikely(ssk_rbuf > sk_rbuf)) {
 654			WRITE_ONCE(sk->sk_rcvbuf, ssk_rbuf);
 655			sk_rbuf = ssk_rbuf;
 656		}
 657	}
 658
 659	pr_debug("msk=%p ssk=%p\n", msk, ssk);
 660	tp = tcp_sk(ssk);
 661	do {
 662		u32 map_remaining, offset;
 663		u32 seq = tp->copied_seq;
 664		struct sk_buff *skb;
 665		bool fin;
 666
 667		/* try to move as much data as available */
 668		map_remaining = subflow->map_data_len -
 669				mptcp_subflow_get_map_offset(subflow);
 670
 671		skb = skb_peek(&ssk->sk_receive_queue);
 672		if (!skb) {
 673			/* With racing move_skbs_to_msk() and __mptcp_move_skbs(),
 674			 * a different CPU can have already processed the pending
 675			 * data, stop here or we can enter an infinite loop
 676			 */
 677			if (!moved)
 678				done = true;
 679			break;
 680		}
 681
 682		if (__mptcp_check_fallback(msk)) {
 683			/* Under fallback skbs have no MPTCP extension and TCP could
 684			 * collapse them between the dummy map creation and the
 685			 * current dequeue. Be sure to adjust the map size.
 686			 */
 687			map_remaining = skb->len;
 688			subflow->map_data_len = skb->len;
 689		}
 690
 691		offset = seq - TCP_SKB_CB(skb)->seq;
 692		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
 693		if (fin) {
 694			done = true;
 695			seq++;
 696		}
 697
 698		if (offset < skb->len) {
 699			size_t len = skb->len - offset;
 700
 701			if (tp->urg_data)
 702				done = true;
 703
 704			if (__mptcp_move_skb(msk, ssk, skb, offset, len))
 705				moved += len;
 706			seq += len;
 707
 708			if (unlikely(map_remaining < len)) {
 709				DEBUG_NET_WARN_ON_ONCE(1);
 710				mptcp_dss_corruption(msk, ssk);
 711			}
 712		} else {
 713			if (unlikely(!fin)) {
 714				DEBUG_NET_WARN_ON_ONCE(1);
 715				mptcp_dss_corruption(msk, ssk);
 716			}
 717
 718			sk_eat_skb(ssk, skb);
 719			done = true;
 720		}
 721
 722		WRITE_ONCE(tp->copied_seq, seq);
 723		more_data_avail = mptcp_subflow_data_available(ssk);
 724
 725		if (atomic_read(&sk->sk_rmem_alloc) > sk_rbuf) {
 726			done = true;
 727			break;
 728		}
 729	} while (more_data_avail);
 730
 731	if (moved > 0)
 732		msk->last_data_recv = tcp_jiffies32;
 733	*bytes += moved;
 734	return done;
 735}
 736
 737static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
 738{
 739	struct sock *sk = (struct sock *)msk;
 740	struct sk_buff *skb, *tail;
 741	bool moved = false;
 742	struct rb_node *p;
 743	u64 end_seq;
 744
 745	p = rb_first(&msk->out_of_order_queue);
 746	pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
 747	while (p) {
 748		skb = rb_to_skb(p);
 749		if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
 750			break;
 751
 752		p = rb_next(p);
 753		rb_erase(&skb->rbnode, &msk->out_of_order_queue);
 754
 755		if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
 756				      msk->ack_seq))) {
 757			mptcp_drop(sk, skb);
 758			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 759			continue;
 760		}
 761
 762		end_seq = MPTCP_SKB_CB(skb)->end_seq;
 763		tail = skb_peek_tail(&sk->sk_receive_queue);
 764		if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
 765			int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
 766
 767			/* skip overlapping data, if any */
 768			pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n",
 769				 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
 770				 delta);
 771			MPTCP_SKB_CB(skb)->offset += delta;
 772			MPTCP_SKB_CB(skb)->map_seq += delta;
 773			__skb_queue_tail(&sk->sk_receive_queue, skb);
 774		}
 775		msk->bytes_received += end_seq - msk->ack_seq;
 776		WRITE_ONCE(msk->ack_seq, end_seq);
 777		moved = true;
 778	}
 779	return moved;
 780}
 781
 782static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
 783{
 784	int err = sock_error(ssk);
 785	int ssk_state;
 786
 787	if (!err)
 788		return false;
 789
 790	/* only propagate errors on fallen-back sockets or
 791	 * on MPC connect
 792	 */
 793	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
 794		return false;
 795
 796	/* We need to propagate only transition to CLOSE state.
 797	 * Orphaned socket will see such state change via
 798	 * subflow_sched_work_if_closed() and that path will properly
 799	 * destroy the msk as needed.
 800	 */
 801	ssk_state = inet_sk_state_load(ssk);
 802	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
 803		mptcp_set_state(sk, ssk_state);
 804	WRITE_ONCE(sk->sk_err, -err);
 805
 806	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
 807	smp_wmb();
 808	sk_error_report(sk);
 809	return true;
 810}
 811
 812void __mptcp_error_report(struct sock *sk)
 813{
 814	struct mptcp_subflow_context *subflow;
 815	struct mptcp_sock *msk = mptcp_sk(sk);
 816
 817	mptcp_for_each_subflow(msk, subflow)
 818		if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
 819			break;
 820}
 821
 822/* In most cases we will be able to lock the mptcp socket.  If its already
 823 * owned, we need to defer to the work queue to avoid ABBA deadlock.
 824 */
 825static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
 826{
 827	struct sock *sk = (struct sock *)msk;
 828	unsigned int moved = 0;
 829
 830	__mptcp_move_skbs_from_subflow(msk, ssk, &moved);
 831	__mptcp_ofo_queue(msk);
 832	if (unlikely(ssk->sk_err)) {
 833		if (!sock_owned_by_user(sk))
 834			__mptcp_error_report(sk);
 835		else
 836			__set_bit(MPTCP_ERROR_REPORT,  &msk->cb_flags);
 837	}
 838
 839	/* If the moves have caught up with the DATA_FIN sequence number
 840	 * it's time to ack the DATA_FIN and change socket state, but
 841	 * this is not a good place to change state. Let the workqueue
 842	 * do it.
 843	 */
 844	if (mptcp_pending_data_fin(sk, NULL))
 845		mptcp_schedule_work(sk);
 846	return moved > 0;
 847}
 848
 849void mptcp_data_ready(struct sock *sk, struct sock *ssk)
 850{
 851	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 852	struct mptcp_sock *msk = mptcp_sk(sk);
 853	int sk_rbuf, ssk_rbuf;
 854
 855	/* The peer can send data while we are shutting down this
 856	 * subflow at msk destruction time, but we must avoid enqueuing
 857	 * more data to the msk receive queue
 858	 */
 859	if (unlikely(subflow->disposable))
 860		return;
 861
 862	ssk_rbuf = READ_ONCE(ssk->sk_rcvbuf);
 863	sk_rbuf = READ_ONCE(sk->sk_rcvbuf);
 864	if (unlikely(ssk_rbuf > sk_rbuf))
 865		sk_rbuf = ssk_rbuf;
 866
 867	/* over limit? can't append more skbs to msk, Also, no need to wake-up*/
 868	if (__mptcp_rmem(sk) > sk_rbuf)
 869		return;
 870
 871	/* Wake-up the reader only for in-sequence data */
 872	mptcp_data_lock(sk);
 873	if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
 874		sk->sk_data_ready(sk);
 875	mptcp_data_unlock(sk);
 876}
 877
 878static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
 879{
 880	mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
 881	WRITE_ONCE(msk->allow_infinite_fallback, false);
 882	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
 883}
 884
 885static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
 886{
 887	struct sock *sk = (struct sock *)msk;
 888
 889	if (sk->sk_state != TCP_ESTABLISHED)
 890		return false;
 891
 892	/* attach to msk socket only after we are sure we will deal with it
 893	 * at close time
 894	 */
 895	if (sk->sk_socket && !ssk->sk_socket)
 896		mptcp_sock_graft(ssk, sk->sk_socket);
 897
 898	mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
 899	mptcp_sockopt_sync_locked(msk, ssk);
 900	mptcp_subflow_joined(msk, ssk);
 901	mptcp_stop_tout_timer(sk);
 902	__mptcp_propagate_sndbuf(sk, ssk);
 903	return true;
 904}
 905
 906static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
 907{
 908	struct mptcp_subflow_context *tmp, *subflow;
 909	struct mptcp_sock *msk = mptcp_sk(sk);
 910
 911	list_for_each_entry_safe(subflow, tmp, join_list, node) {
 912		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 913		bool slow = lock_sock_fast(ssk);
 914
 915		list_move_tail(&subflow->node, &msk->conn_list);
 916		if (!__mptcp_finish_join(msk, ssk))
 917			mptcp_subflow_reset(ssk);
 918		unlock_sock_fast(ssk, slow);
 919	}
 920}
 921
 922static bool mptcp_rtx_timer_pending(struct sock *sk)
 923{
 924	return timer_pending(&inet_csk(sk)->icsk_retransmit_timer);
 925}
 926
 927static void mptcp_reset_rtx_timer(struct sock *sk)
 928{
 929	struct inet_connection_sock *icsk = inet_csk(sk);
 930	unsigned long tout;
 931
 932	/* prevent rescheduling on close */
 933	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
 934		return;
 935
 936	tout = mptcp_sk(sk)->timer_ival;
 937	sk_reset_timer(sk, &icsk->icsk_retransmit_timer, jiffies + tout);
 938}
 939
 940bool mptcp_schedule_work(struct sock *sk)
 941{
 942	if (inet_sk_state_load(sk) != TCP_CLOSE &&
 943	    schedule_work(&mptcp_sk(sk)->work)) {
 944		/* each subflow already holds a reference to the sk, and the
 945		 * workqueue is invoked by a subflow, so sk can't go away here.
 946		 */
 947		sock_hold(sk);
 948		return true;
 949	}
 950	return false;
 951}
 952
 953static struct sock *mptcp_subflow_recv_lookup(const struct mptcp_sock *msk)
 954{
 955	struct mptcp_subflow_context *subflow;
 956
 957	msk_owned_by_me(msk);
 958
 959	mptcp_for_each_subflow(msk, subflow) {
 960		if (READ_ONCE(subflow->data_avail))
 961			return mptcp_subflow_tcp_sock(subflow);
 962	}
 963
 964	return NULL;
 965}
 966
 967static bool mptcp_skb_can_collapse_to(u64 write_seq,
 968				      const struct sk_buff *skb,
 969				      const struct mptcp_ext *mpext)
 970{
 971	if (!tcp_skb_can_collapse_to(skb))
 972		return false;
 973
 974	/* can collapse only if MPTCP level sequence is in order and this
 975	 * mapping has not been xmitted yet
 976	 */
 977	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
 978	       !mpext->frozen;
 979}
 980
 981/* we can append data to the given data frag if:
 982 * - there is space available in the backing page_frag
 983 * - the data frag tail matches the current page_frag free offset
 984 * - the data frag end sequence number matches the current write seq
 985 */
 986static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
 987				       const struct page_frag *pfrag,
 988				       const struct mptcp_data_frag *df)
 989{
 990	return df && pfrag->page == df->page &&
 991		pfrag->size - pfrag->offset > 0 &&
 992		pfrag->offset == (df->offset + df->data_len) &&
 993		df->data_seq + df->data_len == msk->write_seq;
 994}
 995
 996static void dfrag_uncharge(struct sock *sk, int len)
 997{
 998	sk_mem_uncharge(sk, len);
 999	sk_wmem_queued_add(sk, -len);
1000}
1001
1002static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1003{
1004	int len = dfrag->data_len + dfrag->overhead;
1005
1006	list_del(&dfrag->list);
1007	dfrag_uncharge(sk, len);
1008	put_page(dfrag->page);
1009}
1010
1011/* called under both the msk socket lock and the data lock */
1012static void __mptcp_clean_una(struct sock *sk)
1013{
1014	struct mptcp_sock *msk = mptcp_sk(sk);
1015	struct mptcp_data_frag *dtmp, *dfrag;
1016	u64 snd_una;
1017
1018	snd_una = msk->snd_una;
1019	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1020		if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1021			break;
1022
1023		if (unlikely(dfrag == msk->first_pending)) {
1024			/* in recovery mode can see ack after the current snd head */
1025			if (WARN_ON_ONCE(!msk->recovery))
1026				break;
1027
1028			WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1029		}
1030
1031		dfrag_clear(sk, dfrag);
1032	}
1033
1034	dfrag = mptcp_rtx_head(sk);
1035	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1036		u64 delta = snd_una - dfrag->data_seq;
1037
1038		/* prevent wrap around in recovery mode */
1039		if (unlikely(delta > dfrag->already_sent)) {
1040			if (WARN_ON_ONCE(!msk->recovery))
1041				goto out;
1042			if (WARN_ON_ONCE(delta > dfrag->data_len))
1043				goto out;
1044			dfrag->already_sent += delta - dfrag->already_sent;
1045		}
1046
1047		dfrag->data_seq += delta;
1048		dfrag->offset += delta;
1049		dfrag->data_len -= delta;
1050		dfrag->already_sent -= delta;
1051
1052		dfrag_uncharge(sk, delta);
1053	}
1054
1055	/* all retransmitted data acked, recovery completed */
1056	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1057		msk->recovery = false;
1058
1059out:
1060	if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1061		if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1062			mptcp_stop_rtx_timer(sk);
1063	} else {
1064		mptcp_reset_rtx_timer(sk);
1065	}
1066
1067	if (mptcp_pending_data_fin_ack(sk))
1068		mptcp_schedule_work(sk);
1069}
1070
1071static void __mptcp_clean_una_wakeup(struct sock *sk)
1072{
1073	lockdep_assert_held_once(&sk->sk_lock.slock);
1074
1075	__mptcp_clean_una(sk);
1076	mptcp_write_space(sk);
1077}
1078
1079static void mptcp_clean_una_wakeup(struct sock *sk)
1080{
1081	mptcp_data_lock(sk);
1082	__mptcp_clean_una_wakeup(sk);
1083	mptcp_data_unlock(sk);
1084}
1085
1086static void mptcp_enter_memory_pressure(struct sock *sk)
1087{
1088	struct mptcp_subflow_context *subflow;
1089	struct mptcp_sock *msk = mptcp_sk(sk);
1090	bool first = true;
1091
1092	mptcp_for_each_subflow(msk, subflow) {
1093		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1094
1095		if (first)
1096			tcp_enter_memory_pressure(ssk);
1097		sk_stream_moderate_sndbuf(ssk);
1098
1099		first = false;
1100	}
1101	__mptcp_sync_sndbuf(sk);
1102}
1103
1104/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1105 * data
1106 */
1107static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1108{
1109	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1110					pfrag, sk->sk_allocation)))
1111		return true;
1112
1113	mptcp_enter_memory_pressure(sk);
1114	return false;
1115}
1116
1117static struct mptcp_data_frag *
1118mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1119		      int orig_offset)
1120{
1121	int offset = ALIGN(orig_offset, sizeof(long));
1122	struct mptcp_data_frag *dfrag;
1123
1124	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1125	dfrag->data_len = 0;
1126	dfrag->data_seq = msk->write_seq;
1127	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1128	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1129	dfrag->already_sent = 0;
1130	dfrag->page = pfrag->page;
1131
1132	return dfrag;
1133}
1134
1135struct mptcp_sendmsg_info {
1136	int mss_now;
1137	int size_goal;
1138	u16 limit;
1139	u16 sent;
1140	unsigned int flags;
1141	bool data_lock_held;
1142};
1143
1144static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1145				    u64 data_seq, int avail_size)
1146{
1147	u64 window_end = mptcp_wnd_end(msk);
1148	u64 mptcp_snd_wnd;
1149
1150	if (__mptcp_check_fallback(msk))
1151		return avail_size;
1152
1153	mptcp_snd_wnd = window_end - data_seq;
1154	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1155
1156	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1157		tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1158		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1159	}
1160
1161	return avail_size;
1162}
1163
1164static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1165{
1166	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1167
1168	if (!mpext)
1169		return false;
1170	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1171	return true;
1172}
1173
1174static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1175{
1176	struct sk_buff *skb;
1177
1178	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1179	if (likely(skb)) {
1180		if (likely(__mptcp_add_ext(skb, gfp))) {
1181			skb_reserve(skb, MAX_TCP_HEADER);
1182			skb->ip_summed = CHECKSUM_PARTIAL;
1183			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1184			return skb;
1185		}
1186		__kfree_skb(skb);
1187	} else {
1188		mptcp_enter_memory_pressure(sk);
1189	}
1190	return NULL;
1191}
1192
1193static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1194{
1195	struct sk_buff *skb;
1196
1197	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1198	if (!skb)
1199		return NULL;
1200
1201	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1202		tcp_skb_entail(ssk, skb);
1203		return skb;
1204	}
1205	tcp_skb_tsorted_anchor_cleanup(skb);
1206	kfree_skb(skb);
1207	return NULL;
1208}
1209
1210static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1211{
1212	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1213
1214	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1215}
1216
1217/* note: this always recompute the csum on the whole skb, even
1218 * if we just appended a single frag. More status info needed
1219 */
1220static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1221{
1222	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1223	__wsum csum = ~csum_unfold(mpext->csum);
1224	int offset = skb->len - added;
1225
1226	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1227}
1228
1229static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1230				      struct sock *ssk,
1231				      struct mptcp_ext *mpext)
1232{
1233	if (!mpext)
1234		return;
1235
1236	mpext->infinite_map = 1;
1237	mpext->data_len = 0;
1238
1239	MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_INFINITEMAPTX);
1240	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1241	pr_fallback(msk);
1242	mptcp_do_fallback(ssk);
1243}
1244
1245#define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1246
1247static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1248			      struct mptcp_data_frag *dfrag,
1249			      struct mptcp_sendmsg_info *info)
1250{
1251	u64 data_seq = dfrag->data_seq + info->sent;
1252	int offset = dfrag->offset + info->sent;
1253	struct mptcp_sock *msk = mptcp_sk(sk);
1254	bool zero_window_probe = false;
1255	struct mptcp_ext *mpext = NULL;
1256	bool can_coalesce = false;
1257	bool reuse_skb = true;
1258	struct sk_buff *skb;
1259	size_t copy;
1260	int i;
1261
1262	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n",
1263		 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1264
1265	if (WARN_ON_ONCE(info->sent > info->limit ||
1266			 info->limit > dfrag->data_len))
1267		return 0;
1268
1269	if (unlikely(!__tcp_can_send(ssk)))
1270		return -EAGAIN;
1271
1272	/* compute send limit */
1273	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1274		ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1275	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1276	copy = info->size_goal;
1277
1278	skb = tcp_write_queue_tail(ssk);
1279	if (skb && copy > skb->len) {
1280		/* Limit the write to the size available in the
1281		 * current skb, if any, so that we create at most a new skb.
1282		 * Explicitly tells TCP internals to avoid collapsing on later
1283		 * queue management operation, to avoid breaking the ext <->
1284		 * SSN association set here
1285		 */
1286		mpext = mptcp_get_ext(skb);
1287		if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1288			TCP_SKB_CB(skb)->eor = 1;
1289			tcp_mark_push(tcp_sk(ssk), skb);
1290			goto alloc_skb;
1291		}
1292
1293		i = skb_shinfo(skb)->nr_frags;
1294		can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1295		if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1296			tcp_mark_push(tcp_sk(ssk), skb);
1297			goto alloc_skb;
1298		}
1299
1300		copy -= skb->len;
1301	} else {
1302alloc_skb:
1303		skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1304		if (!skb)
1305			return -ENOMEM;
1306
1307		i = skb_shinfo(skb)->nr_frags;
1308		reuse_skb = false;
1309		mpext = mptcp_get_ext(skb);
1310	}
1311
1312	/* Zero window and all data acked? Probe. */
1313	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1314	if (copy == 0) {
1315		u64 snd_una = READ_ONCE(msk->snd_una);
1316
1317		if (snd_una != msk->snd_nxt || tcp_write_queue_tail(ssk)) {
1318			tcp_remove_empty_skb(ssk);
1319			return 0;
1320		}
1321
1322		zero_window_probe = true;
1323		data_seq = snd_una - 1;
1324		copy = 1;
1325	}
1326
1327	copy = min_t(size_t, copy, info->limit - info->sent);
1328	if (!sk_wmem_schedule(ssk, copy)) {
1329		tcp_remove_empty_skb(ssk);
1330		return -ENOMEM;
1331	}
1332
1333	if (can_coalesce) {
1334		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1335	} else {
1336		get_page(dfrag->page);
1337		skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1338	}
1339
1340	skb->len += copy;
1341	skb->data_len += copy;
1342	skb->truesize += copy;
1343	sk_wmem_queued_add(ssk, copy);
1344	sk_mem_charge(ssk, copy);
1345	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1346	TCP_SKB_CB(skb)->end_seq += copy;
1347	tcp_skb_pcount_set(skb, 0);
1348
1349	/* on skb reuse we just need to update the DSS len */
1350	if (reuse_skb) {
1351		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1352		mpext->data_len += copy;
1353		goto out;
1354	}
1355
1356	memset(mpext, 0, sizeof(*mpext));
1357	mpext->data_seq = data_seq;
1358	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1359	mpext->data_len = copy;
1360	mpext->use_map = 1;
1361	mpext->dsn64 = 1;
1362
1363	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n",
1364		 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1365		 mpext->dsn64);
1366
1367	if (zero_window_probe) {
1368		mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1369		mpext->frozen = 1;
1370		if (READ_ONCE(msk->csum_enabled))
1371			mptcp_update_data_checksum(skb, copy);
1372		tcp_push_pending_frames(ssk);
1373		return 0;
1374	}
1375out:
1376	if (READ_ONCE(msk->csum_enabled))
1377		mptcp_update_data_checksum(skb, copy);
1378	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1379		mptcp_update_infinite_map(msk, ssk, mpext);
1380	trace_mptcp_sendmsg_frag(mpext);
1381	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1382	return copy;
1383}
1384
1385#define MPTCP_SEND_BURST_SIZE		((1 << 16) - \
1386					 sizeof(struct tcphdr) - \
1387					 MAX_TCP_OPTION_SPACE - \
1388					 sizeof(struct ipv6hdr) - \
1389					 sizeof(struct frag_hdr))
1390
1391struct subflow_send_info {
1392	struct sock *ssk;
1393	u64 linger_time;
1394};
1395
1396void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1397{
1398	if (!subflow->stale)
1399		return;
1400
1401	subflow->stale = 0;
1402	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1403}
1404
1405bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1406{
1407	if (unlikely(subflow->stale)) {
1408		u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1409
1410		if (subflow->stale_rcv_tstamp == rcv_tstamp)
1411			return false;
1412
1413		mptcp_subflow_set_active(subflow);
1414	}
1415	return __mptcp_subflow_active(subflow);
1416}
1417
1418#define SSK_MODE_ACTIVE	0
1419#define SSK_MODE_BACKUP	1
1420#define SSK_MODE_MAX	2
1421
1422/* implement the mptcp packet scheduler;
1423 * returns the subflow that will transmit the next DSS
1424 * additionally updates the rtx timeout
1425 */
1426struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1427{
1428	struct subflow_send_info send_info[SSK_MODE_MAX];
1429	struct mptcp_subflow_context *subflow;
1430	struct sock *sk = (struct sock *)msk;
1431	u32 pace, burst, wmem;
1432	int i, nr_active = 0;
1433	struct sock *ssk;
1434	u64 linger_time;
1435	long tout = 0;
1436
1437	/* pick the subflow with the lower wmem/wspace ratio */
1438	for (i = 0; i < SSK_MODE_MAX; ++i) {
1439		send_info[i].ssk = NULL;
1440		send_info[i].linger_time = -1;
1441	}
1442
1443	mptcp_for_each_subflow(msk, subflow) {
1444		bool backup = subflow->backup || subflow->request_bkup;
1445
1446		trace_mptcp_subflow_get_send(subflow);
1447		ssk =  mptcp_subflow_tcp_sock(subflow);
1448		if (!mptcp_subflow_active(subflow))
1449			continue;
1450
1451		tout = max(tout, mptcp_timeout_from_subflow(subflow));
1452		nr_active += !backup;
1453		pace = subflow->avg_pacing_rate;
1454		if (unlikely(!pace)) {
1455			/* init pacing rate from socket */
1456			subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1457			pace = subflow->avg_pacing_rate;
1458			if (!pace)
1459				continue;
1460		}
1461
1462		linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1463		if (linger_time < send_info[backup].linger_time) {
1464			send_info[backup].ssk = ssk;
1465			send_info[backup].linger_time = linger_time;
1466		}
1467	}
1468	__mptcp_set_timeout(sk, tout);
1469
1470	/* pick the best backup if no other subflow is active */
1471	if (!nr_active)
1472		send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1473
1474	/* According to the blest algorithm, to avoid HoL blocking for the
1475	 * faster flow, we need to:
1476	 * - estimate the faster flow linger time
1477	 * - use the above to estimate the amount of byte transferred
1478	 *   by the faster flow
1479	 * - check that the amount of queued data is greter than the above,
1480	 *   otherwise do not use the picked, slower, subflow
1481	 * We select the subflow with the shorter estimated time to flush
1482	 * the queued mem, which basically ensure the above. We just need
1483	 * to check that subflow has a non empty cwin.
1484	 */
1485	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1486	if (!ssk || !sk_stream_memory_free(ssk))
1487		return NULL;
1488
1489	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1490	wmem = READ_ONCE(ssk->sk_wmem_queued);
1491	if (!burst)
1492		return ssk;
1493
1494	subflow = mptcp_subflow_ctx(ssk);
1495	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1496					   READ_ONCE(ssk->sk_pacing_rate) * burst,
1497					   burst + wmem);
1498	msk->snd_burst = burst;
1499	return ssk;
1500}
1501
1502static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1503{
1504	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1505	release_sock(ssk);
1506}
1507
1508static void mptcp_update_post_push(struct mptcp_sock *msk,
1509				   struct mptcp_data_frag *dfrag,
1510				   u32 sent)
1511{
1512	u64 snd_nxt_new = dfrag->data_seq;
1513
1514	dfrag->already_sent += sent;
1515
1516	msk->snd_burst -= sent;
1517
1518	snd_nxt_new += dfrag->already_sent;
1519
1520	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1521	 * is recovering after a failover. In that event, this re-sends
1522	 * old segments.
1523	 *
1524	 * Thus compute snd_nxt_new candidate based on
1525	 * the dfrag->data_seq that was sent and the data
1526	 * that has been handed to the subflow for transmission
1527	 * and skip update in case it was old dfrag.
1528	 */
1529	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1530		msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1531		WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1532	}
1533}
1534
1535void mptcp_check_and_set_pending(struct sock *sk)
1536{
1537	if (mptcp_send_head(sk)) {
1538		mptcp_data_lock(sk);
1539		mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1540		mptcp_data_unlock(sk);
1541	}
1542}
1543
1544static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1545				  struct mptcp_sendmsg_info *info)
1546{
1547	struct mptcp_sock *msk = mptcp_sk(sk);
1548	struct mptcp_data_frag *dfrag;
1549	int len, copied = 0, err = 0;
1550
1551	while ((dfrag = mptcp_send_head(sk))) {
1552		info->sent = dfrag->already_sent;
1553		info->limit = dfrag->data_len;
1554		len = dfrag->data_len - dfrag->already_sent;
1555		while (len > 0) {
1556			int ret = 0;
1557
1558			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1559			if (ret <= 0) {
1560				err = copied ? : ret;
1561				goto out;
1562			}
1563
1564			info->sent += ret;
1565			copied += ret;
1566			len -= ret;
1567
1568			mptcp_update_post_push(msk, dfrag, ret);
1569		}
1570		WRITE_ONCE(msk->first_pending, mptcp_send_next(sk));
1571
1572		if (msk->snd_burst <= 0 ||
1573		    !sk_stream_memory_free(ssk) ||
1574		    !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1575			err = copied;
1576			goto out;
1577		}
1578		mptcp_set_timeout(sk);
1579	}
1580	err = copied;
1581
1582out:
1583	if (err > 0)
1584		msk->last_data_sent = tcp_jiffies32;
1585	return err;
1586}
1587
1588void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1589{
1590	struct sock *prev_ssk = NULL, *ssk = NULL;
1591	struct mptcp_sock *msk = mptcp_sk(sk);
1592	struct mptcp_sendmsg_info info = {
1593				.flags = flags,
1594	};
1595	bool do_check_data_fin = false;
1596	int push_count = 1;
1597
1598	while (mptcp_send_head(sk) && (push_count > 0)) {
1599		struct mptcp_subflow_context *subflow;
1600		int ret = 0;
1601
1602		if (mptcp_sched_get_send(msk))
1603			break;
1604
1605		push_count = 0;
1606
1607		mptcp_for_each_subflow(msk, subflow) {
1608			if (READ_ONCE(subflow->scheduled)) {
1609				mptcp_subflow_set_scheduled(subflow, false);
1610
1611				prev_ssk = ssk;
1612				ssk = mptcp_subflow_tcp_sock(subflow);
1613				if (ssk != prev_ssk) {
1614					/* First check. If the ssk has changed since
1615					 * the last round, release prev_ssk
1616					 */
1617					if (prev_ssk)
1618						mptcp_push_release(prev_ssk, &info);
1619
1620					/* Need to lock the new subflow only if different
1621					 * from the previous one, otherwise we are still
1622					 * helding the relevant lock
1623					 */
1624					lock_sock(ssk);
1625				}
1626
1627				push_count++;
1628
1629				ret = __subflow_push_pending(sk, ssk, &info);
1630				if (ret <= 0) {
1631					if (ret != -EAGAIN ||
1632					    (1 << ssk->sk_state) &
1633					     (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1634						push_count--;
1635					continue;
1636				}
1637				do_check_data_fin = true;
1638			}
1639		}
1640	}
1641
1642	/* at this point we held the socket lock for the last subflow we used */
1643	if (ssk)
1644		mptcp_push_release(ssk, &info);
1645
1646	/* ensure the rtx timer is running */
1647	if (!mptcp_rtx_timer_pending(sk))
1648		mptcp_reset_rtx_timer(sk);
1649	if (do_check_data_fin)
1650		mptcp_check_send_data_fin(sk);
1651}
1652
1653static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1654{
1655	struct mptcp_sock *msk = mptcp_sk(sk);
1656	struct mptcp_sendmsg_info info = {
1657		.data_lock_held = true,
1658	};
1659	bool keep_pushing = true;
1660	struct sock *xmit_ssk;
1661	int copied = 0;
1662
1663	info.flags = 0;
1664	while (mptcp_send_head(sk) && keep_pushing) {
1665		struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1666		int ret = 0;
1667
1668		/* check for a different subflow usage only after
1669		 * spooling the first chunk of data
1670		 */
1671		if (first) {
1672			mptcp_subflow_set_scheduled(subflow, false);
1673			ret = __subflow_push_pending(sk, ssk, &info);
1674			first = false;
1675			if (ret <= 0)
1676				break;
1677			copied += ret;
1678			continue;
1679		}
1680
1681		if (mptcp_sched_get_send(msk))
1682			goto out;
1683
1684		if (READ_ONCE(subflow->scheduled)) {
1685			mptcp_subflow_set_scheduled(subflow, false);
1686			ret = __subflow_push_pending(sk, ssk, &info);
1687			if (ret <= 0)
1688				keep_pushing = false;
1689			copied += ret;
1690		}
1691
1692		mptcp_for_each_subflow(msk, subflow) {
1693			if (READ_ONCE(subflow->scheduled)) {
1694				xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1695				if (xmit_ssk != ssk) {
1696					mptcp_subflow_delegate(subflow,
1697							       MPTCP_DELEGATE_SEND);
1698					keep_pushing = false;
1699				}
1700			}
1701		}
1702	}
1703
1704out:
1705	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1706	 * not going to flush it via release_sock()
1707	 */
1708	if (copied) {
1709		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1710			 info.size_goal);
1711		if (!mptcp_rtx_timer_pending(sk))
1712			mptcp_reset_rtx_timer(sk);
1713
1714		if (msk->snd_data_fin_enable &&
1715		    msk->snd_nxt + 1 == msk->write_seq)
1716			mptcp_schedule_work(sk);
1717	}
1718}
1719
1720static int mptcp_disconnect(struct sock *sk, int flags);
1721
1722static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1723				  size_t len, int *copied_syn)
1724{
1725	unsigned int saved_flags = msg->msg_flags;
1726	struct mptcp_sock *msk = mptcp_sk(sk);
1727	struct sock *ssk;
1728	int ret;
1729
1730	/* on flags based fastopen the mptcp is supposed to create the
1731	 * first subflow right now. Otherwise we are in the defer_connect
1732	 * path, and the first subflow must be already present.
1733	 * Since the defer_connect flag is cleared after the first succsful
1734	 * fastopen attempt, no need to check for additional subflow status.
1735	 */
1736	if (msg->msg_flags & MSG_FASTOPEN) {
1737		ssk = __mptcp_nmpc_sk(msk);
1738		if (IS_ERR(ssk))
1739			return PTR_ERR(ssk);
1740	}
1741	if (!msk->first)
1742		return -EINVAL;
1743
1744	ssk = msk->first;
1745
1746	lock_sock(ssk);
1747	msg->msg_flags |= MSG_DONTWAIT;
1748	msk->fastopening = 1;
1749	ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1750	msk->fastopening = 0;
1751	msg->msg_flags = saved_flags;
1752	release_sock(ssk);
1753
1754	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1755	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1756		ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1757					    msg->msg_namelen, msg->msg_flags, 1);
1758
1759		/* Keep the same behaviour of plain TCP: zero the copied bytes in
1760		 * case of any error, except timeout or signal
1761		 */
1762		if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1763			*copied_syn = 0;
1764	} else if (ret && ret != -EINPROGRESS) {
1765		/* The disconnect() op called by tcp_sendmsg_fastopen()/
1766		 * __inet_stream_connect() can fail, due to looking check,
1767		 * see mptcp_disconnect().
1768		 * Attempt it again outside the problematic scope.
1769		 */
1770		if (!mptcp_disconnect(sk, 0)) {
1771			sk->sk_disconnects++;
1772			sk->sk_socket->state = SS_UNCONNECTED;
1773		}
1774	}
1775	inet_clear_bit(DEFER_CONNECT, sk);
1776
1777	return ret;
1778}
1779
1780static int do_copy_data_nocache(struct sock *sk, int copy,
1781				struct iov_iter *from, char *to)
1782{
1783	if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1784		if (!copy_from_iter_full_nocache(to, copy, from))
1785			return -EFAULT;
1786	} else if (!copy_from_iter_full(to, copy, from)) {
1787		return -EFAULT;
1788	}
1789	return 0;
1790}
1791
1792/* open-code sk_stream_memory_free() plus sent limit computation to
1793 * avoid indirect calls in fast-path.
1794 * Called under the msk socket lock, so we can avoid a bunch of ONCE
1795 * annotations.
1796 */
1797static u32 mptcp_send_limit(const struct sock *sk)
1798{
1799	const struct mptcp_sock *msk = mptcp_sk(sk);
1800	u32 limit, not_sent;
1801
1802	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1803		return 0;
1804
1805	limit = mptcp_notsent_lowat(sk);
1806	if (limit == UINT_MAX)
1807		return UINT_MAX;
1808
1809	not_sent = msk->write_seq - msk->snd_nxt;
1810	if (not_sent >= limit)
1811		return 0;
1812
1813	return limit - not_sent;
1814}
1815
1816static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1817{
1818	struct mptcp_sock *msk = mptcp_sk(sk);
1819	struct page_frag *pfrag;
1820	size_t copied = 0;
1821	int ret = 0;
1822	long timeo;
1823
1824	/* silently ignore everything else */
1825	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1826
1827	lock_sock(sk);
1828
1829	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1830		     msg->msg_flags & MSG_FASTOPEN)) {
1831		int copied_syn = 0;
1832
1833		ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1834		copied += copied_syn;
1835		if (ret == -EINPROGRESS && copied_syn > 0)
1836			goto out;
1837		else if (ret)
1838			goto do_error;
1839	}
1840
1841	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1842
1843	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1844		ret = sk_stream_wait_connect(sk, &timeo);
1845		if (ret)
1846			goto do_error;
1847	}
1848
1849	ret = -EPIPE;
1850	if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1851		goto do_error;
1852
1853	pfrag = sk_page_frag(sk);
1854
1855	while (msg_data_left(msg)) {
1856		int total_ts, frag_truesize = 0;
1857		struct mptcp_data_frag *dfrag;
1858		bool dfrag_collapsed;
1859		size_t psize, offset;
1860		u32 copy_limit;
1861
1862		/* ensure fitting the notsent_lowat() constraint */
1863		copy_limit = mptcp_send_limit(sk);
1864		if (!copy_limit)
1865			goto wait_for_memory;
1866
1867		/* reuse tail pfrag, if possible, or carve a new one from the
1868		 * page allocator
1869		 */
1870		dfrag = mptcp_pending_tail(sk);
1871		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1872		if (!dfrag_collapsed) {
1873			if (!mptcp_page_frag_refill(sk, pfrag))
1874				goto wait_for_memory;
1875
1876			dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1877			frag_truesize = dfrag->overhead;
1878		}
1879
1880		/* we do not bound vs wspace, to allow a single packet.
1881		 * memory accounting will prevent execessive memory usage
1882		 * anyway
1883		 */
1884		offset = dfrag->offset + dfrag->data_len;
1885		psize = pfrag->size - offset;
1886		psize = min_t(size_t, psize, msg_data_left(msg));
1887		psize = min_t(size_t, psize, copy_limit);
1888		total_ts = psize + frag_truesize;
1889
1890		if (!sk_wmem_schedule(sk, total_ts))
1891			goto wait_for_memory;
1892
1893		ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1894					   page_address(dfrag->page) + offset);
1895		if (ret)
1896			goto do_error;
1897
1898		/* data successfully copied into the write queue */
1899		sk_forward_alloc_add(sk, -total_ts);
1900		copied += psize;
1901		dfrag->data_len += psize;
1902		frag_truesize += psize;
1903		pfrag->offset += frag_truesize;
1904		WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1905
1906		/* charge data on mptcp pending queue to the msk socket
1907		 * Note: we charge such data both to sk and ssk
1908		 */
1909		sk_wmem_queued_add(sk, frag_truesize);
1910		if (!dfrag_collapsed) {
1911			get_page(dfrag->page);
1912			list_add_tail(&dfrag->list, &msk->rtx_queue);
1913			if (!msk->first_pending)
1914				WRITE_ONCE(msk->first_pending, dfrag);
1915		}
1916		pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk,
1917			 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1918			 !dfrag_collapsed);
1919
1920		continue;
1921
1922wait_for_memory:
1923		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1924		__mptcp_push_pending(sk, msg->msg_flags);
1925		ret = sk_stream_wait_memory(sk, &timeo);
1926		if (ret)
1927			goto do_error;
1928	}
1929
1930	if (copied)
1931		__mptcp_push_pending(sk, msg->msg_flags);
1932
1933out:
1934	release_sock(sk);
1935	return copied;
1936
1937do_error:
1938	if (copied)
1939		goto out;
1940
1941	copied = sk_stream_error(sk, msg->msg_flags, ret);
1942	goto out;
1943}
1944
1945static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied);
1946
1947static int __mptcp_recvmsg_mskq(struct mptcp_sock *msk,
1948				struct msghdr *msg,
1949				size_t len, int flags,
1950				struct scm_timestamping_internal *tss,
1951				int *cmsg_flags)
1952{
1953	struct sk_buff *skb, *tmp;
1954	int copied = 0;
1955
1956	skb_queue_walk_safe(&msk->receive_queue, skb, tmp) {
1957		u32 offset = MPTCP_SKB_CB(skb)->offset;
1958		u32 data_len = skb->len - offset;
1959		u32 count = min_t(size_t, len - copied, data_len);
1960		int err;
1961
1962		if (!(flags & MSG_TRUNC)) {
1963			err = skb_copy_datagram_msg(skb, offset, msg, count);
1964			if (unlikely(err < 0)) {
1965				if (!copied)
1966					return err;
1967				break;
1968			}
1969		}
1970
1971		if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
1972			tcp_update_recv_tstamps(skb, tss);
1973			*cmsg_flags |= MPTCP_CMSG_TS;
1974		}
1975
1976		copied += count;
1977
1978		if (count < data_len) {
1979			if (!(flags & MSG_PEEK)) {
1980				MPTCP_SKB_CB(skb)->offset += count;
1981				MPTCP_SKB_CB(skb)->map_seq += count;
1982				msk->bytes_consumed += count;
1983			}
1984			break;
1985		}
1986
1987		if (!(flags & MSG_PEEK)) {
1988			/* we will bulk release the skb memory later */
1989			skb->destructor = NULL;
1990			WRITE_ONCE(msk->rmem_released, msk->rmem_released + skb->truesize);
1991			__skb_unlink(skb, &msk->receive_queue);
1992			__kfree_skb(skb);
1993			msk->bytes_consumed += count;
1994		}
1995
1996		if (copied >= len)
1997			break;
1998	}
1999
2000	mptcp_rcv_space_adjust(msk, copied);
2001	return copied;
2002}
2003
2004/* receive buffer autotuning.  See tcp_rcv_space_adjust for more information.
2005 *
2006 * Only difference: Use highest rtt estimate of the subflows in use.
2007 */
2008static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
2009{
2010	struct mptcp_subflow_context *subflow;
2011	struct sock *sk = (struct sock *)msk;
2012	u8 scaling_ratio = U8_MAX;
2013	u32 time, advmss = 1;
2014	u64 rtt_us, mstamp;
2015
2016	msk_owned_by_me(msk);
2017
2018	if (copied <= 0)
2019		return;
2020
2021	if (!msk->rcvspace_init)
2022		mptcp_rcv_space_init(msk, msk->first);
2023
2024	msk->rcvq_space.copied += copied;
2025
2026	mstamp = div_u64(tcp_clock_ns(), NSEC_PER_USEC);
2027	time = tcp_stamp_us_delta(mstamp, msk->rcvq_space.time);
2028
2029	rtt_us = msk->rcvq_space.rtt_us;
2030	if (rtt_us && time < (rtt_us >> 3))
2031		return;
2032
2033	rtt_us = 0;
2034	mptcp_for_each_subflow(msk, subflow) {
2035		const struct tcp_sock *tp;
2036		u64 sf_rtt_us;
2037		u32 sf_advmss;
2038
2039		tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2040
2041		sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2042		sf_advmss = READ_ONCE(tp->advmss);
2043
2044		rtt_us = max(sf_rtt_us, rtt_us);
2045		advmss = max(sf_advmss, advmss);
2046		scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2047	}
2048
2049	msk->rcvq_space.rtt_us = rtt_us;
2050	msk->scaling_ratio = scaling_ratio;
2051	if (time < (rtt_us >> 3) || rtt_us == 0)
2052		return;
2053
2054	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2055		goto new_measure;
2056
2057	if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) &&
2058	    !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) {
2059		u64 rcvwin, grow;
2060		int rcvbuf;
2061
2062		rcvwin = ((u64)msk->rcvq_space.copied << 1) + 16 * advmss;
2063
2064		grow = rcvwin * (msk->rcvq_space.copied - msk->rcvq_space.space);
2065
2066		do_div(grow, msk->rcvq_space.space);
2067		rcvwin += (grow << 1);
2068
2069		rcvbuf = min_t(u64, mptcp_space_from_win(sk, rcvwin),
2070			       READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
2071
2072		if (rcvbuf > sk->sk_rcvbuf) {
2073			u32 window_clamp;
2074
2075			window_clamp = mptcp_win_from_space(sk, rcvbuf);
2076			WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
2077
2078			/* Make subflows follow along.  If we do not do this, we
2079			 * get drops at subflow level if skbs can't be moved to
2080			 * the mptcp rx queue fast enough (announced rcv_win can
2081			 * exceed ssk->sk_rcvbuf).
2082			 */
2083			mptcp_for_each_subflow(msk, subflow) {
2084				struct sock *ssk;
2085				bool slow;
2086
2087				ssk = mptcp_subflow_tcp_sock(subflow);
2088				slow = lock_sock_fast(ssk);
2089				WRITE_ONCE(ssk->sk_rcvbuf, rcvbuf);
2090				WRITE_ONCE(tcp_sk(ssk)->window_clamp, window_clamp);
2091				if (tcp_can_send_ack(ssk))
2092					tcp_cleanup_rbuf(ssk, 1);
2093				unlock_sock_fast(ssk, slow);
2094			}
2095		}
2096	}
2097
2098	msk->rcvq_space.space = msk->rcvq_space.copied;
2099new_measure:
2100	msk->rcvq_space.copied = 0;
2101	msk->rcvq_space.time = mstamp;
2102}
2103
2104static void __mptcp_update_rmem(struct sock *sk)
2105{
2106	struct mptcp_sock *msk = mptcp_sk(sk);
2107
2108	if (!msk->rmem_released)
2109		return;
2110
2111	atomic_sub(msk->rmem_released, &sk->sk_rmem_alloc);
2112	mptcp_rmem_uncharge(sk, msk->rmem_released);
2113	WRITE_ONCE(msk->rmem_released, 0);
2114}
2115
2116static void __mptcp_splice_receive_queue(struct sock *sk)
2117{
2118	struct mptcp_sock *msk = mptcp_sk(sk);
2119
2120	skb_queue_splice_tail_init(&sk->sk_receive_queue, &msk->receive_queue);
2121}
2122
2123static bool __mptcp_move_skbs(struct mptcp_sock *msk)
2124{
2125	struct sock *sk = (struct sock *)msk;
2126	unsigned int moved = 0;
2127	bool ret, done;
2128
2129	do {
2130		struct sock *ssk = mptcp_subflow_recv_lookup(msk);
2131		bool slowpath;
2132
2133		/* we can have data pending in the subflows only if the msk
2134		 * receive buffer was full at subflow_data_ready() time,
2135		 * that is an unlikely slow path.
2136		 */
2137		if (likely(!ssk))
2138			break;
2139
2140		slowpath = lock_sock_fast(ssk);
2141		mptcp_data_lock(sk);
2142		__mptcp_update_rmem(sk);
2143		done = __mptcp_move_skbs_from_subflow(msk, ssk, &moved);
2144		mptcp_data_unlock(sk);
2145
2146		if (unlikely(ssk->sk_err))
2147			__mptcp_error_report(sk);
2148		unlock_sock_fast(ssk, slowpath);
2149	} while (!done);
2150
2151	/* acquire the data lock only if some input data is pending */
2152	ret = moved > 0;
2153	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue) ||
2154	    !skb_queue_empty_lockless(&sk->sk_receive_queue)) {
2155		mptcp_data_lock(sk);
2156		__mptcp_update_rmem(sk);
2157		ret |= __mptcp_ofo_queue(msk);
2158		__mptcp_splice_receive_queue(sk);
2159		mptcp_data_unlock(sk);
2160	}
2161	if (ret)
2162		mptcp_check_data_fin((struct sock *)msk);
2163	return !skb_queue_empty(&msk->receive_queue);
2164}
2165
2166static unsigned int mptcp_inq_hint(const struct sock *sk)
2167{
2168	const struct mptcp_sock *msk = mptcp_sk(sk);
2169	const struct sk_buff *skb;
2170
2171	skb = skb_peek(&msk->receive_queue);
2172	if (skb) {
2173		u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2174
2175		if (hint_val >= INT_MAX)
2176			return INT_MAX;
2177
2178		return (unsigned int)hint_val;
2179	}
2180
2181	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2182		return 1;
2183
2184	return 0;
2185}
2186
2187static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2188			 int flags, int *addr_len)
2189{
2190	struct mptcp_sock *msk = mptcp_sk(sk);
2191	struct scm_timestamping_internal tss;
2192	int copied = 0, cmsg_flags = 0;
2193	int target;
2194	long timeo;
2195
2196	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2197	if (unlikely(flags & MSG_ERRQUEUE))
2198		return inet_recv_error(sk, msg, len, addr_len);
2199
2200	lock_sock(sk);
2201	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2202		copied = -ENOTCONN;
2203		goto out_err;
2204	}
2205
2206	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2207
2208	len = min_t(size_t, len, INT_MAX);
2209	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2210
2211	if (unlikely(msk->recvmsg_inq))
2212		cmsg_flags = MPTCP_CMSG_INQ;
2213
2214	while (copied < len) {
2215		int err, bytes_read;
2216
2217		bytes_read = __mptcp_recvmsg_mskq(msk, msg, len - copied, flags, &tss, &cmsg_flags);
2218		if (unlikely(bytes_read < 0)) {
2219			if (!copied)
2220				copied = bytes_read;
2221			goto out_err;
2222		}
2223
2224		copied += bytes_read;
2225
2226		if (skb_queue_empty(&msk->receive_queue) && __mptcp_move_skbs(msk))
2227			continue;
2228
2229		/* only the MPTCP socket status is relevant here. The exit
2230		 * conditions mirror closely tcp_recvmsg()
2231		 */
2232		if (copied >= target)
2233			break;
2234
2235		if (copied) {
2236			if (sk->sk_err ||
2237			    sk->sk_state == TCP_CLOSE ||
2238			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2239			    !timeo ||
2240			    signal_pending(current))
2241				break;
2242		} else {
2243			if (sk->sk_err) {
2244				copied = sock_error(sk);
2245				break;
2246			}
2247
2248			if (sk->sk_shutdown & RCV_SHUTDOWN) {
2249				/* race breaker: the shutdown could be after the
2250				 * previous receive queue check
2251				 */
2252				if (__mptcp_move_skbs(msk))
2253					continue;
2254				break;
2255			}
2256
2257			if (sk->sk_state == TCP_CLOSE) {
2258				copied = -ENOTCONN;
2259				break;
2260			}
2261
2262			if (!timeo) {
2263				copied = -EAGAIN;
2264				break;
2265			}
2266
2267			if (signal_pending(current)) {
2268				copied = sock_intr_errno(timeo);
2269				break;
2270			}
2271		}
2272
2273		pr_debug("block timeout %ld\n", timeo);
2274		mptcp_cleanup_rbuf(msk, copied);
2275		err = sk_wait_data(sk, &timeo, NULL);
2276		if (err < 0) {
2277			err = copied ? : err;
2278			goto out_err;
2279		}
2280	}
2281
2282	mptcp_cleanup_rbuf(msk, copied);
2283
2284out_err:
2285	if (cmsg_flags && copied >= 0) {
2286		if (cmsg_flags & MPTCP_CMSG_TS)
2287			tcp_recv_timestamp(msg, sk, &tss);
2288
2289		if (cmsg_flags & MPTCP_CMSG_INQ) {
2290			unsigned int inq = mptcp_inq_hint(sk);
2291
2292			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2293		}
2294	}
2295
2296	pr_debug("msk=%p rx queue empty=%d:%d copied=%d\n",
2297		 msk, skb_queue_empty_lockless(&sk->sk_receive_queue),
2298		 skb_queue_empty(&msk->receive_queue), copied);
2299
2300	release_sock(sk);
2301	return copied;
2302}
2303
2304static void mptcp_retransmit_timer(struct timer_list *t)
2305{
2306	struct inet_connection_sock *icsk = from_timer(icsk, t,
2307						       icsk_retransmit_timer);
2308	struct sock *sk = &icsk->icsk_inet.sk;
2309	struct mptcp_sock *msk = mptcp_sk(sk);
2310
2311	bh_lock_sock(sk);
2312	if (!sock_owned_by_user(sk)) {
2313		/* we need a process context to retransmit */
2314		if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2315			mptcp_schedule_work(sk);
2316	} else {
2317		/* delegate our work to tcp_release_cb() */
2318		__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2319	}
2320	bh_unlock_sock(sk);
2321	sock_put(sk);
2322}
2323
2324static void mptcp_tout_timer(struct timer_list *t)
2325{
2326	struct sock *sk = from_timer(sk, t, sk_timer);
2327
2328	mptcp_schedule_work(sk);
2329	sock_put(sk);
2330}
2331
2332/* Find an idle subflow.  Return NULL if there is unacked data at tcp
2333 * level.
2334 *
2335 * A backup subflow is returned only if that is the only kind available.
2336 */
2337struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2338{
2339	struct sock *backup = NULL, *pick = NULL;
2340	struct mptcp_subflow_context *subflow;
2341	int min_stale_count = INT_MAX;
2342
2343	mptcp_for_each_subflow(msk, subflow) {
2344		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2345
2346		if (!__mptcp_subflow_active(subflow))
2347			continue;
2348
2349		/* still data outstanding at TCP level? skip this */
2350		if (!tcp_rtx_and_write_queues_empty(ssk)) {
2351			mptcp_pm_subflow_chk_stale(msk, ssk);
2352			min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2353			continue;
2354		}
2355
2356		if (subflow->backup || subflow->request_bkup) {
2357			if (!backup)
2358				backup = ssk;
2359			continue;
2360		}
2361
2362		if (!pick)
2363			pick = ssk;
2364	}
2365
2366	if (pick)
2367		return pick;
2368
2369	/* use backup only if there are no progresses anywhere */
2370	return min_stale_count > 1 ? backup : NULL;
2371}
2372
2373bool __mptcp_retransmit_pending_data(struct sock *sk)
2374{
2375	struct mptcp_data_frag *cur, *rtx_head;
2376	struct mptcp_sock *msk = mptcp_sk(sk);
2377
2378	if (__mptcp_check_fallback(msk))
2379		return false;
2380
2381	/* the closing socket has some data untransmitted and/or unacked:
2382	 * some data in the mptcp rtx queue has not really xmitted yet.
2383	 * keep it simple and re-inject the whole mptcp level rtx queue
2384	 */
2385	mptcp_data_lock(sk);
2386	__mptcp_clean_una_wakeup(sk);
2387	rtx_head = mptcp_rtx_head(sk);
2388	if (!rtx_head) {
2389		mptcp_data_unlock(sk);
2390		return false;
2391	}
2392
2393	msk->recovery_snd_nxt = msk->snd_nxt;
2394	msk->recovery = true;
2395	mptcp_data_unlock(sk);
2396
2397	msk->first_pending = rtx_head;
2398	msk->snd_burst = 0;
2399
2400	/* be sure to clear the "sent status" on all re-injected fragments */
2401	list_for_each_entry(cur, &msk->rtx_queue, list) {
2402		if (!cur->already_sent)
2403			break;
2404		cur->already_sent = 0;
2405	}
2406
2407	return true;
2408}
2409
2410/* flags for __mptcp_close_ssk() */
2411#define MPTCP_CF_PUSH		BIT(1)
2412#define MPTCP_CF_FASTCLOSE	BIT(2)
2413
2414/* be sure to send a reset only if the caller asked for it, also
2415 * clean completely the subflow status when the subflow reaches
2416 * TCP_CLOSE state
2417 */
2418static void __mptcp_subflow_disconnect(struct sock *ssk,
2419				       struct mptcp_subflow_context *subflow,
2420				       unsigned int flags)
2421{
2422	if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2423	    (flags & MPTCP_CF_FASTCLOSE)) {
2424		/* The MPTCP code never wait on the subflow sockets, TCP-level
2425		 * disconnect should never fail
2426		 */
2427		WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2428		mptcp_subflow_ctx_reset(subflow);
2429	} else {
2430		tcp_shutdown(ssk, SEND_SHUTDOWN);
2431	}
2432}
2433
2434/* subflow sockets can be either outgoing (connect) or incoming
2435 * (accept).
2436 *
2437 * Outgoing subflows use in-kernel sockets.
2438 * Incoming subflows do not have their own 'struct socket' allocated,
2439 * so we need to use tcp_close() after detaching them from the mptcp
2440 * parent socket.
2441 */
2442static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2443			      struct mptcp_subflow_context *subflow,
2444			      unsigned int flags)
2445{
2446	struct mptcp_sock *msk = mptcp_sk(sk);
2447	bool dispose_it, need_push = false;
2448
2449	/* If the first subflow moved to a close state before accept, e.g. due
2450	 * to an incoming reset or listener shutdown, the subflow socket is
2451	 * already deleted by inet_child_forget() and the mptcp socket can't
2452	 * survive too.
2453	 */
2454	if (msk->in_accept_queue && msk->first == ssk &&
2455	    (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2456		/* ensure later check in mptcp_worker() will dispose the msk */
2457		sock_set_flag(sk, SOCK_DEAD);
2458		mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2459		lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2460		mptcp_subflow_drop_ctx(ssk);
2461		goto out_release;
2462	}
2463
2464	dispose_it = msk->free_first || ssk != msk->first;
2465	if (dispose_it)
2466		list_del(&subflow->node);
2467
2468	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2469
2470	if ((flags & MPTCP_CF_FASTCLOSE) && !__mptcp_check_fallback(msk)) {
2471		/* be sure to force the tcp_close path
2472		 * to generate the egress reset
2473		 */
2474		ssk->sk_lingertime = 0;
2475		sock_set_flag(ssk, SOCK_LINGER);
2476		subflow->send_fastclose = 1;
2477	}
2478
2479	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2480	if (!dispose_it) {
2481		__mptcp_subflow_disconnect(ssk, subflow, flags);
2482		release_sock(ssk);
2483
2484		goto out;
2485	}
2486
2487	subflow->disposable = 1;
2488
2489	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2490	 * the ssk has been already destroyed, we just need to release the
2491	 * reference owned by msk;
2492	 */
2493	if (!inet_csk(ssk)->icsk_ulp_ops) {
2494		WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2495		kfree_rcu(subflow, rcu);
2496	} else {
2497		/* otherwise tcp will dispose of the ssk and subflow ctx */
2498		__tcp_close(ssk, 0);
2499
2500		/* close acquired an extra ref */
2501		__sock_put(ssk);
2502	}
2503
2504out_release:
2505	__mptcp_subflow_error_report(sk, ssk);
2506	release_sock(ssk);
2507
2508	sock_put(ssk);
2509
2510	if (ssk == msk->first)
2511		WRITE_ONCE(msk->first, NULL);
2512
2513out:
2514	__mptcp_sync_sndbuf(sk);
2515	if (need_push)
2516		__mptcp_push_pending(sk, 0);
2517
2518	/* Catch every 'all subflows closed' scenario, including peers silently
2519	 * closing them, e.g. due to timeout.
2520	 * For established sockets, allow an additional timeout before closing,
2521	 * as the protocol can still create more subflows.
2522	 */
2523	if (list_is_singular(&msk->conn_list) && msk->first &&
2524	    inet_sk_state_load(msk->first) == TCP_CLOSE) {
2525		if (sk->sk_state != TCP_ESTABLISHED ||
2526		    msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2527			mptcp_set_state(sk, TCP_CLOSE);
2528			mptcp_close_wake_up(sk);
2529		} else {
2530			mptcp_start_tout_timer(sk);
2531		}
2532	}
2533}
2534
2535void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2536		     struct mptcp_subflow_context *subflow)
2537{
2538	/* The first subflow can already be closed and still in the list */
2539	if (subflow->close_event_done)
2540		return;
2541
2542	subflow->close_event_done = true;
2543
2544	if (sk->sk_state == TCP_ESTABLISHED)
2545		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2546
2547	/* subflow aborted before reaching the fully_established status
2548	 * attempt the creation of the next subflow
2549	 */
2550	mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2551
2552	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2553}
2554
2555static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2556{
2557	return 0;
2558}
2559
2560static void __mptcp_close_subflow(struct sock *sk)
2561{
2562	struct mptcp_subflow_context *subflow, *tmp;
2563	struct mptcp_sock *msk = mptcp_sk(sk);
2564
2565	might_sleep();
2566
2567	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2568		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2569		int ssk_state = inet_sk_state_load(ssk);
2570
2571		if (ssk_state != TCP_CLOSE &&
2572		    (ssk_state != TCP_CLOSE_WAIT ||
2573		     inet_sk_state_load(sk) != TCP_ESTABLISHED))
2574			continue;
2575
2576		/* 'subflow_data_ready' will re-sched once rx queue is empty */
2577		if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2578			continue;
2579
2580		mptcp_close_ssk(sk, ssk, subflow);
2581	}
2582
2583}
2584
2585static bool mptcp_close_tout_expired(const struct sock *sk)
2586{
2587	if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2588	    sk->sk_state == TCP_CLOSE)
2589		return false;
2590
2591	return time_after32(tcp_jiffies32,
2592		  inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2593}
2594
2595static void mptcp_check_fastclose(struct mptcp_sock *msk)
2596{
2597	struct mptcp_subflow_context *subflow, *tmp;
2598	struct sock *sk = (struct sock *)msk;
2599
2600	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2601		return;
2602
2603	mptcp_token_destroy(msk);
2604
2605	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2606		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2607		bool slow;
2608
2609		slow = lock_sock_fast(tcp_sk);
2610		if (tcp_sk->sk_state != TCP_CLOSE) {
2611			mptcp_send_active_reset_reason(tcp_sk);
2612			tcp_set_state(tcp_sk, TCP_CLOSE);
2613		}
2614		unlock_sock_fast(tcp_sk, slow);
2615	}
2616
2617	/* Mirror the tcp_reset() error propagation */
2618	switch (sk->sk_state) {
2619	case TCP_SYN_SENT:
2620		WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2621		break;
2622	case TCP_CLOSE_WAIT:
2623		WRITE_ONCE(sk->sk_err, EPIPE);
2624		break;
2625	case TCP_CLOSE:
2626		return;
2627	default:
2628		WRITE_ONCE(sk->sk_err, ECONNRESET);
2629	}
2630
2631	mptcp_set_state(sk, TCP_CLOSE);
2632	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2633	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2634	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2635
2636	/* the calling mptcp_worker will properly destroy the socket */
2637	if (sock_flag(sk, SOCK_DEAD))
2638		return;
2639
2640	sk->sk_state_change(sk);
2641	sk_error_report(sk);
2642}
2643
2644static void __mptcp_retrans(struct sock *sk)
2645{
2646	struct mptcp_sock *msk = mptcp_sk(sk);
2647	struct mptcp_subflow_context *subflow;
2648	struct mptcp_sendmsg_info info = {};
2649	struct mptcp_data_frag *dfrag;
2650	struct sock *ssk;
2651	int ret, err;
2652	u16 len = 0;
2653
2654	mptcp_clean_una_wakeup(sk);
2655
2656	/* first check ssk: need to kick "stale" logic */
2657	err = mptcp_sched_get_retrans(msk);
2658	dfrag = mptcp_rtx_head(sk);
2659	if (!dfrag) {
2660		if (mptcp_data_fin_enabled(msk)) {
2661			struct inet_connection_sock *icsk = inet_csk(sk);
2662
2663			icsk->icsk_retransmits++;
2664			mptcp_set_datafin_timeout(sk);
2665			mptcp_send_ack(msk);
2666
2667			goto reset_timer;
2668		}
2669
2670		if (!mptcp_send_head(sk))
2671			return;
2672
2673		goto reset_timer;
2674	}
2675
2676	if (err)
2677		goto reset_timer;
2678
2679	mptcp_for_each_subflow(msk, subflow) {
2680		if (READ_ONCE(subflow->scheduled)) {
2681			u16 copied = 0;
2682
2683			mptcp_subflow_set_scheduled(subflow, false);
2684
2685			ssk = mptcp_subflow_tcp_sock(subflow);
2686
2687			lock_sock(ssk);
2688
2689			/* limit retransmission to the bytes already sent on some subflows */
2690			info.sent = 0;
2691			info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2692								    dfrag->already_sent;
2693			while (info.sent < info.limit) {
2694				ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2695				if (ret <= 0)
2696					break;
2697
2698				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2699				copied += ret;
2700				info.sent += ret;
2701			}
2702			if (copied) {
2703				len = max(copied, len);
2704				tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2705					 info.size_goal);
2706				WRITE_ONCE(msk->allow_infinite_fallback, false);
2707			}
2708
2709			release_sock(ssk);
2710		}
2711	}
2712
2713	msk->bytes_retrans += len;
2714	dfrag->already_sent = max(dfrag->already_sent, len);
2715
2716reset_timer:
2717	mptcp_check_and_set_pending(sk);
2718
2719	if (!mptcp_rtx_timer_pending(sk))
2720		mptcp_reset_rtx_timer(sk);
2721}
2722
2723/* schedule the timeout timer for the relevant event: either close timeout
2724 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2725 */
2726void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2727{
2728	struct sock *sk = (struct sock *)msk;
2729	unsigned long timeout, close_timeout;
2730
2731	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2732		return;
2733
2734	close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp -
2735			tcp_jiffies32 + jiffies + mptcp_close_timeout(sk);
2736
2737	/* the close timeout takes precedence on the fail one, and here at least one of
2738	 * them is active
2739	 */
2740	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2741
2742	sk_reset_timer(sk, &sk->sk_timer, timeout);
2743}
2744
2745static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2746{
2747	struct sock *ssk = msk->first;
2748	bool slow;
2749
2750	if (!ssk)
2751		return;
2752
2753	pr_debug("MP_FAIL doesn't respond, reset the subflow\n");
2754
2755	slow = lock_sock_fast(ssk);
2756	mptcp_subflow_reset(ssk);
2757	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2758	unlock_sock_fast(ssk, slow);
2759}
2760
2761static void mptcp_do_fastclose(struct sock *sk)
2762{
2763	struct mptcp_subflow_context *subflow, *tmp;
2764	struct mptcp_sock *msk = mptcp_sk(sk);
2765
2766	mptcp_set_state(sk, TCP_CLOSE);
2767	mptcp_for_each_subflow_safe(msk, subflow, tmp)
2768		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow),
2769				  subflow, MPTCP_CF_FASTCLOSE);
2770}
2771
2772static void mptcp_worker(struct work_struct *work)
2773{
2774	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2775	struct sock *sk = (struct sock *)msk;
2776	unsigned long fail_tout;
2777	int state;
2778
2779	lock_sock(sk);
2780	state = sk->sk_state;
2781	if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2782		goto unlock;
2783
2784	mptcp_check_fastclose(msk);
2785
2786	mptcp_pm_nl_work(msk);
2787
2788	mptcp_check_send_data_fin(sk);
2789	mptcp_check_data_fin_ack(sk);
2790	mptcp_check_data_fin(sk);
2791
2792	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2793		__mptcp_close_subflow(sk);
2794
2795	if (mptcp_close_tout_expired(sk)) {
2796		mptcp_do_fastclose(sk);
2797		mptcp_close_wake_up(sk);
2798	}
2799
2800	if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2801		__mptcp_destroy_sock(sk);
2802		goto unlock;
2803	}
2804
2805	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2806		__mptcp_retrans(sk);
2807
2808	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2809	if (fail_tout && time_after(jiffies, fail_tout))
2810		mptcp_mp_fail_no_response(msk);
2811
2812unlock:
2813	release_sock(sk);
2814	sock_put(sk);
2815}
2816
2817static void __mptcp_init_sock(struct sock *sk)
2818{
2819	struct mptcp_sock *msk = mptcp_sk(sk);
2820
2821	INIT_LIST_HEAD(&msk->conn_list);
2822	INIT_LIST_HEAD(&msk->join_list);
2823	INIT_LIST_HEAD(&msk->rtx_queue);
2824	INIT_WORK(&msk->work, mptcp_worker);
2825	__skb_queue_head_init(&msk->receive_queue);
2826	msk->out_of_order_queue = RB_ROOT;
2827	msk->first_pending = NULL;
2828	WRITE_ONCE(msk->rmem_fwd_alloc, 0);
2829	WRITE_ONCE(msk->rmem_released, 0);
2830	msk->timer_ival = TCP_RTO_MIN;
2831	msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2832
2833	WRITE_ONCE(msk->first, NULL);
2834	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2835	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2836	WRITE_ONCE(msk->allow_infinite_fallback, true);
2837	msk->recovery = false;
2838	msk->subflow_id = 1;
2839	msk->last_data_sent = tcp_jiffies32;
2840	msk->last_data_recv = tcp_jiffies32;
2841	msk->last_ack_recv = tcp_jiffies32;
2842
2843	mptcp_pm_data_init(msk);
2844
2845	/* re-use the csk retrans timer for MPTCP-level retrans */
2846	timer_setup(&msk->sk.icsk_retransmit_timer, mptcp_retransmit_timer, 0);
2847	timer_setup(&sk->sk_timer, mptcp_tout_timer, 0);
2848}
2849
2850static void mptcp_ca_reset(struct sock *sk)
2851{
2852	struct inet_connection_sock *icsk = inet_csk(sk);
2853
2854	tcp_assign_congestion_control(sk);
2855	strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
2856		sizeof(mptcp_sk(sk)->ca_name));
2857
2858	/* no need to keep a reference to the ops, the name will suffice */
2859	tcp_cleanup_congestion_control(sk);
2860	icsk->icsk_ca_ops = NULL;
2861}
2862
2863static int mptcp_init_sock(struct sock *sk)
2864{
2865	struct net *net = sock_net(sk);
2866	int ret;
2867
2868	__mptcp_init_sock(sk);
2869
2870	if (!mptcp_is_enabled(net))
2871		return -ENOPROTOOPT;
2872
2873	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
2874		return -ENOMEM;
2875
2876	rcu_read_lock();
2877	ret = mptcp_init_sched(mptcp_sk(sk),
2878			       mptcp_sched_find(mptcp_get_scheduler(net)));
2879	rcu_read_unlock();
2880	if (ret)
2881		return ret;
2882
2883	set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
2884
2885	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
2886	 * propagate the correct value
2887	 */
2888	mptcp_ca_reset(sk);
2889
2890	sk_sockets_allocated_inc(sk);
2891	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
2892	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
2893
2894	return 0;
2895}
2896
2897static void __mptcp_clear_xmit(struct sock *sk)
2898{
2899	struct mptcp_sock *msk = mptcp_sk(sk);
2900	struct mptcp_data_frag *dtmp, *dfrag;
2901
2902	WRITE_ONCE(msk->first_pending, NULL);
2903	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
2904		dfrag_clear(sk, dfrag);
2905}
2906
2907void mptcp_cancel_work(struct sock *sk)
2908{
2909	struct mptcp_sock *msk = mptcp_sk(sk);
2910
2911	if (cancel_work_sync(&msk->work))
2912		__sock_put(sk);
2913}
2914
2915void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
2916{
2917	lock_sock(ssk);
2918
2919	switch (ssk->sk_state) {
2920	case TCP_LISTEN:
2921		if (!(how & RCV_SHUTDOWN))
2922			break;
2923		fallthrough;
2924	case TCP_SYN_SENT:
2925		WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
2926		break;
2927	default:
2928		if (__mptcp_check_fallback(mptcp_sk(sk))) {
2929			pr_debug("Fallback\n");
2930			ssk->sk_shutdown |= how;
2931			tcp_shutdown(ssk, how);
2932
2933			/* simulate the data_fin ack reception to let the state
2934			 * machine move forward
2935			 */
2936			WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
2937			mptcp_schedule_work(sk);
2938		} else {
2939			pr_debug("Sending DATA_FIN on subflow %p\n", ssk);
2940			tcp_send_ack(ssk);
2941			if (!mptcp_rtx_timer_pending(sk))
2942				mptcp_reset_rtx_timer(sk);
2943		}
2944		break;
2945	}
2946
2947	release_sock(ssk);
2948}
2949
2950void mptcp_set_state(struct sock *sk, int state)
2951{
2952	int oldstate = sk->sk_state;
2953
2954	switch (state) {
2955	case TCP_ESTABLISHED:
2956		if (oldstate != TCP_ESTABLISHED)
2957			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2958		break;
2959	case TCP_CLOSE_WAIT:
2960		/* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state:
2961		 * MPTCP "accepted" sockets will be created later on. So no
2962		 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT.
2963		 */
2964		break;
2965	default:
2966		if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
2967			MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
2968	}
2969
2970	inet_sk_state_store(sk, state);
2971}
2972
2973static const unsigned char new_state[16] = {
2974	/* current state:     new state:      action:	*/
2975	[0 /* (Invalid) */] = TCP_CLOSE,
2976	[TCP_ESTABLISHED]   = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2977	[TCP_SYN_SENT]      = TCP_CLOSE,
2978	[TCP_SYN_RECV]      = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2979	[TCP_FIN_WAIT1]     = TCP_FIN_WAIT1,
2980	[TCP_FIN_WAIT2]     = TCP_FIN_WAIT2,
2981	[TCP_TIME_WAIT]     = TCP_CLOSE,	/* should not happen ! */
2982	[TCP_CLOSE]         = TCP_CLOSE,
2983	[TCP_CLOSE_WAIT]    = TCP_LAST_ACK  | TCP_ACTION_FIN,
2984	[TCP_LAST_ACK]      = TCP_LAST_ACK,
2985	[TCP_LISTEN]        = TCP_CLOSE,
2986	[TCP_CLOSING]       = TCP_CLOSING,
2987	[TCP_NEW_SYN_RECV]  = TCP_CLOSE,	/* should not happen ! */
2988};
2989
2990static int mptcp_close_state(struct sock *sk)
2991{
2992	int next = (int)new_state[sk->sk_state];
2993	int ns = next & TCP_STATE_MASK;
2994
2995	mptcp_set_state(sk, ns);
2996
2997	return next & TCP_ACTION_FIN;
2998}
2999
3000static void mptcp_check_send_data_fin(struct sock *sk)
3001{
3002	struct mptcp_subflow_context *subflow;
3003	struct mptcp_sock *msk = mptcp_sk(sk);
3004
3005	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n",
3006		 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
3007		 msk->snd_nxt, msk->write_seq);
3008
3009	/* we still need to enqueue subflows or not really shutting down,
3010	 * skip this
3011	 */
3012	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
3013	    mptcp_send_head(sk))
3014		return;
3015
3016	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3017
3018	mptcp_for_each_subflow(msk, subflow) {
3019		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
3020
3021		mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
3022	}
3023}
3024
3025static void __mptcp_wr_shutdown(struct sock *sk)
3026{
3027	struct mptcp_sock *msk = mptcp_sk(sk);
3028
3029	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n",
3030		 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
3031		 !!mptcp_send_head(sk));
3032
3033	/* will be ignored by fallback sockets */
3034	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
3035	WRITE_ONCE(msk->snd_data_fin_enable, 1);
3036
3037	mptcp_check_send_data_fin(sk);
3038}
3039
3040static void __mptcp_destroy_sock(struct sock *sk)
3041{
3042	struct mptcp_sock *msk = mptcp_sk(sk);
3043
3044	pr_debug("msk=%p\n", msk);
3045
3046	might_sleep();
3047
3048	mptcp_stop_rtx_timer(sk);
3049	sk_stop_timer(sk, &sk->sk_timer);
3050	msk->pm.status = 0;
3051	mptcp_release_sched(msk);
3052
3053	sk->sk_prot->destroy(sk);
3054
3055	WARN_ON_ONCE(READ_ONCE(msk->rmem_fwd_alloc));
3056	WARN_ON_ONCE(msk->rmem_released);
3057	sk_stream_kill_queues(sk);
3058	xfrm_sk_free_policy(sk);
3059
3060	sock_put(sk);
3061}
3062
3063void __mptcp_unaccepted_force_close(struct sock *sk)
3064{
3065	sock_set_flag(sk, SOCK_DEAD);
3066	mptcp_do_fastclose(sk);
3067	__mptcp_destroy_sock(sk);
3068}
3069
3070static __poll_t mptcp_check_readable(struct sock *sk)
3071{
3072	return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3073}
3074
3075static void mptcp_check_listen_stop(struct sock *sk)
3076{
3077	struct sock *ssk;
3078
3079	if (inet_sk_state_load(sk) != TCP_LISTEN)
3080		return;
3081
3082	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3083	ssk = mptcp_sk(sk)->first;
3084	if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3085		return;
3086
3087	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3088	tcp_set_state(ssk, TCP_CLOSE);
3089	mptcp_subflow_queue_clean(sk, ssk);
3090	inet_csk_listen_stop(ssk);
3091	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3092	release_sock(ssk);
3093}
3094
3095bool __mptcp_close(struct sock *sk, long timeout)
3096{
3097	struct mptcp_subflow_context *subflow;
3098	struct mptcp_sock *msk = mptcp_sk(sk);
3099	bool do_cancel_work = false;
3100	int subflows_alive = 0;
3101
3102	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3103
3104	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3105		mptcp_check_listen_stop(sk);
3106		mptcp_set_state(sk, TCP_CLOSE);
3107		goto cleanup;
3108	}
3109
3110	if (mptcp_data_avail(msk) || timeout < 0) {
3111		/* If the msk has read data, or the caller explicitly ask it,
3112		 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3113		 */
3114		mptcp_do_fastclose(sk);
3115		timeout = 0;
3116	} else if (mptcp_close_state(sk)) {
3117		__mptcp_wr_shutdown(sk);
3118	}
3119
3120	sk_stream_wait_close(sk, timeout);
3121
3122cleanup:
3123	/* orphan all the subflows */
3124	mptcp_for_each_subflow(msk, subflow) {
3125		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3126		bool slow = lock_sock_fast_nested(ssk);
3127
3128		subflows_alive += ssk->sk_state != TCP_CLOSE;
3129
3130		/* since the close timeout takes precedence on the fail one,
3131		 * cancel the latter
3132		 */
3133		if (ssk == msk->first)
3134			subflow->fail_tout = 0;
3135
3136		/* detach from the parent socket, but allow data_ready to
3137		 * push incoming data into the mptcp stack, to properly ack it
3138		 */
3139		ssk->sk_socket = NULL;
3140		ssk->sk_wq = NULL;
3141		unlock_sock_fast(ssk, slow);
3142	}
3143	sock_orphan(sk);
3144
3145	/* all the subflows are closed, only timeout can change the msk
3146	 * state, let's not keep resources busy for no reasons
3147	 */
3148	if (subflows_alive == 0)
3149		mptcp_set_state(sk, TCP_CLOSE);
3150
3151	sock_hold(sk);
3152	pr_debug("msk=%p state=%d\n", sk, sk->sk_state);
3153	mptcp_pm_connection_closed(msk);
3154
3155	if (sk->sk_state == TCP_CLOSE) {
3156		__mptcp_destroy_sock(sk);
3157		do_cancel_work = true;
3158	} else {
3159		mptcp_start_tout_timer(sk);
3160	}
3161
3162	return do_cancel_work;
3163}
3164
3165static void mptcp_close(struct sock *sk, long timeout)
3166{
3167	bool do_cancel_work;
3168
3169	lock_sock(sk);
3170
3171	do_cancel_work = __mptcp_close(sk, timeout);
3172	release_sock(sk);
3173	if (do_cancel_work)
3174		mptcp_cancel_work(sk);
3175
3176	sock_put(sk);
3177}
3178
3179static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3180{
3181#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3182	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3183	struct ipv6_pinfo *msk6 = inet6_sk(msk);
3184
3185	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3186	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3187
3188	if (msk6 && ssk6) {
3189		msk6->saddr = ssk6->saddr;
3190		msk6->flow_label = ssk6->flow_label;
3191	}
3192#endif
3193
3194	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3195	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3196	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3197	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3198	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3199	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3200}
3201
3202static int mptcp_disconnect(struct sock *sk, int flags)
3203{
3204	struct mptcp_sock *msk = mptcp_sk(sk);
3205
3206	/* We are on the fastopen error path. We can't call straight into the
3207	 * subflows cleanup code due to lock nesting (we are already under
3208	 * msk->firstsocket lock).
3209	 */
3210	if (msk->fastopening)
3211		return -EBUSY;
3212
3213	mptcp_check_listen_stop(sk);
3214	mptcp_set_state(sk, TCP_CLOSE);
3215
3216	mptcp_stop_rtx_timer(sk);
3217	mptcp_stop_tout_timer(sk);
3218
3219	mptcp_pm_connection_closed(msk);
3220
3221	/* msk->subflow is still intact, the following will not free the first
3222	 * subflow
3223	 */
3224	mptcp_destroy_common(msk, MPTCP_CF_FASTCLOSE);
3225	WRITE_ONCE(msk->flags, 0);
3226	msk->cb_flags = 0;
3227	msk->recovery = false;
3228	WRITE_ONCE(msk->can_ack, false);
3229	WRITE_ONCE(msk->fully_established, false);
3230	WRITE_ONCE(msk->rcv_data_fin, false);
3231	WRITE_ONCE(msk->snd_data_fin_enable, false);
3232	WRITE_ONCE(msk->rcv_fastclose, false);
3233	WRITE_ONCE(msk->use_64bit_ack, false);
3234	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3235	mptcp_pm_data_reset(msk);
3236	mptcp_ca_reset(sk);
3237	msk->bytes_consumed = 0;
3238	msk->bytes_acked = 0;
3239	msk->bytes_received = 0;
3240	msk->bytes_sent = 0;
3241	msk->bytes_retrans = 0;
3242	msk->rcvspace_init = 0;
3243
3244	WRITE_ONCE(sk->sk_shutdown, 0);
3245	sk_error_report(sk);
3246	return 0;
3247}
3248
3249#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3250static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3251{
3252	unsigned int offset = sizeof(struct mptcp6_sock) - sizeof(struct ipv6_pinfo);
3253
3254	return (struct ipv6_pinfo *)(((u8 *)sk) + offset);
3255}
3256
3257static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3258{
3259	const struct ipv6_pinfo *np = inet6_sk(sk);
3260	struct ipv6_txoptions *opt;
3261	struct ipv6_pinfo *newnp;
3262
3263	newnp = inet6_sk(newsk);
3264
3265	rcu_read_lock();
3266	opt = rcu_dereference(np->opt);
3267	if (opt) {
3268		opt = ipv6_dup_options(newsk, opt);
3269		if (!opt)
3270			net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3271	}
3272	RCU_INIT_POINTER(newnp->opt, opt);
3273	rcu_read_unlock();
3274}
3275#endif
3276
3277static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3278{
3279	struct ip_options_rcu *inet_opt, *newopt = NULL;
3280	const struct inet_sock *inet = inet_sk(sk);
3281	struct inet_sock *newinet;
3282
3283	newinet = inet_sk(newsk);
3284
3285	rcu_read_lock();
3286	inet_opt = rcu_dereference(inet->inet_opt);
3287	if (inet_opt) {
3288		newopt = sock_kmalloc(newsk, sizeof(*inet_opt) +
3289				      inet_opt->opt.optlen, GFP_ATOMIC);
3290		if (newopt)
3291			memcpy(newopt, inet_opt, sizeof(*inet_opt) +
3292			       inet_opt->opt.optlen);
3293		else
3294			net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3295	}
3296	RCU_INIT_POINTER(newinet->inet_opt, newopt);
3297	rcu_read_unlock();
3298}
3299
3300struct sock *mptcp_sk_clone_init(const struct sock *sk,
3301				 const struct mptcp_options_received *mp_opt,
3302				 struct sock *ssk,
3303				 struct request_sock *req)
3304{
3305	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3306	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3307	struct mptcp_subflow_context *subflow;
3308	struct mptcp_sock *msk;
3309
3310	if (!nsk)
3311		return NULL;
3312
3313#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3314	if (nsk->sk_family == AF_INET6)
3315		inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3316#endif
3317
3318	__mptcp_init_sock(nsk);
3319
3320#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3321	if (nsk->sk_family == AF_INET6)
3322		mptcp_copy_ip6_options(nsk, sk);
3323	else
3324#endif
3325		mptcp_copy_ip_options(nsk, sk);
3326
3327	msk = mptcp_sk(nsk);
3328	WRITE_ONCE(msk->local_key, subflow_req->local_key);
3329	WRITE_ONCE(msk->token, subflow_req->token);
3330	msk->in_accept_queue = 1;
3331	WRITE_ONCE(msk->fully_established, false);
3332	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3333		WRITE_ONCE(msk->csum_enabled, true);
3334
3335	WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3336	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3337	WRITE_ONCE(msk->snd_una, msk->write_seq);
3338	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3339	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3340	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3341
3342	/* passive msk is created after the first/MPC subflow */
3343	msk->subflow_id = 2;
3344
3345	sock_reset_flag(nsk, SOCK_RCU_FREE);
3346	security_inet_csk_clone(nsk, req);
3347
3348	/* this can't race with mptcp_close(), as the msk is
3349	 * not yet exposted to user-space
3350	 */
3351	mptcp_set_state(nsk, TCP_ESTABLISHED);
3352
3353	/* The msk maintain a ref to each subflow in the connections list */
3354	WRITE_ONCE(msk->first, ssk);
3355	subflow = mptcp_subflow_ctx(ssk);
3356	list_add(&subflow->node, &msk->conn_list);
3357	sock_hold(ssk);
3358
3359	/* new mpc subflow takes ownership of the newly
3360	 * created mptcp socket
3361	 */
3362	mptcp_token_accept(subflow_req, msk);
3363
3364	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3365	 * uses the correct data
3366	 */
3367	mptcp_copy_inaddrs(nsk, ssk);
3368	__mptcp_propagate_sndbuf(nsk, ssk);
3369
3370	mptcp_rcv_space_init(msk, ssk);
3371
3372	if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3373		__mptcp_subflow_fully_established(msk, subflow, mp_opt);
3374	bh_unlock_sock(nsk);
3375
3376	/* note: the newly allocated socket refcount is 2 now */
3377	return nsk;
3378}
3379
3380void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3381{
3382	const struct tcp_sock *tp = tcp_sk(ssk);
3383
3384	msk->rcvspace_init = 1;
3385	msk->rcvq_space.copied = 0;
3386	msk->rcvq_space.rtt_us = 0;
3387
3388	msk->rcvq_space.time = tp->tcp_mstamp;
3389
3390	/* initial rcv_space offering made to peer */
3391	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3392				      TCP_INIT_CWND * tp->advmss);
3393	if (msk->rcvq_space.space == 0)
3394		msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3395}
3396
3397void mptcp_destroy_common(struct mptcp_sock *msk, unsigned int flags)
3398{
3399	struct mptcp_subflow_context *subflow, *tmp;
3400	struct sock *sk = (struct sock *)msk;
3401
3402	__mptcp_clear_xmit(sk);
3403
3404	/* join list will be eventually flushed (with rst) at sock lock release time */
3405	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3406		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, flags);
3407
3408	/* move to sk_receive_queue, sk_stream_kill_queues will purge it */
3409	mptcp_data_lock(sk);
3410	skb_queue_splice_tail_init(&msk->receive_queue, &sk->sk_receive_queue);
3411	__skb_queue_purge(&sk->sk_receive_queue);
3412	skb_rbtree_purge(&msk->out_of_order_queue);
3413	mptcp_data_unlock(sk);
3414
3415	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3416	 * inet_sock_destruct() will dispose it
3417	 */
3418	sk_forward_alloc_add(sk, msk->rmem_fwd_alloc);
3419	WRITE_ONCE(msk->rmem_fwd_alloc, 0);
3420	mptcp_token_destroy(msk);
3421	mptcp_pm_free_anno_list(msk);
3422	mptcp_free_local_addr_list(msk);
3423}
3424
3425static void mptcp_destroy(struct sock *sk)
3426{
3427	struct mptcp_sock *msk = mptcp_sk(sk);
3428
3429	/* allow the following to close even the initial subflow */
3430	msk->free_first = 1;
3431	mptcp_destroy_common(msk, 0);
3432	sk_sockets_allocated_dec(sk);
3433}
3434
3435void __mptcp_data_acked(struct sock *sk)
3436{
3437	if (!sock_owned_by_user(sk))
3438		__mptcp_clean_una(sk);
3439	else
3440		__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3441}
3442
3443void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3444{
3445	if (!mptcp_send_head(sk))
3446		return;
3447
3448	if (!sock_owned_by_user(sk))
3449		__mptcp_subflow_push_pending(sk, ssk, false);
3450	else
3451		__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3452}
3453
3454#define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3455				      BIT(MPTCP_RETRANSMIT) | \
3456				      BIT(MPTCP_FLUSH_JOIN_LIST))
3457
3458/* processes deferred events and flush wmem */
3459static void mptcp_release_cb(struct sock *sk)
3460	__must_hold(&sk->sk_lock.slock)
3461{
3462	struct mptcp_sock *msk = mptcp_sk(sk);
3463
3464	for (;;) {
3465		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3466		struct list_head join_list;
3467
3468		if (!flags)
3469			break;
3470
3471		INIT_LIST_HEAD(&join_list);
3472		list_splice_init(&msk->join_list, &join_list);
3473
3474		/* the following actions acquire the subflow socket lock
3475		 *
3476		 * 1) can't be invoked in atomic scope
3477		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3478		 *    datapath acquires the msk socket spinlock while helding
3479		 *    the subflow socket lock
3480		 */
3481		msk->cb_flags &= ~flags;
3482		spin_unlock_bh(&sk->sk_lock.slock);
3483
3484		if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3485			__mptcp_flush_join_list(sk, &join_list);
3486		if (flags & BIT(MPTCP_PUSH_PENDING))
3487			__mptcp_push_pending(sk, 0);
3488		if (flags & BIT(MPTCP_RETRANSMIT))
3489			__mptcp_retrans(sk);
3490
3491		cond_resched();
3492		spin_lock_bh(&sk->sk_lock.slock);
3493	}
3494
3495	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3496		__mptcp_clean_una_wakeup(sk);
3497	if (unlikely(msk->cb_flags)) {
3498		/* be sure to sync the msk state before taking actions
3499		 * depending on sk_state (MPTCP_ERROR_REPORT)
3500		 * On sk release avoid actions depending on the first subflow
3501		 */
3502		if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3503			__mptcp_sync_state(sk, msk->pending_state);
3504		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3505			__mptcp_error_report(sk);
3506		if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3507			__mptcp_sync_sndbuf(sk);
3508	}
3509
3510	__mptcp_update_rmem(sk);
3511}
3512
3513/* MP_JOIN client subflow must wait for 4th ack before sending any data:
3514 * TCP can't schedule delack timer before the subflow is fully established.
3515 * MPTCP uses the delack timer to do 3rd ack retransmissions
3516 */
3517static void schedule_3rdack_retransmission(struct sock *ssk)
3518{
3519	struct inet_connection_sock *icsk = inet_csk(ssk);
3520	struct tcp_sock *tp = tcp_sk(ssk);
3521	unsigned long timeout;
3522
3523	if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established))
3524		return;
3525
3526	/* reschedule with a timeout above RTT, as we must look only for drop */
3527	if (tp->srtt_us)
3528		timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3529	else
3530		timeout = TCP_TIMEOUT_INIT;
3531	timeout += jiffies;
3532
3533	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3534	smp_store_release(&icsk->icsk_ack.pending,
3535			  icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
3536	icsk->icsk_ack.timeout = timeout;
3537	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3538}
3539
3540void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3541{
3542	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3543	struct sock *sk = subflow->conn;
3544
3545	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3546		mptcp_data_lock(sk);
3547		if (!sock_owned_by_user(sk))
3548			__mptcp_subflow_push_pending(sk, ssk, true);
3549		else
3550			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3551		mptcp_data_unlock(sk);
3552	}
3553	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3554		mptcp_data_lock(sk);
3555		if (!sock_owned_by_user(sk))
3556			__mptcp_sync_sndbuf(sk);
3557		else
3558			__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3559		mptcp_data_unlock(sk);
3560	}
3561	if (status & BIT(MPTCP_DELEGATE_ACK))
3562		schedule_3rdack_retransmission(ssk);
3563}
3564
3565static int mptcp_hash(struct sock *sk)
3566{
3567	/* should never be called,
3568	 * we hash the TCP subflows not the MPTCP socket
3569	 */
3570	WARN_ON_ONCE(1);
3571	return 0;
3572}
3573
3574static void mptcp_unhash(struct sock *sk)
3575{
3576	/* called from sk_common_release(), but nothing to do here */
3577}
3578
3579static int mptcp_get_port(struct sock *sk, unsigned short snum)
3580{
3581	struct mptcp_sock *msk = mptcp_sk(sk);
3582
3583	pr_debug("msk=%p, ssk=%p\n", msk, msk->first);
3584	if (WARN_ON_ONCE(!msk->first))
3585		return -EINVAL;
3586
3587	return inet_csk_get_port(msk->first, snum);
3588}
3589
3590void mptcp_finish_connect(struct sock *ssk)
3591{
3592	struct mptcp_subflow_context *subflow;
3593	struct mptcp_sock *msk;
3594	struct sock *sk;
3595
3596	subflow = mptcp_subflow_ctx(ssk);
3597	sk = subflow->conn;
3598	msk = mptcp_sk(sk);
3599
3600	pr_debug("msk=%p, token=%u\n", sk, subflow->token);
3601
3602	subflow->map_seq = subflow->iasn;
3603	subflow->map_subflow_seq = 1;
3604
3605	/* the socket is not connected yet, no msk/subflow ops can access/race
3606	 * accessing the field below
3607	 */
3608	WRITE_ONCE(msk->local_key, subflow->local_key);
3609
3610	mptcp_pm_new_connection(msk, ssk, 0);
3611}
3612
3613void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3614{
3615	write_lock_bh(&sk->sk_callback_lock);
3616	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3617	sk_set_socket(sk, parent);
3618	sk->sk_uid = SOCK_INODE(parent)->i_uid;
3619	write_unlock_bh(&sk->sk_callback_lock);
3620}
3621
3622bool mptcp_finish_join(struct sock *ssk)
3623{
3624	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3625	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3626	struct sock *parent = (void *)msk;
3627	bool ret = true;
3628
3629	pr_debug("msk=%p, subflow=%p\n", msk, subflow);
3630
3631	/* mptcp socket already closing? */
3632	if (!mptcp_is_fully_established(parent)) {
3633		subflow->reset_reason = MPTCP_RST_EMPTCP;
3634		return false;
3635	}
3636
3637	/* active subflow, already present inside the conn_list */
3638	if (!list_empty(&subflow->node)) {
3639		mptcp_subflow_joined(msk, ssk);
3640		mptcp_propagate_sndbuf(parent, ssk);
3641		return true;
3642	}
3643
3644	if (!mptcp_pm_allow_new_subflow(msk))
3645		goto err_prohibited;
3646
3647	/* If we can't acquire msk socket lock here, let the release callback
3648	 * handle it
3649	 */
3650	mptcp_data_lock(parent);
3651	if (!sock_owned_by_user(parent)) {
3652		ret = __mptcp_finish_join(msk, ssk);
3653		if (ret) {
3654			sock_hold(ssk);
3655			list_add_tail(&subflow->node, &msk->conn_list);
3656		}
3657	} else {
3658		sock_hold(ssk);
3659		list_add_tail(&subflow->node, &msk->join_list);
3660		__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3661	}
3662	mptcp_data_unlock(parent);
3663
3664	if (!ret) {
3665err_prohibited:
3666		subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3667		return false;
3668	}
3669
3670	return true;
3671}
3672
3673static void mptcp_shutdown(struct sock *sk, int how)
3674{
3675	pr_debug("sk=%p, how=%d\n", sk, how);
3676
3677	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3678		__mptcp_wr_shutdown(sk);
3679}
3680
3681static int mptcp_forward_alloc_get(const struct sock *sk)
3682{
3683	return READ_ONCE(sk->sk_forward_alloc) +
3684	       READ_ONCE(mptcp_sk(sk)->rmem_fwd_alloc);
3685}
3686
3687static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3688{
3689	const struct sock *sk = (void *)msk;
3690	u64 delta;
3691
3692	if (sk->sk_state == TCP_LISTEN)
3693		return -EINVAL;
3694
3695	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3696		return 0;
3697
3698	delta = msk->write_seq - v;
3699	if (__mptcp_check_fallback(msk) && msk->first) {
3700		struct tcp_sock *tp = tcp_sk(msk->first);
3701
3702		/* the first subflow is disconnected after close - see
3703		 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3704		 * so ignore that status, too.
3705		 */
3706		if (!((1 << msk->first->sk_state) &
3707		      (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3708			delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3709	}
3710	if (delta > INT_MAX)
3711		delta = INT_MAX;
3712
3713	return (int)delta;
3714}
3715
3716static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3717{
3718	struct mptcp_sock *msk = mptcp_sk(sk);
3719	bool slow;
3720
3721	switch (cmd) {
3722	case SIOCINQ:
3723		if (sk->sk_state == TCP_LISTEN)
3724			return -EINVAL;
3725
3726		lock_sock(sk);
3727		__mptcp_move_skbs(msk);
3728		*karg = mptcp_inq_hint(sk);
3729		release_sock(sk);
3730		break;
3731	case SIOCOUTQ:
3732		slow = lock_sock_fast(sk);
3733		*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3734		unlock_sock_fast(sk, slow);
3735		break;
3736	case SIOCOUTQNSD:
3737		slow = lock_sock_fast(sk);
3738		*karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3739		unlock_sock_fast(sk, slow);
3740		break;
3741	default:
3742		return -ENOIOCTLCMD;
3743	}
3744
3745	return 0;
3746}
3747
3748static int mptcp_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
3749{
3750	struct mptcp_subflow_context *subflow;
3751	struct mptcp_sock *msk = mptcp_sk(sk);
3752	int err = -EINVAL;
3753	struct sock *ssk;
3754
3755	ssk = __mptcp_nmpc_sk(msk);
3756	if (IS_ERR(ssk))
3757		return PTR_ERR(ssk);
3758
3759	mptcp_set_state(sk, TCP_SYN_SENT);
3760	subflow = mptcp_subflow_ctx(ssk);
3761#ifdef CONFIG_TCP_MD5SIG
3762	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3763	 * TCP option space.
3764	 */
3765	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3766		mptcp_subflow_early_fallback(msk, subflow);
3767#endif
3768	if (subflow->request_mptcp) {
3769		if (mptcp_active_should_disable(sk)) {
3770			MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEACTIVEDISABLED);
3771			mptcp_subflow_early_fallback(msk, subflow);
3772		} else if (mptcp_token_new_connect(ssk) < 0) {
3773			MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_TOKENFALLBACKINIT);
3774			mptcp_subflow_early_fallback(msk, subflow);
3775		}
3776	}
3777
3778	WRITE_ONCE(msk->write_seq, subflow->idsn);
3779	WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3780	WRITE_ONCE(msk->snd_una, subflow->idsn);
3781	if (likely(!__mptcp_check_fallback(msk)))
3782		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3783
3784	/* if reaching here via the fastopen/sendmsg path, the caller already
3785	 * acquired the subflow socket lock, too.
3786	 */
3787	if (!msk->fastopening)
3788		lock_sock(ssk);
3789
3790	/* the following mirrors closely a very small chunk of code from
3791	 * __inet_stream_connect()
3792	 */
3793	if (ssk->sk_state != TCP_CLOSE)
3794		goto out;
3795
3796	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3797		err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3798		if (err)
3799			goto out;
3800	}
3801
3802	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3803	if (err < 0)
3804		goto out;
3805
3806	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3807
3808out:
3809	if (!msk->fastopening)
3810		release_sock(ssk);
3811
3812	/* on successful connect, the msk state will be moved to established by
3813	 * subflow_finish_connect()
3814	 */
3815	if (unlikely(err)) {
3816		/* avoid leaving a dangling token in an unconnected socket */
3817		mptcp_token_destroy(msk);
3818		mptcp_set_state(sk, TCP_CLOSE);
3819		return err;
3820	}
3821
3822	mptcp_copy_inaddrs(sk, ssk);
3823	return 0;
3824}
3825
3826static struct proto mptcp_prot = {
3827	.name		= "MPTCP",
3828	.owner		= THIS_MODULE,
3829	.init		= mptcp_init_sock,
3830	.connect	= mptcp_connect,
3831	.disconnect	= mptcp_disconnect,
3832	.close		= mptcp_close,
3833	.setsockopt	= mptcp_setsockopt,
3834	.getsockopt	= mptcp_getsockopt,
3835	.shutdown	= mptcp_shutdown,
3836	.destroy	= mptcp_destroy,
3837	.sendmsg	= mptcp_sendmsg,
3838	.ioctl		= mptcp_ioctl,
3839	.recvmsg	= mptcp_recvmsg,
3840	.release_cb	= mptcp_release_cb,
3841	.hash		= mptcp_hash,
3842	.unhash		= mptcp_unhash,
3843	.get_port	= mptcp_get_port,
3844	.forward_alloc_get	= mptcp_forward_alloc_get,
3845	.stream_memory_free	= mptcp_stream_memory_free,
3846	.sockets_allocated	= &mptcp_sockets_allocated,
3847
3848	.memory_allocated	= &tcp_memory_allocated,
3849	.per_cpu_fw_alloc	= &tcp_memory_per_cpu_fw_alloc,
3850
3851	.memory_pressure	= &tcp_memory_pressure,
3852	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_wmem),
3853	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_rmem),
3854	.sysctl_mem	= sysctl_tcp_mem,
3855	.obj_size	= sizeof(struct mptcp_sock),
3856	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
3857	.no_autobind	= true,
3858};
3859
3860static int mptcp_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len)
3861{
3862	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3863	struct sock *ssk, *sk = sock->sk;
3864	int err = -EINVAL;
3865
3866	lock_sock(sk);
3867	ssk = __mptcp_nmpc_sk(msk);
3868	if (IS_ERR(ssk)) {
3869		err = PTR_ERR(ssk);
3870		goto unlock;
3871	}
3872
3873	if (sk->sk_family == AF_INET)
3874		err = inet_bind_sk(ssk, uaddr, addr_len);
3875#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3876	else if (sk->sk_family == AF_INET6)
3877		err = inet6_bind_sk(ssk, uaddr, addr_len);
3878#endif
3879	if (!err)
3880		mptcp_copy_inaddrs(sk, ssk);
3881
3882unlock:
3883	release_sock(sk);
3884	return err;
3885}
3886
3887static int mptcp_listen(struct socket *sock, int backlog)
3888{
3889	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3890	struct sock *sk = sock->sk;
3891	struct sock *ssk;
3892	int err;
3893
3894	pr_debug("msk=%p\n", msk);
3895
3896	lock_sock(sk);
3897
3898	err = -EINVAL;
3899	if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
3900		goto unlock;
3901
3902	ssk = __mptcp_nmpc_sk(msk);
3903	if (IS_ERR(ssk)) {
3904		err = PTR_ERR(ssk);
3905		goto unlock;
3906	}
3907
3908	mptcp_set_state(sk, TCP_LISTEN);
3909	sock_set_flag(sk, SOCK_RCU_FREE);
3910
3911	lock_sock(ssk);
3912	err = __inet_listen_sk(ssk, backlog);
3913	release_sock(ssk);
3914	mptcp_set_state(sk, inet_sk_state_load(ssk));
3915
3916	if (!err) {
3917		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
3918		mptcp_copy_inaddrs(sk, ssk);
3919		mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
3920	}
3921
3922unlock:
3923	release_sock(sk);
3924	return err;
3925}
3926
3927static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
3928			       struct proto_accept_arg *arg)
3929{
3930	struct mptcp_sock *msk = mptcp_sk(sock->sk);
3931	struct sock *ssk, *newsk;
3932
3933	pr_debug("msk=%p\n", msk);
3934
3935	/* Buggy applications can call accept on socket states other then LISTEN
3936	 * but no need to allocate the first subflow just to error out.
3937	 */
3938	ssk = READ_ONCE(msk->first);
3939	if (!ssk)
3940		return -EINVAL;
3941
3942	pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk));
3943	newsk = inet_csk_accept(ssk, arg);
3944	if (!newsk)
3945		return arg->err;
3946
3947	pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk));
3948	if (sk_is_mptcp(newsk)) {
3949		struct mptcp_subflow_context *subflow;
3950		struct sock *new_mptcp_sock;
3951
3952		subflow = mptcp_subflow_ctx(newsk);
3953		new_mptcp_sock = subflow->conn;
3954
3955		/* is_mptcp should be false if subflow->conn is missing, see
3956		 * subflow_syn_recv_sock()
3957		 */
3958		if (WARN_ON_ONCE(!new_mptcp_sock)) {
3959			tcp_sk(newsk)->is_mptcp = 0;
3960			goto tcpfallback;
3961		}
3962
3963		newsk = new_mptcp_sock;
3964		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
3965
3966		newsk->sk_kern_sock = arg->kern;
3967		lock_sock(newsk);
3968		__inet_accept(sock, newsock, newsk);
3969
3970		set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
3971		msk = mptcp_sk(newsk);
3972		msk->in_accept_queue = 0;
3973
3974		/* set ssk->sk_socket of accept()ed flows to mptcp socket.
3975		 * This is needed so NOSPACE flag can be set from tcp stack.
3976		 */
3977		mptcp_for_each_subflow(msk, subflow) {
3978			struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3979
3980			if (!ssk->sk_socket)
3981				mptcp_sock_graft(ssk, newsock);
3982		}
3983
3984		/* Do late cleanup for the first subflow as necessary. Also
3985		 * deal with bad peers not doing a complete shutdown.
3986		 */
3987		if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
3988			__mptcp_close_ssk(newsk, msk->first,
3989					  mptcp_subflow_ctx(msk->first), 0);
3990			if (unlikely(list_is_singular(&msk->conn_list)))
3991				mptcp_set_state(newsk, TCP_CLOSE);
3992		}
3993	} else {
3994tcpfallback:
3995		newsk->sk_kern_sock = arg->kern;
3996		lock_sock(newsk);
3997		__inet_accept(sock, newsock, newsk);
3998		/* we are being invoked after accepting a non-mp-capable
3999		 * flow: sk is a tcp_sk, not an mptcp one.
4000		 *
4001		 * Hand the socket over to tcp so all further socket ops
4002		 * bypass mptcp.
4003		 */
4004		WRITE_ONCE(newsock->sk->sk_socket->ops,
4005			   mptcp_fallback_tcp_ops(newsock->sk));
4006	}
4007	release_sock(newsk);
4008
4009	return 0;
4010}
4011
4012static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
4013{
4014	struct sock *sk = (struct sock *)msk;
4015
4016	if (__mptcp_stream_is_writeable(sk, 1))
4017		return EPOLLOUT | EPOLLWRNORM;
4018
4019	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
4020	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
4021	if (__mptcp_stream_is_writeable(sk, 1))
4022		return EPOLLOUT | EPOLLWRNORM;
4023
4024	return 0;
4025}
4026
4027static __poll_t mptcp_poll(struct file *file, struct socket *sock,
4028			   struct poll_table_struct *wait)
4029{
4030	struct sock *sk = sock->sk;
4031	struct mptcp_sock *msk;
4032	__poll_t mask = 0;
4033	u8 shutdown;
4034	int state;
4035
4036	msk = mptcp_sk(sk);
4037	sock_poll_wait(file, sock, wait);
4038
4039	state = inet_sk_state_load(sk);
4040	pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags);
4041	if (state == TCP_LISTEN) {
4042		struct sock *ssk = READ_ONCE(msk->first);
4043
4044		if (WARN_ON_ONCE(!ssk))
4045			return 0;
4046
4047		return inet_csk_listen_poll(ssk);
4048	}
4049
4050	shutdown = READ_ONCE(sk->sk_shutdown);
4051	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4052		mask |= EPOLLHUP;
4053	if (shutdown & RCV_SHUTDOWN)
4054		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4055
4056	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4057		mask |= mptcp_check_readable(sk);
4058		if (shutdown & SEND_SHUTDOWN)
4059			mask |= EPOLLOUT | EPOLLWRNORM;
4060		else
4061			mask |= mptcp_check_writeable(msk);
4062	} else if (state == TCP_SYN_SENT &&
4063		   inet_test_bit(DEFER_CONNECT, sk)) {
4064		/* cf tcp_poll() note about TFO */
4065		mask |= EPOLLOUT | EPOLLWRNORM;
4066	}
4067
4068	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4069	smp_rmb();
4070	if (READ_ONCE(sk->sk_err))
4071		mask |= EPOLLERR;
4072
4073	return mask;
4074}
4075
4076static const struct proto_ops mptcp_stream_ops = {
4077	.family		   = PF_INET,
4078	.owner		   = THIS_MODULE,
4079	.release	   = inet_release,
4080	.bind		   = mptcp_bind,
4081	.connect	   = inet_stream_connect,
4082	.socketpair	   = sock_no_socketpair,
4083	.accept		   = mptcp_stream_accept,
4084	.getname	   = inet_getname,
4085	.poll		   = mptcp_poll,
4086	.ioctl		   = inet_ioctl,
4087	.gettstamp	   = sock_gettstamp,
4088	.listen		   = mptcp_listen,
4089	.shutdown	   = inet_shutdown,
4090	.setsockopt	   = sock_common_setsockopt,
4091	.getsockopt	   = sock_common_getsockopt,
4092	.sendmsg	   = inet_sendmsg,
4093	.recvmsg	   = inet_recvmsg,
4094	.mmap		   = sock_no_mmap,
4095	.set_rcvlowat	   = mptcp_set_rcvlowat,
4096};
4097
4098static struct inet_protosw mptcp_protosw = {
4099	.type		= SOCK_STREAM,
4100	.protocol	= IPPROTO_MPTCP,
4101	.prot		= &mptcp_prot,
4102	.ops		= &mptcp_stream_ops,
4103	.flags		= INET_PROTOSW_ICSK,
4104};
4105
4106static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4107{
4108	struct mptcp_delegated_action *delegated;
4109	struct mptcp_subflow_context *subflow;
4110	int work_done = 0;
4111
4112	delegated = container_of(napi, struct mptcp_delegated_action, napi);
4113	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4114		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4115
4116		bh_lock_sock_nested(ssk);
4117		if (!sock_owned_by_user(ssk)) {
4118			mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4119		} else {
4120			/* tcp_release_cb_override already processed
4121			 * the action or will do at next release_sock().
4122			 * In both case must dequeue the subflow here - on the same
4123			 * CPU that scheduled it.
4124			 */
4125			smp_wmb();
4126			clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4127		}
4128		bh_unlock_sock(ssk);
4129		sock_put(ssk);
4130
4131		if (++work_done == budget)
4132			return budget;
4133	}
4134
4135	/* always provide a 0 'work_done' argument, so that napi_complete_done
4136	 * will not try accessing the NULL napi->dev ptr
4137	 */
4138	napi_complete_done(napi, 0);
4139	return work_done;
4140}
4141
4142void __init mptcp_proto_init(void)
4143{
4144	struct mptcp_delegated_action *delegated;
4145	int cpu;
4146
4147	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4148
4149	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4150		panic("Failed to allocate MPTCP pcpu counter\n");
4151
4152	init_dummy_netdev(&mptcp_napi_dev);
4153	for_each_possible_cpu(cpu) {
4154		delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4155		INIT_LIST_HEAD(&delegated->head);
4156		netif_napi_add_tx(&mptcp_napi_dev, &delegated->napi,
4157				  mptcp_napi_poll);
4158		napi_enable(&delegated->napi);
4159	}
4160
4161	mptcp_subflow_init();
4162	mptcp_pm_init();
4163	mptcp_sched_init();
4164	mptcp_token_init();
4165
4166	if (proto_register(&mptcp_prot, 1) != 0)
4167		panic("Failed to register MPTCP proto.\n");
4168
4169	inet_register_protosw(&mptcp_protosw);
4170
4171	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4172}
4173
4174#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4175static const struct proto_ops mptcp_v6_stream_ops = {
4176	.family		   = PF_INET6,
4177	.owner		   = THIS_MODULE,
4178	.release	   = inet6_release,
4179	.bind		   = mptcp_bind,
4180	.connect	   = inet_stream_connect,
4181	.socketpair	   = sock_no_socketpair,
4182	.accept		   = mptcp_stream_accept,
4183	.getname	   = inet6_getname,
4184	.poll		   = mptcp_poll,
4185	.ioctl		   = inet6_ioctl,
4186	.gettstamp	   = sock_gettstamp,
4187	.listen		   = mptcp_listen,
4188	.shutdown	   = inet_shutdown,
4189	.setsockopt	   = sock_common_setsockopt,
4190	.getsockopt	   = sock_common_getsockopt,
4191	.sendmsg	   = inet6_sendmsg,
4192	.recvmsg	   = inet6_recvmsg,
4193	.mmap		   = sock_no_mmap,
4194#ifdef CONFIG_COMPAT
4195	.compat_ioctl	   = inet6_compat_ioctl,
4196#endif
4197	.set_rcvlowat	   = mptcp_set_rcvlowat,
4198};
4199
4200static struct proto mptcp_v6_prot;
4201
4202static struct inet_protosw mptcp_v6_protosw = {
4203	.type		= SOCK_STREAM,
4204	.protocol	= IPPROTO_MPTCP,
4205	.prot		= &mptcp_v6_prot,
4206	.ops		= &mptcp_v6_stream_ops,
4207	.flags		= INET_PROTOSW_ICSK,
4208};
4209
4210int __init mptcp_proto_v6_init(void)
4211{
4212	int err;
4213
4214	mptcp_v6_prot = mptcp_prot;
4215	strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4216	mptcp_v6_prot.slab = NULL;
4217	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4218	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4219
4220	err = proto_register(&mptcp_v6_prot, 1);
4221	if (err)
4222		return err;
4223
4224	err = inet6_register_protosw(&mptcp_v6_protosw);
4225	if (err)
4226		proto_unregister(&mptcp_v6_prot);
4227
4228	return err;
4229}
4230#endif