protocol.c

   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/aligned_data.h>
  15#include <net/rps.h>
  16#include <net/sock.h>
  17#include <net/inet_common.h>
  18#include <net/inet_hashtables.h>
  19#include <net/protocol.h>
  20#include <net/tcp_states.h>
  21#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  22#include <net/transp_v6.h>
  23#endif
  24#include <net/mptcp.h>
  25#include <net/hotdata.h>
  26#include <net/xfrm.h>
  27#include <asm/ioctls.h>
  28#include "protocol.h"
  29#include "mib.h"
  30
  31#define CREATE_TRACE_POINTS
  32#include <trace/events/mptcp.h>
  33
  34#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  35struct mptcp6_sock {
  36	struct mptcp_sock msk;
  37	struct ipv6_pinfo np;
  38};
  39#endif
  40
  41enum {
  42	MPTCP_CMSG_TS = BIT(0),
  43	MPTCP_CMSG_INQ = BIT(1),
  44};
  45
  46static struct percpu_counter mptcp_sockets_allocated ____cacheline_aligned_in_smp;
  47
  48static void __mptcp_destroy_sock(struct sock *sk);
  49static void mptcp_check_send_data_fin(struct sock *sk);
  50
  51DEFINE_PER_CPU(struct mptcp_delegated_action, mptcp_delegated_actions) = {
  52	.bh_lock = INIT_LOCAL_LOCK(bh_lock),
  53};
  54static struct net_device *mptcp_napi_dev;
  55
  56/* Returns end sequence number of the receiver's advertised window */
  57static u64 mptcp_wnd_end(const struct mptcp_sock *msk)
  58{
  59	return READ_ONCE(msk->wnd_end);
  60}
  61
  62static const struct proto_ops *mptcp_fallback_tcp_ops(const struct sock *sk)
  63{
  64	unsigned short family = READ_ONCE(sk->sk_family);
  65
  66#if IS_ENABLED(CONFIG_MPTCP_IPV6)
  67	if (family == AF_INET6)
  68		return &inet6_stream_ops;
  69#endif
  70	WARN_ON_ONCE(family != AF_INET);
  71	return &inet_stream_ops;
  72}
  73
  74bool __mptcp_try_fallback(struct mptcp_sock *msk, int fb_mib)
  75{
  76	struct net *net = sock_net((struct sock *)msk);
  77
  78	if (__mptcp_check_fallback(msk))
  79		return true;
  80
  81	/* The caller possibly is not holding the msk socket lock, but
  82	 * in the fallback case only the current subflow is touching
  83	 * the OoO queue.
  84	 */
  85	if (!RB_EMPTY_ROOT(&msk->out_of_order_queue))
  86		return false;
  87
  88	spin_lock_bh(&msk->fallback_lock);
  89	if (!msk->allow_infinite_fallback) {
  90		spin_unlock_bh(&msk->fallback_lock);
  91		return false;
  92	}
  93
  94	msk->allow_subflows = false;
  95	set_bit(MPTCP_FALLBACK_DONE, &msk->flags);
  96	__MPTCP_INC_STATS(net, fb_mib);
  97	spin_unlock_bh(&msk->fallback_lock);
  98	return true;
  99}
 100
 101static int __mptcp_socket_create(struct mptcp_sock *msk)
 102{
 103	struct mptcp_subflow_context *subflow;
 104	struct sock *sk = (struct sock *)msk;
 105	struct socket *ssock;
 106	int err;
 107
 108	err = mptcp_subflow_create_socket(sk, sk->sk_family, &ssock);
 109	if (err)
 110		return err;
 111
 112	msk->scaling_ratio = tcp_sk(ssock->sk)->scaling_ratio;
 113	WRITE_ONCE(msk->first, ssock->sk);
 114	subflow = mptcp_subflow_ctx(ssock->sk);
 115	list_add(&subflow->node, &msk->conn_list);
 116	sock_hold(ssock->sk);
 117	subflow->request_mptcp = 1;
 118	subflow->subflow_id = msk->subflow_id++;
 119
 120	/* This is the first subflow, always with id 0 */
 121	WRITE_ONCE(subflow->local_id, 0);
 122	mptcp_sock_graft(msk->first, sk->sk_socket);
 123	iput(SOCK_INODE(ssock));
 124
 125	return 0;
 126}
 127
 128/* If the MPC handshake is not started, returns the first subflow,
 129 * eventually allocating it.
 130 */
 131struct sock *__mptcp_nmpc_sk(struct mptcp_sock *msk)
 132{
 133	struct sock *sk = (struct sock *)msk;
 134	int ret;
 135
 136	if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
 137		return ERR_PTR(-EINVAL);
 138
 139	if (!msk->first) {
 140		ret = __mptcp_socket_create(msk);
 141		if (ret)
 142			return ERR_PTR(ret);
 143	}
 144
 145	return msk->first;
 146}
 147
 148static void mptcp_drop(struct sock *sk, struct sk_buff *skb)
 149{
 150	sk_drops_skbadd(sk, skb);
 151	__kfree_skb(skb);
 152}
 153
 154static bool __mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
 155				 struct sk_buff *from, bool *fragstolen,
 156				 int *delta)
 157{
 158	int limit = READ_ONCE(sk->sk_rcvbuf);
 159
 160	if (unlikely(MPTCP_SKB_CB(to)->cant_coalesce) ||
 161	    MPTCP_SKB_CB(from)->offset ||
 162	    ((to->len + from->len) > (limit >> 3)) ||
 163	    !skb_try_coalesce(to, from, fragstolen, delta))
 164		return false;
 165
 166	pr_debug("colesced seq %llx into %llx new len %d new end seq %llx\n",
 167		 MPTCP_SKB_CB(from)->map_seq, MPTCP_SKB_CB(to)->map_seq,
 168		 to->len, MPTCP_SKB_CB(from)->end_seq);
 169	MPTCP_SKB_CB(to)->end_seq = MPTCP_SKB_CB(from)->end_seq;
 170	return true;
 171}
 172
 173static bool mptcp_try_coalesce(struct sock *sk, struct sk_buff *to,
 174			       struct sk_buff *from)
 175{
 176	bool fragstolen;
 177	int delta;
 178
 179	if (!__mptcp_try_coalesce(sk, to, from, &fragstolen, &delta))
 180		return false;
 181
 182	/* note the fwd memory can reach a negative value after accounting
 183	 * for the delta, but the later skb free will restore a non
 184	 * negative one
 185	 */
 186	atomic_add(delta, &sk->sk_rmem_alloc);
 187	sk_mem_charge(sk, delta);
 188	kfree_skb_partial(from, fragstolen);
 189
 190	return true;
 191}
 192
 193static bool mptcp_ooo_try_coalesce(struct mptcp_sock *msk, struct sk_buff *to,
 194				   struct sk_buff *from)
 195{
 196	if (MPTCP_SKB_CB(from)->map_seq != MPTCP_SKB_CB(to)->end_seq)
 197		return false;
 198
 199	return mptcp_try_coalesce((struct sock *)msk, to, from);
 200}
 201
 202/* "inspired" by tcp_rcvbuf_grow(), main difference:
 203 * - mptcp does not maintain a msk-level window clamp
 204 * - returns true when  the receive buffer is actually updated
 205 */
 206static bool mptcp_rcvbuf_grow(struct sock *sk, u32 newval)
 207{
 208	struct mptcp_sock *msk = mptcp_sk(sk);
 209	const struct net *net = sock_net(sk);
 210	u32 rcvwin, rcvbuf, cap, oldval;
 211	u64 grow;
 212
 213	oldval = msk->rcvq_space.space;
 214	msk->rcvq_space.space = newval;
 215	if (!READ_ONCE(net->ipv4.sysctl_tcp_moderate_rcvbuf) ||
 216	    (sk->sk_userlocks & SOCK_RCVBUF_LOCK))
 217		return false;
 218
 219	/* DRS is always one RTT late. */
 220	rcvwin = newval << 1;
 221
 222	/* slow start: allow the sender to double its rate. */
 223	grow = (u64)rcvwin * (newval - oldval);
 224	do_div(grow, oldval);
 225	rcvwin += grow << 1;
 226
 227	cap = READ_ONCE(net->ipv4.sysctl_tcp_rmem[2]);
 228
 229	rcvbuf = min_t(u32, mptcp_space_from_win(sk, rcvwin), cap);
 230	if (rcvbuf > sk->sk_rcvbuf) {
 231		WRITE_ONCE(sk->sk_rcvbuf, rcvbuf);
 232		return true;
 233	}
 234	return false;
 235}
 236
 237/* "inspired" by tcp_data_queue_ofo(), main differences:
 238 * - use mptcp seqs
 239 * - don't cope with sacks
 240 */
 241static void mptcp_data_queue_ofo(struct mptcp_sock *msk, struct sk_buff *skb)
 242{
 243	struct sock *sk = (struct sock *)msk;
 244	struct rb_node **p, *parent;
 245	u64 seq, end_seq, max_seq;
 246	struct sk_buff *skb1;
 247
 248	seq = MPTCP_SKB_CB(skb)->map_seq;
 249	end_seq = MPTCP_SKB_CB(skb)->end_seq;
 250	max_seq = atomic64_read(&msk->rcv_wnd_sent);
 251
 252	pr_debug("msk=%p seq=%llx limit=%llx empty=%d\n", msk, seq, max_seq,
 253		 RB_EMPTY_ROOT(&msk->out_of_order_queue));
 254	if (after64(end_seq, max_seq)) {
 255		/* out of window */
 256		mptcp_drop(sk, skb);
 257		pr_debug("oow by %lld, rcv_wnd_sent %llu\n",
 258			 (unsigned long long)end_seq - (unsigned long)max_seq,
 259			 (unsigned long long)atomic64_read(&msk->rcv_wnd_sent));
 260		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_NODSSWINDOW);
 261		return;
 262	}
 263
 264	p = &msk->out_of_order_queue.rb_node;
 265	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUE);
 266	if (RB_EMPTY_ROOT(&msk->out_of_order_queue)) {
 267		rb_link_node(&skb->rbnode, NULL, p);
 268		rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 269		msk->ooo_last_skb = skb;
 270		goto end;
 271	}
 272
 273	/* with 2 subflows, adding at end of ooo queue is quite likely
 274	 * Use of ooo_last_skb avoids the O(Log(N)) rbtree lookup.
 275	 */
 276	if (mptcp_ooo_try_coalesce(msk, msk->ooo_last_skb, skb)) {
 277		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 278		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 279		return;
 280	}
 281
 282	/* Can avoid an rbtree lookup if we are adding skb after ooo_last_skb */
 283	if (!before64(seq, MPTCP_SKB_CB(msk->ooo_last_skb)->end_seq)) {
 284		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOQUEUETAIL);
 285		parent = &msk->ooo_last_skb->rbnode;
 286		p = &parent->rb_right;
 287		goto insert;
 288	}
 289
 290	/* Find place to insert this segment. Handle overlaps on the way. */
 291	parent = NULL;
 292	while (*p) {
 293		parent = *p;
 294		skb1 = rb_to_skb(parent);
 295		if (before64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 296			p = &parent->rb_left;
 297			continue;
 298		}
 299		if (before64(seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 300			if (!after64(end_seq, MPTCP_SKB_CB(skb1)->end_seq)) {
 301				/* All the bits are present. Drop. */
 302				mptcp_drop(sk, skb);
 303				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 304				return;
 305			}
 306			if (after64(seq, MPTCP_SKB_CB(skb1)->map_seq)) {
 307				/* partial overlap:
 308				 *     |     skb      |
 309				 *  |     skb1    |
 310				 * continue traversing
 311				 */
 312			} else {
 313				/* skb's seq == skb1's seq and skb covers skb1.
 314				 * Replace skb1 with skb.
 315				 */
 316				rb_replace_node(&skb1->rbnode, &skb->rbnode,
 317						&msk->out_of_order_queue);
 318				mptcp_drop(sk, skb1);
 319				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 320				goto merge_right;
 321			}
 322		} else if (mptcp_ooo_try_coalesce(msk, skb1, skb)) {
 323			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_OFOMERGE);
 324			return;
 325		}
 326		p = &parent->rb_right;
 327	}
 328
 329insert:
 330	/* Insert segment into RB tree. */
 331	rb_link_node(&skb->rbnode, parent, p);
 332	rb_insert_color(&skb->rbnode, &msk->out_of_order_queue);
 333
 334merge_right:
 335	/* Remove other segments covered by skb. */
 336	while ((skb1 = skb_rb_next(skb)) != NULL) {
 337		if (before64(end_seq, MPTCP_SKB_CB(skb1)->end_seq))
 338			break;
 339		rb_erase(&skb1->rbnode, &msk->out_of_order_queue);
 340		mptcp_drop(sk, skb1);
 341		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 342	}
 343	/* If there is no skb after us, we are the last_skb ! */
 344	if (!skb1)
 345		msk->ooo_last_skb = skb;
 346
 347end:
 348	skb_condense(skb);
 349	skb_set_owner_r(skb, sk);
 350}
 351
 352static void mptcp_init_skb(struct sock *ssk, struct sk_buff *skb, int offset,
 353			   int copy_len)
 354{
 355	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 356	bool has_rxtstamp = TCP_SKB_CB(skb)->has_rxtstamp;
 357
 358	/* the skb map_seq accounts for the skb offset:
 359	 * mptcp_subflow_get_mapped_dsn() is based on the current tp->copied_seq
 360	 * value
 361	 */
 362	MPTCP_SKB_CB(skb)->map_seq = mptcp_subflow_get_mapped_dsn(subflow);
 363	MPTCP_SKB_CB(skb)->end_seq = MPTCP_SKB_CB(skb)->map_seq + copy_len;
 364	MPTCP_SKB_CB(skb)->offset = offset;
 365	MPTCP_SKB_CB(skb)->has_rxtstamp = has_rxtstamp;
 366	MPTCP_SKB_CB(skb)->cant_coalesce = 0;
 367
 368	__skb_unlink(skb, &ssk->sk_receive_queue);
 369
 370	skb_ext_reset(skb);
 371	skb_dst_drop(skb);
 372}
 373
 374static bool __mptcp_move_skb(struct sock *sk, struct sk_buff *skb)
 375{
 376	u64 copy_len = MPTCP_SKB_CB(skb)->end_seq - MPTCP_SKB_CB(skb)->map_seq;
 377	struct mptcp_sock *msk = mptcp_sk(sk);
 378	struct sk_buff *tail;
 379
 380	mptcp_borrow_fwdmem(sk, skb);
 381
 382	if (MPTCP_SKB_CB(skb)->map_seq == msk->ack_seq) {
 383		/* in sequence */
 384		msk->bytes_received += copy_len;
 385		WRITE_ONCE(msk->ack_seq, msk->ack_seq + copy_len);
 386		tail = skb_peek_tail(&sk->sk_receive_queue);
 387		if (tail && mptcp_try_coalesce(sk, tail, skb))
 388			return true;
 389
 390		skb_set_owner_r(skb, sk);
 391		__skb_queue_tail(&sk->sk_receive_queue, skb);
 392		return true;
 393	} else if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq)) {
 394		mptcp_data_queue_ofo(msk, skb);
 395		return false;
 396	}
 397
 398	/* old data, keep it simple and drop the whole pkt, sender
 399	 * will retransmit as needed, if needed.
 400	 */
 401	MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 402	mptcp_drop(sk, skb);
 403	return false;
 404}
 405
 406static void mptcp_stop_rtx_timer(struct sock *sk)
 407{
 408	sk_stop_timer(sk, &sk->mptcp_retransmit_timer);
 409	mptcp_sk(sk)->timer_ival = 0;
 410}
 411
 412static void mptcp_close_wake_up(struct sock *sk)
 413{
 414	if (sock_flag(sk, SOCK_DEAD))
 415		return;
 416
 417	sk->sk_state_change(sk);
 418	if (sk->sk_shutdown == SHUTDOWN_MASK ||
 419	    sk->sk_state == TCP_CLOSE)
 420		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
 421	else
 422		sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
 423}
 424
 425static void mptcp_shutdown_subflows(struct mptcp_sock *msk)
 426{
 427	struct mptcp_subflow_context *subflow;
 428
 429	mptcp_for_each_subflow(msk, subflow) {
 430		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 431		bool slow;
 432
 433		slow = lock_sock_fast(ssk);
 434		tcp_shutdown(ssk, SEND_SHUTDOWN);
 435		unlock_sock_fast(ssk, slow);
 436	}
 437}
 438
 439/* called under the msk socket lock */
 440static bool mptcp_pending_data_fin_ack(struct sock *sk)
 441{
 442	struct mptcp_sock *msk = mptcp_sk(sk);
 443
 444	return ((1 << sk->sk_state) &
 445		(TCPF_FIN_WAIT1 | TCPF_CLOSING | TCPF_LAST_ACK)) &&
 446	       msk->write_seq == READ_ONCE(msk->snd_una);
 447}
 448
 449static void mptcp_check_data_fin_ack(struct sock *sk)
 450{
 451	struct mptcp_sock *msk = mptcp_sk(sk);
 452
 453	/* Look for an acknowledged DATA_FIN */
 454	if (mptcp_pending_data_fin_ack(sk)) {
 455		WRITE_ONCE(msk->snd_data_fin_enable, 0);
 456
 457		switch (sk->sk_state) {
 458		case TCP_FIN_WAIT1:
 459			mptcp_set_state(sk, TCP_FIN_WAIT2);
 460			break;
 461		case TCP_CLOSING:
 462		case TCP_LAST_ACK:
 463			mptcp_shutdown_subflows(msk);
 464			mptcp_set_state(sk, TCP_CLOSE);
 465			break;
 466		}
 467
 468		mptcp_close_wake_up(sk);
 469	}
 470}
 471
 472/* can be called with no lock acquired */
 473static bool mptcp_pending_data_fin(struct sock *sk, u64 *seq)
 474{
 475	struct mptcp_sock *msk = mptcp_sk(sk);
 476
 477	if (READ_ONCE(msk->rcv_data_fin) &&
 478	    ((1 << inet_sk_state_load(sk)) &
 479	     (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2))) {
 480		u64 rcv_data_fin_seq = READ_ONCE(msk->rcv_data_fin_seq);
 481
 482		if (READ_ONCE(msk->ack_seq) == rcv_data_fin_seq) {
 483			if (seq)
 484				*seq = rcv_data_fin_seq;
 485
 486			return true;
 487		}
 488	}
 489
 490	return false;
 491}
 492
 493static void mptcp_set_datafin_timeout(struct sock *sk)
 494{
 495	struct inet_connection_sock *icsk = inet_csk(sk);
 496	u32 retransmits;
 497
 498	retransmits = min_t(u32, icsk->icsk_retransmits,
 499			    ilog2(TCP_RTO_MAX / TCP_RTO_MIN));
 500
 501	mptcp_sk(sk)->timer_ival = TCP_RTO_MIN << retransmits;
 502}
 503
 504static void __mptcp_set_timeout(struct sock *sk, long tout)
 505{
 506	mptcp_sk(sk)->timer_ival = tout > 0 ? tout : TCP_RTO_MIN;
 507}
 508
 509static long mptcp_timeout_from_subflow(const struct mptcp_subflow_context *subflow)
 510{
 511	const struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 512
 513	return inet_csk(ssk)->icsk_pending && !subflow->stale_count ?
 514	       tcp_timeout_expires(ssk) - jiffies : 0;
 515}
 516
 517static void mptcp_set_timeout(struct sock *sk)
 518{
 519	struct mptcp_subflow_context *subflow;
 520	long tout = 0;
 521
 522	mptcp_for_each_subflow(mptcp_sk(sk), subflow)
 523		tout = max(tout, mptcp_timeout_from_subflow(subflow));
 524	__mptcp_set_timeout(sk, tout);
 525}
 526
 527static inline bool tcp_can_send_ack(const struct sock *ssk)
 528{
 529	return !((1 << inet_sk_state_load(ssk)) &
 530	       (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_TIME_WAIT | TCPF_CLOSE | TCPF_LISTEN));
 531}
 532
 533void __mptcp_subflow_send_ack(struct sock *ssk)
 534{
 535	if (tcp_can_send_ack(ssk))
 536		tcp_send_ack(ssk);
 537}
 538
 539static void mptcp_subflow_send_ack(struct sock *ssk)
 540{
 541	bool slow;
 542
 543	slow = lock_sock_fast(ssk);
 544	__mptcp_subflow_send_ack(ssk);
 545	unlock_sock_fast(ssk, slow);
 546}
 547
 548static void mptcp_send_ack(struct mptcp_sock *msk)
 549{
 550	struct mptcp_subflow_context *subflow;
 551
 552	mptcp_for_each_subflow(msk, subflow)
 553		mptcp_subflow_send_ack(mptcp_subflow_tcp_sock(subflow));
 554}
 555
 556static void mptcp_subflow_cleanup_rbuf(struct sock *ssk, int copied)
 557{
 558	bool slow;
 559
 560	slow = lock_sock_fast(ssk);
 561	if (tcp_can_send_ack(ssk))
 562		tcp_cleanup_rbuf(ssk, copied);
 563	unlock_sock_fast(ssk, slow);
 564}
 565
 566static bool mptcp_subflow_could_cleanup(const struct sock *ssk, bool rx_empty)
 567{
 568	const struct inet_connection_sock *icsk = inet_csk(ssk);
 569	u8 ack_pending = READ_ONCE(icsk->icsk_ack.pending);
 570	const struct tcp_sock *tp = tcp_sk(ssk);
 571
 572	return (ack_pending & ICSK_ACK_SCHED) &&
 573		((READ_ONCE(tp->rcv_nxt) - READ_ONCE(tp->rcv_wup) >
 574		  READ_ONCE(icsk->icsk_ack.rcv_mss)) ||
 575		 (rx_empty && ack_pending &
 576			      (ICSK_ACK_PUSHED2 | ICSK_ACK_PUSHED)));
 577}
 578
 579static void mptcp_cleanup_rbuf(struct mptcp_sock *msk, int copied)
 580{
 581	int old_space = READ_ONCE(msk->old_wspace);
 582	struct mptcp_subflow_context *subflow;
 583	struct sock *sk = (struct sock *)msk;
 584	int space =  __mptcp_space(sk);
 585	bool cleanup, rx_empty;
 586
 587	cleanup = (space > 0) && (space >= (old_space << 1)) && copied;
 588	rx_empty = !sk_rmem_alloc_get(sk) && copied;
 589
 590	mptcp_for_each_subflow(msk, subflow) {
 591		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 592
 593		if (cleanup || mptcp_subflow_could_cleanup(ssk, rx_empty))
 594			mptcp_subflow_cleanup_rbuf(ssk, copied);
 595	}
 596}
 597
 598static void mptcp_check_data_fin(struct sock *sk)
 599{
 600	struct mptcp_sock *msk = mptcp_sk(sk);
 601	u64 rcv_data_fin_seq;
 602
 603	/* Need to ack a DATA_FIN received from a peer while this side
 604	 * of the connection is in ESTABLISHED, FIN_WAIT1, or FIN_WAIT2.
 605	 * msk->rcv_data_fin was set when parsing the incoming options
 606	 * at the subflow level and the msk lock was not held, so this
 607	 * is the first opportunity to act on the DATA_FIN and change
 608	 * the msk state.
 609	 *
 610	 * If we are caught up to the sequence number of the incoming
 611	 * DATA_FIN, send the DATA_ACK now and do state transition.  If
 612	 * not caught up, do nothing and let the recv code send DATA_ACK
 613	 * when catching up.
 614	 */
 615
 616	if (mptcp_pending_data_fin(sk, &rcv_data_fin_seq)) {
 617		WRITE_ONCE(msk->ack_seq, msk->ack_seq + 1);
 618		WRITE_ONCE(msk->rcv_data_fin, 0);
 619
 620		WRITE_ONCE(sk->sk_shutdown, sk->sk_shutdown | RCV_SHUTDOWN);
 621		smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
 622
 623		switch (sk->sk_state) {
 624		case TCP_ESTABLISHED:
 625			mptcp_set_state(sk, TCP_CLOSE_WAIT);
 626			break;
 627		case TCP_FIN_WAIT1:
 628			mptcp_set_state(sk, TCP_CLOSING);
 629			break;
 630		case TCP_FIN_WAIT2:
 631			mptcp_shutdown_subflows(msk);
 632			mptcp_set_state(sk, TCP_CLOSE);
 633			break;
 634		default:
 635			/* Other states not expected */
 636			WARN_ON_ONCE(1);
 637			break;
 638		}
 639
 640		if (!__mptcp_check_fallback(msk))
 641			mptcp_send_ack(msk);
 642		mptcp_close_wake_up(sk);
 643	}
 644}
 645
 646static void mptcp_dss_corruption(struct mptcp_sock *msk, struct sock *ssk)
 647{
 648	if (!mptcp_try_fallback(ssk, MPTCP_MIB_DSSCORRUPTIONFALLBACK)) {
 649		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_DSSCORRUPTIONRESET);
 650		mptcp_subflow_reset(ssk);
 651	}
 652}
 653
 654static void __mptcp_add_backlog(struct sock *sk,
 655				struct mptcp_subflow_context *subflow,
 656				struct sk_buff *skb)
 657{
 658	struct mptcp_sock *msk = mptcp_sk(sk);
 659	struct sk_buff *tail = NULL;
 660	struct sock *ssk = skb->sk;
 661	bool fragstolen;
 662	int delta;
 663
 664	if (unlikely(sk->sk_state == TCP_CLOSE)) {
 665		kfree_skb_reason(skb, SKB_DROP_REASON_SOCKET_CLOSE);
 666		return;
 667	}
 668
 669	/* Try to coalesce with the last skb in our backlog */
 670	if (!list_empty(&msk->backlog_list))
 671		tail = list_last_entry(&msk->backlog_list, struct sk_buff, list);
 672
 673	if (tail && MPTCP_SKB_CB(skb)->map_seq == MPTCP_SKB_CB(tail)->end_seq &&
 674	    ssk == tail->sk &&
 675	    __mptcp_try_coalesce(sk, tail, skb, &fragstolen, &delta)) {
 676		skb->truesize -= delta;
 677		kfree_skb_partial(skb, fragstolen);
 678		__mptcp_subflow_lend_fwdmem(subflow, delta);
 679		goto account;
 680	}
 681
 682	list_add_tail(&skb->list, &msk->backlog_list);
 683	mptcp_subflow_lend_fwdmem(subflow, skb);
 684	delta = skb->truesize;
 685
 686account:
 687	WRITE_ONCE(msk->backlog_len, msk->backlog_len + delta);
 688
 689	/* Possibly not accept()ed yet, keep track of memory not CG
 690	 * accounted, mptcp_graft_subflows() will handle it.
 691	 */
 692	if (!mem_cgroup_from_sk(ssk))
 693		msk->backlog_unaccounted += delta;
 694}
 695
 696static bool __mptcp_move_skbs_from_subflow(struct mptcp_sock *msk,
 697					   struct sock *ssk, bool own_msk)
 698{
 699	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 700	struct sock *sk = (struct sock *)msk;
 701	bool more_data_avail;
 702	struct tcp_sock *tp;
 703	bool ret = false;
 704
 705	pr_debug("msk=%p ssk=%p\n", msk, ssk);
 706	tp = tcp_sk(ssk);
 707	do {
 708		u32 map_remaining, offset;
 709		u32 seq = tp->copied_seq;
 710		struct sk_buff *skb;
 711		bool fin;
 712
 713		/* try to move as much data as available */
 714		map_remaining = subflow->map_data_len -
 715				mptcp_subflow_get_map_offset(subflow);
 716
 717		skb = skb_peek(&ssk->sk_receive_queue);
 718		if (unlikely(!skb))
 719			break;
 720
 721		if (__mptcp_check_fallback(msk)) {
 722			/* Under fallback skbs have no MPTCP extension and TCP could
 723			 * collapse them between the dummy map creation and the
 724			 * current dequeue. Be sure to adjust the map size.
 725			 */
 726			map_remaining = skb->len;
 727			subflow->map_data_len = skb->len;
 728		}
 729
 730		offset = seq - TCP_SKB_CB(skb)->seq;
 731		fin = TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN;
 732		if (fin)
 733			seq++;
 734
 735		if (offset < skb->len) {
 736			size_t len = skb->len - offset;
 737
 738			mptcp_init_skb(ssk, skb, offset, len);
 739
 740			if (own_msk && sk_rmem_alloc_get(sk) < sk->sk_rcvbuf) {
 741				mptcp_subflow_lend_fwdmem(subflow, skb);
 742				ret |= __mptcp_move_skb(sk, skb);
 743			} else {
 744				__mptcp_add_backlog(sk, subflow, skb);
 745			}
 746			seq += len;
 747
 748			if (unlikely(map_remaining < len)) {
 749				DEBUG_NET_WARN_ON_ONCE(1);
 750				mptcp_dss_corruption(msk, ssk);
 751			}
 752		} else {
 753			if (unlikely(!fin)) {
 754				DEBUG_NET_WARN_ON_ONCE(1);
 755				mptcp_dss_corruption(msk, ssk);
 756			}
 757
 758			sk_eat_skb(ssk, skb);
 759		}
 760
 761		WRITE_ONCE(tp->copied_seq, seq);
 762		more_data_avail = mptcp_subflow_data_available(ssk);
 763
 764	} while (more_data_avail);
 765
 766	if (ret)
 767		msk->last_data_recv = tcp_jiffies32;
 768	return ret;
 769}
 770
 771static bool __mptcp_ofo_queue(struct mptcp_sock *msk)
 772{
 773	struct sock *sk = (struct sock *)msk;
 774	struct sk_buff *skb, *tail;
 775	bool moved = false;
 776	struct rb_node *p;
 777	u64 end_seq;
 778
 779	p = rb_first(&msk->out_of_order_queue);
 780	pr_debug("msk=%p empty=%d\n", msk, RB_EMPTY_ROOT(&msk->out_of_order_queue));
 781	while (p) {
 782		skb = rb_to_skb(p);
 783		if (after64(MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq))
 784			break;
 785
 786		p = rb_next(p);
 787		rb_erase(&skb->rbnode, &msk->out_of_order_queue);
 788
 789		if (unlikely(!after64(MPTCP_SKB_CB(skb)->end_seq,
 790				      msk->ack_seq))) {
 791			mptcp_drop(sk, skb);
 792			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_DUPDATA);
 793			continue;
 794		}
 795
 796		end_seq = MPTCP_SKB_CB(skb)->end_seq;
 797		tail = skb_peek_tail(&sk->sk_receive_queue);
 798		if (!tail || !mptcp_ooo_try_coalesce(msk, tail, skb)) {
 799			int delta = msk->ack_seq - MPTCP_SKB_CB(skb)->map_seq;
 800
 801			/* skip overlapping data, if any */
 802			pr_debug("uncoalesced seq=%llx ack seq=%llx delta=%d\n",
 803				 MPTCP_SKB_CB(skb)->map_seq, msk->ack_seq,
 804				 delta);
 805			MPTCP_SKB_CB(skb)->offset += delta;
 806			MPTCP_SKB_CB(skb)->map_seq += delta;
 807			__skb_queue_tail(&sk->sk_receive_queue, skb);
 808		}
 809		msk->bytes_received += end_seq - msk->ack_seq;
 810		WRITE_ONCE(msk->ack_seq, end_seq);
 811		moved = true;
 812	}
 813	return moved;
 814}
 815
 816static bool __mptcp_subflow_error_report(struct sock *sk, struct sock *ssk)
 817{
 818	int ssk_state;
 819	int err;
 820
 821	/* only propagate errors on fallen-back sockets or
 822	 * on MPC connect
 823	 */
 824	if (sk->sk_state != TCP_SYN_SENT && !__mptcp_check_fallback(mptcp_sk(sk)))
 825		return false;
 826
 827	err = sock_error(ssk);
 828	if (!err)
 829		return false;
 830
 831	/* We need to propagate only transition to CLOSE state.
 832	 * Orphaned socket will see such state change via
 833	 * subflow_sched_work_if_closed() and that path will properly
 834	 * destroy the msk as needed.
 835	 */
 836	ssk_state = inet_sk_state_load(ssk);
 837	if (ssk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DEAD))
 838		mptcp_set_state(sk, ssk_state);
 839	WRITE_ONCE(sk->sk_err, -err);
 840
 841	/* This barrier is coupled with smp_rmb() in mptcp_poll() */
 842	smp_wmb();
 843	sk_error_report(sk);
 844	return true;
 845}
 846
 847void __mptcp_error_report(struct sock *sk)
 848{
 849	struct mptcp_subflow_context *subflow;
 850	struct mptcp_sock *msk = mptcp_sk(sk);
 851
 852	mptcp_for_each_subflow(msk, subflow)
 853		if (__mptcp_subflow_error_report(sk, mptcp_subflow_tcp_sock(subflow)))
 854			break;
 855}
 856
 857/* In most cases we will be able to lock the mptcp socket.  If its already
 858 * owned, we need to defer to the work queue to avoid ABBA deadlock.
 859 */
 860static bool move_skbs_to_msk(struct mptcp_sock *msk, struct sock *ssk)
 861{
 862	struct sock *sk = (struct sock *)msk;
 863	bool moved;
 864
 865	moved = __mptcp_move_skbs_from_subflow(msk, ssk, true);
 866	__mptcp_ofo_queue(msk);
 867	if (unlikely(ssk->sk_err))
 868		__mptcp_subflow_error_report(sk, ssk);
 869
 870	/* If the moves have caught up with the DATA_FIN sequence number
 871	 * it's time to ack the DATA_FIN and change socket state, but
 872	 * this is not a good place to change state. Let the workqueue
 873	 * do it.
 874	 */
 875	if (mptcp_pending_data_fin(sk, NULL))
 876		mptcp_schedule_work(sk);
 877	return moved;
 878}
 879
 880void mptcp_data_ready(struct sock *sk, struct sock *ssk)
 881{
 882	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
 883	struct mptcp_sock *msk = mptcp_sk(sk);
 884
 885	/* The peer can send data while we are shutting down this
 886	 * subflow at subflow destruction time, but we must avoid enqueuing
 887	 * more data to the msk receive queue
 888	 */
 889	if (unlikely(subflow->closing))
 890		return;
 891
 892	mptcp_data_lock(sk);
 893	if (!sock_owned_by_user(sk)) {
 894		/* Wake-up the reader only for in-sequence data */
 895		if (move_skbs_to_msk(msk, ssk) && mptcp_epollin_ready(sk))
 896			sk->sk_data_ready(sk);
 897	} else {
 898		__mptcp_move_skbs_from_subflow(msk, ssk, false);
 899	}
 900	mptcp_data_unlock(sk);
 901}
 902
 903static void mptcp_subflow_joined(struct mptcp_sock *msk, struct sock *ssk)
 904{
 905	mptcp_subflow_ctx(ssk)->map_seq = READ_ONCE(msk->ack_seq);
 906	msk->allow_infinite_fallback = false;
 907	mptcp_event(MPTCP_EVENT_SUB_ESTABLISHED, msk, ssk, GFP_ATOMIC);
 908}
 909
 910static bool __mptcp_finish_join(struct mptcp_sock *msk, struct sock *ssk)
 911{
 912	struct sock *sk = (struct sock *)msk;
 913
 914	if (sk->sk_state != TCP_ESTABLISHED)
 915		return false;
 916
 917	spin_lock_bh(&msk->fallback_lock);
 918	if (!msk->allow_subflows) {
 919		spin_unlock_bh(&msk->fallback_lock);
 920		return false;
 921	}
 922	mptcp_subflow_joined(msk, ssk);
 923	spin_unlock_bh(&msk->fallback_lock);
 924
 925	mptcp_subflow_ctx(ssk)->subflow_id = msk->subflow_id++;
 926	mptcp_sockopt_sync_locked(msk, ssk);
 927	mptcp_stop_tout_timer(sk);
 928	__mptcp_propagate_sndbuf(sk, ssk);
 929	return true;
 930}
 931
 932static void __mptcp_flush_join_list(struct sock *sk, struct list_head *join_list)
 933{
 934	struct mptcp_subflow_context *tmp, *subflow;
 935	struct mptcp_sock *msk = mptcp_sk(sk);
 936
 937	list_for_each_entry_safe(subflow, tmp, join_list, node) {
 938		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
 939		bool slow = lock_sock_fast(ssk);
 940
 941		list_move_tail(&subflow->node, &msk->conn_list);
 942		if (!__mptcp_finish_join(msk, ssk))
 943			mptcp_subflow_reset(ssk);
 944		unlock_sock_fast(ssk, slow);
 945	}
 946}
 947
 948static bool mptcp_rtx_timer_pending(struct sock *sk)
 949{
 950	return timer_pending(&sk->mptcp_retransmit_timer);
 951}
 952
 953static void mptcp_reset_rtx_timer(struct sock *sk)
 954{
 955	unsigned long tout;
 956
 957	/* prevent rescheduling on close */
 958	if (unlikely(inet_sk_state_load(sk) == TCP_CLOSE))
 959		return;
 960
 961	tout = mptcp_sk(sk)->timer_ival;
 962	sk_reset_timer(sk, &sk->mptcp_retransmit_timer, jiffies + tout);
 963}
 964
 965bool mptcp_schedule_work(struct sock *sk)
 966{
 967	if (inet_sk_state_load(sk) == TCP_CLOSE)
 968		return false;
 969
 970	/* Get a reference on this socket, mptcp_worker() will release it.
 971	 * As mptcp_worker() might complete before us, we can not avoid
 972	 * a sock_hold()/sock_put() if schedule_work() returns false.
 973	 */
 974	sock_hold(sk);
 975
 976	if (schedule_work(&mptcp_sk(sk)->work))
 977		return true;
 978
 979	sock_put(sk);
 980	return false;
 981}
 982
 983static bool mptcp_skb_can_collapse_to(u64 write_seq,
 984				      const struct sk_buff *skb,
 985				      const struct mptcp_ext *mpext)
 986{
 987	if (!tcp_skb_can_collapse_to(skb))
 988		return false;
 989
 990	/* can collapse only if MPTCP level sequence is in order and this
 991	 * mapping has not been xmitted yet
 992	 */
 993	return mpext && mpext->data_seq + mpext->data_len == write_seq &&
 994	       !mpext->frozen;
 995}
 996
 997/* we can append data to the given data frag if:
 998 * - there is space available in the backing page_frag
 999 * - the data frag tail matches the current page_frag free offset
1000 * - the data frag end sequence number matches the current write seq
1001 */
1002static bool mptcp_frag_can_collapse_to(const struct mptcp_sock *msk,
1003				       const struct page_frag *pfrag,
1004				       const struct mptcp_data_frag *df)
1005{
1006	return df && pfrag->page == df->page &&
1007		pfrag->size - pfrag->offset > 0 &&
1008		pfrag->offset == (df->offset + df->data_len) &&
1009		df->data_seq + df->data_len == msk->write_seq;
1010}
1011
1012static void dfrag_uncharge(struct sock *sk, int len)
1013{
1014	sk_mem_uncharge(sk, len);
1015	sk_wmem_queued_add(sk, -len);
1016}
1017
1018static void dfrag_clear(struct sock *sk, struct mptcp_data_frag *dfrag)
1019{
1020	int len = dfrag->data_len + dfrag->overhead;
1021
1022	list_del(&dfrag->list);
1023	dfrag_uncharge(sk, len);
1024	put_page(dfrag->page);
1025}
1026
1027/* called under both the msk socket lock and the data lock */
1028static void __mptcp_clean_una(struct sock *sk)
1029{
1030	struct mptcp_sock *msk = mptcp_sk(sk);
1031	struct mptcp_data_frag *dtmp, *dfrag;
1032	u64 snd_una;
1033
1034	snd_una = msk->snd_una;
1035	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list) {
1036		if (after64(dfrag->data_seq + dfrag->data_len, snd_una))
1037			break;
1038
1039		if (unlikely(dfrag == msk->first_pending)) {
1040			/* in recovery mode can see ack after the current snd head */
1041			if (WARN_ON_ONCE(!msk->recovery))
1042				break;
1043
1044			msk->first_pending = mptcp_send_next(sk);
1045		}
1046
1047		dfrag_clear(sk, dfrag);
1048	}
1049
1050	dfrag = mptcp_rtx_head(sk);
1051	if (dfrag && after64(snd_una, dfrag->data_seq)) {
1052		u64 delta = snd_una - dfrag->data_seq;
1053
1054		/* prevent wrap around in recovery mode */
1055		if (unlikely(delta > dfrag->already_sent)) {
1056			if (WARN_ON_ONCE(!msk->recovery))
1057				goto out;
1058			if (WARN_ON_ONCE(delta > dfrag->data_len))
1059				goto out;
1060			dfrag->already_sent += delta - dfrag->already_sent;
1061		}
1062
1063		dfrag->data_seq += delta;
1064		dfrag->offset += delta;
1065		dfrag->data_len -= delta;
1066		dfrag->already_sent -= delta;
1067
1068		dfrag_uncharge(sk, delta);
1069	}
1070
1071	/* all retransmitted data acked, recovery completed */
1072	if (unlikely(msk->recovery) && after64(msk->snd_una, msk->recovery_snd_nxt))
1073		msk->recovery = false;
1074
1075out:
1076	if (snd_una == msk->snd_nxt && snd_una == msk->write_seq) {
1077		if (mptcp_rtx_timer_pending(sk) && !mptcp_data_fin_enabled(msk))
1078			mptcp_stop_rtx_timer(sk);
1079	} else {
1080		mptcp_reset_rtx_timer(sk);
1081	}
1082
1083	if (mptcp_pending_data_fin_ack(sk))
1084		mptcp_schedule_work(sk);
1085}
1086
1087static void __mptcp_clean_una_wakeup(struct sock *sk)
1088{
1089	lockdep_assert_held_once(&sk->sk_lock.slock);
1090
1091	__mptcp_clean_una(sk);
1092	mptcp_write_space(sk);
1093}
1094
1095static void mptcp_clean_una_wakeup(struct sock *sk)
1096{
1097	mptcp_data_lock(sk);
1098	__mptcp_clean_una_wakeup(sk);
1099	mptcp_data_unlock(sk);
1100}
1101
1102static void mptcp_enter_memory_pressure(struct sock *sk)
1103{
1104	struct mptcp_subflow_context *subflow;
1105	struct mptcp_sock *msk = mptcp_sk(sk);
1106	bool first = true;
1107
1108	mptcp_for_each_subflow(msk, subflow) {
1109		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1110
1111		if (first && !ssk->sk_bypass_prot_mem) {
1112			tcp_enter_memory_pressure(ssk);
1113			first = false;
1114		}
1115
1116		sk_stream_moderate_sndbuf(ssk);
1117	}
1118	__mptcp_sync_sndbuf(sk);
1119}
1120
1121/* ensure we get enough memory for the frag hdr, beyond some minimal amount of
1122 * data
1123 */
1124static bool mptcp_page_frag_refill(struct sock *sk, struct page_frag *pfrag)
1125{
1126	if (likely(skb_page_frag_refill(32U + sizeof(struct mptcp_data_frag),
1127					pfrag, sk->sk_allocation)))
1128		return true;
1129
1130	mptcp_enter_memory_pressure(sk);
1131	return false;
1132}
1133
1134static struct mptcp_data_frag *
1135mptcp_carve_data_frag(const struct mptcp_sock *msk, struct page_frag *pfrag,
1136		      int orig_offset)
1137{
1138	int offset = ALIGN(orig_offset, sizeof(long));
1139	struct mptcp_data_frag *dfrag;
1140
1141	dfrag = (struct mptcp_data_frag *)(page_to_virt(pfrag->page) + offset);
1142	dfrag->data_len = 0;
1143	dfrag->data_seq = msk->write_seq;
1144	dfrag->overhead = offset - orig_offset + sizeof(struct mptcp_data_frag);
1145	dfrag->offset = offset + sizeof(struct mptcp_data_frag);
1146	dfrag->already_sent = 0;
1147	dfrag->page = pfrag->page;
1148
1149	return dfrag;
1150}
1151
1152struct mptcp_sendmsg_info {
1153	int mss_now;
1154	int size_goal;
1155	u16 limit;
1156	u16 sent;
1157	unsigned int flags;
1158	bool data_lock_held;
1159};
1160
1161static int mptcp_check_allowed_size(const struct mptcp_sock *msk, struct sock *ssk,
1162				    u64 data_seq, int avail_size)
1163{
1164	u64 window_end = mptcp_wnd_end(msk);
1165	u64 mptcp_snd_wnd;
1166
1167	if (__mptcp_check_fallback(msk))
1168		return avail_size;
1169
1170	mptcp_snd_wnd = window_end - data_seq;
1171	avail_size = min_t(unsigned int, mptcp_snd_wnd, avail_size);
1172
1173	if (unlikely(tcp_sk(ssk)->snd_wnd < mptcp_snd_wnd)) {
1174		tcp_sk(ssk)->snd_wnd = min_t(u64, U32_MAX, mptcp_snd_wnd);
1175		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_SNDWNDSHARED);
1176	}
1177
1178	return avail_size;
1179}
1180
1181static bool __mptcp_add_ext(struct sk_buff *skb, gfp_t gfp)
1182{
1183	struct skb_ext *mpext = __skb_ext_alloc(gfp);
1184
1185	if (!mpext)
1186		return false;
1187	__skb_ext_set(skb, SKB_EXT_MPTCP, mpext);
1188	return true;
1189}
1190
1191static struct sk_buff *__mptcp_do_alloc_tx_skb(struct sock *sk, gfp_t gfp)
1192{
1193	struct sk_buff *skb;
1194
1195	skb = alloc_skb_fclone(MAX_TCP_HEADER, gfp);
1196	if (likely(skb)) {
1197		if (likely(__mptcp_add_ext(skb, gfp))) {
1198			skb_reserve(skb, MAX_TCP_HEADER);
1199			skb->ip_summed = CHECKSUM_PARTIAL;
1200			INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
1201			return skb;
1202		}
1203		__kfree_skb(skb);
1204	} else {
1205		mptcp_enter_memory_pressure(sk);
1206	}
1207	return NULL;
1208}
1209
1210static struct sk_buff *__mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, gfp_t gfp)
1211{
1212	struct sk_buff *skb;
1213
1214	skb = __mptcp_do_alloc_tx_skb(sk, gfp);
1215	if (!skb)
1216		return NULL;
1217
1218	if (likely(sk_wmem_schedule(ssk, skb->truesize))) {
1219		tcp_skb_entail(ssk, skb);
1220		return skb;
1221	}
1222	tcp_skb_tsorted_anchor_cleanup(skb);
1223	kfree_skb(skb);
1224	return NULL;
1225}
1226
1227static struct sk_buff *mptcp_alloc_tx_skb(struct sock *sk, struct sock *ssk, bool data_lock_held)
1228{
1229	gfp_t gfp = data_lock_held ? GFP_ATOMIC : sk->sk_allocation;
1230
1231	return __mptcp_alloc_tx_skb(sk, ssk, gfp);
1232}
1233
1234/* note: this always recompute the csum on the whole skb, even
1235 * if we just appended a single frag. More status info needed
1236 */
1237static void mptcp_update_data_checksum(struct sk_buff *skb, int added)
1238{
1239	struct mptcp_ext *mpext = mptcp_get_ext(skb);
1240	__wsum csum = ~csum_unfold(mpext->csum);
1241	int offset = skb->len - added;
1242
1243	mpext->csum = csum_fold(csum_block_add(csum, skb_checksum(skb, offset, added, 0), offset));
1244}
1245
1246static void mptcp_update_infinite_map(struct mptcp_sock *msk,
1247				      struct sock *ssk,
1248				      struct mptcp_ext *mpext)
1249{
1250	if (!mpext)
1251		return;
1252
1253	mpext->infinite_map = 1;
1254	mpext->data_len = 0;
1255
1256	if (!mptcp_try_fallback(ssk, MPTCP_MIB_INFINITEMAPTX)) {
1257		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED);
1258		mptcp_subflow_reset(ssk);
1259		return;
1260	}
1261
1262	mptcp_subflow_ctx(ssk)->send_infinite_map = 0;
1263}
1264
1265#define MPTCP_MAX_GSO_SIZE (GSO_LEGACY_MAX_SIZE - (MAX_TCP_HEADER + 1))
1266
1267static int mptcp_sendmsg_frag(struct sock *sk, struct sock *ssk,
1268			      struct mptcp_data_frag *dfrag,
1269			      struct mptcp_sendmsg_info *info)
1270{
1271	u64 data_seq = dfrag->data_seq + info->sent;
1272	int offset = dfrag->offset + info->sent;
1273	struct mptcp_sock *msk = mptcp_sk(sk);
1274	bool zero_window_probe = false;
1275	struct mptcp_ext *mpext = NULL;
1276	bool can_coalesce = false;
1277	bool reuse_skb = true;
1278	struct sk_buff *skb;
1279	size_t copy;
1280	int i;
1281
1282	pr_debug("msk=%p ssk=%p sending dfrag at seq=%llu len=%u already sent=%u\n",
1283		 msk, ssk, dfrag->data_seq, dfrag->data_len, info->sent);
1284
1285	if (WARN_ON_ONCE(info->sent > info->limit ||
1286			 info->limit > dfrag->data_len))
1287		return 0;
1288
1289	if (unlikely(!__tcp_can_send(ssk)))
1290		return -EAGAIN;
1291
1292	/* compute send limit */
1293	if (unlikely(ssk->sk_gso_max_size > MPTCP_MAX_GSO_SIZE))
1294		ssk->sk_gso_max_size = MPTCP_MAX_GSO_SIZE;
1295	info->mss_now = tcp_send_mss(ssk, &info->size_goal, info->flags);
1296	copy = info->size_goal;
1297
1298	skb = tcp_write_queue_tail(ssk);
1299	if (skb && copy > skb->len) {
1300		/* Limit the write to the size available in the
1301		 * current skb, if any, so that we create at most a new skb.
1302		 * Explicitly tells TCP internals to avoid collapsing on later
1303		 * queue management operation, to avoid breaking the ext <->
1304		 * SSN association set here
1305		 */
1306		mpext = mptcp_get_ext(skb);
1307		if (!mptcp_skb_can_collapse_to(data_seq, skb, mpext)) {
1308			TCP_SKB_CB(skb)->eor = 1;
1309			tcp_mark_push(tcp_sk(ssk), skb);
1310			goto alloc_skb;
1311		}
1312
1313		i = skb_shinfo(skb)->nr_frags;
1314		can_coalesce = skb_can_coalesce(skb, i, dfrag->page, offset);
1315		if (!can_coalesce && i >= READ_ONCE(net_hotdata.sysctl_max_skb_frags)) {
1316			tcp_mark_push(tcp_sk(ssk), skb);
1317			goto alloc_skb;
1318		}
1319
1320		copy -= skb->len;
1321	} else {
1322alloc_skb:
1323		skb = mptcp_alloc_tx_skb(sk, ssk, info->data_lock_held);
1324		if (!skb)
1325			return -ENOMEM;
1326
1327		i = skb_shinfo(skb)->nr_frags;
1328		reuse_skb = false;
1329		mpext = mptcp_get_ext(skb);
1330	}
1331
1332	/* Zero window and all data acked? Probe. */
1333	copy = mptcp_check_allowed_size(msk, ssk, data_seq, copy);
1334	if (copy == 0) {
1335		u64 snd_una = READ_ONCE(msk->snd_una);
1336
1337		/* No need for zero probe if there are any data pending
1338		 * either at the msk or ssk level; skb is the current write
1339		 * queue tail and can be empty at this point.
1340		 */
1341		if (snd_una != msk->snd_nxt || skb->len ||
1342		    skb != tcp_send_head(ssk)) {
1343			tcp_remove_empty_skb(ssk);
1344			return 0;
1345		}
1346
1347		zero_window_probe = true;
1348		data_seq = snd_una - 1;
1349		copy = 1;
1350	}
1351
1352	copy = min_t(size_t, copy, info->limit - info->sent);
1353	if (!sk_wmem_schedule(ssk, copy)) {
1354		tcp_remove_empty_skb(ssk);
1355		return -ENOMEM;
1356	}
1357
1358	if (can_coalesce) {
1359		skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1360	} else {
1361		get_page(dfrag->page);
1362		skb_fill_page_desc(skb, i, dfrag->page, offset, copy);
1363	}
1364
1365	skb->len += copy;
1366	skb->data_len += copy;
1367	skb->truesize += copy;
1368	sk_wmem_queued_add(ssk, copy);
1369	sk_mem_charge(ssk, copy);
1370	WRITE_ONCE(tcp_sk(ssk)->write_seq, tcp_sk(ssk)->write_seq + copy);
1371	TCP_SKB_CB(skb)->end_seq += copy;
1372	tcp_skb_pcount_set(skb, 0);
1373
1374	/* on skb reuse we just need to update the DSS len */
1375	if (reuse_skb) {
1376		TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1377		mpext->data_len += copy;
1378		goto out;
1379	}
1380
1381	memset(mpext, 0, sizeof(*mpext));
1382	mpext->data_seq = data_seq;
1383	mpext->subflow_seq = mptcp_subflow_ctx(ssk)->rel_write_seq;
1384	mpext->data_len = copy;
1385	mpext->use_map = 1;
1386	mpext->dsn64 = 1;
1387
1388	pr_debug("data_seq=%llu subflow_seq=%u data_len=%u dsn64=%d\n",
1389		 mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1390		 mpext->dsn64);
1391
1392	if (zero_window_probe) {
1393		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_WINPROBE);
1394		mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1395		mpext->frozen = 1;
1396		if (READ_ONCE(msk->csum_enabled))
1397			mptcp_update_data_checksum(skb, copy);
1398		tcp_push_pending_frames(ssk);
1399		return 0;
1400	}
1401out:
1402	if (READ_ONCE(msk->csum_enabled))
1403		mptcp_update_data_checksum(skb, copy);
1404	if (mptcp_subflow_ctx(ssk)->send_infinite_map)
1405		mptcp_update_infinite_map(msk, ssk, mpext);
1406	trace_mptcp_sendmsg_frag(mpext);
1407	mptcp_subflow_ctx(ssk)->rel_write_seq += copy;
1408	return copy;
1409}
1410
1411#define MPTCP_SEND_BURST_SIZE		((1 << 16) - \
1412					 sizeof(struct tcphdr) - \
1413					 MAX_TCP_OPTION_SPACE - \
1414					 sizeof(struct ipv6hdr) - \
1415					 sizeof(struct frag_hdr))
1416
1417struct subflow_send_info {
1418	struct sock *ssk;
1419	u64 linger_time;
1420};
1421
1422void mptcp_subflow_set_active(struct mptcp_subflow_context *subflow)
1423{
1424	if (!subflow->stale)
1425		return;
1426
1427	subflow->stale = 0;
1428	MPTCP_INC_STATS(sock_net(mptcp_subflow_tcp_sock(subflow)), MPTCP_MIB_SUBFLOWRECOVER);
1429}
1430
1431bool mptcp_subflow_active(struct mptcp_subflow_context *subflow)
1432{
1433	if (unlikely(subflow->stale)) {
1434		u32 rcv_tstamp = READ_ONCE(tcp_sk(mptcp_subflow_tcp_sock(subflow))->rcv_tstamp);
1435
1436		if (subflow->stale_rcv_tstamp == rcv_tstamp)
1437			return false;
1438
1439		mptcp_subflow_set_active(subflow);
1440	}
1441	return __mptcp_subflow_active(subflow);
1442}
1443
1444#define SSK_MODE_ACTIVE	0
1445#define SSK_MODE_BACKUP	1
1446#define SSK_MODE_MAX	2
1447
1448/* implement the mptcp packet scheduler;
1449 * returns the subflow that will transmit the next DSS
1450 * additionally updates the rtx timeout
1451 */
1452struct sock *mptcp_subflow_get_send(struct mptcp_sock *msk)
1453{
1454	struct subflow_send_info send_info[SSK_MODE_MAX];
1455	struct mptcp_subflow_context *subflow;
1456	struct sock *sk = (struct sock *)msk;
1457	u32 pace, burst, wmem;
1458	int i, nr_active = 0;
1459	struct sock *ssk;
1460	u64 linger_time;
1461	long tout = 0;
1462
1463	/* pick the subflow with the lower wmem/wspace ratio */
1464	for (i = 0; i < SSK_MODE_MAX; ++i) {
1465		send_info[i].ssk = NULL;
1466		send_info[i].linger_time = -1;
1467	}
1468
1469	mptcp_for_each_subflow(msk, subflow) {
1470		bool backup = subflow->backup || subflow->request_bkup;
1471
1472		trace_mptcp_subflow_get_send(subflow);
1473		ssk =  mptcp_subflow_tcp_sock(subflow);
1474		if (!mptcp_subflow_active(subflow))
1475			continue;
1476
1477		tout = max(tout, mptcp_timeout_from_subflow(subflow));
1478		nr_active += !backup;
1479		pace = subflow->avg_pacing_rate;
1480		if (unlikely(!pace)) {
1481			/* init pacing rate from socket */
1482			subflow->avg_pacing_rate = READ_ONCE(ssk->sk_pacing_rate);
1483			pace = subflow->avg_pacing_rate;
1484			if (!pace)
1485				continue;
1486		}
1487
1488		linger_time = div_u64((u64)READ_ONCE(ssk->sk_wmem_queued) << 32, pace);
1489		if (linger_time < send_info[backup].linger_time) {
1490			send_info[backup].ssk = ssk;
1491			send_info[backup].linger_time = linger_time;
1492		}
1493	}
1494	__mptcp_set_timeout(sk, tout);
1495
1496	/* pick the best backup if no other subflow is active */
1497	if (!nr_active)
1498		send_info[SSK_MODE_ACTIVE].ssk = send_info[SSK_MODE_BACKUP].ssk;
1499
1500	/* According to the blest algorithm, to avoid HoL blocking for the
1501	 * faster flow, we need to:
1502	 * - estimate the faster flow linger time
1503	 * - use the above to estimate the amount of byte transferred
1504	 *   by the faster flow
1505	 * - check that the amount of queued data is greater than the above,
1506	 *   otherwise do not use the picked, slower, subflow
1507	 * We select the subflow with the shorter estimated time to flush
1508	 * the queued mem, which basically ensure the above. We just need
1509	 * to check that subflow has a non empty cwin.
1510	 */
1511	ssk = send_info[SSK_MODE_ACTIVE].ssk;
1512	if (!ssk || !sk_stream_memory_free(ssk))
1513		return NULL;
1514
1515	burst = min_t(int, MPTCP_SEND_BURST_SIZE, mptcp_wnd_end(msk) - msk->snd_nxt);
1516	wmem = READ_ONCE(ssk->sk_wmem_queued);
1517	if (!burst)
1518		return ssk;
1519
1520	subflow = mptcp_subflow_ctx(ssk);
1521	subflow->avg_pacing_rate = div_u64((u64)subflow->avg_pacing_rate * wmem +
1522					   READ_ONCE(ssk->sk_pacing_rate) * burst,
1523					   burst + wmem);
1524	msk->snd_burst = burst;
1525	return ssk;
1526}
1527
1528static void mptcp_push_release(struct sock *ssk, struct mptcp_sendmsg_info *info)
1529{
1530	tcp_push(ssk, 0, info->mss_now, tcp_sk(ssk)->nonagle, info->size_goal);
1531	release_sock(ssk);
1532}
1533
1534static void mptcp_update_post_push(struct mptcp_sock *msk,
1535				   struct mptcp_data_frag *dfrag,
1536				   u32 sent)
1537{
1538	u64 snd_nxt_new = dfrag->data_seq;
1539
1540	dfrag->already_sent += sent;
1541
1542	msk->snd_burst -= sent;
1543
1544	snd_nxt_new += dfrag->already_sent;
1545
1546	/* snd_nxt_new can be smaller than snd_nxt in case mptcp
1547	 * is recovering after a failover. In that event, this re-sends
1548	 * old segments.
1549	 *
1550	 * Thus compute snd_nxt_new candidate based on
1551	 * the dfrag->data_seq that was sent and the data
1552	 * that has been handed to the subflow for transmission
1553	 * and skip update in case it was old dfrag.
1554	 */
1555	if (likely(after64(snd_nxt_new, msk->snd_nxt))) {
1556		msk->bytes_sent += snd_nxt_new - msk->snd_nxt;
1557		WRITE_ONCE(msk->snd_nxt, snd_nxt_new);
1558	}
1559}
1560
1561void mptcp_check_and_set_pending(struct sock *sk)
1562{
1563	if (mptcp_send_head(sk)) {
1564		mptcp_data_lock(sk);
1565		mptcp_sk(sk)->cb_flags |= BIT(MPTCP_PUSH_PENDING);
1566		mptcp_data_unlock(sk);
1567	}
1568}
1569
1570static int __subflow_push_pending(struct sock *sk, struct sock *ssk,
1571				  struct mptcp_sendmsg_info *info)
1572{
1573	struct mptcp_sock *msk = mptcp_sk(sk);
1574	struct mptcp_data_frag *dfrag;
1575	int len, copied = 0, err = 0;
1576
1577	while ((dfrag = mptcp_send_head(sk))) {
1578		info->sent = dfrag->already_sent;
1579		info->limit = dfrag->data_len;
1580		len = dfrag->data_len - dfrag->already_sent;
1581		while (len > 0) {
1582			int ret = 0;
1583
1584			ret = mptcp_sendmsg_frag(sk, ssk, dfrag, info);
1585			if (ret <= 0) {
1586				err = copied ? : ret;
1587				goto out;
1588			}
1589
1590			info->sent += ret;
1591			copied += ret;
1592			len -= ret;
1593
1594			mptcp_update_post_push(msk, dfrag, ret);
1595		}
1596		msk->first_pending = mptcp_send_next(sk);
1597
1598		if (msk->snd_burst <= 0 ||
1599		    !sk_stream_memory_free(ssk) ||
1600		    !mptcp_subflow_active(mptcp_subflow_ctx(ssk))) {
1601			err = copied;
1602			goto out;
1603		}
1604		mptcp_set_timeout(sk);
1605	}
1606	err = copied;
1607
1608out:
1609	if (err > 0)
1610		msk->last_data_sent = tcp_jiffies32;
1611	return err;
1612}
1613
1614void __mptcp_push_pending(struct sock *sk, unsigned int flags)
1615{
1616	struct sock *prev_ssk = NULL, *ssk = NULL;
1617	struct mptcp_sock *msk = mptcp_sk(sk);
1618	struct mptcp_sendmsg_info info = {
1619				.flags = flags,
1620	};
1621	bool copied = false;
1622	int push_count = 1;
1623
1624	while (mptcp_send_head(sk) && (push_count > 0)) {
1625		struct mptcp_subflow_context *subflow;
1626		int ret = 0;
1627
1628		if (mptcp_sched_get_send(msk))
1629			break;
1630
1631		push_count = 0;
1632
1633		mptcp_for_each_subflow(msk, subflow) {
1634			if (READ_ONCE(subflow->scheduled)) {
1635				mptcp_subflow_set_scheduled(subflow, false);
1636
1637				prev_ssk = ssk;
1638				ssk = mptcp_subflow_tcp_sock(subflow);
1639				if (ssk != prev_ssk) {
1640					/* First check. If the ssk has changed since
1641					 * the last round, release prev_ssk
1642					 */
1643					if (prev_ssk)
1644						mptcp_push_release(prev_ssk, &info);
1645
1646					/* Need to lock the new subflow only if different
1647					 * from the previous one, otherwise we are still
1648					 * helding the relevant lock
1649					 */
1650					lock_sock(ssk);
1651				}
1652
1653				push_count++;
1654
1655				ret = __subflow_push_pending(sk, ssk, &info);
1656				if (ret <= 0) {
1657					if (ret != -EAGAIN ||
1658					    (1 << ssk->sk_state) &
1659					     (TCPF_FIN_WAIT1 | TCPF_FIN_WAIT2 | TCPF_CLOSE))
1660						push_count--;
1661					continue;
1662				}
1663				copied = true;
1664			}
1665		}
1666	}
1667
1668	/* at this point we held the socket lock for the last subflow we used */
1669	if (ssk)
1670		mptcp_push_release(ssk, &info);
1671
1672	/* Avoid scheduling the rtx timer if no data has been pushed; the timer
1673	 * will be updated on positive acks by __mptcp_cleanup_una().
1674	 */
1675	if (copied) {
1676		if (!mptcp_rtx_timer_pending(sk))
1677			mptcp_reset_rtx_timer(sk);
1678		mptcp_check_send_data_fin(sk);
1679	}
1680}
1681
1682static void __mptcp_subflow_push_pending(struct sock *sk, struct sock *ssk, bool first)
1683{
1684	struct mptcp_sock *msk = mptcp_sk(sk);
1685	struct mptcp_sendmsg_info info = {
1686		.data_lock_held = true,
1687	};
1688	bool keep_pushing = true;
1689	struct sock *xmit_ssk;
1690	int copied = 0;
1691
1692	info.flags = 0;
1693	while (mptcp_send_head(sk) && keep_pushing) {
1694		struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1695		int ret = 0;
1696
1697		/* check for a different subflow usage only after
1698		 * spooling the first chunk of data
1699		 */
1700		if (first) {
1701			mptcp_subflow_set_scheduled(subflow, false);
1702			ret = __subflow_push_pending(sk, ssk, &info);
1703			first = false;
1704			if (ret <= 0)
1705				break;
1706			copied += ret;
1707			continue;
1708		}
1709
1710		if (mptcp_sched_get_send(msk))
1711			goto out;
1712
1713		if (READ_ONCE(subflow->scheduled)) {
1714			mptcp_subflow_set_scheduled(subflow, false);
1715			ret = __subflow_push_pending(sk, ssk, &info);
1716			if (ret <= 0)
1717				keep_pushing = false;
1718			copied += ret;
1719		}
1720
1721		mptcp_for_each_subflow(msk, subflow) {
1722			if (READ_ONCE(subflow->scheduled)) {
1723				xmit_ssk = mptcp_subflow_tcp_sock(subflow);
1724				if (xmit_ssk != ssk) {
1725					mptcp_subflow_delegate(subflow,
1726							       MPTCP_DELEGATE_SEND);
1727					keep_pushing = false;
1728				}
1729			}
1730		}
1731	}
1732
1733out:
1734	/* __mptcp_alloc_tx_skb could have released some wmem and we are
1735	 * not going to flush it via release_sock()
1736	 */
1737	if (copied) {
1738		tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
1739			 info.size_goal);
1740		if (!mptcp_rtx_timer_pending(sk))
1741			mptcp_reset_rtx_timer(sk);
1742
1743		if (msk->snd_data_fin_enable &&
1744		    msk->snd_nxt + 1 == msk->write_seq)
1745			mptcp_schedule_work(sk);
1746	}
1747}
1748
1749static int mptcp_disconnect(struct sock *sk, int flags);
1750
1751static int mptcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1752				  size_t len, int *copied_syn)
1753{
1754	unsigned int saved_flags = msg->msg_flags;
1755	struct mptcp_sock *msk = mptcp_sk(sk);
1756	struct sock *ssk;
1757	int ret;
1758
1759	/* on flags based fastopen the mptcp is supposed to create the
1760	 * first subflow right now. Otherwise we are in the defer_connect
1761	 * path, and the first subflow must be already present.
1762	 * Since the defer_connect flag is cleared after the first succsful
1763	 * fastopen attempt, no need to check for additional subflow status.
1764	 */
1765	if (msg->msg_flags & MSG_FASTOPEN) {
1766		ssk = __mptcp_nmpc_sk(msk);
1767		if (IS_ERR(ssk))
1768			return PTR_ERR(ssk);
1769	}
1770	if (!msk->first)
1771		return -EINVAL;
1772
1773	ssk = msk->first;
1774
1775	lock_sock(ssk);
1776	msg->msg_flags |= MSG_DONTWAIT;
1777	msk->fastopening = 1;
1778	ret = tcp_sendmsg_fastopen(ssk, msg, copied_syn, len, NULL);
1779	msk->fastopening = 0;
1780	msg->msg_flags = saved_flags;
1781	release_sock(ssk);
1782
1783	/* do the blocking bits of inet_stream_connect outside the ssk socket lock */
1784	if (ret == -EINPROGRESS && !(msg->msg_flags & MSG_DONTWAIT)) {
1785		ret = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1786					    msg->msg_namelen, msg->msg_flags, 1);
1787
1788		/* Keep the same behaviour of plain TCP: zero the copied bytes in
1789		 * case of any error, except timeout or signal
1790		 */
1791		if (ret && ret != -EINPROGRESS && ret != -ERESTARTSYS && ret != -EINTR)
1792			*copied_syn = 0;
1793	} else if (ret && ret != -EINPROGRESS) {
1794		/* The disconnect() op called by tcp_sendmsg_fastopen()/
1795		 * __inet_stream_connect() can fail, due to looking check,
1796		 * see mptcp_disconnect().
1797		 * Attempt it again outside the problematic scope.
1798		 */
1799		if (!mptcp_disconnect(sk, 0)) {
1800			sk->sk_disconnects++;
1801			sk->sk_socket->state = SS_UNCONNECTED;
1802		}
1803	}
1804	inet_clear_bit(DEFER_CONNECT, sk);
1805
1806	return ret;
1807}
1808
1809static int do_copy_data_nocache(struct sock *sk, int copy,
1810				struct iov_iter *from, char *to)
1811{
1812	if (sk->sk_route_caps & NETIF_F_NOCACHE_COPY) {
1813		if (!copy_from_iter_full_nocache(to, copy, from))
1814			return -EFAULT;
1815	} else if (!copy_from_iter_full(to, copy, from)) {
1816		return -EFAULT;
1817	}
1818	return 0;
1819}
1820
1821/* open-code sk_stream_memory_free() plus sent limit computation to
1822 * avoid indirect calls in fast-path.
1823 * Called under the msk socket lock, so we can avoid a bunch of ONCE
1824 * annotations.
1825 */
1826static u32 mptcp_send_limit(const struct sock *sk)
1827{
1828	const struct mptcp_sock *msk = mptcp_sk(sk);
1829	u32 limit, not_sent;
1830
1831	if (sk->sk_wmem_queued >= READ_ONCE(sk->sk_sndbuf))
1832		return 0;
1833
1834	limit = mptcp_notsent_lowat(sk);
1835	if (limit == UINT_MAX)
1836		return UINT_MAX;
1837
1838	not_sent = msk->write_seq - msk->snd_nxt;
1839	if (not_sent >= limit)
1840		return 0;
1841
1842	return limit - not_sent;
1843}
1844
1845static void mptcp_rps_record_subflows(const struct mptcp_sock *msk)
1846{
1847	struct mptcp_subflow_context *subflow;
1848
1849	if (!rfs_is_needed())
1850		return;
1851
1852	mptcp_for_each_subflow(msk, subflow) {
1853		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
1854
1855		sock_rps_record_flow(ssk);
1856	}
1857}
1858
1859static int mptcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t len)
1860{
1861	struct mptcp_sock *msk = mptcp_sk(sk);
1862	struct page_frag *pfrag;
1863	size_t copied = 0;
1864	int ret = 0;
1865	long timeo;
1866
1867	/* silently ignore everything else */
1868	msg->msg_flags &= MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_FASTOPEN;
1869
1870	lock_sock(sk);
1871
1872	mptcp_rps_record_subflows(msk);
1873
1874	if (unlikely(inet_test_bit(DEFER_CONNECT, sk) ||
1875		     msg->msg_flags & MSG_FASTOPEN)) {
1876		int copied_syn = 0;
1877
1878		ret = mptcp_sendmsg_fastopen(sk, msg, len, &copied_syn);
1879		copied += copied_syn;
1880		if (ret == -EINPROGRESS && copied_syn > 0)
1881			goto out;
1882		else if (ret)
1883			goto do_error;
1884	}
1885
1886	timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
1887
1888	if ((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) {
1889		ret = sk_stream_wait_connect(sk, &timeo);
1890		if (ret)
1891			goto do_error;
1892	}
1893
1894	ret = -EPIPE;
1895	if (unlikely(sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)))
1896		goto do_error;
1897
1898	pfrag = sk_page_frag(sk);
1899
1900	while (msg_data_left(msg)) {
1901		int total_ts, frag_truesize = 0;
1902		struct mptcp_data_frag *dfrag;
1903		bool dfrag_collapsed;
1904		size_t psize, offset;
1905		u32 copy_limit;
1906
1907		/* ensure fitting the notsent_lowat() constraint */
1908		copy_limit = mptcp_send_limit(sk);
1909		if (!copy_limit)
1910			goto wait_for_memory;
1911
1912		/* reuse tail pfrag, if possible, or carve a new one from the
1913		 * page allocator
1914		 */
1915		dfrag = mptcp_pending_tail(sk);
1916		dfrag_collapsed = mptcp_frag_can_collapse_to(msk, pfrag, dfrag);
1917		if (!dfrag_collapsed) {
1918			if (!mptcp_page_frag_refill(sk, pfrag))
1919				goto wait_for_memory;
1920
1921			dfrag = mptcp_carve_data_frag(msk, pfrag, pfrag->offset);
1922			frag_truesize = dfrag->overhead;
1923		}
1924
1925		/* we do not bound vs wspace, to allow a single packet.
1926		 * memory accounting will prevent execessive memory usage
1927		 * anyway
1928		 */
1929		offset = dfrag->offset + dfrag->data_len;
1930		psize = pfrag->size - offset;
1931		psize = min_t(size_t, psize, msg_data_left(msg));
1932		psize = min_t(size_t, psize, copy_limit);
1933		total_ts = psize + frag_truesize;
1934
1935		if (!sk_wmem_schedule(sk, total_ts))
1936			goto wait_for_memory;
1937
1938		ret = do_copy_data_nocache(sk, psize, &msg->msg_iter,
1939					   page_address(dfrag->page) + offset);
1940		if (ret)
1941			goto do_error;
1942
1943		/* data successfully copied into the write queue */
1944		sk_forward_alloc_add(sk, -total_ts);
1945		copied += psize;
1946		dfrag->data_len += psize;
1947		frag_truesize += psize;
1948		pfrag->offset += frag_truesize;
1949		WRITE_ONCE(msk->write_seq, msk->write_seq + psize);
1950
1951		/* charge data on mptcp pending queue to the msk socket
1952		 * Note: we charge such data both to sk and ssk
1953		 */
1954		sk_wmem_queued_add(sk, frag_truesize);
1955		if (!dfrag_collapsed) {
1956			get_page(dfrag->page);
1957			list_add_tail(&dfrag->list, &msk->rtx_queue);
1958			if (!msk->first_pending)
1959				msk->first_pending = dfrag;
1960		}
1961		pr_debug("msk=%p dfrag at seq=%llu len=%u sent=%u new=%d\n", msk,
1962			 dfrag->data_seq, dfrag->data_len, dfrag->already_sent,
1963			 !dfrag_collapsed);
1964
1965		continue;
1966
1967wait_for_memory:
1968		set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1969		__mptcp_push_pending(sk, msg->msg_flags);
1970		ret = sk_stream_wait_memory(sk, &timeo);
1971		if (ret)
1972			goto do_error;
1973	}
1974
1975	if (copied)
1976		__mptcp_push_pending(sk, msg->msg_flags);
1977
1978out:
1979	release_sock(sk);
1980	return copied;
1981
1982do_error:
1983	if (copied)
1984		goto out;
1985
1986	copied = sk_stream_error(sk, msg->msg_flags, ret);
1987	goto out;
1988}
1989
1990static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied);
1991
1992static void mptcp_eat_recv_skb(struct sock *sk, struct sk_buff *skb)
1993{
1994	/* avoid the indirect call, we know the destructor is sock_rfree */
1995	skb->destructor = NULL;
1996	skb->sk = NULL;
1997	atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
1998	sk_mem_uncharge(sk, skb->truesize);
1999	__skb_unlink(skb, &sk->sk_receive_queue);
2000	skb_attempt_defer_free(skb);
2001}
2002
2003static int __mptcp_recvmsg_mskq(struct sock *sk, struct msghdr *msg,
2004				size_t len, int flags, int copied_total,
2005				struct scm_timestamping_internal *tss,
2006				int *cmsg_flags, struct sk_buff **last)
2007{
2008	struct mptcp_sock *msk = mptcp_sk(sk);
2009	struct sk_buff *skb, *tmp;
2010	int total_data_len = 0;
2011	int copied = 0;
2012
2013	skb_queue_walk_safe(&sk->sk_receive_queue, skb, tmp) {
2014		u32 delta, offset = MPTCP_SKB_CB(skb)->offset;
2015		u32 data_len = skb->len - offset;
2016		u32 count;
2017		int err;
2018
2019		if (flags & MSG_PEEK) {
2020			/* skip already peeked skbs */
2021			if (total_data_len + data_len <= copied_total) {
2022				total_data_len += data_len;
2023				*last = skb;
2024				continue;
2025			}
2026
2027			/* skip the already peeked data in the current skb */
2028			delta = copied_total - total_data_len;
2029			offset += delta;
2030			data_len -= delta;
2031		}
2032
2033		count = min_t(size_t, len - copied, data_len);
2034		if (!(flags & MSG_TRUNC)) {
2035			err = skb_copy_datagram_msg(skb, offset, msg, count);
2036			if (unlikely(err < 0)) {
2037				if (!copied)
2038					return err;
2039				break;
2040			}
2041		}
2042
2043		if (MPTCP_SKB_CB(skb)->has_rxtstamp) {
2044			tcp_update_recv_tstamps(skb, tss);
2045			*cmsg_flags |= MPTCP_CMSG_TS;
2046		}
2047
2048		copied += count;
2049
2050		if (!(flags & MSG_PEEK)) {
2051			msk->bytes_consumed += count;
2052			if (count < data_len) {
2053				MPTCP_SKB_CB(skb)->offset += count;
2054				MPTCP_SKB_CB(skb)->map_seq += count;
2055				break;
2056			}
2057
2058			mptcp_eat_recv_skb(sk, skb);
2059		} else {
2060			*last = skb;
2061		}
2062
2063		if (copied >= len)
2064			break;
2065	}
2066
2067	mptcp_rcv_space_adjust(msk, copied);
2068	return copied;
2069}
2070
2071/* receive buffer autotuning.  See tcp_rcv_space_adjust for more information.
2072 *
2073 * Only difference: Use highest rtt estimate of the subflows in use.
2074 */
2075static void mptcp_rcv_space_adjust(struct mptcp_sock *msk, int copied)
2076{
2077	struct mptcp_subflow_context *subflow;
2078	struct sock *sk = (struct sock *)msk;
2079	u8 scaling_ratio = U8_MAX;
2080	u32 time, advmss = 1;
2081	u64 rtt_us, mstamp;
2082
2083	msk_owned_by_me(msk);
2084
2085	if (copied <= 0)
2086		return;
2087
2088	if (!msk->rcvspace_init)
2089		mptcp_rcv_space_init(msk, msk->first);
2090
2091	msk->rcvq_space.copied += copied;
2092
2093	mstamp = mptcp_stamp();
2094	time = tcp_stamp_us_delta(mstamp, READ_ONCE(msk->rcvq_space.time));
2095
2096	rtt_us = msk->rcvq_space.rtt_us;
2097	if (rtt_us && time < (rtt_us >> 3))
2098		return;
2099
2100	rtt_us = 0;
2101	mptcp_for_each_subflow(msk, subflow) {
2102		const struct tcp_sock *tp;
2103		u64 sf_rtt_us;
2104		u32 sf_advmss;
2105
2106		tp = tcp_sk(mptcp_subflow_tcp_sock(subflow));
2107
2108		sf_rtt_us = READ_ONCE(tp->rcv_rtt_est.rtt_us);
2109		sf_advmss = READ_ONCE(tp->advmss);
2110
2111		rtt_us = max(sf_rtt_us, rtt_us);
2112		advmss = max(sf_advmss, advmss);
2113		scaling_ratio = min(tp->scaling_ratio, scaling_ratio);
2114	}
2115
2116	msk->rcvq_space.rtt_us = rtt_us;
2117	msk->scaling_ratio = scaling_ratio;
2118	if (time < (rtt_us >> 3) || rtt_us == 0)
2119		return;
2120
2121	if (msk->rcvq_space.copied <= msk->rcvq_space.space)
2122		goto new_measure;
2123
2124	if (mptcp_rcvbuf_grow(sk, msk->rcvq_space.copied)) {
2125		/* Make subflows follow along.  If we do not do this, we
2126		 * get drops at subflow level if skbs can't be moved to
2127		 * the mptcp rx queue fast enough (announced rcv_win can
2128		 * exceed ssk->sk_rcvbuf).
2129		 */
2130		mptcp_for_each_subflow(msk, subflow) {
2131			struct sock *ssk;
2132			bool slow;
2133
2134			ssk = mptcp_subflow_tcp_sock(subflow);
2135			slow = lock_sock_fast(ssk);
2136			/* subflows can be added before tcp_init_transfer() */
2137			if (tcp_sk(ssk)->rcvq_space.space)
2138				tcp_rcvbuf_grow(ssk, msk->rcvq_space.copied);
2139			unlock_sock_fast(ssk, slow);
2140		}
2141	}
2142
2143new_measure:
2144	msk->rcvq_space.copied = 0;
2145	msk->rcvq_space.time = mstamp;
2146}
2147
2148static bool __mptcp_move_skbs(struct sock *sk, struct list_head *skbs, u32 *delta)
2149{
2150	struct sk_buff *skb = list_first_entry(skbs, struct sk_buff, list);
2151	struct mptcp_sock *msk = mptcp_sk(sk);
2152	bool moved = false;
2153
2154	*delta = 0;
2155	while (1) {
2156		/* If the msk recvbuf is full stop, don't drop */
2157		if (sk_rmem_alloc_get(sk) > sk->sk_rcvbuf)
2158			break;
2159
2160		prefetch(skb->next);
2161		list_del(&skb->list);
2162		*delta += skb->truesize;
2163
2164		moved |= __mptcp_move_skb(sk, skb);
2165		if (list_empty(skbs))
2166			break;
2167
2168		skb = list_first_entry(skbs, struct sk_buff, list);
2169	}
2170
2171	__mptcp_ofo_queue(msk);
2172	if (moved)
2173		mptcp_check_data_fin((struct sock *)msk);
2174	return moved;
2175}
2176
2177static bool mptcp_can_spool_backlog(struct sock *sk, struct list_head *skbs)
2178{
2179	struct mptcp_sock *msk = mptcp_sk(sk);
2180
2181	/* After CG initialization, subflows should never add skb before
2182	 * gaining the CG themself.
2183	 */
2184	DEBUG_NET_WARN_ON_ONCE(msk->backlog_unaccounted && sk->sk_socket &&
2185			       mem_cgroup_from_sk(sk));
2186
2187	/* Don't spool the backlog if the rcvbuf is full. */
2188	if (list_empty(&msk->backlog_list) ||
2189	    sk_rmem_alloc_get(sk) > sk->sk_rcvbuf)
2190		return false;
2191
2192	INIT_LIST_HEAD(skbs);
2193	list_splice_init(&msk->backlog_list, skbs);
2194	return true;
2195}
2196
2197static void mptcp_backlog_spooled(struct sock *sk, u32 moved,
2198				  struct list_head *skbs)
2199{
2200	struct mptcp_sock *msk = mptcp_sk(sk);
2201
2202	WRITE_ONCE(msk->backlog_len, msk->backlog_len - moved);
2203	list_splice(skbs, &msk->backlog_list);
2204}
2205
2206static bool mptcp_move_skbs(struct sock *sk)
2207{
2208	struct list_head skbs;
2209	bool enqueued = false;
2210	u32 moved;
2211
2212	mptcp_data_lock(sk);
2213	while (mptcp_can_spool_backlog(sk, &skbs)) {
2214		mptcp_data_unlock(sk);
2215		enqueued |= __mptcp_move_skbs(sk, &skbs, &moved);
2216
2217		mptcp_data_lock(sk);
2218		mptcp_backlog_spooled(sk, moved, &skbs);
2219	}
2220	mptcp_data_unlock(sk);
2221	return enqueued;
2222}
2223
2224static unsigned int mptcp_inq_hint(const struct sock *sk)
2225{
2226	const struct mptcp_sock *msk = mptcp_sk(sk);
2227	const struct sk_buff *skb;
2228
2229	skb = skb_peek(&sk->sk_receive_queue);
2230	if (skb) {
2231		u64 hint_val = READ_ONCE(msk->ack_seq) - MPTCP_SKB_CB(skb)->map_seq;
2232
2233		if (hint_val >= INT_MAX)
2234			return INT_MAX;
2235
2236		return (unsigned int)hint_val;
2237	}
2238
2239	if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
2240		return 1;
2241
2242	return 0;
2243}
2244
2245static int mptcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len,
2246			 int flags, int *addr_len)
2247{
2248	struct mptcp_sock *msk = mptcp_sk(sk);
2249	struct scm_timestamping_internal tss;
2250	int copied = 0, cmsg_flags = 0;
2251	int target;
2252	long timeo;
2253
2254	/* MSG_ERRQUEUE is really a no-op till we support IP_RECVERR */
2255	if (unlikely(flags & MSG_ERRQUEUE))
2256		return inet_recv_error(sk, msg, len, addr_len);
2257
2258	lock_sock(sk);
2259	if (unlikely(sk->sk_state == TCP_LISTEN)) {
2260		copied = -ENOTCONN;
2261		goto out_err;
2262	}
2263
2264	mptcp_rps_record_subflows(msk);
2265
2266	timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
2267
2268	len = min_t(size_t, len, INT_MAX);
2269	target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2270
2271	if (unlikely(msk->recvmsg_inq))
2272		cmsg_flags = MPTCP_CMSG_INQ;
2273
2274	while (copied < len) {
2275		struct sk_buff *last = NULL;
2276		int err, bytes_read;
2277
2278		bytes_read = __mptcp_recvmsg_mskq(sk, msg, len - copied, flags,
2279						  copied, &tss, &cmsg_flags,
2280						  &last);
2281		if (unlikely(bytes_read < 0)) {
2282			if (!copied)
2283				copied = bytes_read;
2284			goto out_err;
2285		}
2286
2287		copied += bytes_read;
2288
2289		if (!list_empty(&msk->backlog_list) && mptcp_move_skbs(sk))
2290			continue;
2291
2292		/* only the MPTCP socket status is relevant here. The exit
2293		 * conditions mirror closely tcp_recvmsg()
2294		 */
2295		if (copied >= target)
2296			break;
2297
2298		if (copied) {
2299			if (sk->sk_err ||
2300			    sk->sk_state == TCP_CLOSE ||
2301			    (sk->sk_shutdown & RCV_SHUTDOWN) ||
2302			    !timeo ||
2303			    signal_pending(current))
2304				break;
2305		} else {
2306			if (sk->sk_err) {
2307				copied = sock_error(sk);
2308				break;
2309			}
2310
2311			if (sk->sk_shutdown & RCV_SHUTDOWN)
2312				break;
2313
2314			if (sk->sk_state == TCP_CLOSE) {
2315				copied = -ENOTCONN;
2316				break;
2317			}
2318
2319			if (!timeo) {
2320				copied = -EAGAIN;
2321				break;
2322			}
2323
2324			if (signal_pending(current)) {
2325				copied = sock_intr_errno(timeo);
2326				break;
2327			}
2328		}
2329
2330		pr_debug("block timeout %ld\n", timeo);
2331		mptcp_cleanup_rbuf(msk, copied);
2332		err = sk_wait_data(sk, &timeo, last);
2333		if (err < 0) {
2334			err = copied ? : err;
2335			goto out_err;
2336		}
2337	}
2338
2339	mptcp_cleanup_rbuf(msk, copied);
2340
2341out_err:
2342	if (cmsg_flags && copied >= 0) {
2343		if (cmsg_flags & MPTCP_CMSG_TS)
2344			tcp_recv_timestamp(msg, sk, &tss);
2345
2346		if (cmsg_flags & MPTCP_CMSG_INQ) {
2347			unsigned int inq = mptcp_inq_hint(sk);
2348
2349			put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2350		}
2351	}
2352
2353	pr_debug("msk=%p rx queue empty=%d copied=%d\n",
2354		 msk, skb_queue_empty(&sk->sk_receive_queue), copied);
2355
2356	release_sock(sk);
2357	return copied;
2358}
2359
2360static void mptcp_retransmit_timer(struct timer_list *t)
2361{
2362	struct sock *sk = timer_container_of(sk, t, mptcp_retransmit_timer);
2363	struct mptcp_sock *msk = mptcp_sk(sk);
2364
2365	bh_lock_sock(sk);
2366	if (!sock_owned_by_user(sk)) {
2367		/* we need a process context to retransmit */
2368		if (!test_and_set_bit(MPTCP_WORK_RTX, &msk->flags))
2369			mptcp_schedule_work(sk);
2370	} else {
2371		/* delegate our work to tcp_release_cb() */
2372		__set_bit(MPTCP_RETRANSMIT, &msk->cb_flags);
2373	}
2374	bh_unlock_sock(sk);
2375	sock_put(sk);
2376}
2377
2378static void mptcp_tout_timer(struct timer_list *t)
2379{
2380	struct inet_connection_sock *icsk =
2381		timer_container_of(icsk, t, mptcp_tout_timer);
2382	struct sock *sk = &icsk->icsk_inet.sk;
2383
2384	mptcp_schedule_work(sk);
2385	sock_put(sk);
2386}
2387
2388/* Find an idle subflow.  Return NULL if there is unacked data at tcp
2389 * level.
2390 *
2391 * A backup subflow is returned only if that is the only kind available.
2392 */
2393struct sock *mptcp_subflow_get_retrans(struct mptcp_sock *msk)
2394{
2395	struct sock *backup = NULL, *pick = NULL;
2396	struct mptcp_subflow_context *subflow;
2397	int min_stale_count = INT_MAX;
2398
2399	mptcp_for_each_subflow(msk, subflow) {
2400		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2401
2402		if (!__mptcp_subflow_active(subflow))
2403			continue;
2404
2405		/* still data outstanding at TCP level? skip this */
2406		if (!tcp_rtx_and_write_queues_empty(ssk)) {
2407			mptcp_pm_subflow_chk_stale(msk, ssk);
2408			min_stale_count = min_t(int, min_stale_count, subflow->stale_count);
2409			continue;
2410		}
2411
2412		if (subflow->backup || subflow->request_bkup) {
2413			if (!backup)
2414				backup = ssk;
2415			continue;
2416		}
2417
2418		if (!pick)
2419			pick = ssk;
2420	}
2421
2422	if (pick)
2423		return pick;
2424
2425	/* use backup only if there are no progresses anywhere */
2426	return min_stale_count > 1 ? backup : NULL;
2427}
2428
2429bool __mptcp_retransmit_pending_data(struct sock *sk)
2430{
2431	struct mptcp_data_frag *cur, *rtx_head;
2432	struct mptcp_sock *msk = mptcp_sk(sk);
2433
2434	if (__mptcp_check_fallback(msk))
2435		return false;
2436
2437	/* the closing socket has some data untransmitted and/or unacked:
2438	 * some data in the mptcp rtx queue has not really xmitted yet.
2439	 * keep it simple and re-inject the whole mptcp level rtx queue
2440	 */
2441	mptcp_data_lock(sk);
2442	__mptcp_clean_una_wakeup(sk);
2443	rtx_head = mptcp_rtx_head(sk);
2444	if (!rtx_head) {
2445		mptcp_data_unlock(sk);
2446		return false;
2447	}
2448
2449	msk->recovery_snd_nxt = msk->snd_nxt;
2450	msk->recovery = true;
2451	mptcp_data_unlock(sk);
2452
2453	msk->first_pending = rtx_head;
2454	msk->snd_burst = 0;
2455
2456	/* be sure to clear the "sent status" on all re-injected fragments */
2457	list_for_each_entry(cur, &msk->rtx_queue, list) {
2458		if (!cur->already_sent)
2459			break;
2460		cur->already_sent = 0;
2461	}
2462
2463	return true;
2464}
2465
2466/* flags for __mptcp_close_ssk() */
2467#define MPTCP_CF_PUSH		BIT(1)
2468
2469/* be sure to send a reset only if the caller asked for it, also
2470 * clean completely the subflow status when the subflow reaches
2471 * TCP_CLOSE state
2472 */
2473static void __mptcp_subflow_disconnect(struct sock *ssk,
2474				       struct mptcp_subflow_context *subflow,
2475				       bool fastclosing)
2476{
2477	if (((1 << ssk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
2478	    fastclosing) {
2479		/* The MPTCP code never wait on the subflow sockets, TCP-level
2480		 * disconnect should never fail
2481		 */
2482		WARN_ON_ONCE(tcp_disconnect(ssk, 0));
2483		mptcp_subflow_ctx_reset(subflow);
2484	} else {
2485		tcp_shutdown(ssk, SEND_SHUTDOWN);
2486	}
2487}
2488
2489/* subflow sockets can be either outgoing (connect) or incoming
2490 * (accept).
2491 *
2492 * Outgoing subflows use in-kernel sockets.
2493 * Incoming subflows do not have their own 'struct socket' allocated,
2494 * so we need to use tcp_close() after detaching them from the mptcp
2495 * parent socket.
2496 */
2497static void __mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2498			      struct mptcp_subflow_context *subflow,
2499			      unsigned int flags)
2500{
2501	struct mptcp_sock *msk = mptcp_sk(sk);
2502	bool dispose_it, need_push = false;
2503	int fwd_remaining;
2504
2505	/* Do not pass RX data to the msk, even if the subflow socket is not
2506	 * going to be freed (i.e. even for the first subflow on graceful
2507	 * subflow close.
2508	 */
2509	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
2510	subflow->closing = 1;
2511
2512	/* Borrow the fwd allocated page left-over; fwd memory for the subflow
2513	 * could be negative at this point, but will be reach zero soon - when
2514	 * the data allocated using such fragment will be freed.
2515	 */
2516	if (subflow->lent_mem_frag) {
2517		fwd_remaining = PAGE_SIZE - subflow->lent_mem_frag;
2518		sk_forward_alloc_add(sk, fwd_remaining);
2519		sk_forward_alloc_add(ssk, -fwd_remaining);
2520		subflow->lent_mem_frag = 0;
2521	}
2522
2523	/* If the first subflow moved to a close state before accept, e.g. due
2524	 * to an incoming reset or listener shutdown, the subflow socket is
2525	 * already deleted by inet_child_forget() and the mptcp socket can't
2526	 * survive too.
2527	 */
2528	if (msk->in_accept_queue && msk->first == ssk &&
2529	    (sock_flag(sk, SOCK_DEAD) || sock_flag(ssk, SOCK_DEAD))) {
2530		/* ensure later check in mptcp_worker() will dispose the msk */
2531		sock_set_flag(sk, SOCK_DEAD);
2532		mptcp_set_close_tout(sk, tcp_jiffies32 - (mptcp_close_timeout(sk) + 1));
2533		mptcp_subflow_drop_ctx(ssk);
2534		goto out_release;
2535	}
2536
2537	dispose_it = msk->free_first || ssk != msk->first;
2538	if (dispose_it)
2539		list_del(&subflow->node);
2540
2541	if (subflow->send_fastclose && ssk->sk_state != TCP_CLOSE)
2542		tcp_set_state(ssk, TCP_CLOSE);
2543
2544	need_push = (flags & MPTCP_CF_PUSH) && __mptcp_retransmit_pending_data(sk);
2545	if (!dispose_it) {
2546		__mptcp_subflow_disconnect(ssk, subflow, msk->fastclosing);
2547		release_sock(ssk);
2548
2549		goto out;
2550	}
2551
2552	subflow->disposable = 1;
2553
2554	/* if ssk hit tcp_done(), tcp_cleanup_ulp() cleared the related ops
2555	 * the ssk has been already destroyed, we just need to release the
2556	 * reference owned by msk;
2557	 */
2558	if (!inet_csk(ssk)->icsk_ulp_ops) {
2559		WARN_ON_ONCE(!sock_flag(ssk, SOCK_DEAD));
2560		kfree_rcu(subflow, rcu);
2561	} else {
2562		/* otherwise tcp will dispose of the ssk and subflow ctx */
2563		__tcp_close(ssk, 0);
2564
2565		/* close acquired an extra ref */
2566		__sock_put(ssk);
2567	}
2568
2569out_release:
2570	__mptcp_subflow_error_report(sk, ssk);
2571	release_sock(ssk);
2572
2573	sock_put(ssk);
2574
2575	if (ssk == msk->first)
2576		WRITE_ONCE(msk->first, NULL);
2577
2578out:
2579	__mptcp_sync_sndbuf(sk);
2580	if (need_push)
2581		__mptcp_push_pending(sk, 0);
2582
2583	/* Catch every 'all subflows closed' scenario, including peers silently
2584	 * closing them, e.g. due to timeout.
2585	 * For established sockets, allow an additional timeout before closing,
2586	 * as the protocol can still create more subflows.
2587	 */
2588	if (list_is_singular(&msk->conn_list) && msk->first &&
2589	    inet_sk_state_load(msk->first) == TCP_CLOSE) {
2590		if (sk->sk_state != TCP_ESTABLISHED ||
2591		    msk->in_accept_queue || sock_flag(sk, SOCK_DEAD)) {
2592			mptcp_set_state(sk, TCP_CLOSE);
2593			mptcp_close_wake_up(sk);
2594		} else {
2595			mptcp_start_tout_timer(sk);
2596		}
2597	}
2598}
2599
2600void mptcp_close_ssk(struct sock *sk, struct sock *ssk,
2601		     struct mptcp_subflow_context *subflow)
2602{
2603	struct mptcp_sock *msk = mptcp_sk(sk);
2604	struct sk_buff *skb;
2605
2606	/* The first subflow can already be closed or disconnected */
2607	if (subflow->close_event_done || READ_ONCE(subflow->local_id) < 0)
2608		return;
2609
2610	subflow->close_event_done = true;
2611
2612	if (sk->sk_state == TCP_ESTABLISHED)
2613		mptcp_event(MPTCP_EVENT_SUB_CLOSED, mptcp_sk(sk), ssk, GFP_KERNEL);
2614
2615	/* Remove any reference from the backlog to this ssk; backlog skbs consume
2616	 * space in the msk receive queue, no need to touch sk->sk_rmem_alloc
2617	 */
2618	list_for_each_entry(skb, &msk->backlog_list, list) {
2619		if (skb->sk != ssk)
2620			continue;
2621
2622		atomic_sub(skb->truesize, &skb->sk->sk_rmem_alloc);
2623		skb->sk = NULL;
2624	}
2625
2626	/* subflow aborted before reaching the fully_established status
2627	 * attempt the creation of the next subflow
2628	 */
2629	mptcp_pm_subflow_check_next(mptcp_sk(sk), subflow);
2630
2631	__mptcp_close_ssk(sk, ssk, subflow, MPTCP_CF_PUSH);
2632}
2633
2634static unsigned int mptcp_sync_mss(struct sock *sk, u32 pmtu)
2635{
2636	return 0;
2637}
2638
2639static void __mptcp_close_subflow(struct sock *sk)
2640{
2641	struct mptcp_subflow_context *subflow, *tmp;
2642	struct mptcp_sock *msk = mptcp_sk(sk);
2643
2644	might_sleep();
2645
2646	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2647		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2648		int ssk_state = inet_sk_state_load(ssk);
2649
2650		if (ssk_state != TCP_CLOSE &&
2651		    (ssk_state != TCP_CLOSE_WAIT ||
2652		     inet_sk_state_load(sk) != TCP_ESTABLISHED ||
2653		     __mptcp_check_fallback(msk)))
2654			continue;
2655
2656		/* 'subflow_data_ready' will re-sched once rx queue is empty */
2657		if (!skb_queue_empty_lockless(&ssk->sk_receive_queue))
2658			continue;
2659
2660		mptcp_close_ssk(sk, ssk, subflow);
2661	}
2662
2663}
2664
2665static bool mptcp_close_tout_expired(const struct sock *sk)
2666{
2667	if (!inet_csk(sk)->icsk_mtup.probe_timestamp ||
2668	    sk->sk_state == TCP_CLOSE)
2669		return false;
2670
2671	return time_after32(tcp_jiffies32,
2672		  inet_csk(sk)->icsk_mtup.probe_timestamp + mptcp_close_timeout(sk));
2673}
2674
2675static void mptcp_check_fastclose(struct mptcp_sock *msk)
2676{
2677	struct mptcp_subflow_context *subflow, *tmp;
2678	struct sock *sk = (struct sock *)msk;
2679
2680	if (likely(!READ_ONCE(msk->rcv_fastclose)))
2681		return;
2682
2683	mptcp_token_destroy(msk);
2684
2685	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2686		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
2687		bool slow;
2688
2689		slow = lock_sock_fast(tcp_sk);
2690		if (tcp_sk->sk_state != TCP_CLOSE) {
2691			mptcp_send_active_reset_reason(tcp_sk);
2692			tcp_set_state(tcp_sk, TCP_CLOSE);
2693		}
2694		unlock_sock_fast(tcp_sk, slow);
2695	}
2696
2697	/* Mirror the tcp_reset() error propagation */
2698	switch (sk->sk_state) {
2699	case TCP_SYN_SENT:
2700		WRITE_ONCE(sk->sk_err, ECONNREFUSED);
2701		break;
2702	case TCP_CLOSE_WAIT:
2703		WRITE_ONCE(sk->sk_err, EPIPE);
2704		break;
2705	case TCP_CLOSE:
2706		return;
2707	default:
2708		WRITE_ONCE(sk->sk_err, ECONNRESET);
2709	}
2710
2711	mptcp_set_state(sk, TCP_CLOSE);
2712	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
2713	smp_mb__before_atomic(); /* SHUTDOWN must be visible first */
2714	set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags);
2715
2716	/* the calling mptcp_worker will properly destroy the socket */
2717	if (sock_flag(sk, SOCK_DEAD))
2718		return;
2719
2720	sk->sk_state_change(sk);
2721	sk_error_report(sk);
2722}
2723
2724static void __mptcp_retrans(struct sock *sk)
2725{
2726	struct mptcp_sendmsg_info info = { .data_lock_held = true, };
2727	struct mptcp_sock *msk = mptcp_sk(sk);
2728	struct mptcp_subflow_context *subflow;
2729	struct mptcp_data_frag *dfrag;
2730	struct sock *ssk;
2731	int ret, err;
2732	u16 len = 0;
2733
2734	mptcp_clean_una_wakeup(sk);
2735
2736	/* first check ssk: need to kick "stale" logic */
2737	err = mptcp_sched_get_retrans(msk);
2738	dfrag = mptcp_rtx_head(sk);
2739	if (!dfrag) {
2740		if (mptcp_data_fin_enabled(msk)) {
2741			struct inet_connection_sock *icsk = inet_csk(sk);
2742
2743			WRITE_ONCE(icsk->icsk_retransmits,
2744				   icsk->icsk_retransmits + 1);
2745			mptcp_set_datafin_timeout(sk);
2746			mptcp_send_ack(msk);
2747
2748			goto reset_timer;
2749		}
2750
2751		if (!mptcp_send_head(sk))
2752			goto clear_scheduled;
2753
2754		goto reset_timer;
2755	}
2756
2757	if (err)
2758		goto reset_timer;
2759
2760	mptcp_for_each_subflow(msk, subflow) {
2761		if (READ_ONCE(subflow->scheduled)) {
2762			u16 copied = 0;
2763
2764			mptcp_subflow_set_scheduled(subflow, false);
2765
2766			ssk = mptcp_subflow_tcp_sock(subflow);
2767
2768			lock_sock(ssk);
2769
2770			/* limit retransmission to the bytes already sent on some subflows */
2771			info.sent = 0;
2772			info.limit = READ_ONCE(msk->csum_enabled) ? dfrag->data_len :
2773								    dfrag->already_sent;
2774
2775			/*
2776			 * make the whole retrans decision, xmit, disallow
2777			 * fallback atomic, note that we can't retrans even
2778			 * when an infinite fallback is in progress, i.e. new
2779			 * subflows are disallowed.
2780			 */
2781			spin_lock_bh(&msk->fallback_lock);
2782			if (__mptcp_check_fallback(msk) ||
2783			    !msk->allow_subflows) {
2784				spin_unlock_bh(&msk->fallback_lock);
2785				release_sock(ssk);
2786				goto clear_scheduled;
2787			}
2788
2789			while (info.sent < info.limit) {
2790				ret = mptcp_sendmsg_frag(sk, ssk, dfrag, &info);
2791				if (ret <= 0)
2792					break;
2793
2794				MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_RETRANSSEGS);
2795				copied += ret;
2796				info.sent += ret;
2797			}
2798			if (copied) {
2799				len = max(copied, len);
2800				tcp_push(ssk, 0, info.mss_now, tcp_sk(ssk)->nonagle,
2801					 info.size_goal);
2802				msk->allow_infinite_fallback = false;
2803			}
2804			spin_unlock_bh(&msk->fallback_lock);
2805
2806			release_sock(ssk);
2807		}
2808	}
2809
2810	msk->bytes_retrans += len;
2811	dfrag->already_sent = max(dfrag->already_sent, len);
2812
2813reset_timer:
2814	mptcp_check_and_set_pending(sk);
2815
2816	if (!mptcp_rtx_timer_pending(sk))
2817		mptcp_reset_rtx_timer(sk);
2818
2819clear_scheduled:
2820	/* If no rtx data was available or in case of fallback, there
2821	 * could be left-over scheduled subflows; clear them all
2822	 * or later xmit could use bad ones
2823	 */
2824	mptcp_for_each_subflow(msk, subflow)
2825		if (READ_ONCE(subflow->scheduled))
2826			mptcp_subflow_set_scheduled(subflow, false);
2827}
2828
2829/* schedule the timeout timer for the relevant event: either close timeout
2830 * or mp_fail timeout. The close timeout takes precedence on the mp_fail one
2831 */
2832void mptcp_reset_tout_timer(struct mptcp_sock *msk, unsigned long fail_tout)
2833{
2834	struct sock *sk = (struct sock *)msk;
2835	unsigned long timeout, close_timeout;
2836
2837	if (!fail_tout && !inet_csk(sk)->icsk_mtup.probe_timestamp)
2838		return;
2839
2840	close_timeout = (unsigned long)inet_csk(sk)->icsk_mtup.probe_timestamp -
2841			tcp_jiffies32 + jiffies + mptcp_close_timeout(sk);
2842
2843	/* the close timeout takes precedence on the fail one, and here at least one of
2844	 * them is active
2845	 */
2846	timeout = inet_csk(sk)->icsk_mtup.probe_timestamp ? close_timeout : fail_tout;
2847
2848	sk_reset_timer(sk, &inet_csk(sk)->mptcp_tout_timer, timeout);
2849}
2850
2851static void mptcp_mp_fail_no_response(struct mptcp_sock *msk)
2852{
2853	struct sock *ssk = msk->first;
2854	bool slow;
2855
2856	if (!ssk)
2857		return;
2858
2859	pr_debug("MP_FAIL doesn't respond, reset the subflow\n");
2860
2861	slow = lock_sock_fast(ssk);
2862	mptcp_subflow_reset(ssk);
2863	WRITE_ONCE(mptcp_subflow_ctx(ssk)->fail_tout, 0);
2864	unlock_sock_fast(ssk, slow);
2865}
2866
2867static void mptcp_backlog_purge(struct sock *sk)
2868{
2869	struct mptcp_sock *msk = mptcp_sk(sk);
2870	struct sk_buff *tmp, *skb;
2871	LIST_HEAD(backlog);
2872
2873	mptcp_data_lock(sk);
2874	list_splice_init(&msk->backlog_list, &backlog);
2875	msk->backlog_len = 0;
2876	mptcp_data_unlock(sk);
2877
2878	list_for_each_entry_safe(skb, tmp, &backlog, list) {
2879		mptcp_borrow_fwdmem(sk, skb);
2880		kfree_skb_reason(skb, SKB_DROP_REASON_SOCKET_CLOSE);
2881	}
2882	sk_mem_reclaim(sk);
2883}
2884
2885static void mptcp_do_fastclose(struct sock *sk)
2886{
2887	struct mptcp_subflow_context *subflow, *tmp;
2888	struct mptcp_sock *msk = mptcp_sk(sk);
2889
2890	mptcp_set_state(sk, TCP_CLOSE);
2891	mptcp_backlog_purge(sk);
2892	msk->fastclosing = 1;
2893
2894	/* Explicitly send the fastclose reset as need */
2895	if (__mptcp_check_fallback(msk))
2896		return;
2897
2898	mptcp_for_each_subflow_safe(msk, subflow, tmp) {
2899		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
2900
2901		lock_sock(ssk);
2902
2903		/* Some subflow socket states don't allow/need a reset.*/
2904		if ((1 << ssk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE))
2905			goto unlock;
2906
2907		subflow->send_fastclose = 1;
2908
2909		/* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
2910		 * issue in __tcp_select_window(), see tcp_disconnect().
2911		 */
2912		inet_csk(ssk)->icsk_ack.rcv_mss = TCP_MIN_MSS;
2913
2914		tcp_send_active_reset(ssk, ssk->sk_allocation,
2915				      SK_RST_REASON_TCP_ABORT_ON_CLOSE);
2916unlock:
2917		release_sock(ssk);
2918	}
2919}
2920
2921static void mptcp_worker(struct work_struct *work)
2922{
2923	struct mptcp_sock *msk = container_of(work, struct mptcp_sock, work);
2924	struct sock *sk = (struct sock *)msk;
2925	unsigned long fail_tout;
2926	int state;
2927
2928	lock_sock(sk);
2929	state = sk->sk_state;
2930	if (unlikely((1 << state) & (TCPF_CLOSE | TCPF_LISTEN)))
2931		goto unlock;
2932
2933	mptcp_check_fastclose(msk);
2934
2935	mptcp_pm_worker(msk);
2936
2937	mptcp_check_send_data_fin(sk);
2938	mptcp_check_data_fin_ack(sk);
2939	mptcp_check_data_fin(sk);
2940
2941	if (test_and_clear_bit(MPTCP_WORK_CLOSE_SUBFLOW, &msk->flags))
2942		__mptcp_close_subflow(sk);
2943
2944	if (mptcp_close_tout_expired(sk)) {
2945		struct mptcp_subflow_context *subflow, *tmp;
2946
2947		mptcp_do_fastclose(sk);
2948		mptcp_for_each_subflow_safe(msk, subflow, tmp)
2949			__mptcp_close_ssk(sk, subflow->tcp_sock, subflow, 0);
2950		mptcp_close_wake_up(sk);
2951	}
2952
2953	if (sock_flag(sk, SOCK_DEAD) && sk->sk_state == TCP_CLOSE) {
2954		__mptcp_destroy_sock(sk);
2955		goto unlock;
2956	}
2957
2958	if (test_and_clear_bit(MPTCP_WORK_RTX, &msk->flags))
2959		__mptcp_retrans(sk);
2960
2961	fail_tout = msk->first ? READ_ONCE(mptcp_subflow_ctx(msk->first)->fail_tout) : 0;
2962	if (fail_tout && time_after(jiffies, fail_tout))
2963		mptcp_mp_fail_no_response(msk);
2964
2965unlock:
2966	release_sock(sk);
2967	sock_put(sk);
2968}
2969
2970static void __mptcp_init_sock(struct sock *sk)
2971{
2972	struct mptcp_sock *msk = mptcp_sk(sk);
2973
2974	INIT_LIST_HEAD(&msk->conn_list);
2975	INIT_LIST_HEAD(&msk->join_list);
2976	INIT_LIST_HEAD(&msk->rtx_queue);
2977	INIT_LIST_HEAD(&msk->backlog_list);
2978	INIT_WORK(&msk->work, mptcp_worker);
2979	msk->out_of_order_queue = RB_ROOT;
2980	msk->first_pending = NULL;
2981	msk->timer_ival = TCP_RTO_MIN;
2982	msk->scaling_ratio = TCP_DEFAULT_SCALING_RATIO;
2983	msk->backlog_len = 0;
2984
2985	WRITE_ONCE(msk->first, NULL);
2986	inet_csk(sk)->icsk_sync_mss = mptcp_sync_mss;
2987	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
2988	msk->allow_infinite_fallback = true;
2989	msk->allow_subflows = true;
2990	msk->recovery = false;
2991	msk->subflow_id = 1;
2992	msk->last_data_sent = tcp_jiffies32;
2993	msk->last_data_recv = tcp_jiffies32;
2994	msk->last_ack_recv = tcp_jiffies32;
2995
2996	mptcp_pm_data_init(msk);
2997	spin_lock_init(&msk->fallback_lock);
2998
2999	/* re-use the csk retrans timer for MPTCP-level retrans */
3000	timer_setup(&sk->mptcp_retransmit_timer, mptcp_retransmit_timer, 0);
3001	timer_setup(&msk->sk.mptcp_tout_timer, mptcp_tout_timer, 0);
3002}
3003
3004static void mptcp_ca_reset(struct sock *sk)
3005{
3006	struct inet_connection_sock *icsk = inet_csk(sk);
3007
3008	tcp_assign_congestion_control(sk);
3009	strscpy(mptcp_sk(sk)->ca_name, icsk->icsk_ca_ops->name,
3010		sizeof(mptcp_sk(sk)->ca_name));
3011
3012	/* no need to keep a reference to the ops, the name will suffice */
3013	tcp_cleanup_congestion_control(sk);
3014	icsk->icsk_ca_ops = NULL;
3015}
3016
3017static int mptcp_init_sock(struct sock *sk)
3018{
3019	struct net *net = sock_net(sk);
3020	int ret;
3021
3022	__mptcp_init_sock(sk);
3023
3024	if (!mptcp_is_enabled(net))
3025		return -ENOPROTOOPT;
3026
3027	if (unlikely(!net->mib.mptcp_statistics) && !mptcp_mib_alloc(net))
3028		return -ENOMEM;
3029
3030	rcu_read_lock();
3031	ret = mptcp_init_sched(mptcp_sk(sk),
3032			       mptcp_sched_find(mptcp_get_scheduler(net)));
3033	rcu_read_unlock();
3034	if (ret)
3035		return ret;
3036
3037	set_bit(SOCK_CUSTOM_SOCKOPT, &sk->sk_socket->flags);
3038
3039	/* fetch the ca name; do it outside __mptcp_init_sock(), so that clone will
3040	 * propagate the correct value
3041	 */
3042	mptcp_ca_reset(sk);
3043
3044	sk_sockets_allocated_inc(sk);
3045	sk->sk_rcvbuf = READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]);
3046	sk->sk_sndbuf = READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]);
3047
3048	return 0;
3049}
3050
3051static void __mptcp_clear_xmit(struct sock *sk)
3052{
3053	struct mptcp_sock *msk = mptcp_sk(sk);
3054	struct mptcp_data_frag *dtmp, *dfrag;
3055
3056	msk->first_pending = NULL;
3057	list_for_each_entry_safe(dfrag, dtmp, &msk->rtx_queue, list)
3058		dfrag_clear(sk, dfrag);
3059}
3060
3061void mptcp_cancel_work(struct sock *sk)
3062{
3063	struct mptcp_sock *msk = mptcp_sk(sk);
3064
3065	if (cancel_work_sync(&msk->work))
3066		__sock_put(sk);
3067}
3068
3069void mptcp_subflow_shutdown(struct sock *sk, struct sock *ssk, int how)
3070{
3071	lock_sock(ssk);
3072
3073	switch (ssk->sk_state) {
3074	case TCP_LISTEN:
3075		if (!(how & RCV_SHUTDOWN))
3076			break;
3077		fallthrough;
3078	case TCP_SYN_SENT:
3079		WARN_ON_ONCE(tcp_disconnect(ssk, O_NONBLOCK));
3080		break;
3081	default:
3082		if (__mptcp_check_fallback(mptcp_sk(sk))) {
3083			pr_debug("Fallback\n");
3084			ssk->sk_shutdown |= how;
3085			tcp_shutdown(ssk, how);
3086
3087			/* simulate the data_fin ack reception to let the state
3088			 * machine move forward
3089			 */
3090			WRITE_ONCE(mptcp_sk(sk)->snd_una, mptcp_sk(sk)->snd_nxt);
3091			mptcp_schedule_work(sk);
3092		} else {
3093			pr_debug("Sending DATA_FIN on subflow %p\n", ssk);
3094			tcp_send_ack(ssk);
3095			if (!mptcp_rtx_timer_pending(sk))
3096				mptcp_reset_rtx_timer(sk);
3097		}
3098		break;
3099	}
3100
3101	release_sock(ssk);
3102}
3103
3104void mptcp_set_state(struct sock *sk, int state)
3105{
3106	int oldstate = sk->sk_state;
3107
3108	switch (state) {
3109	case TCP_ESTABLISHED:
3110		if (oldstate != TCP_ESTABLISHED)
3111			MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
3112		break;
3113	case TCP_CLOSE_WAIT:
3114		/* Unlike TCP, MPTCP sk would not have the TCP_SYN_RECV state:
3115		 * MPTCP "accepted" sockets will be created later on. So no
3116		 * transition from TCP_SYN_RECV to TCP_CLOSE_WAIT.
3117		 */
3118		break;
3119	default:
3120		if (oldstate == TCP_ESTABLISHED || oldstate == TCP_CLOSE_WAIT)
3121			MPTCP_DEC_STATS(sock_net(sk), MPTCP_MIB_CURRESTAB);
3122	}
3123
3124	inet_sk_state_store(sk, state);
3125}
3126
3127static const unsigned char new_state[16] = {
3128	/* current state:     new state:      action:	*/
3129	[0 /* (Invalid) */] = TCP_CLOSE,
3130	[TCP_ESTABLISHED]   = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3131	[TCP_SYN_SENT]      = TCP_CLOSE,
3132	[TCP_SYN_RECV]      = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
3133	[TCP_FIN_WAIT1]     = TCP_FIN_WAIT1,
3134	[TCP_FIN_WAIT2]     = TCP_FIN_WAIT2,
3135	[TCP_TIME_WAIT]     = TCP_CLOSE,	/* should not happen ! */
3136	[TCP_CLOSE]         = TCP_CLOSE,
3137	[TCP_CLOSE_WAIT]    = TCP_LAST_ACK  | TCP_ACTION_FIN,
3138	[TCP_LAST_ACK]      = TCP_LAST_ACK,
3139	[TCP_LISTEN]        = TCP_CLOSE,
3140	[TCP_CLOSING]       = TCP_CLOSING,
3141	[TCP_NEW_SYN_RECV]  = TCP_CLOSE,	/* should not happen ! */
3142};
3143
3144static int mptcp_close_state(struct sock *sk)
3145{
3146	int next = (int)new_state[sk->sk_state];
3147	int ns = next & TCP_STATE_MASK;
3148
3149	mptcp_set_state(sk, ns);
3150
3151	return next & TCP_ACTION_FIN;
3152}
3153
3154static void mptcp_check_send_data_fin(struct sock *sk)
3155{
3156	struct mptcp_subflow_context *subflow;
3157	struct mptcp_sock *msk = mptcp_sk(sk);
3158
3159	pr_debug("msk=%p snd_data_fin_enable=%d pending=%d snd_nxt=%llu write_seq=%llu\n",
3160		 msk, msk->snd_data_fin_enable, !!mptcp_send_head(sk),
3161		 msk->snd_nxt, msk->write_seq);
3162
3163	/* we still need to enqueue subflows or not really shutting down,
3164	 * skip this
3165	 */
3166	if (!msk->snd_data_fin_enable || msk->snd_nxt + 1 != msk->write_seq ||
3167	    mptcp_send_head(sk))
3168		return;
3169
3170	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3171
3172	mptcp_for_each_subflow(msk, subflow) {
3173		struct sock *tcp_sk = mptcp_subflow_tcp_sock(subflow);
3174
3175		mptcp_subflow_shutdown(sk, tcp_sk, SEND_SHUTDOWN);
3176	}
3177}
3178
3179static void __mptcp_wr_shutdown(struct sock *sk)
3180{
3181	struct mptcp_sock *msk = mptcp_sk(sk);
3182
3183	pr_debug("msk=%p snd_data_fin_enable=%d shutdown=%x state=%d pending=%d\n",
3184		 msk, msk->snd_data_fin_enable, sk->sk_shutdown, sk->sk_state,
3185		 !!mptcp_send_head(sk));
3186
3187	/* will be ignored by fallback sockets */
3188	WRITE_ONCE(msk->write_seq, msk->write_seq + 1);
3189	WRITE_ONCE(msk->snd_data_fin_enable, 1);
3190
3191	mptcp_check_send_data_fin(sk);
3192}
3193
3194static void __mptcp_destroy_sock(struct sock *sk)
3195{
3196	struct mptcp_sock *msk = mptcp_sk(sk);
3197
3198	pr_debug("msk=%p\n", msk);
3199
3200	might_sleep();
3201
3202	mptcp_stop_rtx_timer(sk);
3203	sk_stop_timer(sk, &inet_csk(sk)->mptcp_tout_timer);
3204	msk->pm.status = 0;
3205	mptcp_release_sched(msk);
3206
3207	sk->sk_prot->destroy(sk);
3208
3209	sk_stream_kill_queues(sk);
3210	xfrm_sk_free_policy(sk);
3211
3212	sock_put(sk);
3213}
3214
3215void __mptcp_unaccepted_force_close(struct sock *sk)
3216{
3217	sock_set_flag(sk, SOCK_DEAD);
3218	mptcp_do_fastclose(sk);
3219	__mptcp_destroy_sock(sk);
3220}
3221
3222static __poll_t mptcp_check_readable(struct sock *sk)
3223{
3224	return mptcp_epollin_ready(sk) ? EPOLLIN | EPOLLRDNORM : 0;
3225}
3226
3227static void mptcp_check_listen_stop(struct sock *sk)
3228{
3229	struct sock *ssk;
3230
3231	if (inet_sk_state_load(sk) != TCP_LISTEN)
3232		return;
3233
3234	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
3235	ssk = mptcp_sk(sk)->first;
3236	if (WARN_ON_ONCE(!ssk || inet_sk_state_load(ssk) != TCP_LISTEN))
3237		return;
3238
3239	lock_sock_nested(ssk, SINGLE_DEPTH_NESTING);
3240	tcp_set_state(ssk, TCP_CLOSE);
3241	mptcp_subflow_queue_clean(sk, ssk);
3242	inet_csk_listen_stop(ssk);
3243	mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CLOSED);
3244	release_sock(ssk);
3245}
3246
3247bool __mptcp_close(struct sock *sk, long timeout)
3248{
3249	struct mptcp_subflow_context *subflow;
3250	struct mptcp_sock *msk = mptcp_sk(sk);
3251	bool do_cancel_work = false;
3252	int subflows_alive = 0;
3253
3254	WRITE_ONCE(sk->sk_shutdown, SHUTDOWN_MASK);
3255
3256	if ((1 << sk->sk_state) & (TCPF_LISTEN | TCPF_CLOSE)) {
3257		mptcp_check_listen_stop(sk);
3258		mptcp_set_state(sk, TCP_CLOSE);
3259		goto cleanup;
3260	}
3261
3262	if (mptcp_data_avail(msk) || timeout < 0) {
3263		/* If the msk has read data, or the caller explicitly ask it,
3264		 * do the MPTCP equivalent of TCP reset, aka MPTCP fastclose
3265		 */
3266		mptcp_do_fastclose(sk);
3267		timeout = 0;
3268	} else if (mptcp_close_state(sk)) {
3269		__mptcp_wr_shutdown(sk);
3270	}
3271
3272	sk_stream_wait_close(sk, timeout);
3273
3274cleanup:
3275	/* orphan all the subflows */
3276	mptcp_for_each_subflow(msk, subflow) {
3277		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
3278		bool slow = lock_sock_fast_nested(ssk);
3279
3280		subflows_alive += ssk->sk_state != TCP_CLOSE;
3281
3282		/* since the close timeout takes precedence on the fail one,
3283		 * cancel the latter
3284		 */
3285		if (ssk == msk->first)
3286			subflow->fail_tout = 0;
3287
3288		/* detach from the parent socket, but allow data_ready to
3289		 * push incoming data into the mptcp stack, to properly ack it
3290		 */
3291		ssk->sk_socket = NULL;
3292		ssk->sk_wq = NULL;
3293		unlock_sock_fast(ssk, slow);
3294	}
3295	sock_orphan(sk);
3296
3297	/* all the subflows are closed, only timeout can change the msk
3298	 * state, let's not keep resources busy for no reasons
3299	 */
3300	if (subflows_alive == 0)
3301		mptcp_set_state(sk, TCP_CLOSE);
3302
3303	sock_hold(sk);
3304	pr_debug("msk=%p state=%d\n", sk, sk->sk_state);
3305	mptcp_pm_connection_closed(msk);
3306
3307	if (sk->sk_state == TCP_CLOSE) {
3308		__mptcp_destroy_sock(sk);
3309		do_cancel_work = true;
3310	} else {
3311		mptcp_start_tout_timer(sk);
3312	}
3313
3314	return do_cancel_work;
3315}
3316
3317static void mptcp_close(struct sock *sk, long timeout)
3318{
3319	bool do_cancel_work;
3320
3321	lock_sock(sk);
3322
3323	do_cancel_work = __mptcp_close(sk, timeout);
3324	release_sock(sk);
3325	if (do_cancel_work)
3326		mptcp_cancel_work(sk);
3327
3328	sock_put(sk);
3329}
3330
3331static void mptcp_copy_inaddrs(struct sock *msk, const struct sock *ssk)
3332{
3333#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3334	const struct ipv6_pinfo *ssk6 = inet6_sk(ssk);
3335	struct ipv6_pinfo *msk6 = inet6_sk(msk);
3336
3337	msk->sk_v6_daddr = ssk->sk_v6_daddr;
3338	msk->sk_v6_rcv_saddr = ssk->sk_v6_rcv_saddr;
3339
3340	if (msk6 && ssk6) {
3341		msk6->saddr = ssk6->saddr;
3342		msk6->flow_label = ssk6->flow_label;
3343	}
3344#endif
3345
3346	inet_sk(msk)->inet_num = inet_sk(ssk)->inet_num;
3347	inet_sk(msk)->inet_dport = inet_sk(ssk)->inet_dport;
3348	inet_sk(msk)->inet_sport = inet_sk(ssk)->inet_sport;
3349	inet_sk(msk)->inet_daddr = inet_sk(ssk)->inet_daddr;
3350	inet_sk(msk)->inet_saddr = inet_sk(ssk)->inet_saddr;
3351	inet_sk(msk)->inet_rcv_saddr = inet_sk(ssk)->inet_rcv_saddr;
3352}
3353
3354static void mptcp_destroy_common(struct mptcp_sock *msk)
3355{
3356	struct mptcp_subflow_context *subflow, *tmp;
3357	struct sock *sk = (struct sock *)msk;
3358
3359	__mptcp_clear_xmit(sk);
3360	mptcp_backlog_purge(sk);
3361
3362	/* join list will be eventually flushed (with rst) at sock lock release time */
3363	mptcp_for_each_subflow_safe(msk, subflow, tmp)
3364		__mptcp_close_ssk(sk, mptcp_subflow_tcp_sock(subflow), subflow, 0);
3365
3366	__skb_queue_purge(&sk->sk_receive_queue);
3367	skb_rbtree_purge(&msk->out_of_order_queue);
3368
3369	/* move all the rx fwd alloc into the sk_mem_reclaim_final in
3370	 * inet_sock_destruct() will dispose it
3371	 */
3372	mptcp_token_destroy(msk);
3373	mptcp_pm_destroy(msk);
3374}
3375
3376static int mptcp_disconnect(struct sock *sk, int flags)
3377{
3378	struct mptcp_sock *msk = mptcp_sk(sk);
3379
3380	/* We are on the fastopen error path. We can't call straight into the
3381	 * subflows cleanup code due to lock nesting (we are already under
3382	 * msk->firstsocket lock).
3383	 */
3384	if (msk->fastopening)
3385		return -EBUSY;
3386
3387	mptcp_check_listen_stop(sk);
3388	mptcp_set_state(sk, TCP_CLOSE);
3389
3390	mptcp_stop_rtx_timer(sk);
3391	mptcp_stop_tout_timer(sk);
3392
3393	mptcp_pm_connection_closed(msk);
3394
3395	/* msk->subflow is still intact, the following will not free the first
3396	 * subflow
3397	 */
3398	mptcp_do_fastclose(sk);
3399	mptcp_destroy_common(msk);
3400
3401	/* The first subflow is already in TCP_CLOSE status, the following
3402	 * can't overlap with a fallback anymore
3403	 */
3404	spin_lock_bh(&msk->fallback_lock);
3405	msk->allow_subflows = true;
3406	msk->allow_infinite_fallback = true;
3407	WRITE_ONCE(msk->flags, 0);
3408	spin_unlock_bh(&msk->fallback_lock);
3409
3410	msk->cb_flags = 0;
3411	msk->recovery = false;
3412	WRITE_ONCE(msk->can_ack, false);
3413	WRITE_ONCE(msk->fully_established, false);
3414	WRITE_ONCE(msk->rcv_data_fin, false);
3415	WRITE_ONCE(msk->snd_data_fin_enable, false);
3416	WRITE_ONCE(msk->rcv_fastclose, false);
3417	WRITE_ONCE(msk->use_64bit_ack, false);
3418	WRITE_ONCE(msk->csum_enabled, mptcp_is_checksum_enabled(sock_net(sk)));
3419	mptcp_pm_data_reset(msk);
3420	mptcp_ca_reset(sk);
3421	msk->bytes_consumed = 0;
3422	msk->bytes_acked = 0;
3423	msk->bytes_received = 0;
3424	msk->bytes_sent = 0;
3425	msk->bytes_retrans = 0;
3426	msk->rcvspace_init = 0;
3427	msk->fastclosing = 0;
3428
3429	/* for fallback's sake */
3430	WRITE_ONCE(msk->ack_seq, 0);
3431
3432	WRITE_ONCE(sk->sk_shutdown, 0);
3433	sk_error_report(sk);
3434	return 0;
3435}
3436
3437#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3438static struct ipv6_pinfo *mptcp_inet6_sk(const struct sock *sk)
3439{
3440	struct mptcp6_sock *msk6 = container_of(mptcp_sk(sk), struct mptcp6_sock, msk);
3441
3442	return &msk6->np;
3443}
3444
3445static void mptcp_copy_ip6_options(struct sock *newsk, const struct sock *sk)
3446{
3447	const struct ipv6_pinfo *np = inet6_sk(sk);
3448	struct ipv6_txoptions *opt;
3449	struct ipv6_pinfo *newnp;
3450
3451	newnp = inet6_sk(newsk);
3452
3453	rcu_read_lock();
3454	opt = rcu_dereference(np->opt);
3455	if (opt) {
3456		opt = ipv6_dup_options(newsk, opt);
3457		if (!opt)
3458			net_warn_ratelimited("%s: Failed to copy ip6 options\n", __func__);
3459	}
3460	RCU_INIT_POINTER(newnp->opt, opt);
3461	rcu_read_unlock();
3462}
3463#endif
3464
3465static void mptcp_copy_ip_options(struct sock *newsk, const struct sock *sk)
3466{
3467	struct ip_options_rcu *inet_opt, *newopt = NULL;
3468	const struct inet_sock *inet = inet_sk(sk);
3469	struct inet_sock *newinet;
3470
3471	newinet = inet_sk(newsk);
3472
3473	rcu_read_lock();
3474	inet_opt = rcu_dereference(inet->inet_opt);
3475	if (inet_opt) {
3476		newopt = sock_kmemdup(newsk, inet_opt, sizeof(*inet_opt) +
3477				      inet_opt->opt.optlen, GFP_ATOMIC);
3478		if (!newopt)
3479			net_warn_ratelimited("%s: Failed to copy ip options\n", __func__);
3480	}
3481	RCU_INIT_POINTER(newinet->inet_opt, newopt);
3482	rcu_read_unlock();
3483}
3484
3485struct sock *mptcp_sk_clone_init(const struct sock *sk,
3486				 const struct mptcp_options_received *mp_opt,
3487				 struct sock *ssk,
3488				 struct request_sock *req)
3489{
3490	struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
3491	struct sock *nsk = sk_clone_lock(sk, GFP_ATOMIC);
3492	struct mptcp_subflow_context *subflow;
3493	struct mptcp_sock *msk;
3494
3495	if (!nsk)
3496		return NULL;
3497
3498#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3499	if (nsk->sk_family == AF_INET6)
3500		inet_sk(nsk)->pinet6 = mptcp_inet6_sk(nsk);
3501#endif
3502
3503	__mptcp_init_sock(nsk);
3504
3505#if IS_ENABLED(CONFIG_MPTCP_IPV6)
3506	if (nsk->sk_family == AF_INET6)
3507		mptcp_copy_ip6_options(nsk, sk);
3508	else
3509#endif
3510		mptcp_copy_ip_options(nsk, sk);
3511
3512	msk = mptcp_sk(nsk);
3513	WRITE_ONCE(msk->local_key, subflow_req->local_key);
3514	WRITE_ONCE(msk->token, subflow_req->token);
3515	msk->in_accept_queue = 1;
3516	WRITE_ONCE(msk->fully_established, false);
3517	if (mp_opt->suboptions & OPTION_MPTCP_CSUMREQD)
3518		WRITE_ONCE(msk->csum_enabled, true);
3519
3520	WRITE_ONCE(msk->write_seq, subflow_req->idsn + 1);
3521	WRITE_ONCE(msk->snd_nxt, msk->write_seq);
3522	WRITE_ONCE(msk->snd_una, msk->write_seq);
3523	WRITE_ONCE(msk->wnd_end, msk->snd_nxt + tcp_sk(ssk)->snd_wnd);
3524	msk->setsockopt_seq = mptcp_sk(sk)->setsockopt_seq;
3525	mptcp_init_sched(msk, mptcp_sk(sk)->sched);
3526
3527	/* passive msk is created after the first/MPC subflow */
3528	msk->subflow_id = 2;
3529
3530	sock_reset_flag(nsk, SOCK_RCU_FREE);
3531	security_inet_csk_clone(nsk, req);
3532
3533	/* this can't race with mptcp_close(), as the msk is
3534	 * not yet exposted to user-space
3535	 */
3536	mptcp_set_state(nsk, TCP_ESTABLISHED);
3537
3538	/* The msk maintain a ref to each subflow in the connections list */
3539	WRITE_ONCE(msk->first, ssk);
3540	subflow = mptcp_subflow_ctx(ssk);
3541	list_add(&subflow->node, &msk->conn_list);
3542	sock_hold(ssk);
3543
3544	/* new mpc subflow takes ownership of the newly
3545	 * created mptcp socket
3546	 */
3547	mptcp_token_accept(subflow_req, msk);
3548
3549	/* set msk addresses early to ensure mptcp_pm_get_local_id()
3550	 * uses the correct data
3551	 */
3552	mptcp_copy_inaddrs(nsk, ssk);
3553	__mptcp_propagate_sndbuf(nsk, ssk);
3554
3555	mptcp_rcv_space_init(msk, ssk);
3556	msk->rcvq_space.time = mptcp_stamp();
3557
3558	if (mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
3559		__mptcp_subflow_fully_established(msk, subflow, mp_opt);
3560	bh_unlock_sock(nsk);
3561
3562	/* note: the newly allocated socket refcount is 2 now */
3563	return nsk;
3564}
3565
3566void mptcp_rcv_space_init(struct mptcp_sock *msk, const struct sock *ssk)
3567{
3568	const struct tcp_sock *tp = tcp_sk(ssk);
3569
3570	msk->rcvspace_init = 1;
3571	msk->rcvq_space.copied = 0;
3572	msk->rcvq_space.rtt_us = 0;
3573
3574	/* initial rcv_space offering made to peer */
3575	msk->rcvq_space.space = min_t(u32, tp->rcv_wnd,
3576				      TCP_INIT_CWND * tp->advmss);
3577	if (msk->rcvq_space.space == 0)
3578		msk->rcvq_space.space = TCP_INIT_CWND * TCP_MSS_DEFAULT;
3579}
3580
3581static void mptcp_destroy(struct sock *sk)
3582{
3583	struct mptcp_sock *msk = mptcp_sk(sk);
3584
3585	/* allow the following to close even the initial subflow */
3586	msk->free_first = 1;
3587	mptcp_destroy_common(msk);
3588	sk_sockets_allocated_dec(sk);
3589}
3590
3591void __mptcp_data_acked(struct sock *sk)
3592{
3593	if (!sock_owned_by_user(sk))
3594		__mptcp_clean_una(sk);
3595	else
3596		__set_bit(MPTCP_CLEAN_UNA, &mptcp_sk(sk)->cb_flags);
3597}
3598
3599void __mptcp_check_push(struct sock *sk, struct sock *ssk)
3600{
3601	if (!sock_owned_by_user(sk))
3602		__mptcp_subflow_push_pending(sk, ssk, false);
3603	else
3604		__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3605}
3606
3607#define MPTCP_FLAGS_PROCESS_CTX_NEED (BIT(MPTCP_PUSH_PENDING) | \
3608				      BIT(MPTCP_RETRANSMIT) | \
3609				      BIT(MPTCP_FLUSH_JOIN_LIST))
3610
3611/* processes deferred events and flush wmem */
3612static void mptcp_release_cb(struct sock *sk)
3613	__must_hold(&sk->sk_lock.slock)
3614{
3615	struct mptcp_sock *msk = mptcp_sk(sk);
3616
3617	for (;;) {
3618		unsigned long flags = (msk->cb_flags & MPTCP_FLAGS_PROCESS_CTX_NEED);
3619		struct list_head join_list, skbs;
3620		bool spool_bl;
3621		u32 moved;
3622
3623		spool_bl = mptcp_can_spool_backlog(sk, &skbs);
3624		if (!flags && !spool_bl)
3625			break;
3626
3627		INIT_LIST_HEAD(&join_list);
3628		list_splice_init(&msk->join_list, &join_list);
3629
3630		/* the following actions acquire the subflow socket lock
3631		 *
3632		 * 1) can't be invoked in atomic scope
3633		 * 2) must avoid ABBA deadlock with msk socket spinlock: the RX
3634		 *    datapath acquires the msk socket spinlock while helding
3635		 *    the subflow socket lock
3636		 */
3637		msk->cb_flags &= ~flags;
3638		spin_unlock_bh(&sk->sk_lock.slock);
3639
3640		if (flags & BIT(MPTCP_FLUSH_JOIN_LIST))
3641			__mptcp_flush_join_list(sk, &join_list);
3642		if (flags & BIT(MPTCP_PUSH_PENDING))
3643			__mptcp_push_pending(sk, 0);
3644		if (flags & BIT(MPTCP_RETRANSMIT))
3645			__mptcp_retrans(sk);
3646		if (spool_bl && __mptcp_move_skbs(sk, &skbs, &moved)) {
3647			/* notify ack seq update */
3648			mptcp_cleanup_rbuf(msk, 0);
3649			sk->sk_data_ready(sk);
3650		}
3651
3652		cond_resched();
3653		spin_lock_bh(&sk->sk_lock.slock);
3654		if (spool_bl)
3655			mptcp_backlog_spooled(sk, moved, &skbs);
3656	}
3657
3658	if (__test_and_clear_bit(MPTCP_CLEAN_UNA, &msk->cb_flags))
3659		__mptcp_clean_una_wakeup(sk);
3660	if (unlikely(msk->cb_flags)) {
3661		/* be sure to sync the msk state before taking actions
3662		 * depending on sk_state (MPTCP_ERROR_REPORT)
3663		 * On sk release avoid actions depending on the first subflow
3664		 */
3665		if (__test_and_clear_bit(MPTCP_SYNC_STATE, &msk->cb_flags) && msk->first)
3666			__mptcp_sync_state(sk, msk->pending_state);
3667		if (__test_and_clear_bit(MPTCP_ERROR_REPORT, &msk->cb_flags))
3668			__mptcp_error_report(sk);
3669		if (__test_and_clear_bit(MPTCP_SYNC_SNDBUF, &msk->cb_flags))
3670			__mptcp_sync_sndbuf(sk);
3671	}
3672}
3673
3674/* MP_JOIN client subflow must wait for 4th ack before sending any data:
3675 * TCP can't schedule delack timer before the subflow is fully established.
3676 * MPTCP uses the delack timer to do 3rd ack retransmissions
3677 */
3678static void schedule_3rdack_retransmission(struct sock *ssk)
3679{
3680	struct inet_connection_sock *icsk = inet_csk(ssk);
3681	struct tcp_sock *tp = tcp_sk(ssk);
3682	unsigned long timeout;
3683
3684	if (READ_ONCE(mptcp_subflow_ctx(ssk)->fully_established))
3685		return;
3686
3687	/* reschedule with a timeout above RTT, as we must look only for drop */
3688	if (tp->srtt_us)
3689		timeout = usecs_to_jiffies(tp->srtt_us >> (3 - 1));
3690	else
3691		timeout = TCP_TIMEOUT_INIT;
3692	timeout += jiffies;
3693
3694	WARN_ON_ONCE(icsk->icsk_ack.pending & ICSK_ACK_TIMER);
3695	smp_store_release(&icsk->icsk_ack.pending,
3696			  icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
3697	sk_reset_timer(ssk, &icsk->icsk_delack_timer, timeout);
3698}
3699
3700void mptcp_subflow_process_delegated(struct sock *ssk, long status)
3701{
3702	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3703	struct sock *sk = subflow->conn;
3704
3705	if (status & BIT(MPTCP_DELEGATE_SEND)) {
3706		mptcp_data_lock(sk);
3707		if (!sock_owned_by_user(sk))
3708			__mptcp_subflow_push_pending(sk, ssk, true);
3709		else
3710			__set_bit(MPTCP_PUSH_PENDING, &mptcp_sk(sk)->cb_flags);
3711		mptcp_data_unlock(sk);
3712	}
3713	if (status & BIT(MPTCP_DELEGATE_SNDBUF)) {
3714		mptcp_data_lock(sk);
3715		if (!sock_owned_by_user(sk))
3716			__mptcp_sync_sndbuf(sk);
3717		else
3718			__set_bit(MPTCP_SYNC_SNDBUF, &mptcp_sk(sk)->cb_flags);
3719		mptcp_data_unlock(sk);
3720	}
3721	if (status & BIT(MPTCP_DELEGATE_ACK))
3722		schedule_3rdack_retransmission(ssk);
3723}
3724
3725static int mptcp_hash(struct sock *sk)
3726{
3727	/* should never be called,
3728	 * we hash the TCP subflows not the MPTCP socket
3729	 */
3730	WARN_ON_ONCE(1);
3731	return 0;
3732}
3733
3734static void mptcp_unhash(struct sock *sk)
3735{
3736	/* called from sk_common_release(), but nothing to do here */
3737}
3738
3739static int mptcp_get_port(struct sock *sk, unsigned short snum)
3740{
3741	struct mptcp_sock *msk = mptcp_sk(sk);
3742
3743	pr_debug("msk=%p, ssk=%p\n", msk, msk->first);
3744	if (WARN_ON_ONCE(!msk->first))
3745		return -EINVAL;
3746
3747	return inet_csk_get_port(msk->first, snum);
3748}
3749
3750void mptcp_finish_connect(struct sock *ssk)
3751{
3752	struct mptcp_subflow_context *subflow;
3753	struct mptcp_sock *msk;
3754	struct sock *sk;
3755
3756	subflow = mptcp_subflow_ctx(ssk);
3757	sk = subflow->conn;
3758	msk = mptcp_sk(sk);
3759
3760	pr_debug("msk=%p, token=%u\n", sk, subflow->token);
3761
3762	subflow->map_seq = subflow->iasn;
3763	subflow->map_subflow_seq = 1;
3764
3765	/* the socket is not connected yet, no msk/subflow ops can access/race
3766	 * accessing the field below
3767	 */
3768	WRITE_ONCE(msk->local_key, subflow->local_key);
3769	WRITE_ONCE(msk->rcvq_space.time, mptcp_stamp());
3770
3771	mptcp_pm_new_connection(msk, ssk, 0);
3772}
3773
3774void mptcp_sock_graft(struct sock *sk, struct socket *parent)
3775{
3776	write_lock_bh(&sk->sk_callback_lock);
3777	rcu_assign_pointer(sk->sk_wq, &parent->wq);
3778	sk_set_socket(sk, parent);
3779	write_unlock_bh(&sk->sk_callback_lock);
3780}
3781
3782/* Can be called without holding the msk socket lock; use the callback lock
3783 * to avoid {READ_,WRITE_}ONCE annotations on sk_socket.
3784 */
3785static void mptcp_sock_check_graft(struct sock *sk, struct sock *ssk)
3786{
3787	struct socket *sock;
3788
3789	write_lock_bh(&sk->sk_callback_lock);
3790	sock = sk->sk_socket;
3791	write_unlock_bh(&sk->sk_callback_lock);
3792	if (sock) {
3793		mptcp_sock_graft(ssk, sock);
3794		__mptcp_inherit_cgrp_data(sk, ssk);
3795		__mptcp_inherit_memcg(sk, ssk, GFP_ATOMIC);
3796	}
3797}
3798
3799bool mptcp_finish_join(struct sock *ssk)
3800{
3801	struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
3802	struct mptcp_sock *msk = mptcp_sk(subflow->conn);
3803	struct sock *parent = (void *)msk;
3804	bool ret = true;
3805
3806	pr_debug("msk=%p, subflow=%p\n", msk, subflow);
3807
3808	/* mptcp socket already closing? */
3809	if (!mptcp_is_fully_established(parent)) {
3810		subflow->reset_reason = MPTCP_RST_EMPTCP;
3811		return false;
3812	}
3813
3814	/* Active subflow, already present inside the conn_list; is grafted
3815	 * either by __mptcp_subflow_connect() or accept.
3816	 */
3817	if (!list_empty(&subflow->node)) {
3818		spin_lock_bh(&msk->fallback_lock);
3819		if (!msk->allow_subflows) {
3820			spin_unlock_bh(&msk->fallback_lock);
3821			return false;
3822		}
3823		mptcp_subflow_joined(msk, ssk);
3824		spin_unlock_bh(&msk->fallback_lock);
3825		mptcp_propagate_sndbuf(parent, ssk);
3826		return true;
3827	}
3828
3829	if (!mptcp_pm_allow_new_subflow(msk)) {
3830		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_JOINREJECTED);
3831		goto err_prohibited;
3832	}
3833
3834	/* If we can't acquire msk socket lock here, let the release callback
3835	 * handle it
3836	 */
3837	mptcp_data_lock(parent);
3838	if (!sock_owned_by_user(parent)) {
3839		ret = __mptcp_finish_join(msk, ssk);
3840		if (ret) {
3841			sock_hold(ssk);
3842			list_add_tail(&subflow->node, &msk->conn_list);
3843			mptcp_sock_check_graft(parent, ssk);
3844		}
3845	} else {
3846		sock_hold(ssk);
3847		list_add_tail(&subflow->node, &msk->join_list);
3848		__set_bit(MPTCP_FLUSH_JOIN_LIST, &msk->cb_flags);
3849
3850		/* In case of later failures, __mptcp_flush_join_list() will
3851		 * properly orphan the ssk via mptcp_close_ssk().
3852		 */
3853		mptcp_sock_check_graft(parent, ssk);
3854	}
3855	mptcp_data_unlock(parent);
3856
3857	if (!ret) {
3858err_prohibited:
3859		subflow->reset_reason = MPTCP_RST_EPROHIBIT;
3860		return false;
3861	}
3862
3863	return true;
3864}
3865
3866static void mptcp_shutdown(struct sock *sk, int how)
3867{
3868	pr_debug("sk=%p, how=%d\n", sk, how);
3869
3870	if ((how & SEND_SHUTDOWN) && mptcp_close_state(sk))
3871		__mptcp_wr_shutdown(sk);
3872}
3873
3874static int mptcp_ioctl_outq(const struct mptcp_sock *msk, u64 v)
3875{
3876	const struct sock *sk = (void *)msk;
3877	u64 delta;
3878
3879	if (sk->sk_state == TCP_LISTEN)
3880		return -EINVAL;
3881
3882	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
3883		return 0;
3884
3885	delta = msk->write_seq - v;
3886	if (__mptcp_check_fallback(msk) && msk->first) {
3887		struct tcp_sock *tp = tcp_sk(msk->first);
3888
3889		/* the first subflow is disconnected after close - see
3890		 * __mptcp_close_ssk(). tcp_disconnect() moves the write_seq
3891		 * so ignore that status, too.
3892		 */
3893		if (!((1 << msk->first->sk_state) &
3894		      (TCPF_SYN_SENT | TCPF_SYN_RECV | TCPF_CLOSE)))
3895			delta += READ_ONCE(tp->write_seq) - tp->snd_una;
3896	}
3897	if (delta > INT_MAX)
3898		delta = INT_MAX;
3899
3900	return (int)delta;
3901}
3902
3903static int mptcp_ioctl(struct sock *sk, int cmd, int *karg)
3904{
3905	struct mptcp_sock *msk = mptcp_sk(sk);
3906	bool slow;
3907
3908	switch (cmd) {
3909	case SIOCINQ:
3910		if (sk->sk_state == TCP_LISTEN)
3911			return -EINVAL;
3912
3913		lock_sock(sk);
3914		if (mptcp_move_skbs(sk))
3915			mptcp_cleanup_rbuf(msk, 0);
3916		*karg = mptcp_inq_hint(sk);
3917		release_sock(sk);
3918		break;
3919	case SIOCOUTQ:
3920		slow = lock_sock_fast(sk);
3921		*karg = mptcp_ioctl_outq(msk, READ_ONCE(msk->snd_una));
3922		unlock_sock_fast(sk, slow);
3923		break;
3924	case SIOCOUTQNSD:
3925		slow = lock_sock_fast(sk);
3926		*karg = mptcp_ioctl_outq(msk, msk->snd_nxt);
3927		unlock_sock_fast(sk, slow);
3928		break;
3929	default:
3930		return -ENOIOCTLCMD;
3931	}
3932
3933	return 0;
3934}
3935
3936static int mptcp_connect(struct sock *sk, struct sockaddr_unsized *uaddr,
3937			 int addr_len)
3938{
3939	struct mptcp_subflow_context *subflow;
3940	struct mptcp_sock *msk = mptcp_sk(sk);
3941	int err = -EINVAL;
3942	struct sock *ssk;
3943
3944	ssk = __mptcp_nmpc_sk(msk);
3945	if (IS_ERR(ssk))
3946		return PTR_ERR(ssk);
3947
3948	mptcp_set_state(sk, TCP_SYN_SENT);
3949	subflow = mptcp_subflow_ctx(ssk);
3950#ifdef CONFIG_TCP_MD5SIG
3951	/* no MPTCP if MD5SIG is enabled on this socket or we may run out of
3952	 * TCP option space.
3953	 */
3954	if (rcu_access_pointer(tcp_sk(ssk)->md5sig_info))
3955		mptcp_early_fallback(msk, subflow, MPTCP_MIB_MD5SIGFALLBACK);
3956#endif
3957	if (subflow->request_mptcp) {
3958		if (mptcp_active_should_disable(sk))
3959			mptcp_early_fallback(msk, subflow,
3960					     MPTCP_MIB_MPCAPABLEACTIVEDISABLED);
3961		else if (mptcp_token_new_connect(ssk) < 0)
3962			mptcp_early_fallback(msk, subflow,
3963					     MPTCP_MIB_TOKENFALLBACKINIT);
3964	}
3965
3966	WRITE_ONCE(msk->write_seq, subflow->idsn);
3967	WRITE_ONCE(msk->snd_nxt, subflow->idsn);
3968	WRITE_ONCE(msk->snd_una, subflow->idsn);
3969	if (likely(!__mptcp_check_fallback(msk)))
3970		MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPCAPABLEACTIVE);
3971
3972	/* if reaching here via the fastopen/sendmsg path, the caller already
3973	 * acquired the subflow socket lock, too.
3974	 */
3975	if (!msk->fastopening)
3976		lock_sock(ssk);
3977
3978	/* the following mirrors closely a very small chunk of code from
3979	 * __inet_stream_connect()
3980	 */
3981	if (ssk->sk_state != TCP_CLOSE)
3982		goto out;
3983
3984	if (BPF_CGROUP_PRE_CONNECT_ENABLED(ssk)) {
3985		err = ssk->sk_prot->pre_connect(ssk, uaddr, addr_len);
3986		if (err)
3987			goto out;
3988	}
3989
3990	err = ssk->sk_prot->connect(ssk, uaddr, addr_len);
3991	if (err < 0)
3992		goto out;
3993
3994	inet_assign_bit(DEFER_CONNECT, sk, inet_test_bit(DEFER_CONNECT, ssk));
3995
3996out:
3997	if (!msk->fastopening)
3998		release_sock(ssk);
3999
4000	/* on successful connect, the msk state will be moved to established by
4001	 * subflow_finish_connect()
4002	 */
4003	if (unlikely(err)) {
4004		/* avoid leaving a dangling token in an unconnected socket */
4005		mptcp_token_destroy(msk);
4006		mptcp_set_state(sk, TCP_CLOSE);
4007		return err;
4008	}
4009
4010	mptcp_copy_inaddrs(sk, ssk);
4011	return 0;
4012}
4013
4014static struct proto mptcp_prot = {
4015	.name		= "MPTCP",
4016	.owner		= THIS_MODULE,
4017	.init		= mptcp_init_sock,
4018	.connect	= mptcp_connect,
4019	.disconnect	= mptcp_disconnect,
4020	.close		= mptcp_close,
4021	.setsockopt	= mptcp_setsockopt,
4022	.getsockopt	= mptcp_getsockopt,
4023	.shutdown	= mptcp_shutdown,
4024	.destroy	= mptcp_destroy,
4025	.sendmsg	= mptcp_sendmsg,
4026	.ioctl		= mptcp_ioctl,
4027	.recvmsg	= mptcp_recvmsg,
4028	.release_cb	= mptcp_release_cb,
4029	.hash		= mptcp_hash,
4030	.unhash		= mptcp_unhash,
4031	.get_port	= mptcp_get_port,
4032	.stream_memory_free	= mptcp_stream_memory_free,
4033	.sockets_allocated	= &mptcp_sockets_allocated,
4034
4035	.memory_allocated	= &net_aligned_data.tcp_memory_allocated,
4036	.per_cpu_fw_alloc	= &tcp_memory_per_cpu_fw_alloc,
4037
4038	.memory_pressure	= &tcp_memory_pressure,
4039	.sysctl_wmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_wmem),
4040	.sysctl_rmem_offset	= offsetof(struct net, ipv4.sysctl_tcp_rmem),
4041	.sysctl_mem	= sysctl_tcp_mem,
4042	.obj_size	= sizeof(struct mptcp_sock),
4043	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
4044	.no_autobind	= true,
4045};
4046
4047static int mptcp_bind(struct socket *sock, struct sockaddr_unsized *uaddr, int addr_len)
4048{
4049	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4050	struct sock *ssk, *sk = sock->sk;
4051	int err = -EINVAL;
4052
4053	lock_sock(sk);
4054	ssk = __mptcp_nmpc_sk(msk);
4055	if (IS_ERR(ssk)) {
4056		err = PTR_ERR(ssk);
4057		goto unlock;
4058	}
4059
4060	if (sk->sk_family == AF_INET)
4061		err = inet_bind_sk(ssk, uaddr, addr_len);
4062#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4063	else if (sk->sk_family == AF_INET6)
4064		err = inet6_bind_sk(ssk, uaddr, addr_len);
4065#endif
4066	if (!err)
4067		mptcp_copy_inaddrs(sk, ssk);
4068
4069unlock:
4070	release_sock(sk);
4071	return err;
4072}
4073
4074static int mptcp_listen(struct socket *sock, int backlog)
4075{
4076	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4077	struct sock *sk = sock->sk;
4078	struct sock *ssk;
4079	int err;
4080
4081	pr_debug("msk=%p\n", msk);
4082
4083	lock_sock(sk);
4084
4085	err = -EINVAL;
4086	if (sock->state != SS_UNCONNECTED || sock->type != SOCK_STREAM)
4087		goto unlock;
4088
4089	ssk = __mptcp_nmpc_sk(msk);
4090	if (IS_ERR(ssk)) {
4091		err = PTR_ERR(ssk);
4092		goto unlock;
4093	}
4094
4095	mptcp_set_state(sk, TCP_LISTEN);
4096	sock_set_flag(sk, SOCK_RCU_FREE);
4097
4098	lock_sock(ssk);
4099	err = __inet_listen_sk(ssk, backlog);
4100	release_sock(ssk);
4101	mptcp_set_state(sk, inet_sk_state_load(ssk));
4102
4103	if (!err) {
4104		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
4105		mptcp_copy_inaddrs(sk, ssk);
4106		mptcp_event_pm_listener(ssk, MPTCP_EVENT_LISTENER_CREATED);
4107	}
4108
4109unlock:
4110	release_sock(sk);
4111	return err;
4112}
4113
4114static void mptcp_graft_subflows(struct sock *sk)
4115{
4116	struct mptcp_subflow_context *subflow;
4117	struct mptcp_sock *msk = mptcp_sk(sk);
4118
4119	if (mem_cgroup_sockets_enabled) {
4120		LIST_HEAD(join_list);
4121
4122		/* Subflows joining after __inet_accept() will get the
4123		 * mem CG properly initialized at mptcp_finish_join() time,
4124		 * but subflows pending in join_list need explicit
4125		 * initialization before flushing `backlog_unaccounted`
4126		 * or MPTCP can later unexpectedly observe unaccounted memory.
4127		 */
4128		mptcp_data_lock(sk);
4129		list_splice_init(&msk->join_list, &join_list);
4130		mptcp_data_unlock(sk);
4131
4132		__mptcp_flush_join_list(sk, &join_list);
4133	}
4134
4135	mptcp_for_each_subflow(msk, subflow) {
4136		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4137
4138		lock_sock(ssk);
4139
4140		/* Set ssk->sk_socket of accept()ed flows to mptcp socket.
4141		 * This is needed so NOSPACE flag can be set from tcp stack.
4142		 */
4143		if (!ssk->sk_socket)
4144			mptcp_sock_graft(ssk, sk->sk_socket);
4145
4146		if (!mem_cgroup_sk_enabled(sk))
4147			goto unlock;
4148
4149		__mptcp_inherit_cgrp_data(sk, ssk);
4150		__mptcp_inherit_memcg(sk, ssk, GFP_KERNEL);
4151
4152unlock:
4153		release_sock(ssk);
4154	}
4155
4156	if (mem_cgroup_sk_enabled(sk)) {
4157		gfp_t gfp = GFP_KERNEL | __GFP_NOFAIL;
4158		int amt;
4159
4160		/* Account the backlog memory; prior accept() is aware of
4161		 * fwd and rmem only.
4162		 */
4163		mptcp_data_lock(sk);
4164		amt = sk_mem_pages(sk->sk_forward_alloc +
4165				   msk->backlog_unaccounted +
4166				   atomic_read(&sk->sk_rmem_alloc)) -
4167		      sk_mem_pages(sk->sk_forward_alloc +
4168				   atomic_read(&sk->sk_rmem_alloc));
4169		msk->backlog_unaccounted = 0;
4170		mptcp_data_unlock(sk);
4171
4172		if (amt)
4173			mem_cgroup_sk_charge(sk, amt, gfp);
4174	}
4175}
4176
4177static int mptcp_stream_accept(struct socket *sock, struct socket *newsock,
4178			       struct proto_accept_arg *arg)
4179{
4180	struct mptcp_sock *msk = mptcp_sk(sock->sk);
4181	struct sock *ssk, *newsk;
4182
4183	pr_debug("msk=%p\n", msk);
4184
4185	/* Buggy applications can call accept on socket states other then LISTEN
4186	 * but no need to allocate the first subflow just to error out.
4187	 */
4188	ssk = READ_ONCE(msk->first);
4189	if (!ssk)
4190		return -EINVAL;
4191
4192	pr_debug("ssk=%p, listener=%p\n", ssk, mptcp_subflow_ctx(ssk));
4193	newsk = inet_csk_accept(ssk, arg);
4194	if (!newsk)
4195		return arg->err;
4196
4197	pr_debug("newsk=%p, subflow is mptcp=%d\n", newsk, sk_is_mptcp(newsk));
4198	if (sk_is_mptcp(newsk)) {
4199		struct mptcp_subflow_context *subflow;
4200		struct sock *new_mptcp_sock;
4201
4202		subflow = mptcp_subflow_ctx(newsk);
4203		new_mptcp_sock = subflow->conn;
4204
4205		/* is_mptcp should be false if subflow->conn is missing, see
4206		 * subflow_syn_recv_sock()
4207		 */
4208		if (WARN_ON_ONCE(!new_mptcp_sock)) {
4209			tcp_sk(newsk)->is_mptcp = 0;
4210			goto tcpfallback;
4211		}
4212
4213		newsk = new_mptcp_sock;
4214		MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPCAPABLEPASSIVEACK);
4215
4216		newsk->sk_kern_sock = arg->kern;
4217		lock_sock(newsk);
4218		__inet_accept(sock, newsock, newsk);
4219
4220		set_bit(SOCK_CUSTOM_SOCKOPT, &newsock->flags);
4221		msk = mptcp_sk(newsk);
4222		msk->in_accept_queue = 0;
4223
4224		mptcp_graft_subflows(newsk);
4225		mptcp_rps_record_subflows(msk);
4226
4227		/* Do late cleanup for the first subflow as necessary. Also
4228		 * deal with bad peers not doing a complete shutdown.
4229		 */
4230		if (unlikely(inet_sk_state_load(msk->first) == TCP_CLOSE)) {
4231			if (unlikely(list_is_singular(&msk->conn_list)))
4232				mptcp_set_state(newsk, TCP_CLOSE);
4233			mptcp_close_ssk(newsk, msk->first,
4234					mptcp_subflow_ctx(msk->first));
4235		}
4236	} else {
4237tcpfallback:
4238		newsk->sk_kern_sock = arg->kern;
4239		lock_sock(newsk);
4240		__inet_accept(sock, newsock, newsk);
4241		/* we are being invoked after accepting a non-mp-capable
4242		 * flow: sk is a tcp_sk, not an mptcp one.
4243		 *
4244		 * Hand the socket over to tcp so all further socket ops
4245		 * bypass mptcp.
4246		 */
4247		WRITE_ONCE(newsock->sk->sk_socket->ops,
4248			   mptcp_fallback_tcp_ops(newsock->sk));
4249	}
4250	release_sock(newsk);
4251
4252	return 0;
4253}
4254
4255static __poll_t mptcp_check_writeable(struct mptcp_sock *msk)
4256{
4257	struct sock *sk = (struct sock *)msk;
4258
4259	if (__mptcp_stream_is_writeable(sk, 1))
4260		return EPOLLOUT | EPOLLWRNORM;
4261
4262	set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
4263	smp_mb__after_atomic(); /* NOSPACE is changed by mptcp_write_space() */
4264	if (__mptcp_stream_is_writeable(sk, 1))
4265		return EPOLLOUT | EPOLLWRNORM;
4266
4267	return 0;
4268}
4269
4270static __poll_t mptcp_poll(struct file *file, struct socket *sock,
4271			   struct poll_table_struct *wait)
4272{
4273	struct sock *sk = sock->sk;
4274	struct mptcp_sock *msk;
4275	__poll_t mask = 0;
4276	u8 shutdown;
4277	int state;
4278
4279	msk = mptcp_sk(sk);
4280	sock_poll_wait(file, sock, wait);
4281
4282	state = inet_sk_state_load(sk);
4283	pr_debug("msk=%p state=%d flags=%lx\n", msk, state, msk->flags);
4284	if (state == TCP_LISTEN) {
4285		struct sock *ssk = READ_ONCE(msk->first);
4286
4287		if (WARN_ON_ONCE(!ssk))
4288			return 0;
4289
4290		return inet_csk_listen_poll(ssk);
4291	}
4292
4293	shutdown = READ_ONCE(sk->sk_shutdown);
4294	if (shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
4295		mask |= EPOLLHUP;
4296	if (shutdown & RCV_SHUTDOWN)
4297		mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
4298
4299	if (state != TCP_SYN_SENT && state != TCP_SYN_RECV) {
4300		mask |= mptcp_check_readable(sk);
4301		if (shutdown & SEND_SHUTDOWN)
4302			mask |= EPOLLOUT | EPOLLWRNORM;
4303		else
4304			mask |= mptcp_check_writeable(msk);
4305	} else if (state == TCP_SYN_SENT &&
4306		   inet_test_bit(DEFER_CONNECT, sk)) {
4307		/* cf tcp_poll() note about TFO */
4308		mask |= EPOLLOUT | EPOLLWRNORM;
4309	}
4310
4311	/* This barrier is coupled with smp_wmb() in __mptcp_error_report() */
4312	smp_rmb();
4313	if (READ_ONCE(sk->sk_err))
4314		mask |= EPOLLERR;
4315
4316	return mask;
4317}
4318
4319static const struct proto_ops mptcp_stream_ops = {
4320	.family		   = PF_INET,
4321	.owner		   = THIS_MODULE,
4322	.release	   = inet_release,
4323	.bind		   = mptcp_bind,
4324	.connect	   = inet_stream_connect,
4325	.socketpair	   = sock_no_socketpair,
4326	.accept		   = mptcp_stream_accept,
4327	.getname	   = inet_getname,
4328	.poll		   = mptcp_poll,
4329	.ioctl		   = inet_ioctl,
4330	.gettstamp	   = sock_gettstamp,
4331	.listen		   = mptcp_listen,
4332	.shutdown	   = inet_shutdown,
4333	.setsockopt	   = sock_common_setsockopt,
4334	.getsockopt	   = sock_common_getsockopt,
4335	.sendmsg	   = inet_sendmsg,
4336	.recvmsg	   = inet_recvmsg,
4337	.mmap		   = sock_no_mmap,
4338	.set_rcvlowat	   = mptcp_set_rcvlowat,
4339};
4340
4341static struct inet_protosw mptcp_protosw = {
4342	.type		= SOCK_STREAM,
4343	.protocol	= IPPROTO_MPTCP,
4344	.prot		= &mptcp_prot,
4345	.ops		= &mptcp_stream_ops,
4346	.flags		= INET_PROTOSW_ICSK,
4347};
4348
4349static int mptcp_napi_poll(struct napi_struct *napi, int budget)
4350{
4351	struct mptcp_delegated_action *delegated;
4352	struct mptcp_subflow_context *subflow;
4353	int work_done = 0;
4354
4355	delegated = container_of(napi, struct mptcp_delegated_action, napi);
4356	while ((subflow = mptcp_subflow_delegated_next(delegated)) != NULL) {
4357		struct sock *ssk = mptcp_subflow_tcp_sock(subflow);
4358
4359		bh_lock_sock_nested(ssk);
4360		if (!sock_owned_by_user(ssk)) {
4361			mptcp_subflow_process_delegated(ssk, xchg(&subflow->delegated_status, 0));
4362		} else {
4363			/* tcp_release_cb_override already processed
4364			 * the action or will do at next release_sock().
4365			 * In both case must dequeue the subflow here - on the same
4366			 * CPU that scheduled it.
4367			 */
4368			smp_wmb();
4369			clear_bit(MPTCP_DELEGATE_SCHEDULED, &subflow->delegated_status);
4370		}
4371		bh_unlock_sock(ssk);
4372		sock_put(ssk);
4373
4374		if (++work_done == budget)
4375			return budget;
4376	}
4377
4378	/* always provide a 0 'work_done' argument, so that napi_complete_done
4379	 * will not try accessing the NULL napi->dev ptr
4380	 */
4381	napi_complete_done(napi, 0);
4382	return work_done;
4383}
4384
4385void __init mptcp_proto_init(void)
4386{
4387	struct mptcp_delegated_action *delegated;
4388	int cpu;
4389
4390	mptcp_prot.h.hashinfo = tcp_prot.h.hashinfo;
4391
4392	if (percpu_counter_init(&mptcp_sockets_allocated, 0, GFP_KERNEL))
4393		panic("Failed to allocate MPTCP pcpu counter\n");
4394
4395	mptcp_napi_dev = alloc_netdev_dummy(0);
4396	if (!mptcp_napi_dev)
4397		panic("Failed to allocate MPTCP dummy netdev\n");
4398	for_each_possible_cpu(cpu) {
4399		delegated = per_cpu_ptr(&mptcp_delegated_actions, cpu);
4400		INIT_LIST_HEAD(&delegated->head);
4401		netif_napi_add_tx(mptcp_napi_dev, &delegated->napi,
4402				  mptcp_napi_poll);
4403		napi_enable(&delegated->napi);
4404	}
4405
4406	mptcp_subflow_init();
4407	mptcp_pm_init();
4408	mptcp_sched_init();
4409	mptcp_token_init();
4410
4411	if (proto_register(&mptcp_prot, 1) != 0)
4412		panic("Failed to register MPTCP proto.\n");
4413
4414	inet_register_protosw(&mptcp_protosw);
4415
4416	BUILD_BUG_ON(sizeof(struct mptcp_skb_cb) > sizeof_field(struct sk_buff, cb));
4417}
4418
4419#if IS_ENABLED(CONFIG_MPTCP_IPV6)
4420static const struct proto_ops mptcp_v6_stream_ops = {
4421	.family		   = PF_INET6,
4422	.owner		   = THIS_MODULE,
4423	.release	   = inet6_release,
4424	.bind		   = mptcp_bind,
4425	.connect	   = inet_stream_connect,
4426	.socketpair	   = sock_no_socketpair,
4427	.accept		   = mptcp_stream_accept,
4428	.getname	   = inet6_getname,
4429	.poll		   = mptcp_poll,
4430	.ioctl		   = inet6_ioctl,
4431	.gettstamp	   = sock_gettstamp,
4432	.listen		   = mptcp_listen,
4433	.shutdown	   = inet_shutdown,
4434	.setsockopt	   = sock_common_setsockopt,
4435	.getsockopt	   = sock_common_getsockopt,
4436	.sendmsg	   = inet6_sendmsg,
4437	.recvmsg	   = inet6_recvmsg,
4438	.mmap		   = sock_no_mmap,
4439#ifdef CONFIG_COMPAT
4440	.compat_ioctl	   = inet6_compat_ioctl,
4441#endif
4442	.set_rcvlowat	   = mptcp_set_rcvlowat,
4443};
4444
4445static struct proto mptcp_v6_prot;
4446
4447static struct inet_protosw mptcp_v6_protosw = {
4448	.type		= SOCK_STREAM,
4449	.protocol	= IPPROTO_MPTCP,
4450	.prot		= &mptcp_v6_prot,
4451	.ops		= &mptcp_v6_stream_ops,
4452	.flags		= INET_PROTOSW_ICSK,
4453};
4454
4455int __init mptcp_proto_v6_init(void)
4456{
4457	int err;
4458
4459	mptcp_subflow_v6_init();
4460
4461	mptcp_v6_prot = mptcp_prot;
4462	strscpy(mptcp_v6_prot.name, "MPTCPv6", sizeof(mptcp_v6_prot.name));
4463	mptcp_v6_prot.slab = NULL;
4464	mptcp_v6_prot.obj_size = sizeof(struct mptcp6_sock);
4465	mptcp_v6_prot.ipv6_pinfo_offset = offsetof(struct mptcp6_sock, np);
4466
4467	err = proto_register(&mptcp_v6_prot, 1);
4468	if (err)
4469		return err;
4470
4471	err = inet6_register_protosw(&mptcp_v6_protosw);
4472	if (err)
4473		proto_unregister(&mptcp_v6_prot);
4474
4475	return err;
4476}
4477#endif