Linux Audio

Check our new training course

Linux kernel drivers training

Mar 31-Apr 9, 2025, special US time zones
Register
Loading...
Note: File does not exist in v4.6.
   1// SPDX-License-Identifier: GPL-2.0
   2/* Copyright (c) 2017 - 2018 Covalent IO, Inc. http://covalent.io */
   3
   4#include <linux/bpf.h>
   5#include <linux/filter.h>
   6#include <linux/errno.h>
   7#include <linux/file.h>
   8#include <linux/net.h>
   9#include <linux/workqueue.h>
  10#include <linux/skmsg.h>
  11#include <linux/list.h>
  12#include <linux/jhash.h>
  13#include <linux/sock_diag.h>
  14#include <net/udp.h>
  15
  16struct bpf_stab {
  17	struct bpf_map map;
  18	struct sock **sks;
  19	struct sk_psock_progs progs;
  20	raw_spinlock_t lock;
  21};
  22
  23#define SOCK_CREATE_FLAG_MASK				\
  24	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
  25
  26static struct bpf_map *sock_map_alloc(union bpf_attr *attr)
  27{
  28	struct bpf_stab *stab;
  29	u64 cost;
  30	int err;
  31
  32	if (!capable(CAP_NET_ADMIN))
  33		return ERR_PTR(-EPERM);
  34	if (attr->max_entries == 0 ||
  35	    attr->key_size    != 4 ||
  36	    (attr->value_size != sizeof(u32) &&
  37	     attr->value_size != sizeof(u64)) ||
  38	    attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
  39		return ERR_PTR(-EINVAL);
  40
  41	stab = kzalloc(sizeof(*stab), GFP_USER);
  42	if (!stab)
  43		return ERR_PTR(-ENOMEM);
  44
  45	bpf_map_init_from_attr(&stab->map, attr);
  46	raw_spin_lock_init(&stab->lock);
  47
  48	/* Make sure page count doesn't overflow. */
  49	cost = (u64) stab->map.max_entries * sizeof(struct sock *);
  50	err = bpf_map_charge_init(&stab->map.memory, cost);
  51	if (err)
  52		goto free_stab;
  53
  54	stab->sks = bpf_map_area_alloc(stab->map.max_entries *
  55				       sizeof(struct sock *),
  56				       stab->map.numa_node);
  57	if (stab->sks)
  58		return &stab->map;
  59	err = -ENOMEM;
  60	bpf_map_charge_finish(&stab->map.memory);
  61free_stab:
  62	kfree(stab);
  63	return ERR_PTR(err);
  64}
  65
  66int sock_map_get_from_fd(const union bpf_attr *attr, struct bpf_prog *prog)
  67{
  68	u32 ufd = attr->target_fd;
  69	struct bpf_map *map;
  70	struct fd f;
  71	int ret;
  72
  73	if (attr->attach_flags || attr->replace_bpf_fd)
  74		return -EINVAL;
  75
  76	f = fdget(ufd);
  77	map = __bpf_map_get(f);
  78	if (IS_ERR(map))
  79		return PTR_ERR(map);
  80	ret = sock_map_prog_update(map, prog, NULL, attr->attach_type);
  81	fdput(f);
  82	return ret;
  83}
  84
  85int sock_map_prog_detach(const union bpf_attr *attr, enum bpf_prog_type ptype)
  86{
  87	u32 ufd = attr->target_fd;
  88	struct bpf_prog *prog;
  89	struct bpf_map *map;
  90	struct fd f;
  91	int ret;
  92
  93	if (attr->attach_flags || attr->replace_bpf_fd)
  94		return -EINVAL;
  95
  96	f = fdget(ufd);
  97	map = __bpf_map_get(f);
  98	if (IS_ERR(map))
  99		return PTR_ERR(map);
 100
 101	prog = bpf_prog_get(attr->attach_bpf_fd);
 102	if (IS_ERR(prog)) {
 103		ret = PTR_ERR(prog);
 104		goto put_map;
 105	}
 106
 107	if (prog->type != ptype) {
 108		ret = -EINVAL;
 109		goto put_prog;
 110	}
 111
 112	ret = sock_map_prog_update(map, NULL, prog, attr->attach_type);
 113put_prog:
 114	bpf_prog_put(prog);
 115put_map:
 116	fdput(f);
 117	return ret;
 118}
 119
 120static void sock_map_sk_acquire(struct sock *sk)
 121	__acquires(&sk->sk_lock.slock)
 122{
 123	lock_sock(sk);
 124	preempt_disable();
 125	rcu_read_lock();
 126}
 127
 128static void sock_map_sk_release(struct sock *sk)
 129	__releases(&sk->sk_lock.slock)
 130{
 131	rcu_read_unlock();
 132	preempt_enable();
 133	release_sock(sk);
 134}
 135
 136static void sock_map_add_link(struct sk_psock *psock,
 137			      struct sk_psock_link *link,
 138			      struct bpf_map *map, void *link_raw)
 139{
 140	link->link_raw = link_raw;
 141	link->map = map;
 142	spin_lock_bh(&psock->link_lock);
 143	list_add_tail(&link->list, &psock->link);
 144	spin_unlock_bh(&psock->link_lock);
 145}
 146
 147static void sock_map_del_link(struct sock *sk,
 148			      struct sk_psock *psock, void *link_raw)
 149{
 150	struct sk_psock_link *link, *tmp;
 151	bool strp_stop = false;
 152
 153	spin_lock_bh(&psock->link_lock);
 154	list_for_each_entry_safe(link, tmp, &psock->link, list) {
 155		if (link->link_raw == link_raw) {
 156			struct bpf_map *map = link->map;
 157			struct bpf_stab *stab = container_of(map, struct bpf_stab,
 158							     map);
 159			if (psock->parser.enabled && stab->progs.skb_parser)
 160				strp_stop = true;
 161			list_del(&link->list);
 162			sk_psock_free_link(link);
 163		}
 164	}
 165	spin_unlock_bh(&psock->link_lock);
 166	if (strp_stop) {
 167		write_lock_bh(&sk->sk_callback_lock);
 168		sk_psock_stop_strp(sk, psock);
 169		write_unlock_bh(&sk->sk_callback_lock);
 170	}
 171}
 172
 173static void sock_map_unref(struct sock *sk, void *link_raw)
 174{
 175	struct sk_psock *psock = sk_psock(sk);
 176
 177	if (likely(psock)) {
 178		sock_map_del_link(sk, psock, link_raw);
 179		sk_psock_put(sk, psock);
 180	}
 181}
 182
 183static int sock_map_init_proto(struct sock *sk, struct sk_psock *psock)
 184{
 185	struct proto *prot;
 186
 187	sock_owned_by_me(sk);
 188
 189	switch (sk->sk_type) {
 190	case SOCK_STREAM:
 191		prot = tcp_bpf_get_proto(sk, psock);
 192		break;
 193
 194	case SOCK_DGRAM:
 195		prot = udp_bpf_get_proto(sk, psock);
 196		break;
 197
 198	default:
 199		return -EINVAL;
 200	}
 201
 202	if (IS_ERR(prot))
 203		return PTR_ERR(prot);
 204
 205	sk_psock_update_proto(sk, psock, prot);
 206	return 0;
 207}
 208
 209static struct sk_psock *sock_map_psock_get_checked(struct sock *sk)
 210{
 211	struct sk_psock *psock;
 212
 213	rcu_read_lock();
 214	psock = sk_psock(sk);
 215	if (psock) {
 216		if (sk->sk_prot->close != sock_map_close) {
 217			psock = ERR_PTR(-EBUSY);
 218			goto out;
 219		}
 220
 221		if (!refcount_inc_not_zero(&psock->refcnt))
 222			psock = ERR_PTR(-EBUSY);
 223	}
 224out:
 225	rcu_read_unlock();
 226	return psock;
 227}
 228
 229static int sock_map_link(struct bpf_map *map, struct sk_psock_progs *progs,
 230			 struct sock *sk)
 231{
 232	struct bpf_prog *msg_parser, *skb_parser, *skb_verdict;
 233	struct sk_psock *psock;
 234	bool skb_progs;
 235	int ret;
 236
 237	skb_verdict = READ_ONCE(progs->skb_verdict);
 238	skb_parser = READ_ONCE(progs->skb_parser);
 239	skb_progs = skb_parser && skb_verdict;
 240	if (skb_progs) {
 241		skb_verdict = bpf_prog_inc_not_zero(skb_verdict);
 242		if (IS_ERR(skb_verdict))
 243			return PTR_ERR(skb_verdict);
 244		skb_parser = bpf_prog_inc_not_zero(skb_parser);
 245		if (IS_ERR(skb_parser)) {
 246			bpf_prog_put(skb_verdict);
 247			return PTR_ERR(skb_parser);
 248		}
 249	}
 250
 251	msg_parser = READ_ONCE(progs->msg_parser);
 252	if (msg_parser) {
 253		msg_parser = bpf_prog_inc_not_zero(msg_parser);
 254		if (IS_ERR(msg_parser)) {
 255			ret = PTR_ERR(msg_parser);
 256			goto out;
 257		}
 258	}
 259
 260	psock = sock_map_psock_get_checked(sk);
 261	if (IS_ERR(psock)) {
 262		ret = PTR_ERR(psock);
 263		goto out_progs;
 264	}
 265
 266	if (psock) {
 267		if ((msg_parser && READ_ONCE(psock->progs.msg_parser)) ||
 268		    (skb_progs  && READ_ONCE(psock->progs.skb_parser))) {
 269			sk_psock_put(sk, psock);
 270			ret = -EBUSY;
 271			goto out_progs;
 272		}
 273	} else {
 274		psock = sk_psock_init(sk, map->numa_node);
 275		if (!psock) {
 276			ret = -ENOMEM;
 277			goto out_progs;
 278		}
 279	}
 280
 281	if (msg_parser)
 282		psock_set_prog(&psock->progs.msg_parser, msg_parser);
 283
 284	ret = sock_map_init_proto(sk, psock);
 285	if (ret < 0)
 286		goto out_drop;
 287
 288	write_lock_bh(&sk->sk_callback_lock);
 289	if (skb_progs && !psock->parser.enabled) {
 290		ret = sk_psock_init_strp(sk, psock);
 291		if (ret) {
 292			write_unlock_bh(&sk->sk_callback_lock);
 293			goto out_drop;
 294		}
 295		psock_set_prog(&psock->progs.skb_verdict, skb_verdict);
 296		psock_set_prog(&psock->progs.skb_parser, skb_parser);
 297		sk_psock_start_strp(sk, psock);
 298	}
 299	write_unlock_bh(&sk->sk_callback_lock);
 300	return 0;
 301out_drop:
 302	sk_psock_put(sk, psock);
 303out_progs:
 304	if (msg_parser)
 305		bpf_prog_put(msg_parser);
 306out:
 307	if (skb_progs) {
 308		bpf_prog_put(skb_verdict);
 309		bpf_prog_put(skb_parser);
 310	}
 311	return ret;
 312}
 313
 314static int sock_map_link_no_progs(struct bpf_map *map, struct sock *sk)
 315{
 316	struct sk_psock *psock;
 317	int ret;
 318
 319	psock = sock_map_psock_get_checked(sk);
 320	if (IS_ERR(psock))
 321		return PTR_ERR(psock);
 322
 323	if (!psock) {
 324		psock = sk_psock_init(sk, map->numa_node);
 325		if (!psock)
 326			return -ENOMEM;
 327	}
 328
 329	ret = sock_map_init_proto(sk, psock);
 330	if (ret < 0)
 331		sk_psock_put(sk, psock);
 332	return ret;
 333}
 334
 335static void sock_map_free(struct bpf_map *map)
 336{
 337	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 338	int i;
 339
 340	/* After the sync no updates or deletes will be in-flight so it
 341	 * is safe to walk map and remove entries without risking a race
 342	 * in EEXIST update case.
 343	 */
 344	synchronize_rcu();
 345	for (i = 0; i < stab->map.max_entries; i++) {
 346		struct sock **psk = &stab->sks[i];
 347		struct sock *sk;
 348
 349		sk = xchg(psk, NULL);
 350		if (sk) {
 351			lock_sock(sk);
 352			rcu_read_lock();
 353			sock_map_unref(sk, psk);
 354			rcu_read_unlock();
 355			release_sock(sk);
 356		}
 357	}
 358
 359	/* wait for psock readers accessing its map link */
 360	synchronize_rcu();
 361
 362	bpf_map_area_free(stab->sks);
 363	kfree(stab);
 364}
 365
 366static void sock_map_release_progs(struct bpf_map *map)
 367{
 368	psock_progs_drop(&container_of(map, struct bpf_stab, map)->progs);
 369}
 370
 371static struct sock *__sock_map_lookup_elem(struct bpf_map *map, u32 key)
 372{
 373	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 374
 375	WARN_ON_ONCE(!rcu_read_lock_held());
 376
 377	if (unlikely(key >= map->max_entries))
 378		return NULL;
 379	return READ_ONCE(stab->sks[key]);
 380}
 381
 382static void *sock_map_lookup(struct bpf_map *map, void *key)
 383{
 384	struct sock *sk;
 385
 386	sk = __sock_map_lookup_elem(map, *(u32 *)key);
 387	if (!sk || !sk_fullsock(sk))
 388		return NULL;
 389	if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
 390		return NULL;
 391	return sk;
 392}
 393
 394static void *sock_map_lookup_sys(struct bpf_map *map, void *key)
 395{
 396	struct sock *sk;
 397
 398	if (map->value_size != sizeof(u64))
 399		return ERR_PTR(-ENOSPC);
 400
 401	sk = __sock_map_lookup_elem(map, *(u32 *)key);
 402	if (!sk)
 403		return ERR_PTR(-ENOENT);
 404
 405	sock_gen_cookie(sk);
 406	return &sk->sk_cookie;
 407}
 408
 409static int __sock_map_delete(struct bpf_stab *stab, struct sock *sk_test,
 410			     struct sock **psk)
 411{
 412	struct sock *sk;
 413	int err = 0;
 414
 415	raw_spin_lock_bh(&stab->lock);
 416	sk = *psk;
 417	if (!sk_test || sk_test == sk)
 418		sk = xchg(psk, NULL);
 419
 420	if (likely(sk))
 421		sock_map_unref(sk, psk);
 422	else
 423		err = -EINVAL;
 424
 425	raw_spin_unlock_bh(&stab->lock);
 426	return err;
 427}
 428
 429static void sock_map_delete_from_link(struct bpf_map *map, struct sock *sk,
 430				      void *link_raw)
 431{
 432	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 433
 434	__sock_map_delete(stab, sk, link_raw);
 435}
 436
 437static int sock_map_delete_elem(struct bpf_map *map, void *key)
 438{
 439	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 440	u32 i = *(u32 *)key;
 441	struct sock **psk;
 442
 443	if (unlikely(i >= map->max_entries))
 444		return -EINVAL;
 445
 446	psk = &stab->sks[i];
 447	return __sock_map_delete(stab, NULL, psk);
 448}
 449
 450static int sock_map_get_next_key(struct bpf_map *map, void *key, void *next)
 451{
 452	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 453	u32 i = key ? *(u32 *)key : U32_MAX;
 454	u32 *key_next = next;
 455
 456	if (i == stab->map.max_entries - 1)
 457		return -ENOENT;
 458	if (i >= stab->map.max_entries)
 459		*key_next = 0;
 460	else
 461		*key_next = i + 1;
 462	return 0;
 463}
 464
 465static bool sock_map_redirect_allowed(const struct sock *sk);
 466
 467static int sock_map_update_common(struct bpf_map *map, u32 idx,
 468				  struct sock *sk, u64 flags)
 469{
 470	struct bpf_stab *stab = container_of(map, struct bpf_stab, map);
 471	struct sk_psock_link *link;
 472	struct sk_psock *psock;
 473	struct sock *osk;
 474	int ret;
 475
 476	WARN_ON_ONCE(!rcu_read_lock_held());
 477	if (unlikely(flags > BPF_EXIST))
 478		return -EINVAL;
 479	if (unlikely(idx >= map->max_entries))
 480		return -E2BIG;
 481	if (inet_csk_has_ulp(sk))
 482		return -EINVAL;
 483
 484	link = sk_psock_init_link();
 485	if (!link)
 486		return -ENOMEM;
 487
 488	/* Only sockets we can redirect into/from in BPF need to hold
 489	 * refs to parser/verdict progs and have their sk_data_ready
 490	 * and sk_write_space callbacks overridden.
 491	 */
 492	if (sock_map_redirect_allowed(sk))
 493		ret = sock_map_link(map, &stab->progs, sk);
 494	else
 495		ret = sock_map_link_no_progs(map, sk);
 496	if (ret < 0)
 497		goto out_free;
 498
 499	psock = sk_psock(sk);
 500	WARN_ON_ONCE(!psock);
 501
 502	raw_spin_lock_bh(&stab->lock);
 503	osk = stab->sks[idx];
 504	if (osk && flags == BPF_NOEXIST) {
 505		ret = -EEXIST;
 506		goto out_unlock;
 507	} else if (!osk && flags == BPF_EXIST) {
 508		ret = -ENOENT;
 509		goto out_unlock;
 510	}
 511
 512	sock_map_add_link(psock, link, map, &stab->sks[idx]);
 513	stab->sks[idx] = sk;
 514	if (osk)
 515		sock_map_unref(osk, &stab->sks[idx]);
 516	raw_spin_unlock_bh(&stab->lock);
 517	return 0;
 518out_unlock:
 519	raw_spin_unlock_bh(&stab->lock);
 520	if (psock)
 521		sk_psock_put(sk, psock);
 522out_free:
 523	sk_psock_free_link(link);
 524	return ret;
 525}
 526
 527static bool sock_map_op_okay(const struct bpf_sock_ops_kern *ops)
 528{
 529	return ops->op == BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB ||
 530	       ops->op == BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB ||
 531	       ops->op == BPF_SOCK_OPS_TCP_LISTEN_CB;
 532}
 533
 534static bool sk_is_tcp(const struct sock *sk)
 535{
 536	return sk->sk_type == SOCK_STREAM &&
 537	       sk->sk_protocol == IPPROTO_TCP;
 538}
 539
 540static bool sk_is_udp(const struct sock *sk)
 541{
 542	return sk->sk_type == SOCK_DGRAM &&
 543	       sk->sk_protocol == IPPROTO_UDP;
 544}
 545
 546static bool sock_map_redirect_allowed(const struct sock *sk)
 547{
 548	return sk_is_tcp(sk) && sk->sk_state != TCP_LISTEN;
 549}
 550
 551static bool sock_map_sk_is_suitable(const struct sock *sk)
 552{
 553	return sk_is_tcp(sk) || sk_is_udp(sk);
 554}
 555
 556static bool sock_map_sk_state_allowed(const struct sock *sk)
 557{
 558	if (sk_is_tcp(sk))
 559		return (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_LISTEN);
 560	else if (sk_is_udp(sk))
 561		return sk_hashed(sk);
 562
 563	return false;
 564}
 565
 566static int sock_map_update_elem(struct bpf_map *map, void *key,
 567				void *value, u64 flags)
 568{
 569	u32 idx = *(u32 *)key;
 570	struct socket *sock;
 571	struct sock *sk;
 572	int ret;
 573	u64 ufd;
 574
 575	if (map->value_size == sizeof(u64))
 576		ufd = *(u64 *)value;
 577	else
 578		ufd = *(u32 *)value;
 579	if (ufd > S32_MAX)
 580		return -EINVAL;
 581
 582	sock = sockfd_lookup(ufd, &ret);
 583	if (!sock)
 584		return ret;
 585	sk = sock->sk;
 586	if (!sk) {
 587		ret = -EINVAL;
 588		goto out;
 589	}
 590	if (!sock_map_sk_is_suitable(sk)) {
 591		ret = -EOPNOTSUPP;
 592		goto out;
 593	}
 594
 595	sock_map_sk_acquire(sk);
 596	if (!sock_map_sk_state_allowed(sk))
 597		ret = -EOPNOTSUPP;
 598	else
 599		ret = sock_map_update_common(map, idx, sk, flags);
 600	sock_map_sk_release(sk);
 601out:
 602	fput(sock->file);
 603	return ret;
 604}
 605
 606BPF_CALL_4(bpf_sock_map_update, struct bpf_sock_ops_kern *, sops,
 607	   struct bpf_map *, map, void *, key, u64, flags)
 608{
 609	WARN_ON_ONCE(!rcu_read_lock_held());
 610
 611	if (likely(sock_map_sk_is_suitable(sops->sk) &&
 612		   sock_map_op_okay(sops)))
 613		return sock_map_update_common(map, *(u32 *)key, sops->sk,
 614					      flags);
 615	return -EOPNOTSUPP;
 616}
 617
 618const struct bpf_func_proto bpf_sock_map_update_proto = {
 619	.func		= bpf_sock_map_update,
 620	.gpl_only	= false,
 621	.pkt_access	= true,
 622	.ret_type	= RET_INTEGER,
 623	.arg1_type	= ARG_PTR_TO_CTX,
 624	.arg2_type	= ARG_CONST_MAP_PTR,
 625	.arg3_type	= ARG_PTR_TO_MAP_KEY,
 626	.arg4_type	= ARG_ANYTHING,
 627};
 628
 629BPF_CALL_4(bpf_sk_redirect_map, struct sk_buff *, skb,
 630	   struct bpf_map *, map, u32, key, u64, flags)
 631{
 632	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
 633	struct sock *sk;
 634
 635	if (unlikely(flags & ~(BPF_F_INGRESS)))
 636		return SK_DROP;
 637
 638	sk = __sock_map_lookup_elem(map, key);
 639	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 640		return SK_DROP;
 641
 642	tcb->bpf.flags = flags;
 643	tcb->bpf.sk_redir = sk;
 644	return SK_PASS;
 645}
 646
 647const struct bpf_func_proto bpf_sk_redirect_map_proto = {
 648	.func           = bpf_sk_redirect_map,
 649	.gpl_only       = false,
 650	.ret_type       = RET_INTEGER,
 651	.arg1_type	= ARG_PTR_TO_CTX,
 652	.arg2_type      = ARG_CONST_MAP_PTR,
 653	.arg3_type      = ARG_ANYTHING,
 654	.arg4_type      = ARG_ANYTHING,
 655};
 656
 657BPF_CALL_4(bpf_msg_redirect_map, struct sk_msg *, msg,
 658	   struct bpf_map *, map, u32, key, u64, flags)
 659{
 660	struct sock *sk;
 661
 662	if (unlikely(flags & ~(BPF_F_INGRESS)))
 663		return SK_DROP;
 664
 665	sk = __sock_map_lookup_elem(map, key);
 666	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
 667		return SK_DROP;
 668
 669	msg->flags = flags;
 670	msg->sk_redir = sk;
 671	return SK_PASS;
 672}
 673
 674const struct bpf_func_proto bpf_msg_redirect_map_proto = {
 675	.func           = bpf_msg_redirect_map,
 676	.gpl_only       = false,
 677	.ret_type       = RET_INTEGER,
 678	.arg1_type	= ARG_PTR_TO_CTX,
 679	.arg2_type      = ARG_CONST_MAP_PTR,
 680	.arg3_type      = ARG_ANYTHING,
 681	.arg4_type      = ARG_ANYTHING,
 682};
 683
 684static int sock_map_btf_id;
 685const struct bpf_map_ops sock_map_ops = {
 686	.map_alloc		= sock_map_alloc,
 687	.map_free		= sock_map_free,
 688	.map_get_next_key	= sock_map_get_next_key,
 689	.map_lookup_elem_sys_only = sock_map_lookup_sys,
 690	.map_update_elem	= sock_map_update_elem,
 691	.map_delete_elem	= sock_map_delete_elem,
 692	.map_lookup_elem	= sock_map_lookup,
 693	.map_release_uref	= sock_map_release_progs,
 694	.map_check_btf		= map_check_no_btf,
 695	.map_btf_name		= "bpf_stab",
 696	.map_btf_id		= &sock_map_btf_id,
 697};
 698
 699struct bpf_shtab_elem {
 700	struct rcu_head rcu;
 701	u32 hash;
 702	struct sock *sk;
 703	struct hlist_node node;
 704	u8 key[];
 705};
 706
 707struct bpf_shtab_bucket {
 708	struct hlist_head head;
 709	raw_spinlock_t lock;
 710};
 711
 712struct bpf_shtab {
 713	struct bpf_map map;
 714	struct bpf_shtab_bucket *buckets;
 715	u32 buckets_num;
 716	u32 elem_size;
 717	struct sk_psock_progs progs;
 718	atomic_t count;
 719};
 720
 721static inline u32 sock_hash_bucket_hash(const void *key, u32 len)
 722{
 723	return jhash(key, len, 0);
 724}
 725
 726static struct bpf_shtab_bucket *sock_hash_select_bucket(struct bpf_shtab *htab,
 727							u32 hash)
 728{
 729	return &htab->buckets[hash & (htab->buckets_num - 1)];
 730}
 731
 732static struct bpf_shtab_elem *
 733sock_hash_lookup_elem_raw(struct hlist_head *head, u32 hash, void *key,
 734			  u32 key_size)
 735{
 736	struct bpf_shtab_elem *elem;
 737
 738	hlist_for_each_entry_rcu(elem, head, node) {
 739		if (elem->hash == hash &&
 740		    !memcmp(&elem->key, key, key_size))
 741			return elem;
 742	}
 743
 744	return NULL;
 745}
 746
 747static struct sock *__sock_hash_lookup_elem(struct bpf_map *map, void *key)
 748{
 749	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 750	u32 key_size = map->key_size, hash;
 751	struct bpf_shtab_bucket *bucket;
 752	struct bpf_shtab_elem *elem;
 753
 754	WARN_ON_ONCE(!rcu_read_lock_held());
 755
 756	hash = sock_hash_bucket_hash(key, key_size);
 757	bucket = sock_hash_select_bucket(htab, hash);
 758	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 759
 760	return elem ? elem->sk : NULL;
 761}
 762
 763static void sock_hash_free_elem(struct bpf_shtab *htab,
 764				struct bpf_shtab_elem *elem)
 765{
 766	atomic_dec(&htab->count);
 767	kfree_rcu(elem, rcu);
 768}
 769
 770static void sock_hash_delete_from_link(struct bpf_map *map, struct sock *sk,
 771				       void *link_raw)
 772{
 773	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 774	struct bpf_shtab_elem *elem_probe, *elem = link_raw;
 775	struct bpf_shtab_bucket *bucket;
 776
 777	WARN_ON_ONCE(!rcu_read_lock_held());
 778	bucket = sock_hash_select_bucket(htab, elem->hash);
 779
 780	/* elem may be deleted in parallel from the map, but access here
 781	 * is okay since it's going away only after RCU grace period.
 782	 * However, we need to check whether it's still present.
 783	 */
 784	raw_spin_lock_bh(&bucket->lock);
 785	elem_probe = sock_hash_lookup_elem_raw(&bucket->head, elem->hash,
 786					       elem->key, map->key_size);
 787	if (elem_probe && elem_probe == elem) {
 788		hlist_del_rcu(&elem->node);
 789		sock_map_unref(elem->sk, elem);
 790		sock_hash_free_elem(htab, elem);
 791	}
 792	raw_spin_unlock_bh(&bucket->lock);
 793}
 794
 795static int sock_hash_delete_elem(struct bpf_map *map, void *key)
 796{
 797	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 798	u32 hash, key_size = map->key_size;
 799	struct bpf_shtab_bucket *bucket;
 800	struct bpf_shtab_elem *elem;
 801	int ret = -ENOENT;
 802
 803	hash = sock_hash_bucket_hash(key, key_size);
 804	bucket = sock_hash_select_bucket(htab, hash);
 805
 806	raw_spin_lock_bh(&bucket->lock);
 807	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 808	if (elem) {
 809		hlist_del_rcu(&elem->node);
 810		sock_map_unref(elem->sk, elem);
 811		sock_hash_free_elem(htab, elem);
 812		ret = 0;
 813	}
 814	raw_spin_unlock_bh(&bucket->lock);
 815	return ret;
 816}
 817
 818static struct bpf_shtab_elem *sock_hash_alloc_elem(struct bpf_shtab *htab,
 819						   void *key, u32 key_size,
 820						   u32 hash, struct sock *sk,
 821						   struct bpf_shtab_elem *old)
 822{
 823	struct bpf_shtab_elem *new;
 824
 825	if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
 826		if (!old) {
 827			atomic_dec(&htab->count);
 828			return ERR_PTR(-E2BIG);
 829		}
 830	}
 831
 832	new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 833			   htab->map.numa_node);
 834	if (!new) {
 835		atomic_dec(&htab->count);
 836		return ERR_PTR(-ENOMEM);
 837	}
 838	memcpy(new->key, key, key_size);
 839	new->sk = sk;
 840	new->hash = hash;
 841	return new;
 842}
 843
 844static int sock_hash_update_common(struct bpf_map *map, void *key,
 845				   struct sock *sk, u64 flags)
 846{
 847	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 848	u32 key_size = map->key_size, hash;
 849	struct bpf_shtab_elem *elem, *elem_new;
 850	struct bpf_shtab_bucket *bucket;
 851	struct sk_psock_link *link;
 852	struct sk_psock *psock;
 853	int ret;
 854
 855	WARN_ON_ONCE(!rcu_read_lock_held());
 856	if (unlikely(flags > BPF_EXIST))
 857		return -EINVAL;
 858	if (inet_csk_has_ulp(sk))
 859		return -EINVAL;
 860
 861	link = sk_psock_init_link();
 862	if (!link)
 863		return -ENOMEM;
 864
 865	/* Only sockets we can redirect into/from in BPF need to hold
 866	 * refs to parser/verdict progs and have their sk_data_ready
 867	 * and sk_write_space callbacks overridden.
 868	 */
 869	if (sock_map_redirect_allowed(sk))
 870		ret = sock_map_link(map, &htab->progs, sk);
 871	else
 872		ret = sock_map_link_no_progs(map, sk);
 873	if (ret < 0)
 874		goto out_free;
 875
 876	psock = sk_psock(sk);
 877	WARN_ON_ONCE(!psock);
 878
 879	hash = sock_hash_bucket_hash(key, key_size);
 880	bucket = sock_hash_select_bucket(htab, hash);
 881
 882	raw_spin_lock_bh(&bucket->lock);
 883	elem = sock_hash_lookup_elem_raw(&bucket->head, hash, key, key_size);
 884	if (elem && flags == BPF_NOEXIST) {
 885		ret = -EEXIST;
 886		goto out_unlock;
 887	} else if (!elem && flags == BPF_EXIST) {
 888		ret = -ENOENT;
 889		goto out_unlock;
 890	}
 891
 892	elem_new = sock_hash_alloc_elem(htab, key, key_size, hash, sk, elem);
 893	if (IS_ERR(elem_new)) {
 894		ret = PTR_ERR(elem_new);
 895		goto out_unlock;
 896	}
 897
 898	sock_map_add_link(psock, link, map, elem_new);
 899	/* Add new element to the head of the list, so that
 900	 * concurrent search will find it before old elem.
 901	 */
 902	hlist_add_head_rcu(&elem_new->node, &bucket->head);
 903	if (elem) {
 904		hlist_del_rcu(&elem->node);
 905		sock_map_unref(elem->sk, elem);
 906		sock_hash_free_elem(htab, elem);
 907	}
 908	raw_spin_unlock_bh(&bucket->lock);
 909	return 0;
 910out_unlock:
 911	raw_spin_unlock_bh(&bucket->lock);
 912	sk_psock_put(sk, psock);
 913out_free:
 914	sk_psock_free_link(link);
 915	return ret;
 916}
 917
 918static int sock_hash_update_elem(struct bpf_map *map, void *key,
 919				 void *value, u64 flags)
 920{
 921	struct socket *sock;
 922	struct sock *sk;
 923	int ret;
 924	u64 ufd;
 925
 926	if (map->value_size == sizeof(u64))
 927		ufd = *(u64 *)value;
 928	else
 929		ufd = *(u32 *)value;
 930	if (ufd > S32_MAX)
 931		return -EINVAL;
 932
 933	sock = sockfd_lookup(ufd, &ret);
 934	if (!sock)
 935		return ret;
 936	sk = sock->sk;
 937	if (!sk) {
 938		ret = -EINVAL;
 939		goto out;
 940	}
 941	if (!sock_map_sk_is_suitable(sk)) {
 942		ret = -EOPNOTSUPP;
 943		goto out;
 944	}
 945
 946	sock_map_sk_acquire(sk);
 947	if (!sock_map_sk_state_allowed(sk))
 948		ret = -EOPNOTSUPP;
 949	else
 950		ret = sock_hash_update_common(map, key, sk, flags);
 951	sock_map_sk_release(sk);
 952out:
 953	fput(sock->file);
 954	return ret;
 955}
 956
 957static int sock_hash_get_next_key(struct bpf_map *map, void *key,
 958				  void *key_next)
 959{
 960	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
 961	struct bpf_shtab_elem *elem, *elem_next;
 962	u32 hash, key_size = map->key_size;
 963	struct hlist_head *head;
 964	int i = 0;
 965
 966	if (!key)
 967		goto find_first_elem;
 968	hash = sock_hash_bucket_hash(key, key_size);
 969	head = &sock_hash_select_bucket(htab, hash)->head;
 970	elem = sock_hash_lookup_elem_raw(head, hash, key, key_size);
 971	if (!elem)
 972		goto find_first_elem;
 973
 974	elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&elem->node)),
 975				     struct bpf_shtab_elem, node);
 976	if (elem_next) {
 977		memcpy(key_next, elem_next->key, key_size);
 978		return 0;
 979	}
 980
 981	i = hash & (htab->buckets_num - 1);
 982	i++;
 983find_first_elem:
 984	for (; i < htab->buckets_num; i++) {
 985		head = &sock_hash_select_bucket(htab, i)->head;
 986		elem_next = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
 987					     struct bpf_shtab_elem, node);
 988		if (elem_next) {
 989			memcpy(key_next, elem_next->key, key_size);
 990			return 0;
 991		}
 992	}
 993
 994	return -ENOENT;
 995}
 996
 997static struct bpf_map *sock_hash_alloc(union bpf_attr *attr)
 998{
 999	struct bpf_shtab *htab;
1000	int i, err;
1001	u64 cost;
1002
1003	if (!capable(CAP_NET_ADMIN))
1004		return ERR_PTR(-EPERM);
1005	if (attr->max_entries == 0 ||
1006	    attr->key_size    == 0 ||
1007	    (attr->value_size != sizeof(u32) &&
1008	     attr->value_size != sizeof(u64)) ||
1009	    attr->map_flags & ~SOCK_CREATE_FLAG_MASK)
1010		return ERR_PTR(-EINVAL);
1011	if (attr->key_size > MAX_BPF_STACK)
1012		return ERR_PTR(-E2BIG);
1013
1014	htab = kzalloc(sizeof(*htab), GFP_USER);
1015	if (!htab)
1016		return ERR_PTR(-ENOMEM);
1017
1018	bpf_map_init_from_attr(&htab->map, attr);
1019
1020	htab->buckets_num = roundup_pow_of_two(htab->map.max_entries);
1021	htab->elem_size = sizeof(struct bpf_shtab_elem) +
1022			  round_up(htab->map.key_size, 8);
1023	if (htab->buckets_num == 0 ||
1024	    htab->buckets_num > U32_MAX / sizeof(struct bpf_shtab_bucket)) {
1025		err = -EINVAL;
1026		goto free_htab;
1027	}
1028
1029	cost = (u64) htab->buckets_num * sizeof(struct bpf_shtab_bucket) +
1030	       (u64) htab->elem_size * htab->map.max_entries;
1031	if (cost >= U32_MAX - PAGE_SIZE) {
1032		err = -EINVAL;
1033		goto free_htab;
1034	}
1035	err = bpf_map_charge_init(&htab->map.memory, cost);
1036	if (err)
1037		goto free_htab;
1038
1039	htab->buckets = bpf_map_area_alloc(htab->buckets_num *
1040					   sizeof(struct bpf_shtab_bucket),
1041					   htab->map.numa_node);
1042	if (!htab->buckets) {
1043		bpf_map_charge_finish(&htab->map.memory);
1044		err = -ENOMEM;
1045		goto free_htab;
1046	}
1047
1048	for (i = 0; i < htab->buckets_num; i++) {
1049		INIT_HLIST_HEAD(&htab->buckets[i].head);
1050		raw_spin_lock_init(&htab->buckets[i].lock);
1051	}
1052
1053	return &htab->map;
1054free_htab:
1055	kfree(htab);
1056	return ERR_PTR(err);
1057}
1058
1059static void sock_hash_free(struct bpf_map *map)
1060{
1061	struct bpf_shtab *htab = container_of(map, struct bpf_shtab, map);
1062	struct bpf_shtab_bucket *bucket;
1063	struct hlist_head unlink_list;
1064	struct bpf_shtab_elem *elem;
1065	struct hlist_node *node;
1066	int i;
1067
1068	/* After the sync no updates or deletes will be in-flight so it
1069	 * is safe to walk map and remove entries without risking a race
1070	 * in EEXIST update case.
1071	 */
1072	synchronize_rcu();
1073	for (i = 0; i < htab->buckets_num; i++) {
1074		bucket = sock_hash_select_bucket(htab, i);
1075
1076		/* We are racing with sock_hash_delete_from_link to
1077		 * enter the spin-lock critical section. Every socket on
1078		 * the list is still linked to sockhash. Since link
1079		 * exists, psock exists and holds a ref to socket. That
1080		 * lets us to grab a socket ref too.
1081		 */
1082		raw_spin_lock_bh(&bucket->lock);
1083		hlist_for_each_entry(elem, &bucket->head, node)
1084			sock_hold(elem->sk);
1085		hlist_move_list(&bucket->head, &unlink_list);
1086		raw_spin_unlock_bh(&bucket->lock);
1087
1088		/* Process removed entries out of atomic context to
1089		 * block for socket lock before deleting the psock's
1090		 * link to sockhash.
1091		 */
1092		hlist_for_each_entry_safe(elem, node, &unlink_list, node) {
1093			hlist_del(&elem->node);
1094			lock_sock(elem->sk);
1095			rcu_read_lock();
1096			sock_map_unref(elem->sk, elem);
1097			rcu_read_unlock();
1098			release_sock(elem->sk);
1099			sock_put(elem->sk);
1100			sock_hash_free_elem(htab, elem);
1101		}
1102	}
1103
1104	/* wait for psock readers accessing its map link */
1105	synchronize_rcu();
1106
1107	bpf_map_area_free(htab->buckets);
1108	kfree(htab);
1109}
1110
1111static void *sock_hash_lookup_sys(struct bpf_map *map, void *key)
1112{
1113	struct sock *sk;
1114
1115	if (map->value_size != sizeof(u64))
1116		return ERR_PTR(-ENOSPC);
1117
1118	sk = __sock_hash_lookup_elem(map, key);
1119	if (!sk)
1120		return ERR_PTR(-ENOENT);
1121
1122	sock_gen_cookie(sk);
1123	return &sk->sk_cookie;
1124}
1125
1126static void *sock_hash_lookup(struct bpf_map *map, void *key)
1127{
1128	struct sock *sk;
1129
1130	sk = __sock_hash_lookup_elem(map, key);
1131	if (!sk || !sk_fullsock(sk))
1132		return NULL;
1133	if (sk_is_refcounted(sk) && !refcount_inc_not_zero(&sk->sk_refcnt))
1134		return NULL;
1135	return sk;
1136}
1137
1138static void sock_hash_release_progs(struct bpf_map *map)
1139{
1140	psock_progs_drop(&container_of(map, struct bpf_shtab, map)->progs);
1141}
1142
1143BPF_CALL_4(bpf_sock_hash_update, struct bpf_sock_ops_kern *, sops,
1144	   struct bpf_map *, map, void *, key, u64, flags)
1145{
1146	WARN_ON_ONCE(!rcu_read_lock_held());
1147
1148	if (likely(sock_map_sk_is_suitable(sops->sk) &&
1149		   sock_map_op_okay(sops)))
1150		return sock_hash_update_common(map, key, sops->sk, flags);
1151	return -EOPNOTSUPP;
1152}
1153
1154const struct bpf_func_proto bpf_sock_hash_update_proto = {
1155	.func		= bpf_sock_hash_update,
1156	.gpl_only	= false,
1157	.pkt_access	= true,
1158	.ret_type	= RET_INTEGER,
1159	.arg1_type	= ARG_PTR_TO_CTX,
1160	.arg2_type	= ARG_CONST_MAP_PTR,
1161	.arg3_type	= ARG_PTR_TO_MAP_KEY,
1162	.arg4_type	= ARG_ANYTHING,
1163};
1164
1165BPF_CALL_4(bpf_sk_redirect_hash, struct sk_buff *, skb,
1166	   struct bpf_map *, map, void *, key, u64, flags)
1167{
1168	struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
1169	struct sock *sk;
1170
1171	if (unlikely(flags & ~(BPF_F_INGRESS)))
1172		return SK_DROP;
1173
1174	sk = __sock_hash_lookup_elem(map, key);
1175	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1176		return SK_DROP;
1177
1178	tcb->bpf.flags = flags;
1179	tcb->bpf.sk_redir = sk;
1180	return SK_PASS;
1181}
1182
1183const struct bpf_func_proto bpf_sk_redirect_hash_proto = {
1184	.func           = bpf_sk_redirect_hash,
1185	.gpl_only       = false,
1186	.ret_type       = RET_INTEGER,
1187	.arg1_type	= ARG_PTR_TO_CTX,
1188	.arg2_type      = ARG_CONST_MAP_PTR,
1189	.arg3_type      = ARG_PTR_TO_MAP_KEY,
1190	.arg4_type      = ARG_ANYTHING,
1191};
1192
1193BPF_CALL_4(bpf_msg_redirect_hash, struct sk_msg *, msg,
1194	   struct bpf_map *, map, void *, key, u64, flags)
1195{
1196	struct sock *sk;
1197
1198	if (unlikely(flags & ~(BPF_F_INGRESS)))
1199		return SK_DROP;
1200
1201	sk = __sock_hash_lookup_elem(map, key);
1202	if (unlikely(!sk || !sock_map_redirect_allowed(sk)))
1203		return SK_DROP;
1204
1205	msg->flags = flags;
1206	msg->sk_redir = sk;
1207	return SK_PASS;
1208}
1209
1210const struct bpf_func_proto bpf_msg_redirect_hash_proto = {
1211	.func           = bpf_msg_redirect_hash,
1212	.gpl_only       = false,
1213	.ret_type       = RET_INTEGER,
1214	.arg1_type	= ARG_PTR_TO_CTX,
1215	.arg2_type      = ARG_CONST_MAP_PTR,
1216	.arg3_type      = ARG_PTR_TO_MAP_KEY,
1217	.arg4_type      = ARG_ANYTHING,
1218};
1219
1220static int sock_hash_map_btf_id;
1221const struct bpf_map_ops sock_hash_ops = {
1222	.map_alloc		= sock_hash_alloc,
1223	.map_free		= sock_hash_free,
1224	.map_get_next_key	= sock_hash_get_next_key,
1225	.map_update_elem	= sock_hash_update_elem,
1226	.map_delete_elem	= sock_hash_delete_elem,
1227	.map_lookup_elem	= sock_hash_lookup,
1228	.map_lookup_elem_sys_only = sock_hash_lookup_sys,
1229	.map_release_uref	= sock_hash_release_progs,
1230	.map_check_btf		= map_check_no_btf,
1231	.map_btf_name		= "bpf_shtab",
1232	.map_btf_id		= &sock_hash_map_btf_id,
1233};
1234
1235static struct sk_psock_progs *sock_map_progs(struct bpf_map *map)
1236{
1237	switch (map->map_type) {
1238	case BPF_MAP_TYPE_SOCKMAP:
1239		return &container_of(map, struct bpf_stab, map)->progs;
1240	case BPF_MAP_TYPE_SOCKHASH:
1241		return &container_of(map, struct bpf_shtab, map)->progs;
1242	default:
1243		break;
1244	}
1245
1246	return NULL;
1247}
1248
1249int sock_map_prog_update(struct bpf_map *map, struct bpf_prog *prog,
1250			 struct bpf_prog *old, u32 which)
1251{
1252	struct sk_psock_progs *progs = sock_map_progs(map);
1253	struct bpf_prog **pprog;
1254
1255	if (!progs)
1256		return -EOPNOTSUPP;
1257
1258	switch (which) {
1259	case BPF_SK_MSG_VERDICT:
1260		pprog = &progs->msg_parser;
1261		break;
1262	case BPF_SK_SKB_STREAM_PARSER:
1263		pprog = &progs->skb_parser;
1264		break;
1265	case BPF_SK_SKB_STREAM_VERDICT:
1266		pprog = &progs->skb_verdict;
1267		break;
1268	default:
1269		return -EOPNOTSUPP;
1270	}
1271
1272	if (old)
1273		return psock_replace_prog(pprog, prog, old);
1274
1275	psock_set_prog(pprog, prog);
1276	return 0;
1277}
1278
1279static void sock_map_unlink(struct sock *sk, struct sk_psock_link *link)
1280{
1281	switch (link->map->map_type) {
1282	case BPF_MAP_TYPE_SOCKMAP:
1283		return sock_map_delete_from_link(link->map, sk,
1284						 link->link_raw);
1285	case BPF_MAP_TYPE_SOCKHASH:
1286		return sock_hash_delete_from_link(link->map, sk,
1287						  link->link_raw);
1288	default:
1289		break;
1290	}
1291}
1292
1293static void sock_map_remove_links(struct sock *sk, struct sk_psock *psock)
1294{
1295	struct sk_psock_link *link;
1296
1297	while ((link = sk_psock_link_pop(psock))) {
1298		sock_map_unlink(sk, link);
1299		sk_psock_free_link(link);
1300	}
1301}
1302
1303void sock_map_unhash(struct sock *sk)
1304{
1305	void (*saved_unhash)(struct sock *sk);
1306	struct sk_psock *psock;
1307
1308	rcu_read_lock();
1309	psock = sk_psock(sk);
1310	if (unlikely(!psock)) {
1311		rcu_read_unlock();
1312		if (sk->sk_prot->unhash)
1313			sk->sk_prot->unhash(sk);
1314		return;
1315	}
1316
1317	saved_unhash = psock->saved_unhash;
1318	sock_map_remove_links(sk, psock);
1319	rcu_read_unlock();
1320	saved_unhash(sk);
1321}
1322
1323void sock_map_close(struct sock *sk, long timeout)
1324{
1325	void (*saved_close)(struct sock *sk, long timeout);
1326	struct sk_psock *psock;
1327
1328	lock_sock(sk);
1329	rcu_read_lock();
1330	psock = sk_psock(sk);
1331	if (unlikely(!psock)) {
1332		rcu_read_unlock();
1333		release_sock(sk);
1334		return sk->sk_prot->close(sk, timeout);
1335	}
1336
1337	saved_close = psock->saved_close;
1338	sock_map_remove_links(sk, psock);
1339	rcu_read_unlock();
1340	release_sock(sk);
1341	saved_close(sk, timeout);
1342}