Linux Audio

Check our new training course

Loading...
v4.17
 
   1/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   2 * Copyright (c) 2016 Facebook
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of version 2 of the GNU General Public
   6 * License as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful, but
   9 * WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  11 * General Public License for more details.
  12 */
  13#include <linux/bpf.h>
 
  14#include <linux/jhash.h>
  15#include <linux/filter.h>
  16#include <linux/rculist_nulls.h>
 
 
  17#include "percpu_freelist.h"
  18#include "bpf_lru_list.h"
  19#include "map_in_map.h"
  20
  21#define HTAB_CREATE_FLAG_MASK						\
  22	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
  23	 BPF_F_RDONLY | BPF_F_WRONLY)
 
 
 
 
 
 
 
 
 
 
  24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  25struct bucket {
  26	struct hlist_nulls_head head;
  27	raw_spinlock_t lock;
 
 
 
  28};
  29
  30struct bpf_htab {
  31	struct bpf_map map;
  32	struct bucket *buckets;
  33	void *elems;
  34	union {
  35		struct pcpu_freelist freelist;
  36		struct bpf_lru lru;
  37	};
  38	struct htab_elem *__percpu *extra_elems;
  39	atomic_t count;	/* number of elements in this hashtable */
  40	u32 n_buckets;	/* number of hash buckets */
  41	u32 elem_size;	/* size of each element in bytes */
 
  42};
  43
  44/* each htab element is struct htab_elem + key + value */
  45struct htab_elem {
  46	union {
  47		struct hlist_nulls_node hash_node;
  48		struct {
  49			void *padding;
  50			union {
  51				struct bpf_htab *htab;
  52				struct pcpu_freelist_node fnode;
 
  53			};
  54		};
  55	};
  56	union {
  57		struct rcu_head rcu;
  58		struct bpf_lru_node lru_node;
  59	};
  60	u32 hash;
  61	char key[0] __aligned(8);
  62};
  63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  64static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
  65
  66static bool htab_is_lru(const struct bpf_htab *htab)
  67{
  68	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
  69		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  70}
  71
  72static bool htab_is_percpu(const struct bpf_htab *htab)
  73{
  74	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
  75		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
  76}
  77
  78static bool htab_is_prealloc(const struct bpf_htab *htab)
  79{
  80	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
  81}
  82
  83static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
  84				     void __percpu *pptr)
  85{
  86	*(void __percpu **)(l->key + key_size) = pptr;
  87}
  88
  89static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
  90{
  91	return *(void __percpu **)(l->key + key_size);
  92}
  93
  94static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
  95{
  96	return *(void **)(l->key + roundup(map->key_size, 8));
  97}
  98
  99static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 100{
 101	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 102}
 103
 104static void htab_free_elems(struct bpf_htab *htab)
 105{
 106	int i;
 107
 108	if (!htab_is_percpu(htab))
 109		goto free_elems;
 110
 111	for (i = 0; i < htab->map.max_entries; i++) {
 112		void __percpu *pptr;
 113
 114		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 115					 htab->map.key_size);
 116		free_percpu(pptr);
 117		cond_resched();
 118	}
 119free_elems:
 120	bpf_map_area_free(htab->elems);
 121}
 122
 
 
 
 
 
 
 
 
 
 
 
 123static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 124					  u32 hash)
 125{
 126	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 127	struct htab_elem *l;
 128
 129	if (node) {
 130		l = container_of(node, struct htab_elem, lru_node);
 131		memcpy(l->key, key, htab->map.key_size);
 132		return l;
 133	}
 134
 135	return NULL;
 136}
 137
 138static int prealloc_init(struct bpf_htab *htab)
 139{
 140	u32 num_entries = htab->map.max_entries;
 141	int err = -ENOMEM, i;
 142
 143	if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 144		num_entries += num_possible_cpus();
 145
 146	htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 147					 htab->map.numa_node);
 148	if (!htab->elems)
 149		return -ENOMEM;
 150
 151	if (!htab_is_percpu(htab))
 152		goto skip_percpu_elems;
 153
 154	for (i = 0; i < num_entries; i++) {
 155		u32 size = round_up(htab->map.value_size, 8);
 156		void __percpu *pptr;
 157
 158		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 159		if (!pptr)
 160			goto free_elems;
 161		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 162				  pptr);
 163		cond_resched();
 164	}
 165
 166skip_percpu_elems:
 167	if (htab_is_lru(htab))
 168		err = bpf_lru_init(&htab->lru,
 169				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 170				   offsetof(struct htab_elem, hash) -
 171				   offsetof(struct htab_elem, lru_node),
 172				   htab_lru_map_delete_node,
 173				   htab);
 174	else
 175		err = pcpu_freelist_init(&htab->freelist);
 176
 177	if (err)
 178		goto free_elems;
 179
 180	if (htab_is_lru(htab))
 181		bpf_lru_populate(&htab->lru, htab->elems,
 182				 offsetof(struct htab_elem, lru_node),
 183				 htab->elem_size, num_entries);
 184	else
 185		pcpu_freelist_populate(&htab->freelist,
 186				       htab->elems + offsetof(struct htab_elem, fnode),
 187				       htab->elem_size, num_entries);
 188
 189	return 0;
 190
 191free_elems:
 192	htab_free_elems(htab);
 193	return err;
 194}
 195
 196static void prealloc_destroy(struct bpf_htab *htab)
 197{
 198	htab_free_elems(htab);
 199
 200	if (htab_is_lru(htab))
 201		bpf_lru_destroy(&htab->lru);
 202	else
 203		pcpu_freelist_destroy(&htab->freelist);
 204}
 205
 206static int alloc_extra_elems(struct bpf_htab *htab)
 207{
 208	struct htab_elem *__percpu *pptr, *l_new;
 209	struct pcpu_freelist_node *l;
 210	int cpu;
 211
 212	pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 213				  GFP_USER | __GFP_NOWARN);
 214	if (!pptr)
 215		return -ENOMEM;
 216
 217	for_each_possible_cpu(cpu) {
 218		l = pcpu_freelist_pop(&htab->freelist);
 219		/* pop will succeed, since prealloc_init()
 220		 * preallocated extra num_possible_cpus elements
 221		 */
 222		l_new = container_of(l, struct htab_elem, fnode);
 223		*per_cpu_ptr(pptr, cpu) = l_new;
 224	}
 225	htab->extra_elems = pptr;
 226	return 0;
 227}
 228
 229/* Called from syscall */
 230static int htab_map_alloc_check(union bpf_attr *attr)
 231{
 232	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 233		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 234	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 235		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 236	/* percpu_lru means each cpu has its own LRU list.
 237	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 238	 * the map's value itself is percpu.  percpu_lru has
 239	 * nothing to do with the map's value.
 240	 */
 241	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 242	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 
 243	int numa_node = bpf_map_attr_numa_node(attr);
 244
 245	BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 246		     offsetof(struct htab_elem, hash_node.pprev));
 247	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 248		     offsetof(struct htab_elem, hash_node.pprev));
 249
 250	if (lru && !capable(CAP_SYS_ADMIN))
 251		/* LRU implementation is much complicated than other
 252		 * maps.  Hence, limit to CAP_SYS_ADMIN for now.
 253		 */
 254		return -EPERM;
 255
 256	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
 257		/* reserved bits should not be used */
 
 
 
 
 258		return -EINVAL;
 259
 260	if (!lru && percpu_lru)
 261		return -EINVAL;
 262
 263	if (lru && !prealloc)
 264		return -ENOTSUPP;
 265
 266	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 267		return -EINVAL;
 268
 269	/* check sanity of attributes.
 270	 * value_size == 0 may be allowed in the future to use map as a set
 271	 */
 272	if (attr->max_entries == 0 || attr->key_size == 0 ||
 273	    attr->value_size == 0)
 274		return -EINVAL;
 275
 276	if (attr->key_size > MAX_BPF_STACK)
 277		/* eBPF programs initialize keys on stack, so they cannot be
 278		 * larger than max stack size
 279		 */
 280		return -E2BIG;
 281
 282	if (attr->value_size >= KMALLOC_MAX_SIZE -
 283	    MAX_BPF_STACK - sizeof(struct htab_elem))
 284		/* if value_size is bigger, the user space won't be able to
 285		 * access the elements via bpf syscall. This check also makes
 286		 * sure that the elem_size doesn't overflow and it's
 287		 * kmalloc-able later in htab_map_update_elem()
 288		 */
 289		return -E2BIG;
 290
 291	return 0;
 292}
 293
 294static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 295{
 296	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 297		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 298	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 299		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 300	/* percpu_lru means each cpu has its own LRU list.
 301	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 302	 * the map's value itself is percpu.  percpu_lru has
 303	 * nothing to do with the map's value.
 304	 */
 305	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 306	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 307	struct bpf_htab *htab;
 308	int err, i;
 309	u64 cost;
 
 310
 311	htab = kzalloc(sizeof(*htab), GFP_USER);
 312	if (!htab)
 313		return ERR_PTR(-ENOMEM);
 314
 315	bpf_map_init_from_attr(&htab->map, attr);
 316
 317	if (percpu_lru) {
 318		/* ensure each CPU's lru list has >=1 elements.
 319		 * since we are at it, make each lru list has the same
 320		 * number of elements.
 321		 */
 322		htab->map.max_entries = roundup(attr->max_entries,
 323						num_possible_cpus());
 324		if (htab->map.max_entries < attr->max_entries)
 325			htab->map.max_entries = rounddown(attr->max_entries,
 326							  num_possible_cpus());
 327	}
 328
 329	/* hash table size must be power of 2 */
 330	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 331
 332	htab->elem_size = sizeof(struct htab_elem) +
 333			  round_up(htab->map.key_size, 8);
 334	if (percpu)
 335		htab->elem_size += sizeof(void *);
 336	else
 337		htab->elem_size += round_up(htab->map.value_size, 8);
 338
 339	err = -E2BIG;
 340	/* prevent zero size kmalloc and check for u32 overflow */
 341	if (htab->n_buckets == 0 ||
 342	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
 343		goto free_htab;
 344
 345	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 346	       (u64) htab->elem_size * htab->map.max_entries;
 347
 348	if (percpu)
 349		cost += (u64) round_up(htab->map.value_size, 8) *
 350			num_possible_cpus() * htab->map.max_entries;
 351	else
 352	       cost += (u64) htab->elem_size * num_possible_cpus();
 353
 354	if (cost >= U32_MAX - PAGE_SIZE)
 355		/* make sure page count doesn't overflow */
 356		goto free_htab;
 357
 358	htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
 359
 360	/* if map size is larger than memlock limit, reject it early */
 361	err = bpf_map_precharge_memlock(htab->map.pages);
 362	if (err)
 363		goto free_htab;
 364
 365	err = -ENOMEM;
 366	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 367					   sizeof(struct bucket),
 368					   htab->map.numa_node);
 369	if (!htab->buckets)
 370		goto free_htab;
 371
 372	for (i = 0; i < htab->n_buckets; i++) {
 373		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 374		raw_spin_lock_init(&htab->buckets[i].lock);
 375	}
 
 
 376
 377	if (prealloc) {
 378		err = prealloc_init(htab);
 379		if (err)
 380			goto free_buckets;
 381
 382		if (!percpu && !lru) {
 383			/* lru itself can remove the least used element, so
 384			 * there is no need for an extra elem during map_update.
 385			 */
 386			err = alloc_extra_elems(htab);
 387			if (err)
 388				goto free_prealloc;
 389		}
 390	}
 391
 392	return &htab->map;
 393
 394free_prealloc:
 395	prealloc_destroy(htab);
 396free_buckets:
 397	bpf_map_area_free(htab->buckets);
 
 
 398free_htab:
 399	kfree(htab);
 400	return ERR_PTR(err);
 401}
 402
 403static inline u32 htab_map_hash(const void *key, u32 key_len)
 404{
 405	return jhash(key, key_len, 0);
 406}
 407
 408static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 409{
 410	return &htab->buckets[hash & (htab->n_buckets - 1)];
 411}
 412
 413static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 414{
 415	return &__select_bucket(htab, hash)->head;
 416}
 417
 418/* this lookup function can only be called with bucket lock taken */
 419static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 420					 void *key, u32 key_size)
 421{
 422	struct hlist_nulls_node *n;
 423	struct htab_elem *l;
 424
 425	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 426		if (l->hash == hash && !memcmp(&l->key, key, key_size))
 427			return l;
 428
 429	return NULL;
 430}
 431
 432/* can be called without bucket lock. it will repeat the loop in
 433 * the unlikely event when elements moved from one bucket into another
 434 * while link list is being walked
 435 */
 436static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 437					       u32 hash, void *key,
 438					       u32 key_size, u32 n_buckets)
 439{
 440	struct hlist_nulls_node *n;
 441	struct htab_elem *l;
 442
 443again:
 444	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 445		if (l->hash == hash && !memcmp(&l->key, key, key_size))
 446			return l;
 447
 448	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 449		goto again;
 450
 451	return NULL;
 452}
 453
 454/* Called from syscall or from eBPF program directly, so
 455 * arguments have to match bpf_map_lookup_elem() exactly.
 456 * The return value is adjusted by BPF instructions
 457 * in htab_map_gen_lookup().
 458 */
 459static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 460{
 461	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 462	struct hlist_nulls_head *head;
 463	struct htab_elem *l;
 464	u32 hash, key_size;
 465
 466	/* Must be called with rcu_read_lock. */
 467	WARN_ON_ONCE(!rcu_read_lock_held());
 468
 469	key_size = map->key_size;
 470
 471	hash = htab_map_hash(key, key_size);
 472
 473	head = select_bucket(htab, hash);
 474
 475	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 476
 477	return l;
 478}
 479
 480static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 481{
 482	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 483
 484	if (l)
 485		return l->key + round_up(map->key_size, 8);
 486
 487	return NULL;
 488}
 489
 490/* inline bpf_map_lookup_elem() call.
 491 * Instead of:
 492 * bpf_prog
 493 *   bpf_map_lookup_elem
 494 *     map->ops->map_lookup_elem
 495 *       htab_map_lookup_elem
 496 *         __htab_map_lookup_elem
 497 * do:
 498 * bpf_prog
 499 *   __htab_map_lookup_elem
 500 */
 501static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 502{
 503	struct bpf_insn *insn = insn_buf;
 504	const int ret = BPF_REG_0;
 505
 506	*insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 
 
 507	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 508	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 509				offsetof(struct htab_elem, key) +
 510				round_up(map->key_size, 8));
 511	return insn - insn_buf;
 512}
 513
 514static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 
 515{
 516	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 517
 518	if (l) {
 519		bpf_lru_node_set_ref(&l->lru_node);
 
 520		return l->key + round_up(map->key_size, 8);
 521	}
 522
 523	return NULL;
 524}
 525
 
 
 
 
 
 
 
 
 
 
 526static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 527				   struct bpf_insn *insn_buf)
 528{
 529	struct bpf_insn *insn = insn_buf;
 530	const int ret = BPF_REG_0;
 531	const int ref_reg = BPF_REG_1;
 532
 533	*insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 
 
 534	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 535	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 536			      offsetof(struct htab_elem, lru_node) +
 537			      offsetof(struct bpf_lru_node, ref));
 538	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 539	*insn++ = BPF_ST_MEM(BPF_B, ret,
 540			     offsetof(struct htab_elem, lru_node) +
 541			     offsetof(struct bpf_lru_node, ref),
 542			     1);
 543	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 544				offsetof(struct htab_elem, key) +
 545				round_up(map->key_size, 8));
 546	return insn - insn_buf;
 547}
 548
 549/* It is called from the bpf_lru_list when the LRU needs to delete
 550 * older elements from the htab.
 551 */
 552static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 553{
 554	struct bpf_htab *htab = (struct bpf_htab *)arg;
 555	struct htab_elem *l = NULL, *tgt_l;
 556	struct hlist_nulls_head *head;
 557	struct hlist_nulls_node *n;
 558	unsigned long flags;
 559	struct bucket *b;
 560
 561	tgt_l = container_of(node, struct htab_elem, lru_node);
 562	b = __select_bucket(htab, tgt_l->hash);
 563	head = &b->head;
 564
 565	raw_spin_lock_irqsave(&b->lock, flags);
 566
 567	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 568		if (l == tgt_l) {
 569			hlist_nulls_del_rcu(&l->hash_node);
 570			break;
 571		}
 572
 573	raw_spin_unlock_irqrestore(&b->lock, flags);
 574
 575	return l == tgt_l;
 576}
 577
 578/* Called from syscall */
 579static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 580{
 581	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 582	struct hlist_nulls_head *head;
 583	struct htab_elem *l, *next_l;
 584	u32 hash, key_size;
 585	int i = 0;
 586
 587	WARN_ON_ONCE(!rcu_read_lock_held());
 588
 589	key_size = map->key_size;
 590
 591	if (!key)
 592		goto find_first_elem;
 593
 594	hash = htab_map_hash(key, key_size);
 595
 596	head = select_bucket(htab, hash);
 597
 598	/* lookup the key */
 599	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 600
 601	if (!l)
 602		goto find_first_elem;
 603
 604	/* key was found, get next key in the same bucket */
 605	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 606				  struct htab_elem, hash_node);
 607
 608	if (next_l) {
 609		/* if next elem in this hash list is non-zero, just return it */
 610		memcpy(next_key, next_l->key, key_size);
 611		return 0;
 612	}
 613
 614	/* no more elements in this hash list, go to the next bucket */
 615	i = hash & (htab->n_buckets - 1);
 616	i++;
 617
 618find_first_elem:
 619	/* iterate over buckets */
 620	for (; i < htab->n_buckets; i++) {
 621		head = select_bucket(htab, i);
 622
 623		/* pick first element in the bucket */
 624		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 625					  struct htab_elem, hash_node);
 626		if (next_l) {
 627			/* if it's not empty, just return it */
 628			memcpy(next_key, next_l->key, key_size);
 629			return 0;
 630		}
 631	}
 632
 633	/* iterated over all buckets and all elements */
 634	return -ENOENT;
 635}
 636
 637static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 638{
 639	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 640		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 641	kfree(l);
 642}
 643
 644static void htab_elem_free_rcu(struct rcu_head *head)
 645{
 646	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 647	struct bpf_htab *htab = l->htab;
 648
 649	/* must increment bpf_prog_active to avoid kprobe+bpf triggering while
 650	 * we're calling kfree, otherwise deadlock is possible if kprobes
 651	 * are placed somewhere inside of slub
 652	 */
 653	preempt_disable();
 654	__this_cpu_inc(bpf_prog_active);
 655	htab_elem_free(htab, l);
 656	__this_cpu_dec(bpf_prog_active);
 657	preempt_enable();
 658}
 659
 660static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 661{
 662	struct bpf_map *map = &htab->map;
 
 663
 664	if (map->ops->map_fd_put_ptr) {
 665		void *ptr = fd_htab_map_get_ptr(map, l);
 666
 667		map->ops->map_fd_put_ptr(ptr);
 668	}
 
 
 
 
 
 669
 670	if (htab_is_prealloc(htab)) {
 671		pcpu_freelist_push(&htab->freelist, &l->fnode);
 672	} else {
 673		atomic_dec(&htab->count);
 674		l->htab = htab;
 675		call_rcu(&l->rcu, htab_elem_free_rcu);
 676	}
 677}
 678
 679static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 680			    void *value, bool onallcpus)
 681{
 682	if (!onallcpus) {
 683		/* copy true value_size bytes */
 684		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 685	} else {
 686		u32 size = round_up(htab->map.value_size, 8);
 687		int off = 0, cpu;
 688
 689		for_each_possible_cpu(cpu) {
 690			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 691					value + off, size);
 692			off += size;
 693		}
 694	}
 695}
 696
 697static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 698{
 699	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 700	       BITS_PER_LONG == 64;
 701}
 702
 703static u32 htab_size_value(const struct bpf_htab *htab, bool percpu)
 704{
 705	u32 size = htab->map.value_size;
 706
 707	if (percpu || fd_htab_map_needs_adjust(htab))
 708		size = round_up(size, 8);
 709	return size;
 710}
 711
 712static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 713					 void *value, u32 key_size, u32 hash,
 714					 bool percpu, bool onallcpus,
 715					 struct htab_elem *old_elem)
 716{
 717	u32 size = htab_size_value(htab, percpu);
 718	bool prealloc = htab_is_prealloc(htab);
 719	struct htab_elem *l_new, **pl_new;
 720	void __percpu *pptr;
 721
 722	if (prealloc) {
 723		if (old_elem) {
 724			/* if we're updating the existing element,
 725			 * use per-cpu extra elems to avoid freelist_pop/push
 726			 */
 727			pl_new = this_cpu_ptr(htab->extra_elems);
 728			l_new = *pl_new;
 
 729			*pl_new = old_elem;
 730		} else {
 731			struct pcpu_freelist_node *l;
 732
 733			l = pcpu_freelist_pop(&htab->freelist);
 734			if (!l)
 735				return ERR_PTR(-E2BIG);
 736			l_new = container_of(l, struct htab_elem, fnode);
 737		}
 738	} else {
 739		if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 740			if (!old_elem) {
 741				/* when map is full and update() is replacing
 742				 * old element, it's ok to allocate, since
 743				 * old element will be freed immediately.
 744				 * Otherwise return an error
 745				 */
 746				atomic_dec(&htab->count);
 747				return ERR_PTR(-E2BIG);
 748			}
 749		l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 750				     htab->map.numa_node);
 751		if (!l_new)
 752			return ERR_PTR(-ENOMEM);
 
 
 
 
 753	}
 754
 755	memcpy(l_new->key, key, key_size);
 756	if (percpu) {
 
 757		if (prealloc) {
 758			pptr = htab_elem_get_ptr(l_new, key_size);
 759		} else {
 760			/* alloc_percpu zero-fills */
 761			pptr = __alloc_percpu_gfp(size, 8,
 762						  GFP_ATOMIC | __GFP_NOWARN);
 763			if (!pptr) {
 764				kfree(l_new);
 765				return ERR_PTR(-ENOMEM);
 
 766			}
 767		}
 768
 769		pcpu_copy_value(htab, pptr, value, onallcpus);
 770
 771		if (!prealloc)
 772			htab_elem_set_ptr(l_new, key_size, pptr);
 773	} else {
 
 774		memcpy(l_new->key + round_up(key_size, 8), value, size);
 
 
 
 
 775	}
 776
 777	l_new->hash = hash;
 778	return l_new;
 
 
 
 779}
 780
 781static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 782		       u64 map_flags)
 783{
 784	if (l_old && map_flags == BPF_NOEXIST)
 785		/* elem already exists */
 786		return -EEXIST;
 787
 788	if (!l_old && map_flags == BPF_EXIST)
 789		/* elem doesn't exist, cannot update it */
 790		return -ENOENT;
 791
 792	return 0;
 793}
 794
 795/* Called from syscall or from eBPF program */
 796static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 797				u64 map_flags)
 798{
 799	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 800	struct htab_elem *l_new = NULL, *l_old;
 801	struct hlist_nulls_head *head;
 802	unsigned long flags;
 803	struct bucket *b;
 804	u32 key_size, hash;
 805	int ret;
 806
 807	if (unlikely(map_flags > BPF_EXIST))
 808		/* unknown flags */
 809		return -EINVAL;
 810
 811	WARN_ON_ONCE(!rcu_read_lock_held());
 812
 813	key_size = map->key_size;
 814
 815	hash = htab_map_hash(key, key_size);
 816
 817	b = __select_bucket(htab, hash);
 818	head = &b->head;
 819
 820	/* bpf_map_update_elem() can be called in_irq() */
 821	raw_spin_lock_irqsave(&b->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 822
 823	l_old = lookup_elem_raw(head, hash, key, key_size);
 824
 825	ret = check_flags(htab, l_old, map_flags);
 826	if (ret)
 827		goto err;
 828
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 829	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 830				l_old);
 831	if (IS_ERR(l_new)) {
 832		/* all pre-allocated elements are in use or memory exhausted */
 833		ret = PTR_ERR(l_new);
 834		goto err;
 835	}
 836
 837	/* add new element to the head of the list, so that
 838	 * concurrent search will find it before old elem
 839	 */
 840	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 841	if (l_old) {
 842		hlist_nulls_del_rcu(&l_old->hash_node);
 843		if (!htab_is_prealloc(htab))
 844			free_htab_elem(htab, l_old);
 845	}
 846	ret = 0;
 847err:
 848	raw_spin_unlock_irqrestore(&b->lock, flags);
 849	return ret;
 850}
 851
 852static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
 853				    u64 map_flags)
 854{
 855	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 856	struct htab_elem *l_new, *l_old = NULL;
 857	struct hlist_nulls_head *head;
 858	unsigned long flags;
 859	struct bucket *b;
 860	u32 key_size, hash;
 861	int ret;
 862
 863	if (unlikely(map_flags > BPF_EXIST))
 864		/* unknown flags */
 865		return -EINVAL;
 866
 867	WARN_ON_ONCE(!rcu_read_lock_held());
 868
 869	key_size = map->key_size;
 870
 871	hash = htab_map_hash(key, key_size);
 872
 873	b = __select_bucket(htab, hash);
 874	head = &b->head;
 875
 876	/* For LRU, we need to alloc before taking bucket's
 877	 * spinlock because getting free nodes from LRU may need
 878	 * to remove older elements from htab and this removal
 879	 * operation will need a bucket lock.
 880	 */
 881	l_new = prealloc_lru_pop(htab, key, hash);
 882	if (!l_new)
 883		return -ENOMEM;
 884	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
 885
 886	/* bpf_map_update_elem() can be called in_irq() */
 887	raw_spin_lock_irqsave(&b->lock, flags);
 888
 889	l_old = lookup_elem_raw(head, hash, key, key_size);
 890
 891	ret = check_flags(htab, l_old, map_flags);
 892	if (ret)
 893		goto err;
 894
 895	/* add new element to the head of the list, so that
 896	 * concurrent search will find it before old elem
 897	 */
 898	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 899	if (l_old) {
 900		bpf_lru_node_set_ref(&l_new->lru_node);
 901		hlist_nulls_del_rcu(&l_old->hash_node);
 902	}
 903	ret = 0;
 904
 905err:
 906	raw_spin_unlock_irqrestore(&b->lock, flags);
 907
 908	if (ret)
 909		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
 910	else if (l_old)
 911		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
 912
 913	return ret;
 914}
 915
 916static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
 917					 void *value, u64 map_flags,
 918					 bool onallcpus)
 919{
 920	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 921	struct htab_elem *l_new = NULL, *l_old;
 922	struct hlist_nulls_head *head;
 923	unsigned long flags;
 924	struct bucket *b;
 925	u32 key_size, hash;
 926	int ret;
 927
 928	if (unlikely(map_flags > BPF_EXIST))
 929		/* unknown flags */
 930		return -EINVAL;
 931
 932	WARN_ON_ONCE(!rcu_read_lock_held());
 933
 934	key_size = map->key_size;
 935
 936	hash = htab_map_hash(key, key_size);
 937
 938	b = __select_bucket(htab, hash);
 939	head = &b->head;
 940
 941	/* bpf_map_update_elem() can be called in_irq() */
 942	raw_spin_lock_irqsave(&b->lock, flags);
 943
 944	l_old = lookup_elem_raw(head, hash, key, key_size);
 945
 946	ret = check_flags(htab, l_old, map_flags);
 947	if (ret)
 948		goto err;
 949
 950	if (l_old) {
 951		/* per-cpu hash map can update value in-place */
 952		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
 953				value, onallcpus);
 954	} else {
 955		l_new = alloc_htab_elem(htab, key, value, key_size,
 956					hash, true, onallcpus, NULL);
 957		if (IS_ERR(l_new)) {
 958			ret = PTR_ERR(l_new);
 959			goto err;
 960		}
 961		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
 962	}
 963	ret = 0;
 964err:
 965	raw_spin_unlock_irqrestore(&b->lock, flags);
 966	return ret;
 967}
 968
 969static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
 970					     void *value, u64 map_flags,
 971					     bool onallcpus)
 972{
 973	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 974	struct htab_elem *l_new = NULL, *l_old;
 975	struct hlist_nulls_head *head;
 976	unsigned long flags;
 977	struct bucket *b;
 978	u32 key_size, hash;
 979	int ret;
 980
 981	if (unlikely(map_flags > BPF_EXIST))
 982		/* unknown flags */
 983		return -EINVAL;
 984
 985	WARN_ON_ONCE(!rcu_read_lock_held());
 986
 987	key_size = map->key_size;
 988
 989	hash = htab_map_hash(key, key_size);
 990
 991	b = __select_bucket(htab, hash);
 992	head = &b->head;
 993
 994	/* For LRU, we need to alloc before taking bucket's
 995	 * spinlock because LRU's elem alloc may need
 996	 * to remove older elem from htab and this removal
 997	 * operation will need a bucket lock.
 998	 */
 999	if (map_flags != BPF_EXIST) {
1000		l_new = prealloc_lru_pop(htab, key, hash);
1001		if (!l_new)
1002			return -ENOMEM;
1003	}
1004
1005	/* bpf_map_update_elem() can be called in_irq() */
1006	raw_spin_lock_irqsave(&b->lock, flags);
1007
1008	l_old = lookup_elem_raw(head, hash, key, key_size);
1009
1010	ret = check_flags(htab, l_old, map_flags);
1011	if (ret)
1012		goto err;
1013
1014	if (l_old) {
1015		bpf_lru_node_set_ref(&l_old->lru_node);
1016
1017		/* per-cpu hash map can update value in-place */
1018		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1019				value, onallcpus);
1020	} else {
1021		pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1022				value, onallcpus);
1023		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1024		l_new = NULL;
1025	}
1026	ret = 0;
1027err:
1028	raw_spin_unlock_irqrestore(&b->lock, flags);
1029	if (l_new)
1030		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1031	return ret;
1032}
1033
1034static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1035				       void *value, u64 map_flags)
1036{
1037	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1038}
1039
1040static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1041					   void *value, u64 map_flags)
1042{
1043	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1044						 false);
1045}
1046
1047/* Called from syscall or from eBPF program */
1048static int htab_map_delete_elem(struct bpf_map *map, void *key)
1049{
1050	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1051	struct hlist_nulls_head *head;
1052	struct bucket *b;
1053	struct htab_elem *l;
1054	unsigned long flags;
1055	u32 hash, key_size;
1056	int ret = -ENOENT;
1057
1058	WARN_ON_ONCE(!rcu_read_lock_held());
1059
1060	key_size = map->key_size;
1061
1062	hash = htab_map_hash(key, key_size);
1063	b = __select_bucket(htab, hash);
1064	head = &b->head;
1065
1066	raw_spin_lock_irqsave(&b->lock, flags);
1067
1068	l = lookup_elem_raw(head, hash, key, key_size);
1069
1070	if (l) {
1071		hlist_nulls_del_rcu(&l->hash_node);
1072		free_htab_elem(htab, l);
1073		ret = 0;
1074	}
1075
1076	raw_spin_unlock_irqrestore(&b->lock, flags);
1077	return ret;
1078}
1079
1080static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1081{
1082	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1083	struct hlist_nulls_head *head;
1084	struct bucket *b;
1085	struct htab_elem *l;
1086	unsigned long flags;
1087	u32 hash, key_size;
1088	int ret = -ENOENT;
1089
1090	WARN_ON_ONCE(!rcu_read_lock_held());
1091
1092	key_size = map->key_size;
1093
1094	hash = htab_map_hash(key, key_size);
1095	b = __select_bucket(htab, hash);
1096	head = &b->head;
1097
1098	raw_spin_lock_irqsave(&b->lock, flags);
1099
1100	l = lookup_elem_raw(head, hash, key, key_size);
1101
1102	if (l) {
1103		hlist_nulls_del_rcu(&l->hash_node);
1104		ret = 0;
1105	}
1106
1107	raw_spin_unlock_irqrestore(&b->lock, flags);
1108	if (l)
1109		bpf_lru_push_free(&htab->lru, &l->lru_node);
1110	return ret;
1111}
1112
1113static void delete_all_elements(struct bpf_htab *htab)
1114{
1115	int i;
1116
1117	for (i = 0; i < htab->n_buckets; i++) {
1118		struct hlist_nulls_head *head = select_bucket(htab, i);
1119		struct hlist_nulls_node *n;
1120		struct htab_elem *l;
1121
1122		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1123			hlist_nulls_del_rcu(&l->hash_node);
1124			htab_elem_free(htab, l);
1125		}
1126	}
1127}
1128
1129/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1130static void htab_map_free(struct bpf_map *map)
1131{
1132	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1133
1134	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1135	 * so the programs (can be more than one that used this map) were
1136	 * disconnected from events. Wait for outstanding critical sections in
1137	 * these programs to complete
1138	 */
1139	synchronize_rcu();
1140
1141	/* some of free_htab_elem() callbacks for elements of this map may
1142	 * not have executed. Wait for them.
1143	 */
1144	rcu_barrier();
1145	if (!htab_is_prealloc(htab))
1146		delete_all_elements(htab);
1147	else
1148		prealloc_destroy(htab);
1149
1150	free_percpu(htab->extra_elems);
1151	bpf_map_area_free(htab->buckets);
1152	kfree(htab);
1153}
1154
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1155const struct bpf_map_ops htab_map_ops = {
1156	.map_alloc_check = htab_map_alloc_check,
1157	.map_alloc = htab_map_alloc,
1158	.map_free = htab_map_free,
1159	.map_get_next_key = htab_map_get_next_key,
1160	.map_lookup_elem = htab_map_lookup_elem,
1161	.map_update_elem = htab_map_update_elem,
1162	.map_delete_elem = htab_map_delete_elem,
1163	.map_gen_lookup = htab_map_gen_lookup,
 
 
 
 
 
1164};
1165
 
1166const struct bpf_map_ops htab_lru_map_ops = {
1167	.map_alloc_check = htab_map_alloc_check,
1168	.map_alloc = htab_map_alloc,
1169	.map_free = htab_map_free,
1170	.map_get_next_key = htab_map_get_next_key,
1171	.map_lookup_elem = htab_lru_map_lookup_elem,
 
1172	.map_update_elem = htab_lru_map_update_elem,
1173	.map_delete_elem = htab_lru_map_delete_elem,
1174	.map_gen_lookup = htab_lru_map_gen_lookup,
 
 
 
 
 
1175};
1176
1177/* Called from eBPF program */
1178static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1179{
1180	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1181
1182	if (l)
1183		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1184	else
1185		return NULL;
1186}
1187
1188static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1189{
1190	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1191
1192	if (l) {
1193		bpf_lru_node_set_ref(&l->lru_node);
1194		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1195	}
1196
1197	return NULL;
1198}
1199
1200int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1201{
1202	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1203	struct htab_elem *l;
1204	void __percpu *pptr;
1205	int ret = -ENOENT;
1206	int cpu, off = 0;
1207	u32 size;
1208
1209	/* per_cpu areas are zero-filled and bpf programs can only
1210	 * access 'value_size' of them, so copying rounded areas
1211	 * will not leak any kernel data
1212	 */
1213	size = round_up(map->value_size, 8);
1214	rcu_read_lock();
1215	l = __htab_map_lookup_elem(map, key);
1216	if (!l)
1217		goto out;
1218	if (htab_is_lru(htab))
1219		bpf_lru_node_set_ref(&l->lru_node);
 
1220	pptr = htab_elem_get_ptr(l, map->key_size);
1221	for_each_possible_cpu(cpu) {
1222		bpf_long_memcpy(value + off,
1223				per_cpu_ptr(pptr, cpu), size);
1224		off += size;
1225	}
1226	ret = 0;
1227out:
1228	rcu_read_unlock();
1229	return ret;
1230}
1231
1232int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1233			   u64 map_flags)
1234{
1235	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1236	int ret;
1237
1238	rcu_read_lock();
1239	if (htab_is_lru(htab))
1240		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1241							map_flags, true);
1242	else
1243		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1244						    true);
1245	rcu_read_unlock();
1246
1247	return ret;
1248}
1249
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1250const struct bpf_map_ops htab_percpu_map_ops = {
1251	.map_alloc_check = htab_map_alloc_check,
1252	.map_alloc = htab_map_alloc,
1253	.map_free = htab_map_free,
1254	.map_get_next_key = htab_map_get_next_key,
1255	.map_lookup_elem = htab_percpu_map_lookup_elem,
1256	.map_update_elem = htab_percpu_map_update_elem,
1257	.map_delete_elem = htab_map_delete_elem,
 
 
 
 
 
1258};
1259
 
1260const struct bpf_map_ops htab_lru_percpu_map_ops = {
1261	.map_alloc_check = htab_map_alloc_check,
1262	.map_alloc = htab_map_alloc,
1263	.map_free = htab_map_free,
1264	.map_get_next_key = htab_map_get_next_key,
1265	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1266	.map_update_elem = htab_lru_percpu_map_update_elem,
1267	.map_delete_elem = htab_lru_map_delete_elem,
 
 
 
 
 
1268};
1269
1270static int fd_htab_map_alloc_check(union bpf_attr *attr)
1271{
1272	if (attr->value_size != sizeof(u32))
1273		return -EINVAL;
1274	return htab_map_alloc_check(attr);
1275}
1276
1277static void fd_htab_map_free(struct bpf_map *map)
1278{
1279	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1280	struct hlist_nulls_node *n;
1281	struct hlist_nulls_head *head;
1282	struct htab_elem *l;
1283	int i;
1284
1285	for (i = 0; i < htab->n_buckets; i++) {
1286		head = select_bucket(htab, i);
1287
1288		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1289			void *ptr = fd_htab_map_get_ptr(map, l);
1290
1291			map->ops->map_fd_put_ptr(ptr);
1292		}
1293	}
1294
1295	htab_map_free(map);
1296}
1297
1298/* only called from syscall */
1299int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1300{
1301	void **ptr;
1302	int ret = 0;
1303
1304	if (!map->ops->map_fd_sys_lookup_elem)
1305		return -ENOTSUPP;
1306
1307	rcu_read_lock();
1308	ptr = htab_map_lookup_elem(map, key);
1309	if (ptr)
1310		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1311	else
1312		ret = -ENOENT;
1313	rcu_read_unlock();
1314
1315	return ret;
1316}
1317
1318/* only called from syscall */
1319int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1320				void *key, void *value, u64 map_flags)
1321{
1322	void *ptr;
1323	int ret;
1324	u32 ufd = *(u32 *)value;
1325
1326	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1327	if (IS_ERR(ptr))
1328		return PTR_ERR(ptr);
1329
1330	ret = htab_map_update_elem(map, key, &ptr, map_flags);
1331	if (ret)
1332		map->ops->map_fd_put_ptr(ptr);
1333
1334	return ret;
1335}
1336
1337static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1338{
1339	struct bpf_map *map, *inner_map_meta;
1340
1341	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1342	if (IS_ERR(inner_map_meta))
1343		return inner_map_meta;
1344
1345	map = htab_map_alloc(attr);
1346	if (IS_ERR(map)) {
1347		bpf_map_meta_free(inner_map_meta);
1348		return map;
1349	}
1350
1351	map->inner_map_meta = inner_map_meta;
1352
1353	return map;
1354}
1355
1356static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1357{
1358	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1359
1360	if (!inner_map)
1361		return NULL;
1362
1363	return READ_ONCE(*inner_map);
1364}
1365
1366static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1367				  struct bpf_insn *insn_buf)
1368{
1369	struct bpf_insn *insn = insn_buf;
1370	const int ret = BPF_REG_0;
1371
1372	*insn++ = BPF_EMIT_CALL((u64 (*)(u64, u64, u64, u64, u64))__htab_map_lookup_elem);
 
 
1373	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1374	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1375				offsetof(struct htab_elem, key) +
1376				round_up(map->key_size, 8));
1377	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1378
1379	return insn - insn_buf;
1380}
1381
1382static void htab_of_map_free(struct bpf_map *map)
1383{
1384	bpf_map_meta_free(map->inner_map_meta);
1385	fd_htab_map_free(map);
1386}
1387
 
1388const struct bpf_map_ops htab_of_maps_map_ops = {
1389	.map_alloc_check = fd_htab_map_alloc_check,
1390	.map_alloc = htab_of_map_alloc,
1391	.map_free = htab_of_map_free,
1392	.map_get_next_key = htab_map_get_next_key,
1393	.map_lookup_elem = htab_of_map_lookup_elem,
1394	.map_delete_elem = htab_map_delete_elem,
1395	.map_fd_get_ptr = bpf_map_fd_get_ptr,
1396	.map_fd_put_ptr = bpf_map_fd_put_ptr,
1397	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1398	.map_gen_lookup = htab_of_map_gen_lookup,
 
 
 
1399};
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
   3 * Copyright (c) 2016 Facebook
 
 
 
 
 
 
 
 
 
   4 */
   5#include <linux/bpf.h>
   6#include <linux/btf.h>
   7#include <linux/jhash.h>
   8#include <linux/filter.h>
   9#include <linux/rculist_nulls.h>
  10#include <linux/random.h>
  11#include <uapi/linux/btf.h>
  12#include "percpu_freelist.h"
  13#include "bpf_lru_list.h"
  14#include "map_in_map.h"
  15
  16#define HTAB_CREATE_FLAG_MASK						\
  17	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
  18	 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
  19
  20#define BATCH_OPS(_name)			\
  21	.map_lookup_batch =			\
  22	_name##_map_lookup_batch,		\
  23	.map_lookup_and_delete_batch =		\
  24	_name##_map_lookup_and_delete_batch,	\
  25	.map_update_batch =			\
  26	generic_map_update_batch,		\
  27	.map_delete_batch =			\
  28	generic_map_delete_batch
  29
  30/*
  31 * The bucket lock has two protection scopes:
  32 *
  33 * 1) Serializing concurrent operations from BPF programs on differrent
  34 *    CPUs
  35 *
  36 * 2) Serializing concurrent operations from BPF programs and sys_bpf()
  37 *
  38 * BPF programs can execute in any context including perf, kprobes and
  39 * tracing. As there are almost no limits where perf, kprobes and tracing
  40 * can be invoked from the lock operations need to be protected against
  41 * deadlocks. Deadlocks can be caused by recursion and by an invocation in
  42 * the lock held section when functions which acquire this lock are invoked
  43 * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
  44 * variable bpf_prog_active, which prevents BPF programs attached to perf
  45 * events, kprobes and tracing to be invoked before the prior invocation
  46 * from one of these contexts completed. sys_bpf() uses the same mechanism
  47 * by pinning the task to the current CPU and incrementing the recursion
  48 * protection accross the map operation.
  49 *
  50 * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
  51 * operations like memory allocations (even with GFP_ATOMIC) from atomic
  52 * contexts. This is required because even with GFP_ATOMIC the memory
  53 * allocator calls into code pathes which acquire locks with long held lock
  54 * sections. To ensure the deterministic behaviour these locks are regular
  55 * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
  56 * true atomic contexts on an RT kernel are the low level hardware
  57 * handling, scheduling, low level interrupt handling, NMIs etc. None of
  58 * these contexts should ever do memory allocations.
  59 *
  60 * As regular device interrupt handlers and soft interrupts are forced into
  61 * thread context, the existing code which does
  62 *   spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*();
  63 * just works.
  64 *
  65 * In theory the BPF locks could be converted to regular spinlocks as well,
  66 * but the bucket locks and percpu_freelist locks can be taken from
  67 * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
  68 * atomic contexts even on RT. These mechanisms require preallocated maps,
  69 * so there is no need to invoke memory allocations within the lock held
  70 * sections.
  71 *
  72 * BPF maps which need dynamic allocation are only used from (forced)
  73 * thread context on RT and can therefore use regular spinlocks which in
  74 * turn allows to invoke memory allocations from the lock held section.
  75 *
  76 * On a non RT kernel this distinction is neither possible nor required.
  77 * spinlock maps to raw_spinlock and the extra code is optimized out by the
  78 * compiler.
  79 */
  80struct bucket {
  81	struct hlist_nulls_head head;
  82	union {
  83		raw_spinlock_t raw_lock;
  84		spinlock_t     lock;
  85	};
  86};
  87
  88struct bpf_htab {
  89	struct bpf_map map;
  90	struct bucket *buckets;
  91	void *elems;
  92	union {
  93		struct pcpu_freelist freelist;
  94		struct bpf_lru lru;
  95	};
  96	struct htab_elem *__percpu *extra_elems;
  97	atomic_t count;	/* number of elements in this hashtable */
  98	u32 n_buckets;	/* number of hash buckets */
  99	u32 elem_size;	/* size of each element in bytes */
 100	u32 hashrnd;
 101};
 102
 103/* each htab element is struct htab_elem + key + value */
 104struct htab_elem {
 105	union {
 106		struct hlist_nulls_node hash_node;
 107		struct {
 108			void *padding;
 109			union {
 110				struct bpf_htab *htab;
 111				struct pcpu_freelist_node fnode;
 112				struct htab_elem *batch_flink;
 113			};
 114		};
 115	};
 116	union {
 117		struct rcu_head rcu;
 118		struct bpf_lru_node lru_node;
 119	};
 120	u32 hash;
 121	char key[] __aligned(8);
 122};
 123
 124static inline bool htab_is_prealloc(const struct bpf_htab *htab)
 125{
 126	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
 127}
 128
 129static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
 130{
 131	return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
 132}
 133
 134static void htab_init_buckets(struct bpf_htab *htab)
 135{
 136	unsigned i;
 137
 138	for (i = 0; i < htab->n_buckets; i++) {
 139		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
 140		if (htab_use_raw_lock(htab))
 141			raw_spin_lock_init(&htab->buckets[i].raw_lock);
 142		else
 143			spin_lock_init(&htab->buckets[i].lock);
 144	}
 145}
 146
 147static inline unsigned long htab_lock_bucket(const struct bpf_htab *htab,
 148					     struct bucket *b)
 149{
 150	unsigned long flags;
 151
 152	if (htab_use_raw_lock(htab))
 153		raw_spin_lock_irqsave(&b->raw_lock, flags);
 154	else
 155		spin_lock_irqsave(&b->lock, flags);
 156	return flags;
 157}
 158
 159static inline void htab_unlock_bucket(const struct bpf_htab *htab,
 160				      struct bucket *b,
 161				      unsigned long flags)
 162{
 163	if (htab_use_raw_lock(htab))
 164		raw_spin_unlock_irqrestore(&b->raw_lock, flags);
 165	else
 166		spin_unlock_irqrestore(&b->lock, flags);
 167}
 168
 169static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
 170
 171static bool htab_is_lru(const struct bpf_htab *htab)
 172{
 173	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
 174		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 175}
 176
 177static bool htab_is_percpu(const struct bpf_htab *htab)
 178{
 179	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 180		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
 181}
 182
 
 
 
 
 
 183static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
 184				     void __percpu *pptr)
 185{
 186	*(void __percpu **)(l->key + key_size) = pptr;
 187}
 188
 189static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
 190{
 191	return *(void __percpu **)(l->key + key_size);
 192}
 193
 194static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
 195{
 196	return *(void **)(l->key + roundup(map->key_size, 8));
 197}
 198
 199static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
 200{
 201	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
 202}
 203
 204static void htab_free_elems(struct bpf_htab *htab)
 205{
 206	int i;
 207
 208	if (!htab_is_percpu(htab))
 209		goto free_elems;
 210
 211	for (i = 0; i < htab->map.max_entries; i++) {
 212		void __percpu *pptr;
 213
 214		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
 215					 htab->map.key_size);
 216		free_percpu(pptr);
 217		cond_resched();
 218	}
 219free_elems:
 220	bpf_map_area_free(htab->elems);
 221}
 222
 223/* The LRU list has a lock (lru_lock). Each htab bucket has a lock
 224 * (bucket_lock). If both locks need to be acquired together, the lock
 225 * order is always lru_lock -> bucket_lock and this only happens in
 226 * bpf_lru_list.c logic. For example, certain code path of
 227 * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
 228 * will acquire lru_lock first followed by acquiring bucket_lock.
 229 *
 230 * In hashtab.c, to avoid deadlock, lock acquisition of
 231 * bucket_lock followed by lru_lock is not allowed. In such cases,
 232 * bucket_lock needs to be released first before acquiring lru_lock.
 233 */
 234static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
 235					  u32 hash)
 236{
 237	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
 238	struct htab_elem *l;
 239
 240	if (node) {
 241		l = container_of(node, struct htab_elem, lru_node);
 242		memcpy(l->key, key, htab->map.key_size);
 243		return l;
 244	}
 245
 246	return NULL;
 247}
 248
 249static int prealloc_init(struct bpf_htab *htab)
 250{
 251	u32 num_entries = htab->map.max_entries;
 252	int err = -ENOMEM, i;
 253
 254	if (!htab_is_percpu(htab) && !htab_is_lru(htab))
 255		num_entries += num_possible_cpus();
 256
 257	htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
 258					 htab->map.numa_node);
 259	if (!htab->elems)
 260		return -ENOMEM;
 261
 262	if (!htab_is_percpu(htab))
 263		goto skip_percpu_elems;
 264
 265	for (i = 0; i < num_entries; i++) {
 266		u32 size = round_up(htab->map.value_size, 8);
 267		void __percpu *pptr;
 268
 269		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
 270		if (!pptr)
 271			goto free_elems;
 272		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
 273				  pptr);
 274		cond_resched();
 275	}
 276
 277skip_percpu_elems:
 278	if (htab_is_lru(htab))
 279		err = bpf_lru_init(&htab->lru,
 280				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
 281				   offsetof(struct htab_elem, hash) -
 282				   offsetof(struct htab_elem, lru_node),
 283				   htab_lru_map_delete_node,
 284				   htab);
 285	else
 286		err = pcpu_freelist_init(&htab->freelist);
 287
 288	if (err)
 289		goto free_elems;
 290
 291	if (htab_is_lru(htab))
 292		bpf_lru_populate(&htab->lru, htab->elems,
 293				 offsetof(struct htab_elem, lru_node),
 294				 htab->elem_size, num_entries);
 295	else
 296		pcpu_freelist_populate(&htab->freelist,
 297				       htab->elems + offsetof(struct htab_elem, fnode),
 298				       htab->elem_size, num_entries);
 299
 300	return 0;
 301
 302free_elems:
 303	htab_free_elems(htab);
 304	return err;
 305}
 306
 307static void prealloc_destroy(struct bpf_htab *htab)
 308{
 309	htab_free_elems(htab);
 310
 311	if (htab_is_lru(htab))
 312		bpf_lru_destroy(&htab->lru);
 313	else
 314		pcpu_freelist_destroy(&htab->freelist);
 315}
 316
 317static int alloc_extra_elems(struct bpf_htab *htab)
 318{
 319	struct htab_elem *__percpu *pptr, *l_new;
 320	struct pcpu_freelist_node *l;
 321	int cpu;
 322
 323	pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
 324				  GFP_USER | __GFP_NOWARN);
 325	if (!pptr)
 326		return -ENOMEM;
 327
 328	for_each_possible_cpu(cpu) {
 329		l = pcpu_freelist_pop(&htab->freelist);
 330		/* pop will succeed, since prealloc_init()
 331		 * preallocated extra num_possible_cpus elements
 332		 */
 333		l_new = container_of(l, struct htab_elem, fnode);
 334		*per_cpu_ptr(pptr, cpu) = l_new;
 335	}
 336	htab->extra_elems = pptr;
 337	return 0;
 338}
 339
 340/* Called from syscall */
 341static int htab_map_alloc_check(union bpf_attr *attr)
 342{
 343	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 344		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 345	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 346		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 347	/* percpu_lru means each cpu has its own LRU list.
 348	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 349	 * the map's value itself is percpu.  percpu_lru has
 350	 * nothing to do with the map's value.
 351	 */
 352	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 353	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 354	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
 355	int numa_node = bpf_map_attr_numa_node(attr);
 356
 357	BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
 358		     offsetof(struct htab_elem, hash_node.pprev));
 359	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
 360		     offsetof(struct htab_elem, hash_node.pprev));
 361
 362	if (lru && !bpf_capable())
 363		/* LRU implementation is much complicated than other
 364		 * maps.  Hence, limit to CAP_BPF.
 365		 */
 366		return -EPERM;
 367
 368	if (zero_seed && !capable(CAP_SYS_ADMIN))
 369		/* Guard against local DoS, and discourage production use. */
 370		return -EPERM;
 371
 372	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
 373	    !bpf_map_flags_access_ok(attr->map_flags))
 374		return -EINVAL;
 375
 376	if (!lru && percpu_lru)
 377		return -EINVAL;
 378
 379	if (lru && !prealloc)
 380		return -ENOTSUPP;
 381
 382	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
 383		return -EINVAL;
 384
 385	/* check sanity of attributes.
 386	 * value_size == 0 may be allowed in the future to use map as a set
 387	 */
 388	if (attr->max_entries == 0 || attr->key_size == 0 ||
 389	    attr->value_size == 0)
 390		return -EINVAL;
 391
 392	if (attr->key_size > MAX_BPF_STACK)
 393		/* eBPF programs initialize keys on stack, so they cannot be
 394		 * larger than max stack size
 395		 */
 396		return -E2BIG;
 397
 398	if (attr->value_size >= KMALLOC_MAX_SIZE -
 399	    MAX_BPF_STACK - sizeof(struct htab_elem))
 400		/* if value_size is bigger, the user space won't be able to
 401		 * access the elements via bpf syscall. This check also makes
 402		 * sure that the elem_size doesn't overflow and it's
 403		 * kmalloc-able later in htab_map_update_elem()
 404		 */
 405		return -E2BIG;
 406
 407	return 0;
 408}
 409
 410static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
 411{
 412	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
 413		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 414	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
 415		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
 416	/* percpu_lru means each cpu has its own LRU list.
 417	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
 418	 * the map's value itself is percpu.  percpu_lru has
 419	 * nothing to do with the map's value.
 420	 */
 421	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
 422	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
 423	struct bpf_htab *htab;
 
 424	u64 cost;
 425	int err;
 426
 427	htab = kzalloc(sizeof(*htab), GFP_USER);
 428	if (!htab)
 429		return ERR_PTR(-ENOMEM);
 430
 431	bpf_map_init_from_attr(&htab->map, attr);
 432
 433	if (percpu_lru) {
 434		/* ensure each CPU's lru list has >=1 elements.
 435		 * since we are at it, make each lru list has the same
 436		 * number of elements.
 437		 */
 438		htab->map.max_entries = roundup(attr->max_entries,
 439						num_possible_cpus());
 440		if (htab->map.max_entries < attr->max_entries)
 441			htab->map.max_entries = rounddown(attr->max_entries,
 442							  num_possible_cpus());
 443	}
 444
 445	/* hash table size must be power of 2 */
 446	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
 447
 448	htab->elem_size = sizeof(struct htab_elem) +
 449			  round_up(htab->map.key_size, 8);
 450	if (percpu)
 451		htab->elem_size += sizeof(void *);
 452	else
 453		htab->elem_size += round_up(htab->map.value_size, 8);
 454
 455	err = -E2BIG;
 456	/* prevent zero size kmalloc and check for u32 overflow */
 457	if (htab->n_buckets == 0 ||
 458	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
 459		goto free_htab;
 460
 461	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
 462	       (u64) htab->elem_size * htab->map.max_entries;
 463
 464	if (percpu)
 465		cost += (u64) round_up(htab->map.value_size, 8) *
 466			num_possible_cpus() * htab->map.max_entries;
 467	else
 468	       cost += (u64) htab->elem_size * num_possible_cpus();
 469
 470	/* if map size is larger than memlock limit, reject it */
 471	err = bpf_map_charge_init(&htab->map.memory, cost);
 
 
 
 
 
 
 472	if (err)
 473		goto free_htab;
 474
 475	err = -ENOMEM;
 476	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
 477					   sizeof(struct bucket),
 478					   htab->map.numa_node);
 479	if (!htab->buckets)
 480		goto free_charge;
 481
 482	if (htab->map.map_flags & BPF_F_ZERO_SEED)
 483		htab->hashrnd = 0;
 484	else
 485		htab->hashrnd = get_random_int();
 486
 487	htab_init_buckets(htab);
 488
 489	if (prealloc) {
 490		err = prealloc_init(htab);
 491		if (err)
 492			goto free_buckets;
 493
 494		if (!percpu && !lru) {
 495			/* lru itself can remove the least used element, so
 496			 * there is no need for an extra elem during map_update.
 497			 */
 498			err = alloc_extra_elems(htab);
 499			if (err)
 500				goto free_prealloc;
 501		}
 502	}
 503
 504	return &htab->map;
 505
 506free_prealloc:
 507	prealloc_destroy(htab);
 508free_buckets:
 509	bpf_map_area_free(htab->buckets);
 510free_charge:
 511	bpf_map_charge_finish(&htab->map.memory);
 512free_htab:
 513	kfree(htab);
 514	return ERR_PTR(err);
 515}
 516
 517static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
 518{
 519	return jhash(key, key_len, hashrnd);
 520}
 521
 522static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
 523{
 524	return &htab->buckets[hash & (htab->n_buckets - 1)];
 525}
 526
 527static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
 528{
 529	return &__select_bucket(htab, hash)->head;
 530}
 531
 532/* this lookup function can only be called with bucket lock taken */
 533static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
 534					 void *key, u32 key_size)
 535{
 536	struct hlist_nulls_node *n;
 537	struct htab_elem *l;
 538
 539	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 540		if (l->hash == hash && !memcmp(&l->key, key, key_size))
 541			return l;
 542
 543	return NULL;
 544}
 545
 546/* can be called without bucket lock. it will repeat the loop in
 547 * the unlikely event when elements moved from one bucket into another
 548 * while link list is being walked
 549 */
 550static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
 551					       u32 hash, void *key,
 552					       u32 key_size, u32 n_buckets)
 553{
 554	struct hlist_nulls_node *n;
 555	struct htab_elem *l;
 556
 557again:
 558	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 559		if (l->hash == hash && !memcmp(&l->key, key, key_size))
 560			return l;
 561
 562	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
 563		goto again;
 564
 565	return NULL;
 566}
 567
 568/* Called from syscall or from eBPF program directly, so
 569 * arguments have to match bpf_map_lookup_elem() exactly.
 570 * The return value is adjusted by BPF instructions
 571 * in htab_map_gen_lookup().
 572 */
 573static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
 574{
 575	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 576	struct hlist_nulls_head *head;
 577	struct htab_elem *l;
 578	u32 hash, key_size;
 579
 580	/* Must be called with rcu_read_lock. */
 581	WARN_ON_ONCE(!rcu_read_lock_held());
 582
 583	key_size = map->key_size;
 584
 585	hash = htab_map_hash(key, key_size, htab->hashrnd);
 586
 587	head = select_bucket(htab, hash);
 588
 589	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 590
 591	return l;
 592}
 593
 594static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
 595{
 596	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 597
 598	if (l)
 599		return l->key + round_up(map->key_size, 8);
 600
 601	return NULL;
 602}
 603
 604/* inline bpf_map_lookup_elem() call.
 605 * Instead of:
 606 * bpf_prog
 607 *   bpf_map_lookup_elem
 608 *     map->ops->map_lookup_elem
 609 *       htab_map_lookup_elem
 610 *         __htab_map_lookup_elem
 611 * do:
 612 * bpf_prog
 613 *   __htab_map_lookup_elem
 614 */
 615static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
 616{
 617	struct bpf_insn *insn = insn_buf;
 618	const int ret = BPF_REG_0;
 619
 620	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 621		     (void *(*)(struct bpf_map *map, void *key))NULL));
 622	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 623	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
 624	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 625				offsetof(struct htab_elem, key) +
 626				round_up(map->key_size, 8));
 627	return insn - insn_buf;
 628}
 629
 630static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
 631							void *key, const bool mark)
 632{
 633	struct htab_elem *l = __htab_map_lookup_elem(map, key);
 634
 635	if (l) {
 636		if (mark)
 637			bpf_lru_node_set_ref(&l->lru_node);
 638		return l->key + round_up(map->key_size, 8);
 639	}
 640
 641	return NULL;
 642}
 643
 644static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
 645{
 646	return __htab_lru_map_lookup_elem(map, key, true);
 647}
 648
 649static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
 650{
 651	return __htab_lru_map_lookup_elem(map, key, false);
 652}
 653
 654static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
 655				   struct bpf_insn *insn_buf)
 656{
 657	struct bpf_insn *insn = insn_buf;
 658	const int ret = BPF_REG_0;
 659	const int ref_reg = BPF_REG_1;
 660
 661	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
 662		     (void *(*)(struct bpf_map *map, void *key))NULL));
 663	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
 664	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
 665	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
 666			      offsetof(struct htab_elem, lru_node) +
 667			      offsetof(struct bpf_lru_node, ref));
 668	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
 669	*insn++ = BPF_ST_MEM(BPF_B, ret,
 670			     offsetof(struct htab_elem, lru_node) +
 671			     offsetof(struct bpf_lru_node, ref),
 672			     1);
 673	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
 674				offsetof(struct htab_elem, key) +
 675				round_up(map->key_size, 8));
 676	return insn - insn_buf;
 677}
 678
 679/* It is called from the bpf_lru_list when the LRU needs to delete
 680 * older elements from the htab.
 681 */
 682static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
 683{
 684	struct bpf_htab *htab = (struct bpf_htab *)arg;
 685	struct htab_elem *l = NULL, *tgt_l;
 686	struct hlist_nulls_head *head;
 687	struct hlist_nulls_node *n;
 688	unsigned long flags;
 689	struct bucket *b;
 690
 691	tgt_l = container_of(node, struct htab_elem, lru_node);
 692	b = __select_bucket(htab, tgt_l->hash);
 693	head = &b->head;
 694
 695	flags = htab_lock_bucket(htab, b);
 696
 697	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
 698		if (l == tgt_l) {
 699			hlist_nulls_del_rcu(&l->hash_node);
 700			break;
 701		}
 702
 703	htab_unlock_bucket(htab, b, flags);
 704
 705	return l == tgt_l;
 706}
 707
 708/* Called from syscall */
 709static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
 710{
 711	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 712	struct hlist_nulls_head *head;
 713	struct htab_elem *l, *next_l;
 714	u32 hash, key_size;
 715	int i = 0;
 716
 717	WARN_ON_ONCE(!rcu_read_lock_held());
 718
 719	key_size = map->key_size;
 720
 721	if (!key)
 722		goto find_first_elem;
 723
 724	hash = htab_map_hash(key, key_size, htab->hashrnd);
 725
 726	head = select_bucket(htab, hash);
 727
 728	/* lookup the key */
 729	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
 730
 731	if (!l)
 732		goto find_first_elem;
 733
 734	/* key was found, get next key in the same bucket */
 735	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
 736				  struct htab_elem, hash_node);
 737
 738	if (next_l) {
 739		/* if next elem in this hash list is non-zero, just return it */
 740		memcpy(next_key, next_l->key, key_size);
 741		return 0;
 742	}
 743
 744	/* no more elements in this hash list, go to the next bucket */
 745	i = hash & (htab->n_buckets - 1);
 746	i++;
 747
 748find_first_elem:
 749	/* iterate over buckets */
 750	for (; i < htab->n_buckets; i++) {
 751		head = select_bucket(htab, i);
 752
 753		/* pick first element in the bucket */
 754		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
 755					  struct htab_elem, hash_node);
 756		if (next_l) {
 757			/* if it's not empty, just return it */
 758			memcpy(next_key, next_l->key, key_size);
 759			return 0;
 760		}
 761	}
 762
 763	/* iterated over all buckets and all elements */
 764	return -ENOENT;
 765}
 766
 767static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
 768{
 769	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
 770		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
 771	kfree(l);
 772}
 773
 774static void htab_elem_free_rcu(struct rcu_head *head)
 775{
 776	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
 777	struct bpf_htab *htab = l->htab;
 778
 
 
 
 
 
 
 779	htab_elem_free(htab, l);
 
 
 780}
 781
 782static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
 783{
 784	struct bpf_map *map = &htab->map;
 785	void *ptr;
 786
 787	if (map->ops->map_fd_put_ptr) {
 788		ptr = fd_htab_map_get_ptr(map, l);
 
 789		map->ops->map_fd_put_ptr(ptr);
 790	}
 791}
 792
 793static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
 794{
 795	htab_put_fd_value(htab, l);
 796
 797	if (htab_is_prealloc(htab)) {
 798		__pcpu_freelist_push(&htab->freelist, &l->fnode);
 799	} else {
 800		atomic_dec(&htab->count);
 801		l->htab = htab;
 802		call_rcu(&l->rcu, htab_elem_free_rcu);
 803	}
 804}
 805
 806static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
 807			    void *value, bool onallcpus)
 808{
 809	if (!onallcpus) {
 810		/* copy true value_size bytes */
 811		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
 812	} else {
 813		u32 size = round_up(htab->map.value_size, 8);
 814		int off = 0, cpu;
 815
 816		for_each_possible_cpu(cpu) {
 817			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
 818					value + off, size);
 819			off += size;
 820		}
 821	}
 822}
 823
 824static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
 825{
 826	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
 827	       BITS_PER_LONG == 64;
 828}
 829
 
 
 
 
 
 
 
 
 
 830static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
 831					 void *value, u32 key_size, u32 hash,
 832					 bool percpu, bool onallcpus,
 833					 struct htab_elem *old_elem)
 834{
 835	u32 size = htab->map.value_size;
 836	bool prealloc = htab_is_prealloc(htab);
 837	struct htab_elem *l_new, **pl_new;
 838	void __percpu *pptr;
 839
 840	if (prealloc) {
 841		if (old_elem) {
 842			/* if we're updating the existing element,
 843			 * use per-cpu extra elems to avoid freelist_pop/push
 844			 */
 845			pl_new = this_cpu_ptr(htab->extra_elems);
 846			l_new = *pl_new;
 847			htab_put_fd_value(htab, old_elem);
 848			*pl_new = old_elem;
 849		} else {
 850			struct pcpu_freelist_node *l;
 851
 852			l = __pcpu_freelist_pop(&htab->freelist);
 853			if (!l)
 854				return ERR_PTR(-E2BIG);
 855			l_new = container_of(l, struct htab_elem, fnode);
 856		}
 857	} else {
 858		if (atomic_inc_return(&htab->count) > htab->map.max_entries)
 859			if (!old_elem) {
 860				/* when map is full and update() is replacing
 861				 * old element, it's ok to allocate, since
 862				 * old element will be freed immediately.
 863				 * Otherwise return an error
 864				 */
 865				l_new = ERR_PTR(-E2BIG);
 866				goto dec_count;
 867			}
 868		l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
 869				     htab->map.numa_node);
 870		if (!l_new) {
 871			l_new = ERR_PTR(-ENOMEM);
 872			goto dec_count;
 873		}
 874		check_and_init_map_lock(&htab->map,
 875					l_new->key + round_up(key_size, 8));
 876	}
 877
 878	memcpy(l_new->key, key, key_size);
 879	if (percpu) {
 880		size = round_up(size, 8);
 881		if (prealloc) {
 882			pptr = htab_elem_get_ptr(l_new, key_size);
 883		} else {
 884			/* alloc_percpu zero-fills */
 885			pptr = __alloc_percpu_gfp(size, 8,
 886						  GFP_ATOMIC | __GFP_NOWARN);
 887			if (!pptr) {
 888				kfree(l_new);
 889				l_new = ERR_PTR(-ENOMEM);
 890				goto dec_count;
 891			}
 892		}
 893
 894		pcpu_copy_value(htab, pptr, value, onallcpus);
 895
 896		if (!prealloc)
 897			htab_elem_set_ptr(l_new, key_size, pptr);
 898	} else if (fd_htab_map_needs_adjust(htab)) {
 899		size = round_up(size, 8);
 900		memcpy(l_new->key + round_up(key_size, 8), value, size);
 901	} else {
 902		copy_map_value(&htab->map,
 903			       l_new->key + round_up(key_size, 8),
 904			       value);
 905	}
 906
 907	l_new->hash = hash;
 908	return l_new;
 909dec_count:
 910	atomic_dec(&htab->count);
 911	return l_new;
 912}
 913
 914static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
 915		       u64 map_flags)
 916{
 917	if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
 918		/* elem already exists */
 919		return -EEXIST;
 920
 921	if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
 922		/* elem doesn't exist, cannot update it */
 923		return -ENOENT;
 924
 925	return 0;
 926}
 927
 928/* Called from syscall or from eBPF program */
 929static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
 930				u64 map_flags)
 931{
 932	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
 933	struct htab_elem *l_new = NULL, *l_old;
 934	struct hlist_nulls_head *head;
 935	unsigned long flags;
 936	struct bucket *b;
 937	u32 key_size, hash;
 938	int ret;
 939
 940	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
 941		/* unknown flags */
 942		return -EINVAL;
 943
 944	WARN_ON_ONCE(!rcu_read_lock_held());
 945
 946	key_size = map->key_size;
 947
 948	hash = htab_map_hash(key, key_size, htab->hashrnd);
 949
 950	b = __select_bucket(htab, hash);
 951	head = &b->head;
 952
 953	if (unlikely(map_flags & BPF_F_LOCK)) {
 954		if (unlikely(!map_value_has_spin_lock(map)))
 955			return -EINVAL;
 956		/* find an element without taking the bucket lock */
 957		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
 958					      htab->n_buckets);
 959		ret = check_flags(htab, l_old, map_flags);
 960		if (ret)
 961			return ret;
 962		if (l_old) {
 963			/* grab the element lock and update value in place */
 964			copy_map_value_locked(map,
 965					      l_old->key + round_up(key_size, 8),
 966					      value, false);
 967			return 0;
 968		}
 969		/* fall through, grab the bucket lock and lookup again.
 970		 * 99.9% chance that the element won't be found,
 971		 * but second lookup under lock has to be done.
 972		 */
 973	}
 974
 975	flags = htab_lock_bucket(htab, b);
 976
 977	l_old = lookup_elem_raw(head, hash, key, key_size);
 978
 979	ret = check_flags(htab, l_old, map_flags);
 980	if (ret)
 981		goto err;
 982
 983	if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
 984		/* first lookup without the bucket lock didn't find the element,
 985		 * but second lookup with the bucket lock found it.
 986		 * This case is highly unlikely, but has to be dealt with:
 987		 * grab the element lock in addition to the bucket lock
 988		 * and update element in place
 989		 */
 990		copy_map_value_locked(map,
 991				      l_old->key + round_up(key_size, 8),
 992				      value, false);
 993		ret = 0;
 994		goto err;
 995	}
 996
 997	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
 998				l_old);
 999	if (IS_ERR(l_new)) {
1000		/* all pre-allocated elements are in use or memory exhausted */
1001		ret = PTR_ERR(l_new);
1002		goto err;
1003	}
1004
1005	/* add new element to the head of the list, so that
1006	 * concurrent search will find it before old elem
1007	 */
1008	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1009	if (l_old) {
1010		hlist_nulls_del_rcu(&l_old->hash_node);
1011		if (!htab_is_prealloc(htab))
1012			free_htab_elem(htab, l_old);
1013	}
1014	ret = 0;
1015err:
1016	htab_unlock_bucket(htab, b, flags);
1017	return ret;
1018}
1019
1020static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1021				    u64 map_flags)
1022{
1023	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1024	struct htab_elem *l_new, *l_old = NULL;
1025	struct hlist_nulls_head *head;
1026	unsigned long flags;
1027	struct bucket *b;
1028	u32 key_size, hash;
1029	int ret;
1030
1031	if (unlikely(map_flags > BPF_EXIST))
1032		/* unknown flags */
1033		return -EINVAL;
1034
1035	WARN_ON_ONCE(!rcu_read_lock_held());
1036
1037	key_size = map->key_size;
1038
1039	hash = htab_map_hash(key, key_size, htab->hashrnd);
1040
1041	b = __select_bucket(htab, hash);
1042	head = &b->head;
1043
1044	/* For LRU, we need to alloc before taking bucket's
1045	 * spinlock because getting free nodes from LRU may need
1046	 * to remove older elements from htab and this removal
1047	 * operation will need a bucket lock.
1048	 */
1049	l_new = prealloc_lru_pop(htab, key, hash);
1050	if (!l_new)
1051		return -ENOMEM;
1052	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1053
1054	flags = htab_lock_bucket(htab, b);
 
1055
1056	l_old = lookup_elem_raw(head, hash, key, key_size);
1057
1058	ret = check_flags(htab, l_old, map_flags);
1059	if (ret)
1060		goto err;
1061
1062	/* add new element to the head of the list, so that
1063	 * concurrent search will find it before old elem
1064	 */
1065	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1066	if (l_old) {
1067		bpf_lru_node_set_ref(&l_new->lru_node);
1068		hlist_nulls_del_rcu(&l_old->hash_node);
1069	}
1070	ret = 0;
1071
1072err:
1073	htab_unlock_bucket(htab, b, flags);
1074
1075	if (ret)
1076		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1077	else if (l_old)
1078		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1079
1080	return ret;
1081}
1082
1083static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1084					 void *value, u64 map_flags,
1085					 bool onallcpus)
1086{
1087	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1088	struct htab_elem *l_new = NULL, *l_old;
1089	struct hlist_nulls_head *head;
1090	unsigned long flags;
1091	struct bucket *b;
1092	u32 key_size, hash;
1093	int ret;
1094
1095	if (unlikely(map_flags > BPF_EXIST))
1096		/* unknown flags */
1097		return -EINVAL;
1098
1099	WARN_ON_ONCE(!rcu_read_lock_held());
1100
1101	key_size = map->key_size;
1102
1103	hash = htab_map_hash(key, key_size, htab->hashrnd);
1104
1105	b = __select_bucket(htab, hash);
1106	head = &b->head;
1107
1108	flags = htab_lock_bucket(htab, b);
 
1109
1110	l_old = lookup_elem_raw(head, hash, key, key_size);
1111
1112	ret = check_flags(htab, l_old, map_flags);
1113	if (ret)
1114		goto err;
1115
1116	if (l_old) {
1117		/* per-cpu hash map can update value in-place */
1118		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1119				value, onallcpus);
1120	} else {
1121		l_new = alloc_htab_elem(htab, key, value, key_size,
1122					hash, true, onallcpus, NULL);
1123		if (IS_ERR(l_new)) {
1124			ret = PTR_ERR(l_new);
1125			goto err;
1126		}
1127		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1128	}
1129	ret = 0;
1130err:
1131	htab_unlock_bucket(htab, b, flags);
1132	return ret;
1133}
1134
1135static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1136					     void *value, u64 map_flags,
1137					     bool onallcpus)
1138{
1139	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1140	struct htab_elem *l_new = NULL, *l_old;
1141	struct hlist_nulls_head *head;
1142	unsigned long flags;
1143	struct bucket *b;
1144	u32 key_size, hash;
1145	int ret;
1146
1147	if (unlikely(map_flags > BPF_EXIST))
1148		/* unknown flags */
1149		return -EINVAL;
1150
1151	WARN_ON_ONCE(!rcu_read_lock_held());
1152
1153	key_size = map->key_size;
1154
1155	hash = htab_map_hash(key, key_size, htab->hashrnd);
1156
1157	b = __select_bucket(htab, hash);
1158	head = &b->head;
1159
1160	/* For LRU, we need to alloc before taking bucket's
1161	 * spinlock because LRU's elem alloc may need
1162	 * to remove older elem from htab and this removal
1163	 * operation will need a bucket lock.
1164	 */
1165	if (map_flags != BPF_EXIST) {
1166		l_new = prealloc_lru_pop(htab, key, hash);
1167		if (!l_new)
1168			return -ENOMEM;
1169	}
1170
1171	flags = htab_lock_bucket(htab, b);
 
1172
1173	l_old = lookup_elem_raw(head, hash, key, key_size);
1174
1175	ret = check_flags(htab, l_old, map_flags);
1176	if (ret)
1177		goto err;
1178
1179	if (l_old) {
1180		bpf_lru_node_set_ref(&l_old->lru_node);
1181
1182		/* per-cpu hash map can update value in-place */
1183		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1184				value, onallcpus);
1185	} else {
1186		pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1187				value, onallcpus);
1188		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1189		l_new = NULL;
1190	}
1191	ret = 0;
1192err:
1193	htab_unlock_bucket(htab, b, flags);
1194	if (l_new)
1195		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1196	return ret;
1197}
1198
1199static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1200				       void *value, u64 map_flags)
1201{
1202	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1203}
1204
1205static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1206					   void *value, u64 map_flags)
1207{
1208	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1209						 false);
1210}
1211
1212/* Called from syscall or from eBPF program */
1213static int htab_map_delete_elem(struct bpf_map *map, void *key)
1214{
1215	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1216	struct hlist_nulls_head *head;
1217	struct bucket *b;
1218	struct htab_elem *l;
1219	unsigned long flags;
1220	u32 hash, key_size;
1221	int ret = -ENOENT;
1222
1223	WARN_ON_ONCE(!rcu_read_lock_held());
1224
1225	key_size = map->key_size;
1226
1227	hash = htab_map_hash(key, key_size, htab->hashrnd);
1228	b = __select_bucket(htab, hash);
1229	head = &b->head;
1230
1231	flags = htab_lock_bucket(htab, b);
1232
1233	l = lookup_elem_raw(head, hash, key, key_size);
1234
1235	if (l) {
1236		hlist_nulls_del_rcu(&l->hash_node);
1237		free_htab_elem(htab, l);
1238		ret = 0;
1239	}
1240
1241	htab_unlock_bucket(htab, b, flags);
1242	return ret;
1243}
1244
1245static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1246{
1247	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1248	struct hlist_nulls_head *head;
1249	struct bucket *b;
1250	struct htab_elem *l;
1251	unsigned long flags;
1252	u32 hash, key_size;
1253	int ret = -ENOENT;
1254
1255	WARN_ON_ONCE(!rcu_read_lock_held());
1256
1257	key_size = map->key_size;
1258
1259	hash = htab_map_hash(key, key_size, htab->hashrnd);
1260	b = __select_bucket(htab, hash);
1261	head = &b->head;
1262
1263	flags = htab_lock_bucket(htab, b);
1264
1265	l = lookup_elem_raw(head, hash, key, key_size);
1266
1267	if (l) {
1268		hlist_nulls_del_rcu(&l->hash_node);
1269		ret = 0;
1270	}
1271
1272	htab_unlock_bucket(htab, b, flags);
1273	if (l)
1274		bpf_lru_push_free(&htab->lru, &l->lru_node);
1275	return ret;
1276}
1277
1278static void delete_all_elements(struct bpf_htab *htab)
1279{
1280	int i;
1281
1282	for (i = 0; i < htab->n_buckets; i++) {
1283		struct hlist_nulls_head *head = select_bucket(htab, i);
1284		struct hlist_nulls_node *n;
1285		struct htab_elem *l;
1286
1287		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1288			hlist_nulls_del_rcu(&l->hash_node);
1289			htab_elem_free(htab, l);
1290		}
1291	}
1292}
1293
1294/* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1295static void htab_map_free(struct bpf_map *map)
1296{
1297	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1298
1299	/* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1300	 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1301	 * There is no need to synchronize_rcu() here to protect map elements.
 
1302	 */
 
1303
1304	/* some of free_htab_elem() callbacks for elements of this map may
1305	 * not have executed. Wait for them.
1306	 */
1307	rcu_barrier();
1308	if (!htab_is_prealloc(htab))
1309		delete_all_elements(htab);
1310	else
1311		prealloc_destroy(htab);
1312
1313	free_percpu(htab->extra_elems);
1314	bpf_map_area_free(htab->buckets);
1315	kfree(htab);
1316}
1317
1318static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1319				   struct seq_file *m)
1320{
1321	void *value;
1322
1323	rcu_read_lock();
1324
1325	value = htab_map_lookup_elem(map, key);
1326	if (!value) {
1327		rcu_read_unlock();
1328		return;
1329	}
1330
1331	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1332	seq_puts(m, ": ");
1333	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1334	seq_puts(m, "\n");
1335
1336	rcu_read_unlock();
1337}
1338
1339static int
1340__htab_map_lookup_and_delete_batch(struct bpf_map *map,
1341				   const union bpf_attr *attr,
1342				   union bpf_attr __user *uattr,
1343				   bool do_delete, bool is_lru_map,
1344				   bool is_percpu)
1345{
1346	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1347	u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1348	void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1349	void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1350	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1351	void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1352	u32 batch, max_count, size, bucket_size;
1353	struct htab_elem *node_to_free = NULL;
1354	u64 elem_map_flags, map_flags;
1355	struct hlist_nulls_head *head;
1356	struct hlist_nulls_node *n;
1357	unsigned long flags = 0;
1358	bool locked = false;
1359	struct htab_elem *l;
1360	struct bucket *b;
1361	int ret = 0;
1362
1363	elem_map_flags = attr->batch.elem_flags;
1364	if ((elem_map_flags & ~BPF_F_LOCK) ||
1365	    ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1366		return -EINVAL;
1367
1368	map_flags = attr->batch.flags;
1369	if (map_flags)
1370		return -EINVAL;
1371
1372	max_count = attr->batch.count;
1373	if (!max_count)
1374		return 0;
1375
1376	if (put_user(0, &uattr->batch.count))
1377		return -EFAULT;
1378
1379	batch = 0;
1380	if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1381		return -EFAULT;
1382
1383	if (batch >= htab->n_buckets)
1384		return -ENOENT;
1385
1386	key_size = htab->map.key_size;
1387	roundup_key_size = round_up(htab->map.key_size, 8);
1388	value_size = htab->map.value_size;
1389	size = round_up(value_size, 8);
1390	if (is_percpu)
1391		value_size = size * num_possible_cpus();
1392	total = 0;
1393	/* while experimenting with hash tables with sizes ranging from 10 to
1394	 * 1000, it was observed that a bucket can have upto 5 entries.
1395	 */
1396	bucket_size = 5;
1397
1398alloc:
1399	/* We cannot do copy_from_user or copy_to_user inside
1400	 * the rcu_read_lock. Allocate enough space here.
1401	 */
1402	keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1403	values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1404	if (!keys || !values) {
1405		ret = -ENOMEM;
1406		goto after_loop;
1407	}
1408
1409again:
1410	bpf_disable_instrumentation();
1411	rcu_read_lock();
1412again_nocopy:
1413	dst_key = keys;
1414	dst_val = values;
1415	b = &htab->buckets[batch];
1416	head = &b->head;
1417	/* do not grab the lock unless need it (bucket_cnt > 0). */
1418	if (locked)
1419		flags = htab_lock_bucket(htab, b);
1420
1421	bucket_cnt = 0;
1422	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1423		bucket_cnt++;
1424
1425	if (bucket_cnt && !locked) {
1426		locked = true;
1427		goto again_nocopy;
1428	}
1429
1430	if (bucket_cnt > (max_count - total)) {
1431		if (total == 0)
1432			ret = -ENOSPC;
1433		/* Note that since bucket_cnt > 0 here, it is implicit
1434		 * that the locked was grabbed, so release it.
1435		 */
1436		htab_unlock_bucket(htab, b, flags);
1437		rcu_read_unlock();
1438		bpf_enable_instrumentation();
1439		goto after_loop;
1440	}
1441
1442	if (bucket_cnt > bucket_size) {
1443		bucket_size = bucket_cnt;
1444		/* Note that since bucket_cnt > 0 here, it is implicit
1445		 * that the locked was grabbed, so release it.
1446		 */
1447		htab_unlock_bucket(htab, b, flags);
1448		rcu_read_unlock();
1449		bpf_enable_instrumentation();
1450		kvfree(keys);
1451		kvfree(values);
1452		goto alloc;
1453	}
1454
1455	/* Next block is only safe to run if you have grabbed the lock */
1456	if (!locked)
1457		goto next_batch;
1458
1459	hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1460		memcpy(dst_key, l->key, key_size);
1461
1462		if (is_percpu) {
1463			int off = 0, cpu;
1464			void __percpu *pptr;
1465
1466			pptr = htab_elem_get_ptr(l, map->key_size);
1467			for_each_possible_cpu(cpu) {
1468				bpf_long_memcpy(dst_val + off,
1469						per_cpu_ptr(pptr, cpu), size);
1470				off += size;
1471			}
1472		} else {
1473			value = l->key + roundup_key_size;
1474			if (elem_map_flags & BPF_F_LOCK)
1475				copy_map_value_locked(map, dst_val, value,
1476						      true);
1477			else
1478				copy_map_value(map, dst_val, value);
1479			check_and_init_map_lock(map, dst_val);
1480		}
1481		if (do_delete) {
1482			hlist_nulls_del_rcu(&l->hash_node);
1483
1484			/* bpf_lru_push_free() will acquire lru_lock, which
1485			 * may cause deadlock. See comments in function
1486			 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1487			 * after releasing the bucket lock.
1488			 */
1489			if (is_lru_map) {
1490				l->batch_flink = node_to_free;
1491				node_to_free = l;
1492			} else {
1493				free_htab_elem(htab, l);
1494			}
1495		}
1496		dst_key += key_size;
1497		dst_val += value_size;
1498	}
1499
1500	htab_unlock_bucket(htab, b, flags);
1501	locked = false;
1502
1503	while (node_to_free) {
1504		l = node_to_free;
1505		node_to_free = node_to_free->batch_flink;
1506		bpf_lru_push_free(&htab->lru, &l->lru_node);
1507	}
1508
1509next_batch:
1510	/* If we are not copying data, we can go to next bucket and avoid
1511	 * unlocking the rcu.
1512	 */
1513	if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1514		batch++;
1515		goto again_nocopy;
1516	}
1517
1518	rcu_read_unlock();
1519	bpf_enable_instrumentation();
1520	if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1521	    key_size * bucket_cnt) ||
1522	    copy_to_user(uvalues + total * value_size, values,
1523	    value_size * bucket_cnt))) {
1524		ret = -EFAULT;
1525		goto after_loop;
1526	}
1527
1528	total += bucket_cnt;
1529	batch++;
1530	if (batch >= htab->n_buckets) {
1531		ret = -ENOENT;
1532		goto after_loop;
1533	}
1534	goto again;
1535
1536after_loop:
1537	if (ret == -EFAULT)
1538		goto out;
1539
1540	/* copy # of entries and next batch */
1541	ubatch = u64_to_user_ptr(attr->batch.out_batch);
1542	if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1543	    put_user(total, &uattr->batch.count))
1544		ret = -EFAULT;
1545
1546out:
1547	kvfree(keys);
1548	kvfree(values);
1549	return ret;
1550}
1551
1552static int
1553htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1554			     union bpf_attr __user *uattr)
1555{
1556	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1557						  false, true);
1558}
1559
1560static int
1561htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1562					const union bpf_attr *attr,
1563					union bpf_attr __user *uattr)
1564{
1565	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1566						  false, true);
1567}
1568
1569static int
1570htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1571		      union bpf_attr __user *uattr)
1572{
1573	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1574						  false, false);
1575}
1576
1577static int
1578htab_map_lookup_and_delete_batch(struct bpf_map *map,
1579				 const union bpf_attr *attr,
1580				 union bpf_attr __user *uattr)
1581{
1582	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1583						  false, false);
1584}
1585
1586static int
1587htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1588				 const union bpf_attr *attr,
1589				 union bpf_attr __user *uattr)
1590{
1591	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1592						  true, true);
1593}
1594
1595static int
1596htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1597					    const union bpf_attr *attr,
1598					    union bpf_attr __user *uattr)
1599{
1600	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1601						  true, true);
1602}
1603
1604static int
1605htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1606			  union bpf_attr __user *uattr)
1607{
1608	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1609						  true, false);
1610}
1611
1612static int
1613htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1614				     const union bpf_attr *attr,
1615				     union bpf_attr __user *uattr)
1616{
1617	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1618						  true, false);
1619}
1620
1621struct bpf_iter_seq_hash_map_info {
1622	struct bpf_map *map;
1623	struct bpf_htab *htab;
1624	void *percpu_value_buf; // non-zero means percpu hash
1625	u32 bucket_id;
1626	u32 skip_elems;
1627};
1628
1629static struct htab_elem *
1630bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1631			   struct htab_elem *prev_elem)
1632{
1633	const struct bpf_htab *htab = info->htab;
1634	u32 skip_elems = info->skip_elems;
1635	u32 bucket_id = info->bucket_id;
1636	struct hlist_nulls_head *head;
1637	struct hlist_nulls_node *n;
1638	struct htab_elem *elem;
1639	struct bucket *b;
1640	u32 i, count;
1641
1642	if (bucket_id >= htab->n_buckets)
1643		return NULL;
1644
1645	/* try to find next elem in the same bucket */
1646	if (prev_elem) {
1647		/* no update/deletion on this bucket, prev_elem should be still valid
1648		 * and we won't skip elements.
1649		 */
1650		n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1651		elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1652		if (elem)
1653			return elem;
1654
1655		/* not found, unlock and go to the next bucket */
1656		b = &htab->buckets[bucket_id++];
1657		rcu_read_unlock();
1658		skip_elems = 0;
1659	}
1660
1661	for (i = bucket_id; i < htab->n_buckets; i++) {
1662		b = &htab->buckets[i];
1663		rcu_read_lock();
1664
1665		count = 0;
1666		head = &b->head;
1667		hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1668			if (count >= skip_elems) {
1669				info->bucket_id = i;
1670				info->skip_elems = count;
1671				return elem;
1672			}
1673			count++;
1674		}
1675
1676		rcu_read_unlock();
1677		skip_elems = 0;
1678	}
1679
1680	info->bucket_id = i;
1681	info->skip_elems = 0;
1682	return NULL;
1683}
1684
1685static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1686{
1687	struct bpf_iter_seq_hash_map_info *info = seq->private;
1688	struct htab_elem *elem;
1689
1690	elem = bpf_hash_map_seq_find_next(info, NULL);
1691	if (!elem)
1692		return NULL;
1693
1694	if (*pos == 0)
1695		++*pos;
1696	return elem;
1697}
1698
1699static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1700{
1701	struct bpf_iter_seq_hash_map_info *info = seq->private;
1702
1703	++*pos;
1704	++info->skip_elems;
1705	return bpf_hash_map_seq_find_next(info, v);
1706}
1707
1708static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1709{
1710	struct bpf_iter_seq_hash_map_info *info = seq->private;
1711	u32 roundup_key_size, roundup_value_size;
1712	struct bpf_iter__bpf_map_elem ctx = {};
1713	struct bpf_map *map = info->map;
1714	struct bpf_iter_meta meta;
1715	int ret = 0, off = 0, cpu;
1716	struct bpf_prog *prog;
1717	void __percpu *pptr;
1718
1719	meta.seq = seq;
1720	prog = bpf_iter_get_info(&meta, elem == NULL);
1721	if (prog) {
1722		ctx.meta = &meta;
1723		ctx.map = info->map;
1724		if (elem) {
1725			roundup_key_size = round_up(map->key_size, 8);
1726			ctx.key = elem->key;
1727			if (!info->percpu_value_buf) {
1728				ctx.value = elem->key + roundup_key_size;
1729			} else {
1730				roundup_value_size = round_up(map->value_size, 8);
1731				pptr = htab_elem_get_ptr(elem, map->key_size);
1732				for_each_possible_cpu(cpu) {
1733					bpf_long_memcpy(info->percpu_value_buf + off,
1734							per_cpu_ptr(pptr, cpu),
1735							roundup_value_size);
1736					off += roundup_value_size;
1737				}
1738				ctx.value = info->percpu_value_buf;
1739			}
1740		}
1741		ret = bpf_iter_run_prog(prog, &ctx);
1742	}
1743
1744	return ret;
1745}
1746
1747static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1748{
1749	return __bpf_hash_map_seq_show(seq, v);
1750}
1751
1752static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1753{
1754	if (!v)
1755		(void)__bpf_hash_map_seq_show(seq, NULL);
1756	else
1757		rcu_read_unlock();
1758}
1759
1760static int bpf_iter_init_hash_map(void *priv_data,
1761				  struct bpf_iter_aux_info *aux)
1762{
1763	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1764	struct bpf_map *map = aux->map;
1765	void *value_buf;
1766	u32 buf_size;
1767
1768	if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1769	    map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1770		buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1771		value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1772		if (!value_buf)
1773			return -ENOMEM;
1774
1775		seq_info->percpu_value_buf = value_buf;
1776	}
1777
1778	seq_info->map = map;
1779	seq_info->htab = container_of(map, struct bpf_htab, map);
1780	return 0;
1781}
1782
1783static void bpf_iter_fini_hash_map(void *priv_data)
1784{
1785	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1786
1787	kfree(seq_info->percpu_value_buf);
1788}
1789
1790static const struct seq_operations bpf_hash_map_seq_ops = {
1791	.start	= bpf_hash_map_seq_start,
1792	.next	= bpf_hash_map_seq_next,
1793	.stop	= bpf_hash_map_seq_stop,
1794	.show	= bpf_hash_map_seq_show,
1795};
1796
1797static const struct bpf_iter_seq_info iter_seq_info = {
1798	.seq_ops		= &bpf_hash_map_seq_ops,
1799	.init_seq_private	= bpf_iter_init_hash_map,
1800	.fini_seq_private	= bpf_iter_fini_hash_map,
1801	.seq_priv_size		= sizeof(struct bpf_iter_seq_hash_map_info),
1802};
1803
1804static int htab_map_btf_id;
1805const struct bpf_map_ops htab_map_ops = {
1806	.map_alloc_check = htab_map_alloc_check,
1807	.map_alloc = htab_map_alloc,
1808	.map_free = htab_map_free,
1809	.map_get_next_key = htab_map_get_next_key,
1810	.map_lookup_elem = htab_map_lookup_elem,
1811	.map_update_elem = htab_map_update_elem,
1812	.map_delete_elem = htab_map_delete_elem,
1813	.map_gen_lookup = htab_map_gen_lookup,
1814	.map_seq_show_elem = htab_map_seq_show_elem,
1815	BATCH_OPS(htab),
1816	.map_btf_name = "bpf_htab",
1817	.map_btf_id = &htab_map_btf_id,
1818	.iter_seq_info = &iter_seq_info,
1819};
1820
1821static int htab_lru_map_btf_id;
1822const struct bpf_map_ops htab_lru_map_ops = {
1823	.map_alloc_check = htab_map_alloc_check,
1824	.map_alloc = htab_map_alloc,
1825	.map_free = htab_map_free,
1826	.map_get_next_key = htab_map_get_next_key,
1827	.map_lookup_elem = htab_lru_map_lookup_elem,
1828	.map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1829	.map_update_elem = htab_lru_map_update_elem,
1830	.map_delete_elem = htab_lru_map_delete_elem,
1831	.map_gen_lookup = htab_lru_map_gen_lookup,
1832	.map_seq_show_elem = htab_map_seq_show_elem,
1833	BATCH_OPS(htab_lru),
1834	.map_btf_name = "bpf_htab",
1835	.map_btf_id = &htab_lru_map_btf_id,
1836	.iter_seq_info = &iter_seq_info,
1837};
1838
1839/* Called from eBPF program */
1840static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1841{
1842	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1843
1844	if (l)
1845		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1846	else
1847		return NULL;
1848}
1849
1850static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1851{
1852	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1853
1854	if (l) {
1855		bpf_lru_node_set_ref(&l->lru_node);
1856		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1857	}
1858
1859	return NULL;
1860}
1861
1862int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1863{
 
1864	struct htab_elem *l;
1865	void __percpu *pptr;
1866	int ret = -ENOENT;
1867	int cpu, off = 0;
1868	u32 size;
1869
1870	/* per_cpu areas are zero-filled and bpf programs can only
1871	 * access 'value_size' of them, so copying rounded areas
1872	 * will not leak any kernel data
1873	 */
1874	size = round_up(map->value_size, 8);
1875	rcu_read_lock();
1876	l = __htab_map_lookup_elem(map, key);
1877	if (!l)
1878		goto out;
1879	/* We do not mark LRU map element here in order to not mess up
1880	 * eviction heuristics when user space does a map walk.
1881	 */
1882	pptr = htab_elem_get_ptr(l, map->key_size);
1883	for_each_possible_cpu(cpu) {
1884		bpf_long_memcpy(value + off,
1885				per_cpu_ptr(pptr, cpu), size);
1886		off += size;
1887	}
1888	ret = 0;
1889out:
1890	rcu_read_unlock();
1891	return ret;
1892}
1893
1894int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1895			   u64 map_flags)
1896{
1897	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1898	int ret;
1899
1900	rcu_read_lock();
1901	if (htab_is_lru(htab))
1902		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1903							map_flags, true);
1904	else
1905		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1906						    true);
1907	rcu_read_unlock();
1908
1909	return ret;
1910}
1911
1912static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1913					  struct seq_file *m)
1914{
1915	struct htab_elem *l;
1916	void __percpu *pptr;
1917	int cpu;
1918
1919	rcu_read_lock();
1920
1921	l = __htab_map_lookup_elem(map, key);
1922	if (!l) {
1923		rcu_read_unlock();
1924		return;
1925	}
1926
1927	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1928	seq_puts(m, ": {\n");
1929	pptr = htab_elem_get_ptr(l, map->key_size);
1930	for_each_possible_cpu(cpu) {
1931		seq_printf(m, "\tcpu%d: ", cpu);
1932		btf_type_seq_show(map->btf, map->btf_value_type_id,
1933				  per_cpu_ptr(pptr, cpu), m);
1934		seq_puts(m, "\n");
1935	}
1936	seq_puts(m, "}\n");
1937
1938	rcu_read_unlock();
1939}
1940
1941static int htab_percpu_map_btf_id;
1942const struct bpf_map_ops htab_percpu_map_ops = {
1943	.map_alloc_check = htab_map_alloc_check,
1944	.map_alloc = htab_map_alloc,
1945	.map_free = htab_map_free,
1946	.map_get_next_key = htab_map_get_next_key,
1947	.map_lookup_elem = htab_percpu_map_lookup_elem,
1948	.map_update_elem = htab_percpu_map_update_elem,
1949	.map_delete_elem = htab_map_delete_elem,
1950	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1951	BATCH_OPS(htab_percpu),
1952	.map_btf_name = "bpf_htab",
1953	.map_btf_id = &htab_percpu_map_btf_id,
1954	.iter_seq_info = &iter_seq_info,
1955};
1956
1957static int htab_lru_percpu_map_btf_id;
1958const struct bpf_map_ops htab_lru_percpu_map_ops = {
1959	.map_alloc_check = htab_map_alloc_check,
1960	.map_alloc = htab_map_alloc,
1961	.map_free = htab_map_free,
1962	.map_get_next_key = htab_map_get_next_key,
1963	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1964	.map_update_elem = htab_lru_percpu_map_update_elem,
1965	.map_delete_elem = htab_lru_map_delete_elem,
1966	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1967	BATCH_OPS(htab_lru_percpu),
1968	.map_btf_name = "bpf_htab",
1969	.map_btf_id = &htab_lru_percpu_map_btf_id,
1970	.iter_seq_info = &iter_seq_info,
1971};
1972
1973static int fd_htab_map_alloc_check(union bpf_attr *attr)
1974{
1975	if (attr->value_size != sizeof(u32))
1976		return -EINVAL;
1977	return htab_map_alloc_check(attr);
1978}
1979
1980static void fd_htab_map_free(struct bpf_map *map)
1981{
1982	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1983	struct hlist_nulls_node *n;
1984	struct hlist_nulls_head *head;
1985	struct htab_elem *l;
1986	int i;
1987
1988	for (i = 0; i < htab->n_buckets; i++) {
1989		head = select_bucket(htab, i);
1990
1991		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1992			void *ptr = fd_htab_map_get_ptr(map, l);
1993
1994			map->ops->map_fd_put_ptr(ptr);
1995		}
1996	}
1997
1998	htab_map_free(map);
1999}
2000
2001/* only called from syscall */
2002int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2003{
2004	void **ptr;
2005	int ret = 0;
2006
2007	if (!map->ops->map_fd_sys_lookup_elem)
2008		return -ENOTSUPP;
2009
2010	rcu_read_lock();
2011	ptr = htab_map_lookup_elem(map, key);
2012	if (ptr)
2013		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2014	else
2015		ret = -ENOENT;
2016	rcu_read_unlock();
2017
2018	return ret;
2019}
2020
2021/* only called from syscall */
2022int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2023				void *key, void *value, u64 map_flags)
2024{
2025	void *ptr;
2026	int ret;
2027	u32 ufd = *(u32 *)value;
2028
2029	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2030	if (IS_ERR(ptr))
2031		return PTR_ERR(ptr);
2032
2033	ret = htab_map_update_elem(map, key, &ptr, map_flags);
2034	if (ret)
2035		map->ops->map_fd_put_ptr(ptr);
2036
2037	return ret;
2038}
2039
2040static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2041{
2042	struct bpf_map *map, *inner_map_meta;
2043
2044	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2045	if (IS_ERR(inner_map_meta))
2046		return inner_map_meta;
2047
2048	map = htab_map_alloc(attr);
2049	if (IS_ERR(map)) {
2050		bpf_map_meta_free(inner_map_meta);
2051		return map;
2052	}
2053
2054	map->inner_map_meta = inner_map_meta;
2055
2056	return map;
2057}
2058
2059static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2060{
2061	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2062
2063	if (!inner_map)
2064		return NULL;
2065
2066	return READ_ONCE(*inner_map);
2067}
2068
2069static u32 htab_of_map_gen_lookup(struct bpf_map *map,
2070				  struct bpf_insn *insn_buf)
2071{
2072	struct bpf_insn *insn = insn_buf;
2073	const int ret = BPF_REG_0;
2074
2075	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2076		     (void *(*)(struct bpf_map *map, void *key))NULL));
2077	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2078	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2079	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2080				offsetof(struct htab_elem, key) +
2081				round_up(map->key_size, 8));
2082	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2083
2084	return insn - insn_buf;
2085}
2086
2087static void htab_of_map_free(struct bpf_map *map)
2088{
2089	bpf_map_meta_free(map->inner_map_meta);
2090	fd_htab_map_free(map);
2091}
2092
2093static int htab_of_maps_map_btf_id;
2094const struct bpf_map_ops htab_of_maps_map_ops = {
2095	.map_alloc_check = fd_htab_map_alloc_check,
2096	.map_alloc = htab_of_map_alloc,
2097	.map_free = htab_of_map_free,
2098	.map_get_next_key = htab_map_get_next_key,
2099	.map_lookup_elem = htab_of_map_lookup_elem,
2100	.map_delete_elem = htab_map_delete_elem,
2101	.map_fd_get_ptr = bpf_map_fd_get_ptr,
2102	.map_fd_put_ptr = bpf_map_fd_put_ptr,
2103	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2104	.map_gen_lookup = htab_of_map_gen_lookup,
2105	.map_check_btf = map_check_no_btf,
2106	.map_btf_name = "bpf_htab",
2107	.map_btf_id = &htab_of_maps_map_btf_id,
2108};